lkml.org 
[lkml]   [2013]   [Oct]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH v3 00/15] KVM: MMU: locklessly write-protect
Date
Changelog v3:
- the changes from Gleb's review:
1) drop the patch which fixed the count of spte number in rmap since it
can not be easily fixed and it has gone after applying this patchset

- ideas from Gleb and discussion with Marcelo is also very appreciated:
2) change the way to locklessly access shadow page - use SLAB_DESTROY_BY_RCU
to protect shadow page instead of conditionally using call_rcu()
3) improve is_last_spte() that checks last spte by only using some bits on
the spte, then it is safely used when we locklessly write-protect the
shadow page table

Changelog v2:

- the changes from Gleb's review:
1) fix calculating the number of spte in the pte_list_add()
2) set iter->desc to NULL if meet a nulls desc to cleanup the code of
rmap_get_next()
3) fix hlist corruption due to accessing sp->hlish out of mmu-lock
4) use rcu functions to access the rcu protected pointer
5) spte will be missed in lockless walker if the spte is moved in a desc
(remove a spte from the rmap using only one desc). Fix it by bottom-up
walking the desc

- the changes from Paolo's review
1) make the order and memory barriers between update spte / add spte into
rmap and dirty-log more clear

- the changes from Marcelo's review:
1) let fast page fault only fix the spts on the last level (level = 1)
2) improve some changelogs and comments

- the changes from Takuya's review:
move the patch "flush tlb if the spte can be locklessly modified" forward
to make it's more easily merged

Thank all of you very much for your time and patience on this patchset!

Since we use rcu_assign_pointer() to update the points in desc even if dirty
log is disabled, i have measured the performance:
Host: Intel(R) Xeon(R) CPU X5690 @ 3.47GHz * 12 + 36G memory

- migrate-perf (benchmark the time of get-dirty-log)
before: Run 10 times, Avg time:9009483 ns.
after: Run 10 times, Avg time:4807343 ns.

- kerbench
Guest: 12 VCPUs + 8G memory
before:
EPT is enabled:
# cat 09-05-origin-ept | grep real
real 85.58
real 83.47
real 82.95

EPT is disabled:
# cat 09-05-origin-shadow | grep real
real 138.77
real 138.99
real 139.55

after:
EPT is enabled:
# cat 09-05-lockless-ept | grep real
real 83.40
real 82.81
real 83.39

EPT is disabled:
# cat 09-05-lockless-shadow | grep real
real 138.91
real 139.71
real 138.94

No performance regression!



Background
==========
Currently, when mark memslot dirty logged or get dirty page, we need to
write-protect large guest memory, it is the heavy work, especially, we need to
hold mmu-lock which is also required by vcpu to fix its page table fault and
mmu-notifier when host page is being changed. In the extreme cpu / memory used
guest, it becomes a scalability issue.

This patchset introduces a way to locklessly write-protect guest memory.

Idea
==========
There are the challenges we meet and the ideas to resolve them.

1) How to locklessly walk rmap?
The first idea we got to prevent "desc" being freed when we are walking the
rmap is using RCU. But when vcpu runs on shadow page mode or nested mmu mode,
it updates the rmap really frequently.

So we uses SLAB_DESTROY_BY_RCU to manage "desc" instead, it allows the object
to be reused more quickly. We also store a "nulls" in the last "desc"
(desc->more) which can help us to detect whether the "desc" is moved to anther
rmap then we can re-walk the rmap if that happened. I learned this idea from
nulls-list.

Another issue is, when a spte is deleted from the "desc", another spte in the
last "desc" will be moved to this position to replace the deleted one. If the
deleted one has been accessed and we do not access the replaced one, the
replaced one is missed when we do lockless walk.
To fix this case, we do not backward move the spte, instead, we forward move
the entry: when a spte is deleted, we move the entry in the first desc to that
position.

2) How to locklessly access shadow page table?
It is easy if the handler is in the vcpu context, in that case we can use
walk_shadow_page_lockless_begin() and walk_shadow_page_lockless_end() that
disable interrupt to stop shadow page be freed. But we are on the ioctl context
and the paths we are optimizing for have heavy workload, disabling interrupt is
not good for the system performance.

We add a indicator into kvm struct (kvm->arch.rcu_free_shadow_page), then use
call_rcu() to free the shadow page if that indicator is set. Set/Clear the
indicator are protected by slot-lock, so it need not be atomic and does not
hurt the performance and the scalability.

3) How to locklessly write-protect guest memory?
Currently, there are two behaviors when we write-protect guest memory, one is
clearing the Writable bit on spte and the another one is dropping spte when it
points to large page. The former is easy we only need to atomicly clear a bit
but the latter is hard since we need to remove the spte from rmap. so we unify
these two behaviors that only make the spte readonly. Making large spte
readonly instead of nonpresent is also good for reducing jitter.

And we need to pay more attention on the order of making spte writable, adding
spte into rmap and setting the corresponding bit on dirty bitmap since
kvm_vm_ioctl_get_dirty_log() write-protects the spte based on the dirty bitmap,
we should ensure the writable spte can be found in rmap before the dirty bitmap
is visible. Otherwise, we cleared the dirty bitmap and failed to write-protect
the page.

Performance result
====================
The performance result and the benchmark can be found at:
http://permalink.gmane.org/gmane.linux.kernel/1534876

Xiao Guangrong (15):
KVM: MMU: properly check last spte in fast_page_fault()
KVM: MMU: lazily drop large spte
KVM: MMU: flush tlb if the spte can be locklessly modified
KVM: MMU: flush tlb out of mmu lock when write-protect the sptes
KVM: MMU: update spte and add it into rmap before dirty log
KVM: MMU: redesign the algorithm of pte_list
KVM: MMU: introduce nulls desc
KVM: MMU: introduce pte-list lockless walker
KVM: MMU: initialize the pointers in pte_list_desc properly
KVM: MMU: allocate shadow pages from slab
KVM: MMU: locklessly access shadow page under rcu protection
KVM: MMU: check last spte with unawareness of mapping level
KVM: MMU: locklessly write-protect the page
KVM: MMU: clean up spte_write_protect
KVM: MMU: use rcu functions to access the pointer

arch/x86/include/asm/kvm_host.h | 7 +-
arch/x86/kvm/mmu.c | 586 ++++++++++++++++++++++++++++++----------
arch/x86/kvm/mmu.h | 6 +
arch/x86/kvm/mmu_audit.c | 6 +-
arch/x86/kvm/paging_tmpl.h | 6 +-
arch/x86/kvm/x86.c | 34 ++-
6 files changed, 475 insertions(+), 170 deletions(-)

--
1.8.1.4



\
 
 \ /
  Last update: 2013-10-23 15:41    [W:0.352 / U:0.228 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site