lkml.org 
[lkml]   [2018]   [Apr]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v9 00/24] Speculative page faults
On Tue, Mar 13, 2018 at 06:59:30PM +0100, Laurent Dufour wrote:
> This is a port on kernel 4.16 of the work done by Peter Zijlstra to
> handle page fault without holding the mm semaphore [1].
>
> The idea is to try to handle user space page faults without holding the
> mmap_sem. This should allow better concurrency for massively threaded
> process since the page fault handler will not wait for other threads memory
> layout change to be done, assuming that this change is done in another part
> of the process's memory space. This type page fault is named speculative
> page fault. If the speculative page fault fails because of a concurrency is
> detected or because underlying PMD or PTE tables are not yet allocating, it
> is failing its processing and a classic page fault is then tried.
>
> The speculative page fault (SPF) has to look for the VMA matching the fault
> address without holding the mmap_sem, this is done by introducing a rwlock
> which protects the access to the mm_rb tree. Previously this was done using
> SRCU but it was introducing a lot of scheduling to process the VMA's
> freeing
> operation which was hitting the performance by 20% as reported by Kemi Wang
> [2].Using a rwlock to protect access to the mm_rb tree is limiting the
> locking contention to these operations which are expected to be in a O(log
> n)
> order. In addition to ensure that the VMA is not freed in our back a
> reference count is added and 2 services (get_vma() and put_vma()) are
> introduced to handle the reference count. When a VMA is fetch from the RB
> tree using get_vma() is must be later freeed using put_vma(). Furthermore,
> to allow the VMA to be used again by the classic page fault handler a
> service is introduced can_reuse_spf_vma(). This service is expected to be
> called with the mmap_sem hold. It checked that the VMA is still matching
> the specified address and is releasing its reference count as the mmap_sem
> is hold it is ensure that it will not be freed in our back. In general, the
> VMA's reference count could be decremented when holding the mmap_sem but it
> should not be increased as holding the mmap_sem is ensuring that the VMA is
> stable. I can't see anymore the overhead I got while will-it-scale
> benchmark anymore.
>
> The VMA's attributes checked during the speculative page fault processing
> have to be protected against parallel changes. This is done by using a per
> VMA sequence lock. This sequence lock allows the speculative page fault
> handler to fast check for parallel changes in progress and to abort the
> speculative page fault in that case.
>
> Once the VMA is found, the speculative page fault handler would check for
> the VMA's attributes to verify that the page fault has to be handled
> correctly or not. Thus the VMA is protected through a sequence lock which
> allows fast detection of concurrent VMA changes. If such a change is
> detected, the speculative page fault is aborted and a *classic* page fault
> is tried. VMA sequence lockings are added when VMA attributes which are
> checked during the page fault are modified.
>
> When the PTE is fetched, the VMA is checked to see if it has been changed,
> so once the page table is locked, the VMA is valid, so any other changes
> leading to touching this PTE will need to lock the page table, so no
> parallel change is possible at this time.

What would have been nice is some pseudo highlevel code before all the
above detailed description. Something like:
speculative_fault(addr) {
mm_lock_for_vma_snapshot()
vma_snapshot = snapshot_vma_infos(addr)
mm_unlock_for_vma_snapshot()
...
if (!vma_can_speculatively_fault(vma_snapshot, addr))
return;
...
/* Do fault ie alloc memory, read from file ... */
page = ...;

preempt_disable();
if (vma_snapshot_still_valid(vma_snapshot, addr) &&
vma_pte_map_lock(vma_snapshot, addr)) {
if (pte_same(ptep, orig_pte)) {
/* Setup new pte */
page = NULL;
}
}
preempt_enable();
if (page)
put(page)
}

I just find pseudo code easier for grasping the highlevel view of the
expected code flow.


>
> The locking of the PTE is done with interrupts disabled, this allows to
> check for the PMD to ensure that there is not an ongoing collapsing
> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
> waiting for the other CPU to have catch the IPI interrupt, if the pmd is
> valid at the time the PTE is locked, we have the guarantee that the
> collapsing opertion will have to wait on the PTE lock to move foward. This
> allows the SPF handler to map the PTE safely. If the PMD value is different
> than the one recorded at the beginning of the SPF operation, the classic
> page fault handler will be called to handle the operation while holding the
> mmap_sem. As the PTE lock is done with the interrupts disabled, the lock is
> done using spin_trylock() to avoid dead lock when handling a page fault
> while a TLB invalidate is requested by an other CPU holding the PTE.
>
> Support for THP is not done because when checking for the PMD, we can be
> confused by an in progress collapsing operation done by khugepaged. The
> issue is that pmd_none() could be true either if the PMD is not already
> populated or if the underlying PTE are in the way to be collapsed. So we
> cannot safely allocate a PMD if pmd_none() is true.

Might be a good topic fo LSF/MM, should we set the pmd to something
else then 0 when collapsing pmd (apply to pud too) ? This would allow
support THP.

[...]

>
> Ebizzy:
> -------
> The test is counting the number of records per second it can manage, the
> higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
> result I repeated the test 100 times and measure the average result. The
> number is the record processes per second, the higher is the best.
>
> BASE SPF delta
> 16 CPUs x86 VM 14902.6 95905.16 543.55%
> 80 CPUs P8 node 37240.24 78185.67 109.95%

I find those results interesting as it seems that SPF do not scale well
on big configuration. Note that it still have a sizeable improvement so
it is still a very interesting feature i believe.

Still understanding what is happening here might a good idea. From the
numbers below it seems there is 2 causes to the scaling issue. First
pte lock contention (kind of expected i guess). Second changes to vma
while faulting.

Have you thought about this ? Do i read those numbers in the wrong way ?

>
> Here are the performance counter read during a run on a 16 CPUs x86 VM:
> Performance counter stats for './ebizzy -mRTp':
> 888157 faults
> 884773 spf
> 92 pagefault:spf_pte_lock
> 2379 pagefault:spf_vma_changed
> 0 pagefault:spf_vma_noanon
> 80 pagefault:spf_vma_notsup
> 0 pagefault:spf_vma_access
> 0 pagefault:spf_pmd_changed
>
> And the ones captured during a run on a 80 CPUs Power node:
> Performance counter stats for './ebizzy -mRTp':
> 762134 faults
> 728663 spf
> 19101 pagefault:spf_pte_lock
> 13969 pagefault:spf_vma_changed
> 0 pagefault:spf_vma_noanon
> 272 pagefault:spf_vma_notsup
> 0 pagefault:spf_vma_access
> 0 pagefault:spf_pmd_changed


There is one aspect that i would like to see cover. Maybe i am not
understanding something fundamental, but it seems to me that SPF can
trigger OOM or at very least over stress page allocation.

Assume you have a lot of concurrent SPF to anonymous vma and they all
allocate new pages, then you might overallocate for a single address
by a factor correlated with the number of CPUs in your system. Now,
multiply this for several distinc address and you might be allocating
a lot of memory transiently ie just for a short period time. While
the fact that you quickly free when you fail should prevent the OOM
reaper. But still this might severly stress the memory allocation
path.

Am i missing something in how this all work ? Or is the above some-
thing that might be of concern ? Should there be some boundary on the
maximum number of concurrent SPF (and thus boundary on maximum page
temporary page allocation) ?

Cheers,
Jérôme

\
 
 \ /
  Last update: 2018-04-03 22:38    [W:0.542 / U:0.628 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site