VM forking
PATCH v2 10/20 x86/mem_sharing: Make add_to_physmap static and shorten name
In reply to Tamas K LengyelIt's not being called from outside mem_sharing.c
--- xen/arch/x86/mm/mem_sharing.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
xen/arch/x86/mm/mem_sharing.c +4 -3
@@ -1064,8 +1064,9 @@ err_out:return ret; } -int mem_sharing_add_to_physmap (struct domain *sd,unsigned long sgfn,shr_handle_t sh, -struct domain *cd,unsigned long cgfn,bool lock) +static +int add_to_physmap (struct domain *sd,unsigned long sgfn,shr_handle_t sh, +struct domain *cd,unsigned long cgfn,bool lock) {struct page_info *spage;int ret = -EINVAL ;
@@ -1558,7 +1559,7 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg) sh = mso.u.share.source_handle; cgfn = mso.u.share.client_gfn; - rc =mem_sharing_add_to_physmap (d, sgfn, sh, cd, cgfn, true); + rc =add_to_physmap (d, sgfn, sh, cd, cgfn, true);rcu_unlock_domain (cd); }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
v2 cover VM forking
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.
At runtime the forked VM starts running with an empty p2m which gets lazily populated when the VM generates EPT faults, similar to how altp2m views are populated. If the memory access is a read-only access, the p2m entry is populated with a memory shared entry with its parent. For write memory accesses or in case memory sharing wasn't possible (for example in case a reference is held by a third party), a new page is allocated and the page contents are copied over from the parent VM. Forks can be further forked if needed, thus allowing for further memory savings.
A VM fork reset hypercall is also added that allows the fork to be reset to the state it was just after a fork. This is an optimization for cases where the forks are very short-lived and run without a device model, so resetting saves some time compared to creating a brand new fork.
The series has been tested with both Linux and Windows VMs and functions as expected. VM forking time has been measured to be 0.018s, device model launch to be around 1s depending largely on the number of devices being emulated.
Patches 1-2 implement changes to existing internal Xen APIs to make VM forking possible.
Patches 3-4 are simple code-formatting fixes for the toolstack and Xen for the memory sharing paths with no functional changes.
Patches 5-16 are code-cleanups and adjustments of to Xen memory sharing subsystem with no functional changes.
Patch 17 adds the hypervisor-side code implementing VM forking.
Patch 18 is integration of mem_access with forked VMs.
Patch 19 implements the VM fork reset operation hypervisor side bits.
Patch 20 adds the toolstack-side code implementing VM forking and reset.
Tamas K Lengyel (20):
x86: make hvm_{get/set}_param accessible
xen/x86: Make hap_get_allocation accessible tools/libxc: clean up memory sharing files x86/mem_sharing: cleanup code and comments in various locations x86/mem_sharing: make get_two_gfns take locks conditionally x86/mem_sharing: drop flags from mem_sharing_unshare_page x86/mem_sharing: don't try to unshare twice during page fault x86/mem_sharing: define mem_sharing_domain to hold some scattered
variables
x86/mem_sharing: Use INVALID_MFN and p2m_is_shared in
relinquish_shared_pages
x86/mem_sharing: Make add_to_physmap static and shorten name x86/mem_sharing: Convert MEM_SHARING_DESTROY_GFN to a bool x86/mem_sharing: Replace MEM_SHARING_DEBUG with gdprintk x86/mem_sharing: ASSERT that p2m_set_entry succeeds x86/mem_sharing: Enable mem_sharing on first memop x86/mem_sharing: Skip xen heap pages in memshr nominate x86/mem_sharing: check page type count earlier xen/mem_sharing: VM forking
xen/mem_access: Use __get_gfn_type_access in set_mem_access x86/mem_sharing: reset a fork
xen/tools: VM forking toolstack side
tools/libxc/include/xenctrl.h | 30 +- tools/libxc/xc_memshr.c | 34 +- tools/libxl/libxl.h | 7 + tools/libxl/libxl_create.c | 237 +++++--- tools/libxl/libxl_dm.c | 2 +- tools/libxl/libxl_dom.c | 83 ++- tools/libxl/libxl_internal.h | 1 + tools/libxl/libxl_types.idl | 1 + tools/xl/xl.h | 5 + tools/xl/xl_cmdtable.c | 22 + tools/xl/xl_saverestore.c | 96 ++++ tools/xl/xl_vmcontrol.c | 8 + xen/arch/x86/hvm/hvm.c | 206 ++++--- xen/arch/x86/mm/hap/hap.c | 3 +- xen/arch/x86/mm/mem_access.c | 5 +- xen/arch/x86/mm/mem_sharing.c | 875 +++++++++++++++++++++--------- xen/arch/x86/mm/p2m.c | 34 +- xen/common/memory.c | 2 +- xen/drivers/passthrough/pci.c | 3 +- xen/include/asm-x86/hap.h | 1 + xen/include/asm-x86/hvm/domain.h | 6 +- xen/include/asm-x86/hvm/hvm.h | 4 + xen/include/asm-x86/mem_sharing.h | 84 ++- xen/include/asm-x86/p2m.h | 14 +- xen/include/public/memory.h | 6 + xen/include/xen/sched.h | 1 + 26 files changed, 1258 insertions(+), 512 deletions(-)
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 3/20 tools/libxc: clean up memory sharing files
No functional changes.
--- tools/libxc/include/xenctrl.h | 24 ++++++++++++------------ tools/libxc/xc_memshr.c | 12 ++++++------ 2 files changed, 18 insertions(+), 18 deletions(-)
tools/libxc/include/xenctrl.h +12 -12
@@ -2060,7 +2060,7 @@ int xc_monitor_emulate_each_rep(xc_interface *xch, uint32_t domain_id,* * Sharing is supported only on the x86 architecture in 64 bit mode, with * Hardware-Assisted Paging (i.e. Intel EPT, AMD NPT). Moreover, AMD NPT -* support is considered experimental. +* support is considered experimental. * Calls below return ENOSYS if not in the x86_64 architecture. * Calls below return ENODEV if the domain does not support HAP.
@@ -2107,13 +2107,13 @@ int xc_memshr_control(xc_interface *xch,* EINVAL or EACCESS if the request is denied by the security policy */ -int xc_memshr_ring_enable (xc_interface *xch, +int xc_memshr_ring_enable (xc_interface *xch,uint32_t domid,uint32_t *port);/* Disable the ring for ENOMEM communication. * May fail with EINVAL if the ring was not enabled in the first place. */ -int xc_memshr_ring_disable (xc_interface *xch, +int xc_memshr_ring_disable (xc_interface *xch,uint32_t domid);/*
@@ -2126,7 +2126,7 @@ int xc_memshr_ring_disable(xc_interface *xch,int xc_memshr_domain_resume (xc_interface *xch,uint32_t domid); -/* Select a page for sharing. +/* Select a page for sharing. * * A 64 bit opaque handle will be stored in handle. The hypervisor ensures * that if the page is modified, the handle will be invalidated, and future
@@ -2155,7 +2155,7 @@ int xc_memshr_nominate_gref(xc_interface *xch,/* The three calls below may fail with * 10 (or -XENMEM_SHARING_OP_S_HANDLE_INVALID) if the handle passed as source -* is invalid. +* is invalid. * 9 (or -XENMEM_SHARING_OP_C_HANDLE_INVALID) if the handle passed as client is * invalid. */
@@ -2168,7 +2168,7 @@ int xc_memshr_nominate_gref(xc_interface *xch,* * After successful sharing, the client handle becomes invalid. Both <domain, * gfn> tuples point to the same mfn with the same handle, the one specified as -* source. Either 3-tuple can be specified later for further re-sharing. +* source. Either 3-tuple can be specified later for further re-sharing. */ int xc_memshr_share_gfns(xc_interface *xch, uint32_t source_domain,
@@ -2193,7 +2193,7 @@ int xc_memshr_share_grefs(xc_interface *xch,/* Allows to add to the guest physmap of the client domain a shared frame * directly. * -* May additionally fail with +* May additionally fail with * 9 (-XENMEM_SHARING_OP_C_HANDLE_INVALID) if the physmap entry for the gfn is * not suitable. * ENOMEM if internal data structures cannot be allocated.
@@ -2222,7 +2222,7 @@ int xc_memshr_range_share(xc_interface *xch, uint64_t last_gfn);/* Debug calls: return the number of pages referencing the shared frame backing -* the input argument. Should be one or greater. +* the input argument. Should be one or greater. * * May fail with EINVAL if there is no backing shared frame for the input * argument.
@@ -2235,9 +2235,9 @@ int xc_memshr_debug_gref(xc_interface *xch,uint32_t domid, grant_ref_t gref); -/* Audits the share subsystem. -* -* Returns ENOSYS if not supported (may not be compiled into the hypervisor). +/* Audits the share subsystem. +* +* Returns ENOSYS if not supported (may not be compiled into the hypervisor). * * Returns the number of errors found during auditing otherwise. May be (should * be!) zero.
@@ -2273,7 +2273,7 @@ long xc_sharing_freed_pages(xc_interface *xch);* should return 1. (And dominfo(d) for each of the two domains should return 1 * as well). * -* Note that some of these sharing_used_frames may be referenced by +* Note that some of these sharing_used_frames may be referenced by * a single domain page, and thus not realize any savings. The same * applies to some of the pages counted in dominfo(d)->shr_pages. */
tools/libxc/xc_memshr.c +6 -6
@@ -41,7 +41,7 @@ int xc_memshr_control(xc_interface *xch,return do_domctl (xch, &domctl); } -int xc_memshr_ring_enable (xc_interface *xch, +int xc_memshr_ring_enable (xc_interface *xch,uint32_t domid,uint32_t *port) {
@@ -57,7 +57,7 @@ int xc_memshr_ring_enable(xc_interface *xch, port); } -int xc_memshr_ring_disable (xc_interface *xch, +int xc_memshr_ring_disable (xc_interface *xch,uint32_t domid) { return xc_vm_event_control(xch, domid,
@@ -85,11 +85,11 @@ int xc_memshr_nominate_gfn(xc_interface *xch,memset (&mso,0 ,sizeof (mso)); mso.op = XENMEM_sharing_op_nominate_gfn; - mso.u.nominate.u.gfn = gfn; + mso.u.nominate.u.gfn = gfn; rc =xc_memshr_memop (xch, domid, &mso); -if (!rc) *handle = mso.u.nominate.handle; +if (!rc) *handle = mso.u.nominate.handle;return rc; }
@@ -105,11 +105,11 @@ int xc_memshr_nominate_gref(xc_interface *xch,memset (&mso,0 ,sizeof (mso)); mso.op = XENMEM_sharing_op_nominate_gref; - mso.u.nominate.u.grant_ref = gref; + mso.u.nominate.u.grant_ref = gref; rc =xc_memshr_memop (xch, domid, &mso); -if (!rc) *handle = mso.u.nominate.handle; +if (!rc) *handle = mso.u.nominate.handle;return rc; }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 1/20 x86: make hvm_{get/set}_param accessible
In reply to Tamas K LengyelCurrently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
--- xen/arch/x86/hvm/hvm.c | 169 ++++++++++++++++++++-------------- xen/include/asm-x86/hvm/hvm.h | 4 + 2 files changed, 106 insertions(+), 67 deletions(-)
xen/arch/x86/hvm/hvm.c +102 -67
@@ -4072,16 +4072,17 @@ static int hvmop_set_evtchn_upcall_vector( }static int hvm_allow_set_param (struct domain *d, -const struct xen_hvm_param *a) +uint32_t index, +uint64_t new_value) { -uint64_t value = d->arch.hvm.params[a->index]; +uint64_t value = d->arch.hvm.params[index];int rc; rc =xsm_hvm_param (XSM_TARGET , d, HVMOP_set_param);if ( rc )return rc; -switch ( a->index ) +switch ( index ) {/* The following parameters can be set by the guest. */ case HVM_PARAM_CALLBACK_IRQ :
@@ -4114,7 +4115,7 @@ static int hvm_allow_set_param(struct domain *d,if ( rc )return rc; -switch ( a->index ) +switch ( index ) {/* The following parameters should only be changed once. */ case HVM_PARAM_VIRIDIAN :
@@ -4124,7 +4125,7 @@ static int hvm_allow_set_param(struct domain *d,case HVM_PARAM_NR_IOREQ_SERVER_PAGES :case HVM_PARAM_ALTP2M :case HVM_PARAM_MCA_CAP : -if ( value !=0 && a->value != value ) +if ( value !=0 && new_value != value ) rc = -EEXIST ;break ;default :
@@ -4134,13 +4135,11 @@ static int hvm_allow_set_param(struct domain *d,return rc; } -static int hvmop_set_param ( +int hvmop_set_param (XEN_GUEST_HANDLE_PARAM (xen_hvm_param_t) arg) { -struct domain *curr_d = current->domain;struct xen_hvm_param a;struct domain *d; -struct vcpu *v;int rc;if (copy_from_guest (&a, arg,1 ) )
@@ -4160,23 +4159,42 @@ static int hvmop_set_param(if ( !is_hvm_domain (d) ) goto out; - rc =hvm_allow_set_param (d, &a); + rc =hvm_set_param (d, a.index, a.value); + + out: +rcu_unlock_domain (d); +return rc; +} + +int hvm_set_param ( +struct domain *d, +uint32_t index, +uint64_t value) +{ +struct domain *curr_d = current->domain; +int rc; +struct vcpu *v; + +if ( index >=HVM_NR_PARAMS ) +return -EINVAL ; + + rc =hvm_allow_set_param (d, index, value);if ( rc ) goto out; -switch ( a.index ) +switch ( index ) {case HVM_PARAM_CALLBACK_IRQ : -hvm_set_callback_via (d, a.value); +hvm_set_callback_via (d, value);hvm_latch_shinfo_size (d);break ;case HVM_PARAM_TIMER_MODE : -if ( a.value > HVMPTM_one_missed_tick_pending ) +if ( value > HVMPTM_one_missed_tick_pending ) rc = -EINVAL ;break ;case HVM_PARAM_VIRIDIAN : -if ( (a.value & ~HVMPV_feature_mask) || - !(a.value & HVMPV_base_freq) ) +if ( (value & ~HVMPV_feature_mask) || + !(value & HVMPV_base_freq) ) rc = -EINVAL ;break ;case HVM_PARAM_IDENT_PT :
@@ -4186,7 +4204,7 @@ static int hvmop_set_param(*/ if ( !paging_mode_hap (d) || !cpu_has_vmx ) { - d->arch.hvm.params[a.index] = a.value; + d->arch.hvm.params[index] = value;break ; }
@@ -4201,7 +4219,7 @@ static int hvmop_set_param( rc =0 ;domain_pause (d); - d->arch.hvm.params[a.index] = a.value; + d->arch.hvm.params[index] = value; for_each_vcpu ( d, v )paging_update_cr3 (v, false);domain_unpause (d);
@@ -4210,23 +4228,23 @@ static int hvmop_set_param(break ;case HVM_PARAM_DM_DOMAIN :/* The only value this should ever be set to is DOMID_SELF */ -if ( a.value !=DOMID_SELF ) +if ( value !=DOMID_SELF ) rc = -EINVAL ; - a.value = curr_d->domain_id; + value = curr_d->domain_id;break ;case HVM_PARAM_ACPI_S_STATE : rc =0 ; -if ( a.value ==3 ) +if ( value ==3 )hvm_s3_suspend (d); -else if ( a.value ==0 ) +else if ( value ==0 )hvm_s3_resume (d);else rc = -EINVAL ;break ;case HVM_PARAM_ACPI_IOPORTS_LOCATION : - rc =pmtimer_change_ioport (d, a.value); + rc =pmtimer_change_ioport (d, value);break ;case HVM_PARAM_MEMORY_EVENT_CR0 :case HVM_PARAM_MEMORY_EVENT_CR3 :
@@ -4241,24 +4259,24 @@ static int hvmop_set_param( rc =xsm_hvm_param_nested (XSM_PRIV , d);if ( rc )break ; -if ( a.value >1 ) +if ( value >1 ) rc = -EINVAL ;/* * Remove the check below once we have * shadow-on-shadow. */ -if ( !paging_mode_hap (d) && a.value ) +if ( !paging_mode_hap (d) && value ) rc = -EINVAL ; -if ( a.value && +if ( value && d->arch.hvm.params[HVM_PARAM_ALTP2M ] ) rc = -EINVAL ;/* Set up NHVM state for any vcpus that are already up. */ -if ( a.value && +if ( value && !d->arch.hvm.params[HVM_PARAM_NESTEDHVM ] )for_each_vcpu (d, v)if ( rc ==0 ) rc =nestedhvm_vcpu_initialise (v); -if ( !a.value || rc ) +if ( !value || rc )for_each_vcpu (d, v)nestedhvm_vcpu_destroy (v);break ;
@@ -4266,30 +4284,30 @@ static int hvmop_set_param( rc =xsm_hvm_param_altp2mhvm (XSM_PRIV , d);if ( rc )break ; -if ( a.value > XEN_ALTP2M_limited ) +if ( value > XEN_ALTP2M_limited ) rc = -EINVAL ; -if ( a.value && +if ( value && d->arch.hvm.params[HVM_PARAM_NESTEDHVM ] ) rc = -EINVAL ;break ;case HVM_PARAM_TRIPLE_FAULT_REASON : -if ( a.value >SHUTDOWN_MAX ) +if ( value >SHUTDOWN_MAX ) rc = -EINVAL ;break ;case HVM_PARAM_IOREQ_SERVER_PFN : - d->arch.hvm.ioreq_gfn.base = a.value; + d->arch.hvm.ioreq_gfn.base = value;break ;case HVM_PARAM_NR_IOREQ_SERVER_PAGES : {unsigned int i; -if ( a.value ==0 || - a.value >sizeof (d->arch.hvm.ioreq_gfn.mask) *8 ) +if ( value ==0 || + value >sizeof (d->arch.hvm.ioreq_gfn.mask) *8 ) { rc = -EINVAL ;break ; } -for ( i =0 ; i < a.value; i++ ) +for ( i =0 ; i < value; i++ )set_bit (i, &d->arch.hvm.ioreq_gfn.mask);break ;
@@ -4301,35 +4319,35 @@ static int hvmop_set_param( sizeof(d->arch.hvm.ioreq_gfn.legacy_mask) * 8);BUILD_BUG_ON (HVM_PARAM_BUFIOREQ_PFN >sizeof (d->arch.hvm.ioreq_gfn.legacy_mask) *8 ); -if ( a.value ) -set_bit (a.index, &d->arch.hvm.ioreq_gfn.legacy_mask); +if ( value ) +set_bit (index, &d->arch.hvm.ioreq_gfn.legacy_mask);break ;case HVM_PARAM_X87_FIP_WIDTH : -if ( a.value !=0 && a.value !=4 && a.value !=8 ) +if ( value !=0 && value !=4 && value !=8 ) { rc = -EINVAL ;break ; } - d->arch.x87_fip_width = a.value; + d->arch.x87_fip_width = value;break ;case HVM_PARAM_VM86_TSS :/* Hardware would silently truncate high bits. */ -if ( a.value != (uint32_t )a.value ) +if ( value != (uint32_t )value ) {if ( d == curr_d )domain_crash (d); rc = -EINVAL ; }/* Old hvmloader binaries hardcode the size to 128 bytes. */ -if ( a.value ) - a.value |= (128ULL <<32 ) |VM86_TSS_UPDATED ; - a.index =HVM_PARAM_VM86_TSS_SIZED ; +if ( value ) + value |= (128ULL <<32 ) |VM86_TSS_UPDATED ; + index =HVM_PARAM_VM86_TSS_SIZED ;break ;case HVM_PARAM_VM86_TSS_SIZED : -if ( (a.value >>32 ) <sizeof (struct tss32 ) ) +if ( (value >>32 ) <sizeof (struct tss32 ) ) {if ( d == curr_d )domain_crash (d);
@@ -4340,34 +4358,33 @@ static int hvmop_set_param(* 256 bits interrupt redirection bitmap + 64k bits I/O bitmap * plus one padding byte). */ -if ( (a.value >>32 ) >sizeof (struct tss32 ) + +if ( (value >>32 ) >sizeof (struct tss32 ) + (0x100 /8 ) + (0x10000 /8 ) +1 ) - a.value = (uint32_t )a.value | + value = (uint32_t )value | ((sizeof (struct tss32 ) + (0x100 /8 ) + (0x10000 /8 ) +1 ) <<32 ); - a.value |=VM86_TSS_UPDATED ; + value |=VM86_TSS_UPDATED ;break ;case HVM_PARAM_MCA_CAP : - rc =vmce_enable_mca_cap (d, a.value); + rc =vmce_enable_mca_cap (d, value);break ; }if ( rc !=0 ) goto out; - d->arch.hvm.params[a.index] = a.value; + d->arch.hvm.params[index] = value;HVM_DBG_LOG (DBG_LEVEL_HCALL ,"set param %u = %" PRIx64, - a.index, a.value); + index, value); out: -rcu_unlock_domain (d);return rc; }static int hvm_allow_get_param (struct domain *d, -const struct xen_hvm_param *a) +uint32_t index) {int rc;
@@ -4375,7 +4392,7 @@ static int hvm_allow_get_param(struct domain *d,if ( rc )return rc; -switch ( a->index ) +switch ( index ) {/* The following parameters can be read by the guest. */ case HVM_PARAM_CALLBACK_IRQ :
@@ -4429,42 +4446,60 @@ static int hvmop_get_param(if ( !is_hvm_domain (d) ) goto out; - rc =hvm_allow_get_param (d, &a); + rc =hvm_get_param (d, a.index, &a.value);if ( rc ) goto out; -switch ( a.index ) + rc =__copy_to_guest (arg, &a,1 ) ? -EFAULT :0 ; + +HVM_DBG_LOG (DBG_LEVEL_HCALL ,"get param %u = %" PRIx64, + a.index, a.value); + + out: +rcu_unlock_domain (d); +return rc; +} + +int hvm_get_param ( +struct domain *d, +uint32_t index, +uint64_t *value) +{ +int rc; + +if ( index >=HVM_NR_PARAMS || !value ) +return -EINVAL ; + + rc =hvm_allow_get_param (d, index); +if ( rc ) +return rc; + +switch ( index ) {case HVM_PARAM_ACPI_S_STATE : - a.value = d->arch.hvm.is_s3_suspended ?3 :0 ; + *value = d->arch.hvm.is_s3_suspended ?3 :0 ;break ;case HVM_PARAM_VM86_TSS : - a.value = (uint32_t )d->arch.hvm.params[HVM_PARAM_VM86_TSS_SIZED ]; + *value = (uint32_t )d->arch.hvm.params[HVM_PARAM_VM86_TSS_SIZED ];break ;case HVM_PARAM_VM86_TSS_SIZED : - a.value = d->arch.hvm.params[HVM_PARAM_VM86_TSS_SIZED ] & - ~VM86_TSS_UPDATED ; + *value = d->arch.hvm.params[HVM_PARAM_VM86_TSS_SIZED ] & + ~VM86_TSS_UPDATED ;break ;case HVM_PARAM_X87_FIP_WIDTH : - a.value = d->arch.x87_fip_width; + *value = d->arch.x87_fip_width;break ;default : - a.value = d->arch.hvm.params[a.index]; + *value = d->arch.hvm.params[index];break ; } - rc =__copy_to_guest (arg, &a,1 ) ? -EFAULT :0 ; - -HVM_DBG_LOG (DBG_LEVEL_HCALL ,"get param %u = %" PRIx64, - a.index, a.value); +return 0 ; +}; - out: -rcu_unlock_domain (d); -return rc; -}/* * altp2m operations are envisioned as being used in several different
xen/include/asm-x86/hvm/hvm.h +4 -0
@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */ +int hvm_get_param (struct domain *d,uint32_t index,uint64_t *value); +int hvm_set_param (struct domain *d,uint32_t index,uint64_t value); +#ifdef CONFIG_HVM #define hvm_get_guest_tsc (v) hvm_get_guest_tsc_fixed(v, 0)
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 2/20 xen/x86: Make hap_get_allocation accessible
In reply to Tamas K LengyelDuring VM forking we'll copy the parent domain's parameters to the client, including the HAP shadow memory setting that is used for storing the domain's EPT. We'll copy this in the hypervisor instead doing it during toolstack launch to allow the domain to start executing and unsharing memory before (or even completely without) the toolstack.
--- xen/arch/x86/mm/hap/hap.c | 3 +-- xen/include/asm-x86/hap.h | 1 + 2 files changed, 2 insertions(+), 2 deletions(-)
xen/arch/x86/mm/hap/hap.c +1 -2
@@ -321,8 +321,7 @@ static void hap_free_p2m_page(struct domain *d, struct page_info *pg) }/* Return the size of the pool, rounded up to the nearest MB */ -static unsigned int -hap_get_allocation (struct domain *d) +unsigned int hap_get_allocation (struct domain *d) {unsigned int pg = d->arch.paging.hap.total_pages + d->arch.paging.hap.p2m_pages;
xen/include/asm-x86/hap.h +1 -0
@@ -45,6 +45,7 @@ int hap_track_dirty_vram(struct domain *d,extern const struct paging_mode *hap_paging_get_mode (struct vcpu *);int hap_set_allocation (struct domain *d,unsigned int pages,bool *preempted); +unsigned int hap_get_allocation (struct domain *d);#endif /* XEN_HAP_H */
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 4/20 x86/mem_sharing: cleanup code and comments in various locations
In reply to Tamas K LengyelNo functional changes.
--- xen/arch/x86/hvm/hvm.c | 11 +- xen/arch/x86/mm/mem_sharing.c | 342 +++++++++++++++++------------- xen/arch/x86/mm/p2m.c | 17 +- xen/include/asm-x86/mem_sharing.h | 51 +++-- 4 files changed, 236 insertions(+), 185 deletions(-)
xen/arch/x86/hvm/hvm.c +6 -5
@@ -1902,12 +1902,11 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,if ( npfec.write_access && (p2mt == p2m_ram_shared) ) {ASSERT (p2m_is_hostp2m (p2m)); - sharing_enomem = - (mem_sharing_unshare_page (currd, gfn,0 ) <0 ); + sharing_enomem =mem_sharing_unshare_page (currd, gfn,0 ); rc =1 ; goto out_put_gfn; } - +/* Spurious fault? PoD and log-dirty also take this path. */ if (p2m_is_ram (p2mt) ) {
@@ -1953,9 +1952,11 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,__put_gfn (p2m, gfn);__put_gfn (hostp2m, gfn); out: -/* All of these are delayed until we exit, since we might +/* +* All of these are delayed until we exit, since we might * sleep on event ring wait queues, and we must not hold -* locks in such circumstance */ +* locks in such circumstance. +*/ if ( paged )p2m_mem_paging_populate (currd, gfn);if ( sharing_enomem )
xen/arch/x86/mm/mem_sharing.c +191 -151
763 lines not on this page · raw message
xen/arch/x86/mm/p2m.c +10 -7
@@ -511,8 +511,10 @@ mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l,if ( (q &P2M_UNSHARE ) &&p2m_is_shared (*t) ) {ASSERT (p2m_is_hostp2m (p2m)); -/* Try to unshare. If we fail, communicate ENOMEM without -* sleeping. */ +/* +* Try to unshare. If we fail, communicate ENOMEM without +* sleeping. +*/ if (mem_sharing_unshare_page (p2m->domain, gfn_l,0 ) <0 )mem_sharing_notify_enomem (p2m->domain, gfn_l, false); mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL );
@@ -892,15 +894,15 @@ guest_physmap_add_entry(struct domain *d, gfn_t gfn, mfn_t mfn, &a, 0, NULL, NULL);if (p2m_is_shared (ot) ) { -/* Do an unshare to cleanly take care of all corner -* cases. */ +/* Do an unshare to cleanly take care of all corner cases. */ int rc; rc =mem_sharing_unshare_page (p2m->domain,gfn_x (gfn_add (gfn, i)),0 );if ( rc ) {p2m_unlock (p2m); -/* NOTE: Should a guest domain bring this upon itself, +/* +* NOTE: Should a guest domain bring this upon itself, * there is not a whole lot we can do. We are buried * deep in locks from most code paths by now. So, fail * the call and don't try to sleep on a wait queue
@@ -909,8 +911,9 @@ guest_physmap_add_entry(struct domain *d, gfn_t gfn, mfn_t mfn,* However, all current (changeset 3432abcf9380) code * paths avoid this unsavoury situation. For now. * -* Foreign domains are okay to place an event as they -* won't go to sleep. */ +* Foreign domains are okay to place an event as they +* won't go to sleep. +*/ (void )mem_sharing_notify_enomem (p2m->domain,gfn_x (gfn_add (gfn, i)), false);return rc;
xen/include/asm-x86/mem_sharing.h +29 -22
103 lines not on this page · raw message
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 5/20 x86/mem_sharing: make get_two_gfns take locks conditionally
In reply to Tamas K LengyelDuring VM forking the client lock will already be taken.
--- xen/arch/x86/mm/mem_sharing.c | 11 ++++++----- xen/include/asm-x86/p2m.h | 10 +++++----- 2 files changed, 11 insertions(+), 10 deletions(-)
xen/arch/x86/mm/mem_sharing.c +6 -5
@@ -954,7 +954,7 @@ static int share_pages(struct domain *sd, gfn_t sgfn, shr_handle_t sh,unsigned long put_count =0 ;get_two_gfns (sd, sgfn, &smfn_type,NULL , &smfn, - cd, cgfn, &cmfn_type,NULL , &cmfn,0 , &tg); + cd, cgfn, &cmfn_type,NULL , &cmfn,0 , &tg, true);/* * This tricky business is to avoid two callers deadlocking if
@@ -1068,7 +1068,7 @@ err_out: }int mem_sharing_add_to_physmap (struct domain *sd,unsigned long sgfn,shr_handle_t sh, -struct domain *cd,unsigned long cgfn) +struct domain *cd,unsigned long cgfn,bool lock) {struct page_info *spage;int ret = -EINVAL ;
@@ -1080,7 +1080,7 @@ int mem_sharing_add_to_physmap(struct domain *sd, unsigned long sgfn, shr_handlestruct two_gfns tg;get_two_gfns (sd,_gfn (sgfn), &smfn_type,NULL , &smfn, - cd,_gfn (cgfn), &cmfn_type, &a, &cmfn,0 , &tg); + cd,_gfn (cgfn), &cmfn_type, &a, &cmfn,0 , &tg, lock);/* Get the source shared page, check and lock */ ret =XENMEM_SHARING_OP_S_HANDLE_INVALID ;
@@ -1155,7 +1155,8 @@ int mem_sharing_add_to_physmap(struct domain *sd, unsigned long sgfn, shr_handle err_unlock:mem_sharing_page_unlock (spage); err_out: -put_two_gfns (&tg); +if ( lock ) +put_two_gfns (&tg);return ret; }
@@ -1574,7 +1575,7 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg) sh = mso.u.share.source_handle; cgfn = mso.u.share.client_gfn; - rc =mem_sharing_add_to_physmap (d, sgfn, sh, cd, cgfn); + rc =mem_sharing_add_to_physmap (d, sgfn, sh, cd, cgfn, true);rcu_unlock_domain (cd); }
xen/include/asm-x86/p2m.h +5 -5
@@ -539,7 +539,7 @@ struct two_gfns {static inline void get_two_gfns (struct domain *rd,gfn_t rgfn,p2m_type_t *rt,p2m_access_t *ra,mfn_t *rmfn,struct domain *ld,gfn_t lgfn,p2m_type_t *lt,p2m_access_t *la,mfn_t *lmfn, -p2m_query_t q,struct two_gfns *rval) +p2m_query_t q,struct two_gfns *rval,bool lock) {mfn_t *first_mfn, *second_mfn, scratch_mfn;p2m_access_t *first_a, *second_a, scratch_a;
@@ -569,10 +569,10 @@ do { \#undef assign_pointers/* Now do the gets */ - *first_mfn =get_gfn_type_access (p2m_get_hostp2m (rval->first_domain), -gfn_x (rval->first_gfn), first_t, first_a, q,NULL ); - *second_mfn =get_gfn_type_access (p2m_get_hostp2m (rval->second_domain), -gfn_x (rval->second_gfn), second_t, second_a, q,NULL ); + *first_mfn =__get_gfn_type_access (p2m_get_hostp2m (rval->first_domain), +gfn_x (rval->first_gfn), first_t, first_a, q,NULL , lock); + *second_mfn =__get_gfn_type_access (p2m_get_hostp2m (rval->second_domain), +gfn_x (rval->second_gfn), second_t, second_a, q,NULL , lock); }static inline void put_two_gfns (struct two_gfns *arg)
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 16/20 x86/mem_sharing: check page type count earlier
In reply to Tamas K Lengyel--- xen/arch/x86/mm/mem_sharing.c | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-)
xen/arch/x86/mm/mem_sharing.c +6 -7
@@ -649,19 +649,18 @@ static int page_make_sharable(struct domain *d,return -EBUSY ; } -/* Change page type and count atomically */ -if ( !get_page_and_type (page, d, PGT_shared_page) ) +/* Check if page is already typed and bail early if it is */ +if ( (page->u.inuse.type_info & PGT_count_mask) !=1 ) {spin_unlock (&d->page_alloc_lock); -return -EINVAL ; +return -EEXIST ; } -/* Check it wasn't already sharable and undo if it was */ -if ( (page->u.inuse.type_info & PGT_count_mask) !=1 ) +/* Change page type and count atomically */ +if ( !get_page_and_type (page, d, PGT_shared_page) ) {spin_unlock (&d->page_alloc_lock); -put_page_and_type (page); -return -EEXIST ; +return -EINVAL ; }/*
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 17/20 xen/mem_sharing: VM forking
In reply to Tamas K LengyelVM forking is the process of creating a domain with an empty memory space and a parent domain specified from which to populate the memory when necessary. For the new domain to be functional the VM state is copied over as part of the fork operation (HVM params, hap allocation, etc).
--- xen/arch/x86/hvm/hvm.c | 2 +- xen/arch/x86/mm/mem_sharing.c | 228 ++++++++++++++++++++++++++++++ xen/arch/x86/mm/p2m.c | 11 +- xen/include/asm-x86/mem_sharing.h | 20 ++- xen/include/public/memory.h | 5 + xen/include/xen/sched.h | 1 + 6 files changed, 263 insertions(+), 4 deletions(-)
xen/arch/x86/hvm/hvm.c +1 -1
@@ -1913,7 +1913,7 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla, }#endif -/* Spurious fault? PoD and log-dirty also take this path. */ +/* Spurious fault? PoD, log-dirty and VM forking also take this path. */ if (p2m_is_ram (p2mt) ) { rc =1 ;
xen/arch/x86/mm/mem_sharing.c +228 -0
257 lines not on this page · raw message
xen/arch/x86/mm/p2m.c +10 -1
@@ -508,6 +508,14 @@ mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l, mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL ); +/* Check if we need to fork the page */ +if ( (q &P2M_ALLOC ) &&p2m_is_hole (*t) && + !mem_sharing_fork_page (p2m->domain, gfn, !!(q &P2M_UNSHARE )) ) + { + mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL ); + } + +/* Check if we need to unshare the page */ if ( (q &P2M_UNSHARE ) &&p2m_is_shared (*t) ) {ASSERT (p2m_is_hostp2m (p2m));
@@ -586,7 +594,8 @@ struct page_info *p2m_get_page_from_gfn(return page;/* Error path: not a suitable GFN at all */ -if ( !p2m_is_ram (*t) && !p2m_is_paging (*t) && !p2m_is_pod (*t) ) +if ( !p2m_is_ram (*t) && !p2m_is_paging (*t) && !p2m_is_pod (*t) && + !mem_sharing_is_fork (p2m->domain) )return NULL ; }
xen/include/asm-x86/mem_sharing.h +18 -2
@@ -26,8 +26,7 @@ #ifdef CONFIG_MEM_SHARING -struct mem_sharing_domain -{ +struct mem_sharing_domain { bool enabled; /*
@@ -40,6 +39,9 @@ struct mem_sharing_domain#define mem_sharing_enabled (d) \ (hap_enabled(d) && (d)->arch.hvm.mem_sharing.enabled) +#define mem_sharing_is_fork (d) \ + (mem_sharing_enabled(d) && !!((d)->parent)) +/* Auditing of memory sharing code? */ #ifndef NDEBUG #define MEM_SHARING_AUDIT 1
@@ -90,6 +92,9 @@ int mem_sharing_unshare_page(struct domain *d,return rc; } +int mem_sharing_fork_page (struct domain *d,gfn_t gfn, +bool unsharing); +/* * If called by a foreign domain, possible errors are * -EBUSY -> ring full
@@ -119,6 +124,7 @@ int relinquish_shared_pages(struct domain *d);#else #define mem_sharing_enabled (d) false +#define mem_sharing_is_fork (p2m) falsestatic inline unsigned int mem_sharing_get_nr_saved_mfns (void ) {
@@ -145,6 +151,16 @@ int mem_sharing_notify_enomem(struct domain *d, unsigned long gfn,return -EOPNOTSUPP ; } +static inline int mem_sharing_fork (struct domain *d,struct domain *cd,bool vcpu) +{ +return -EOPNOTSUPP ; +} + +static inline int mem_sharing_fork_page (struct domain *d,gfn_t gfn,bool lock) +{ +return -EOPNOTSUPP ; +} +#endif #endif /* __MEM_SHARING_H__ */
xen/include/public/memory.h +5 -0
@@ -482,6 +482,7 @@ DEFINE_XEN_GUEST_HANDLE(xen_mem_access_op_t);#define XENMEM_sharing_op_add_physmap 6#define XENMEM_sharing_op_audit 7#define XENMEM_sharing_op_range_share 8 +#define XENMEM_sharing_op_fork 9#define XENMEM_SHARING_OP_S_HANDLE_INVALID (-10)#define XENMEM_SHARING_OP_C_HANDLE_INVALID (-9)
@@ -532,6 +533,10 @@ struct xen_mem_sharing_op {uint32_t gref;/* IN: gref to debug */ } u; } debug; +struct mem_sharing_op_fork { +domid_t parent_domain; +uint16_t _pad[3 ];/* Must be set to 0 */ + } fork; } u; };typedef struct xen_mem_sharing_op xen_mem_sharing_op_t ;
xen/include/xen/sched.h +1 -0
@@ -501,6 +501,7 @@ struct domain/* Memory sharing support */ #ifdef CONFIG_MEM_SHARING struct vm_event_domain *vm_event_share; +struct domain *parent;/* VM fork parent */ #endif /* Memory paging support */ #ifdef CONFIG_HAS_MEM_PAGING
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 6/20 x86/mem_sharing: drop flags from mem_sharing_unshare_page
In reply to Tamas K LengyelAll callers pass 0 in.
--- xen/arch/x86/hvm/hvm.c | 2 +- xen/arch/x86/mm/p2m.c | 5 ++--- xen/include/asm-x86/mem_sharing.h | 8 +++----- 3 files changed, 6 insertions(+), 9 deletions(-)
xen/arch/x86/hvm/hvm.c +1 -1
@@ -1902,7 +1902,7 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,if ( npfec.write_access && (p2mt == p2m_ram_shared) ) {ASSERT (p2m_is_hostp2m (p2m)); - sharing_enomem =mem_sharing_unshare_page (currd, gfn,0 ); + sharing_enomem =mem_sharing_unshare_page (currd, gfn); rc =1 ; goto out_put_gfn; }
xen/arch/x86/mm/p2m.c +2 -3
@@ -515,7 +515,7 @@ mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l,* Try to unshare. If we fail, communicate ENOMEM without * sleeping. */ -if (mem_sharing_unshare_page (p2m->domain, gfn_l,0 ) <0 ) +if (mem_sharing_unshare_page (p2m->domain, gfn_l) <0 )mem_sharing_notify_enomem (p2m->domain, gfn_l, false); mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL ); }
@@ -896,8 +896,7 @@ guest_physmap_add_entry(struct domain *d, gfn_t gfn, mfn_t mfn, {/* Do an unshare to cleanly take care of all corner cases. */ int rc; - rc =mem_sharing_unshare_page (p2m->domain, -gfn_x (gfn_add (gfn, i)),0 ); + rc =mem_sharing_unshare_page (p2m->domain,gfn_x (gfn_add (gfn, i)));if ( rc ) {p2m_unlock (p2m);
xen/include/asm-x86/mem_sharing.h +3 -5
@@ -70,10 +70,9 @@ int __mem_sharing_unshare_page(struct domain *d,static inline int mem_sharing_unshare_page (struct domain *d, -unsigned long gfn, -uint16_t flags) +unsigned long gfn) { -int rc =__mem_sharing_unshare_page (d, gfn, flags); +int rc =__mem_sharing_unshare_page (d, gfn,0 );BUG_ON (rc && (rc != -ENOMEM ));return rc; }
@@ -117,8 +116,7 @@ static inline unsigned int mem_sharing_get_nr_shared_mfns(void) }static inline -int mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags) +int mem_sharing_unshare_page (struct domain *d,unsigned long gfn) {ASSERT_UNREACHABLE ();return -EOPNOTSUPP ;
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 7/20 x86/mem_sharing: don't try to unshare twice during page fault
In reply to Tamas K LengyelThe page was already tried to be unshared in get_gfn_type_access. If that didn't work, then trying again is pointless. Don't try to send vm_event again either, simply check if there is a ring or not.
--- xen/arch/x86/hvm/hvm.c | 26 +++++++++++++++++--------- 1 file changed, 17 insertions(+), 9 deletions(-)
xen/arch/x86/hvm/hvm.c +17 -9
@@ -38,6 +38,7 @@#include <xen/warning.h> #include <xen/vpci.h> #include <xen/nospec.h> +#include <xen/vm_event.h> #include <asm/shadow.h> #include <asm/hap.h> #include <asm/current.h>
@@ -1706,11 +1707,14 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,struct domain *currd = curr->domain;struct p2m_domain *p2m, *hostp2m;int rc, fall_through =0 , paged =0 ; -int sharing_enomem =0 ;vm_event_request_t *req_ptr =NULL ;bool sync = false;unsigned int page_order; +#ifdef CONFIG_MEM_SHARING +bool sharing_enomem = false; +#endif +/* On Nested Virtualization, walk the guest page table. * If this succeeds, all is fine. * If this fails, inject a nested page fault into the guest.
@@ -1898,14 +1902,16 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,if (p2m_is_paged (p2mt) || (p2mt == p2m_ram_paging_out) ) paged =1 ; -/* Mem sharing: unshare the page and try again */ -if ( npfec.write_access && (p2mt == p2m_ram_shared) ) +#ifdef CONFIG_MEM_SHARING +/* Mem sharing: if still shared on write access then its enomem */ +if ( npfec.write_access &&p2m_is_shared (p2mt) ) {ASSERT (p2m_is_hostp2m (p2m)); - sharing_enomem =mem_sharing_unshare_page (currd, gfn); + sharing_enomem = true; rc =1 ; goto out_put_gfn; } +#endif /* Spurious fault? PoD and log-dirty also take this path. */ if (p2m_is_ram (p2mt) )
@@ -1959,19 +1965,21 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,*/ if ( paged )p2m_mem_paging_populate (currd, gfn); + +#ifdef CONFIG_MEM_SHARING if ( sharing_enomem ) { -int rv; - -if ( (rv =mem_sharing_notify_enomem (currd, gfn, true)) <0 ) +if ( !vm_event_check_ring (currd->vm_event_share) ) {gdprintk (XENLOG_ERR ,"Domain %hu attempt to unshare " -"gfn %lx, ENOMEM and no helper (rc %d)\n" , - currd->domain_id, gfn, rv); +"gfn %lx, ENOMEM and no helper\n" , + currd->domain_id, gfn);/* Crash the domain */ rc =0 ; } } +#endif +if ( req_ptr ) {if (monitor_traps (curr, sync, req_ptr) <0 )
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 8/20 x86/mem_sharing: define mem_sharing_domain to hold some scattered variables
In reply to Tamas K LengyelCreate struct mem_sharing_domain under hvm_domain and move mem sharing variables into it from p2m_domain and hvm_domain.
Expose the mem_sharing_enabled macro to be used consistently across Xen.
Remove some duplicate calls to mem_sharing_enabled in mem_sharing.c
--- xen/arch/x86/mm/mem_sharing.c | 30 +++++------------------------- xen/drivers/passthrough/pci.c | 3 +-- xen/include/asm-x86/hvm/domain.h | 6 +++++- xen/include/asm-x86/mem_sharing.h | 16 ++++++++++++++++ xen/include/asm-x86/p2m.h | 4 ---- 5 files changed, 27 insertions(+), 32 deletions(-)
xen/arch/x86/mm/mem_sharing.c +5 -25
@@ -197,9 +197,6 @@ static inline shr_handle_t get_next_handle(void)return x +1 ; } -#define mem_sharing_enabled (d) \ - (is_hvm_domain(d) && (d)->arch.hvm.mem_sharing_enabled) -static atomic_t nr_saved_mfns =ATOMIC_INIT (0 );static atomic_t nr_shared_mfns =ATOMIC_INIT (0 );
@@ -1300,6 +1297,7 @@ int __mem_sharing_unshare_page(struct domain *d,int relinquish_shared_pages (struct domain *d) {int rc =0 ; +struct mem_sharing_domain *msd = &d->arch.hvm.mem_sharing;struct p2m_domain *p2m =p2m_get_hostp2m (d);unsigned long gfn, count =0 ;
@@ -1307,7 +1305,7 @@ int relinquish_shared_pages(struct domain *d)return 0 ;p2m_lock (p2m); -for ( gfn = p2m->next_shared_gfn_to_relinquish; +for ( gfn = msd->next_shared_gfn_to_relinquish; gfn <= p2m->max_mapped_pfn; gfn++ ) {p2m_access_t a;
@@ -1342,7 +1340,7 @@ int relinquish_shared_pages(struct domain *d) {if (hypercall_preempt_check () ) { - p2m->next_shared_gfn_to_relinquish = gfn +1 ; + msd->next_shared_gfn_to_relinquish = gfn +1 ; rc = -ERESTART ;break ; }
@@ -1428,7 +1426,7 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)/* Only HAP is supported */ rc = -ENODEV ; -if ( !hap_enabled (d) || !d->arch.hvm.mem_sharing_enabled ) +if ( !mem_sharing_enabled (d) ) goto out; switch ( mso.op )
@@ -1437,10 +1435,6 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg) {shr_handle_t handle; - rc = -EINVAL ; -if ( !mem_sharing_enabled (d) ) - goto out; - rc =nominate_page (d,_gfn (mso.u.nominate.u.gfn),0 , &handle); mso.u.nominate.handle = handle; }
@@ -1452,9 +1446,6 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)gfn_t gfn;shr_handle_t handle; - rc = -EINVAL ; -if ( !mem_sharing_enabled (d) ) - goto out; rc =mem_sharing_gref_to_gfn (d->grant_table, gref, &gfn,NULL );if ( rc <0 ) goto out;
@@ -1470,10 +1461,6 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)struct domain *cd;shr_handle_t sh, ch; - rc = -EINVAL ; -if ( !mem_sharing_enabled (d) ) - goto out; - rc =rcu_lock_live_remote_domain_by_id (mso.u.share.client_domain, &cd);if ( rc )
@@ -1540,10 +1527,6 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)struct domain *cd;shr_handle_t sh; - rc = -EINVAL ; -if ( !mem_sharing_enabled (d) ) - goto out; - rc =rcu_lock_live_remote_domain_by_id (mso.u.share.client_domain, &cd);if ( rc )
@@ -1602,9 +1585,6 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg) mso.u.range.opaque > mso.u.range.last_gfn) ) goto out; -if ( !mem_sharing_enabled (d) ) - goto out; - rc =rcu_lock_live_remote_domain_by_id (mso.u.range.client_domain, &cd);if ( rc )
@@ -1708,7 +1688,7 @@ int mem_sharing_domctl(struct domain *d, struct xen_domctl_mem_sharing_op *mec)if (unlikely (is_iommu_enabled (d) && mec->u.enable) ) rc = -EXDEV ;else - d->arch.hvm.mem_sharing_enabled = mec->u.enable; + d->arch.hvm.mem_sharing.enabled = mec->u.enable; }break ;
xen/drivers/passthrough/pci.c +1 -2
@@ -1498,8 +1498,7 @@ static int assign_device(struct domain *d, u16 seg, u8 bus, u8 devfn, u32 flag)/* Prevent device assign if mem paging or mem sharing have been * enabled for this domain */ if ( d != dom_io && -unlikely ((is_hvm_domain (d) && - d->arch.hvm.mem_sharing_enabled) || +unlikely (mem_sharing_enabled (d) ||vm_event_check_ring (d->vm_event_paging) ||p2m_get_hostp2m (d)->global_logdirty) )return -EXDEV ;
xen/include/asm-x86/hvm/domain.h +5 -1
@@ -29,6 +29,7 @@#include <asm/hvm/viridian.h> #include <asm/hvm/vmx/vmcs.h> #include <asm/hvm/svm/vmcb.h> +#include <asm/mem_sharing.h> #include <public/grant_table.h> #include <public/hvm/params.h> #include <public/hvm/save.h>
@@ -156,7 +157,6 @@ struct hvm_domain {struct viridian_domain *viridian; -bool_t mem_sharing_enabled;bool_t qemu_mapcache_invalidate;bool_t is_s3_suspended;
@@ -192,6 +192,10 @@ struct hvm_domain {struct vmx_domain vmx; struct svm_domain svm; }; + +#ifdef CONFIG_MEM_SHARING +struct mem_sharing_domain mem_sharing; +#endif };#endif /* __ASM_X86_HVM_DOMAIN_H__ */
xen/include/asm-x86/mem_sharing.h +16 -0
@@ -26,6 +26,20 @@#ifdef CONFIG_MEM_SHARING +struct mem_sharing_domain +{ +bool enabled; + +/* +* When releasing shared gfn's in a preemptible manner, recall where +* to resume the search. +*/ +unsigned long next_shared_gfn_to_relinquish; +}; + +#define mem_sharing_enabled (d) \ + (hap_enabled(d) && (d)->arch.hvm.mem_sharing.enabled) +/* Auditing of memory sharing code? */ #ifndef NDEBUG #define MEM_SHARING_AUDIT 1
@@ -105,6 +119,8 @@ int relinquish_shared_pages(struct domain *d);#else +#define mem_sharing_enabled (d) false +static inline unsigned int mem_sharing_get_nr_saved_mfns (void ) {return 0 ;
xen/include/asm-x86/p2m.h +0 -4
@@ -305,10 +305,6 @@ struct p2m_domain {unsigned long min_remapped_gfn;unsigned long max_remapped_gfn; -/* When releasing shared gfn's in a preemptible manner, recall where -* to resume the search */ -unsigned long next_shared_gfn_to_relinquish; -#ifdef CONFIG_HVM /* Populate-on-demand variables * All variables are protected with the pod lock. We cannot rely on
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 9/20 x86/mem_sharing: Use INVALID_MFN and p2m_is_shared in relinquish_shared_pages
In reply to Tamas K LengyelWhile using _mfn(0) is of no consequence during teardown, INVALID_MFN is the correct value that should be used.
--- xen/arch/x86/mm/mem_sharing.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
xen/arch/x86/mm/mem_sharing.c +2 -2
@@ -1317,7 +1317,7 @@ int relinquish_shared_pages(struct domain *d)break ; mfn = p2m->get_entry (p2m,_gfn (gfn), &t, &a,0 ,NULL ,NULL ); -if (mfn_valid (mfn) && t == p2m_ram_shared ) +if (mfn_valid (mfn) &&p2m_is_shared (t) ) {/* Does not fail with ENOMEM given the DESTROY flag */ BUG_ON(__mem_sharing_unshare_page(d, gfn,
@@ -1327,7 +1327,7 @@ int relinquish_shared_pages(struct domain *d)* unshare. Must succeed: we just read the old entry and * we hold the p2m lock. */ - set_rc = p2m->set_entry (p2m,_gfn (gfn),_mfn (0 ),PAGE_ORDER_4K , + set_rc = p2m->set_entry (p2m,_gfn (gfn),INVALID_MFN ,PAGE_ORDER_4K , p2m_invalid, p2m_access_rwx,-1 );ASSERT (!set_rc); count +=0x10 ;
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 11/20 x86/mem_sharing: Convert MEM_SHARING_DESTROY_GFN to a bool
In reply to Tamas K LengyelMEM_SHARING_DESTROY_GFN is used on the 'flags' bitfield during unsharing. However, the bitfield is not used for anything else, so just convert it to a bool instead.
--- xen/arch/x86/mm/mem_sharing.c | 7 +++---- xen/arch/x86/mm/p2m.c | 1 + xen/common/memory.c | 2 +- xen/include/asm-x86/mem_sharing.h | 5 ++--- 4 files changed, 7 insertions(+), 8 deletions(-)
xen/arch/x86/mm/mem_sharing.c +3 -4
@@ -1175,7 +1175,7 @@ err_out:*/ int __mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags) +bool destroy) {p2m_type_t p2mt;mfn_t mfn;
@@ -1231,7 +1231,7 @@ int __mem_sharing_unshare_page(struct domain *d,* If the GFN is getting destroyed drop the references to MFN * (possibly freeing the page), and exit early. */ -if ( flags &MEM_SHARING_DESTROY_GFN ) +if ( destroy ) {if ( !last_gfn )mem_sharing_gfn_destroy (page, d, gfn_info);
@@ -1321,8 +1321,7 @@ int relinquish_shared_pages(struct domain *d)if (mfn_valid (mfn) &&p2m_is_shared (t) ) {/* Does not fail with ENOMEM given the DESTROY flag */ -BUG_ON (__mem_sharing_unshare_page (d, gfn, -MEM_SHARING_DESTROY_GFN )); +BUG_ON (__mem_sharing_unshare_page (d, gfn, true));/* * Clear out the p2m entry so no one else may try to * unshare. Must succeed: we just read the old entry and
xen/arch/x86/mm/p2m.c +1 -0
@@ -517,6 +517,7 @@ mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l,*/ if (mem_sharing_unshare_page (p2m->domain, gfn_l) <0 )mem_sharing_notify_enomem (p2m->domain, gfn_l, false); + mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL ); }
xen/common/memory.c +1 -1
@@ -352,7 +352,7 @@ int guest_remove_page(struct domain *d, unsigned long gmfn)* might be the only one using this shared page, and we need to * trigger proper cleanup. Once done, this is like any other page. */ - rc =mem_sharing_unshare_page (d, gmfn,0 ); + rc =mem_sharing_unshare_page (d, gmfn);if ( rc ) {mem_sharing_notify_enomem (d, gmfn, false);
xen/include/asm-x86/mem_sharing.h +2 -3
@@ -76,17 +76,16 @@ struct page_sharing_infounsigned int mem_sharing_get_nr_saved_mfns (void );unsigned int mem_sharing_get_nr_shared_mfns (void ); -#define MEM_SHARING_DESTROY_GFN (1<<1)/* Only fails with -ENOMEM. Enforce it with a BUG_ON wrapper. */ int __mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags); +bool destroy);static inline int mem_sharing_unshare_page (struct domain *d,unsigned long gfn) { -int rc =__mem_sharing_unshare_page (d, gfn,0 ); +int rc =__mem_sharing_unshare_page (d, gfn, false);BUG_ON (rc && (rc != -ENOMEM ));return rc; }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 13/20 x86/mem_sharing: ASSERT that p2m_set_entry succeeds
In reply to Tamas K Lengyel--- xen/arch/x86/mm/mem_sharing.c | 46 +++++++++++++++++------------------ 1 file changed, 22 insertions(+), 24 deletions(-)
xen/arch/x86/mm/mem_sharing.c +22 -24
@@ -1113,39 +1113,37 @@ int add_to_physmap(struct domain *sd, unsigned long sgfn, shr_handle_t sh, goto err_unlock; } +/* +* Must succeed, we just read the entry and hold the p2m lock +* via get_two_gfns. +*/ ret =p2m_set_entry (p2m,_gfn (cgfn), smfn,PAGE_ORDER_4K , p2m_ram_shared, a); +ASSERT (!ret); -/* Tempted to turn this into an assert */ -if ( ret ) +/* +* There is a chance we're plugging a hole where a paged out +* page was. +*/ +if (p2m_is_paging (cmfn_type) && (cmfn_type != p2m_ram_paging_out) ) { -mem_sharing_gfn_destroy (spage, cd, gfn_info); -put_page_and_type (spage); - }else { +atomic_dec (&cd->paged_pages);/* -* There is a chance we're plugging a hole where a paged out -* page was. +* Further, there is a chance this was a valid page. +* Don't leak it. */ -if (p2m_is_paging (cmfn_type) && (cmfn_type != p2m_ram_paging_out) ) +if (mfn_valid (cmfn) ) { -atomic_dec (&cd->paged_pages); -/* -* Further, there is a chance this was a valid page. -* Don't leak it. -*/ -if (mfn_valid (cmfn) ) - { -struct page_info *cpage =mfn_to_page (cmfn); +struct page_info *cpage =mfn_to_page (cmfn); -if ( !get_page (cpage, cd) ) - { -domain_crash (cd); - ret = -EOVERFLOW ; - goto err_unlock; - } -put_page_alloc_ref (cpage); -put_page (cpage); +if ( !get_page (cpage, cd) ) + { +domain_crash (cd); + ret = -EOVERFLOW ; + goto err_unlock; } +put_page_alloc_ref (cpage); +put_page (cpage); } }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 12/20 x86/mem_sharing: Replace MEM_SHARING_DEBUG with gdprintk
In reply to Tamas K LengyelUsing XENLOG_ERR level since this is only used in debug paths (ie. it's expected the user already has loglvl=all set).
--- xen/arch/x86/mm/mem_sharing.c | 81 ++++++++++++++++++----------------- 1 file changed, 41 insertions(+), 40 deletions(-)
xen/arch/x86/mm/mem_sharing.c +41 -40
@@ -49,9 +49,6 @@ typedef struct pg_lock_data { static DEFINE_PER_CPU(pg_lock_data_t, __pld); -#define MEM_SHARING_DEBUG(_f, _a...) \ - debugtrace_printk("mem_sharing_debug: %s(): " _f, __func__, ##_a) -/* Reverse map defines */ #define RMAP_HASHTAB_ORDER 0#define RMAP_HASHTAB_SIZE \
@@ -491,8 +488,9 @@ static int audit(void)/* If we can't lock it, it's definitely not a shared page */ if ( !mem_sharing_page_lock (pg) ) { -MEM_SHARING_DEBUG ("mfn %lx in audit list, but cannot be locked (%lx)!\n" , -mfn_x (mfn), pg->u.inuse.type_info); +gdprintk (XENLOG_ERR , +"mfn %lx in audit list, but cannot be locked (%lx)!\n" , +mfn_x (mfn), pg->u.inuse.type_info); errors++;continue ; }
@@ -500,8 +498,9 @@ static int audit(void)/* Check if the MFN has correct type, owner and handle. */ if ( (pg->u.inuse.type_info & PGT_type_mask) != PGT_shared_page ) { -MEM_SHARING_DEBUG ("mfn %lx in audit list, but not PGT_shared_page (%lx)!\n" , -mfn_x (mfn), pg->u.inuse.type_info & PGT_type_mask); +gdprintk (XENLOG_ERR , +"mfn %lx in audit list, but not PGT_shared_page (%lx)!\n" , +mfn_x (mfn), pg->u.inuse.type_info & PGT_type_mask); errors++;continue ; }
@@ -509,24 +508,24 @@ static int audit(void)/* Check the page owner. */ if (page_get_owner (pg) != dom_cow ) { -MEM_SHARING_DEBUG ("mfn %lx shared, but wrong owner (%hu)!\n" , -mfn_x (mfn),page_get_owner (pg)->domain_id); +gdprintk (XENLOG_ERR ,"mfn %lx shared, but wrong owner (%hu)!\n" , +mfn_x (mfn),page_get_owner (pg)->domain_id); errors++; }/* Check the m2p entry */ if ( !SHARED_M2P (get_gpfn_from_mfn (mfn_x (mfn))) ) { -MEM_SHARING_DEBUG ("mfn %lx shared, but wrong m2p entry (%lx)!\n" , -mfn_x (mfn),get_gpfn_from_mfn (mfn_x (mfn))); +gdprintk (XENLOG_ERR ,"mfn %lx shared, but wrong m2p entry (%lx)!\n" , +mfn_x (mfn),get_gpfn_from_mfn (mfn_x (mfn))); errors++; }/* Check we have a list */ if ( (!pg->sharing) || !rmap_has_entries (pg) ) { -MEM_SHARING_DEBUG ("mfn %lx shared, but empty gfn list!\n" , -mfn_x (mfn)); +gdprintk (XENLOG_ERR ,"mfn %lx shared, but empty gfn list!\n" , +mfn_x (mfn)); errors++;continue ; }
@@ -545,24 +544,26 @@ static int audit(void) d =get_domain_by_id (g->domain);if ( d ==NULL ) { -MEM_SHARING_DEBUG ("Unknown dom: %hu, for PFN=%lx, MFN=%lx\n" , - g->domain, g->gfn,mfn_x (mfn)); +gdprintk (XENLOG_ERR , +"Unknown dom: %hu, for PFN=%lx, MFN=%lx\n" , + g->domain, g->gfn,mfn_x (mfn)); errors++;continue ; } o_mfn =get_gfn_query_unlocked (d, g->gfn, &t);if ( !mfn_eq (o_mfn, mfn) ) { -MEM_SHARING_DEBUG ("Incorrect P2M for d=%hu, PFN=%lx." -"Expecting MFN=%lx, got %lx\n" , - g->domain, g->gfn,mfn_x (mfn),mfn_x (o_mfn)); +gdprintk (XENLOG_ERR ,"Incorrect P2M for d=%hu, PFN=%lx." +"Expecting MFN=%lx, got %lx\n" , + g->domain, g->gfn,mfn_x (mfn),mfn_x (o_mfn)); errors++; }if ( t != p2m_ram_shared ) { -MEM_SHARING_DEBUG ("Incorrect P2M type for d=%hu, PFN=%lx MFN=%lx." -"Expecting t=%d, got %d\n" , - g->domain, g->gfn,mfn_x (mfn), p2m_ram_shared, t); +gdprintk (XENLOG_ERR , +"Incorrect P2M type for d=%hu, PFN=%lx MFN=%lx." +"Expecting t=%d, got %d\n" , + g->domain, g->gfn,mfn_x (mfn), p2m_ram_shared, t); errors++; }put_domain (d);
@@ -571,10 +572,10 @@ static int audit(void)/* The type count has an extra ref because we have locked the page */ if ( (nr_gfns +1 ) != (pg->u.inuse.type_info & PGT_count_mask) ) { -MEM_SHARING_DEBUG ("Mismatched counts for MFN=%lx." -"nr_gfns in list %lu, in type_info %lx\n" , -mfn_x (mfn), nr_gfns, - (pg->u.inuse.type_info & PGT_count_mask)); +gdprintk (XENLOG_ERR ,"Mismatched counts for MFN=%lx." +"nr_gfns in list %lu, in type_info %lx\n" , +mfn_x (mfn), nr_gfns, + (pg->u.inuse.type_info & PGT_count_mask)); errors++; }
@@ -585,8 +586,8 @@ static int audit(void)if ( count_found != count_expected ) { -MEM_SHARING_DEBUG ("Expected %ld shared mfns, found %ld." , - count_expected, count_found); +gdprintk (XENLOG_ERR ,"Expected %ld shared mfns, found %ld." , + count_expected, count_found); errors++; }
@@ -765,12 +766,12 @@ static int debug_mfn(mfn_t mfn)return -EINVAL ; } -MEM_SHARING_DEBUG ( -"Debug page: MFN=%lx is ci=%lx, ti=%lx, owner_id=%d\n" , -mfn_x (page_to_mfn (page)), - page->count_info, - page->u.inuse.type_info, -page_get_owner (page)->domain_id); +gdprintk (XENLOG_ERR , +"Debug page: MFN=%lx is ci=%lx, ti=%lx, owner_id=%d\n" , +mfn_x (page_to_mfn (page)), + page->count_info, + page->u.inuse.type_info, +page_get_owner (page)->domain_id);/* -1 because the page is locked and that's an additional type ref */ num_refs = ((int ) (page->u.inuse.type_info & PGT_count_mask)) -1 ;
@@ -786,8 +787,9 @@ static int debug_gfn(struct domain *d, gfn_t gfn) mfn =get_gfn_query (d,gfn_x (gfn), &p2mt); -MEM_SHARING_DEBUG ("Debug for dom%d, gfn=%" PRI_gfn"\n" , - d->domain_id,gfn_x (gfn)); +gdprintk (XENLOG_ERR ,"Debug for dom%d, gfn=%" PRI_gfn"\n" , + d->domain_id,gfn_x (gfn)); + num_refs =debug_mfn (mfn);put_gfn (d,gfn_x (gfn));
@@ -803,14 +805,13 @@ static int debug_gref(struct domain *d, grant_ref_t ref) rc =mem_sharing_gref_to_gfn (d->grant_table, ref, &gfn, &status);if ( rc ) { -MEM_SHARING_DEBUG ("Asked to debug [dom=%d,gref=%u]: error %d.\n" , - d->domain_id, ref, rc); +gdprintk (XENLOG_ERR ,"Asked to debug [dom=%d,gref=%u]: error %d.\n" , + d->domain_id, ref, rc);return rc; } -MEM_SHARING_DEBUG ( -"==> Grant [dom=%d,ref=%d], status=%x. " , - d->domain_id, ref, status); +gdprintk (XENLOG_ERR ," ==> Grant [dom=%d,ref=%d], status=%x. " , + d->domain_id, ref, status);return debug_gfn (d, gfn); }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 14/20 x86/mem_sharing: Enable mem_sharing on first memop
In reply to Tamas K LengyelIt is wasteful to require separate hypercalls to enable sharing on both the parent and the client domain during VM forking. To speed things up we enable sharing on the first memop in case it wasn't already enabled.
--- xen/arch/x86/mm/mem_sharing.c | 39 +++++++++++++++++++++-------------- 1 file changed, 23 insertions(+), 16 deletions(-)
xen/arch/x86/mm/mem_sharing.c +23 -16
@@ -1402,6 +1402,24 @@ static int range_share(struct domain *d, struct domain *cd,return rc; } +static inline int mem_sharing_control (struct domain *d,bool enable) +{ +if ( enable ) + { +if (unlikely (!is_hvm_domain (d)) ) +return -ENOSYS ; + +if (unlikely (!hap_enabled (d)) ) +return -ENODEV ; + +if (unlikely (is_iommu_enabled (d)) ) +return -EXDEV ; + } + + d->arch.hvm.mem_sharing.enabled = enable; +return 0 ; +} +int mem_sharing_memop (XEN_GUEST_HANDLE_PARAM (xen_mem_sharing_op_t) arg) {int rc;
@@ -1423,10 +1441,8 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)if ( rc ) goto out; -/* Only HAP is supported */ - rc = -ENODEV ; -if ( !mem_sharing_enabled (d) ) - goto out; +if ( !mem_sharing_enabled (d) && (rc =mem_sharing_control (d, true)) ) +return rc;switch ( mso.op ) {
@@ -1675,24 +1691,15 @@ int mem_sharing_domctl(struct domain *d, struct xen_domctl_mem_sharing_op *mec) {int rc; -/* Only HAP is supported */ -if ( !hap_enabled (d) ) -return -ENODEV ; -switch (mec->op) {case XEN_DOMCTL_MEM_SHARING_CONTROL : - { - rc =0 ; -if (unlikely (is_iommu_enabled (d) && mec->u.enable) ) - rc = -EXDEV ; -else - d->arch.hvm.mem_sharing.enabled = mec->u.enable; - } -break ; + rc =mem_sharing_control (d, mec->u.enable); +break ;default : rc = -ENOSYS ; +break ; }return rc;
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 15/20 x86/mem_sharing: Skip xen heap pages in memshr nominate
In reply to Tamas K LengyelTrying to share these would fail anyway, better to skip them early.
--- xen/arch/x86/mm/mem_sharing.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
xen/arch/x86/mm/mem_sharing.c +5 -1
@@ -852,6 +852,11 @@ static int nominate_page(struct domain *d, gfn_t gfn,if ( !p2m_is_sharable (p2mt) ) goto out; +/* Skip xen heap pages */ + page =mfn_to_page (mfn); +if ( !page ||is_xen_heap_page (page) ) + goto out; +/* Check if there are mem_access/remapped altp2m entries for this page */ if (altp2m_active (d) ) {
@@ -882,7 +887,6 @@ static int nominate_page(struct domain *d, gfn_t gfn, }/* Try to convert the mfn to the sharable type */ - page =mfn_to_page (mfn); ret =page_make_sharable (d, page, expected_refcnt);if ( ret ) goto out;
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 18/20 xen/mem_access: Use __get_gfn_type_access in set_mem_access
In reply to Tamas K LengyelUse __get_gfn_type_access instead of p2m->get_entry to trigger page-forking when the mem_access permission is being set on a page that has not yet been copied over from the parent.
--- xen/arch/x86/mm/mem_access.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
xen/arch/x86/mm/mem_access.c +2 -3
@@ -303,11 +303,10 @@ static int set_mem_access(struct domain *d, struct p2m_domain *p2m,ASSERT (!ap2m);#endif { -mfn_t mfn;p2m_access_t _a;p2m_type_t t; - - mfn = p2m->get_entry (p2m, gfn, &t, &_a,0 ,NULL ,NULL ); +mfn_t mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &t, &_a, +P2M_ALLOC ,NULL , false); rc = p2m->set_entry (p2m, gfn, mfn,PAGE_ORDER_4K , t, a,-1 ); }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 19/20 x86/mem_sharing: reset a fork
In reply to Tamas K LengyelImplement hypercall that allows a fork to shed all memory that got allocated for it during its execution and re-load its vCPU context from the parent VM. This allows the forked VM to reset into the same state the parent VM is in a faster way then creating a new fork would be. Measurements show about a 2x speedup during normal fuzzing operations. Performance may vary depending how much memory got allocated for the forked VM. If it has been completely deduplicated from the parent VM then creating a new fork would likely be more performant.
--- xen/arch/x86/mm/mem_sharing.c | 105 ++++++++++++++++++++++++++++++++++ xen/include/public/memory.h | 1 + 2 files changed, 106 insertions(+)
xen/arch/x86/mm/mem_sharing.c +105 -0
@@ -1622,6 +1622,87 @@ static int mem_sharing_fork(struct domain *d, struct domain *cd)return 0 ; } +struct gfn_free ; +struct gfn_free { +struct gfn_free *next; +struct page_info *page; +gfn_t gfn; +}; + +static int mem_sharing_fork_reset (struct domain *d,struct domain *cd) +{ +int rc; + +struct p2m_domain * p2m =p2m_get_hostp2m (cd); +struct gfn_free *list =NULL ; +struct page_info *page; + +page_list_for_each (page, &cd->page_list) + { +mfn_t mfn =page_to_mfn (page); +if (mfn_valid (mfn) ) + { +p2m_type_t p2mt; +p2m_access_t p2ma; +gfn_t gfn =mfn_to_gfn (cd, mfn); + mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &p2mt, &p2ma, +0 ,NULL , false); +if (p2m_is_ram (p2mt) ) + { +struct gfn_free *gfn_free; +if ( !get_page (page, cd) ) + goto err_reset; + +/* +* We can't free the page while iterating over the page_list +* so we build a separate list to loop over. +* +* We want to iterate over the page_list instead of checking +* gfn from 0 to max_gfn because this is ~10x faster. +*/ + gfn_free = xmalloc(struct gfn_free); +if ( !gfn_free ) + goto err_reset; + + gfn_free->gfn = gfn; + gfn_free->page = page; + gfn_free->next = list; + list = gfn_free; + } + } + } + +while ( list ) + { +struct gfn_free *next = list->next; + + rc = p2m->set_entry (p2m, list->gfn,INVALID_MFN ,PAGE_ORDER_4K , + p2m_invalid, p2m_access_rwx,-1 ); +put_page_alloc_ref (list->page); +put_page (list->page); + +xfree (list); + list = next; + +ASSERT (!rc); + } + +if ( (rc =fork_hvm (d, cd)) ) +return rc; + + err_reset: +while ( list ) + { +struct gfn_free *next = list->next; + +put_page (list->page); +xfree (list); + list = next; + } + +return 0 ; +} +int mem_sharing_memop (XEN_GUEST_HANDLE_PARAM (xen_mem_sharing_op_t) arg) {int rc;
@@ -1905,6 +1986,30 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)rcu_unlock_domain (pd);break ; } + +case XENMEM_sharing_op_fork_reset: + { +struct domain *pd; + + rc = -EINVAL ; +if ( mso.u.fork._pad[0 ] || mso.u.fork._pad[1 ] || + mso.u.fork._pad[2 ] ) + goto out; + + rc = -ENOSYS ; +if ( !d->parent ) + goto out; + + rc =rcu_lock_live_remote_domain_by_id (d->parent->domain_id, &pd); +if ( rc ) + goto out; + + rc =mem_sharing_fork_reset (pd, d); + +rcu_unlock_domain (pd); +break ; + } +default : rc = -ENOSYS ;break ;
xen/include/public/memory.h +1 -0
@@ -483,6 +483,7 @@ DEFINE_XEN_GUEST_HANDLE(xen_mem_access_op_t);#define XENMEM_sharing_op_audit 7#define XENMEM_sharing_op_range_share 8#define XENMEM_sharing_op_fork 9 +#define XENMEM_sharing_op_fork_reset 10#define XENMEM_SHARING_OP_S_HANDLE_INVALID (-10)#define XENMEM_SHARING_OP_C_HANDLE_INVALID (-9)
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 20/20 xen/tools: VM forking toolstack side
In reply to Tamas K LengyelAdd necessary bits to implement "xl fork-vm", "xl fork-launch-dm" and "xl fork-reset" commands. The process is split in two to allow tools needing access to the new VM as fast as possible after it was forked. It is expected that under certain use-cases the second command that launches QEMU will be skipped entirely.
--- tools/libxc/include/xenctrl.h | 6 + tools/libxc/xc_memshr.c | 22 ++++ tools/libxl/libxl.h | 7 + tools/libxl/libxl_create.c | 237 +++++++++++++++++++++++----------- tools/libxl/libxl_dm.c | 2 +- tools/libxl/libxl_dom.c | 83 ++++++++---- tools/libxl/libxl_internal.h | 1 + tools/libxl/libxl_types.idl | 1 + tools/xl/xl.h | 5 + tools/xl/xl_cmdtable.c | 22 ++++ tools/xl/xl_saverestore.c | 96 ++++++++++++++ tools/xl/xl_vmcontrol.c | 8 ++ 12 files changed, 386 insertions(+), 104 deletions(-)
tools/libxc/include/xenctrl.h +6 -0
@@ -2221,6 +2221,12 @@ int xc_memshr_range_share(xc_interface *xch,uint64_t first_gfn, uint64_t last_gfn); +int xc_memshr_fork (xc_interface *xch, +uint32_t source_domain, +uint32_t client_domain); + +int xc_memshr_fork_reset (xc_interface *xch,uint32_t forked_domain); +/* Debug calls: return the number of pages referencing the shared frame backing * the input argument. Should be one or greater. *
tools/libxc/xc_memshr.c +22 -0
@@ -237,6 +237,28 @@ int xc_memshr_debug_gref(xc_interface *xch,return xc_memshr_memop (xch, domid, &mso); } +int xc_memshr_fork (xc_interface *xch,uint32_t pdomid,uint32_t domid) +{ +xen_mem_sharing_op_t mso; + +memset (&mso,0 ,sizeof (mso)); + + mso.op = XENMEM_sharing_op_fork; + mso.u.fork.parent_domain = pdomid; + +return xc_memshr_memop (xch, domid, &mso); +} + +int xc_memshr_fork_reset (xc_interface *xch,uint32_t domid) +{ +xen_mem_sharing_op_t mso; + +memset (&mso,0 ,sizeof (mso)); + mso.op = XENMEM_sharing_op_fork_reset; + +return xc_memshr_memop (xch, domid, &mso); +} +int xc_memshr_audit (xc_interface *xch) {xen_mem_sharing_op_t mso;
tools/libxl/libxl.h +7 -0
@@ -1536,6 +1536,13 @@ int libxl_domain_create_new(libxl_ctx *ctx, libxl_domain_config *d_config,const libxl_asyncop_how *ao_how, const libxl_asyncprogress_how *aop_console_how)LIBXL_EXTERNAL_CALLERS_ONLY ; +int libxl_domain_fork_vm(libxl_ctx *ctx, uint32_t pdomid, uint32_t *domid) + LIBXL_EXTERNAL_CALLERS_ONLY; +int libxl_domain_fork_launch_dm(libxl_ctx *ctx, libxl_domain_config *d_config, + uint32_t domid, + const libxl_asyncprogress_how *aop_console_how) + LIBXL_EXTERNAL_CALLERS_ONLY; +int libxl_domain_fork_reset (libxl_ctx *ctx,uint32_t domid); int libxl_domain_create_restore(libxl_ctx *ctx, libxl_domain_config *d_config, uint32_t *domid, int restore_fd, int send_back_fd,
tools/libxl/libxl_create.c +161 -76
350 lines not on this page · raw message
tools/libxl/libxl_dm.c +1 -1
@@ -2784,7 +2784,7 @@ static void device_model_spawn_outcome(libxl__egc *egc,libxl__domain_build_state *state = dmss->build_state; -if (state->saved_state) { +if (state->saved_state && !state->forked_vm) { ret2 =unlink (state->saved_state);if (ret2) { LOGED(ERROR, dmss->guest_domid, "%s: failed to remove device-model state %s",
tools/libxl/libxl_dom.c +56 -27
141 lines not on this page · raw message
tools/libxl/libxl_internal.h +1 -0
@@ -1360,6 +1360,7 @@ typedef struct {char *saved_state;int dm_monitor_fd; +bool forked_vm;libxl__file_reference pv_kernel;libxl__file_reference pv_ramdisk;
tools/libxl/libxl_types.idl +1 -0
@@ -956,6 +956,7 @@ libxl_domain_config = Struct("domain_config", [ ("on_watchdog", libxl_action_on_shutdown), ("on_crash", libxl_action_on_shutdown), ("on_soft_reset", libxl_action_on_shutdown), + ("dm_restore_file", string, {'const': True}), ], dir=DIR_IN) libxl_diskinfo = Struct("diskinfo", [
tools/xl/xl.h +5 -0
@@ -31,6 +31,7 @@ struct cmd_spec { };struct domain_create { +uint32_t ddomid;/* fork launch dm for this domid */ int debug;int daemonize;int monitor;/* handle guest reboots etc */
@@ -45,6 +46,7 @@ struct domain_create {const char *config_file;char *extra_config;/* extra config string */ const char *restore_file; +const char *dm_restore_file;char *colo_proxy_script;bool userspace_colo_proxy;int migrate_fd;/* -1 means none */
@@ -127,6 +129,9 @@ int main_pciassignable_remove(int argc, char **argv);int main_pciassignable_list (int argc,char **argv);#ifndef LIBXL_HAVE_NO_SUSPEND_RESUME int main_restore (int argc,char **argv); +int main_fork_vm (int argc,char **argv); +int main_fork_launch_dm (int argc,char **argv); +int main_fork_reset (int argc,char **argv);int main_migrate_receive (int argc,char **argv);int main_save (int argc,char **argv);int main_migrate (int argc,char **argv);
tools/xl/xl_cmdtable.c +22 -0
@@ -180,6 +180,28 @@ struct cmd_spec cmd_table[] = {"-V, --vncviewer Connect to the VNC display after the domain is created.\n" "-A, --vncviewer-autopass Pass VNC password to viewer via stdin." }, + { "fork-vm", + &main_fork_vm,0 ,1 , +"Fork a domain from the running parent domid" , +"[options] <ParentDomid>" , +"-h Print this help.\n" +"-d Enable debug messages.\n" + }, + { "fork-launch-dm", + &main_fork_launch_dm,0 ,1 , +"Launch the device model for a forked VM" , +"[options] <ConfigFile> <DmRestoreFile> <Domid>" , +"-h Print this help.\n" +"-p Do not unpause domain after restoring it.\n" +"-d Enable debug messages.\n" + }, + { "fork-reset", + &main_fork_reset,0 ,1 , +"Launch the device model for a forked VM" , +"[options] <Domid>" , +"-h Print this help.\n" +"-d Enable debug messages.\n" + }, { "migrate-receive", &main_migrate_receive,0 ,1 , "Restore a domain from a saved state",
tools/xl/xl_saverestore.c +96 -0
102 lines not on this page · raw message
tools/xl/xl_vmcontrol.c +8 -0
@@ -645,6 +645,7 @@ int create_domain(struct domain_create *dom_info)libxl_domain_config d_config; +uint32_t ddomid = dom_info->ddomid;// launch dm for this domain iff set int debug = dom_info->debug;int daemonize = dom_info->daemonize;int monitor = dom_info->monitor;
@@ -655,6 +656,7 @@ int create_domain(struct domain_create *dom_info)const char *restore_file = dom_info->restore_file;const char *config_source =NULL ;const char *restore_source =NULL ; +const char *dm_restore_file = dom_info->dm_restore_file;int migrate_fd = dom_info->migrate_fd;bool config_in_json;
@@ -923,6 +925,12 @@ start:* restore/migrate-receive it again. */ restoring =0 ; + }else if ( ddomid ) { + d_config.dm_restore_file = dm_restore_file; + ret =libxl_domain_fork_launch_dm (ctx, &d_config, ddomid, + autoconnect_console_how); + domid = ddomid; + ddomid =INVALID_DOMID ; }else if (domid_soft_reset !=INVALID_DOMID ) {/* Do soft reset. */ ret = libxl_domain_soft_reset(ctx, &d_config, domid_soft_reset,
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 11/20] x86/mem_sharing: Convert MEM_SHARING_DESTROY_GFN to a bool
In reply to Tamas K LengyelHi Tamas,
Quoted 49 lines · Tamas K Lengyel
MEM_SHARING_DESTROY_GFNis used on the 'flags' bitfield during unsharing. However, the bitfield is not used for anything else, so just convert it to a bool instead.Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 7 +++---- xen/arch/x86/mm/p2m.c | 1 + xen/common/memory.c | 2 +- xen/include/asm-x86/mem_sharing.h | 5 ++--- 4 files changed, 7 insertions(+), 8 deletions(-)xen/arch/x86/mm/mem_sharing.c +3
-4@@ -1175,7 +1175,7 @@ err_out:*/ int __mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags) +bool destroy) {p2m_type_t p2mt;mfn_t mfn;@@ -1231,7 +1231,7 @@ int __mem_sharing_unshare_page(struct domain *d,* If the GFN is getting destroyed drop the references to MFN * (possibly freeing the page), and exit early. */ -if ( flags &MEM_SHARING_DESTROY_GFN ) +if ( destroy ) {if ( !last_gfn )mem_sharing_gfn_destroy (page, d, gfn_info);@@ -1321,8 +1321,7 @@ int relinquish_shared_pages(struct domain *d)if (mfn_valid (mfn) &&p2m_is_shared (t) ) {/* Does not fail with ENOMEM given the DESTROY flag */ -BUG_ON (__mem_sharing_unshare_page (d, gfn, -MEM_SHARING_DESTROY_GFN )); +BUG_ON (__mem_sharing_unshare_page (d, gfn, true));/* * Clear out the p2m entry so no one else may try to * unshare. Must succeed: we just read the old entry and xen/arch/x86/mm/p2m.c +0
-0
xen/arch/x86/mm/p2m.c mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l, · line 520
Quoted 4 lines
xen/arch/x86/mm/p2m.c +1
-0@@ -517,6 +517,7 @@ mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l,*/ if (mem_sharing_unshare_page (p2m->domain, gfn_l) <0 )mem_sharing_notify_enomem (p2m->domain, gfn_l, false); +
This line looks spurious.
xen/common/memory.c int guest_remove_page(struct domain *d, unsigned long gmfn) · line 355
Quoted 13 lines
xen/arch/x86/mm/p2m.c +0
-0mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL ); }xen/common/memory.c +1
-1@@ -352,7 +352,7 @@ int guest_remove_page(struct domain *d, unsigned long gmfn)* might be the only one using this shared page, and we need to * trigger proper cleanup. Once done, this is like any other page. */ - rc =mem_sharing_unshare_page (d, gmfn,0 ); + rc =mem_sharing_unshare_page (d, gmfn);
AFAICT, this patch does not reduce the number of parameters for mem_sharing_unshare_page(). Did you intend to make this change in another patch?
Quoted 18 lines
xen/common/memory.c +0
-0if ( rc ) {mem_sharing_notify_enomem (d, gmfn, false);xen/include/asm-x86/mem_sharing.h +1
-2@@ -76,17 +76,16 @@ struct page_sharing_infounsigned int mem_sharing_get_nr_saved_mfns (void );unsigned int mem_sharing_get_nr_shared_mfns (void ); -#define MEM_SHARING_DESTROY_GFN (1<<1)/* Only fails with -ENOMEM. Enforce it with a BUG_ON wrapper. */ int __mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags); +bool destroy);
xen/include/asm-x86/mem_sharing.h struct page_sharing_info · line 91
Quoted 10 lines
xen/include/asm-x86/mem_sharing.h +1
-1static inline int mem_sharing_unshare_page (struct domain *d,unsigned long gfn) { -int rc =__mem_sharing_unshare_page (d, gfn,0 ); +int rc =__mem_sharing_unshare_page (d, gfn, false);BUG_ON (rc && (rc != -ENOMEM ));return rc; }
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Tamas K LengyelHi Tamas,
Quoted 29 lines · Tamas K Lengyel
Implement hypercall that allows a fork to shed all memory that got allocated for it during its execution and re-load its vCPU context from the parent VM. This allows the forked VM to reset into the same state the parent VM is in a faster way then creating a new fork would be. Measurements show about a 2x speedup during normal fuzzing operations. Performance may vary depending how much memory got allocated for the forked VM. If it has been completely deduplicated from the parent VM then creating a new fork would likely be more performant.
Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 105 ++++++++++++++++++++++++++++++++++ xen/include/public/memory.h | 1 + 2 files changed, 106 insertions(+)xen/arch/x86/mm/mem_sharing.c +6
-0@@ -1622,6 +1622,87 @@ static int mem_sharing_fork(struct domain *d, struct domain *cd)return 0 ; } +struct gfn_free ; +struct gfn_free { +struct gfn_free *next; +struct page_info *page; +gfn_t gfn; +};
xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1640
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+ +static int mem_sharing_fork_reset (struct domain *d,struct domain *cd) +{ +int rc; + +struct p2m_domain * p2m =p2m_get_hostp2m (cd); +struct gfn_free *list =NULL ; +struct page_info *page; + +page_list_for_each (page, &cd->page_list)
AFAICT, your domain is not paused, so it would be possible to have page added/remove in that list behind your back.
You also have multiple loop on the page_list in this function. Given the number of page_list can be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.
Quoted 13 lines
xen/arch/x86/mm/mem_sharing.c +13
-0+ { +mfn_t mfn =page_to_mfn (page); +if (mfn_valid (mfn) ) + { +p2m_type_t p2mt; +p2m_access_t p2ma; +gfn_t gfn =mfn_to_gfn (cd, mfn); + mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &p2mt, &p2ma, +0 ,NULL , false); +if (p2m_is_ram (p2mt) ) + { +struct gfn_free *gfn_free; +if ( !get_page (page, cd) )
xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1663
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+ goto err_reset; + +/* +* We can't free the page while iterating over the page_list +* so we build a separate list to loop over. +* +* We want to iterate over the page_list instead of checking +* gfn from 0 to max_gfn because this is ~10x faster. +*/ + gfn_free = xmalloc(struct gfn_free);
If I did the math right, for a 4G guest this will require at ~24MB of memory. Actually, is it really necessary to do the allocation for a short period of time?
What are you trying to achieve by iterating twice on the GFN? Wouldn't it be easier to pause the domain?
Quoted 32 lines
xen/arch/x86/mm/mem_sharing.c +32
-0+if ( !gfn_free ) + goto err_reset; + + gfn_free->gfn = gfn; + gfn_free->page = page; + gfn_free->next = list; + list = gfn_free; + } + } + } + +while ( list ) + { +struct gfn_free *next = list->next; + + rc = p2m->set_entry (p2m, list->gfn,INVALID_MFN ,PAGE_ORDER_4K , + p2m_invalid, p2m_access_rwx,-1 ); +put_page_alloc_ref (list->page); +put_page (list->page); + +xfree (list); + list = next; + +ASSERT (!rc); + } + +if ( (rc =fork_hvm (d, cd)) ) +return rc; + + err_reset: +while ( list ) + {
xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1705
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+struct gfn_free *next = list->next; + +put_page (list->page); +xfree (list); + list = next; + } + +return 0 ; +} +
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 11/20] x86/mem_sharing: Convert MEM_SHARING_DESTROY_GFN to a bool
In reply to Julien GrallQuoted 51 lines · Julien Grall
Hi Tamas,
Quoted 47 lines · Tamas K Lengyel
MEM_SHARING_DESTROY_GFNis used on the 'flags' bitfield during unsharing. However, the bitfield is not used for anything else, so just convert it to a bool instead.Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 7 +++---- xen/arch/x86/mm/p2m.c | 1 + xen/common/memory.c | 2 +- xen/include/asm-x86/mem_sharing.h | 5 ++--- 4 files changed, 7 insertions(+), 8 deletions(-)xen/arch/x86/mm/mem_sharing.c +3
-4@@ -1175,7 +1175,7 @@ err_out:*/ int __mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags) +bool destroy) {p2m_type_t p2mt;mfn_t mfn;@@ -1231,7 +1231,7 @@ int __mem_sharing_unshare_page(struct domain *d,* If the GFN is getting destroyed drop the references to MFN * (possibly freeing the page), and exit early. */ -if ( flags &MEM_SHARING_DESTROY_GFN ) +if ( destroy ) {if ( !last_gfn )mem_sharing_gfn_destroy (page, d, gfn_info);@@ -1321,8 +1321,7 @@ int relinquish_shared_pages(struct domain *d)if (mfn_valid (mfn) &&p2m_is_shared (t) ) {/* Does not fail with ENOMEM given the DESTROY flag */ -BUG_ON (__mem_sharing_unshare_page (d, gfn, -MEM_SHARING_DESTROY_GFN )); +BUG_ON (__mem_sharing_unshare_page (d, gfn, true));/* * Clear out the p2m entry so no one else may try to * unshare. Must succeed: we just read the old entry and xen/arch/x86/mm/p2m.c +0
-1--- a/xen/arch/x86/mm/p2m.c
Quoted 8 lines
xen/arch/x86/mm/p2m.c mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l, · line 520
Quoted 6 lines
+++
b/xen/arch/x86/mm/p2m.cxen/arch/x86/mm/p2m.c +1
-0@@ -517,6 +517,7 @@ mfn_t __get_gfn_type_access(struct p2m_domain *p2m, unsigned long gfn_l,*/ if (mem_sharing_unshare_page (p2m->domain, gfn_l) <0 )mem_sharing_notify_enomem (p2m->domain, gfn_l, false); +This line looks spurious.
Yeap.
Quoted 4 lines
Quoted 4 lines
xen/arch/x86/mm/p2m.c +0
-0mfn = p2m->get_entry (p2m, gfn, t, a, q, page_order,NULL ); }
Quoted 14 lines
xen/common/memory.c int guest_remove_page(struct domain *d, unsigned long gmfn) · line 355
Quoted 10 lines
xen/common/memory.c +1
-1@@ -352,7 +352,7 @@ int guest_remove_page(struct domain *d, unsigned long gmfn)* might be the only one using this shared page, and we need to * trigger proper cleanup. Once done, this is like any other page. */ - rc =mem_sharing_unshare_page (d, gmfn,0 ); + rc =mem_sharing_unshare_page (d, gmfn);AFAICT, this patch does not reduce the number of parameters for
mem_sharing_unshare_page(). Did you intend to make this change in another patch?
Ah yea, it should have been dropped in patch 6 of the series.
Quoted 19 lines
Quoted 19 lines
xen/common/memory.c +0
-0if ( rc ) {mem_sharing_notify_enomem (d, gmfn, false);xen/include/asm-x86/mem_sharing.h +1
-2@@ -76,17 +76,16 @@ struct page_sharing_infounsigned int mem_sharing_get_nr_saved_mfns (void );unsigned int mem_sharing_get_nr_shared_mfns (void ); -#define MEM_SHARING_DESTROY_GFN (1<<1)/* Only fails with -ENOMEM. Enforce it with a BUG_ON wrapper. */ int __mem_sharing_unshare_page (struct domain *d,unsigned long gfn, -uint16_t flags); +bool destroy);
Quoted 12 lines
xen/include/asm-x86/mem_sharing.h struct page_sharing_info · line 91
Quoted 10 lines
xen/include/asm-x86/mem_sharing.h +1
-1static inline int mem_sharing_unshare_page (struct domain *d,unsigned long gfn) { -int rc =__mem_sharing_unshare_page (d, gfn,0 ); +int rc =__mem_sharing_unshare_page (d, gfn, false);BUG_ON (rc && (rc != -ENOMEM ));return rc; }Cheers,
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Julien GrallQuoted 38 lines · Julien Grall
Hi Tamas,
Quoted 34 lines · Tamas K Lengyel
Implement hypercall that allows a fork to shed all memory that got allocated for it during its execution and re-load its vCPU context from the parent VM. This allows the forked VM to reset into the same state the parent VM is in a faster way then creating a new fork would be. Measurements show about a 2x speedup during normal fuzzing operations. Performance may vary depending how much memory got allocated for the forked VM. If it has been completely deduplicated from the parent VM then creating a new fork would likely be more performant.
Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 105 ++++++++++++++++++++++++++++++++++ xen/include/public/memory.h | 1 + 2 files changed, 106 insertions(+)xen/arch/x86/mm/mem_sharing.c +11
-0@@ -1622,6 +1622,87 @@ static int mem_sharing_fork(struct domain *d, struct domain *cd)return 0 ; } +struct gfn_free ; +struct gfn_free { +struct gfn_free *next; +struct page_info *page; +gfn_t gfn; +}; + +static int mem_sharing_fork_reset (struct domain *d,struct domain *cd) +{ +int rc; +
Quoted 8 lines
xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1640
Quoted 5 lines
xen/arch/x86/mm/mem_sharing.c +5
-0+struct p2m_domain * p2m =p2m_get_hostp2m (cd); +struct gfn_free *list =NULL ; +struct page_info *page; + +page_list_for_each (page, &cd->page_list)AFAICT, your domain is not paused, so it would be possible to have page added/remove in that list behind your back.
Well, it's not that it's not paused, it's just that I haven't added a sanity check to make sure it is. The toolstack can (and should) pause it, so that sanity check would be warranted.
Quoted 4 lines
You also have multiple loop on the
page_listin this function. Given the number ofpage_listcan be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.
There is just one loop over page_list itself, the second loop is on the internal list that is being built here which will be a subset. The list itself in fact should be small (in our tests usually <100). Granted the list can grow larger, but in those cases its likely better to just discard the fork and create a new one. So in my opinion adding a hypercall continuation to this not needed.
Quoted 25 lines
Quoted 14 lines
xen/arch/x86/mm/mem_sharing.c +13
-0+ { +mfn_t mfn =page_to_mfn (page); +if (mfn_valid (mfn) ) + { +p2m_type_t p2mt; +p2m_access_t p2ma; +gfn_t gfn =mfn_to_gfn (cd, mfn); + mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &p2mt, &p2ma, +0 ,NULL , false); +if (p2m_is_ram (p2mt) ) + { +struct gfn_free *gfn_free; +if ( !get_page (page, cd) )xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1663
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+ goto err_reset; + +/* +* We can't free the page while iterating over the page_list +* so we build a separate list to loop over. +* +* We want to iterate over the page_list instead of checking +* gfn from 0 to max_gfn because this is ~10x faster. +*/ + gfn_free = xmalloc(struct gfn_free);
Quoted 3 lines
If I did the math right, for a 4G guest this will require at ~24MB of memory. Actually, is it really necessary to do the allocation for a short period of time?
If you have a fully deduplicated fork then you should not be using this function to begin with. You get better performance my throwing that one away and creating a new one. As for using xmalloc here, I'm not sure what other way I have to build a list of pages that need to be freed. I can't free the page itself while I'm iterating on page_list (that I'm aware of). The only other option available is calling __get_gfn_type_access with gfn=0..max_gfn which will be extremely slow because you have to loop over a lot of holes.
Quoted 3 lines
What are you trying to achieve by iterating twice on the GFN? Wouldn't it be easier to pause the domain?
I'm not sure what you mean, where do you see me iterating twice on the gfn? And what does pausing have to do with it?
Than
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
Hi,
Quoted 48 lines · Tamas K Lengyel
Quoted 38 lines · Julien Grall
Hi Tamas,
Quoted 34 lines · Tamas K Lengyel
Implement hypercall that allows a fork to shed all memory that got allocated for it during its execution and re-load its vCPU context from the parent VM. This allows the forked VM to reset into the same state the parent VM is in a faster way then creating a new fork would be. Measurements show about a 2x speedup during normal fuzzing operations. Performance may vary depending how much memory got allocated for the forked VM. If it has been completely deduplicated from the parent VM then creating a new fork would likely be more performant.
Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 105 ++++++++++++++++++++++++++++++++++ xen/include/public/memory.h | 1 + 2 files changed, 106 insertions(+)xen/arch/x86/mm/mem_sharing.c +11
-0@@ -1622,6 +1622,87 @@ static int mem_sharing_fork(struct domain *d, struct domain *cd)return 0 ; } +struct gfn_free ; +struct gfn_free { +struct gfn_free *next; +struct page_info *page; +gfn_t gfn; +}; + +static int mem_sharing_fork_reset (struct domain *d,struct domain *cd) +{ +int rc; +Quoted 8 lines
xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1640
Quoted 5 lines
xen/arch/x86/mm/mem_sharing.c +5
-0+struct p2m_domain * p2m =p2m_get_hostp2m (cd); +struct gfn_free *list =NULL ; +struct page_info *page; + +page_list_for_each (page, &cd->page_list)AFAICT, your domain is not paused, so it would be possible to have page added/remove in that list behind your back.
Quoted 3 lines
Well, it's not that it's not paused, it's just that I haven't added a sanity check to make sure it is. The toolstack can (and should) pause it, so that sanity check would be warranted.
I have only read the hypervisor part, so I didn't know what the toolstack has done.
Quoted 9 lines
Quoted 5 lines
You also have multiple loop on the
page_listin this function. Given the number ofpage_listcan be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.There is just one loop over
page_listitself, the second loop is on the internal list that is being built here which will be a subset. The list itself in fact should be small (in our tests usually <100).
For a first, nothing in this function tells me that there will be only 100 pages. But then, I don't think this is right to implement your hypercall based only the "normal" scenario. You should also think about the "worst" case scenario.
In this case the worst case scenario is have hundreds of page in page_list.
Quoted 3 lines
Granted the list can grow larger, but in those cases its likely better to just discard the fork and create a new one. So in my opinion adding a hypercall continuation to this not needed
How would the caller know it? What would happen if the caller ends up to call this with a growing list.
Quoted 26 lines
Quoted 26 lines
Quoted 25 lines
xen/arch/x86/mm/mem_sharing.c +23
-0+ { +mfn_t mfn =page_to_mfn (page); +if (mfn_valid (mfn) ) + { +p2m_type_t p2mt; +p2m_access_t p2ma; +gfn_t gfn =mfn_to_gfn (cd, mfn); + mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &p2mt, &p2ma, +0 ,NULL , false); +if (p2m_is_ram (p2mt) ) + { +struct gfn_free *gfn_free; +if ( !get_page (page, cd) ) + goto err_reset; + +/* +* We can't free the page while iterating over the page_list +* so we build a separate list to loop over. +* +* We want to iterate over the page_list instead of checking +* gfn from 0 to max_gfn because this is ~10x faster. +*/ + gfn_free = xmalloc(struct gfn_free);
Quoted 7 lines
Quoted 3 lines
If I did the math right, for a 4G guest this will require at ~24MB of memory. Actually, is it really necessary to do the allocation for a short period of time?
If you have a fully deduplicated fork then you should not be using this function to begin with. You get better performance my throwing that one away and creating a new one.
How a user knows when/how this can be called? But then, as said above, this may be called by mistake... So I still think you need to be prepare for the worst case.
Quoted 6 lines
As for using xmalloc here, I'm
not sure what other way I have to build a list of pages that need to be freed. I can't free the page itself while I'm iterating onpage_list(that I'm aware of). The only other option available is calling__get_gfn_type_accesswith gfn=0..max_gfn which will be extremely slow because you have to loop over a lot of holes.
You can use page_list_for_each_safe(). This is already used by function such as relinquish_memory().
Quoted 7 lines
Quoted 4 lines
What are you trying to achieve by iterating twice on the GFN? Wouldn't it be easier to pause the domain?
I'm not sure what you mean, where do you see me iterating twice on the gfn? And what does pausing have to do with it?
It was unclear why you decided to use a double loop here. You explained it above, so this can be discarded.
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
Quoted 52 lines · Julien Grall
Hi,
Quoted 48 lines · Tamas K Lengyel
Quoted 46 lines · Julien Grall
Hi Tamas,
Quoted 29 lines · Tamas K Lengyel
Implement hypercall that allows a fork to shed all memory that got allocated for it during its execution and re-load its vCPU context from the parent VM. This allows the forked VM to reset into the same state the parent VM is in a faster way then creating a new fork would be. Measurements show about a 2x speedup during normal fuzzing operations. Performance may vary depending how much memory got allocated for the forked VM. If it has been completely deduplicated from the parent VM then creating a new fork would likely be more performant.
Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 105 ++++++++++++++++++++++++++++++++++ xen/include/public/memory.h | 1 + 2 files changed, 106 insertions(+)xen/arch/x86/mm/mem_sharing.c +6
-0@@ -1622,6 +1622,87 @@ static int mem_sharing_fork(struct domain *d, struct domain *cd)return 0 ; } +struct gfn_free ; +struct gfn_free { +struct gfn_free *next; +struct page_info *page; +gfn_t gfn; +};xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1640
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+ +static int mem_sharing_fork_reset (struct domain *d,struct domain *cd) +{ +int rc; + +struct p2m_domain * p2m =p2m_get_hostp2m (cd); +struct gfn_free *list =NULL ; +struct page_info *page; + +page_list_for_each (page, &cd->page_list)AFAICT, your domain is not paused, so it would be possible to have page added/remove in that list behind your back.
Quoted 5 lines
Quoted 3 lines
Well, it's not that it's not paused, it's just that I haven't added a sanity check to make sure it is. The toolstack can (and should) pause it, so that sanity check would be warranted.
I have only read the hypervisor part, so I didn't know what the toolstack has done.
I've added the same enforced VM paused operation that is present for the fork hypercall handler.
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines
You also have multiple loop on the
page_listin this function. Given the number ofpage_listcan be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.There is just one loop over
page_listitself, the second loop is on the internal list that is being built here which will be a subset. The list itself in fact should be small (in our tests usually <100).
Quoted 6 lines
For a first, nothing in this function tells me that there will be only 100 pages. But then, I don't think this is right to implement your hypercall based only the "normal" scenario. You should also think about the "worst" case scenario.
In this case the worst case scenario is have hundreds of page in
page_list.
Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
Quoted 7 lines
Quoted 4 lines
Granted the list can grow larger, but in those cases its likely better to just discard the fork and create a new one. So in my opinion adding a hypercall continuation to this not needed
How would the caller know it? What would happen if the caller ends up to call this with a growing list.
The caller knows by virtue of knowing how long the VM was executed for. In the usecase this is targeted at the VM was executing only for a couple seconds at most. Usually much less then that (we get about ~80 resets/s with AFL). During that time its extremely unlikely you get more then a ~100 pages deduplicated (that is, written to). But even if there are more pages, it just means the hypercall might take a bit longer to run for that iteration. I don't see any issue with not breaking up this hypercall with continuation even under the worst case situation though. But if others feel that strongly as well about having to have continuation for this I don't really mind adding it.
Quoted 31 lines
Quoted 31 lines
Quoted 30 lines
Quoted 16 lines
xen/arch/x86/mm/mem_sharing.c +13
-0+ { +mfn_t mfn =page_to_mfn (page); +if (mfn_valid (mfn) ) + { +p2m_type_t p2mt; +p2m_access_t p2ma; +gfn_t gfn =mfn_to_gfn (cd, mfn); + mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &p2mt, &p2ma, +0 ,NULL , false); +if (p2m_is_ram (p2mt) ) + { +struct gfn_free *gfn_free; +if ( !get_page (page, cd) )xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1663
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+ goto err_reset; + +/* +* We can't free the page while iterating over the page_list +* so we build a separate list to loop over. +* +* We want to iterate over the page_list instead of checking +* gfn from 0 to max_gfn because this is ~10x faster. +*/ + gfn_free = xmalloc(struct gfn_free);If I did the math right, for a 4G guest this will require at ~24MB of memory. Actually, is it really necessary to do the allocation for a short period of time?
Quoted 7 lines
Quoted 3 lines
If you have a fully deduplicated fork then you should not be using this function to begin with. You get better performance my throwing that one away and creating a new one.
How a user knows when/how this can be called? But then, as said above, this may be called by mistake... So I still think you need to be prepare for the worst case.
See my answer above.
Quoted 9 lines
Quoted 7 lines
As for using xmalloc here, I'm
not sure what other way I have to build a list of pages that need to be freed. I can't free the page itself while I'm iterating onpage_list(that I'm aware of). The only other option available is calling__get_gfn_type_accesswith gfn=0..max_gfn which will be extremely slow because you have to loop over a lot of holes.You can use
page_list_for_each_safe(). This is already used by function such asrelinquish_memory().
Neat, will check it out! Would certainly simplify things not having to build a private list.
Quoted 11 lines
Quoted 8 lines
Quoted 5 lines
What are you trying to achieve by iterating twice on the GFN? Wouldn't it be easier to pause the domain?
I'm not sure what you mean, where do you see me iterating twice on the gfn? And what does pausing have to do with it?
It was unclear why you decided to use a double loop here. You explained it above, so this can be discarded.
OK, cool.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
Hi Tamas,
Quoted 78 lines · Tamas K Lengyel
Quoted 52 lines · Julien Grall
Hi,
Quoted 48 lines · Tamas K Lengyel
Quoted 46 lines · Julien Grall
Hi Tamas,
Quoted 29 lines · Tamas K Lengyel
Implement hypercall that allows a fork to shed all memory that got allocated for it during its execution and re-load its vCPU context from the parent VM. This allows the forked VM to reset into the same state the parent VM is in a faster way then creating a new fork would be. Measurements show about a 2x speedup during normal fuzzing operations. Performance may vary depending how much memory got allocated for the forked VM. If it has been completely deduplicated from the parent VM then creating a new fork would likely be more performant.
Signed-off-by Tamas K Lengyel <[email protected]>--- xen/arch/x86/mm/mem_sharing.c | 105 ++++++++++++++++++++++++++++++++++ xen/include/public/memory.h | 1 + 2 files changed, 106 insertions(+)xen/arch/x86/mm/mem_sharing.c +6
-0@@ -1622,6 +1622,87 @@ static int mem_sharing_fork(struct domain *d, struct domain *cd)return 0 ; } +struct gfn_free ; +struct gfn_free { +struct gfn_free *next; +struct page_info *page; +gfn_t gfn; +};xen/arch/x86/mm/mem_sharing.c static int mem_sharing_fork(struct domain *d, struct domain *cd) · line 1640
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +10
-0+ +static int mem_sharing_fork_reset (struct domain *d,struct domain *cd) +{ +int rc; + +struct p2m_domain * p2m =p2m_get_hostp2m (cd); +struct gfn_free *list =NULL ; +struct page_info *page; + +page_list_for_each (page, &cd->page_list)AFAICT, your domain is not paused, so it would be possible to have page added/remove in that list behind your back.
Quoted 5 lines
Quoted 3 lines
Well, it's not that it's not paused, it's just that I haven't added a sanity check to make sure it is. The toolstack can (and should) pause it, so that sanity check would be warranted.
I have only read the hypervisor part, so I didn't know what the toolstack has done.
I've added the same enforced VM paused operation that is present for the fork hypercall handler.
Quoted 16 lines
Quoted 10 lines
Quoted 6 lines
You also have multiple loop on the
page_listin this function. Given the number ofpage_listcan be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.There is just one loop over
page_listitself, the second loop is on the internal list that is being built here which will be a subset. The list itself in fact should be small (in our tests usually <100).For a first, nothing in this function tells me that there will be only 100 pages. But then, I don't think this is right to implement your hypercall based only the "normal" scenario. You should also think about the "worst" case scenario.
Quoted 5 lines
Quoted 1 line
In this case the worst case scenario is have hundreds of page in
page_list.Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
I assume that if you submit to upstream this new hypercall then there is longer plan to have more people to use it and potentially making "stable". If not, then it raises the question why this is pushed upstream...
In any case, all the known assumptions should be documented so they can be fixed rather than forgotten until it is rediscovered via an XSA.
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
Granted the list can grow larger, but in those cases its likely better to just discard the fork and create a new one. So in my opinion adding a hypercall continuation to this not needed
How would the caller know it? What would happen if the caller ends up to call this with a growing list.
Quoted 7 lines
The caller knows by virtue of knowing how long the VM was executed for. In the usecase this is targeted at the VM was executing only for a couple seconds at most. Usually much less then that (we get about ~80 resets/s with AFL). During that time its extremely unlikely you get more then a ~100 pages deduplicated (that is, written to). But even if there are more pages, it just means the hypercall might take a bit longer to run for that iteration.
I assume if you upstream the code then you want more people to use it (otherwise what's the point?). In this case, you will likely have people that heard about the feature, wants to test but don't know the internal.
Such users need to know how this can be call safely without reading the implementation. In other words, some documentation for your hypercall is needed.
Quoted 3 lines
I don't see any issue with not
breaking up this hypercall with continuation even under the worst case situation though.
Xen only supports voluntary preemption, this means that an hypercall can only be preempted if there is code for it. Otherwise the preemption will mostly only happen when returning to the guest.
In other words, the vCPU executing the hypercall may go past its timeslice and prevent other vCPU to run.
There are other problem with long running hypercalls. Anyway, in short, if you ever want to get you code supported then you will need the hypercall to be broken down.
Quoted 2 lines
But if others feel that strongly as well about
having to have continuation for this I don't really mind adding it.
I don't think the continuation work is going to be difficult, but if you want to delay it, then the minimum is to document such assumptions for your users.
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 18/20] xen/mem_access: Use __get_gfn_type_access in set_mem_access
In reply to Tamas K LengyelQuoted 15 lines · Tamas K Lengyel
Use
__get_gfn_type_accessinstead ofp2m->get_entryto trigger page-forking when themem_accesspermission is being set on a page that has not yet been copied over from the parent.Signed-off-by Tamas K Lengyel <[email protected]: Alexandru Isaila <[email protected]>--- xen/arch/x86/mm/mem_access.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)xen/arch/x86/mm/mem_access.c +0
-0
xen/arch/x86/mm/mem_access.c static int set_mem_access(struct domain *d, struct p2m_domain *p2m, · line 309
Quoted 10 lines
xen/arch/x86/mm/mem_access.c +2
-3@@ -303,11 +303,10 @@ static int set_mem_access(struct domain *d, struct p2m_domain *p2m,ASSERT (!ap2m);#endif { -mfn_t mfn;p2m_access_t _a;p2m_type_t t; - - mfn = p2m->get_entry (p2m, gfn, &t, &_a,0 ,NULL ,NULL ); +mfn_t mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &t, &_a, +P2M_ALLOC ,NULL , false);
Don't you want p2m_query_t to be 0 as it was in the p2m->get_entry() call ?
xen/arch/x86/mm/mem_access.c static int set_mem_access(struct domain *d, struct p2m_domain *p2m, · line 312
Quoted 4 lines
xen/arch/x86/mm/mem_access.c +0
-0rc = p2m->set_entry (p2m, gfn, mfn,PAGE_ORDER_4K , t, a,-1 ); }
Alex
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 12 lines · Tamas K Lengyel
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.
Quoted 3 lines
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking
^ than
Quoted 9 lines
operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Quoted 7 lines
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.
IMO having two different commands is confusing for the end user, I would rather have something like:
xl fork-vm [-d] ...
Where '-d' would prevent forking any user-space emulators. I don't thinks there's a need for a separate command to fork the underlying user-space emulators.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Tamas K LengyelQuoted 16 lines · Tamas K Lengyel
Quoted 15 lines · Julien Grall
Quoted 8 lines · Tamas K Lengyel
Quoted 3 lines · Julien Grall
You also have multiple loop on the
page_listin this function. Given the number ofpage_listcan be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.There is just one loop over
page_listitself, the second loop is on the internal list that is being built here which will be a subset. The list itself in fact should be small (in our tests usually <100).For a first, nothing in this function tells me that there will be only 100 pages. But then, I don't think this is right to implement your hypercall based only the "normal" scenario. You should also think about the "worst" case scenario.
Quoted 5 lines
Quoted 1 line
In this case the worst case scenario is have hundreds of page in
page_list.Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
FWIW I'm with Julien here: The preferred course of action is to make the operation safe against abuse. The minimum requirement is to document obvious missing pieces for this to become supported code.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 04/20] x86/mem_sharing: cleanup code and comments in various locations
In reply to Tamas K Lengyelxen/arch/x86/hvm/hvm.c int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla, · line 1905
Quoted 11 lines · Tamas K Lengyel
xen/arch/x86/hvm/hvm.c +1
-2@@ -1902,12 +1902,11 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,if ( npfec.write_access && (p2mt == p2m_ram_shared) ) {ASSERT (p2m_is_hostp2m (p2m)); - sharing_enomem = - (mem_sharing_unshare_page (currd, gfn,0 ) <0 ); + sharing_enomem =mem_sharing_unshare_page (currd, gfn,0 );
This is a logical change. Is it intended to be in a later patch?
Quoted 25 lines
xen/arch/x86/mm/mem_sharing.c +7
-4@@ -198,24 +200,26 @@ static inline shr_handle_t get_next_handle(void)#define mem_sharing_enabled (d) \ (is_hvm_domain(d) && (d)->arch.hvm.mem_sharing_enabled) -static atomic_t nr_saved_mfns =ATOMIC_INIT (0 ); +static atomic_t nr_saved_mfns =ATOMIC_INIT (0 );static atomic_t nr_shared_mfns =ATOMIC_INIT (0 ); -/** Reverse map **/ -/* Every shared frame keeps a reverse map (rmap) of <domain, gfn> tuples that +/* +* Reverse map +* +* Every shared frame keeps a reverse map (rmap) of <domain, gfn> tuples that * this shared frame backs. For pages with a low degree of sharing, a O(n) * search linked list is good enough. For pages with higher degree of sharing, -* we use a hash table instead. */ +* we use a hash table instead. +*/ typedef struct gfn_info
xen/arch/x86/mm/mem_sharing.c static inline shr_handle_t get_next_handle(void) · line 222
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +2
-3{unsigned long gfn; -domid_t domain; +domid_t domain;struct list_head list; } gfn_info_t; -static inline void -rmap_init (struct page_info *page) +static inline void rmap_init (struct page_info *page)
Seeing as you're folding this, the inline can be dropped. In .c files, functions should just be static.
xen/arch/x86/mm/mem_sharing.c static inline void mem_sharing_gfn_destroy(struct page_info *page, · line 444
Quoted 6 lines
xen/arch/x86/mm/mem_sharing.c +1
-1@@ -437,25 +441,29 @@ static inline void mem_sharing_gfn_destroy(struct page_info *page,xfree (gfn_info); } -static struct page_info *mem_sharing_lookup (unsigned long mfn) +static inline struct page_info *mem_sharing_lookup (unsigned long mfn)
Seeing as this is cleanup, the position of the * can move. Similarly, it shouldn't gain an inline.
I can fix all of this up on commit (and a few other brace position issues) if you want, to save reworking a trivial v2.
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 35 lines · Roger Pau Monné
Quoted 12 lines · Tamas K Lengyel
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.Quoted 3 lines
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking
^ than
Quoted 9 lines
operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Quoted 7 lines
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.
Quoted 8 lines
IMO having two different commands is confusing for the end user, I would rather have something like:
xl fork-vm [-d] ...
Where '-d' would prevent forking any user-space emulators. I don't thinks there's a need for a separate command to fork the underlying user-space emulators.
Keeping it as two commands allows you to start up the fork and let it run immediately and only start up QEMU when you notice it is needed. The idea being that you can monitor the kernel and see when it tries to do some I/O that would require the QEMU backend. If you combine the commands that option goes away. Also, QEMU itself isn't getting forked right now, we just start a new QEMU process with the saved-state getting loaded into it. I looked into implementing a QEMU fork command but it turns out that for the vast majority of our use-cases QEMU isn't needed at all, so developing that addition was tabled.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Jan BeulichQuoted 20 lines · Jan Beulich
Quoted 18 lines · Tamas K Lengyel
Quoted 16 lines · Julien Grall
Quoted 8 lines · Tamas K Lengyel
Quoted 3 lines · Julien Grall
You also have multiple loop on the
page_listin this function. Given the number ofpage_listcan be quite big, this is a call for hogging the pCPU and an RCU lock on the domain vCPU running this call.There is just one loop over
page_listitself, the second loop is on the internal list that is being built here which will be a subset. The list itself in fact should be small (in our tests usually <100).For a first, nothing in this function tells me that there will be only 100 pages. But then, I don't think this is right to implement your hypercall based only the "normal" scenario. You should also think about the "worst" case scenario.
In this case the worst case scenario is have hundreds of page in
page_list.
Quoted 7 lines
Quoted 3 lines
Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
FWIW I'm with Julien here: The preferred course of action is to make the operation safe against abuse. The minimum requirement is to document obvious missing pieces for this to become supported code.
That's perfectly fine by me.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Julien GrallQuoted 7 lines
Quoted 3 lines
Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
I assume that if you submit to upstream this new hypercall then there is longer plan to have more people to use it and potentially making "stable". If not, then it raises the question why this is pushed upstream...
It's being pushed upstream so other people can make use of it, even if it's not "production quality". Beyond what is being sent here there are no longer term plans from Intel (at this point) to support this in any way. The alternative would be that we just release a fork (or just the patches) and walk away. If the Xen community wants to make the announcement that only code that will have long term support and is "stable" is accepted upstream that's IMHO drastically going to reduce people's interest to share anything.
Quoted 2 lines
In any case, all the known assumptions should be documented so they can be fixed rather than forgotten until it is rediscovered via an XSA.
Fair enough.
Quoted 17 lines
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
Granted the list can grow larger, but in those cases its likely better to just discard the fork and create a new one. So in my opinion adding a hypercall continuation to this not needed
How would the caller know it? What would happen if the caller ends up to call this with a growing list.
Quoted 7 lines
The caller knows by virtue of knowing how long the VM was executed for. In the usecase this is targeted at the VM was executing only for a couple seconds at most. Usually much less then that (we get about ~80 resets/s with AFL). During that time its extremely unlikely you get more then a ~100 pages deduplicated (that is, written to). But even if there are more pages, it just means the hypercall might take a bit longer to run for that iteration.
Quoted 7 lines
I assume if you upstream the code then you want more people to use it (otherwise what's the point?). In this case, you will likely have people that heard about the feature, wants to test but don't know the internal.
Such users need to know how this can be call safely without reading the implementation. In other words, some documentation for your hypercall is needed.
Sure.
Quoted 16 lines
Quoted 4 lines
I don't see any issue with not
breaking up this hypercall with continuation even under the worst case situation though.Xen only supports voluntary preemption, this means that an hypercall can only be preempted if there is code for it. Otherwise the preemption will mostly only happen when returning to the guest.
In other words, the vCPU executing the hypercall may go past its timeslice and prevent other vCPU to run.
There are other problem with long running hypercalls. Anyway, in short, if you ever want to get you code supported then you will need the hypercall to be broken down.
Quoted 6 lines
Quoted 2 lines
But if others feel that strongly as well about
having to have continuation for this I don't really mind adding it.I don't think the continuation work is going to be difficult, but if you want to delay it, then the minimum is to document such assumptions for your users.
I just don't see a use for it because it will never actually execute. So to me it just looks like unnecessary dead glue. But documenting the assumption under which this hypercall should execute is perfectly fair.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
Hi,
Quoted 4 lines · Tamas K Lengyel
Quoted 4 lines
Quoted 3 lines
Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
Quoted 8 lines
Quoted 3 lines
I assume that if you submit to upstream this new hypercall then there is longer plan to have more people to use it and potentially making "stable". If not, then it raises the question why this is pushed upstream...
It's being pushed upstream so other people can make use of it, even if it's not "production quality". Beyond what is being sent here there are no longer term plans from Intel (at this point) to support this in any way.
So if this is merged, who is going to maintain it?
Quoted 6 lines
The alternative would be that we just release a fork (or just the patches) and walk away.
If the Xen community wants to make the
announcement that only code that will have long term support and is "stable" is accepted upstream that's IMHO drastically going to reduce people's interest to share anything.
Sharing is one thing, but if this code is not at least a minimum maintained then it is likely the code will not be functional in a year time.
Don't get me wrong, this is a cool feature to have but you make it sounds like this is going to be dumped in Xen and never going to be touched again. How is this going to be beneficial for the community?
Quoted 49 lines
Quoted 3 lines
In any case, all the known assumptions should be documented so they can be fixed rather than forgotten until it is rediscovered via an XSA.
Fair enough.
Quoted 17 lines
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
Granted the list can grow larger, but in those cases its likely better to just discard the fork and create a new one. So in my opinion adding a hypercall continuation to this not needed
How would the caller know it? What would happen if the caller ends up to call this with a growing list.
Quoted 7 lines
The caller knows by virtue of knowing how long the VM was executed for. In the usecase this is targeted at the VM was executing only for a couple seconds at most. Usually much less then that (we get about ~80 resets/s with AFL). During that time its extremely unlikely you get more then a ~100 pages deduplicated (that is, written to). But even if there are more pages, it just means the hypercall might take a bit longer to run for that iteration.
Quoted 7 lines
I assume if you upstream the code then you want more people to use it (otherwise what's the point?). In this case, you will likely have people that heard about the feature, wants to test but don't know the internal.
Such users need to know how this can be call safely without reading the implementation. In other words, some documentation for your hypercall is needed.
Sure.
Quoted 16 lines
Quoted 4 lines
I don't see any issue with not
breaking up this hypercall with continuation even under the worst case situation though.Xen only supports voluntary preemption, this means that an hypercall can only be preempted if there is code for it. Otherwise the preemption will mostly only happen when returning to the guest.
In other words, the vCPU executing the hypercall may go past its timeslice and prevent other vCPU to run.
There are other problem with long running hypercalls. Anyway, in short, if you ever want to get you code supported then you will need the hypercall to be broken down.
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
But if others feel that strongly as well about
having to have continuation for this I don't really mind adding it.I don't think the continuation work is going to be difficult, but if you want to delay it, then the minimum is to document such assumptions for your users.
I just don't see a use for it because it will never actually execute.
That's a very narrow view of how your hypercall can be used. I know that you said that should only be called early, but imagine for a moment the user decide to call it much later in the fork process.
Quoted 1 line
So to me it just looks like unnecessary dead glue.
Try to call the hypercall after enough deduplication happen (maybe 20min). Alternatively, give me access to your machine with the code and I can show how it can be misused ;).
Quoted 3 lines
But documenting the
assumption under which this hypercall should execute is perfectly fair.
You might want to think how the interface would look like with the preemption. So the day you decide to introduce preemption you don't have to create a new hypercall.
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 18/20] xen/mem_access: Use __get_gfn_type_access in set_mem_access
In reply to Alexandru Stefan ISAILAQuoted 23 lines · Alexandru Stefan ISAILA
Quoted 19 lines · Tamas K Lengyel
Use
__get_gfn_type_accessinstead ofp2m->get_entryto trigger page-forking when themem_accesspermission is being set on a page that has not yet been copied over from the parent.Signed-off-by Tamas K Lengyel <[email protected]: Alexandru Isaila <[email protected]>--- xen/arch/x86/mm/mem_access.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)xen/arch/x86/mm/mem_access.c +0
-1@@ -303,11 +303,10 @@ static int set_mem_access(struct domain *d, struct p2m_domain *p2m,ASSERT (!ap2m);#endif { -mfn_t mfn;
Quoted 8 lines
xen/arch/x86/mm/mem_access.c static int set_mem_access(struct domain *d, struct p2m_domain *p2m, · line 309
Quoted 6 lines
xen/arch/x86/mm/mem_access.c +2
-2p2m_access_t _a;p2m_type_t t; - - mfn = p2m->get_entry (p2m, gfn, &t, &_a,0 ,NULL ,NULL ); +mfn_t mfn =__get_gfn_type_access (p2m,gfn_x (gfn), &t, &_a, +P2M_ALLOC ,NULL , false);Don't you want
p2m_query_tto be 0 as it was in thep2m->get_entry()call ?
No, the entire point of the patch is to have the P2M_ALLOC query flag set. That triggers the fork's p2m population.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 04/20] x86/mem_sharing: cleanup code and comments in various locations
In reply to Andrew CooperQuoted 59 lines · Andrew Cooper
xen/arch/x86/hvm/hvm.c int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla, · line 1905
Quoted 11 lines · Tamas K Lengyel
xen/arch/x86/hvm/hvm.c +1
-2@@ -1902,12 +1902,11 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,if ( npfec.write_access && (p2mt == p2m_ram_shared) ) {ASSERT (p2m_is_hostp2m (p2m)); - sharing_enomem = - (mem_sharing_unshare_page (currd, gfn,0 ) <0 ); + sharing_enomem =mem_sharing_unshare_page (currd, gfn,0 );This is a logical change. Is it intended to be in a later patch?
Quoted 25 lines
xen/arch/x86/mm/mem_sharing.c +7
-4@@ -198,24 +200,26 @@ static inline shr_handle_t get_next_handle(void)#define mem_sharing_enabled (d) \ (is_hvm_domain(d) && (d)->arch.hvm.mem_sharing_enabled) -static atomic_t nr_saved_mfns =ATOMIC_INIT (0 ); +static atomic_t nr_saved_mfns =ATOMIC_INIT (0 );static atomic_t nr_shared_mfns =ATOMIC_INIT (0 ); -/** Reverse map **/ -/* Every shared frame keeps a reverse map (rmap) of <domain, gfn> tuples that +/* +* Reverse map +* +* Every shared frame keeps a reverse map (rmap) of <domain, gfn> tuples that * this shared frame backs. For pages with a low degree of sharing, a O(n) * search linked list is good enough. For pages with higher degree of sharing, -* we use a hash table instead. */ +* we use a hash table instead. +*/ typedef struct gfn_infoxen/arch/x86/mm/mem_sharing.c static inline shr_handle_t get_next_handle(void) · line 222
Quoted 10 lines
xen/arch/x86/mm/mem_sharing.c +2
-3{unsigned long gfn; -domid_t domain; +domid_t domain;struct list_head list; } gfn_info_t; -static inline void -rmap_init (struct page_info *page) +static inline void rmap_init (struct page_info *page)Seeing as you're folding this, the inline can be dropped. In .c files, functions should just be static.
Quoted 4 lines
xen/arch/x86/mm/mem_sharing.c +0
-0@@ -437,25 +441,29 @@ static inline void mem_sharing_gfn_destroy(struct page_info *page,xfree (gfn_info); }
Quoted 9 lines
xen/arch/x86/mm/mem_sharing.c static inline void mem_sharing_gfn_destroy(struct page_info *page, · line 444
Quoted 2 lines
xen/arch/x86/mm/mem_sharing.c +1
-1-static struct page_info *mem_sharing_lookup (unsigned long mfn) +static inline struct page_info *mem_sharing_lookup (unsigned long mfn)Seeing as this is cleanup, the position of the * can move. Similarly, it shouldn't gain an inline.
I can fix all of this up on commit (and a few other brace position issues) if you want, to save reworking a trivial v2.
Provided nothing else is outstanding with the patch and it can be committed I would certainly appreciate that.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Tamas K LengyelQuoted 32 lines · Tamas K Lengyel
Quoted 12 lines · Julien Grall
Hi,
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines
Quoted 3 lines
Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
I assume that if you submit to upstream this new hypercall then there is longer plan to have more people to use it and potentially making "stable". If not, then it raises the question why this is pushed upstream...
Quoted 6 lines
Quoted 4 lines
It's being pushed upstream so other people can make use of it, even if it's not "production quality". Beyond what is being sent here there are no longer term plans from Intel (at this point) to support this in any way.
So if this is merged, who is going to maintain it?
It falls under
mem_sharingfor which I'm listed as "Odd Fixes" maintainer, which I do in my personal free time. The status there isn't changing with this new feature.Quoted 8 lines
Quoted 7 lines
The alternative would be that we just release a fork (or just the patches) and walk away.
If the Xen community wants to make the
announcement that only code that will have long term support and is "stable" is accepted upstream that's IMHO drastically going to reduce people's interest to share anything.
Quoted 5 lines
Quoted 2 lines
Sharing is one thing, but if this code is not at least a minimum maintained then it is likely the code will not be functional in a year time.
Surprisingly
mem_sharinghad only minor bitrots in the last ~5 years in which time it has been pretty much abandoned.
This falls under a "minimum maintained". This wasn't clear from your previous statement stating there will be no support.
[...]
Quoted 20 lines
Quoted 14 lines
Quoted 9 lines
Quoted 7 lines
Quoted 3 lines
But if others feel that strongly as well about
having to have continuation for this I don't really mind adding it.I don't think the continuation work is going to be difficult, but if you want to delay it, then the minimum is to document such assumptions for your users.
I just don't see a use for it because it will never actually execute.
That's a very narrow view of how your hypercall can be used. I know that you said that should only be called early, but imagine for a moment the user decide to call it much later in the fork process.
Quoted 5 lines
Quoted 1 line
So to me it just looks like unnecessary dead glue.
Try to call the hypercall after enough deduplication happen (maybe 20min). Alternatively, give me access to your machine with the code and I can show how it can be misused ;).
Quoted 5 lines
It will hang for a bit for sure and Linux in dom0 will complain that a CPU is stuck. But it will eventually finish. It's not like it's doing all that much. And anyway, if you notice that happening when you call it it will be an obvious clue that you shouldn't be using it under the situation you are using it under. Having continuation would hide that.
I am not going to argue more as this is an experimental feature. But this will be a showstopper if we ever consider mem_sharing to be supported (or even security supported).
Meanwhile please document the assumption.
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
But documenting the
assumption under which this hypercall should execute is perfectly fair.
Quoted 7 lines
Quoted 3 lines
You might want to think how the interface would look like with the preemption. So the day you decide to introduce preemption you don't have to create a new hypercall.
Why would you need to introduce a new hypercall if preemption becomes necessary? This is not a stable interfaces. It can be removed/changed completely from one Xen release to the next.
Sorry, I didn't realize the mem_sharing code was not a stable interfaces.
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Julien GrallQuoted 12 lines · Julien Grall
Hi,
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines
Quoted 3 lines
Well, this is only an experimental system that's completely disabled by default. Making the assumption that people who make use of it will know what they are doing I think is fair.
I assume that if you submit to upstream this new hypercall then there is longer plan to have more people to use it and potentially making "stable". If not, then it raises the question why this is pushed upstream...
Quoted 6 lines
Quoted 4 lines
It's being pushed upstream so other people can make use of it, even if it's not "production quality". Beyond what is being sent here there are no longer term plans from Intel (at this point) to support this in any way.
So if this is merged, who is going to maintain it?
It falls under mem_sharing for which I'm listed as "Odd Fixes" maintainer, which I do in my personal free time. The status there isn't changing with this new feature.
Quoted 10 lines
Quoted 7 lines
The alternative would be that we just release a fork (or just the patches) and walk away.
If the Xen community wants to make the
announcement that only code that will have long term support and is "stable" is accepted upstream that's IMHO drastically going to reduce people's interest to share anything.Sharing is one thing, but if this code is not at least a minimum maintained then it is likely the code will not be functional in a year time.
Surprisingly mem_sharing had only minor bitrots in the last ~5 years in which time it has been pretty much abandoned.
Quoted 4 lines
Don't get me wrong, this is a cool feature to have but you make it sounds like this is going to be dumped in Xen and never going to be touched again. How is this going to be beneficial for the community?
It adds an experimental feature to Xen that people can try out and well experiment with! It may be useful, it may not be. You yourself said that this is a cool feature. The fact that there is a JIRA ticket tracking this also shows there is community interest in having such a feature available. If down the road that changes and this proves to be dead code, it can be removed. I don't think that will be necessary since this isn't even compiled by default anymore though. But anyway, it makes more sense to get it upstream then carry it out of tree because it gets more exposure that way, more people may try it out. Having it in-tree also means that in a couple releases people who are interested in the feature don't have to go through the process of rebasing the patchset on newer versions of Xen since it's already in-tree.
Quoted 13 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
But if others feel that strongly as well about
having to have continuation for this I don't really mind adding it.I don't think the continuation work is going to be difficult, but if you want to delay it, then the minimum is to document such assumptions for your users.
I just don't see a use for it because it will never actually execute.
That's a very narrow view of how your hypercall can be used. I know that you said that should only be called early, but imagine for a moment the user decide to call it much later in the fork process.
Quoted 5 lines
Quoted 1 line
So to me it just looks like unnecessary dead glue.
Try to call the hypercall after enough deduplication happen (maybe 20min). Alternatively, give me access to your machine with the code and I can show how it can be misused ;).
It will hang for a bit for sure and Linux in dom0 will complain that a CPU is stuck. But it will eventually finish. It's not like it's doing all that much. And anyway, if you notice that happening when you call it it will be an obvious clue that you shouldn't be using it under the situation you are using it under. Having continuation would hide that.
Quoted 8 lines
Quoted 4 lines
But documenting the
assumption under which this hypercall should execute is perfectly fair.You might want to think how the interface would look like with the preemption. So the day you decide to introduce preemption you don't have to create a new hypercall.
Why would you need to introduce a new hypercall if preemption becomes necessary? This is not a stable interfaces. It can be removed/changed completely from one Xen release to the next.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 19/20] x86/mem_sharing: reset a fork
In reply to Julien GrallQuoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
The alternative would be that we just release a fork (or just the patches) and walk away.
If the Xen community wants to make the
announcement that only code that will have long term support and is "stable" is accepted upstream that's IMHO drastically going to reduce people's interest to share anything.
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Sharing is one thing, but if this code is not at least a minimum maintained then it is likely the code will not be functional in a year time.
Surprisingly
mem_sharinghad only minor bitrots in the last ~5 years in which time it has been pretty much abandoned.This falls under a "minimum maintained". This wasn't clear from your previous statement stating there will be no support.
Sure, I meant there is no support from Intel (ie. it's not part of my job-description nor do I get payed to support this long-term). I usually check during the RC test days that it's at least functional by doing some testing manually. But it's pretty ad-hoc when and if I do that (hence the Odd fixes status).
Quoted 26 lines
Quoted 20 lines
Quoted 14 lines
Quoted 9 lines
Quoted 7 lines
Quoted 3 lines
But if others feel that strongly as well about
having to have continuation for this I don't really mind adding it.I don't think the continuation work is going to be difficult, but if you want to delay it, then the minimum is to document such assumptions for your users.
I just don't see a use for it because it will never actually execute.
That's a very narrow view of how your hypercall can be used. I know that you said that should only be called early, but imagine for a moment the user decide to call it much later in the fork process.
Quoted 5 lines
Quoted 1 line
So to me it just looks like unnecessary dead glue.
Try to call the hypercall after enough deduplication happen (maybe 20min). Alternatively, give me access to your machine with the code and I can show how it can be misused ;).
Quoted 5 lines
It will hang for a bit for sure and Linux in dom0 will complain that a CPU is stuck. But it will eventually finish. It's not like it's doing all that much. And anyway, if you notice that happening when you call it it will be an obvious clue that you shouldn't be using it under the situation you are using it under. Having continuation would hide that.
Quoted 5 lines
I am not going to argue more as this is an experimental feature. But this will be a showstopper if we ever consider
mem_sharingto be supported (or even security supported).Meanwhile please document the assumption.
Ack, already did.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 04/20] x86/mem_sharing: cleanup code and comments in various locations
In reply to Andrew CooperQuoted 15 lines · Andrew Cooper
xen/arch/x86/hvm/hvm.c int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla, · line 1905
Quoted 11 lines · Tamas K Lengyel
xen/arch/x86/hvm/hvm.c +1
-2@@ -1902,12 +1902,11 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,if ( npfec.write_access && (p2mt == p2m_ram_shared) ) {ASSERT (p2m_is_hostp2m (p2m)); - sharing_enomem = - (mem_sharing_unshare_page (currd, gfn,0 ) <0 ); + sharing_enomem =mem_sharing_unshare_page (currd, gfn,0 );This is a logical change. Is it intended to be in a later patch?
While it may look like one it's actually not. The variable sharing_enomem is declared as an int and the function only has two possible return values, 0 and -ENOMEM.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 04/20] x86/mem_sharing: cleanup code and comments in various locations
In reply to Tamas K LengyelQuoted 9 lines · Tamas K Lengyel
Quoted 9 lines
xen/arch/x86/mm/mem_sharing.c static inline void mem_sharing_gfn_destroy(struct page_info *page, · line 444
Quoted 6 lines
xen/arch/x86/mm/mem_sharing.c +1
-1@@ -437,25 +441,29 @@ static inline void mem_sharing_gfn_destroy(struct page_info *page,xfree (gfn_info); } -static struct page_info *mem_sharing_lookup (unsigned long mfn) +static inline struct page_info *mem_sharing_lookup (unsigned long mfn)Seeing as this is cleanup, the position of the * can move. Similarly, it shouldn't gain an inline.
Quoted 5 lines
Quoted 3 lines
I can fix all of this up on commit (and a few other brace position issues) if you want, to save reworking a trivial v2.
Provided nothing else is outstanding with the patch and it can be committed I would certainly appreciate that.
I've pushed this and the previous patch, with some further fixups that I spotted.
Hopefully the rebase won't be an issue.
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
In reply to Tamas K Lengyelxen/arch/x86/hvm/hvm.c static int hvmop_set_evtchn_upcall_vector( · line 4076
Quoted 11 lines · Tamas K Lengyel
xen/arch/x86/hvm/hvm.c +2
-1@@ -4072,16 +4072,17 @@ static int hvmop_set_evtchn_upcall_vector( }static int hvm_allow_set_param (struct domain *d, -const struct xen_hvm_param *a) +uint32_t index, +uint64_t new_value)
These two lines can be folded together.
xen/arch/x86/hvm/hvm.c static int hvm_allow_set_param(struct domain *d, · line 4139
Quoted 7 lines
xen/arch/x86/hvm/hvm.c +1
-1@@ -4134,13 +4135,11 @@ static int hvm_allow_set_param(struct domain *d,return rc; } -static int hvmop_set_param ( +int hvmop_set_param (XEN_GUEST_HANDLE_PARAM (xen_hvm_param_t) arg)
and here.
Quoted 6 lines
xen/arch/x86/hvm/hvm.c +1
-1@@ -4160,23 +4159,42 @@ static int hvmop_set_param(if ( !is_hvm_domain (d) ) goto out; - rc =hvm_allow_set_param (d, &a); + rc =hvm_set_param (d, a.index, a.value);
xen/arch/x86/hvm/hvm.c static int hvmop_set_param( · line 4172
Quoted 10 lines
xen/arch/x86/hvm/hvm.c +10
-0+ + out: +rcu_unlock_domain (d); +return rc; +} + +int hvm_set_param ( +struct domain *d, +uint32_t index, +uint64_t value)
and again.
xen/arch/x86/hvm/hvm.c static int hvmop_set_param( · line 4365
Quoted 11 lines
xen/arch/x86/hvm/hvm.c +2
-2@@ -4340,34 +4358,33 @@ static int hvmop_set_param(* 256 bits interrupt redirection bitmap + 64k bits I/O bitmap * plus one padding byte). */ -if ( (a.value >>32 ) >sizeof (struct tss32 ) + +if ( (value >>32 ) >sizeof (struct tss32 ) + (0x100 /8 ) + (0x10000 /8 ) +1 ) - a.value = (uint32_t )a.value | + value = (uint32_t )value | ((sizeof (struct tss32 ) + (0x100 /8 ) + (0x10000 /8 ) +1 ) <<32 );
Can you reindent the surrounding lines so they line up again?
Quoted 14 lines
xen/arch/x86/hvm/hvm.c +5
-2@@ -4429,42 +4446,60 @@ static int hvmop_get_param(if ( !is_hvm_domain (d) ) goto out; - rc =hvm_allow_get_param (d, &a); + rc =hvm_get_param (d, a.index, &a.value);if ( rc ) goto out; - switch ( a.index ) + rc =__copy_to_guest (arg, &a,1 ) ? -EFAULT :0 ; + +HVM_DBG_LOG (DBG_LEVEL_HCALL ,"get param %u = %" PRIx64, + a.index, a.value);
xen/arch/x86/hvm/hvm.c static int hvmop_get_param( · line 4466
Quoted 10 lines
xen/arch/x86/hvm/hvm.c +10
-0+ + out: +rcu_unlock_domain (d); +return rc; +} + +int hvm_get_param ( +struct domain *d, +uint32_t index, +uint64_t *value)
Fold.
xen/arch/x86/hvm/hvm.c static int hvmop_get_param( · line 4470
Quoted 4 lines
xen/arch/x86/hvm/hvm.c +4
-0+{ +int rc; + +if ( index >=HVM_NR_PARAMS || !value )
No need for !value. It had better only ever point onto the hypervisor stack now that this is an internal function.
xen/include/asm-x86/hvm/hvm.h unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore); · line 338
Quoted 9 lines
xen/include/asm-x86/hvm/hvm.h +1
-0@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */
How about asserting so?
No major problems so Acked-by: Andrew Cooper <[email protected]>
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 02/20] xen/x86: Make hap_get_allocation accessible
In reply to Tamas K LengyelQuoted 7 lines · Tamas K Lengyel
During VM forking we'll copy the parent domain's parameters to the client, including the HAP shadow memory setting that is used for storing the domain's EPT. We'll copy this in the hypervisor instead doing it during toolstack launch to allow the domain to start executing and unsharing memory before (or even completely without) the toolstack.
Signed-off-by Tamas K Lengyel <[email protected]>
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 05/20] x86/mem_sharing: make get_two_gfns take locks conditionally
In reply to Tamas K LengyelQuoted 3 lines · Tamas K Lengyel
During VM forking the client lock will already be taken.
Signed-off-by Tamas K Lengyel <[email protected]>
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 07/20] x86/mem_sharing: don't try to unshare twice during page fault
In reply to Tamas K LengyelQuoted 8 lines · Tamas K Lengyel
xen/arch/x86/hvm/hvm.c +2
-0@@ -1959,19 +1965,21 @@ int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla,*/ if ( paged )p2m_mem_paging_populate (currd, gfn); + +#ifdef CONFIG_MEM_SHARING if ( sharing_enomem ) {
xen/arch/x86/hvm/hvm.c int hvm_hap_nested_page_fault(paddr_t gpa, unsigned long gla, · line 1976
Quoted 10 lines
xen/arch/x86/hvm/hvm.c +3
-5-int rv; - -if ( (rv =mem_sharing_notify_enomem (currd, gfn, true)) <0 ) +if ( !vm_event_check_ring (currd->vm_event_share) ) {gdprintk (XENLOG_ERR ,"Domain %hu attempt to unshare " -"gfn %lx, ENOMEM and no helper (rc %d)\n" , - currd->domain_id, gfn, rv); +"gfn %lx, ENOMEM and no helper\n" , + currd->domain_id, gfn);
As observations on this and later patches. Use %pd (gets d%u or d[NAME] as applicable), especially were d[COW] is a liable domid to get.
Also, any action which crashes a domain must not be a gdprintk(), because otherwise you end up with a domain_crash() and no context/symptoms in release builds of Xen.
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 04/20] x86/mem_sharing: cleanup code and comments in various locations
In reply to Andrew CooperQuoted 14 lines · Andrew Cooper
Quoted 12 lines · Tamas K Lengyel
Quoted 12 lines
xen/arch/x86/mm/mem_sharing.c static inline void mem_sharing_gfn_destroy(struct page_info *page, · line 444
Quoted 6 lines
xen/arch/x86/mm/mem_sharing.c +1
-1@@ -437,25 +441,29 @@ static inline void mem_sharing_gfn_destroy(struct page_info *page,xfree (gfn_info); } -static struct page_info *mem_sharing_lookup (unsigned long mfn) +static inline struct page_info *mem_sharing_lookup (unsigned long mfn)Seeing as this is cleanup, the position of the * can move. Similarly, it shouldn't gain an inline.
I can fix all of this up on commit (and a few other brace position issues) if you want, to save reworking a trivial v2.
Quoted 7 lines
Quoted 2 lines
Provided nothing else is outstanding with the patch and it can be committed I would certainly appreciate that.
I've pushed this and the previous patch, with some further fixups that I spotted.
Hopefully the rebase won't be an issue.
Great! There were quite a few conflicts in the rebase so I hope I don't accidentally revert some of your new fixes :)
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
In reply to Andrew CooperQuoted 9 lines
xen/include/asm-x86/hvm/hvm.h unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore); · line 338
Quoted 7 lines
xen/include/asm-x86/hvm/hvm.h +1
-0@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */ How about asserting so?
AFAICT there is no "domain_locked_by_me" function, only paging/p2m/gfn_locked_by_me. So any suggestion on how to achieve that?
Thanks!
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 10 lines · Tamas K Lengyel
Quoted 8 lines
xen/include/asm-x86/hvm/hvm.h unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore); · line 338
Quoted 7 lines
xen/include/asm-x86/hvm/hvm.h +1
-0@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */ How about asserting so?
AFAICT there is no "
domain_locked_by_me" function, only paging/p2m/gfn_locked_by_me. So any suggestion on how to achieve that?
spin_is_locked() gets you most of the way, and would be a start.
But yes - now you say this, I remember that we don't currently have suitable infrastructure.
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 10 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 8 lines
xen/include/asm-x86/hvm/hvm.h unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore); · line 338
Quoted 7 lines
xen/include/asm-x86/hvm/hvm.h +1
-0@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */ How about asserting so?
Quoted 7 lines
Quoted 2 lines
AFAICT there is no "
domain_locked_by_me" function, only paging/p2m/gfn_locked_by_me. So any suggestion on how to achieve that?
spin_is_locked()gets you most of the way, and would be a start.But yes - now you say this, I remember that we don't currently have suitable infrastructure.
Is the domain lock even a spin lock (the on you use by rcu_lock_live_remote_domain_by_id)? Looks to me it just goes down to "rcu_read_lock" which only does a preempt_disable() call o.O
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 13 lines · Tamas K Lengyel
Quoted 11 lines · Andrew Cooper
Quoted 10 lines · Tamas K Lengyel
Quoted 8 lines
xen/include/asm-x86/hvm/hvm.h unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore); · line 338
Quoted 7 lines
xen/include/asm-x86/hvm/hvm.h +1
-0@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */ How about asserting so?
AFAICT there is no "
domain_locked_by_me" function, only paging/p2m/gfn_locked_by_me. So any suggestion on how to achieve that?
Quoted 7 lines
Quoted 4 lines
spin_is_locked()gets you most of the way, and would be a start.But yes - now you say this, I remember that we don't currently have suitable infrastructure.
Is the domain lock even a spin lock (the on you use by
rcu_lock_live_remote_domain_by_id)? Looks to me it just goes down to "rcu_read_lock" which only does apreempt_disable()call o.O
/sigh - of course we have two things called the domain lock.
The RCU one is to ensure that the struct domain can't get freed behind our back, and shouldn't be interesting in this context (obtaining the d pointer in the first place causes it to be safe). If that is the only one which matters, I would drop the comment.
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 23 lines · Andrew Cooper
Quoted 13 lines · Tamas K Lengyel
Quoted 11 lines · Andrew Cooper
Quoted 10 lines · Tamas K Lengyel
Quoted 8 lines
xen/include/asm-x86/hvm/hvm.h unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore); · line 338
Quoted 7 lines
xen/include/asm-x86/hvm/hvm.h +1
-0@@ -335,6 +335,10 @@ unsigned long hvm_cr4_guest_valid_bits(const struct domain *d, bool restore);bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt); +/* Caller must hold domain locks */ How about asserting so?
AFAICT there is no "
domain_locked_by_me" function, only paging/p2m/gfn_locked_by_me. So any suggestion on how to achieve that?Quoted 7 lines
Quoted 4 lines
spin_is_locked()gets you most of the way, and would be a start.But yes - now you say this, I remember that we don't currently have suitable infrastructure.
Is the domain lock even a spin lock (the on you use by
rcu_lock_live_remote_domain_by_id)? Looks to me it just goes down to "rcu_read_lock" which only does apreempt_disable()call o.O
Quoted 6 lines
/sigh - of course we have two things called the domain lock.
The RCU one is to ensure that the struct domain can't get freed behind our back, and shouldn't be interesting in this context (obtaining the d pointer in the first place causes it to be safe). If that is the only one which matters, I would drop the comment.
Yes, only the RCU lock gets taken right now by all callers, so I'll drop the comment.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
In reply to Tamas K LengyelQuoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 12 lines
xen/arch/x86/hvm/hvm.c +3
-2@@ -4429,42 +4446,60 @@ static int hvmop_get_param(if ( !is_hvm_domain (d) ) goto out; - rc =hvm_allow_get_param (d, &a); + rc =hvm_get_param (d, a.index, &a.value);if ( rc ) goto out; - switch ( a.index ) + rc =__copy_to_guest (arg, &a,1 ) ? -EFAULT :0 ; +
xen/arch/x86/hvm/hvm.c static int hvmop_get_param( · line 4464
Quoted 10 lines
xen/arch/x86/hvm/hvm.c +10
-0+HVM_DBG_LOG (DBG_LEVEL_HCALL ,"get param %u = %" PRIx64, + a.index, a.value); + + out: +rcu_unlock_domain (d); +return rc; +} + +int hvm_get_param( + struct domain *d,
If this is to be non-static, I think it would be quite nice if this parameter was const. This will take a prereq patch to constify the XSM path involved, but other than this I can't see anything getting in the way.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 02/20] xen/x86: Make hap_get_allocation accessible
In reply to Tamas K Lengyelxen/arch/x86/mm/hap/hap.c static void hap_free_p2m_page(struct domain *d, struct page_info *pg) · line 324
Quoted 9 lines · Tamas K Lengyel
xen/arch/x86/mm/hap/hap.c +1
-2@@ -321,8 +321,7 @@ static void hap_free_p2m_page(struct domain *d, struct page_info *pg) }/* Return the size of the pool, rounded up to the nearest MB */ -static unsigned int -hap_get_allocation (struct domain *d) +unsigned int hap_get_allocation (struct domain *d)
Along the lines of the comment on patch 1, please take the opportunity and constify the parameter (which looks entirely straightforward here).
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
In reply to Jan BeulichQuoted 9 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 19 lines
Quoted 19 lines
xen/arch/x86/hvm/hvm.c +10
-2@@ -4429,42 +4446,60 @@ static int hvmop_get_param(if ( !is_hvm_domain (d) ) goto out; - rc =hvm_allow_get_param (d, &a); + rc =hvm_get_param (d, a.index, &a.value);if ( rc ) goto out; - switch ( a.index ) + rc =__copy_to_guest (arg, &a,1 ) ? -EFAULT :0 ; + +HVM_DBG_LOG (DBG_LEVEL_HCALL ,"get param %u = %" PRIx64, + a.index, a.value); + + out: +rcu_unlock_domain (d); +return rc; +}
Quoted 8 lines
xen/arch/x86/hvm/hvm.c static int hvmop_get_param( · line 4464
Quoted 3 lines
xen/arch/x86/hvm/hvm.c +3
-0+ +int hvm_get_param( + struct domain *d,If this is to be non-static, I think it would be quite nice if this parameter was const. This will take a prereq patch to constify the XSM path involved, but other than this I can't see anything getting in the way.
Sure.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
In particular, a number of the set operations are distinctly non-trivial. (OTOH, those are not long for this world, and should be creation X86_EMU_* constants instead).
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 17 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 4 lines
In particular, a number of the set operations are distinctly non-trivial. (OTOH, those are not long for this world, and should be creation X86_EMU_* constants instead).
I actually wanted to ask about that. In
https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=tools/libxc/xc_sr_save_x86_hvm.c;h=97a8c49807f192c47209525f51e4d79a50c66cec;hb=HEAD#l61
the toolstack only selects certain HVM params to be saved (and restored later). I originally followed the same logic in the fork code-path but after a lot of experiments it looks like it's actually OK to grab all params but only call set_param on the ones that have a non-zero value. So setting some params with a zero value has certainly lead to crashes, but otherwise it seems to "just work" to copy all the rest.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 21 lines · Tamas K Lengyel
Quoted 15 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 4 lines
In particular, a number of the set operations are distinctly non-trivial. (OTOH, those are not long for this world, and should be creation X86_EMU_* constants instead).
Quoted 9 lines
I actually wanted to ask about that. In
https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=tools/libxc/xc_sr_save_x86_hvm.c;h=97a8c49807f192c47209525f51e4d79a50c66cec;hb=HEAD#l61
the toolstack only selects certain HVM params to be saved (and restored later). I originally followed the same logic in the fork code-path but after a lot of experiments it looks like it's actually OK to grab all params but only callset_paramon the ones that have a non-zero value. So setting some params with a zero value has certainly lead to crashes, but otherwise it seems to "just work" to copy all the rest.
I think you're trying to ascribe any form of design/plan to a system which had none. :)
The code you quote was like that because that is how legacy migration worked. That said, eliding empty records was an effort-saving exercise (avoid redundant hypercalls on destination side), not because there was any suggestion that attempting to explicitly set 0 would crash.
Do you have any idea which param was causing problems?
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 36 lines · Andrew Cooper
Quoted 21 lines · Tamas K Lengyel
Quoted 15 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 4 lines
In particular, a number of the set operations are distinctly non-trivial. (OTOH, those are not long for this world, and should be creation X86_EMU_* constants instead).
Quoted 9 lines
I actually wanted to ask about that. In
https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=tools/libxc/xc_sr_save_x86_hvm.c;h=97a8c49807f192c47209525f51e4d79a50c66cec;hb=HEAD#l61
the toolstack only selects certain HVM params to be saved (and restored later). I originally followed the same logic in the fork code-path but after a lot of experiments it looks like it's actually OK to grab all params but only callset_paramon the ones that have a non-zero value. So setting some params with a zero value has certainly lead to crashes, but otherwise it seems to "just work" to copy all the rest.I think you're trying to ascribe any form of design/plan to a system which had none. :)
Quoted 6 lines
The code you quote was like that because that is how legacy migration worked. That said, eliding empty records was an effort-saving exercise (avoid redundant hypercalls on destination side), not because there was any suggestion that attempting to explicitly set 0 would crash.
Do you have any idea which param was causing problems?
Yes, HVM_PARAM_IDENT_PT was one sure. There may have been others (I don't recall now) but simply checking for non-zero value before calling set_param resolved everything.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 41 lines · Tamas K Lengyel
Quoted 39 lines · Andrew Cooper
Quoted 21 lines · Tamas K Lengyel
Quoted 15 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 4 lines
In particular, a number of the set operations are distinctly non-trivial. (OTOH, those are not long for this world, and should be creation X86_EMU_* constants instead).
Quoted 9 lines
I actually wanted to ask about that. In
https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=tools/libxc/xc_sr_save_x86_hvm.c;h=97a8c49807f192c47209525f51e4d79a50c66cec;hb=HEAD#l61
the toolstack only selects certain HVM params to be saved (and restored later). I originally followed the same logic in the fork code-path but after a lot of experiments it looks like it's actually OK to grab all params but only callset_paramon the ones that have a non-zero value. So setting some params with a zero value has certainly lead to crashes, but otherwise it seems to "just work" to copy all the rest.I think you're trying to ascribe any form of design/plan to a system which had none. :)
The code you quote was like that because that is how legacy migration worked. That said, eliding empty records was an effort-saving exercise (avoid redundant hypercalls on destination side), not because there was any suggestion that attempting to explicitly set 0 would crash.
Quoted 4 lines
Quoted 1 line
Do you have any idea which param was causing problems?
Yes,
HVM_PARAM_IDENT_PTwas one sure. There may have been others (I don't recall now) but simply checking for non-zero value before callingset_paramresolved everything.
IDENT_PT is an Westmere(?) wrinkle.
There was one processor back in those days which supported EPT, but didn't support VT-x running in unpaged mode. Therefore, we had to fake up unpaged mode by pointing vCR3 at an identity pagetable inside the guests physical address space.
The crash won't be from the IDENT_PT itself, but the paging_update_cr3() side effect. Was it a host crash, or guest crash?
~Andrew
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 48 lines · Andrew Cooper
Quoted 41 lines · Tamas K Lengyel
Quoted 39 lines · Andrew Cooper
Quoted 21 lines · Tamas K Lengyel
Quoted 15 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 4 lines
In particular, a number of the set operations are distinctly non-trivial. (OTOH, those are not long for this world, and should be creation X86_EMU_* constants instead).
Quoted 9 lines
I actually wanted to ask about that. In
https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=tools/libxc/xc_sr_save_x86_hvm.c;h=97a8c49807f192c47209525f51e4d79a50c66cec;hb=HEAD#l61
the toolstack only selects certain HVM params to be saved (and restored later). I originally followed the same logic in the fork code-path but after a lot of experiments it looks like it's actually OK to grab all params but only callset_paramon the ones that have a non-zero value. So setting some params with a zero value has certainly lead to crashes, but otherwise it seems to "just work" to copy all the rest.I think you're trying to ascribe any form of design/plan to a system which had none. :)
The code you quote was like that because that is how legacy migration worked. That said, eliding empty records was an effort-saving exercise (avoid redundant hypercalls on destination side), not because there was any suggestion that attempting to explicitly set 0 would crash.
Quoted 4 lines
Quoted 1 line
Do you have any idea which param was causing problems?
Yes,
HVM_PARAM_IDENT_PTwas one sure. There may have been others (I don't recall now) but simply checking for non-zero value before callingset_paramresolved everything.
Quoted 6 lines
IDENT_PTis an Westmere(?) wrinkle.There was one processor back in those days which supported EPT, but didn't support VT-x running in unpaged mode. Therefore, we had to fake up unpaged mode by pointing vCR3 at an identity pagetable inside the guests physical address space.
Eh, yikes.
Quoted 4 lines
The crash won't be from the
IDENT_PTitself, but thepaging_update_cr3()side effect. Was it a host crash, or guest crash?
Yes, that's what I recall after I looked into it. It was a guest a crash as I remember.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
In reply to Andrew CooperQuoted 16 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 2 lines
In particular, a number of the set operations are distinctly non-trivial.
How is trivial or not related to there being one function doing the looping wanted here vs the looping being done by the caller around the two per-entity calls?
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 18 lines · Jan Beulich
Quoted 16 lines · Andrew Cooper
Quoted 14 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
Quoted 6 lines
Quoted 2 lines
In particular, a number of the set operations are distinctly non-trivial.
How is trivial or not related to there being one function doing the looping wanted here vs the looping being done by the caller around the two per-entity calls?
I don't really get why would it matter where the looping is being done? Even if I were to add a single function to do this, it would do the same looping and just call the now internally kept get/set params functions.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
(re-sending, as I still don't see the mail having appeared on the list)
Quoted 22 lines · Tamas K Lengyel
Quoted 21 lines · Jan Beulich
Quoted 18 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
In particular, a number of the set operations are distinctly non-trivial.
Quoted 8 lines
Quoted 3 lines
How is trivial or not related to there being one function doing the looping wanted here vs the looping being done by the caller around the two per-entity calls?
I don't really get why would it matter where the looping is being done? Even if I were to add a single function to do this, it would do the same looping and just call the now internally kept get/set params functions.
The difference (to me) is what level of control gets exposed outside of the file. For example I also dislike external code doing this somewhat odd (but necessary as per your communication with Andrew) checking against zero values. Such are implementation details which would better not be scatter around.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 35 lines · Jan Beulich
(re-sending, as I still don't see the mail having appeared on the list)
Quoted 22 lines · Tamas K Lengyel
Quoted 21 lines · Jan Beulich
Quoted 18 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
In particular, a number of the set operations are distinctly non-trivial.
Quoted 8 lines
Quoted 3 lines
How is trivial or not related to there being one function doing the looping wanted here vs the looping being done by the caller around the two per-entity calls?
I don't really get why would it matter where the looping is being done? Even if I were to add a single function to do this, it would do the same looping and just call the now internally kept get/set params functions.
Quoted 5 lines
The difference (to me) is what level of control gets exposed outside of the file. For example I also dislike external code doing this somewhat odd (but necessary as per your communication with Andrew) checking against zero values. Such are implementation details which would better not be scatter around.
But you do realize that both of these functions are already exposed via hypercalls? So it's OK to call them from the toolstack but not from other parts of Xen itself? I don't see much reason there. But to me it makes 0 difference where the loop that copies the params is done, as long as it's within Xen. So if you really want it to be in hvm.c I can do that, I just find it silly.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 42 lines · Tamas K Lengyel
Quoted 35 lines · Jan Beulich
(re-sending, as I still don't see the mail having appeared on the list)
Quoted 22 lines · Tamas K Lengyel
Quoted 21 lines · Jan Beulich
Quoted 18 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
In particular, a number of the set operations are distinctly non-trivial.
Quoted 8 lines
Quoted 3 lines
How is trivial or not related to there being one function doing the looping wanted here vs the looping being done by the caller around the two per-entity calls?
I don't really get why would it matter where the looping is being done? Even if I were to add a single function to do this, it would do the same looping and just call the now internally kept get/set params functions.
Quoted 5 lines
The difference (to me) is what level of control gets exposed outside of the file. For example I also dislike external code doing this somewhat odd (but necessary as per your communication with Andrew) checking against zero values. Such are implementation details which would better not be scatter around.
Quoted 3 lines
But you do realize that both of these functions are already exposed via hypercalls? So it's OK to call them from the toolstack but not from other parts of Xen itself? I don't see much reason there.
Right now we have exactly one path each allowing this get/set. Adding a 2nd (from outside of hvm.c) opens the door for possible races between the various (for now just two) possible call sites. Noticing a possible problem when adding new code is imo quite a bit more likely if everything lives centralized in one place. IOW "exposure" here isn't meant so much in the sense of what entity in the system gets to drive the data, but which entities within Xen may play with it.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 01/20] x86: make hvm_{get/set}_param accessible
Quoted 48 lines · Jan Beulich
Quoted 42 lines · Tamas K Lengyel
Quoted 35 lines · Jan Beulich
(re-sending, as I still don't see the mail having appeared on the list)
Quoted 22 lines · Tamas K Lengyel
Quoted 21 lines · Jan Beulich
Quoted 18 lines · Andrew Cooper
Quoted 8 lines · Tamas K Lengyel
Quoted 7 lines · Jan Beulich
Quoted 3 lines · Tamas K Lengyel
Currently the hvm parameters are only accessible via the HVMOP hypercalls. By exposing hvm_{get/set}_param it will be possible for VM forking to copy the parameters directly into the clone domain.
Having peeked ahead at patch 17, where this gets used, I wonder why you want a pair of one-by-one functions, rather than a copy-all one. This then wouldn't require exposure of the functions you touch here.
Quoted 6 lines
Well, provided there is no such function in existence today it was just easier to use what's already available. I still wouldn't want to implement a one-shot function like that because this same code-path is shared by the save-restore operations on the toolstack side, so at least I have a reasonable assumption that it won't break on me in the future.
In particular, a number of the set operations are distinctly non-trivial.
Quoted 8 lines
Quoted 3 lines
How is trivial or not related to there being one function doing the looping wanted here vs the looping being done by the caller around the two per-entity calls?
I don't really get why would it matter where the looping is being done? Even if I were to add a single function to do this, it would do the same looping and just call the now internally kept get/set params functions.
Quoted 5 lines
The difference (to me) is what level of control gets exposed outside of the file. For example I also dislike external code doing this somewhat odd (but necessary as per your communication with Andrew) checking against zero values. Such are implementation details which would better not be scatter around.
Quoted 3 lines
But you do realize that both of these functions are already exposed via hypercalls? So it's OK to call them from the toolstack but not from other parts of Xen itself? I don't see much reason there.
Quoted 8 lines
Right now we have exactly one path each allowing this get/set. Adding a 2nd (from outside of
hvm.c) opens the door for possible races between the various (for now just two) possible call sites. Noticing a possible problem when adding new code is imo quite a bit more likely if everything lives centralized in one place. IOW "exposure" here isn't meant so much in the sense of what entity in the system gets to drive the data, but which entities within Xen may play with it.
Sure, I'll move the loop to hvm.c then.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 45 lines · Tamas K Lengyel
Quoted 35 lines · Roger Pau Monné
Quoted 12 lines · Tamas K Lengyel
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.Quoted 3 lines
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking
^ than
Quoted 9 lines
operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Quoted 7 lines
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.Quoted 8 lines
IMO having two different commands is confusing for the end user, I would rather have something like:
xl fork-vm [-d] ...
Where '-d' would prevent forking any user-space emulators. I don't thinks there's a need for a separate command to fork the underlying user-space emulators.
Quoted 5 lines
Keeping it as two commands allows you to start up the fork and let it run immediately and only start up QEMU when you notice it is needed. The idea being that you can monitor the kernel and see when it tries to do some I/O that would require the QEMU backend. If you combine the commands that option goes away.
I'm not sure I see why, you could still provide a xl fork-vm [-c] ... that would just lunch a QEMU instance. End users using xl have AFAICT no way to tell whether or when a QEMU is needed or not, and hence the default behavior should be a fully functional one.
IMO I think fork-vm without any options should do a complete fork of a VM, rather than a partial one without a device model clone.
Quoted 5 lines
Also, QEMU itself isn't getting forked
right now, we just start a new QEMU process with the saved-state getting loaded into it. I looked into implementing a QEMU fork command but it turns out that for the vast majority of our use-cases QEMU isn't needed at all, so developing that addition was tabled.
Starting a new process with the saved-state looks fine to me, it can always be improved afterwards if there's a need for it.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 53 lines · Roger Pau Monné
Quoted 45 lines · Tamas K Lengyel
Quoted 35 lines · Roger Pau Monné
Quoted 12 lines · Tamas K Lengyel
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.Quoted 3 lines
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking
^ than
Quoted 9 lines
operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Quoted 7 lines
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.Quoted 8 lines
IMO having two different commands is confusing for the end user, I would rather have something like:
xl fork-vm [-d] ...
Where '-d' would prevent forking any user-space emulators. I don't thinks there's a need for a separate command to fork the underlying user-space emulators.
Quoted 5 lines
Keeping it as two commands allows you to start up the fork and let it run immediately and only start up QEMU when you notice it is needed. The idea being that you can monitor the kernel and see when it tries to do some I/O that would require the QEMU backend. If you combine the commands that option goes away.
Quoted 7 lines
I'm not sure I see why, you could still provide a
xl fork-vm [-c] ...that would just lunch a QEMU instance. End users using xl have AFAICT no way to tell whether or when a QEMU is needed or not, and hence the default behavior should be a fully functional one.IMO I think fork-vm without any options should do a complete fork of a VM, rather than a partial one without a device model clone.
I understand your point but implementing that is outside the scope of what we are doing right now. There are a lot more steps involved if you want to create a fully functional VM fork with QEMU, for example you also have to create a separate disk so you don't clobber the parent VM's disk. Also, saving the QEMU device state is currently hard-wired into the save/migration operation, so changing that plumbing in libxl is quite involved. I actually found it way easier to just write a script that connects to the socket and saves it to a target file then going through the pain of adjusting libxl. So while this could be implemented at this time it won't be.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Hi Tamas,
Quoted 62 lines · Tamas K Lengyel
Quoted 53 lines · Roger Pau Monné
Quoted 45 lines · Tamas K Lengyel
Quoted 35 lines · Roger Pau Monné
Quoted 12 lines · Tamas K Lengyel
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.Quoted 3 lines
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking
^ than
Quoted 9 lines
operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Quoted 7 lines
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.Quoted 8 lines
IMO having two different commands is confusing for the end user, I would rather have something like:
xl fork-vm [-d] ...
Where '-d' would prevent forking any user-space emulators. I don't thinks there's a need for a separate command to fork the underlying user-space emulators.
Quoted 5 lines
Keeping it as two commands allows you to start up the fork and let it run immediately and only start up QEMU when you notice it is needed. The idea being that you can monitor the kernel and see when it tries to do some I/O that would require the QEMU backend. If you combine the commands that option goes away.
Quoted 7 lines
I'm not sure I see why, you could still provide a
xl fork-vm [-c] ...that would just lunch a QEMU instance. End users using xl have AFAICT no way to tell whether or when a QEMU is needed or not, and hence the default behavior should be a fully functional one.IMO I think fork-vm without any options should do a complete fork of a VM, rather than a partial one without a device model clone.
Quoted 10 lines
I understand your point but implementing that is outside the scope of what we are doing right now. There are a lot more steps involved if you want to create a fully functional VM fork with QEMU, for example you also have to create a separate disk so you don't clobber the parent VM's disk. Also, saving the QEMU device state is currently hard-wired into the save/migration operation, so changing that plumbing in libxl is quite involved. I actually found it way easier to just write a script that connects to the socket and saves it to a target file then going through the pain of adjusting libxl. So while this could be implemented at this time it won't be.
That's fine to not implement it right now, however the user interface should be able to cater it.
In this case, I agree with Roger that it is more intuitive to think that fork means a complete fork, not a partial one.
You could impose the user to always pass that option to not clone the device model and return an error if it is not there.
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 76 lines · Julien Grall
Hi Tamas,
Quoted 62 lines · Tamas K Lengyel
Quoted 53 lines · Roger Pau Monné
Quoted 45 lines · Tamas K Lengyel
Quoted 35 lines · Roger Pau Monné
Quoted 12 lines · Tamas K Lengyel
The following series implements VM forking for Intel HVM guests to allow for the fast creation of identical VMs without the assosciated high startup costs of booting or restoring the VM from a savefile.
JIRA issue: https://xenproject.atlassian.net/browse/XEN-89
The main design goal with this series has been to reduce the time of creating the VM fork as much as possible. To achieve this the VM forking process is split into two steps:
1) forking the VM on the hypervisor side;
2) starting QEMU to handle the backed for emulated devices.Quoted 3 lines
Step 1) involves creating a VM using the new "xl fork-vm" command. The parent VM is expected to remain paused after forks are created from it (which is different then what process forking normally entails). During this forking
^ than
Quoted 9 lines
operation the HVM context and VM settings are copied over to the new forked VM. This operation is fast and it allows the forked VM to be unpaused and to be monitored and accessed via VMI. Note however that without its device model running (depending on what is executing in the VM) it is bound to misbehave/crash when its trying to access devices that would be emulated by QEMU. We anticipate that for certain use-cases this would be an acceptable situation, in case for example when fuzzing is performed of code segments that don't access such devices.
Quoted 7 lines
Step 2) involves launching QEMU to support the forked VM, which requires the QEMU Xen savefile to be generated manually from the parent VM. This can be accomplished simply by connecting to its QMP socket and issuing the "xen-save-devices-state" command as documented by QEMU:
https://github.com/qemu/qemu/blob/master/docs/xen-save-devices-state.txt
Once the QEMU Xen savefile is generated the new "xl fork-launch-dm" command is used to launch QEMU and load the specified savefile for it.Quoted 8 lines
IMO having two different commands is confusing for the end user, I would rather have something like:
xl fork-vm [-d] ...
Where '-d' would prevent forking any user-space emulators. I don't thinks there's a need for a separate command to fork the underlying user-space emulators.
Quoted 5 lines
Keeping it as two commands allows you to start up the fork and let it run immediately and only start up QEMU when you notice it is needed. The idea being that you can monitor the kernel and see when it tries to do some I/O that would require the QEMU backend. If you combine the commands that option goes away.
Quoted 7 lines
I'm not sure I see why, you could still provide a
xl fork-vm [-c] ...that would just lunch a QEMU instance. End users using xl have AFAICT no way to tell whether or when a QEMU is needed or not, and hence the default behavior should be a fully functional one.IMO I think fork-vm without any options should do a complete fork of a VM, rather than a partial one without a device model clone.
Quoted 10 lines
I understand your point but implementing that is outside the scope of what we are doing right now. There are a lot more steps involved if you want to create a fully functional VM fork with QEMU, for example you also have to create a separate disk so you don't clobber the parent VM's disk. Also, saving the QEMU device state is currently hard-wired into the save/migration operation, so changing that plumbing in libxl is quite involved. I actually found it way easier to just write a script that connects to the socket and saves it to a target file then going through the pain of adjusting libxl. So while this could be implemented at this time it won't be.
Quoted 8 lines
That's fine to not implement it right now, however the user interface should be able to cater it.
In this case, I agree with Roger that it is more intuitive to think that fork means a complete fork, not a partial one.
You could impose the user to always pass that option to not clone the device model and return an error if it is not there.
Just to be clear, I can add the option to the "fork-vm" command to load the QEMU state with it, effectively combining the "fork-vm" and "fork-launch-dm" into one. But I still need the separate "fork-launch-dm" command since in our model we need to be able to launch the VM and run it without QEMU for a while, only launching QEMU when it is determined to be necessary. So if that's what you are asking, sure, I can do that.
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc. So if your concern is that the "fork-vm" command's name will imply that it is going to be producing fully functional VM on its own I would rather just rename the command because by itself it will never create a fully functional VM.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Hi,
Quoted 5 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.
If you fork then the configuration should be very similar. Right?
So why does the user requires to provide a new config rather than the command to update the existing one? To me, it feels this is a call to make mistake when forking.
How is the new config different from the original VM?
As a side note, I can't see any patch adding documentation.
Cheers,
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 12 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.
Quoted 5 lines
If you fork then the configuration should be very similar. Right?
So why does the user requires to provide a new config rather than the command to update the existing one? To me, it feels this is a call to make mistake when forking.
How is the new config different from the original VM?
The config must be different at least by giving the fork a different name. That's the minimum and it's enough only if the VM you are forking has no disk at all. If it has a disk, you also have to update the config to point to where the new disk is. I'm using LVM snapshots but you could also use qcow2, or whatever else there is for disk-CoW. The fork can also have different options enabled than it's parent. For example in our test-case, the forks have altp2m enabled while the parent VM doesn't. There could be other options like that someone might want to enable for the fork(s). If there is networking involved you likely also have to attach the fork to a new VLAN as to avoid MAC-address collision on the bridge. So there are quite a lot of variation possible, hence its better to have the user generate the new config they want instead of xl coming up with something on its own.
Quoted 2 lines
As a side note, I can't see any patch adding documentation.
It's only an experimental feature so adding documentation was not a priority. The documentation is pretty much in the cover letter. I'm happy to add its content as a file under docs in a patch (with the above extra information).
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.
IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.
We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
Quoted 10 lines
Quoted 6 lines
If you fork then the configuration should be very similar. Right?
So why does the user requires to provide a new config rather than the command to update the existing one? To me, it feels this is a call to make mistake when forking.
How is the new config different from the original VM?
The config must be different at least by giving the fork a different name. That's the minimum and it's enough only if the VM you are forking has no disk at all.
Adding an option to pass an explicit name for the fork would be handy, or else xl could come up with a name by itself, like it's done for migration, ie: <orignal name>--fork<digit>.
Quoted 11 lines
If it has a disk, you also have to update
the config to point to where the new disk is. I'm using LVM snapshots but you could also use qcow2, or whatever else there is for disk-CoW. The fork can also have different options enabled than it's parent. For example in our test-case, the forks have altp2m enabled while the parent VM doesn't. There could be other options like that someone might want to enable for the fork(s). If there is networking involved you likely also have to attach the fork to a new VLAN as to avoid MAC-address collision on the bridge. So there are quite a lot of variation possible, hence its better to have the user generate the new config they want instead of xl coming up with something on its own.
Passing a new config file for the fork is indeed fine, but maybe we don't want this to be the default behavior, as said above I think it's possible to fork a VM without passing a new config file.
Quoted 7 lines
Quoted 2 lines
As a side note, I can't see any patch adding documentation.
It's only an experimental feature so adding documentation was not a priority. The documentation is pretty much in the cover letter. I'm happy to add its content as a file under docs in a patch (with the above extra information).
Please also document the new xl command(s) in the man page [0].
Thanks, Roger.
[0] https://xenbits.xen.org/docs/unstable/man/xl.1.html
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?
Quoted 33 lines
Quoted 11 lines
Quoted 7 lines
If you fork then the configuration should be very similar. Right?
So why does the user requires to provide a new config rather than the command to update the existing one? To me, it feels this is a call to make mistake when forking.
How is the new config different from the original VM?
The config must be different at least by giving the fork a different name. That's the minimum and it's enough only if the VM you are forking has no disk at all.
Adding an option to pass an explicit name for the fork would be handy, or else xl could come up with a name by itself, like it's done for migration, ie: <orignal name>--fork<digit>.
Quoted 11 lines
If it has a disk, you also have to update
the config to point to where the new disk is. I'm using LVM snapshots but you could also use qcow2, or whatever else there is for disk-CoW. The fork can also have different options enabled than it's parent. For example in our test-case, the forks have altp2m enabled while the parent VM doesn't. There could be other options like that someone might want to enable for the fork(s). If there is networking involved you likely also have to attach the fork to a new VLAN as to avoid MAC-address collision on the bridge. So there are quite a lot of variation possible, hence its better to have the user generate the new config they want instead of xl coming up with something on its own.Passing a new config file for the fork is indeed fine, but maybe we don't want this to be the default behavior, as said above I think it's possible to fork a VM without passing a new config file.
Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
As a side note, I can't see any patch adding documentation.
It's only an experimental feature so adding documentation was not a priority. The documentation is pretty much in the cover letter. I'm happy to add its content as a file under docs in a patch (with the above extra information).
Please also document the new xl command(s) in the man page [0].
Ack.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 29 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
Quoted 6 lines
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 38 lines · Roger Pau Monné
Quoted 29 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
Quoted 6 lines
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.
It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?
Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"
"-C <config> Use config file for VM fork.\n"
"-Q <qemu-save-file> Use qemu save file for VM fork.\n"
"--launch-dm <yes|no|late> Launch device model (QEMU) for VM fork.\n"
"--fork-reset Reset VM fork.\n"
"-p Do not unpause VMs after fork."
"-h Print this help.\n"
"-d Enable debug messages.\n"
},
Currently the parts that are implemented would look like:
xl fork-vm -p --launch-dm no <parent_domid>
xl fork-vm -p --launch-dm late -C <config> -Q <qemu-save-file> <fork_domid> xl fork-vm -p --fork-reset <fork_domid>
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Hi Tamas,
Quoted 70 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"
"-C <config> Use config file for VM fork.\n"
"-Q <qemu-save-file> Use qemu save file for VM fork.\n"
"--launch-dm <yes|no|late> Launch device model (QEMU) for VM fork.\n"
"--fork-reset Reset VM fork.\n"
"-p Do not unpause VMs after fork."
"-h Print this help.\n"
"-d Enable debug messages.\n"
},
Quoted 4 lines
Currently the parts that are implemented would look like:
xl fork-vm -p --launch-dm no <parent_domid>
xl fork-vm -p --launch-dm late -C <config> -Q <qemu-save-file> <fork_domid> xl fork-vm -p --fork-reset <fork_domid>
The interface looks good to me. Note that I don't think you need to describre all the unimplemented option in the help. It would be sufficient to describe what you only support and bail out if the user gives something different.
What matters is we are able to extend the command line option over the time.
Cheers,
Signature · 6 lines
Julien Grall _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 52 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
Quoted 9 lines
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"
IMO I think the name of fork is the only useful option. Being able to assign disks or bridges from the command line seems quite complicated. What about VMs with multiple disks? Or VMs with multiple nics on different bridges?
I think it's easier for both the implementation and the user to just use a config file in that case.
Quoted 5 lines
"-C <config> Use config file for VM fork.\n"
"-Q <qemu-save-file> Use qemu save file for VM fork.\n"
"--launch-dm <yes|no|late> Launch device model (QEMU) for VM fork.\n"
"--fork-reset Reset VM fork.\n"
"-p Do not unpause VMs after fork."I think the default behaviour should be to leave the original VM paused and the forked one running, and hence this should be:
"-p Leave forked VM paused."
"-u Leave parent VM unpaused."
Quoted 7 lines
"-h Print this help.\n"
"-d Enable debug messages.\n"
},Currently the parts that are implemented would look like:
xl fork-vm -p --launch-dm no <parent_domid>
xl fork-vm -p --launch-dm late -C <config> -Q <qemu-save-file> <fork_domid>
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
I think you also need something like:
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>
So that a user doesn't need to pass a qemu-save-file?
Can you also list the command used to get the Qemu save-file from the parent? (just for completeness purposes).
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 64 lines · Roger Pau Monné
Quoted 52 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
Quoted 9 lines
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"
Quoted 7 lines
IMO I think the name of fork is the only useful option. Being able to assign disks or bridges from the command line seems quite complicated. What about VMs with multiple disks? Or VMs with multiple nics on different bridges?
I think it's easier for both the implementation and the user to just use a config file in that case.
I agree but it sounded to me you guys wanted to have a "complete" interface even if it's unimplemented. This is what a complete interface would look to me.
Quoted 9 lines
Quoted 6 lines
"-C <config> Use config file for VM fork.\n" "-Q <qemu-save-file> Use qemu save file for VM fork.\n" "--launch-dm <yes|no|late> Launch device model (QEMU) for VM fork.\n" "--fork-reset Reset VM fork.\n" "-p Do not unpause VMs after fork."I think the default behaviour should be to leave the original VM paused and the forked one running, and hence this should be:
That is the default. I guess the text saying VMs was not correctly worded, it just means don't unpause fork after it's created. The parent remains always paused.
Quoted 3 lines
"-p Leave forked VM paused."
"-u Leave parent VM unpaused."
But you shouldn't unpause the parent VM at all. It should remain paused as long as there are forks running that were split from it. Unpausing it will lead to subtle and unexplainable crashes in the fork since the fork now will use pages that are from a different execution path. Technically in the future it would be possible to unpause the VM but it requires to fully populate the pagetables in all forks made from it with mem_shared entries and deduplicate to the forks all the pages that's can't be be shared. That was what I originally tried to do but it was extremely slow, hence the lazy-population of the pagetable in the forks.
Quoted 11 lines
Quoted 8 lines
"-h Print this help.\n"
"-d Enable debug messages.\n"
},Currently the parts that are implemented would look like:
xl fork-vm -p --launch-dm no <parent_domid>
xl fork-vm -p --launch-dm late -C <config> -Q <qemu-save-file> <fork_domid>Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
Quoted 6 lines
I think you also need something like:
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
Quoted 3 lines
Can you also list the command used to get the Qemu save-file from the parent? (just for completeness purposes).
It's explained in the cover letter. You connect to the QEMU socket and issue the xen-save-devices-state QMP command.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 37 lines · Roger Pau Monné
Quoted 29 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
Quoted 6 lines
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.
So first of all, Tamas -- do you actually need to exec xl here? Would it make sense for these to start out simply as libxl functions that are called by your system?
I actually disagree that we want a single command to do all of these. If we did want exec xl to be one of the supported interfaces, I think it would break down something like this:
xl fork-domain: Only forks the domain.xl fork-launch-dm: (or attach-dm?): Start up and attach the devicemodel to the domain
Then xl fork (or maybe xl fork-vm) would be something implemented in the future that would fork the entire domain.
(This is similar to how git am works for instance; internally it runs several steps, including git mailsplit, git mailinfo, and git apply-patch, each of which can be called individually.)
I think I would also have:
xl fork-save-dm: Connect over qmp to the parent domain and save the dm file
Then have xl fork-launch-dm either take a filename (saved from the previous step) or a parent domain id (in which case it would arrange to save the file itself).
Although in fact, is there any reason we couldn't store the parent domain ID in xenstore, so that xl fork-launch-dm could find the parent by itself? (Although that, of course, is something that could be added later if it's not something Tamas needs.)
Thoughts?
-George
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 39 lines · George Dunlap
Quoted 37 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?
Quoted 7 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.So first of all, Tamas -- do you actually need to exec xl here? Would it make sense for these to start out simply as libxl functions that are called by your system?
For my current tools & tests - no. I don't start QEMU for the forks at all. So at this point I don't even need libxl. But I can foresee that at some point in the future it may become necessary in case we want allow the forked VM to touch emulated devices. Wiring QEMU up and making the system functional as a whole I found it easier to do it via xl. There is just way too many moving components involved to do that any other way.
Quoted 11 lines
I actually disagree that we want a single command to do all of these. If we did want
exec xlto be one of the supported interfaces, I think it would break down something like this:
xl fork-domain: Only forks the domain.xl fork-launch-dm: (or attach-dm?): Start up and attach the devicemodel to the domainThen
xl fork(or maybexl fork-vm) would be something implemented in the future that would fork the entire domain.
I really don't have a strong opinion about this either way. I can see it working either way. Having them all bundled under a single top-level comment doesn't pollute the help text when someone is just looking at what xl can do in general. Makes that command a lot more complex for sure but I don't think it's too bad.
Quoted 9 lines
(This is similar to how
git amworks for instance; internally it runs several steps, includinggit mailsplit,git mailinfo, andgit apply-patch, each of which can be called individually.)I think I would also have:
xl fork-save-dm: Connect over qmp to the parent domain and save the dm file
Aye, could be done. For now I didn't bother since its trivial to do manually already.
Quoted 9 lines
Then have
xl fork-launch-dmeither take a filename (saved from the previous step) or a parent domain id (in which case it would arrange to save the file itself).Although in fact, is there any reason we couldn't store the parent domain ID in xenstore, so that
xl fork-launch-dmcould find the parent by itself? (Although that, of course, is something that could be added later if it's not something Tamas needs.)
Could be done. But I store ID internally in my tools anyway since I need it to initialize VMI. So having it in Xenstore is not required for me. In fact I would prefer to leave Xenstore out of these operations as much as possible cause it would slow things down. In my latest tests forking is down to 0.0007s, having to touch Xenstore for each would slow things down considerably.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 65 lines · Tamas K Lengyel
Quoted 39 lines · George Dunlap
Quoted 37 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 7 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.So first of all, Tamas -- do you actually need to exec xl here? Would it make sense for these to start out simply as libxl functions that are called by your system?
For my current tools & tests - no. I don't start QEMU for the forks at all. So at this point I don't even need libxl. But I can foresee that at some point in the future it may become necessary in case we want allow the forked VM to touch emulated devices. Wiring QEMU up and making the system functional as a whole I found it easier to do it via xl. There is just way too many moving components involved to do that any other way.
Quoted 9 lines
I actually disagree that we want a single command to do all of these. If we did want
exec xlto be one of the supported interfaces, I think it would break down something like this:
xl fork-domain: Only forks the domain.xl fork-launch-dm: (or attach-dm?): Start up and attach the devicemodel to the domain
Quoted 8 lines
Quoted 2 lines
Then
xl fork(or maybexl fork-vm) would be something implemented in the future that would fork the entire domain.I really don't have a strong opinion about this either way. I can see it working either way. Having them all bundled under a single top-level comment doesn't pollute the help text when someone is just looking at what xl can do in general. Makes that command a lot more complex for sure but I don't think it's too bad.
One thing I don't like about having a single command is that since you're not planning on implementing the end-to-end "vm fork" command, then when running the base "fork-vm" command, you'll have to print an error message that says "This command is not available in its completeness; you'll have to implement your own via fork-vm --domain, fork-vm --save-dm, and fork-vm --launch-dm."
Which we could do, but seem a bit strange. :-)
Quoted 9 lines
Quoted 8 lines
Then have
xl fork-launch-dmeither take a filename (saved from the previous step) or a parent domain id (in which case it would arrange to save the file itself).Although in fact, is there any reason we couldn't store the parent domain ID in xenstore, so that
xl fork-launch-dmcould find the parent by itself? (Although that, of course, is something that could be added later if it's not something Tamas needs.)
Quoted 6 lines
Could be done. But I store ID internally in my tools anyway since I need it to initialize VMI. So having it in Xenstore is not required for me. In fact I would prefer to leave Xenstore out of these operations as much as possible cause it would slow things down. In my latest tests forking is down to 0.0007s, having to touch Xenstore for each would slow things down considerably.
Right, that makes sense.
-George
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 17 lines
Quoted 8 lines
Quoted 8 lines
I actually disagree that we want a single command to do all of these. If we did want
exec xlto be one of the supported interfaces, I think it would break down something like this:
xl fork-domain: Only forks the domain.xl fork-launch-dm: (or attach-dm?): Start up and attach the devicemodel to the domainQuoted 8 lines
Quoted 2 lines
Then
xl fork(or maybexl fork-vm) would be something implemented in the future that would fork the entire domain.I really don't have a strong opinion about this either way. I can see it working either way. Having them all bundled under a single top-level comment doesn't pollute the help text when someone is just looking at what xl can do in general. Makes that command a lot more complex for sure but I don't think it's too bad.
Quoted 8 lines
One thing I don't like about having a single command is that since you're not planning on implementing the end-to-end "vm fork" command, then when running the base "fork-vm" command, you'll have to print an error message that says "This command is not available in its completeness; you'll have to implement your own via fork-vm --domain, fork-vm --save-dm, and fork-vm --launch-dm."
Which we could do, but seem a bit strange. :-)
Yea, it's not a single step to get to a fully functional fork but its close:
1. pause parent vm
2. generate qemu_save_file
3. xl fork-vm -C config -Q qemu_save_file <parent_domid>
For the second fork - provided it has its own config file ready to go
- it is enough to just run the 3. step. Technically we could integrate all these three steps into one and then the user would only have to generate the new config file. But I found this setup to be "good enough" already.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 70 lines · Tamas K Lengyel
Quoted 69 lines · Roger Pau Monné
Quoted 52 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
Quoted 9 lines
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"IMO I think the name of fork is the only useful option. Being able to assign disks or bridges from the command line seems quite complicated. What about VMs with multiple disks? Or VMs with multiple nics on different bridges?
Quoted 6 lines
Quoted 2 lines
I think it's easier for both the implementation and the user to just use a config file in that case.
I agree but it sounded to me you guys wanted to have a "complete" interface even if it's unimplemented. This is what a complete interface would look to me.
I would add those options afterwards if there's a need for them. I was mainly concerned about introducing a top level command (ie: fork-vm) that would require calling other commands in order to get a functional fork. I'm not so concerned about having all the possible options listed now, as long as the default behavior of fork-vm is something sane that produces a working fork, even if not fully implemented at this stage.
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
"-C <config> Use config file for VM fork.\n" "-Q <qemu-save-file> Use qemu save file for VM fork.\n" "--launch-dm <yes|no|late> Launch device model (QEMU) for VM fork.\n" "--fork-reset Reset VM fork.\n" "-p Do not unpause VMs after fork."
Quoted 6 lines
Quoted 2 lines
I think the default behaviour should be to leave the original VM paused and the forked one running, and hence this should be:
That is the default. I guess the text saying VMs was not correctly worded, it just means don't unpause fork after it's created. The parent remains always paused.
Ack.
Quoted 4 lines
Quoted 3 lines
"-p Leave forked VM paused."
"-u Leave parent VM unpaused."
Quoted 8 lines
But you shouldn't unpause the parent VM at all. It should remain paused as long as there are forks running that were split from it. Unpausing it will lead to subtle and unexplainable crashes in the fork since the fork now will use pages that are from a different execution path. Technically in the future it would be possible to unpause the VM but it requires to fully populate the pagetables in all forks made from it with
mem_sharedentries and deduplicate to the forks all the pages that's can't be be shared.
Oh, OK, since I didn't look at the implementation yet I assumed that the parent would also be switched to trap on memory writes, so that you could duplicate the pages before the parent writes to them, and hence the parent could be left running.
Anyway, let's forget about the "leave parent unpaused" option then.
Quoted 13 lines
That was what I originally tried to
do but it was extremely slow, hence the lazy-population of the pagetable in the forks.Quoted 9 lines
Quoted 8 lines
"-h Print this help.\n"
"-d Enable debug messages.\n"
},Currently the parts that are implemented would look like:
xl fork-vm -p --launch-dm no <parent_domid>
xl fork-vm -p --launch-dm late -C <config> -Q <qemu-save-file> <fork_domid>
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
Anyway, I'm not going to insist on this but the workflow of the Qemu forking seems to not be very user friendly unless you know exactly how to use it.
Quoted 4 lines
Quoted 4 lines
I think you also need something like:
Quoted 8 lines
Quoted 3 lines
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
You could get all this information from the parent VM, IIRC libxl has a json version of the config. For example for migration there's no need to pass any config file, since the incoming VM can be recreated from the data in the source VM.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 50 lines · George Dunlap
Quoted 37 lines · Roger Pau Monné
Quoted 29 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
Quoted 6 lines
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.So first of all, Tamas -- do you actually need to exec xl here? Would it make sense for these to start out simply as libxl functions that are called by your system?
I actually disagree that we want a single command to do all of these. If we did want
exec xlto be one of the supported interfaces, I think it would break down something like this:
Quoted 6 lines
xl fork-domain: Only forks the domain.xl fork-launch-dm: (or attach-dm?): Start up and attach the devicemodel to the domainThen
xl fork(or maybexl fork-vm) would be something implemented in the future that would fork the entire domain.
I don't have a strong opinion on whether we should have a bunch of fork-* commands or a single one. My preference would be for a single one because I think other commands can be implemented as options.
What I would like to prevent is ending up with something like fork-domain and fork-vm commands, which look like aliases, and can lead to confusion.
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 79 lines · Roger Pau Monné
Quoted 70 lines · Tamas K Lengyel
Quoted 69 lines · Roger Pau Monné
Quoted 52 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
Quoted 9 lines
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"IMO I think the name of fork is the only useful option. Being able to assign disks or bridges from the command line seems quite complicated. What about VMs with multiple disks? Or VMs with multiple nics on different bridges?
Quoted 6 lines
Quoted 2 lines
I think it's easier for both the implementation and the user to just use a config file in that case.
I agree but it sounded to me you guys wanted to have a "complete" interface even if it's unimplemented. This is what a complete interface would look to me.
Quoted 7 lines
I would add those options afterwards if there's a need for them. I was mainly concerned about introducing a top level command (ie: fork-vm) that would require calling other commands in order to get a functional fork. I'm not so concerned about having all the possible options listed now, as long as the default behavior of fork-vm is something sane that produces a working fork, even if not fully implemented at this stage.
OK
Quoted 8 lines
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
Quoted 4 lines
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
The forked VM has no "name" when it's created. For performance reasons when the VM fork is created with "--launch-dm no" we explicitly want to avoid touching Xenstore. Even parsing the config file would be unneeded overhead at that stage.
Quoted 18 lines
Anyway, I'm not going to insist on this but the workflow of the Qemu forking seems to not be very user friendly unless you know exactly how to use it.
Quoted 4 lines
Quoted 4 lines
I think you also need something like:
Quoted 8 lines
Quoted 3 lines
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
Quoted 5 lines
You could get all this information from the parent VM, IIRC libxl has a json version of the config. For example for migration there's no need to pass any config file, since the incoming VM can be recreated from the data in the source VM.
But again, creating a fork with the exact config of the parent is not possible. Even if the tool would rename the fork on-the-fly as it does during the migration, the fork would end up thrashing the parent VM's disk and making it impossible to create any additional forks. It would also mean that at no point can the original VM be unpaused after the forks are gone. I don't see any usecase in which that would make any sense at all.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 13 lines
Quoted 8 lines
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
Quoted 4 lines
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
Quoted 4 lines
The forked VM has no "name" when it's created. For performance reasons when the VM fork is created with "--launch-dm no" we explicitly want to avoid touching Xenstore. Even parsing the config file would be unneeded overhead at that stage.
And to answer your question, no, the parent VM's name is not recorded anywhere for the fork. Technically not even the parent's domain id is kept by Xen. The fork only keeps a pointer to the parent's "struct domain". So right now there is no hypercall interface to retrieve a fork's parent's ID - it is assumed the tools using the interface are keeping track of that. Could this information be dumped into Xenstore as well? Yes. But we specifically want be able to create the fork as fast possible without any unnecessary overhead.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 103 lines · Tamas K Lengyel
Quoted 79 lines · Roger Pau Monné
Quoted 70 lines · Tamas K Lengyel
Quoted 69 lines · Roger Pau Monné
Quoted 52 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update would still not produce for you a "fully functional" VM on its own. The user still has to produce a new VM config file, create the new disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
Quoted 9 lines
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"IMO I think the name of fork is the only useful option. Being able to assign disks or bridges from the command line seems quite complicated. What about VMs with multiple disks? Or VMs with multiple nics on different bridges?
Quoted 6 lines
Quoted 2 lines
I think it's easier for both the implementation and the user to just use a config file in that case.
I agree but it sounded to me you guys wanted to have a "complete" interface even if it's unimplemented. This is what a complete interface would look to me.
Quoted 7 lines
I would add those options afterwards if there's a need for them. I was mainly concerned about introducing a top level command (ie: fork-vm) that would require calling other commands in order to get a functional fork. I'm not so concerned about having all the possible options listed now, as long as the default behavior of fork-vm is something sane that produces a working fork, even if not fully implemented at this stage.
OK
Quoted 8 lines
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
Quoted 4 lines
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
Quoted 4 lines
The forked VM has no "name" when it's created. For performance reasons when the VM fork is created with "--launch-dm no" we explicitly want to avoid touching Xenstore. Even parsing the config file would be unneeded overhead at that stage.
I think you need another option to tell xl to not name the forked VM, abusing of "--launch-dm no" to not set a name is not expected I think.
Quoted 24 lines
Quoted 18 lines
Anyway, I'm not going to insist on this but the workflow of the Qemu forking seems to not be very user friendly unless you know exactly how to use it.
Quoted 4 lines
Quoted 4 lines
I think you also need something like:
Quoted 8 lines
Quoted 3 lines
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
Quoted 5 lines
You could get all this information from the parent VM, IIRC libxl has a json version of the config. For example for migration there's no need to pass any config file, since the incoming VM can be recreated from the data in the source VM.
Quoted 7 lines
But again, creating a fork with the exact config of the parent is not possible. Even if the tool would rename the fork on-the-fly as it does during the migration, the fork would end up thrashing the parent VM's disk and making it impossible to create any additional forks. It would also mean that at no point can the original VM be unpaused after the forks are gone. I don't see any usecase in which that would make any sense at all.
You could have the disk(s) as read-only and the VM running completely from RAM. Alpine-linux has (or had) a mode where it was completely stateless and running from RAM. I think it's fine to require passing a config file for the time being, we can look at other options afterwards.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 18 lines · Tamas K Lengyel
Quoted 13 lines
Quoted 8 lines
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
Quoted 4 lines
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
Quoted 4 lines
The forked VM has no "name" when it's created. For performance reasons when the VM fork is created with "--launch-dm no" we explicitly want to avoid touching Xenstore. Even parsing the config file would be unneeded overhead at that stage.
Quoted 4 lines
And to answer your question, no, the parent VM's name is not recorded anywhere for the fork. Technically not even the parent's domain id is kept by Xen. The fork only keeps a pointer to the parent's "struct domain"
There's the domain_id field inside of struct domain, so it seems quite easy to get the parent domid from the fork if there's a pointer to the parent's struct domain.
Quoted 5 lines
So right now there is no hypercall interface to retrieve a fork's parent's ID - it is assumed the tools using the interface are keeping track of that. Could this information be dumped into Xenstore as well? Yes. But we specifically want be able to create the fork as fast possible without any unnecessary overhead.
I think it would be nice to identify forked domains usingXEN_DOMCTL_getdomaininfo: you could add a parent_domid field to xen_domctl_getdomaininfo and if it's set to something different than DOMID_INVALID then the domain is a fork of the given domid.
Not saying it should be done now, but AFAICT getting the parent's domid is feasible and doesn't require xenstore.
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 35 lines · Roger Pau Monné
Quoted 18 lines · Tamas K Lengyel
Quoted 13 lines
Quoted 8 lines
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
Quoted 4 lines
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
Quoted 4 lines
The forked VM has no "name" when it's created. For performance reasons when the VM fork is created with "--launch-dm no" we explicitly want to avoid touching Xenstore. Even parsing the config file would be unneeded overhead at that stage.
Quoted 4 lines
And to answer your question, no, the parent VM's name is not recorded anywhere for the fork. Technically not even the parent's domain id is kept by Xen. The fork only keeps a pointer to the parent's "struct domain"
There's the
domain_idfield inside of struct domain, so it seems quite easy to get the parent domid from the fork if there's a pointer to the parent's struct domain.Quoted 5 lines
So right now there is no hypercall interface to retrieve a fork's parent's ID - it is assumed the tools using the interface are keeping track of that. Could this information be dumped into Xenstore as well? Yes. But we specifically want be able to create the fork as fast possible without any unnecessary overhead.
Quoted 8 lines
I think it would be nice to identify forked domains using
XEN_DOMCTL_getdomaininfo: you could add aparent_domidfield toxen_domctl_getdomaininfoand if it's set to something different thanDOMID_INVALIDthen the domain is a fork of the given domid.Not saying it should be done now, but AFAICT getting the parent's domid is feasible and doesn't require xenstore.
Of course it could be done. I was just pointing out that it's not currently kept separately and there is no interface to retrieve it. But TBH I have lost the train the though why we would need that in the first place? When QEMU is being launched the fork is already created and QEMU doesn't need to know anything about the parent.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 105 lines · Roger Pau Monné
Quoted 103 lines · Tamas K Lengyel
Quoted 79 lines · Roger Pau Monné
Quoted 70 lines · Tamas K Lengyel
Quoted 69 lines · Roger Pau Monné
Quoted 52 lines · Tamas K Lengyel
Quoted 40 lines · Tamas K Lengyel
Quoted 38 lines · Roger Pau Monné
Quoted 35 lines · Tamas K Lengyel
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 10 lines · Julien Grall
Hi,
Quoted 6 lines · On Mon, 30 Dec 2019, 20:49 Tamas K Lengyel, <[email protected]> wrote:
On Mon, Dec 30, 2019 at 11:43 AM Julien Grall <[email protected]> wrote:
But keep in mind that the "fork-vm" command even with this update
would still not produce for you a "fully functional" VM on its own.
The user still has to produce a new VM config file, create the new
disk, save the QEMU state, etc.IMO the default behavior of the fork command should be to leave the original VM paused, so that you can continue using the same disk and network config in the fork and you won't need to pass a new config file.
Quoted 8 lines
As Julien already said, maybe I wasn't clear in my previous replies:
I'm not asking you to implement all this, it's fine if the implementation of the fork-vm xl command requires you to pass certain options, and that the default behavior is not implemented.We need an interface that's sane, and that's designed to be easy and comprehensive to use, not an interface built around what's currently implemented.
OK, so I think that would look like "xl fork-vm <
parent_domid>" with additional options for things like name, disk, vlan, or a completely new config, all of which are currently not implemented, + an additional option to not launch QEMU at all, which would be the only one currently working. Also keeping the separate "xl fork-launch-dm" as is. Is that what we are talking about?Quoted 8 lines
Quoted 3 lines
I think fork-launch-vm should just be an option of fork-vm (ie:
--launch-dm-only or some such). I don't think there's a reason to have a separate top-level command to just launch the device model.It's just that the fork-launch-dm needs the domid of the fork, while the fork-vm needs the parent's domid. But I guess we can interpret the "domid" required input differently depending on which sub-option is specified for the command. Let's see how it pans out.
How does the following look for the interface?
Quoted 9 lines
{ "fork-vm",
&main_fork_vm, 0, 1,
"Fork a domain from the running parent domid", "[options] <Domid>",
"-h Print this help.\n"
"-N <name> Assign name to VM fork.\n"
"-D <disk> Assign disk to VM fork.\n"
"-B <bridge Assign bridge to VM fork.\n"
"-V <vlan> Assign vlan to VM fork.\n"IMO I think the name of fork is the only useful option. Being able to assign disks or bridges from the command line seems quite complicated. What about VMs with multiple disks? Or VMs with multiple nics on different bridges?
Quoted 6 lines
Quoted 2 lines
I think it's easier for both the implementation and the user to just use a config file in that case.
I agree but it sounded to me you guys wanted to have a "complete" interface even if it's unimplemented. This is what a complete interface would look to me.
Quoted 7 lines
I would add those options afterwards if there's a need for them. I was mainly concerned about introducing a top level command (ie: fork-vm) that would require calling other commands in order to get a functional fork. I'm not so concerned about having all the possible options listed now, as long as the default behavior of fork-vm is something sane that produces a working fork, even if not fully implemented at this stage.
OK
Quoted 12 lines
Quoted 7 lines
Quoted 2 lines
Why do you need a config file for launching the Qemu device model? Doesn't the save-file contain all the information?
The config is used to populate xenstore, not just for QEMU. The QEMU save file doesn't contain the xl config. This is not a full VM save file, it is only the QEMU state that gets dumped with xen-save-devices-state.
TBH I think it would be easier to have something like my proposal below, where you tell xl the parent and the forked VM names and xl does the rest. Even better would be to not have to tell xl the parent VM name (since I guess this is already tracked internally somewhere?).
Quoted 7 lines
Quoted 4 lines
The forked VM has no "name" when it's created. For performance reasons when the VM fork is created with "--launch-dm no" we explicitly want to avoid touching Xenstore. Even parsing the config file would be unneeded overhead at that stage.
I think you need another option to tell xl to not name the forked VM, abusing of "--launch-dm no" to not set a name is not expected I think.
See my reply below.
Quoted 33 lines
Quoted 25 lines
Quoted 19 lines
Anyway, I'm not going to insist on this but the workflow of the Qemu forking seems to not be very user friendly unless you know exactly how to use it.
Quoted 4 lines
Quoted 4 lines
I think you also need something like:
Quoted 8 lines
Quoted 3 lines
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
Quoted 5 lines
You could get all this information from the parent VM, IIRC libxl has a json version of the config. For example for migration there's no need to pass any config file, since the incoming VM can be recreated from the data in the source VM.
Quoted 7 lines
But again, creating a fork with the exact config of the parent is not possible. Even if the tool would rename the fork on-the-fly as it does during the migration, the fork would end up thrashing the parent VM's disk and making it impossible to create any additional forks. It would also mean that at no point can the original VM be unpaused after the forks are gone. I don't see any usecase in which that would make any sense at all.
Quoted 6 lines
You could have the disk(s) as read-only and the VM running completely from RAM. Alpine-linux has (or had) a mode where it was completely stateless and running from RAM. I think it's fine to require passing a config file for the time being, we can look at other options afterwards.
OK, there is that. But I would say that's a fairly niche use-case. You wouldn't have any network access in that fork, no disk, no way to get information in or out beside the serial console. So I wouldn't want that setup to be considered the default. If someone wants to that I would rather have an option that tells xl to automatically name the fork for you instead of the other way around.
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 39 lines · Tamas K Lengyel
Quoted 31 lines · Roger Pau Monné
Quoted 21 lines · Tamas K Lengyel
Quoted 14 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 5 lines · Roger Pau Monné
I think you also need something like:
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
Quoted 5 lines
You could get all this information from the parent VM, IIRC libxl has a json version of the config. For example for migration there's no need to pass any config file, since the incoming VM can be recreated from the data in the source VM.
Quoted 7 lines
But again, creating a fork with the exact config of the parent is not possible. Even if the tool would rename the fork on-the-fly as it does during the migration, the fork would end up thrashing the parent VM's disk and making it impossible to create any additional forks. It would also mean that at no point can the original VM be unpaused after the forks are gone. I don't see any usecase in which that would make any sense at all.
Quoted 6 lines
You could have the disk(s) as read-only and the VM running completely from RAM. Alpine-linux has (or had) a mode where it was completely stateless and running from RAM. I think it's fine to require passing a config file for the time being, we can look at other options afterwards.
Quoted 3 lines
OK, there is that. But I would say that's a fairly niche use-case. You wouldn't have any network access in that fork, no disk, no way to get information in or out beside the serial console.
Why won't the fork have network access?
If the parent VM is left paused the fork should behave like a local migration regarding network access, and thus be fully functional.
Quoted 4 lines
So I wouldn't want
that setup to be considered the default. If someone wants to that I would rather have an option that tells xl to automatically name the fork for you instead of the other way around.
Ack, I just want to make sure that whatever interface we end up using is designed taking into account other use cases apart from the one at hand.
On an unrelated note, does forking work when using PV interfaces?
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Quoted 41 lines · Roger Pau Monné
Quoted 39 lines · Tamas K Lengyel
Quoted 37 lines · Roger Pau Monné
Quoted 21 lines · Tamas K Lengyel
Quoted 14 lines · Roger Pau Monné
Quoted 11 lines · Tamas K Lengyel
Quoted 5 lines · Roger Pau Monné
I think you also need something like:
# xl fork-vm --launch-dm late <parent_domid> <fork_domid>So that a user doesn't need to pass a qemu-save-file?
This doesn't make much sense to me. To launch QEMU you need the config file to wire things up correctly. Like in order to launch QEMU you need to tell it the name of the VM, disk path, etc. that are all contained in the config.
Quoted 5 lines
You could get all this information from the parent VM, IIRC libxl has a json version of the config. For example for migration there's no need to pass any config file, since the incoming VM can be recreated from the data in the source VM.
Quoted 7 lines
But again, creating a fork with the exact config of the parent is not possible. Even if the tool would rename the fork on-the-fly as it does during the migration, the fork would end up thrashing the parent VM's disk and making it impossible to create any additional forks. It would also mean that at no point can the original VM be unpaused after the forks are gone. I don't see any usecase in which that would make any sense at all.
You could have the disk(s) as read-only and the VM running completely from RAM. Alpine-linux has (or had) a mode where it was completely stateless and running from RAM. I think it's fine to require passing a config file for the time being, we can look at other options afterwards.
Quoted 5 lines
Quoted 3 lines
OK, there is that. But I would say that's a fairly niche use-case. You wouldn't have any network access in that fork, no disk, no way to get information in or out beside the serial console.
Why won't the fork have network access?
If you have multiple forks you end up having MAC-address collision. I don't see what would be the point of creating a single fork when the parent remains paused - you could just keep running the parent since you aren't gaining anything by creating the fork. The main reason to create a fork would be to create multiples of them.
Quoted 9 lines
If the parent VM is left paused the fork should behave like a local migration regarding network access, and thus be fully functional.
Quoted 4 lines
So I wouldn't want
that setup to be considered the default. If someone wants to that I would rather have an option that tells xl to automatically name the fork for you instead of the other way around.
Quoted 5 lines
Ack, I just want to make sure that whatever interface we end up using is designed taking into account other use cases apart from the one at hand.
On an unrelated note, does forking work when using PV interfaces?
As I recall yes, but In my Linux tests these were the config options I tested and work with the fork. Not sure if the vif device by default is PV or emulated:
vnc=1
vnclisten="0.0.0.0:1"
usb=1
usbdevice=['tablet']
disk = ['phy:/dev/t0vg/debian-stretch,xvda,w']
vif = ['bridge=xenbr0,mac=00:07:5B:BB:00:01']
Tamas
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel