When we invalidate a THP page, we call the handler with the same
rmap_pde argument 512 times in the following loop:
for each guest page in the range
for each level
unmap using rmap
This patch avoids these extra handler calls by changing the loop order
like this:
for each level
for each rmap in the range
unmap using rmap
With the preceding patches in the patch series, this made THP page
invalidation more than 5 times faster on our x86 host: the host became
more responsive during swapping the guest's memory as a result.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This restricts the tracing to page aging and makes it possible to
optimize kvm_handle_hva_range() further in the following patch.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This is needed to push trace_kvm_age_page() into kvm_age_rmapp() in the
following patch.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This makes it possible to loop over rmap_pde arrays in the same way as
we do over rmap so that we can optimize kvm_handle_hva_range() easily in
the following patch.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
When we tested KVM under memory pressure, with THP enabled on the host,
we noticed that MMU notifier took a long time to invalidate huge pages.
Since the invalidation was done with mmu_lock held, it not only wasted
the CPU but also made the host harder to respond.
This patch mitigates this by using kvm_handle_hva_range().
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Cc: Alexander Graf <agraf@suse.de>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
When guest's memory is backed by THP pages, MMU notifier needs to call
kvm_unmap_hva(), which in turn leads to kvm_handle_hva(), in a loop to
invalidate a range of pages which constitute one huge page:
for each page
for each memslot
if page is in memslot
unmap using rmap
This means although every page in that range is expected to be found in
the same memslot, we are forced to check unrelated memslots many times.
If the guest has more memslots, the situation will become worse.
Furthermore, if the range does not include any pages in the guest's
memory, the loop over the pages will just consume extra time.
This patch, together with the following patches, solves this problem by
introducing kvm_handle_hva_range() which makes the loop look like this:
for each memslot
for each page in memslot
unmap using rmap
In this new processing, the actual work is converted to a loop over rmap
which is much more cache friendly than before.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Cc: Alexander Graf <agraf@suse.de>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This restricts hva handling in mmu code and makes it easier to extend
kvm_handle_hva() so that it can treat a range of addresses later in this
patch series.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Cc: Alexander Graf <agraf@suse.de>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
We can treat every level uniformly.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
To see what happen on this path and help us to optimize it
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
If the the present bit of page fault error code is set, it indicates
the shadow page is populated on all levels, it means what we do is
only modify the access bit which can be done out of mmu-lock
Currently, in order to simplify the code, we only fix the page fault
caused by write-protect on the fast path
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This bit indicates whether the spte can be writable on MMU, that means
the corresponding gpte is writable and the corresponding gfn is not
protected by shadow page protection
In the later path, SPTE_MMU_WRITEABLE will indicates whether the spte
can be locklessly updated
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
mmu_spte_update() is the common function, we can easily audit the path
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Use __drop_large_spte to cleanup this function and comment spte_write_protect
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Introduce a common function to abstract spte write-protect to
cleanup the code
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The reture value of __rmap_write_protect is either 1 or 0, use
true/false instead of these
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently the MMU's ->new_cr3() callback does nothing when guest paging
is disabled or when two-dimentional paging (e.g. EPT on Intel) is active.
This means that an emulated write to cr3 can be lost; kvm_set_cr3() will
write vcpu-arch.cr3, but the GUEST_CR3 field in the VMCS will retain its
old value and this is what the guest sees.
This bug did not have any effect until now because:
- with unrestricted guest, or with svm, we never emulate a mov cr3 instruction
- without unrestricted guest, and with paging enabled, we also never emulate a
mov cr3 instruction
- without unrestricted guest, but with paging disabled, the guest's cr3 is
ignored until the guest enables paging; at this point the value from arch.cr3
is loaded correctly my the mov cr0 instruction which turns on paging
However, the patchset that enables big real mode causes us to emulate mov cr3
instructions in protected mode sometimes (when guest state is not virtualizable
by vmx); this mov cr3 is effectively ignored and will crash the guest.
The fix is to make nonpaging_new_cr3() call mmu_free_roots() to force a cr3
reload. This is awkward because now all the new_cr3 callbacks to the same
thing, and because mmu_free_roots() is somewhat of an overkill; but fixing
that is more complicated and will be done after this minimal fix.
Observed in the Window XP 32-bit installer while bringing up secondary vcpus.
Signed-off-by: Avi Kivity <avi@redhat.com>
EPT Dirty bit use bit 9 as Intel SDM definition, to avoid conflict, change
PT_FIRST_AVAIL_BITS_SHIFT to 10.
Signed-off-by: Xudong Hao <xudong.hao@intel.com>
Signed-off-by: Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Size is not needed to return one from pre-allocated objects.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
I see this in 3.5-rc1:
arch/x86/kvm/mmu.c: In function ‘kvm_test_age_rmapp’:
arch/x86/kvm/mmu.c:1271: warning: ‘iter.desc’ may be used uninitialized in this function
The line in question was introduced by commit
1e3f42f03c
static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
unsigned long data)
{
- u64 *spte;
+ u64 *sptep;
+ struct rmap_iterator iter; <- line 1271
int young = 0;
/*
The reason I think is that the compiler assumes that
the rmap value could be 0, so
static u64 *rmap_get_first(unsigned long rmap, struct rmap_iterator
*iter)
{
if (!rmap)
return NULL;
if (!(rmap & 1)) {
iter->desc = NULL;
return (u64 *)rmap;
}
iter->desc = (struct pte_list_desc *)(rmap & ~1ul);
iter->pos = 0;
return iter->desc->sptes[iter->pos];
}
will not initialize iter.desc, but the compiler isn't
smart enough to see that
for (sptep = rmap_get_first(*rmapp, &iter); sptep;
sptep = rmap_get_next(&iter)) {
will immediately exit in this case.
I checked by adding
if (!*rmapp)
goto out;
on top which is clearly equivalent but disables the warning.
This patch uses uninitialized_var to disable the warning without
increasing code size.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
mmu_shrink() needlessly iterates over all VMs even though it will not
attempt to free mmu pages from more than one on them. Fix that and also
check used mmu pages count outside of VM lock to skip inactive VMs faster.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The huge page size is 4M on non-PAE host, but 2M page size is used in
transparent_hugepage_adjust(), so the page we get after adjust the
mapping level is not the head page, the BUG_ON() will be triggered
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Using RCU for lockless shadow walking can increase the amount of memory
in use by the system, since RCU grace periods are unpredictable. We also
have an unconditional write to a shared variable (reader_counter), which
isn't good for scaling.
Replace that with a scheme similar to x86's get_user_pages_fast(): disable
interrupts during lockless shadow walk to force the freer
(kvm_mmu_commit_zap_page()) to wait for the TLB flush IPI to find the
processor with interrupts enabled.
We also add a new vcpu->mode, READING_SHADOW_PAGE_TABLES, to prevent
kvm_flush_remote_tlbs() from avoiding the IPI.
Signed-off-by: Avi Kivity <avi@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Its much cleaner to use PT_PAGE_TABLE_LEVEL than its numeric value.
Signed-off-by: Davidlohr Bueso <dave@gnu.org>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Iteration using rmap_next(), the actual body is pte_list_next(), is
inefficient: every time we call it we start from checking whether rmap
holds a single spte or points to a descriptor which links more sptes.
In the case of shadow paging, this quadratic total iteration cost is a
problem. Even for two dimensional paging, with EPT/NPT on, in which we
almost always have a single mapping, the extra checks at the end of the
iteration should be eliminated.
This patch fixes this by introducing rmap_iterator which keeps the
iteration context for the next search. Furthermore the implementation
of rmap_next() is splitted into two functions, rmap_get_first() and
rmap_get_next(), to avoid repeatedly checking whether the rmap being
iterated on has only one spte.
Although there seemed to be only a slight change for EPT/NPT, the actual
improvement was significant: we observed that GET_DIRTY_LOG for 1GB
dirty memory became 15% faster than before. This is probably because
the new code is easy to make branch predictions.
Note: we just remove pte_list_next() because we can think of parent_ptes
as a reverse mapping.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
We have PTE_LIST_EXT + 1 pointers in this structure and these 40/20
bytes do not fit cache lines well. Furthermore, some allocators may
use 64/32-byte objects for the pte_list_desc cache.
This patch solves this problem by changing PTE_LIST_EXT from 4 to 3.
For shadow paging, the new size is still large enough to hold both the
kernel and process mappings for usual anonymous pages. For file
mappings, there may be a slight change in the cache usage.
Note: with EPT/NPT we almost always have a single spte in each reverse
mapping and we will not see any change by this.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
Dropped such mappings when we enabled dirty logging and we will never
create new ones until we stop the logging.
For this we introduce a new function which can be used to write protect
a range of PT level pages: although we do not need to care about a range
of pages at this point, the following patch will need this feature to
optimize the write protection of many pages.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
We will use this in the following patch to implement another function
which needs to write protect pages using the rmap information.
Note that there is a small change in debug printing for large pages:
we do not differentiate them from others to avoid duplicating code.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
The reset_rsvds_bits_mask() function can use the guest walker's root level
number instead of using a separate 'level' variable.
Signed-off-by: Davidlohr Bueso <dave@gnu.org>
Signed-off-by: Avi Kivity <avi@redhat.com>
Some members of kvm_memory_slot are not used by every architecture.
This patch is the first step to make this difference clear by
introducing kvm_memory_slot::arch; lpage_info is moved into it.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This patch cleans up the code and removes the "(void)level;" warning
suppressor.
Note that we can also use this for PT_PAGE_TABLE_LEVEL to treat every
level uniformly later.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
We can remove the first ->nx state assignment since it is assigned afterwards anyways.
Signed-off-by: Davidlohr Bueso <dave@gnu.org>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
get_written_sptes is called twice in kvm_mmu_pte_write, one of them can be
removed
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
There is only one user of it and for_each_set_bit() does the same.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
module_param(bool) used to counter-intuitively take an int. In
fddd5201 (mid-2009) we allowed bool or int/unsigned int using a messy
trick.
It's time to remove the int/unsigned int option. For this version
it'll simply give a warning, but it'll break next kernel version.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
freed_pages is never evaluated, so remove it as well as the return code
kvm_mmu_remove_some_alloc_mmu_pages so far delivered to its only user.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The unsync code should be stable now, maybe it is the time to remove this
parameter to cleanup the code a little bit
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Move the mmu code in kvm_arch_vcpu_init() to kvm_mmu_create()
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The tracepoint is only used to audit mmu code, it should not be exposed to
user, let us replace it with jump-label.
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Introduce kvm_for_each_memslot to walk all valid memslot
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently, write protecting a slot needs to walk all the shadow pages
and checks ones which have a pte mapping a page in it.
The walk is overly heavy when dirty pages in that slot are not so many
and checking the shadow pages would result in unwanted cache pollution.
To mitigate this problem, we use rmap_write_protect() and check only
the sptes which can be reached from gfns marked in the dirty bitmap
when the number of dirty pages are less than that of shadow pages.
This criterion is reasonable in its meaning and worked well in our test:
write protection became some times faster than before when the ratio of
dirty pages are low and was not worse even when the ratio was near the
criterion.
Note that the locking for this write protection becomes fine grained.
The reason why this is safe is descripted in the comments.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
rmap_write_protect() calls gfn_to_rmap() for each level with gfn fixed.
This results in calling gfn_to_memslot() repeatedly with that gfn.
This patch introduces __gfn_to_rmap() which takes the slot as an
argument to avoid this.
This is also needed for the following dirty logging optimization.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
Remove redundant checks and use is_large_pte() macro.
Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>
The host side pv mmu support has been marked for feature removal in
January 2011. It's not in use, is slower than shadow or hardware
assisted paging, and a maintenance burden. It's November 2011, time to
remove it.
Signed-off-by: Chris Wright <chrisw@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Detecting write-flooding does not work well, when we handle page written, if
the last speculative spte is not accessed, we treat the page is
write-flooding, however, we can speculative spte on many path, such as pte
prefetch, page synced, that means the last speculative spte may be not point
to the written page and the written page can be accessed via other sptes, so
depends on the Accessed bit of the last speculative spte is not enough
Instead of detected page accessed, we can detect whether the spte is accessed
after it is written, if the spte is not accessed but it is written frequently,
we treat is not a page table or it not used for a long time
Signed-off-by: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
