x86: improve assisted tlb flush and use it in guest mode
v5 cover x86: improve assisted tlb flush and use it in guest mode
Hello,
The following series aims to improve the TLB flush times when running nested Xen, and it's specially beneficial when running in shim mode.
Only the HAP guest TLB flush is improved, the shadow paging TLB flush is left as-is, and can be improved later if there's interest.
For a reference on the performance improvement see patch #7, as it's a huge increase which can benefit other guests using assisted TLB flushes, and also the ones using the viridian TLB flush assist (ie: Windows).
All patches have at least one Reviewed-by or Acked-by tag, the only code change vs v4 is a fix to extract the order from the flags field in patch #7.
Thanks, Roger.
Roger Pau Monne (7):
x86/hvm: allow ASID flush when v != current x86/paging: add TLB flush hooks
x86/hap: improve hypervisor assisted guest TLB flush x86/tlb: introduce a flush guests TLB flag x86/tlb: allow disabling the TLB clock xen/guest: prepare hypervisor ops to use alternative calls x86/tlb: use Xen L0 assisted TLB flush when available
xen/arch/x86/flushtlb.c | 24 ++++++--- xen/arch/x86/guest/hyperv/hyperv.c | 2 +- xen/arch/x86/guest/hypervisor.c | 51 ++++++++++-------- xen/arch/x86/guest/xen/xen.c | 8 ++- xen/arch/x86/hvm/asid.c | 6 +-- xen/arch/x86/hvm/hvm.c | 51 ++---------------- xen/arch/x86/mm/hap/hap.c | 52 +++++++++++++++++++ xen/arch/x86/mm/shadow/common.c | 71 +++++++++++++++++++++++--- xen/arch/x86/mm/shadow/hvm.c | 2 +- xen/arch/x86/mm/shadow/multi.c | 16 +++--- xen/arch/x86/smp.c | 11 ++++ xen/include/asm-x86/flushtlb.h | 19 ++++++- xen/include/asm-x86/guest/hypervisor.h | 17 ++++++ xen/include/asm-x86/hap.h | 3 ++ xen/include/asm-x86/shadow.h | 12 +++++ 15 files changed, 246 insertions(+), 99 deletions(-)
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 1/7 x86/hvm: allow ASID flush when v != current
Current implementation of hvm_asid_flush_vcpu is not safe to use unless the target vCPU is either paused or the currently running one, as it modifies the generation without any locking.
Fix this by using atomic operations when accessing the generation field, both in hvm_asid_flush_vcpu_asid and other ASID functions. This allows to safely flush the current ASID generation. Note that for the flush to take effect if the vCPU is currently running a vmexit is required.
Note the same could be achieved by introducing an extra field to hvm_vcpu_asid that signals hvm_asid_handle_vmenter the need to call hvm_asid_flush_vcpu on the given vCPU before vmentry, this however seems unnecessary as hvm_asid_flush_vcpu itself only sets two vCPU fields to 0, so there's no need to delay this to the vmentry ASID helper.
This is not a bugfix as no callers that would violate the assumptions listed in the first paragraph have been found, but a preparatory change in order to allow remote flushing of HVM vCPUs.
--- xen/arch/x86/hvm/asid.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
xen/arch/x86/hvm/asid.c +3 -3
@@ -83,7 +83,7 @@ void hvm_asid_init(int nasids)void hvm_asid_flush_vcpu_asid (struct hvm_vcpu_asid *asid) { - asid->generation =0 ; +write_atomic (&asid->generation,0 ); }void hvm_asid_flush_vcpu (struct vcpu *v)
@@ -121,7 +121,7 @@ bool_t hvm_asid_handle_vmenter(struct hvm_vcpu_asid *asid) goto disabled;/* Test if VCPU has valid ASID. */ -if ( asid->generation == data->core_asid_generation ) +if (read_atomic (&asid->generation) == data->core_asid_generation )return 0 ;/* If there are no free ASIDs, need to go to a new generation */
@@ -135,7 +135,7 @@ bool_t hvm_asid_handle_vmenter(struct hvm_vcpu_asid *asid)/* Now guaranteed to be a free ASID. */ asid->asid = data->next_asid++; - asid->generation = data->core_asid_generation; +write_atomic (&asid->generation, data->core_asid_generation);/* * When we assign ASID 1, flush all TLB entries as we are starting a new
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 1/7] x86/hvm: allow ASID flush when v != current
Quoted 3 lines · Roger Pau Monne
Current implementation of
hvm_asid_flush_vcpuis not safe to use unless the target vCPU is either paused or the currently running one, as it modifies the generation without any locking.
Indeed, but the issue you're taking care of is highly theoretical:
I don't think any sane compiler will split writes of the fields to multiple insns. It would be nice if this was made clear here.
Quoted 13 lines
Fix this by using atomic operations when accessing the generation field, both in
hvm_asid_flush_vcpu_asidand other ASID functions. This allows to safely flush the current ASID generation. Note that for the flush to take effect if the vCPU is currently running a vmexit is required.Note the same could be achieved by introducing an extra field to
hvm_vcpu_asidthat signalshvm_asid_handle_vmenterthe need to callhvm_asid_flush_vcpuon the given vCPU before vmentry, this however seems unnecessary ashvm_asid_flush_vcpuitself only sets two vCPU fields to 0, so there's no need to delay this to the vmentry ASID helper.
Quoted 6 lines
This is not a bugfix as no callers that would violate the assumptions listed in the first paragraph have been found, but a preparatory change in order to allow remote flushing of HVM vCPUs.
Signed-off-by Roger Pau Monné <[email protected]>Reviewed-by Wei Liu <[email protected]>
With a suitable clarification added
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 1/7] x86/hvm: allow ASID flush when v != current
Quoted 8 lines · Jan Beulich
Quoted 3 lines · Roger Pau Monne
Current implementation of
hvm_asid_flush_vcpuis not safe to use unless the target vCPU is either paused or the currently running one, as it modifies the generation without any locking.Indeed, but the issue you're taking care of is highly theoretical:
I don't think any sane compiler will split writes of the fields to multiple insns. It would be nice if this was made clear here.
What about adding:
Quoted 5 lines
Quoted 5 lines
Fix this by using atomic operations when accessing the generation field, both in
hvm_asid_flush_vcpu_asidand other ASID functions. This allows to safely flush the current ASID generation. Note that for the flush to take effect if the vCPU is currently running a vmexit is required.
"Most compilers will already do such writes and reads as a single instruction, so the usage of atomic operations is mostly used as a safety measure."
Here?
Quoted 11 lines
Quoted 11 lines
Note the same could be achieved by introducing an extra field to
hvm_vcpu_asidthat signalshvm_asid_handle_vmenterthe need to callhvm_asid_flush_vcpuon the given vCPU before vmentry, this however seems unnecessary ashvm_asid_flush_vcpuitself only sets two vCPU fields to 0, so there's no need to delay this to the vmentry ASID helper.This is not a bugfix as no callers that would violate the assumptions listed in the first paragraph have been found, but a preparatory change in order to allow remote flushing of HVM vCPUs.
Quoted 5 lines
Quoted 2 lines
Signed-off-by Roger Pau Monné <[email protected]>Reviewed-by Wei Liu <[email protected]>With a suitable clarification added
Acked-by Jan Beulich <[email protected]>
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 1/7] x86/hvm: allow ASID flush when v != current
Quoted 18 lines · Roger Pau Monné
Quoted 8 lines · Jan Beulich
Quoted 3 lines · Roger Pau Monne
Current implementation of
hvm_asid_flush_vcpuis not safe to use unless the target vCPU is either paused or the currently running one, as it modifies the generation without any locking.Indeed, but the issue you're taking care of is highly theoretical:
I don't think any sane compiler will split writes of the fields to multiple insns. It would be nice if this was made clear here.What about adding:
Quoted 5 lines
Quoted 5 lines
Fix this by using atomic operations when accessing the generation field, both in
hvm_asid_flush_vcpu_asidand other ASID functions. This allows to safely flush the current ASID generation. Note that for the flush to take effect if the vCPU is currently running a vmexit is required.
Quoted 5 lines
"Most compilers will already do such writes and reads as a single instruction, so the usage of atomic operations is mostly used as a safety measure."
Here?
Could you perhaps start with "Compilers will normally ..." I'm fine with the rest, it's just that "most compilers" still feels like an understatement.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 1/7] x86/hvm: allow ASID flush when v != current
Quoted 23 lines · Jan Beulich
Quoted 22 lines · Roger Pau Monné
Quoted 8 lines · Jan Beulich
Quoted 3 lines · Roger Pau Monne
Current implementation of
hvm_asid_flush_vcpuis not safe to use unless the target vCPU is either paused or the currently running one, as it modifies the generation without any locking.Indeed, but the issue you're taking care of is highly theoretical:
I don't think any sane compiler will split writes of the fields to multiple insns. It would be nice if this was made clear here.What about adding:
Quoted 5 lines
Quoted 5 lines
Fix this by using atomic operations when accessing the generation field, both in
hvm_asid_flush_vcpu_asidand other ASID functions. This allows to safely flush the current ASID generation. Note that for the flush to take effect if the vCPU is currently running a vmexit is required."Most compilers will already do such writes and reads as a single instruction, so the usage of atomic operations is mostly used as a safety measure."
Quoted 5 lines
Quoted 1 line
Here?
Could you perhaps start with "Compilers will normally ..." I'm fine with the rest, it's just that "most compilers" still feels like an understatement.
Sure, that's fine.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 2/7 x86/paging: add TLB flush hooks
In reply to Roger Pau MonneAdd shadow and hap implementation specific helpers to perform guest TLB flushes. Note that the code for both is exactly the same at the moment, and is copied from hvm_flush_vcpu_tlb. This will be changed by further patches that will add implementation specific optimizations to them.
No functional change intended.
---
Changes since v3:
- Fix stray newline removal.
- Fix return of shadow_flush_tlb dummy function.
--- xen/arch/x86/hvm/hvm.c | 51 ++---------------------------- xen/arch/x86/mm/hap/hap.c | 54 ++++++++++++++++++++++++++++++++ xen/arch/x86/mm/shadow/common.c | 55 +++++++++++++++++++++++++++++++++ xen/include/asm-x86/hap.h | 3 ++ xen/include/asm-x86/shadow.h | 12 +++++++ 5 files changed, 127 insertions(+), 48 deletions(-)
xen/arch/x86/hvm/hvm.c +3 -48
@@ -3990,55 +3990,10 @@ static void hvm_s3_resume(struct domain *d)bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt) { - static DEFINE_PER_CPU(cpumask_t, flush_cpumask); -cpumask_t *mask = &this_cpu (flush_cpumask); -struct domain *d = current->domain; -struct vcpu *v; - -/* Avoid deadlock if more than one vcpu tries this at the same time. */ -if ( !spin_trylock (&d->hypercall_deadlock_mutex) ) -return false; - -/* Pause all other vcpus. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_pause_nosync (v); - -/* Now that all VCPUs are signalled to deschedule, we wait... */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -while ( !vcpu_runnable (v) && v->is_running ) -cpu_relax (); - -/* All other vcpus are paused, safe to unlock now. */ -spin_unlock (&d->hypercall_deadlock_mutex); - -cpumask_clear (mask); - -/* Flush paging-mode soft state (e.g., va->gfn cache; PAE PDPE cache). */ - for_each_vcpu ( d, v ) - { -unsigned int cpu; - -if ( !flush_vcpu (ctxt, v) ) -continue ; - -paging_update_cr3 (v, false); +struct domain *currd = current->domain; - cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) -__cpumask_set_cpu (cpu, mask); - } - -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_unpause (v); - -return true; +return shadow_mode_enabled (currd) ?shadow_flush_tlb (flush_vcpu, ctxt) + :hap_flush_tlb (flush_vcpu, ctxt); }static bool always_flush (void *ctxt,struct vcpu *v)
xen/arch/x86/mm/hap/hap.c +54 -0
60 lines not on this page · raw message
xen/arch/x86/mm/shadow/common.c +55 -0
61 lines not on this page · raw message
xen/include/asm-x86/hap.h +3 -0
@@ -46,6 +46,9 @@ 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); +bool hap_flush_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v), +void *ctxt); +#endif /* XEN_HAP_H */ /*
xen/include/asm-x86/shadow.h +12 -0
@@ -95,6 +95,10 @@ void shadow_blow_tables_per_domain(struct domain *d);int shadow_set_allocation (struct domain *d,unsigned int pages,bool *preempted); +/* Flush the TLB of the selected vCPUs. */ +bool shadow_flush_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v), +void *ctxt); +#else /* !CONFIG_SHADOW_PAGING */ #define shadow_teardown (d, p) ASSERT(is_pv_domain(d))
@@ -106,6 +110,14 @@ int shadow_set_allocation(struct domain *d, unsigned int pages,#define shadow_set_allocation (d, pages, preempted) \ ({ ASSERT_UNREACHABLE(); -EOPNOTSUPP; }) +static inline bool shadow_flush_tlb (bool (*flush_vcpu)(void *ctxt, +struct vcpu *v), +void *ctxt) +{ +ASSERT_UNREACHABLE (); +return false; +} +static inline void sh_remove_shadows (struct domain *d,mfn_t gmfn,int fast,int all) {}
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 2/7] x86/paging: add TLB flush hooks
Quoted 11 lines · Roger Pau Monne
Add shadow and hap implementation specific helpers to perform guest TLB flushes. Note that the code for both is exactly the same at the moment, and is copied from
hvm_flush_vcpu_tlb. This will be changed by further patches that will add implementation specific optimizations to them.No functional change intended.
Signed-off-by Roger Pau Monné <[email protected]>Reviewed-by Wei Liu <[email protected]>Acked-by Tim Deegan <[email protected]>
This looks good in principle, with one possible anomaly:
Quoted 49 lines
xen/arch/x86/hvm/hvm.c +1
-41@@ -3990,55 +3990,10 @@ static void hvm_s3_resume(struct domain *d)bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt) { - static DEFINE_PER_CPU(cpumask_t, flush_cpumask); -cpumask_t *mask = &this_cpu (flush_cpumask); -struct domain *d = current->domain; -struct vcpu *v; - -/* Avoid deadlock if more than one vcpu tries this at the same time. */ -if ( !spin_trylock (&d->hypercall_deadlock_mutex) ) -return false; - -/* Pause all other vcpus. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_pause_nosync (v); - -/* Now that all VCPUs are signalled to deschedule, we wait... */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -while ( !vcpu_runnable (v) && v->is_running ) -cpu_relax (); - -/* All other vcpus are paused, safe to unlock now. */ -spin_unlock (&d->hypercall_deadlock_mutex); - -cpumask_clear (mask); - -/* Flush paging-mode soft state (e.g., va->gfn cache; PAE PDPE cache). */ - for_each_vcpu ( d, v ) - { -unsigned int cpu; - -if ( !flush_vcpu (ctxt, v) ) -continue ; - -paging_update_cr3 (v, false); +struct domain *currd = current->domain; - cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) -__cpumask_set_cpu (cpu, mask); - } - -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask);
xen/arch/x86/hvm/hvm.c static void hvm_s3_resume(struct domain *d) · line 3997
Quoted 10 lines
xen/arch/x86/hvm/hvm.c +2
-7- -/* Done. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_unpause (v); - -return true; +return shadow_mode_enabled (currd) ?shadow_flush_tlb (flush_vcpu, ctxt) + :hap_flush_tlb (flush_vcpu, ctxt); }
Following our current model I think this should be a new pointer in struct paging_mode (then truly fitting "hooks" in the title). I can see the desire to avoid the indirect call though, but I also think that if we were to go that route, we should settle on switching around others as well which are paging mode dependent. (FAOD this is nothing I ask you to do here.) Andrew, thoughts?
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 2/7] x86/paging: add TLB flush hooks
Quoted 69 lines · Jan Beulich
Quoted 11 lines · Roger Pau Monne
Add shadow and hap implementation specific helpers to perform guest TLB flushes. Note that the code for both is exactly the same at the moment, and is copied from
hvm_flush_vcpu_tlb. This will be changed by further patches that will add implementation specific optimizations to them.No functional change intended.
Signed-off-by Roger Pau Monné <[email protected]>Reviewed-by Wei Liu <[email protected]>Acked-by Tim Deegan <[email protected]>This looks good in principle, with one possible anomaly:
Quoted 54 lines
xen/arch/x86/hvm/hvm.c +1
-46@@ -3990,55 +3990,10 @@ static void hvm_s3_resume(struct domain *d)bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt) { - static DEFINE_PER_CPU(cpumask_t, flush_cpumask); -cpumask_t *mask = &this_cpu (flush_cpumask); -struct domain *d = current->domain; -struct vcpu *v; - -/* Avoid deadlock if more than one vcpu tries this at the same time. */ -if ( !spin_trylock (&d->hypercall_deadlock_mutex) ) -return false; - -/* Pause all other vcpus. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_pause_nosync (v); - -/* Now that all VCPUs are signalled to deschedule, we wait... */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -while ( !vcpu_runnable (v) && v->is_running ) -cpu_relax (); - -/* All other vcpus are paused, safe to unlock now. */ -spin_unlock (&d->hypercall_deadlock_mutex); - -cpumask_clear (mask); - -/* Flush paging-mode soft state (e.g., va->gfn cache; PAE PDPE cache). */ - for_each_vcpu ( d, v ) - { -unsigned int cpu; - -if ( !flush_vcpu (ctxt, v) ) -continue ; - -paging_update_cr3 (v, false); +struct domain *currd = current->domain; - cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) -__cpumask_set_cpu (cpu, mask); - } - -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_unpause (v);
Quoted 8 lines
xen/arch/x86/hvm/hvm.c static void hvm_s3_resume(struct domain *d) · line 3997
Quoted 5 lines
xen/arch/x86/hvm/hvm.c +2
-2- -return true; +return shadow_mode_enabled (currd) ?shadow_flush_tlb (flush_vcpu, ctxt) + :hap_flush_tlb (flush_vcpu, ctxt); }Following our current model I think this should be a new pointer in struct
paging_mode(then truly fitting "hooks" in the title).
I tried doing it that way, but there was something weird about it, the paging mode is per-vcpu, and hence I needed to do something like:
paging_get_hostmode(current)->flush(current->domain, ...)
I can try to move it to being a paging_mode hook if you prefer.
Quoted 4 lines
I can see the desire to avoid the indirect call though, but I also think that if we were to go that route, we should settle on switching around others as well which are paging mode dependent. (FAOD this is nothing I ask you to do here.) Andrew, thoughts?
I think it's already quite of a mixed bag, see track_dirty_vram for example which uses a similar model.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 2/7] x86/paging: add TLB flush hooks
Quoted 79 lines · Roger Pau Monné
Quoted 69 lines · Jan Beulich
Quoted 11 lines · Roger Pau Monne
Add shadow and hap implementation specific helpers to perform guest TLB flushes. Note that the code for both is exactly the same at the moment, and is copied from
hvm_flush_vcpu_tlb. This will be changed by further patches that will add implementation specific optimizations to them.No functional change intended.
Signed-off-by Roger Pau Monné <[email protected]>Reviewed-by Wei Liu <[email protected]>Acked-by Tim Deegan <[email protected]>This looks good in principle, with one possible anomaly:
Quoted 54 lines
xen/arch/x86/hvm/hvm.c +1
-46@@ -3990,55 +3990,10 @@ static void hvm_s3_resume(struct domain *d)bool hvm_flush_vcpu_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt) { - static DEFINE_PER_CPU(cpumask_t, flush_cpumask); -cpumask_t *mask = &this_cpu (flush_cpumask); -struct domain *d = current->domain; -struct vcpu *v; - -/* Avoid deadlock if more than one vcpu tries this at the same time. */ -if ( !spin_trylock (&d->hypercall_deadlock_mutex) ) -return false; - -/* Pause all other vcpus. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_pause_nosync (v); - -/* Now that all VCPUs are signalled to deschedule, we wait... */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -while ( !vcpu_runnable (v) && v->is_running ) -cpu_relax (); - -/* All other vcpus are paused, safe to unlock now. */ -spin_unlock (&d->hypercall_deadlock_mutex); - -cpumask_clear (mask); - -/* Flush paging-mode soft state (e.g., va->gfn cache; PAE PDPE cache). */ - for_each_vcpu ( d, v ) - { -unsigned int cpu; - -if ( !flush_vcpu (ctxt, v) ) -continue ; - -paging_update_cr3 (v, false); +struct domain *currd = current->domain; - cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) -__cpumask_set_cpu (cpu, mask); - } - -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_unpause (v);Quoted 8 lines
xen/arch/x86/hvm/hvm.c static void hvm_s3_resume(struct domain *d) · line 3997
Quoted 5 lines
xen/arch/x86/hvm/hvm.c +2
-2- -return true; +return shadow_mode_enabled (currd) ?shadow_flush_tlb (flush_vcpu, ctxt) + :hap_flush_tlb (flush_vcpu, ctxt); }Following our current model I think this should be a new pointer in struct
paging_mode(then truly fitting "hooks" in the title).
Quoted 4 lines
I tried doing it that way, but there was something weird about it, the paging mode is per-vcpu, and hence I needed to do something like:
paging_get_hostmode(current)->flush(current->domain, ...)
I don't see anything wrong with the left side of the -> (it parallels what is needed for the write_p2m_entry() hook). For the right I can't see why you'd want to have current->domain there when both functions want flush_vcpu and ctxt. Ultimately we probably want per-vCPU and per-domain hooks in separate structures (the former for hooks where the current paging mode matters, the latter for those where it doesn't matter), but of course that's nothing I'm meaning to ask you to do.
Quoted 1 line
I can try to move it to being a
paging_modehook if you prefer.
It would seem cleaner to me, but ...
Quoted 4 lines
Quoted 4 lines
I can see the desire to avoid the indirect call though, but I also think that if we were to go that route, we should settle on switching around others as well which are paging mode dependent. (FAOD this is nothing I ask you to do here.) Andrew, thoughts?
... as said I'd prefer to also know Andrew's opinion, in particular to settle where we would want this to move in the mid to long term. Whereas ...
Quoted 2 lines
I think it's already quite of a mixed bag, see
track_dirty_vramfor example which uses a similar model.
... you probably know my typical response to something like
this: Bad examples aren't to be taken as excuse to introduce further inconsistencies.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 3/7 x86/hap: improve hypervisor assisted guest TLB flush
In reply to Roger Pau MonneThe current implementation of the hypervisor assisted flush for HAP is extremely inefficient.
First of all there's no need to call paging_update_cr3, as the only relevant part of that function when doing a flush is the ASID vCPU flush, so just call that function directly.
Since hvm_asid_flush_vcpu is protected against concurrent callers by using atomic operations there's no need anymore to pause the affected vCPUs.
Finally the global TLB flush performed by flush_tlb_mask is also not necessary, since we only want to flush the guest TLB state it's enough to trigger a vmexit on the pCPUs currently holding any vCPU state, as such vmexit will already perform an ASID/VPID update, and thus clear the guest TLB.
---
Changes since v3:
- s/do_flush/handle_flush/.
- Add comment about handle_flush usage.
- Fix VPID typo in comment.
--- xen/arch/x86/mm/hap/hap.c | 52 +++++++++++++++++++-------------------- 1 file changed, 25 insertions(+), 27 deletions(-)
xen/arch/x86/mm/hap/hap.c +25 -27
@@ -669,32 +669,28 @@ static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush)hvm_update_guest_cr3 (v, noflush); } +/* +* NB: doesn't actually perform any flush, used just to clear the CPU from the +* mask and hence signal that the guest TLB flush has been done. +*/ +static void handle_flush (void *data) +{ +cpumask_t *mask = data; +unsigned int cpu =smp_processor_id (); + +ASSERT (cpumask_test_cpu (cpu, mask)); +cpumask_clear_cpu (cpu, mask); +} +bool hap_flush_tlb (bool (*flush_vcpu)(void *ctxt,struct vcpu *v),void *ctxt) { static DEFINE_PER_CPU(cpumask_t, flush_cpumask);cpumask_t *mask = &this_cpu (flush_cpumask);struct domain *d = current->domain; +unsigned int this_cpu =smp_processor_id ();struct vcpu *v; -/* Avoid deadlock if more than one vcpu tries this at the same time. */ -if ( !spin_trylock (&d->hypercall_deadlock_mutex) ) -return false; - -/* Pause all other vcpus. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_pause_nosync (v); - -/* Now that all VCPUs are signalled to deschedule, we wait... */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -while ( !vcpu_runnable (v) && v->is_running ) -cpu_relax (); - -/* All other vcpus are paused, safe to unlock now. */ -spin_unlock (&d->hypercall_deadlock_mutex); -cpumask_clear (mask);/* Flush paging-mode soft state (e.g., va->gfn cache; PAE PDPE cache). */
@@ -705,20 +701,22 @@ bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v),if ( !flush_vcpu (ctxt, v) )continue ; -paging_update_cr3 (v, false); +hvm_asid_flush_vcpu (v); cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) +if ( cpu != this_cpu &&is_vcpu_dirty_cpu (cpu) )__cpumask_set_cpu (cpu, mask); } -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ - for_each_vcpu ( d, v ) -if ( v != current &&flush_vcpu (ctxt, v) ) -vcpu_unpause (v); +/* +* Trigger a vmexit on all pCPUs with dirty vCPU state in order to force an +* ASID/VPID change and hence accomplish a guest TLB flush. Note that vCPUs +* not currently running will already be flushed when scheduled because of +* the ASID tickle done in the loop above. +*/ +on_selected_cpus (mask, handle_flush, mask,0 ); +while ( !cpumask_empty (mask) ) +cpu_relax ();return true; }
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 3/7] x86/hap: improve hypervisor assisted guest TLB flush
xen/arch/x86/mm/hap/hap.c static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush) · line 675
Quoted 10 lines · Roger Pau Monne
xen/arch/x86/mm/hap/hap.c +4
-0@@ -669,32 +669,28 @@ static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush)hvm_update_guest_cr3 (v, noflush); } +/* +* NB: doesn't actually perform any flush, used just to clear the CPU from the +* mask and hence signal that the guest TLB flush has been done. +*/
"has been done" isn't correct, as the flush may happen only later on (upon next VM entry). I think wording here needs to be as precise as possible, however the comment may turn out unnecessary altogether:
Quoted 19 lines
xen/arch/x86/mm/hap/hap.c +2
-8@@ -705,20 +701,22 @@ bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v),if ( !flush_vcpu (ctxt, v) )continue ; -paging_update_cr3 (v, false); +hvm_asid_flush_vcpu (v); cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) +if ( cpu != this_cpu &&is_vcpu_dirty_cpu (cpu) )__cpumask_set_cpu (cpu, mask); } -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ -for_each_vcpu ( d, v ) - if ( v != current && flush_vcpu(ctxt, v) )
xen/arch/x86/mm/hap/hap.c bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v), · line 719
Quoted 10 lines
xen/arch/x86/mm/hap/hap.c +9
-1-vcpu_unpause (v); +/* +* Trigger a vmexit on all pCPUs with dirty vCPU state in order to force an +* ASID/VPID change and hence accomplish a guest TLB flush. Note that vCPUs +* not currently running will already be flushed when scheduled because of +* the ASID tickle done in the loop above. +*/ +on_selected_cpus (mask, handle_flush, mask,0 ); +while ( !cpumask_empty (mask) ) +cpu_relax ();
Why do you pass 0 as last argument? If you passed 1, you wouldn't need the while() here, handle_flush() could be empty, and you'd save a perhaps large amount of CPUs to all try to modify two cache lines (the one used by on_selected_cpus() itself and the one here) instead of just one. Yes, latency from the last CPU clearing its bit to you being able to exit from here may be a little larger this way, but overall latency may shrink with the cut by half amount of atomic ops issued to the bus.
(Forcing an empty function to be called may seem odd, but sending
just some IPI [like the event check one] wouldn't be enough, as you want to be sure the other side has actually received the request.)
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 3/7] x86/hap: improve hypervisor assisted guest TLB flush
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Roger Pau Monne
xen/arch/x86/mm/hap/hap.c +0
-0@@ -669,32 +669,28 @@ static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush)hvm_update_guest_cr3 (v, noflush); }
Quoted 9 lines
xen/arch/x86/mm/hap/hap.c static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush) · line 675
Quoted 4 lines
xen/arch/x86/mm/hap/hap.c +4
-0+/* +* NB: doesn't actually perform any flush, used just to clear the CPU from the +* mask and hence signal that the guest TLB flush has been done. +*/ "has been done" isn't correct, as the flush may happen only later on (upon next VM entry). I think wording here needs to be as precise as possible, however the comment may turn out unnecessary altogether:
What about:
/*
* NB: Dummy function exclusively used as a way to trigger a vmexit,
* and thus force an ASID/VPID update on vmentry (that will flush the
* cache).
*/
Quoted 30 lines
Quoted 19 lines
xen/arch/x86/mm/hap/hap.c +2
-8@@ -705,20 +701,22 @@ bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v),if ( !flush_vcpu (ctxt, v) )continue ; -paging_update_cr3 (v, false); +hvm_asid_flush_vcpu (v); cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) +if ( cpu != this_cpu &&is_vcpu_dirty_cpu (cpu) )__cpumask_set_cpu (cpu, mask); } -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ -for_each_vcpu ( d, v ) - if ( v != current && flush_vcpu(ctxt, v) )xen/arch/x86/mm/hap/hap.c bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v), · line 719
Quoted 10 lines
xen/arch/x86/mm/hap/hap.c +9
-1-vcpu_unpause (v); +/* +* Trigger a vmexit on all pCPUs with dirty vCPU state in order to force an +* ASID/VPID change and hence accomplish a guest TLB flush. Note that vCPUs +* not currently running will already be flushed when scheduled because of +* the ASID tickle done in the loop above. +*/ +on_selected_cpus (mask, handle_flush, mask,0 ); +while ( !cpumask_empty (mask) ) +cpu_relax ();
Quoted 8 lines
Why do you pass 0 as last argument? If you passed 1, you wouldn't need the
while()here,handle_flush()could be empty, and you'd save a perhaps large amount of CPUs to all try to modify two cache lines (the one used byon_selected_cpus()itself and the one here) instead of just one. Yes, latency from the last CPU clearing its bit to you being able to exit from here may be a little larger this way, but overall latency may shrink with the cut by half amount of atomic ops issued to the bus.
In fact I think passing 0 as the last argument is fine, and handle_flush can be empty in that case anyway. We don't really care whether on_selected_cpus returns before all CPUs have executed the dummy function, as long as all of them have taken a vmexit. Using 0 already guarantees that AFAICT.
Quoted 4 lines
(Forcing an empty function to be called may seem odd, but sending
just some IPI [like the event check one] wouldn't be enough, as you want to be sure the other side has actually received the request.)
Exactly, that's the reason for the empty function.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 3/7] x86/hap: improve hypervisor assisted guest TLB flush
Quoted 18 lines · Roger Pau Monné
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Roger Pau Monne
xen/arch/x86/mm/hap/hap.c +0
-0@@ -669,32 +669,28 @@ static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush)hvm_update_guest_cr3 (v, noflush); }Quoted 9 lines
xen/arch/x86/mm/hap/hap.c static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush) · line 675
Quoted 4 lines
xen/arch/x86/mm/hap/hap.c +4
-0+/* +* NB: doesn't actually perform any flush, used just to clear the CPU from the +* mask and hence signal that the guest TLB flush has been done. +*/ "has been done" isn't correct, as the flush may happen only later on (upon next VM entry). I think wording here needs to be as precise as possible, however the comment may turn out unnecessary altogether:
Quoted 7 lines
What about:
/*
* NB: Dummy function exclusively used as a way to trigger a vmexit,
* and thus force an ASID/VPID update on vmentry (that will flush the
* cache).
*/
Once it's empty - yes, looks okay (with s/cache/TLB/).
Quoted 39 lines
Quoted 30 lines
Quoted 19 lines
xen/arch/x86/mm/hap/hap.c +2
-8@@ -705,20 +701,22 @@ bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v),if ( !flush_vcpu (ctxt, v) )continue ; -paging_update_cr3 (v, false); +hvm_asid_flush_vcpu (v); cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) +if ( cpu != this_cpu &&is_vcpu_dirty_cpu (cpu) )__cpumask_set_cpu (cpu, mask); } -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ -for_each_vcpu ( d, v ) - if ( v != current && flush_vcpu(ctxt, v) )xen/arch/x86/mm/hap/hap.c bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v), · line 719
Quoted 10 lines
xen/arch/x86/mm/hap/hap.c +9
-1-vcpu_unpause (v); +/* +* Trigger a vmexit on all pCPUs with dirty vCPU state in order to force an +* ASID/VPID change and hence accomplish a guest TLB flush. Note that vCPUs +* not currently running will already be flushed when scheduled because of +* the ASID tickle done in the loop above. +*/ +on_selected_cpus (mask, handle_flush, mask,0 ); +while ( !cpumask_empty (mask) ) +cpu_relax ();Quoted 8 lines
Why do you pass 0 as last argument? If you passed 1, you wouldn't need the
while()here,handle_flush()could be empty, and you'd save a perhaps large amount of CPUs to all try to modify two cache lines (the one used byon_selected_cpus()itself and the one here) instead of just one. Yes, latency from the last CPU clearing its bit to you being able to exit from here may be a little larger this way, but overall latency may shrink with the cut by half amount of atomic ops issued to the bus.
Quoted 5 lines
In fact I think passing 0 as the last argument is fine, and
handle_flushcan be empty in that case anyway. We don't really care whetheron_selected_cpusreturns before all CPUs have executed the dummy function, as long as all of them have taken a vmexit. Using 0 already guarantees that AFAICT.
Isn't it that the caller ants to be guaranteed that the flush has actually occurred (or at least that no further insns can be executed prior to the flush happening, i.e. at least the VM exit having occurred)?
Quoted 6 lines
Quoted 4 lines
(Forcing an empty function to be called may seem odd, but sending
just some IPI [like the event check one] wouldn't be enough, as you want to be sure the other side has actually received the request.)Exactly, that's the reason for the empty function.
But the function isn't empty.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 3/7] x86/hap: improve hypervisor assisted guest TLB flush
Quoted 21 lines · Jan Beulich
Quoted 20 lines · Roger Pau Monné
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Roger Pau Monne
xen/arch/x86/mm/hap/hap.c +0
-0@@ -669,32 +669,28 @@ static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush)hvm_update_guest_cr3 (v, noflush); }Quoted 9 lines
xen/arch/x86/mm/hap/hap.c static void hap_update_cr3(struct vcpu *v, int do_locking, bool noflush) · line 675
Quoted 4 lines
xen/arch/x86/mm/hap/hap.c +4
-0+/* +* NB: doesn't actually perform any flush, used just to clear the CPU from the +* mask and hence signal that the guest TLB flush has been done. +*/ "has been done" isn't correct, as the flush may happen only later on (upon next VM entry). I think wording here needs to be as precise as possible, however the comment may turn out unnecessary altogether:
What about:
Quoted 7 lines
Quoted 5 lines
/*
* NB: Dummy function exclusively used as a way to trigger a vmexit,
* and thus force an ASID/VPID update on vmentry (that will flush the
* cache).
*/Once it's empty - yes, looks okay (with s/cache/TLB/).
Ack.
Quoted 45 lines
Quoted 39 lines
Quoted 30 lines
Quoted 19 lines
xen/arch/x86/mm/hap/hap.c +2
-8@@ -705,20 +701,22 @@ bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v),if ( !flush_vcpu (ctxt, v) )continue ; -paging_update_cr3 (v, false); +hvm_asid_flush_vcpu (v); cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) +if ( cpu != this_cpu &&is_vcpu_dirty_cpu (cpu) )__cpumask_set_cpu (cpu, mask); } -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ -for_each_vcpu ( d, v ) - if ( v != current && flush_vcpu(ctxt, v) )xen/arch/x86/mm/hap/hap.c bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v), · line 719
Quoted 10 lines
xen/arch/x86/mm/hap/hap.c +9
-1-vcpu_unpause (v); +/* +* Trigger a vmexit on all pCPUs with dirty vCPU state in order to force an +* ASID/VPID change and hence accomplish a guest TLB flush. Note that vCPUs +* not currently running will already be flushed when scheduled because of +* the ASID tickle done in the loop above. +*/ +on_selected_cpus (mask, handle_flush, mask,0 ); +while ( !cpumask_empty (mask) ) +cpu_relax ();Quoted 8 lines
Why do you pass 0 as last argument? If you passed 1, you wouldn't need the
while()here,handle_flush()could be empty, and you'd save a perhaps large amount of CPUs to all try to modify two cache lines (the one used byon_selected_cpus()itself and the one here) instead of just one. Yes, latency from the last CPU clearing its bit to you being able to exit from here may be a little larger this way, but overall latency may shrink with the cut by half amount of atomic ops issued to the bus.Quoted 5 lines
In fact I think passing 0 as the last argument is fine, and
handle_flushcan be empty in that case anyway. We don't really care whetheron_selected_cpusreturns before all CPUs have executed the dummy function, as long as all of them have taken a vmexit. Using 0 already guarantees that AFAICT.
Quoted 4 lines
Isn't it that the caller ants to be guaranteed that the flush has actually occurred (or at least that no further insns can be executed prior to the flush happening, i.e. at least the VM exit having occurred)?
Yes, but that would happen with 0 as the last argument already, see the code in smp_call_function_interrupt:
if ( call_data.wait )
{
(*func)(info);
smp_mb();
cpumask_clear_cpu(cpu, &call_data.selected);
}
else
{
smp_mb();
cpumask_clear_cpu(cpu, &call_data.selected);
(*func)(info);
}The only difference is whether on_selected_cpus can return before all the functions have been executed on all CPUs, or whether all CPUs have taken a vmexit. The later is fine for our use-case.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
(Forcing an empty function to be called may seem odd, but sending
just some IPI [like the event check one] wouldn't be enough, as you want to be sure the other side has actually received the request.)Exactly, that's the reason for the empty function.
But the function isn't empty.
It will be now, since on_selected_cpus already does the needed synchronization as you pointed out.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 3/7] x86/hap: improve hypervisor assisted guest TLB flush
Quoted 70 lines · Roger Pau Monné
Quoted 48 lines · Jan Beulich
Quoted 41 lines · Roger Pau Monné
Quoted 31 lines · Jan Beulich
Quoted 19 lines · Roger Pau Monne
xen/arch/x86/mm/hap/hap.c +2
-8@@ -705,20 +701,22 @@ bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v),if ( !flush_vcpu (ctxt, v) )continue ; -paging_update_cr3 (v, false); +hvm_asid_flush_vcpu (v); cpu =read_atomic (&v->dirty_cpu); -if (is_vcpu_dirty_cpu (cpu) ) +if ( cpu != this_cpu &&is_vcpu_dirty_cpu (cpu) )__cpumask_set_cpu (cpu, mask); } -/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); - -/* Done. */ -for_each_vcpu ( d, v ) - if ( v != current && flush_vcpu(ctxt, v) )xen/arch/x86/mm/hap/hap.c bool hap_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v), · line 719
Quoted 10 lines
xen/arch/x86/mm/hap/hap.c +9
-1-vcpu_unpause (v); +/* +* Trigger a vmexit on all pCPUs with dirty vCPU state in order to force an +* ASID/VPID change and hence accomplish a guest TLB flush. Note that vCPUs +* not currently running will already be flushed when scheduled because of +* the ASID tickle done in the loop above. +*/ +on_selected_cpus (mask, handle_flush, mask,0 ); +while ( !cpumask_empty (mask) ) +cpu_relax ();Quoted 8 lines
Why do you pass 0 as last argument? If you passed 1, you wouldn't need the
while()here,handle_flush()could be empty, and you'd save a perhaps large amount of CPUs to all try to modify two cache lines (the one used byon_selected_cpus()itself and the one here) instead of just one. Yes, latency from the last CPU clearing its bit to you being able to exit from here may be a little larger this way, but overall latency may shrink with the cut by half amount of atomic ops issued to the bus.Quoted 5 lines
In fact I think passing 0 as the last argument is fine, and
handle_flushcan be empty in that case anyway. We don't really care whetheron_selected_cpusreturns before all CPUs have executed the dummy function, as long as all of them have taken a vmexit. Using 0 already guarantees that AFAICT.Quoted 4 lines
Isn't it that the caller ants to be guaranteed that the flush has actually occurred (or at least that no further insns can be executed prior to the flush happening, i.e. at least the VM exit having occurred)?
Yes, but that would happen with 0 as the last argument already, see the code in
smp_call_function_interrupt:if ( call_data.wait ) { (*func)(info); smp_mb(); cpumask_clear_cpu(cpu, &call_data.selected); } else { smp_mb(); cpumask_clear_cpu(cpu, &call_data.selected); (*func)(info); }
Quoted 3 lines
The only difference is whether
on_selected_cpuscan return before all the functions have been executed on all CPUs, or whether all CPUs have taken a vmexit. The later is fine for our use-case.
Oh, yes - right you are.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 4/7 x86/tlb: introduce a flush guests TLB flag
In reply to Roger Pau MonneIntroduce a specific flag to request a HVM guest TLB flush, which is an ASID/VPID tickle that forces a linear TLB flush for all HVM guests.
This was previously unconditionally done in each pre_flush call, but that's not required: HVM guests not using shadow don't require linear TLB flushes as Xen doesn't modify the guest page tables in that case (ie: when using HAP).
Modify all shadow code TLB flushes to also flush the guest TLB, in order to keep the previous behavior. I haven't looked at each specific shadow code TLB flush in order to figure out whether it actually requires a guest TLB flush or not, so there might be room for improvement in that regard.
--- xen/arch/x86/flushtlb.c | 5 +++-- xen/arch/x86/mm/shadow/common.c | 18 +++++++++--------- xen/arch/x86/mm/shadow/hvm.c | 2 +- xen/arch/x86/mm/shadow/multi.c | 16 ++++++++-------- xen/include/asm-x86/flushtlb.h | 2 ++ 5 files changed, 23 insertions(+), 20 deletions(-)
xen/arch/x86/flushtlb.c +3 -2
@@ -59,8 +59,6 @@ static u32 pre_flush(void)raise_softirq (NEW_TLBFLUSH_CLOCK_PERIOD_SOFTIRQ ); skip_clocktick: -hvm_flush_guest_tlbs (); -return t2; }
@@ -221,6 +219,9 @@ unsigned int flush_area_local(const void *va, unsigned int flags)do_tlb_flush (); } +if ( flags &FLUSH_GUESTS_TLB ) +hvm_flush_guest_tlbs (); +if ( flags &FLUSH_CACHE ) {const struct cpuinfo_x86 *c = ¤t_cpu_data;
xen/arch/x86/mm/shadow/common.c +9 -9
@@ -363,7 +363,7 @@ static int oos_remove_write_access(struct vcpu *v, mfn_t gmfn, }if ( ftlb ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );return 0 ; }
@@ -939,7 +939,7 @@ static void _shadow_prealloc(struct domain *d, unsigned int pages)/* See if that freed up enough space */ if ( d->arch.paging.shadow.free_pages >= pages ) { -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );return ; } }
@@ -993,7 +993,7 @@ static void shadow_blow_tables(struct domain *d) pagetable_get_mfn(v->arch.shadow_table[i]), 0);/* Make sure everyone sees the unshadowings */ -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB ); }void shadow_blow_tables_per_domain (struct domain *d)
@@ -1102,7 +1102,7 @@ mfn_t shadow_alloc(struct domain *d,if (unlikely (!cpumask_empty (&mask)) ) {perfc_incr (shadow_alloc_tlbflush); -flush_tlb_mask (&mask); +flush_mask (&mask,FLUSH_TLB |FLUSH_GUESTS_TLB ); }/* Now safe to clear the page for reuse */ clear_domain_page (page_to_mfn (sp));
@@ -2290,7 +2290,7 @@ void sh_remove_shadows(struct domain *d, mfn_t gmfn, int fast, int all)/* Need to flush TLBs now, so that linear maps are safe next time we * take a fault. */ -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );paging_unlock (d); }
@@ -3005,7 +3005,7 @@ static void sh_unshadow_for_p2m_change(struct domain *d, unsigned long gfn, {sh_remove_all_shadows_and_parents (d, mfn);if (sh_remove_all_mappings (d, mfn,_gfn (gfn)) ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB ); } }
@@ -3045,7 +3045,7 @@ static void sh_unshadow_for_p2m_change(struct domain *d, unsigned long gfn, } omfn =mfn_add (omfn,1 ); } -flush_tlb_mask (&flushmask); +flush_mask (&flushmask,FLUSH_TLB |FLUSH_GUESTS_TLB );if ( npte )unmap_domain_page (npte);
@@ -3332,7 +3332,7 @@ int shadow_track_dirty_vram(struct domain *d, } }if ( flush_tlb ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB ); goto out; out_sl1ma:
@@ -3402,7 +3402,7 @@ bool shadow_flush_tlb(bool (*flush_vcpu)(void *ctxt, struct vcpu *v), }/* Flush TLBs on all CPUs with dirty vcpu state. */ -flush_tlb_mask (mask); +flush_mask (mask,FLUSH_TLB |FLUSH_GUESTS_TLB );/* Done. */ for_each_vcpu ( d, v )
xen/arch/x86/mm/shadow/hvm.c +1 -1
@@ -590,7 +590,7 @@ static void validate_guest_pt_write(struct vcpu *v, mfn_t gmfn,if ( rc &SHADOW_SET_FLUSH )/* Need to flush TLBs to pick up shadow PT changes */ -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );if ( rc &SHADOW_SET_ERROR ) {
xen/arch/x86/mm/shadow/multi.c +8 -8
@@ -3066,7 +3066,7 @@ static int sh_page_fault(struct vcpu *v,perfc_incr (shadow_rm_write_flush_tlb);smp_wmb ();atomic_inc (&d->arch.paging.shadow.gtable_dirty_version); -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB ); }#if (SHADOW_OPTIMIZATIONS &SHOPT_OUT_OF_SYNC )
@@ -3575,7 +3575,7 @@ static bool sh_invlpg(struct vcpu *v, unsigned long linear)if (mfn_to_page (sl1mfn)->u.sh.type == SH_type_fl1_shadow ) { -flush_tlb_local (); +flush_local (FLUSH_TLB |FLUSH_GUESTS_TLB );return false; }
@@ -3810,7 +3810,7 @@ sh_update_linear_entries(struct vcpu *v)* table entry. But, without this change, it would fetch the wrong * value due to a stale TLB. */ -flush_tlb_local (); +flush_local (FLUSH_TLB |FLUSH_GUESTS_TLB ); } }
@@ -4011,7 +4011,7 @@ sh_update_cr3(struct vcpu *v, int do_locking, bool noflush)* (old) shadow linear maps in the writeable mapping heuristics. */ #if GUEST_PAGING_LEVELS ==2 if (sh_remove_write_access (d, gmfn,2 ,0 ) !=0 ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );sh_set_toplevel_shadow (v,0 , gmfn, SH_type_l2_shadow);#elif GUEST_PAGING_LEVELS ==3 /* PAE guests have four shadow_table entries, based on the
@@ -4035,7 +4035,7 @@ sh_update_cr3(struct vcpu *v, int do_locking, bool noflush) } }if ( flush ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );/* Now install the new shadows. */ for ( i =0 ; i <4 ; i++ ) {
@@ -4056,7 +4056,7 @@ sh_update_cr3(struct vcpu *v, int do_locking, bool noflush) }#elif GUEST_PAGING_LEVELS == 4if (sh_remove_write_access (d, gmfn,4 ,0 ) !=0 ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );sh_set_toplevel_shadow (v,0 , gmfn, SH_type_l4_shadow);if ( !shadow_mode_external (d) && !is_pv_32bit_domain (d) ) {
@@ -4502,7 +4502,7 @@ static void sh_pagetable_dying(paddr_t gpa) } }if ( flush ) -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB );/* Remember that we've seen the guest use this interface, so we * can rely on it using it in future, instead of guessing at
@@ -4539,7 +4539,7 @@ static void sh_pagetable_dying(paddr_t gpa)mfn_to_page (gmfn)->pagetable_dying = true;shadow_unhook_mappings (d, smfn,1 /* user pages only */ );/* Now flush the TLB: we removed toplevel mappings. */ -flush_tlb_mask (d->dirty_cpumask); +flush_mask (d->dirty_cpumask,FLUSH_TLB |FLUSH_GUESTS_TLB ); }/* Remember that we've seen the guest use this interface, so we
xen/include/asm-x86/flushtlb.h +2 -0
@@ -105,6 +105,8 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);#define FLUSH_VCPU_STATE 0x1000/* Flush the per-cpu root page table */ #define FLUSH_ROOT_PGTBL 0x2000 +/* Flush all HVM guests linear TLB (using ASID/VPID) */ +#define FLUSH_GUESTS_TLB 0x4000/* Flush local TLBs/caches. */ unsigned int flush_area_local (const void *va,unsigned int flags);
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 4/7] x86/tlb: introduce a flush guests TLB flag
Quoted 2 lines · Roger Pau Monne
Introduce a specific flag to request a HVM guest TLB flush, which is an ASID/VPID tickle that forces a linear TLB flush for all HVM guests.
Here and below, what do you mean by "linear"? I guess you mean TLBs holding translations from guest linear to guest physical, but I think this could do with then also saying so, even if it's more words.
Quoted 4 lines
This was previously unconditionally done in each
pre_flushcall, but that's not required: HVM guests not using shadow don't require linear TLB flushes as Xen doesn't modify the guest page tables in that case (ie: when using HAP).
This explains the correctness in one direction. What about the removal of this from the switch_cr3_cr4() path? And what about our safety assumptions from the ticking of tlbflush_clock, where we then imply that pages e.g. about to be freed can't have any translations left in any TLBs anymore?
xen/include/asm-x86/flushtlb.h void switch_cr3_cr4(unsigned long cr3, unsigned long cr4); · line 109
Quoted 8 lines
xen/include/asm-x86/flushtlb.h +2
-0@@ -105,6 +105,8 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);#define FLUSH_VCPU_STATE 0x1000/* Flush the per-cpu root page table */ #define FLUSH_ROOT_PGTBL 0x2000 +/* Flush all HVM guests linear TLB (using ASID/VPID) */ +#define FLUSH_GUESTS_TLB 0x4000
For one, the "all" is pretty misleading. A single such request doesn't do this for all vCPU-s of all HVM guests, does it? I'm also struggling with the 'S' in "GUESTS" - why is it not just "GUEST"? I admit the names of the involved functions (hvm_flush_guest_tlbs(), hvm_asid_flush_core()) are somewhat misleading, as they don't actually do any flushing, they merely arrange for what is in the TLB to no longer be able to be used, so giving this a suitable name is "historically" complicated. What if we did away with the hvm_flush_guest_tlbs() wrapper, naming the constant here then after hvm_asid_flush_core(), e.g. FLUSH_ASID_CORE?
I also think this constant might better be zero when !HVM.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 4/7] x86/tlb: introduce a flush guests TLB flag
Quoted 8 lines · Jan Beulich
Quoted 2 lines · Roger Pau Monne
Introduce a specific flag to request a HVM guest TLB flush, which is an ASID/VPID tickle that forces a linear TLB flush for all HVM guests.
Here and below, what do you mean by "linear"? I guess you mean TLBs holding translations from guest linear to guest physical, but I think this could do with then also saying so, even if it's more words.
Yes, will change.
Quoted 7 lines
Quoted 4 lines
This was previously unconditionally done in each
pre_flushcall, but that's not required: HVM guests not using shadow don't require linear TLB flushes as Xen doesn't modify the guest page tables in that case (ie: when using HAP).This explains the correctness in one direction. What about the removal of this from the
switch_cr3_cr4()path?
AFAICT that's never used by shadow code to modify cr3 or cr4, and hence doesn't require a guest linear TLB flush.
Quoted 4 lines
And what about
our safety assumptions from the ticking oftlbflush_clock, where we then imply that pages e.g. about to be freed can't have any translations left in any TLBs anymore?
I'm slightly confused. That flush only affects HVM guests linear TLB, and hence is not under Xen control unless shadow mode is used. Pages to be freed in the HAP case need to be flushed from the EPT/NPT, but whether there are references left in the guest TLB to point to that gfn really doesn't matter to Xen at all, since the guest is in full control of it's MMU and TLB in that case.
For shadow any change in the guest page-tables should already be accompanied by a guest TLB flush, or else the guest could access no longer present entries, regardless of whether the affected pages are freed or not.
Quoted 11 lines
xen/include/asm-x86/flushtlb.h void switch_cr3_cr4(unsigned long cr3, unsigned long cr4); · line 109
Quoted 8 lines
xen/include/asm-x86/flushtlb.h +2
-0@@ -105,6 +105,8 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);#define FLUSH_VCPU_STATE 0x1000/* Flush the per-cpu root page table */ #define FLUSH_ROOT_PGTBL 0x2000 +/* Flush all HVM guests linear TLB (using ASID/VPID) */ +#define FLUSH_GUESTS_TLB 0x4000For one, the "all" is pretty misleading. A single such request doesn't do this for all vCPU-s of all HVM guests, does it?
It kind of does because it tickles the pCPU ASID/VPID generation ID, so any vCPU scheduled on the selected pCPUs will get a new ASID/VPID allocated and thus a clean TLB.
Quoted 3 lines
I'm
also struggling with the 'S' in "GUESTS" - why is it not just "GUEST"?
Any guest vCPUs running on the selected pCPUs will get a new ASID/VPID ID and thus a clean TLB.
Quoted 8 lines
I admit the names of the involved functions (
hvm_flush_guest_tlbs(),hvm_asid_flush_core()) are somewhat misleading, as they don't actually do any flushing, they merely arrange for what is in the TLB to no longer be able to be used, so giving this a suitable name is "historically" complicated. What if we did away with thehvm_flush_guest_tlbs()wrapper, naming the constant here then afterhvm_asid_flush_core(), e.g.FLUSH_ASID_CORE?
I'm not opposed to renaming. The comment before the definition was also meant to clarify it's usage, and hence the explicit mention of ASID/VPID.
Quoted 1 line
I also think this constant might better be zero when !HVM.
Yes, that's a good idea.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 4/7] x86/tlb: introduce a flush guests TLB flag
Quoted 7 lines · Roger Pau Monné
Quoted 6 lines · Jan Beulich
Quoted 4 lines · Roger Pau Monne
This was previously unconditionally done in each
pre_flushcall, but that's not required: HVM guests not using shadow don't require linear TLB flushes as Xen doesn't modify the guest page tables in that case (ie: when using HAP).
Quoted 5 lines
Quoted 2 lines
This explains the correctness in one direction. What about the removal of this from the
switch_cr3_cr4()path?AFAICT that's never used by shadow code to modify cr3 or cr4, and hence doesn't require a guest linear TLB flush.
XSA-294 tells me to be very conservative here. It is not necessarily the direct use by shadow code that matters; toggle_guest_*() isn't used directly by it, either.
Quoted 11 lines
Quoted 4 lines
And what about
our safety assumptions from the ticking oftlbflush_clock, where we then imply that pages e.g. about to be freed can't have any translations left in any TLBs anymore?I'm slightly confused. That flush only affects HVM guests linear TLB, and hence is not under Xen control unless shadow mode is used. Pages to be freed in the HAP case need to be flushed from the EPT/NPT, but whether there are references left in the guest TLB to point to that gfn really doesn't matter to Xen at all, since the guest is in full control of it's MMU and TLB in that case.
Ah yes, sorry, I probably didn't get my thinking right around combined mappings and when they get invalidated.
Quoted 16 lines
Quoted 11 lines
xen/include/asm-x86/flushtlb.h void switch_cr3_cr4(unsigned long cr3, unsigned long cr4); · line 109
Quoted 8 lines
xen/include/asm-x86/flushtlb.h +2
-0@@ -105,6 +105,8 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);#define FLUSH_VCPU_STATE 0x1000/* Flush the per-cpu root page table */ #define FLUSH_ROOT_PGTBL 0x2000 +/* Flush all HVM guests linear TLB (using ASID/VPID) */ +#define FLUSH_GUESTS_TLB 0x4000For one, the "all" is pretty misleading. A single such request doesn't do this for all vCPU-s of all HVM guests, does it?
It kind of does because it tickles the pCPU ASID/VPID generation ID, so any vCPU scheduled on the selected pCPUs will get a new ASID/VPID allocated and thus a clean TLB.
Quoted 6 lines
Quoted 3 lines
I'm
also struggling with the 'S' in "GUESTS" - why is it not just "GUEST"?Any guest vCPUs running on the selected pCPUs will get a new ASID/VPID ID and thus a clean TLB.
Right, I see. Yet ...
Quoted 9 lines
Quoted 8 lines
I admit the names of the involved functions (
hvm_flush_guest_tlbs(),hvm_asid_flush_core()) are somewhat misleading, as they don't actually do any flushing, they merely arrange for what is in the TLB to no longer be able to be used, so giving this a suitable name is "historically" complicated. What if we did away with thehvm_flush_guest_tlbs()wrapper, naming the constant here then afterhvm_asid_flush_core(), e.g.FLUSH_ASID_CORE?
Quoted 3 lines
I'm not opposed to renaming. The comment before the definition was also meant to clarify it's usage, and hence the explicit mention of ASID/VPID.
... there's also one more argument for renaming: The present name doesn't convey at all that this operation is HVM-only (i.e. PV guests wouldn't have their TLBs [as far as one can call them "their"] flushed).
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 4/7] x86/tlb: introduce a flush guests TLB flag
Quoted 11 lines · Jan Beulich
Quoted 10 lines · Roger Pau Monné
Quoted 8 lines · Jan Beulich
Quoted 4 lines · Roger Pau Monne
This was previously unconditionally done in each
pre_flushcall, but that's not required: HVM guests not using shadow don't require linear TLB flushes as Xen doesn't modify the guest page tables in that case (ie: when using HAP).This explains the correctness in one direction. What about the removal of this from the
switch_cr3_cr4()path?
Quoted 6 lines
Quoted 2 lines
AFAICT that's never used by shadow code to modify cr3 or cr4, and hence doesn't require a guest linear TLB flush.
XSA-294 tells me to be very conservative here. It is not necessarily the direct use by shadow code that matters; toggle_guest_*() isn't used directly by it, either.
toggle_guest_{mode/pt} seems to be exclusively used by PV guests. I'm fine with adding extra flushes to be on the safe side, but those functions are never used against a HVM guest AFAICT. The only reason to flush a HVM guest 'internal' TLB is when using shadow, and in that case the shadow code must already take care of issuing such flushes.
Quoted 9 lines
Quoted 9 lines
Quoted 8 lines
I admit the names of the involved functions (
hvm_flush_guest_tlbs(),hvm_asid_flush_core()) are somewhat misleading, as they don't actually do any flushing, they merely arrange for what is in the TLB to no longer be able to be used, so giving this a suitable name is "historically" complicated. What if we did away with thehvm_flush_guest_tlbs()wrapper, naming the constant here then afterhvm_asid_flush_core(), e.g.FLUSH_ASID_CORE?
Quoted 8 lines
Quoted 3 lines
I'm not opposed to renaming. The comment before the definition was also meant to clarify it's usage, and hence the explicit mention of ASID/VPID.
... there's also one more argument for renaming: The present name doesn't convey at all that this operation is HVM-only (i.e. PV guests wouldn't have their TLBs [as far as one can call them "their"] flushed).
Do you think FLUSH_ASID_CORE is clear enough, or would you prefer FLUSH_HVM_ASID_CORE?
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 4/7] x86/tlb: introduce a flush guests TLB flag
Quoted 19 lines · Roger Pau Monné
Quoted 11 lines · Jan Beulich
Quoted 10 lines · Roger Pau Monné
Quoted 8 lines · Jan Beulich
Quoted 4 lines · Roger Pau Monne
This was previously unconditionally done in each
pre_flushcall, but that's not required: HVM guests not using shadow don't require linear TLB flushes as Xen doesn't modify the guest page tables in that case (ie: when using HAP).This explains the correctness in one direction. What about the removal of this from the
switch_cr3_cr4()path?Quoted 6 lines
Quoted 2 lines
AFAICT that's never used by shadow code to modify cr3 or cr4, and hence doesn't require a guest linear TLB flush.
XSA-294 tells me to be very conservative here. It is not necessarily the direct use by shadow code that matters; toggle_guest_*() isn't used directly by it, either.
Quoted 5 lines
toggle_guest_{mode/pt} seems to be exclusively used by PV guests. I'm fine with adding extra flushes to be on the safe side, but those functions are never used against a HVM guest AFAICT. The only reason to flush a HVM guest 'internal' TLB is when using shadow, and in that case the shadow code must already take care of issuing such flushes.
What I'm asking for primarily is to extend the description. If it is clear enough, it ought to also be clear enough that no flushes need inserting anywhere.
Quoted 13 lines
Quoted 13 lines
Quoted 12 lines
Quoted 8 lines
I admit the names of the involved functions (
hvm_flush_guest_tlbs(),hvm_asid_flush_core()) are somewhat misleading, as they don't actually do any flushing, they merely arrange for what is in the TLB to no longer be able to be used, so giving this a suitable name is "historically" complicated. What if we did away with thehvm_flush_guest_tlbs()wrapper, naming the constant here then afterhvm_asid_flush_core(), e.g.FLUSH_ASID_CORE?I'm not opposed to renaming. The comment before the definition was also meant to clarify it's usage, and hence the explicit mention of ASID/VPID.
Quoted 7 lines
Quoted 4 lines
... there's also one more argument for renaming: The present name doesn't convey at all that this operation is HVM-only (i.e. PV guests wouldn't have their TLBs [as far as one can call them "their"] flushed).
Do you think
FLUSH_ASID_COREis clear enough, or would you preferFLUSH_HVM_ASID_CORE?
While in principle in our code base ASID implies HVM, perhaps the latter would still be even slightly better.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 5/7 x86/tlb: allow disabling the TLB clock
In reply to Roger Pau MonneThe TLB clock is helpful when running Xen on bare metal because when doing a TLB flush each CPU is IPI'ed and can keep a timestamp of the last flush.
This is not the case however when Xen is running virtualized, and the underlying hypervisor provides mechanism to assist in performing TLB
flushes: Xen itself for example offers a HVMOP_flush_tlbs hypercall in order to perform a TLB flush without having to IPI each CPU. When using such mechanisms it's no longer possible to keep a timestamp of the flushes on each CPU, as they are performed by the underlying hypervisor.
Offer a boolean in order to signal Xen that the timestamped TLB shouldn't be used. This avoids keeping the timestamps of the flushes, and also forces NEED_FLUSH to always return true.
No functional change intended, as this change doesn't introduce any user that disables the timestamped TLB.
--- xen/arch/x86/flushtlb.c | 19 +++++++++++++------ xen/include/asm-x86/flushtlb.h | 17 ++++++++++++++++- 2 files changed, 29 insertions(+), 7 deletions(-)
xen/arch/x86/flushtlb.c +13 -6
@@ -32,6 +32,9 @@u32 tlbflush_clock =1U ;DEFINE_PER_CPU (u32, tlbflush_time); +/* Signals whether the TLB flush clock is in use. */ +bool __read_mostly tlb_clk_enabled = true; +/* * pre_flush(): Increment the virtual TLB-flush clock. Returns new clock value. *
@@ -82,12 +85,13 @@ static void post_flush(u32 t)static void do_tlb_flush (void ) {unsigned long flags, cr4; -u32 t; +u32 t =0 ;/* This non-reentrant function is sometimes called in interrupt context. */ local_irq_save (flags); - t =pre_flush (); +if ( tlb_clk_enabled ) + t =pre_flush ();if ( use_invpcid )invpcid_flush_all ();
@@ -99,7 +103,8 @@ static void do_tlb_flush(void)else write_cr3 (read_cr3 ()); -post_flush (t); +if ( tlb_clk_enabled ) +post_flush (t);local_irq_restore (flags); }
@@ -107,7 +112,7 @@ static void do_tlb_flush(void)void switch_cr3_cr4 (unsigned long cr3,unsigned long cr4) {unsigned long flags, old_cr4; -u32 t; +u32 t =0 ;/* Throughout this function we make this assumption: */ ASSERT (!(cr4 &X86_CR4_PCIDE ) || !(cr4 &X86_CR4_PGE ));
@@ -115,7 +120,8 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4)/* This non-reentrant function is sometimes called in interrupt context. */ local_irq_save (flags); - t =pre_flush (); +if ( tlb_clk_enabled ) + t =pre_flush (); old_cr4 =read_cr4 ();ASSERT (!(old_cr4 &X86_CR4_PCIDE ) || !(old_cr4 &X86_CR4_PGE ));
@@ -167,7 +173,8 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4)if ( cr4 &X86_CR4_PCIDE )invpcid_flush_all_nonglobals (); -post_flush (t); +if ( tlb_clk_enabled ) +post_flush (t);local_irq_restore (flags); }
xen/include/asm-x86/flushtlb.h +16 -1
@@ -21,10 +21,21 @@ extern u32 tlbflush_clock;/* Time at which each CPU's TLB was last flushed. */ DECLARE_PER_CPU (u32, tlbflush_time); -#define tlbflush_current_time () tlbflush_clock +/* TLB clock is in use. */ +extern bool tlb_clk_enabled; + +static inline uint32_t tlbflush_current_time (void ) +{ +/* Returning 0 from tlbflush_current_time will always force a flush. */ +return tlb_clk_enabled ? tlbflush_clock :0 ; +}static inline void page_set_tlbflush_timestamp (struct page_info *page) { +/* Avoid the write if the TLB clock is disabled. */ +if ( !tlb_clk_enabled ) +return ; +/* * Prevent storing a stale time stamp, which could happen if an update * to tlbflush_clock plus a subsequent flush IPI happen between the
@@ -67,6 +78,10 @@ static inline void tlbflush_filter(cpumask_t *mask, uint32_t page_timestamp) {unsigned int cpu; +/* Short-circuit: there's no need to iterate if the clock is disabled. */ +if ( !tlb_clk_enabled ) +return ; + for_each_cpu ( cpu, mask )if ( !NEED_FLUSH (per_cpu (tlbflush_time, cpu), page_timestamp) )__cpumask_clear_cpu (cpu, mask);
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 5/7] x86/tlb: allow disabling the TLB clock
Quoted 16 lines · Roger Pau Monne
The TLB clock is helpful when running Xen on bare metal because when doing a TLB flush each CPU is IPI'ed and can keep a timestamp of the last flush.
This is not the case however when Xen is running virtualized, and the underlying hypervisor provides mechanism to assist in performing TLB
flushes: Xen itself for example offers aHVMOP_flush_tlbshypercall in order to perform a TLB flush without having to IPI each CPU. When using such mechanisms it's no longer possible to keep a timestamp of the flushes on each CPU, as they are performed by the underlying hypervisor.Offer a boolean in order to signal Xen that the timestamped TLB shouldn't be used. This avoids keeping the timestamps of the flushes, and also forces
NEED_FLUSHto always return true.
Quoted 5 lines
No functional change intended, as this change doesn't introduce any user that disables the timestamped TLB.
Signed-off-by Roger Pau Monné <[email protected]>Reviewed-by Wei Liu <[email protected]>
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 6/7 xen/guest: prepare hypervisor ops to use alternative calls
In reply to Roger Pau MonneAdapt the hypervisor ops framework so it can be used with the alternative calls framework. So far no hooks are modified to make use of the alternatives patching, as they are not in any hot path.
No functional change intended.
---
Changes since v3:
- New in this version.
--- xen/arch/x86/guest/hyperv/hyperv.c | 2 +- xen/arch/x86/guest/hypervisor.c | 41 +++++++++++++++--------------- xen/arch/x86/guest/xen/xen.c | 2 +- 3 files changed, 23 insertions(+), 22 deletions(-)
xen/arch/x86/guest/hyperv/hyperv.c +1 -1
@@ -199,7 +199,7 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); } -static const struct hypervisor_ops ops = { +static const struct hypervisor_ops __initdata ops = { .name ="Hyper-V" , .setup = setup, .ap_setup = ap_setup,
xen/arch/x86/guest/hypervisor.c +21 -20
73 lines not on this page · raw message
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 const struct hypervisor_ops ops = { +static const struct hypervisor_ops __initdata ops = { .name ="Xen" , .setup = setup, .ap_setup = ap_setup,
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 6/7] xen/guest: prepare hypervisor ops to use alternative calls
xen/arch/x86/guest/hyperv/hyperv.c static void __init e820_fixup(struct e820map *e820) · line 202
Quoted 8 lines · Roger Pau Monne
xen/arch/x86/guest/hyperv/hyperv.c +1
-1@@ -199,7 +199,7 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); } -static const struct hypervisor_ops ops = { +static const struct hypervisor_ops __initdata ops = {
This needs to be __initconstrel in order to avoid triggering (possibly only in the future) section mismatch warnings with at least some gcc versions. With this and the other instance adjusted
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 6/7] xen/guest: prepare hypervisor ops to use alternative calls
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Roger Pau Monne
xen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -199,7 +199,7 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); }
Quoted 7 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init e820_fixup(struct e820map *e820) · line 202
Quoted 2 lines
xen/arch/x86/guest/hyperv/hyperv.c +1
-1-static const struct hypervisor_ops ops = { +static const struct hypervisor_ops __initdata ops = {This needs to be
__initconstrelin order to avoid triggering (possibly only in the future) section mismatch warnings with at least some gcc versions. With this and the other instance adjusted
I can do that when posting a new version, unless you want to pick this earlier and adjust on commit.
Quoted 1 line
Reviewed-by Jan Beulich <[email protected]>
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 6/7] xen/guest: prepare hypervisor ops to use alternative calls
Quoted 16 lines · Roger Pau Monné
Quoted 7 lines · Jan Beulich
Quoted 6 lines · Roger Pau Monne
xen/arch/x86/guest/hyperv/hyperv.c +0
-0@@ -199,7 +199,7 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); }Quoted 7 lines
xen/arch/x86/guest/hyperv/hyperv.c static void __init e820_fixup(struct e820map *e820) · line 202
Quoted 2 lines
xen/arch/x86/guest/hyperv/hyperv.c +1
-1-static const struct hypervisor_ops ops = { +static const struct hypervisor_ops __initdata ops = {This needs to be
__initconstrelin order to avoid triggering (possibly only in the future) section mismatch warnings with at least some gcc versions. With this and the other instance adjusted
Quoted 2 lines
I can do that when posting a new version, unless you want to pick this earlier and adjust on commit.
Is this to mean that this 2nd to last patch in the series is fully independent of the earlier five (also contextually)? Then of course I'd be fine to make the adjustments and commit. Please confirm if so.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 6/7] xen/guest: prepare hypervisor ops to use alternative calls
Quoted 17 lines · Jan Beulich
Quoted 16 lines · Roger Pau Monné
Quoted 14 lines · Jan Beulich
xen/arch/x86/guest/hyperv/hyperv.c static void __init e820_fixup(struct e820map *e820) · line 202
Quoted 8 lines · Roger Pau Monne
xen/arch/x86/guest/hyperv/hyperv.c +1
-1@@ -199,7 +199,7 @@ static void __init e820_fixup(struct e820map *e820)panic ("Unable to reserve Hyper-V hypercall range\n" ); } -static const struct hypervisor_ops ops = { +static const struct hypervisor_ops __initdata ops = {This needs to be
__initconstrelin order to avoid triggering (possibly only in the future) section mismatch warnings with at least some gcc versions. With this and the other instance adjusted
Quoted 5 lines
Quoted 2 lines
I can do that when posting a new version, unless you want to pick this earlier and adjust on commit.
Is this to mean that this 2nd to last patch in the series is fully independent of the earlier five (also contextually)?
Right, patches 5 to 7 should be completely independent, as they only modify code related to Xen running as a guest.
Quoted 2 lines
Then of course I'd be fine to make the adjustments and commit. Please confirm if so.
Yes please, do the adjustments on commit if you don't mind.
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
PATCH v5 7/7 x86/tlb: use Xen L0 assisted TLB flush when available
In reply to Roger Pau MonneUse Xen's L0 HVMOP_flush_tlbs hypercall in order to perform flushes. This greatly increases the performance of TLB flushes when running with a high amount of vCPUs as a Xen guest, and is specially important when running in shim mode.
The following figures are from a PV guest running make -j32 xen in shim mode with 32 vCPUs and HAP.
Using x2APIC and ALLBUT shorthand:
real 4m35.973s
user 4m35.110s
sys 36m24.117s
Using L0 assisted flush:
real 1m2.596s user 4m34.818s sys 5m16.374s
The implementation adds a new hook to hypervisor_ops so other enlightenments can also implement such assisted flush just by filling the hook. Note that the Xen implementation completely ignores the dirty CPU mask and the linear address passed in, and always performs a global TLB flush on all vCPUs.
---
Changes since v4:
- Adjust order calculation.
Changes since v3:
- Use an alternative call for the flush hook.
Changes since v1:
- Add a L0 assisted hook to hypervisor ops.
--- xen/arch/x86/guest/hypervisor.c | 10 ++++++++++ xen/arch/x86/guest/xen/xen.c | 6 ++++++ xen/arch/x86/smp.c | 11 +++++++++++ xen/include/asm-x86/guest/hypervisor.h | 17 +++++++++++++++++ 4 files changed, 44 insertions(+)
xen/arch/x86/guest/hypervisor.c +10 -0
@@ -18,6 +18,7 @@* * Copyright (c) 2019 Microsoft. */ +#include <xen/cpumask.h> #include <xen/init.h> #include <xen/types.h>
@@ -73,6 +74,15 @@ void __init hypervisor_e820_fixup(struct e820map *e820) ops.e820_fixup (e820); } +int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int order) +{ +if ( ops.flush_tlb ) +return alternative_call (ops.flush_tlb, mask, va, order); + +return -ENOSYS ; +} +/* * Local variables: * mode: C
xen/arch/x86/guest/xen/xen.c +6 -0
@@ -324,12 +324,18 @@ 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) +{ +return xen_hypercall_hvm_op (HVMOP_flush_tlbs,NULL ); +} +static const struct hypervisor_ops __initdata ops = { .name ="Xen" , .setup = setup, .ap_setup = ap_setup, .resume = resume, .e820_fixup = e820_fixup, + .flush_tlb = flush_tlb, };const struct hypervisor_ops *__init xg_probe (void )
xen/arch/x86/smp.c +11 -0
@@ -15,6 +15,7 @@#include <xen/perfc.h> #include <xen/spinlock.h> #include <asm/current.h> +#include <asm/guest.h> #include <asm/smp.h> #include <asm/mc146818rtc.h> #include <asm/flushtlb.h>
@@ -256,6 +257,16 @@ void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags)if ( (flags & ~FLUSH_ORDER_MASK ) && !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 -1 ) &FLUSH_ORDER_MASK ) ) + { +if ( tlb_clk_enabled ) + tlb_clk_enabled = false; +return ; + } +spin_lock (&flush_lock);cpumask_and (&flush_cpumask, mask, &cpu_online_map);cpumask_clear_cpu (cpu, &flush_cpumask);
xen/include/asm-x86/guest/hypervisor.h +17 -0
@@ -19,6 +19,8 @@#ifndef __X86_HYPERVISOR_H__#define __X86_HYPERVISOR_H__ +#include <xen/cpumask.h> +#include <asm/e820.h> struct hypervisor_ops {
@@ -32,6 +34,8 @@ struct hypervisor_ops {void (*resume)(void );/* 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); };#ifdef CONFIG_GUEST
@@ -41,6 +45,14 @@ void hypervisor_setup(void);int hypervisor_ap_setup (void );void hypervisor_resume (void );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. +*/ +int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int order);#else
@@ -52,6 +64,11 @@ static inline void hypervisor_setup(void) { ASSERT_UNREACHABLE(); }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) +{ +return -ENOSYS ; +}#endif /* CONFIG_GUEST */
Signature · 7 lines
2.25.0 _______________________________________________ Xen-devel mailing list [email protected] https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 7/7] x86/tlb: use Xen L0 assisted TLB flush when available
Quoted 18 lines · Roger Pau Monne
Use Xen's L0
HVMOP_flush_tlbshypercall in order to perform flushes. This greatly increases the performance of TLB flushes when running with a high amount of vCPUs as a Xen guest, and is specially important when running in shim mode.The following figures are from a PV guest running
make -j32 xenin shim mode with 32 vCPUs and HAP.Using x2APIC and ALLBUT shorthand:
real 4m35.973s
user 4m35.110s
sys 36m24.117sUsing L0 assisted flush:
real 1m2.596s user 4m34.818s sys 5m16.374s
Quoted 5 lines
The implementation adds a new hook to
hypervisor_opsso other enlightenments can also implement such assisted flush just by filling the hook. Note that the Xen implementation completely ignores the dirty CPU mask and the linear address passed in, and always performs a global TLB flush on all vCPUs.
This isn't because of an implementation choice of yours, but because of how HVMOP_flush_tlbs works. I think the statement should somehow express this. I also think it wants clarifying that using the hypercall is indeed faster even in the case of single-page, single- CPU flush (which I suspect may not be the case especially as vCPU count grows). The stats above prove a positive overall effect, but they don't say whether the effect could be even bigger by being at least a little selective.
xen/arch/x86/guest/hypervisor.c void __init hypervisor_e820_fixup(struct e820map *e820) · line 84
Quoted 12 lines
xen/arch/x86/guest/hypervisor.c +8
-0@@ -73,6 +74,15 @@ void __init hypervisor_e820_fixup(struct e820map *e820) ops.e820_fixup (e820); } +int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int order) +{ +if ( ops.flush_tlb ) +return alternative_call (ops.flush_tlb, mask, va, order); + +return -ENOSYS ; +}
Please no new -ENOSYS anywhere (except in new ports' top level hypercall handlers).
xen/arch/x86/smp.c void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags) · line 266
Quoted 11 lines
xen/arch/x86/smp.c +7
-0@@ -256,6 +257,16 @@ void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags)if ( (flags & ~FLUSH_ORDER_MASK ) && !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 -1 ) &FLUSH_ORDER_MASK ) ) + { +if ( tlb_clk_enabled ) + tlb_clk_enabled = false;
Why does this need doing here? Couldn't Xen guest setup code clear the flag?
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 7/7] x86/tlb: use Xen L0 assisted TLB flush when available
Quoted 25 lines · Jan Beulich
Quoted 18 lines · Roger Pau Monne
Use Xen's L0
HVMOP_flush_tlbshypercall in order to perform flushes. This greatly increases the performance of TLB flushes when running with a high amount of vCPUs as a Xen guest, and is specially important when running in shim mode.The following figures are from a PV guest running
make -j32 xenin shim mode with 32 vCPUs and HAP.Using x2APIC and ALLBUT shorthand:
real 4m35.973s
user 4m35.110s
sys 36m24.117sUsing L0 assisted flush:
real 1m2.596s user 4m34.818s sys 5m16.374sQuoted 5 lines
The implementation adds a new hook to
hypervisor_opsso other enlightenments can also implement such assisted flush just by filling the hook. Note that the Xen implementation completely ignores the dirty CPU mask and the linear address passed in, and always performs a global TLB flush on all vCPUs.
Quoted 5 lines
This isn't because of an implementation choice of yours, but because of how
HVMOP_flush_tlbsworks. I think the statement should somehow express this. I also think it wants clarifying that using the hypercall is indeed faster even in the case of single-page, single- CPU flush
Are you sure about this? I think taking a vmexit is going to be more costly than executing a local invlpg?
Quoted 4 lines
(which I suspect may not be the case especially as vCPU count grows). The stats above prove a positive overall effect, but they don't say whether the effect could be even bigger by being at least a little selective.
I assume that being able to provide a bitmap with the vCPUs on whether the TLB flush should be performed would give us some more performance, but I haven't looked into it yet.
Quoted 4 lines
Quoted 4 lines
xen/arch/x86/guest/hypervisor.c +0
-0@@ -73,6 +74,15 @@ void __init hypervisor_e820_fixup(struct e820map *e820) ops.e820_fixup (e820); }
Quoted 11 lines
xen/arch/x86/guest/hypervisor.c void __init hypervisor_e820_fixup(struct e820map *e820) · line 84
Quoted 8 lines
xen/arch/x86/guest/hypervisor.c +8
-0+int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int order) +{ +if ( ops.flush_tlb ) +return alternative_call (ops.flush_tlb, mask, va, order); + +return -ENOSYS ; +}Please no new
-ENOSYSanywhere (except in new ports' top level hypercall handlers).
Ack. Is EOPNOTSUPP OK?
Quoted 6 lines
Quoted 6 lines
xen/arch/x86/smp.c +2
-0@@ -256,6 +257,16 @@ void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags)if ( (flags & ~FLUSH_ORDER_MASK ) && !cpumask_subset (mask,cpumask_of (cpu)) ) { + if ( cpu_has_hypervisor && + !(flags & ~(FLUSH_TLB | FLUSH_TLB_GLOBAL | FLUSH_VA_VALID |
Quoted 8 lines
xen/arch/x86/smp.c void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags) · line 266
Quoted 5 lines
xen/arch/x86/smp.c +5
-0+ FLUSH_ORDER_MASK)) && + !hypervisor_flush_tlb(mask, va, (flags - 1) & FLUSH_ORDER_MASK) ) + { +if ( tlb_clk_enabled ) + tlb_clk_enabled = false;Why does this need doing here? Couldn't Xen guest setup code clear the flag?
Because it's possible that the hypercalls fails, and hence the tlb clock should be kept enabled. There's no reason to disable it until Xen knows the assisted flush is indeed available.
I don't mind moving it to Xen guest setup code, but I'm not sure I see why it would be any better than doing it here. The only reason I guess is to avoid checking tlb_clk_enabled on every successful assisted flush?
Thanks, Roger.
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel
Re: [Xen-devel] [PATCH v5 7/7] x86/tlb: use Xen L0 assisted TLB flush when available
Quoted 26 lines · Roger Pau Monné
Quoted 25 lines · Jan Beulich
Quoted 23 lines · Roger Pau Monne
Use Xen's L0
HVMOP_flush_tlbshypercall in order to perform flushes. This greatly increases the performance of TLB flushes when running with a high amount of vCPUs as a Xen guest, and is specially important when running in shim mode.The following figures are from a PV guest running
make -j32 xenin shim mode with 32 vCPUs and HAP.Using x2APIC and ALLBUT shorthand:
real 4m35.973s
user 4m35.110s
sys 36m24.117sUsing L0 assisted flush:
real 1m2.596s user 4m34.818s sys 5m16.374sThe implementation adds a new hook to
hypervisor_opsso other enlightenments can also implement such assisted flush just by filling the hook. Note that the Xen implementation completely ignores the dirty CPU mask and the linear address passed in, and always performs a global TLB flush on all vCPUs.
Quoted 8 lines
Quoted 5 lines
This isn't because of an implementation choice of yours, but because of how
HVMOP_flush_tlbsworks. I think the statement should somehow express this. I also think it wants clarifying that using the hypercall is indeed faster even in the case of single-page, single- CPU flushAre you sure about this? I think taking a vmexit is going to be more costly than executing a local invlpg?
The answer to this was already ...
Quoted 2 lines
Quoted 2 lines
(which I suspect may not be the case especially as vCPU count grows).
... here (or at least it was meant to address questions back like this one).
Quoted 24 lines
Quoted 3 lines
The stats above prove a positive overall effect, but they don't say whether the effect could be even bigger by being at least a little selective.
I assume that being able to provide a bitmap with the vCPUs on whether the TLB flush should be performed would give us some more performance, but I haven't looked into it yet.
Quoted 4 lines
Quoted 4 lines
xen/arch/x86/guest/hypervisor.c +0
-0@@ -73,6 +74,15 @@ void __init hypervisor_e820_fixup(struct e820map *e820) ops.e820_fixup (e820); }Quoted 11 lines
xen/arch/x86/guest/hypervisor.c void __init hypervisor_e820_fixup(struct e820map *e820) · line 84
Quoted 8 lines
xen/arch/x86/guest/hypervisor.c +8
-0+int hypervisor_flush_tlb (const cpumask_t *mask,const void *va, +unsigned int order) +{ +if ( ops.flush_tlb ) +return alternative_call (ops.flush_tlb, mask, va, order); + +return -ENOSYS ; +}Please no new
-ENOSYSanywhere (except in new ports' top level hypercall handlers).
Quoted 1 line
Ack. Is EOPNOTSUPP OK?
Sure.
Quoted 15 lines
Quoted 6 lines
Quoted 6 lines
xen/arch/x86/smp.c +2
-0@@ -256,6 +257,16 @@ void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags)if ( (flags & ~FLUSH_ORDER_MASK ) && !cpumask_subset (mask,cpumask_of (cpu)) ) { + if ( cpu_has_hypervisor && + !(flags & ~(FLUSH_TLB | FLUSH_TLB_GLOBAL | FLUSH_VA_VALID |Quoted 8 lines
xen/arch/x86/smp.c void flush_area_mask(const cpumask_t *mask, const void *va, unsigned int flags) · line 266
Quoted 5 lines
xen/arch/x86/smp.c +5
-0+ FLUSH_ORDER_MASK)) && + !hypervisor_flush_tlb(mask, va, (flags - 1) & FLUSH_ORDER_MASK) ) + { +if ( tlb_clk_enabled ) + tlb_clk_enabled = false;Why does this need doing here? Couldn't Xen guest setup code clear the flag?
Quoted 8 lines
Because it's possible that the hypercalls fails, and hence the tlb clock should be kept enabled. There's no reason to disable it until Xen knows the assisted flush is indeed available.
I don't mind moving it to Xen guest setup code, but I'm not sure I see why it would be any better than doing it here. The only reason I guess is to avoid checking
tlb_clk_enabledon every successful assisted flush?
Well, iirc there had already been a question on why the if() here is needed. I second the reason, but the whole construct looks misplaced, i.e. is prone to cause further questions down the road. I think if it was to stay here, a comment would be needed to address any such possible questions. But I still think it should live in the init code. This isn't a feature that simply lacks a CPUID bit (or alike) and hence needs probing (unless of course we expect people to want to put a modern Xen [shim] on top of a pre-3.<whatever> Xen; and even then you could probe the underlying hypercall once at boot time).
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel