Xen on Hyper-V: Implement L0 assisted TLB flush
v2 cover Xen on Hyper-V: Implement L0 assisted TLB flush
Hi all
This seris is based on Roger's L0 assisted flush series.
I have done some testing against a Linux on Hyper-V in a 32-vcpu VM. All builds were done with -j32.
Building Xen on Linux:
real 0m45.376s user 2m28.156s sys 0m51.672s
Building Xen on Linux on Xen on Hyper-V, no assisted flush:
real 3m8.762s user 10m46.787s sys 30m14.492s
Building Xen on Linux on Xen on Hyper-V, with assisted flush:
real 0m44.369s user 3m16.231s sys 3m3.330s
Building Linux x86_64_defconfig on Linux:
real 0m59.698s user 21m14.014s sys 2m58.742s
Building Linux x86_64_defconfig on Linux on Xen on Hyper-V, no assisted flush:
real 2m6.284s user 31m18.706s sys 20m31.106s
Building Linux x86_64_defconfig on Linux on Xen on Hyper-V, with assisted flush:
real 1m38.968s user 28m40.398s sys 11m20.151s
There are various degrees of improvement depending on the workload. Xen can perhaps be optmised a bit more because it currently doesn't pass the address space id (cr3) to Hyper-V, but that requires reworking TLB flush APIs within Xen.
Wei.
Wei Liu (3):
x86/hypervisor: pass flags to hypervisor_flush_tlb x86/hyperv: skeleton for L0 assisted TLB flush x86/hyperv: L0 assisted TLB flush
xen/arch/x86/guest/hyperv/Makefile | 2 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++ xen/arch/x86/guest/hyperv/private.h | 13 ++ xen/arch/x86/guest/hyperv/tlb.c | 211 +++++++++++++++++++++++++ xen/arch/x86/guest/hyperv/util.c | 74 +++++++++ xen/arch/x86/guest/hypervisor.c | 7 +- xen/arch/x86/guest/xen/xen.c | 2 +- xen/arch/x86/smp.c | 5 +- xen/include/asm-x86/flushtlb.h | 3 + xen/include/asm-x86/guest/hypervisor.h | 10 +- 10 files changed, 333 insertions(+), 11 deletions(-) create mode 100644 xen/arch/x86/guest/hyperv/tlb.c create mode 100644 xen/arch/x86/guest/hyperv/util.c
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 1/3 x86/hypervisor: pass flags to hypervisor_flush_tlb
Hyper-V's L0 assisted flush has fine-grained control over what gets flushed. We need all the flags available to make the best decisions possible.
No functional change because Xen's implementation doesn't care about what is passed to it.
---
v2:
1. Introduce FLUSH_TLB_FLAGS_MASK
--- xen/arch/x86/guest/hypervisor.c | 7 +++++-- xen/arch/x86/guest/xen/xen.c | 2 +- xen/arch/x86/smp.c | 5 ++--- xen/include/asm-x86/flushtlb.h | 3 +++ xen/include/asm-x86/guest/hypervisor.h | 10 +++++----- 5 files changed, 16 insertions(+), 11 deletions(-)
xen/arch/x86/guest/hypervisor.c +5 -2
@@ -75,10 +75,13 @@ void __init hypervisor_e820_fixup(struct e820map *e820) }int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order) +unsigned int flags) { +if ( flags & ~FLUSH_TLB_FLAGS_MASK ) +return -EINVAL ; +if ( ops.flush_tlb ) -return alternative_call (ops.flush_tlb, mask, va, order); +return alternative_call (ops.flush_tlb, mask, va, flags);return -ENOSYS ; }
xen/arch/x86/guest/xen/xen.c +1 -1
@@ -324,7 +324,7 @@ static void __init e820_fixup(struct e820map *e820)pv_shim_fixup_e820 (e820); } -static int flush_tlb (const cpumask_t *mask,const void *va,unsigned int order) +static int flush_tlb (const cpumask_t *mask,const void *va,unsigned int flags) {return xen_hypercall_hvm_op (HVMOP_flush_tlbs,NULL ); }
xen/arch/x86/smp.c +2 -3
@@ -258,9 +258,8 @@ void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags) !cpumask_subset(mask, cpumask_of(cpu)) ) {if ( cpu_has_hypervisor && - !(flags & ~(FLUSH_TLB |FLUSH_TLB_GLOBAL |FLUSH_VA_VALID | -FLUSH_ORDER_MASK )) && - !hypervisor_flush_tlb (mask, va, flags &FLUSH_ORDER_MASK ) ) + !(flags & ~FLUSH_TLB_FLAGS_MASK ) && + !hypervisor_flush_tlb (mask, va, flags) ) {if ( tlb_clk_enabled ) tlb_clk_enabled = false;
xen/include/asm-x86/flushtlb.h +3 -0
@@ -123,6 +123,9 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);/* Flush all HVM guests linear TLB (using ASID/VPID) */ #define FLUSH_GUESTS_TLB 0x4000 +#define FLUSH_TLB_FLAGS_MASK (FLUSH_TLB | FLUSH_TLB_GLOBAL | FLUSH_VA_VALID | \ + FLUSH_ORDER_MASK) +/* Flush local TLBs/caches. */ unsigned int flush_area_local (const void *va,unsigned int flags);#define flush_local (flags) flush_area_local(NULL, flags)
xen/include/asm-x86/guest/hypervisor.h +5 -5
@@ -35,7 +35,7 @@ struct hypervisor_ops {/* Fix up e820 map */ void (*e820_fixup)(struct e820map *e820);/* L0 assisted TLB flush */ -int (*flush_tlb)(const cpumask_t *mask,const void *va,unsigned int order); +int (*flush_tlb)(const cpumask_t *mask,const void *va,unsigned int flags); };#ifdef CONFIG_GUEST
@@ -48,11 +48,11 @@ void hypervisor_e820_fixup(struct e820map *e820);/* * L0 assisted TLB flush. * mask: cpumask of the dirty vCPUs that should be flushed. -* va: linear address to flush, or NULL for global flushes. -* order: order of the linear address pointed by va. +* va: linear address to flush, or NULL for entire address space. +* flags: flags for flushing, including the order of va. */ int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order); +unsigned int flags);#else
@@ -65,7 +65,7 @@ static inline int hypervisor_ap_setup(void) { return 0; }static inline void hypervisor_resume (void ) {ASSERT_UNREACHABLE (); }static inline void hypervisor_e820_fixup (struct e820map *e820) {}static inline int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order) +unsigned int flags) {return -ENOSYS ; }
Signature · 7 lines
2.20.1 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v2 2/3 x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Wei LiuImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.c
xen/arch/x86/guest/hyperv/Makefile +1 -0
@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.o
xen/arch/x86/guest/hyperv/hyperv.c +17 -0
@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};
@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT ); + + hv_hcall_page_ready = true; }static int setup_hypercall_pcpu_arg (void )
@@ -199,11 +203,24 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); } +static int flush_tlb (const cpumask_t *mask,const void *va, +unsigned int flags) +{ +if ( !(ms_hyperv.hints &HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED ) ) +return -EOPNOTSUPP ; + +if ( !hv_hcall_page_ready || !this_cpu (hv_input_page) ) +return -ENXIO ; + +return hyperv_flush_tlb (mask, va, flags); +} +static const struct hypervisor_ops __initdata ops = { .name ="Hyper-V" , .setup = setup, .ap_setup = ap_setup, .e820_fixup = e820_fixup, + .flush_tlb = flush_tlb, };/*
xen/arch/x86/guest/hyperv/private.h +4 -0
@@ -22,10 +22,14 @@#ifndef __XEN_HYPERV_PRIVIATE_H__#define __XEN_HYPERV_PRIVIATE_H__ +#include <xen/cpumask.h> #include <xen/percpu.h> DECLARE_PER_CPU(void *, hv_input_page); DECLARE_PER_CPU(void *, hv_vp_assist); DECLARE_PER_CPU(unsigned int, hv_vp_index); +int hyperv_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int flags); +#endif /* __XEN_HYPERV_PRIVIATE_H__ */
xen/arch/x86/guest/hyperv/tlb.c new file +41 -0
41 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 3/3 x86/hyperv: L0 assisted TLB flush
In reply to Wei LiuImplement L0 assisted TLB flush for Xen on Hyper-V. It takes advantage of several hypercalls:
* HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST
* HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX
* HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE
* HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX
Pick the most efficient hypercalls available.
---
v2:
1. Address Roger and Jan's comments re types etc.
2. Fix pointer arithmetic.
3. Misc improvement to code.
--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/private.h | 9 ++ xen/arch/x86/guest/hyperv/tlb.c | 172 +++++++++++++++++++++++++++- xen/arch/x86/guest/hyperv/util.c | 74 ++++++++++++ 4 files changed, 255 insertions(+), 1 deletion(-) create mode 100644 xen/arch/x86/guest/hyperv/util.c
xen/arch/x86/guest/hyperv/Makefile +1 -0
@@ -1,2 +1,3 @@ obj-y += hyperv.o obj-y += tlb.o +obj-y += util.o
xen/arch/x86/guest/hyperv/private.h +9 -0
@@ -24,12 +24,21 @@#include <xen/cpumask.h> #include <xen/percpu.h> +#include <xen/types.h> DECLARE_PER_CPU(void *, hv_input_page); DECLARE_PER_CPU(void *, hv_vp_assist); DECLARE_PER_CPU(unsigned int, hv_vp_index); +static inline unsigned int hv_vp_index (unsigned int cpu) +{ +return per_cpu (hv_vp_index, cpu); +} +int hyperv_flush_tlb (const cpumask_t *mask,const void *va,unsigned int flags); +/* Returns number of banks, -ev if error */ +int cpumask_to_vpset (struct hv_vpset *vpset,const cpumask_t *mask); +#endif /* __XEN_HYPERV_PRIVIATE_H__ */
xen/arch/x86/guest/hyperv/tlb.c +171 -1
186 lines not on this page · raw message
xen/arch/x86/guest/hyperv/util.c new file +74 -0
74 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
Re: [Xen-devel] [PATCH v2 3/3] x86/hyperv: L0 assisted TLB flush
Quoted 11 lines · Wei Liu
Implement L0 assisted TLB flush for Xen on Hyper-V. It takes advantage of several hypercalls:
*
HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST
*HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX
*HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE
*HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EXPick the most efficient hypercalls available.
Signed-off-by Wei Liu <[email protected]>
Thanks! LGTM, I've just got a couple of comments below.
Quoted 83 lines
---
v2:
1. Address Roger and Jan's comments re types etc.
2. Fix pointer arithmetic.
3. Misc improvement to code.--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/private.h | 9 ++ xen/arch/x86/guest/hyperv/tlb.c | 172 +++++++++++++++++++++++++++- xen/arch/x86/guest/hyperv/util.c | 74 ++++++++++++ 4 files changed, 255 insertions(+), 1 deletion(-) create mode 100644 xen/arch/x86/guest/hyperv/util.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1,2 +1,3 @@ obj-y += hyperv.o obj-y += tlb.o +obj-y += util.oxen/arch/x86/guest/hyperv/private.h +9
-0@@ -24,12 +24,21 @@#include <xen/cpumask.h> #include <xen/percpu.h> +#include <xen/types.h> DECLARE_PER_CPU(void *, hv_input_page); DECLARE_PER_CPU(void *, hv_vp_assist); DECLARE_PER_CPU(unsigned int, hv_vp_index); +static inline unsigned int hv_vp_index (unsigned int cpu) +{ +return per_cpu (hv_vp_index, cpu); +} +int hyperv_flush_tlb (const cpumask_t *mask,const void *va,unsigned int flags); +/* Returns number of banks, -ev if error */ +int cpumask_to_vpset (struct hv_vpset *vpset,const cpumask_t *mask); +#endif /* __XEN_HYPERV_PRIVIATE_H__ */ xen/arch/x86/guest/hyperv/tlb.c +23
-0@@ -19,15 +19,185 @@* Copyright (c) 2020 Microsoft. */ +#include <xen/cpu.h> #include <xen/cpumask.h> #include <xen/errno.h> +#include <asm/guest/hyperv.h> +#include <asm/guest/hyperv-hcall.h> +#include <asm/guest/hyperv-tlfs.h> +#include "private.h" +/* +* It is possible to encode up to 4096 pages using the lower 12 bits +* in an element of gva_list +*/ +#define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE) + +static unsigned int fill_gva_list (uint64_t *gva_list,const void *va, +unsigned int order) +{ +unsigned long start = (unsigned long )va; +unsigned long end = start + (PAGE_SIZE << order) -1 ; +unsigned int n =0 ; + + do { + unsigned long remain = end - start; + + gva_list[n] = start & PAGE_MASK; +
xen/arch/x86/guest/hyperv/tlb.c line 59
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+/* +* Use lower 12 bits to encode the number of additional pages +* to flush +*/ +if ( remain >=HV_TLB_FLUSH_UNIT ) + { + gva_list[n] |= ~PAGE_MASK ; + start +=HV_TLB_FLUSH_UNIT ; + } +else if ( remain )
remain is always going to be > 0, since the loop condition is end > start, and hence this can be a plain else.
Quoted 26 lines
xen/arch/x86/guest/hyperv/tlb.c +26
-0+ { + gva_list[n] |= (remain -1 ) >>PAGE_SHIFT ; + start = end; + } + + n++; + }while ( start < end ); + +return n; +} + +static uint64_t flush_tlb_ex (const cpumask_t *mask,const void *va, +unsigned int flags) +{ +struct hv_tlb_flush_ex *flush =this_cpu (hv_input_page); +int nr_banks; +unsigned int max_gvas, order = flags &FLUSH_ORDER_MASK ; +uint64_t ret; + +if ( !flush ||local_irq_is_enabled () ) + { +ASSERT_UNREACHABLE (); +return ~0ULL ; + } + +if ( !(ms_hyperv.hints &HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED ) )
xen/arch/x86/guest/hyperv/tlb.c line 95
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+return ~0ULL ; + + flush->address_space =0 ; + flush->flags =HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES ; +if ( !(flags &FLUSH_TLB_GLOBAL ) ) + flush->flags |=HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY ; + + nr_banks =cpumask_to_vpset (&flush->hv_vp_set, mask); +if ( nr_banks <0 ) +return ~0ULL ;
It would be nice to propagate the error code from cpumask_to_vpset, but since the function can also return HyperV error codes this doesn't make much sense.
xen/arch/x86/guest/hyperv/tlb.c line 108
Quoted 13 lines
xen/arch/x86/guest/hyperv/tlb.c +13
-0+ + max_gvas = + (PAGE_SIZE -sizeof (*flush) - nr_banks * +sizeof (flush->hv_vp_set.bank_contents[0 ])) / +sizeof (uint64_t );/* gva is represented as uint64_t */ + +/* +* Flush the entire address space if va is NULL or if there is not +* enough space for gva_list. +*/ +if ( !va || (PAGE_SIZE << order) /HV_TLB_FLUSH_UNIT > max_gvas ) + ret =hv_do_rep_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX ,0 , + nr_banks,virt_to_maddr (flush),0 );
You could just return hv_do_rep_hypercall(...); here, which will avoid the else branch below and the indentation.
Quoted 16 lines
xen/arch/x86/guest/hyperv/tlb.c +15
-0+else + { +uint64_t *gva_list = + (uint64_t *)flush +sizeof (*flush) /sizeof (uint64_t ) + nr_banks; +unsigned int gvas =fill_gva_list (gva_list, va, order); + +BUILD_BUG_ON (sizeof (*flush) %sizeof (uint64_t )); + + ret =hv_do_rep_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX , + gvas, nr_banks,virt_to_maddr (flush),0 ); + } + +return ret; +} + int hyperv_flush_tlb(const cpumask_t *mask, const void *va,
xen/arch/x86/guest/hyperv/tlb.c line 133
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +7
-1unsigned int flags) { -return -EOPNOTSUPP ; +unsigned long irq_flags; +struct hv_tlb_flush *flush =this_cpu (hv_input_page); +unsigned int max_gvas, order = flags &FLUSH_ORDER_MASK ; +uint64_t ret; + +ASSERT (flush); +ASSERT (!cpumask_empty (mask));
I would also turn this into an if ( ... ) { ASSERT; return -EFOO; }
Quoted 17 lines
xen/arch/x86/guest/hyperv/tlb.c +17
-0+ +local_irq_save (irq_flags); + + flush->address_space =0 ; + flush->flags =HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES ; + flush->processor_mask =0 ; +if ( !(flags &FLUSH_TLB_GLOBAL ) ) + flush->flags |=HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY ; + +if (cpumask_equal (mask, &cpu_online_map) ) + flush->flags |=HV_FLUSH_ALL_PROCESSORS ; +else + { +unsigned int cpu; + +/* +* Normally VP indices are in ascending order and match Xen's
xen/arch/x86/guest/hyperv/tlb.c line 160
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+* idea of CPU ids. Check the last index to see if VP index is +* >= 64. If so, we can skip setting up parameters for +* non-applicable hypercalls without looking further. +*/ +if (hv_vp_index (cpumask_last (mask)) >=64 ) + goto do_ex_hypercall; + + for_each_cpu ( cpu, mask ) + { +uint32_t vpid =hv_vp_index (cpu);
This should be unsigned int now.
xen/arch/x86/guest/hyperv/tlb.c line 172
Quoted 12 lines
xen/arch/x86/guest/hyperv/tlb.c +12
-0+ +if ( vpid > ms_hyperv.max_vp_index ) + { +local_irq_restore (irq_flags); +return -ENXIO ; + } + +if ( vpid >=64 ) + goto do_ex_hypercall; + +__set_bit (vpid, &flush->processor_mask); + }
Would it make sense to abstract this as cpumask_to_processor_mask, since you are adding cpumask_to_vpset below.
xen/arch/x86/guest/hyperv/tlb.c line 175
Quoted 3 lines
xen/arch/x86/guest/hyperv/tlb.c +3
-0+ } + + max_gvas = (PAGE_SIZE -sizeof (*flush)) /sizeof (flush->gva_list[0 ]);
You could init this at declaration, and make it const static since the value can be calculated at compile time AFAICT. Or create a define with it (HV_TLB_FLUSH_MAX_GVAS?). There's no need to store it on the stack.
xen/arch/x86/guest/hyperv/tlb.c line 186
Quoted 11 lines
xen/arch/x86/guest/hyperv/tlb.c +11
-0+ +/* +* Flush the entire address space if va is NULL or if there is not +* enough space for gva_list. +*/ +if ( !va || (PAGE_SIZE << order) /HV_TLB_FLUSH_UNIT > max_gvas ) + ret =hv_do_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE , +virt_to_maddr (flush),0 ); +else + { +unsigned int gvas =fill_gva_list (flush->gva_list, va, order);
No need for the gvas variable, you can just call fill_gva_list at hv_do_rep_hypercall.
Quoted 56 lines
xen/arch/x86/guest/hyperv/tlb.c +14
-0+ + ret =hv_do_rep_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST , gvas,0 , +virt_to_maddr (flush),0 ); + } + + goto done; + + do_ex_hypercall: + ret =flush_tlb_ex (mask, va, flags); + + done: +local_irq_restore (irq_flags); + +return ret &HV_HYPERCALL_RESULT_MASK ? -ENXIO :0 ; }/* xen/arch/x86/guest/hyperv/util.c new file +33
-0@@ -0,0 +1,74 @@ +/****************************************************************************** +* arch/x86/guest/hyperv/util.c +* +* Hyper-V utility functions +* +* This program is free software; you can redistribute it and/or modify +* it under the terms of the GNU General Public License as published by +* the Free Software Foundation; either version 2 of the License, or +* (at your option) any later version. +* +* This program is distributed in the hope that it will be useful, +* but WITHOUT ANY WARRANTY; without even the implied warranty of +* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +* GNU General Public License for more details. +* +* You should have received a copy of the GNU General Public License +* along with this program; If not, see <http://www.gnu.org/licenses/>. +* +* Copyright (c) 2020 Microsoft. +*/ + +#include <xen/cpu.h> +#include <xen/cpumask.h> +#include <xen/errno.h> + +#include <asm/guest/hyperv.h> +#include <asm/guest/hyperv-tlfs.h> + +#include "private.h" + +int cpumask_to_vpset (struct hv_vpset *vpset, +const cpumask_t *mask) +{
xen/arch/x86/guest/hyperv/util.c line 43
Quoted 10 lines
xen/arch/x86/guest/hyperv/util.c new file +10
-0+int nr =1 ; +unsigned int cpu, vcpu_bank, vcpu_offset; +unsigned int max_banks = ms_hyperv.max_vp_index /64 ; + +/* Up to 64 banks can be represented by valid_bank_mask */ +if ( max_banks >=64 ) +return -E2BIG ; + +/* Clear all banks to avoid flushing unwanted CPUs */ +for ( vcpu_bank =0 ; vcpu_bank <= max_banks; vcpu_bank++ )
I think this is off by one and should be vcpu_bank < max_banks? Or else you are clearing one extra bank.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 1/3] x86/hypervisor: pass flags to hypervisor_flush_tlb
In reply to Wei LiuQuoted 8 lines · Wei Liu
Hyper-V's L0 assisted flush has fine-grained control over what gets flushed. We need all the flags available to make the best decisions possible.
No functional change because Xen's implementation doesn't care about what is passed to it.
Signed-off-by Wei Liu <[email protected]>
LGTM:
Just one comment below.
Quoted 16 lines
---
v2:
1. IntroduceFLUSH_TLB_FLAGS_MASK--- xen/arch/x86/guest/hypervisor.c | 7 +++++-- xen/arch/x86/guest/xen/xen.c | 2 +- xen/arch/x86/smp.c | 5 ++--- xen/include/asm-x86/flushtlb.h | 3 +++ xen/include/asm-x86/guest/hypervisor.h | 10 +++++----- 5 files changed, 16 insertions(+), 11 deletions(-)xen/arch/x86/guest/hypervisor.c +0
-0
xen/arch/x86/guest/hypervisor.c void __init hypervisor_e820_fixup(struct e820map *e820) · line 80
Quoted 7 lines
xen/arch/x86/guest/hypervisor.c +2
-1@@ -75,10 +75,13 @@ void __init hypervisor_e820_fixup(struct e820map *e820) }int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order) +unsigned int flags) { +if ( flags & ~FLUSH_TLB_FLAGS_MASK )
I think an ASSERT_UNREACHABLE() would be good here, since you are not supposed to call hypervisor_flush_tlb with non TLB related flags.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Wei LiuQuoted 7 lines · Wei Liu
Implement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>
Just one comment below.
Quoted 20 lines
--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +0
-0
xen/arch/x86/guest/hyperv/hyperv.c DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); · line 36
Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c +1
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready;
Since this is static, I would drop the hv_ prefix.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 3/3] x86/hyperv: L0 assisted TLB flush
In reply to Roger Pau MonnéOn Fri, Feb 14, 2020 at 03:42:17PM +0100, Roger Pau Monné wrote:
[...]
Quoted 42 lines
Quoted 42 lines
#endif /*
__XEN_HYPERV_PRIVIATE_H__*/xen/arch/x86/guest/hyperv/tlb.c +28
-0@@ -19,15 +19,185 @@* Copyright (c) 2020 Microsoft. */ +#include <xen/cpu.h> #include <xen/cpumask.h> #include <xen/errno.h> +#include <asm/guest/hyperv.h> +#include <asm/guest/hyperv-hcall.h> +#include <asm/guest/hyperv-tlfs.h> +#include "private.h" +/* +* It is possible to encode up to 4096 pages using the lower 12 bits +* in an element of gva_list +*/ +#define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE) + +static unsigned int fill_gva_list (uint64_t *gva_list,const void *va, +unsigned int order) +{ +unsigned long start = (unsigned long )va; +unsigned long end = start + (PAGE_SIZE << order) -1 ; +unsigned int n =0 ; + + do { + unsigned long remain = end - start; + + gva_list[n] = start & PAGE_MASK; + + /* + * Use lower 12 bits to encode the number of additional pages + * to flush + */ + if ( remain >= HV_TLB_FLUSH_UNIT )
Quoted 8 lines
xen/arch/x86/guest/hyperv/tlb.c line 59
Quoted 5 lines
xen/arch/x86/guest/hyperv/tlb.c +5
-0+ { + gva_list[n] |= ~PAGE_MASK ; + start +=HV_TLB_FLUSH_UNIT ; + } + else if ( remain )remain is always going to be > 0, since the loop condition is end > start, and hence this can be a plain else.
Ack.
Quoted 42 lines
Quoted 27 lines
xen/arch/x86/guest/hyperv/tlb.c +26
-0+ { + gva_list[n] |= (remain -1 ) >>PAGE_SHIFT ; + start = end; + } + + n++; + }while ( start < end ); + +return n; +} + +static uint64_t flush_tlb_ex (const cpumask_t *mask,const void *va, +unsigned int flags) +{ +struct hv_tlb_flush_ex *flush =this_cpu (hv_input_page); +int nr_banks; +unsigned int max_gvas, order = flags &FLUSH_ORDER_MASK ; +uint64_t ret; + +if ( !flush ||local_irq_is_enabled () ) + { +ASSERT_UNREACHABLE (); +return ~0ULL ; + } + +if ( !(ms_hyperv.hints &HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED ) )xen/arch/x86/guest/hyperv/tlb.c line 95
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+return ~0ULL ; + + flush->address_space =0 ; + flush->flags =HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES ; +if ( !(flags &FLUSH_TLB_GLOBAL ) ) + flush->flags |=HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY ; + + nr_banks =cpumask_to_vpset (&flush->hv_vp_set, mask); +if ( nr_banks <0 ) +return ~0ULL ;It would be nice to propagate the error code from
cpumask_to_vpset, but since the function can also return HyperV error codes this doesn't make much sense.
Quoted 16 lines
xen/arch/x86/guest/hyperv/tlb.c line 108
Quoted 13 lines
xen/arch/x86/guest/hyperv/tlb.c +13
-0+ + max_gvas = + (PAGE_SIZE -sizeof (*flush) - nr_banks * +sizeof (flush->hv_vp_set.bank_contents[0 ])) / +sizeof (uint64_t );/* gva is represented as uint64_t */ + +/* +* Flush the entire address space if va is NULL or if there is not +* enough space for gva_list. +*/ +if ( !va || (PAGE_SIZE << order) /HV_TLB_FLUSH_UNIT > max_gvas ) + ret =hv_do_rep_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX ,0 , + nr_banks,virt_to_maddr (flush),0 );You could just return
hv_do_rep_hypercall(...); here, which will avoid the else branch below and the indentation.
Ack.
Quoted 21 lines
Quoted 21 lines
xen/arch/x86/guest/hyperv/tlb.c +16
-1+else + { +uint64_t *gva_list = + (uint64_t *)flush +sizeof (*flush) /sizeof (uint64_t ) + nr_banks; +unsigned int gvas =fill_gva_list (gva_list, va, order); + +BUILD_BUG_ON (sizeof (*flush) %sizeof (uint64_t )); + + ret =hv_do_rep_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX , + gvas, nr_banks,virt_to_maddr (flush),0 ); + } + +return ret; +} +int hyperv_flush_tlb (const cpumask_t *mask,const void *va,unsigned int flags) { -return -EOPNOTSUPP ; +unsigned long irq_flags;
Quoted 8 lines
xen/arch/x86/guest/hyperv/tlb.c line 133
Quoted 6 lines
xen/arch/x86/guest/hyperv/tlb.c +6
-0+struct hv_tlb_flush *flush =this_cpu (hv_input_page); +unsigned int max_gvas, order = flags &FLUSH_ORDER_MASK ; +uint64_t ret; + +ASSERT (flush); +ASSERT (!cpumask_empty (mask));I would also turn this into an if ( ... ) { ASSERT; return
-EFOO; }
Ack.
Quoted 22 lines
Quoted 22 lines
xen/arch/x86/guest/hyperv/tlb.c +21
-0+ +local_irq_save (irq_flags); + + flush->address_space =0 ; + flush->flags =HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES ; + flush->processor_mask =0 ; +if ( !(flags &FLUSH_TLB_GLOBAL ) ) + flush->flags |=HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY ; + +if (cpumask_equal (mask, &cpu_online_map) ) + flush->flags |=HV_FLUSH_ALL_PROCESSORS ; +else + { +unsigned int cpu; + +/* +* Normally VP indices are in ascending order and match Xen's +* idea of CPU ids. Check the last index to see if VP index is +* >= 64. If so, we can skip setting up parameters for +* non-applicable hypercalls without looking further. +*/
Quoted 8 lines
xen/arch/x86/guest/hyperv/tlb.c line 160
Quoted 6 lines
xen/arch/x86/guest/hyperv/tlb.c +6
-0+if (hv_vp_index (cpumask_last (mask)) >=64 ) + goto do_ex_hypercall; + + for_each_cpu ( cpu, mask ) + { +uint32_t vpid =hv_vp_index (cpu);This should be unsigned int now.
Good catch.
Quoted 17 lines
xen/arch/x86/guest/hyperv/tlb.c line 172
Quoted 13 lines
xen/arch/x86/guest/hyperv/tlb.c +12
-0+ +if ( vpid > ms_hyperv.max_vp_index ) + { +local_irq_restore (irq_flags); +return -ENXIO ; + } + +if ( vpid >=64 ) + goto do_ex_hypercall; + +__set_bit (vpid, &flush->processor_mask); + }Would it make sense to abstract this as
cpumask_to_processor_mask, since you are addingcpumask_to_vpsetbelow.
There is only one usage so far, so I don't think it is necessary.
Quoted 8 lines
xen/arch/x86/guest/hyperv/tlb.c line 175
Quoted 3 lines
xen/arch/x86/guest/hyperv/tlb.c +3
-0+ } + + max_gvas = (PAGE_SIZE -sizeof (*flush)) /sizeof (flush->gva_list[0 ]);You could init this at declaration, and make it const static since the value can be calculated at compile time AFAICT. Or create a define with it (
HV_TLB_FLUSH_MAX_GVAS?). There's no need to store it on the stack.
I can introduce a define, but the name you suggested is too generic. If there is variable sized header, the calculation is going to be different.
Quoted 13 lines
xen/arch/x86/guest/hyperv/tlb.c line 186
Quoted 12 lines
xen/arch/x86/guest/hyperv/tlb.c +11
-0+ +/* +* Flush the entire address space if va is NULL or if there is not +* enough space for gva_list. +*/ +if ( !va || (PAGE_SIZE << order) /HV_TLB_FLUSH_UNIT > max_gvas ) + ret =hv_do_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE , +virt_to_maddr (flush),0 ); +else + { +unsigned int gvas =fill_gva_list (flush->gva_list, va, order);
Quoted 3 lines
No need for the gvas variable, you can just call
fill_gva_listathv_do_rep_hypercall.
Sure.
Quoted 61 lines
Quoted 61 lines
xen/arch/x86/guest/hyperv/tlb.c +14
-0+ + ret =hv_do_rep_hypercall (HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST , gvas,0 , +virt_to_maddr (flush),0 ); + } + + goto done; + + do_ex_hypercall: + ret =flush_tlb_ex (mask, va, flags); + + done: +local_irq_restore (irq_flags); + +return ret &HV_HYPERCALL_RESULT_MASK ? -ENXIO :0 ; }/* xen/arch/x86/guest/hyperv/util.c new file +38
-0@@ -0,0 +1,74 @@ +/****************************************************************************** +* arch/x86/guest/hyperv/util.c +* +* Hyper-V utility functions +* +* This program is free software; you can redistribute it and/or modify +* it under the terms of the GNU General Public License as published by +* the Free Software Foundation; either version 2 of the License, or +* (at your option) any later version. +* +* This program is distributed in the hope that it will be useful, +* but WITHOUT ANY WARRANTY; without even the implied warranty of +* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +* GNU General Public License for more details. +* +* You should have received a copy of the GNU General Public License +* along with this program; If not, see <http://www.gnu.org/licenses/>. +* +* Copyright (c) 2020 Microsoft. +*/ + +#include <xen/cpu.h> +#include <xen/cpumask.h> +#include <xen/errno.h> + +#include <asm/guest/hyperv.h> +#include <asm/guest/hyperv-tlfs.h> + +#include "private.h" + +int cpumask_to_vpset (struct hv_vpset *vpset, +const cpumask_t *mask) +{ +int nr =1 ; +unsigned int cpu, vcpu_bank, vcpu_offset; +unsigned int max_banks = ms_hyperv.max_vp_index /64 ; + +/* Up to 64 banks can be represented by valid_bank_mask */
Quoted 8 lines
xen/arch/x86/guest/hyperv/util.c line 43
Quoted 5 lines
xen/arch/x86/guest/hyperv/util.c new file +5
-0+if ( max_banks >=64 ) +return -E2BIG ; + +/* Clear all banks to avoid flushing unwanted CPUs */ +for ( vcpu_bank =0 ; vcpu_bank <= max_banks; vcpu_bank++ )I think this is off by one and should be
vcpu_bank<max_banks? Or else you are clearing one extra bank.
If max_vp_index were the maximum vp index, max_banks would be inclusive (for example, max_vp_index is 63, 63/64=0). The code should be correct in that case.
However, when I did another round of self-review, I discovered that the value stored in max_vp_index was actually "the maximum number of virtual processors supported", so it wasn't really the maximum index.
This means I will need to adjust places where max_vp_index is used.
Here, max_banks is not inclusive anymore.
And an earlier place
+ if ( vpid > ms_hyperv.max_vp_index )
+ {
also requires fixing.
I will also see about changing max_{l,v}p_index to more sensible names or add some comments to their definitions.
Wei.
Quoted 2 lines
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 3/3] x86/hyperv: L0 assisted TLB flush
In reply to Wei LiuFrom: Wei Liu <[email protected]> On Behalf Of Wei Liu Sent: Friday, February 14, 2020 4:35 AM
Quoted 82 lines
Implement L0 assisted TLB flush for Xen on Hyper-V. It takes advantage of several hypercalls:
*
HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST
*HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX
*HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE
*HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EXPick the most efficient hypercalls available.
Signed-off-by Wei Liu <[email protected]>---
v2:
1. Address Roger and Jan's comments re types etc.
2. Fix pointer arithmetic.
3. Misc improvement to code.--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/private.h | 9 ++ xen/arch/x86/guest/hyperv/tlb.c | 172 +++++++++++++++++++++++++++- xen/arch/x86/guest/hyperv/util.c | 74 ++++++++++++ 4 files changed, 255 insertions(+), 1 deletion(-) create mode 100644 xen/arch/x86/guest/hyperv/util.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1,2 +1,3 @@ obj-y += hyperv.o obj-y += tlb.o +obj-y += util.oxen/arch/x86/guest/hyperv/private.h +9
-0@@ -24,12 +24,21 @@#include <xen/cpumask.h> #include <xen/percpu.h> +#include <xen/types.h> DECLARE_PER_CPU(void *, hv_input_page); DECLARE_PER_CPU(void *, hv_vp_assist); DECLARE_PER_CPU(unsigned int, hv_vp_index); +static inline unsigned int hv_vp_index (unsigned int cpu) +{ +return per_cpu (hv_vp_index, cpu); +} +int hyperv_flush_tlb (const cpumask_t *mask,const void *va,unsigned int flags); +/* Returns number of banks, -ev if error */ +int cpumask_to_vpset (struct hv_vpset *vpset,const cpumask_t *mask); +#endif /* __XEN_HYPERV_PRIVIATE_H__ */ xen/arch/x86/guest/hyperv/tlb.c +10
-0@@ -19,15 +19,185 @@* Copyright (c) 2020 Microsoft. */ +#include <xen/cpu.h> #include <xen/cpumask.h> #include <xen/errno.h> +#include <asm/guest/hyperv.h> +#include <asm/guest/hyperv-hcall.h> +#include <asm/guest/hyperv-tlfs.h> +#include "private.h" +/* +* It is possible to encode up to 4096 pages using the lower 12 bits +* in an element of gva_list +*/ +#define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
xen/arch/x86/guest/hyperv/tlb.c line 46
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+ +static unsigned int fill_gva_list (uint64_t *gva_list,const void *va, +unsigned int order) +{ +unsigned long start = (unsigned long )va; +unsigned long end = start + (PAGE_SIZE << order) -1 ; +unsigned int n =0 ; + + do { + unsigned long remain = end - start;
The calculated value here isn't actually the remaining bytes in the range to flush -- it's one less than the remaining bytes in the range
to flush because of the -1 in the calculation of 'end'. That difference
will mess up the comparison below against HV_TLB_FLUSH_UNIT in the case that there are exactly 4096 page remaining to be flushed. It should take the "=" case, but won't. Also, the '-1' in 'remain - 1' in the else clause becomes unneeded, and the 'start = end' assignment then propagates the error.
In the parallel code in Linux, if you follow the call sequence to get to fill_gav_list(), the 'end' argument is really the address of the first byte of the first page that isn't in the flush range (i.e., one beyond the true 'end') and so is a bit misnamed.
I think the calculation of 'end' should drop the -1, and perhaps 'end' should be renamed.
Michael
Quoted 14 lines
xen/arch/x86/guest/hyperv/tlb.c +14
-0+ + gva_list[n] = start &PAGE_MASK ; + +/* +* Use lower 12 bits to encode the number of additional pages +* to flush +*/ +if ( remain >=HV_TLB_FLUSH_UNIT ) + { + gva_list[n] |= ~PAGE_MASK ; + start +=HV_TLB_FLUSH_UNIT ; + } +else if ( remain ) + {
xen/arch/x86/guest/hyperv/tlb.c line 70
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+ gva_list[n] |= (remain -1 ) >>PAGE_SHIFT ; + start = end; + } + + n++; + }while ( start < end ); + +return n; +} +
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 1/3] x86/hypervisor: pass flags to hypervisor_flush_tlb
In reply to Wei LiuQuoted 10 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Wei Liu <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 1/3] x86/hypervisor: pass flags tohypervisor_flush_tlb
Quoted 8 lines
Hyper-V's L0 assisted flush has fine-grained control over what gets flushed. We need all the flags available to make the best decisions possible.
No functional change because Xen's implementation doesn't care about what is passed to it.
Signed-off-by Wei Liu <[email protected]>
Quoted 123 lines
---
v2:
1. IntroduceFLUSH_TLB_FLAGS_MASK--- xen/arch/x86/guest/hypervisor.c | 7 +++++-- xen/arch/x86/guest/xen/xen.c | 2 +- xen/arch/x86/smp.c | 5 ++--- xen/include/asm-x86/flushtlb.h | 3 +++ xen/include/asm-x86/guest/hypervisor.h | 10 +++++----- 5 files changed, 16 insertions(+), 11 deletions(-)xen/arch/x86/guest/hypervisor.c +5
-2@@ -75,10 +75,13 @@ void __init hypervisor_e820_fixup(struct e820map *e820) }int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order) +unsigned int flags) { +if ( flags & ~FLUSH_TLB_FLAGS_MASK ) +return -EINVAL ; +if ( ops.flush_tlb ) -return alternative_call (ops.flush_tlb, mask, va, order); +return alternative_call (ops.flush_tlb, mask, va, flags);return -ENOSYS ;}
xen/arch/x86/guest/xen/xen.c +1
-1@@ -324,7 +324,7 @@ static void __init e820_fixup(struct e820map *e820)pv_shim_fixup_e820 (e820); } -static int flush_tlb (const cpumask_t *mask,const void *va,unsigned int order) +static int flush_tlb(const cpumask_t *mask, const void *va, unsigned int flags) {return xen_hypercall_hvm_op(HVMOP_flush_tlbs, NULL);}
xen/arch/x86/smp.c +2
-3@@ -258,9 +258,8 @@ void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags) !cpumask_subset(mask, cpumask_of(cpu)) ) {if ( cpu_has_hypervisor && - !(flags & ~(FLUSH_TLB |FLUSH_TLB_GLOBAL |FLUSH_VA_VALID | -FLUSH_ORDER_MASK )) && - !hypervisor_flush_tlb (mask, va, flags &FLUSH_ORDER_MASK ) ) + !(flags & ~FLUSH_TLB_FLAGS_MASK ) && + !hypervisor_flush_tlb (mask, va, flags) ) {if ( tlb_clk_enabled )tlb_clk_enabled = false;xen/include/asm- renamed from xen/include/asm-x86/flushtlb.h +0
-0
x86/flushtlb.h
index 9773014320..a4de317452 100644xen/include/asm-x86/flushtlb.h +3
-0@@ -123,6 +123,9 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);/* Flush all HVM guests linear TLB (using ASID/VPID) */ #define FLUSH_GUESTS_TLB 0x4000 +#define FLUSH_TLB_FLAGS_MASK (FLUSH_TLB | FLUSH_TLB_GLOBAL |FLUSH_VA_VALID | \ + FLUSH_ORDER_MASK) +/* Flush local TLBs/caches. */ unsigned int
flush_area_local(const void *va, unsigned int flags); #defineflush_local(flags)flush_area_local(NULL, flags)xen/include/asm- renamed from xen/include/asm-x86/guest/hypervisor.h +0
-0
x86/guest/hypervisor.h
index432e57c2a0..48d54735d2100644xen/include/asm-x86/guest/hypervisor.h +1
-1@@ -35,7 +35,7 @@ struct hypervisor_ops {/* Fix up e820 map */ void (*e820_fixup)(struct e820map *e820);/* L0 assisted TLB flush */ -int (*flush_tlb)(const cpumask_t *mask,const void *va,unsigned int order); +int (*flush_tlb)(const cpumask_t *mask,const void *va,unsigned int flags); };#ifdef
CONFIG_GUESTxen/include/asm-x86/guest/hypervisor.h +4
-4@@ -48,11 +48,11 @@ void hypervisor_e820_fixup(struct e820map *e820);/* * L0 assisted TLB flush. * mask: cpumask of the dirty vCPUs that should be flushed. -* va: linear address to flush, or NULL for global flushes. -* order: order of the linear address pointed by va. +* va: linear address to flush, or NULL for entire address space. +* flags: flags for flushing, including the order of va. */ int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order); +unsigned int flags);#else @@ -65,7 +65,7 @@ static inline int hypervisor_ap_setup(void) { return 0; }static inline void hypervisor_resume (void ) {ASSERT_UNREACHABLE (); }static inline void hypervisor_e820_fixup (struct e820map *e820) {}static inline int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, -unsigned int order) +unsigned int flags) {return -ENOSYS;
Quoted 3 lines
}
--
2.20.1
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Wei LiuQuoted 48 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 3 lines
xen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );
Shouldn't this have at least a compiler barrier here?
Paul
Quoted 100 lines
xen/arch/x86/guest/hyperv/hyperv.c +15
-0+ + hv_hcall_page_ready = true; }static int setup_hypercall_pcpu_arg (void )@@ -199,11 +203,24 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); } +static int flush_tlb (const cpumask_t *mask,const void *va, +unsigned int flags) +{ +if ( !(ms_hyperv.hints &HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED ) ) +return -EOPNOTSUPP ; + +if ( !hv_hcall_page_ready || !this_cpu (hv_input_page) ) +return -ENXIO ; + +return hyperv_flush_tlb (mask, va, flags); +} +static const struct hypervisor_ops __initdata ops = { .name ="Hyper-V" , .setup = setup, .ap_setup = ap_setup, .e820_fixup = e820_fixup, + .flush_tlb = flush_tlb, };/* xen/arch/x86/guest/hyperv/private.h +4
-0@@ -22,10 +22,14 @@#ifndef __XEN_HYPERV_PRIVIATE_H__#define __XEN_HYPERV_PRIVIATE_H__ +#include <xen/cpumask.h> #include <xen/percpu.h> DECLARE_PER_CPU(void *, hv_input_page); DECLARE_PER_CPU(void *, hv_vp_assist); DECLARE_PER_CPU(unsigned int, hv_vp_index); +int hyperv_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int flags); +#endif /* __XEN_HYPERV_PRIVIATE_H__ */ xen/arch/x86/guest/hyperv/tlb.c new file +1
-0@@ -0,0 +1,41 @@ +/************************************************************************ ******
+ *arch/x86/guest/hyperv/tlb.c
+ *
+ * Support for TLB management using hypercalls
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; If not, see <http://www.gnu.org/licenses/>.
+ *
+ * Copyright (c) 2020 Microsoft.
+ */xen/arch/x86/guest/hyperv/tlb.c new file +21
-1+ +#include <xen/cpumask.h> +#include <xen/errno.h> + +#include "private.h" + +int hyperv_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int flags) +{ +return -EOPNOTSUPP ; +} + +/* +* Local variables: +* mode: C +* c-file-style: "BSD" +* c-basic-offset: 4 +* tab-width: 4 +* indent-tabs-mode: nil +* End: +*/ --
Quoted 1 line
2.20.1
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
Quoted 50 lines · Paul
Quoted 50 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);
Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Wei.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
Quoted 51 lines · Wei Liu
Quoted 50 lines · Paul
Quoted 50 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);
Quoted 6 lines
Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of set_fixmap_x itself?
Note that map_pages_to_xen already performs atomic writes.
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
Quoted 57 lines · Roger Pau Monné
Quoted 56 lines · Wei Liu
Quoted 54 lines · Paul
Quoted 48 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 3 lines
xen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
Quoted 5 lines
Quoted 1 line
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Note that
map_pages_to_xenalready performs atomic writes.
I don't mind making things more explicit though. However unlikely, I may end up putting something in between set_fixmap_x and setting hcall_page_ready, I will need the barrier by then, I may as well put it in now.
Wei.
Quoted 2 lines
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
Quoted 60 lines · Wei Liu
Quoted 59 lines · Roger Pau Monné
Quoted 57 lines · Wei Liu
Quoted 50 lines · Paul
Quoted 50 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Quoted 8 lines
Quoted 3 lines
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Note that
map_pages_to_xenalready performs atomic writes.I don't mind making things more explicit though. However unlikely, I may end up putting something in between
set_fixmap_xand settinghcall_page_ready, I will need the barrier by then, I may as well put it in now.
IMO set_fixmap_x should have the necessary barriers (or other synchronization methods) so that on return the address is correctly mapped across all processors, and that it prevents the compiler from moving accesses past it. I would consider a bug of set_fixmap_x not having this behavior and requiring callers to do extra work in order to ensure this.
Ie: something like the snipped below should not require an extra barrier IMO:
set_fixmap_x(FIX_X_HYPERV_HCALL, mfn << PAGE_SHIFT); *((unsigned int *)fix_x_to_virt(FIX_X_HYPERV_HCALL)) = 0;
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Roger Pau MonnéQuoted 65 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 11:41
To: Wei Liu <[email protected]>
Cc: Durrant, Paul <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushQuoted 10 lines · Wei Liu
Quoted 9 lines · Paul
Quoted 9 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLBflush
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Implement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an
error
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41+++++++++++++++++++++++++++++
Quoted 28 lines
Quoted 28 lines
Quoted 28 lines
4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);
Quoted 9 lines
Quoted 6 lines
Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of
set_fixmap_xitself?
Not really, for the purpose I had in mind. The hv_hcall_page_ready global is specific to this code and we need to make sure the page is actually ready before the code says it is.
Paul
Quoted 3 lines
Note that
map_pages_to_xenalready performs atomic writes.Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 3/3] x86/hyperv: L0 assisted TLB flush
In reply to Michael KelleyQuoted 103 lines · Michael Kelley
From: Wei Liu <[email protected]> On Behalf Of Wei Liu Sent: Friday, February 14, 2020 4:35 AM
Quoted 82 lines
Implement L0 assisted TLB flush for Xen on Hyper-V. It takes advantage of several hypercalls:
*
HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST
*HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX
*HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE
*HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EXPick the most efficient hypercalls available.
Signed-off-by Wei Liu <[email protected]>---
v2:
1. Address Roger and Jan's comments re types etc.
2. Fix pointer arithmetic.
3. Misc improvement to code.--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/private.h | 9 ++ xen/arch/x86/guest/hyperv/tlb.c | 172 +++++++++++++++++++++++++++- xen/arch/x86/guest/hyperv/util.c | 74 ++++++++++++ 4 files changed, 255 insertions(+), 1 deletion(-) create mode 100644 xen/arch/x86/guest/hyperv/util.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1,2 +1,3 @@ obj-y += hyperv.o obj-y += tlb.o +obj-y += util.oxen/arch/x86/guest/hyperv/private.h +9
-0@@ -24,12 +24,21 @@#include <xen/cpumask.h> #include <xen/percpu.h> +#include <xen/types.h> DECLARE_PER_CPU(void *, hv_input_page); DECLARE_PER_CPU(void *, hv_vp_assist); DECLARE_PER_CPU(unsigned int, hv_vp_index); +static inline unsigned int hv_vp_index (unsigned int cpu) +{ +return per_cpu (hv_vp_index, cpu); +} +int hyperv_flush_tlb (const cpumask_t *mask,const void *va,unsigned int flags); +/* Returns number of banks, -ev if error */ +int cpumask_to_vpset (struct hv_vpset *vpset,const cpumask_t *mask); +#endif /* __XEN_HYPERV_PRIVIATE_H__ */ xen/arch/x86/guest/hyperv/tlb.c +10
-0@@ -19,15 +19,185 @@* Copyright (c) 2020 Microsoft. */ +#include <xen/cpu.h> #include <xen/cpumask.h> #include <xen/errno.h> +#include <asm/guest/hyperv.h> +#include <asm/guest/hyperv-hcall.h> +#include <asm/guest/hyperv-tlfs.h> +#include "private.h" +/* +* It is possible to encode up to 4096 pages using the lower 12 bits +* in an element of gva_list +*/ +#define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)xen/arch/x86/guest/hyperv/tlb.c line 46
Quoted 10 lines
xen/arch/x86/guest/hyperv/tlb.c +10
-0+ +static unsigned int fill_gva_list (uint64_t *gva_list,const void *va, +unsigned int order) +{ +unsigned long start = (unsigned long )va; +unsigned long end = start + (PAGE_SIZE << order) -1 ; +unsigned int n =0 ; + + do { + unsigned long remain = end - start;The calculated value here isn't actually the remaining bytes in the range to flush -- it's one less than the remaining bytes in the range
to flush because of the -1 in the calculation of 'end'. That difference
will mess up the comparison below againstHV_TLB_FLUSH_UNITin the case that there are exactly 4096 page remaining to be flushed. It should take the "=" case, but won't. Also, the '-1' in 'remain - 1' in the else clause becomes unneeded, and the 'start = end' assignment then propagates the error.
Quoted 7 lines
In the parallel code in Linux, if you follow the call sequence to get to
fill_gav_list(), the 'end' argument is really the address of the first byte of the first page that isn't in the flush range (i.e., one beyond the true 'end') and so is a bit misnamed.I think the calculation of 'end' should drop the -1, and perhaps 'end' should be renamed.
Thanks for the detailed review. Let me fix this.
Wei.
Quoted 3 lines
Michael
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Roger Pau MonnéQuoted 77 lines · Roger Pau Monné
Quoted 60 lines · Wei Liu
Quoted 59 lines · Roger Pau Monné
Quoted 57 lines · Wei Liu
Quoted 50 lines · Paul
Quoted 50 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Quoted 8 lines
Quoted 3 lines
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Note that
map_pages_to_xenalready performs atomic writes.I don't mind making things more explicit though. However unlikely, I may end up putting something in between
set_fixmap_xand settinghcall_page_ready, I will need the barrier by then, I may as well put it in now.IMO
set_fixmap_xshould have the necessary barriers (or other synchronization methods) so that on return the address is correctly mapped across all processors, and that it prevents the compiler from moving accesses past it. I would consider a bug ofset_fixmap_xnot having this behavior and requiring callers to do extra work in order to ensure this.
Quoted 5 lines
Ie: something like the snipped below should not require an extra barrier IMO:
set_fixmap_x(FIX_X_HYPERV_HCALL, mfn << PAGE_SHIFT); *((unsigned int *)fix_x_to_virt(FIX_X_HYPERV_HCALL)) = 0;
That's different though. Compiler can't make the connection between hcall_page_ready and the address returned by set_fixmap_x.
Wei.
Quoted 2 lines
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Durrant, PaulQuoted 70 lines · Paul
Quoted 70 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 11:41
To: Wei Liu <[email protected]>
Cc: Durrant, Paul <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushQuoted 10 lines · Wei Liu
Quoted 9 lines · Paul
Quoted 9 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLBflush
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Implement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an
error
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41+++++++++++++++++++++++++++++
Quoted 33 lines
Quoted 32 lines
Quoted 26 lines
4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 3 lines
xen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Not really, for the purpose I had in mind. The
hv_hcall_page_readyglobal is specific to this code and we need to make sure the page is actually ready before the code says it is.
But anything that modifies the page tables should already have a barrier if required in order to prevent accesses from being moved ahead of it, or else things would certainly go wrong in many other places?
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Wei LiuQuoted 78 lines · Wei Liu
Quoted 77 lines · Roger Pau Monné
Quoted 60 lines · Wei Liu
Quoted 59 lines · Roger Pau Monné
Quoted 57 lines · Wei Liu
Quoted 50 lines · Paul
Quoted 50 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Quoted 8 lines
Quoted 3 lines
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Note that
map_pages_to_xenalready performs atomic writes.I don't mind making things more explicit though. However unlikely, I may end up putting something in between
set_fixmap_xand settinghcall_page_ready, I will need the barrier by then, I may as well put it in now.IMO
set_fixmap_xshould have the necessary barriers (or other synchronization methods) so that on return the address is correctly mapped across all processors, and that it prevents the compiler from moving accesses past it. I would consider a bug ofset_fixmap_xnot having this behavior and requiring callers to do extra work in order to ensure this.
Quoted 8 lines
Quoted 5 lines
Ie: something like the snipped below should not require an extra barrier IMO:
set_fixmap_x(FIX_X_HYPERV_HCALL, mfn << PAGE_SHIFT); *((unsigned int *)fix_x_to_virt(FIX_X_HYPERV_HCALL)) = 0;That's different though. Compiler can't make the connection between
hcall_page_readyand the address returned byset_fixmap_x.
I'm not sure the compiler can make a connection between set_fixmap_x and fix_x_to_virt either (as fix_x_to_virt is a simple mathematical operation and FIX_X_HYPERV_HCALL is a constant known at build time).
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Roger Pau MonnéQuoted 94 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 12:08
To: Durrant, Paul <[email protected]>
Cc: Wei Liu <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushQuoted 6 lines · Paul
Quoted 6 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 11:41
To: Wei Liu <[email protected]>
Cc: Durrant, Paul <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>;Wei
Quoted 1 line
Quoted 1 line
Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew
Cooper
Quoted 2 lines
Quoted 2 lines
<[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLBflush
Quoted 11 lines
Quoted 11 lines · Wei Liu
Quoted 9 lines · Paul
Quoted 8 lines
Quoted 8 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau MonnéQuoted 4 lines
Quoted 4 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB
flush
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Implement a basic hook for L0 assisted TLB flush. The hook needs
to
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
check if prerequisites are met. If they are not met, it returns
an
Quoted 15 lines
Quoted 15 lines
error
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41+++++++++++++++++++++++++++++
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cQuoted 14 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
xen/arch/x86/guest/hyperv/Makefile +0
-0@@ -1 +1,2 @@Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
xen/arch/x86/guest/hyperv/hyperv.c DEFINE_PER_CPU_READ_MOSTLY(void *,
Quoted 8 lines
xen/arch/x86/guest/hyperv/Makefile +1
-0obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *,
hv_input_page);Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init
Quoted 9 lines
xen/arch/x86/guest/hyperv/hyperv.c +2
-0DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init
setup_hypercall_page(void)Quoted 12 lines
Quoted 11 lines
Quoted 8 lines
Quoted 6 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 3 lines
xen/arch/x86/guest/hyperv/hyperv.c +0
-0BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of
set_fixmap_xitself?
Quoted 8 lines
Quoted 1 line
Not really, for the purpose I had in mind. The
hv_hcall_page_readyglobal is specific to this code and we need to make sure the page is actually ready before the code says it is.
But anything that modifies the page tables should already have a barrier if required in order to prevent accesses from being moved ahead of it, or else things would certainly go wrong in many other places?
Oh. I'm not saying that we don't need a barrier there too (and more than a compiler one in that case).
Paul
Quoted 2 lines
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Roger Pau MonnéQuoted 85 lines · Roger Pau Monné
Quoted 84 lines · Wei Liu
Quoted 82 lines · Roger Pau Monné
Quoted 60 lines · Wei Liu
Quoted 59 lines · Roger Pau Monné
Quoted 57 lines · Wei Liu
Quoted 50 lines · Paul
Quoted 50 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné <[email protected]>
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushImplement a basic hook for L0 assisted TLB flush. The hook needs to check if prerequisites are met. If they are not met, it returns an error number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41 +++++++++++++++++++++++++++++ 4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cxen/arch/x86/guest/hyperv/Makefile +1
-0@@ -1 +1,2 @@ obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +2
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *, hv_input_page); DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init setup_hypercall_page(void)BUG_ON (!hypercall_msr.enable);Quoted 4 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init setup_hypercall_page(void) · line 123
Quoted 1 line
xen/arch/x86/guest/hyperv/hyperv.c +0
-0set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
OK. I will add a write barrier here.
Quoted 8 lines
Quoted 3 lines
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Note that
map_pages_to_xenalready performs atomic writes.I don't mind making things more explicit though. However unlikely, I may end up putting something in between
set_fixmap_xand settinghcall_page_ready, I will need the barrier by then, I may as well put it in now.IMO
set_fixmap_xshould have the necessary barriers (or other synchronization methods) so that on return the address is correctly mapped across all processors, and that it prevents the compiler from moving accesses past it. I would consider a bug ofset_fixmap_xnot having this behavior and requiring callers to do extra work in order to ensure this.Ie: something like the snipped below should not require an extra barrier IMO:
set_fixmap_x(FIX_X_HYPERV_HCALL, mfn << PAGE_SHIFT); *((unsigned int *)fix_x_to_virt(FIX_X_HYPERV_HCALL)) = 0;
Quoted 6 lines
Quoted 2 lines
That's different though. Compiler can't make the connection between
hcall_page_readyand the address returned byset_fixmap_x.I'm not sure the compiler can make a connection between
set_fixmap_xandfix_x_to_virteither (asfix_x_to_virtis a simple mathematical operation andFIX_X_HYPERV_HCALLis a constant known at build time).
Oh, I misread your example, sorry.
Wei.
Quoted 2 lines
Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
In reply to Durrant, PaulQuoted 98 lines · Paul
Quoted 98 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 12:08
To: Durrant, Paul <[email protected]>
Cc: Wei Liu <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushQuoted 6 lines · Paul
Quoted 6 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 11:41
To: Wei Liu <[email protected]>
Cc: Durrant, Paul <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>;Wei
Quoted 1 line
Quoted 1 line
Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew
Cooper
Quoted 2 lines
Quoted 2 lines
<[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLBflush
Quoted 11 lines
Quoted 11 lines · Wei Liu
Quoted 9 lines · Paul
Quoted 8 lines
Quoted 8 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau MonnéQuoted 4 lines
Quoted 4 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB
flush
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Implement a basic hook for L0 assisted TLB flush. The hook needs
to
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
check if prerequisites are met. If they are not met, it returns
an
Quoted 15 lines
Quoted 15 lines
error
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is ready.
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41+++++++++++++++++++++++++++++
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cQuoted 14 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
xen/arch/x86/guest/hyperv/Makefile +0
-0@@ -1 +1,2 @@Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
xen/arch/x86/guest/hyperv/hyperv.c DEFINE_PER_CPU_READ_MOSTLY(void *,
Quoted 8 lines
xen/arch/x86/guest/hyperv/Makefile +1
-0obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *,
hv_input_page);Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init
Quoted 9 lines
xen/arch/x86/guest/hyperv/hyperv.c +2
-0DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init
setup_hypercall_page(void)Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init · line 123
Quoted 3 lines
xen/arch/x86/guest/hyperv/hyperv.c +0
-0BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Not really, for the purpose I had in mind. The
hv_hcall_page_readyglobal is specific to this code and we need to make sure the page is actually ready before the code says it is.
Quoted 8 lines
Quoted 4 lines
But anything that modifies the page tables should already have a barrier if required in order to prevent accesses from being moved ahead of it, or else things would certainly go wrong in many other places?
Oh. I'm not saying that we don't need a barrier there too (and more than a compiler one in that case).
The argument Roger has is that set_fixmap_x also contains strong enough barriers to prevent hcall_page_ready to be set before page table is correctly set up.
Since you asked for it, there must be something on your mind that prompted this (maybe it is simply because you were bitten by similar things and wants to be extra sure, maybe you think it is harder to grasp the side effect of set_fixmap_x, maybe something else).
Code is written to be read by humans after all. I would rather be more explicit / redundant to make humans happy than to save a potential barrier / some typing in a code path.
Wei.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flush
Quoted 133 lines
-----Original Message-----
From: Wei Liu <[email protected]>
Sent: 17 February 2020 12:48
To: Durrant, Paul <[email protected]>
Cc: Roger Pau Monné <[email protected]>; Wei Liu <[email protected]>; Xen Development List <[email protected]>; Michael Kelley <[email protected]>; Wei Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLB flushQuoted 6 lines · Paul
Quoted 6 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 12:08
To: Durrant, Paul <[email protected]>
Cc: Wei Liu <[email protected]>; Xen Development List <xen- [email protected]>; Michael Kelley <[email protected]>;Wei
Quoted 1 line
Quoted 1 line
Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew
Cooper
Quoted 2 lines
Quoted 2 lines
<[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted TLBflush
Quoted 7 lines
Quoted 7 lines · Paul
Quoted 5 lines
Quoted 5 lines
-----Original Message-----
From: Roger Pau Monné <[email protected]>
Sent: 17 February 2020 11:41
To: Wei Liu <[email protected]>
Cc: Durrant, Paul <[email protected]>; Xen Development List<xen-
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
[email protected]>; Michael Kelley
Quoted 5 lines
Quoted 5 lines
Wei
Quoted 1 line
Quoted 1 line
Liu <[email protected]>; Jan Beulich <[email protected]>; Andrew
Cooper
Quoted 2 lines
Quoted 2 lines
<[email protected]>
Subject: Re: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assistedTLB
Quoted 10 lines
Quoted 10 lines
flush
Quoted 9 lines
Quoted 9 lines · Wei Liu
Quoted 7 lines · Paul
Quoted 6 lines
Quoted 6 lines
-----Original Message-----
From: Wei Liu <[email protected]> On Behalf Of Wei Liu
Sent: 14 February 2020 13:34
To: Xen Development List <[email protected]>
Cc: Michael Kelley <[email protected]>; Durrant, Paul <[email protected]>; Wei Liu <[email protected]>; WeiLiu
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
<[email protected]>; Jan Beulich <[email protected]>; Andrew Cooper <[email protected]>; Roger Pau Monné
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Subject: [PATCH v2 2/3] x86/hyperv: skeleton for L0 assisted
TLB
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
flush
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Implement a basic hook for L0 assisted TLB flush. The hook
needs
Quoted 2 lines
Quoted 2 lines
to
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
check if prerequisites are met. If they are not met, it
returns
Quoted 5 lines
Quoted 5 lines
an
Quoted 4 lines
Quoted 4 lines
error
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
number to fall back to native flushes.
Introduce a new variable to indicate if hypercall page is
ready.
Quoted 52 lines
Quoted 52 lines
Quoted 11 lines
Quoted 11 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Signed-off-by Wei Liu <[email protected]>--- xen/arch/x86/guest/hyperv/Makefile | 1 + xen/arch/x86/guest/hyperv/hyperv.c | 17 ++++++++++++ xen/arch/x86/guest/hyperv/private.h | 4 +++ xen/arch/x86/guest/hyperv/tlb.c | 41+++++++++++++++++++++++++++++
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
4 files changed, 63 insertions(+) create mode 100644 xen/arch/x86/guest/hyperv/tlb.cQuoted 14 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
xen/arch/x86/guest/hyperv/Makefile +0
-0@@ -1 +1,2 @@Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
xen/arch/x86/guest/hyperv/hyperv.c DEFINE_PER_CPU_READ_MOSTLY(void *,
Quoted 8 lines
xen/arch/x86/guest/hyperv/Makefile +1
-0obj-y += hyperv.o +obj-y += tlb.oxen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -33,6 +33,8 @@ DEFINE_PER_CPU_READ_MOSTLY(void *,
hv_input_page);Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init
Quoted 9 lines
xen/arch/x86/guest/hyperv/hyperv.c +2
-0DEFINE_PER_CPU_READ_MOSTLY(void *, hv_vp_assist); DEFINE_PER_CPU_READ_MOSTLY(unsigned int, hv_vp_index); +static bool __read_mostly hv_hcall_page_ready; +static uint64_t generate_guest_id (void ) {union hv_guest_os_id id = {};@@ -119,6 +121,8 @@ static void __init
setup_hypercall_page(void)Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init · line 123
Quoted 3 lines
xen/arch/x86/guest/hyperv/hyperv.c +0
-0BUG_ON (!hypercall_msr.enable);set_fixmap_x (FIX_X_HYPERV_HCALL , mfn <<PAGE_SHIFT );Shouldn't this have at least a compiler barrier here?
Quoted 6 lines
Quoted 4 lines
Quoted 1 line
OK. I will add a write barrier here.
Hm, shouldn't such barrier be part of
set_fixmap_xitself?Not really, for the purpose I had in mind. The
hv_hcall_page_readyglobal is specific to this code and we need to make sure the page is actually ready before the code says it is.
Quoted 8 lines
Quoted 4 lines
But anything that modifies the page tables should already have a barrier if required in order to prevent accesses from being moved ahead of it, or else things would certainly go wrong in many other places?
Oh. I'm not saying that we don't need a barrier there too (and more than a compiler one in that case).
The argument Roger has is that
set_fixmap_xalso contains strong enough barriers to preventhcall_page_readyto be set before page table is correctly set up.
Quoted 8 lines
Since you asked for it, there must be something on your mind that prompted this (maybe it is simply because you were bitten by similar things and wants to be extra sure, maybe you think it is harder to grasp the side effect of
set_fixmap_x, maybe something else).Code is written to be read by humans after all. I would rather be more explicit / redundant to make humans happy than to save a potential barrier / some typing in a code path.
If set_fixmap_x() is a barriering operation, and that is reasonably obvious (from comments or cursory examination of the code) then that's fine. I have indeed been bitten by this kind of thing in writing PV drivers so just wanted to ensure we didn't run into it here.
Paul
Quoted 2 lines
Wei.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel