On Tue, Feb 6, 2024 at 6:21 PM Kairui Song <[email protected]> wrote:
>
> On Wed, Feb 7, 2024 at 10:03 AM Chris Li <[email protected]> wrote:
> >
> > On Tue, Feb 6, 2024 at 4:43 PM Barry Song <[email protected]> wrote:
> > >
> > > On Wed, Feb 7, 2024 at 7:18 AM Chris Li <[email protected]> wrote:
> > > >
> > > > Hi Kairui,
> > > >
> > > > Sorry replying to your patch V1 late, I will reply on the V2 thread.
> > > >
> > > > On Tue, Feb 6, 2024 at 10:28 AM Kairui Song <[email protected]> wrote:
> > > > >
> > > > > From: Kairui Song <[email protected]>
> > > > >
> > > > > When skipping swapcache for SWP_SYNCHRONOUS_IO, if two or more threads
> > > > > swapin the same entry at the same time, they get different pages (A, B).
> > > > > Before one thread (T0) finishes the swapin and installs page (A)
> > > > > to the PTE, another thread (T1) could finish swapin of page (B),
> > > > > swap_free the entry, then swap out the possibly modified page
> > > > > reusing the same entry. It breaks the pte_same check in (T0) because
> > > > > PTE value is unchanged, causing ABA problem. Thread (T0) will
> > > > > install a stalled page (A) into the PTE and cause data corruption.
> > > > >
> > > > > One possible callstack is like this:
> > > > >
> > > > > CPU0 CPU1
> > > > > ---- ----
> > > > > do_swap_page() do_swap_page() with same entry
> > > > > <direct swapin path> <direct swapin path>
> > > > > <alloc page A> <alloc page B>
> > > > > swap_read_folio() <- read to page A swap_read_folio() <- read to page B
> > > > > <slow on later locks or interrupt> <finished swapin first>
> > > > > ... set_pte_at()
> > > > > swap_free() <- entry is free
> > > > > <write to page B, now page A stalled>
> > > > > <swap out page B to same swap entry>
> > > > > pte_same() <- Check pass, PTE seems
> > > > > unchanged, but page A
> > > > > is stalled!
> > > > > swap_free() <- page B content lost!
> > > > > set_pte_at() <- staled page A installed!
> > > > >
> > > > > And besides, for ZRAM, swap_free() allows the swap device to discard
> > > > > the entry content, so even if page (B) is not modified, if
> > > > > swap_read_folio() on CPU0 happens later than swap_free() on CPU1,
> > > > > it may also cause data loss.
> > > > >
> > > > > To fix this, reuse swapcache_prepare which will pin the swap entry using
> > > > > the cache flag, and allow only one thread to pin it. Release the pin
> > > > > after PT unlocked. Racers will simply busy wait since it's a rare
> > > > > and very short event.
> > > > >
> > > > > Other methods like increasing the swap count don't seem to be a good
> > > > > idea after some tests, that will cause racers to fall back to use the
> > > > > swap cache again. Parallel swapin using different methods leads to
> > > > > a much more complex scenario.
> > > > >
> > > > > Reproducer:
> > > > >
> > > > > This race issue can be triggered easily using a well constructed
> > > > > reproducer and patched brd (with a delay in read path) [1]:
> > > > >
> > > > > With latest 6.8 mainline, race caused data loss can be observed easily:
> > > > > $ gcc -g -lpthread test-thread-swap-race.c && ./a.out
> > > > > Polulating 32MB of memory region...
> > > > > Keep swapping out...
> > > > > Starting round 0...
> > > > > Spawning 65536 workers...
> > > > > 32746 workers spawned, wait for done...
> > > > > Round 0: Error on 0x5aa00, expected 32746, got 32743, 3 data loss!
> > > > > Round 0: Error on 0x395200, expected 32746, got 32743, 3 data loss!
> > > > > Round 0: Error on 0x3fd000, expected 32746, got 32737, 9 data loss!
> > > > > Round 0 Failed, 15 data loss!
> > > > >
> > > > > This reproducer spawns multiple threads sharing the same memory region
> > > > > using a small swap device. Every two threads updates mapped pages one by
> > > > > one in opposite direction trying to create a race, with one dedicated
> > > > > thread keep swapping out the data out using madvise.
> > > > >
> > > > > The reproducer created a reproduce rate of about once every 5 minutes,
> > > > > so the race should be totally possible in production.
> > > > >
> > > > > After this patch, I ran the reproducer for over a few hundred rounds
> > > > > and no data loss observed.
> > > > >
> > > > > Performance overhead is minimal, microbenchmark swapin 10G from 32G
> > > > > zram:
> > > > >
> > > > > Before: 10934698 us
> > > > > After: 11157121 us
> > > > > Non-direct: 13155355 us (Dropping SWP_SYNCHRONOUS_IO flag)
> > > > >
> > > > > Fixes: 0bcac06f27d7 ("mm, swap: skip swapcache for swapin of synchronous device")
> > > > > Reported-by: "Huang, Ying" <[email protected]>
> > > > > Closes: https://lore.kernel.org/lkml/[email protected]/
> > > > > Link: https://github.com/ryncsn/emm-test-project/tree/master/swap-stress-race [1]
> > > > > Signed-off-by: Kairui Song <[email protected]>
> > > > > Reviewed-by: "Huang, Ying" <[email protected]>
> > > > > Acked-by: Yu Zhao <[email protected]>
> > > > >
> > > > > ---
> > > > > Update from V1:
> > > > > - Add some words on ZRAM case, it will discard swap content on swap_free so the race window is a bit different but cure is the same. [Barry Song]
> > > > > - Update comments make it cleaner [Huang, Ying]
> > > > > - Add a function place holder to fix CONFIG_SWAP=n built [SeongJae Park]
> > > > > - Update the commit message and summary, refer to SWP_SYNCHRONOUS_IO instead of "direct swapin path" [Yu Zhao]
> > > > > - Update commit message.
> > > > > - Collect Review and Acks.
> > > > >
> > > > > include/linux/swap.h | 5 +++++
> > > > > mm/memory.c | 15 +++++++++++++++
> > > > > mm/swap.h | 5 +++++
> > > > > mm/swapfile.c | 13 +++++++++++++
> > > > > 4 files changed, 38 insertions(+)
> > > > >
> > > > > diff --git a/include/linux/swap.h b/include/linux/swap.h
> > > > > index 4db00ddad261..8d28f6091a32 100644
> > > > > --- a/include/linux/swap.h
> > > > > +++ b/include/linux/swap.h
> > > > > @@ -549,6 +549,11 @@ static inline int swap_duplicate(swp_entry_t swp)
> > > > > return 0;
> > > > > }
> > > > >
> > > > > +static inline int swapcache_prepare(swp_entry_t swp)
> > > > > +{
> > > > > + return 0;
> > > > > +}
> > > > > +
> > > > > static inline void swap_free(swp_entry_t swp)
> > > > > {
> > > > > }
> > > > > diff --git a/mm/memory.c b/mm/memory.c
> > > > > index 7e1f4849463a..1749c700823d 100644
> > > > > --- a/mm/memory.c
> > > > > +++ b/mm/memory.c
> > > > > @@ -3867,6 +3867,16 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
> > > > > if (!folio) {
> > > > > if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
> > > > > __swap_count(entry) == 1) {
> > > > > + /*
> > > > > + * Prevent parallel swapin from proceeding with
> > > > > + * the cache flag. Otherwise, another thread may
> > > > > + * finish swapin first, free the entry, and swapout
> > > > > + * reusing the same entry. It's undetectable as
> > > > > + * pte_same() returns true due to entry reuse.
> > > > > + */
> > > > > + if (swapcache_prepare(entry))
> > > > > + goto out;
> > > > > +
> > > >
> > > > I am puzzled by this "goto out". If I understand this correctly, you
> > > > have two threads CPU1 and CPU2 racing to set the flag SWAP_HAS_CACHE.
> > > > The CPU1 will succeed in adding the flag and the CPU2 will get
> > > > "-EEXIST" from "swapcache_prepare(entry)". Am I understanding it
> > > > correctly so far?
> > > >
> > > > Then the goto out seems wrong to me. For the CPU2, the page fault will
> > > > return *unhandled*. Even worse, the "-EEXIST" error is not preserved,
> > > > CPU2 does not even know the page fault is not handled, it will resume
> > > > from the page fault instruction, possibly generate another page fault
> > > > at the exact same location. That page fault loop will repeat until
> > > > CPU1 install the new pte on that faulting virtual address and pick up
> > > > by CPU2.
> > > >
> > > > Am I missing something obvious there?
> > >
> > > I feel you are right. any concurrent page faults at the same pte
> > > will increase the count of page faults for a couple of times now.
> > >
> > > >
> > > > I just re-read your comment: "Racers will simply busy wait since it's
> > > > a rare and very short event." That might be referring to the above
> > > > CPU2 page fault looping situation. I consider the page fault looping
> > > > on CPU2 not acceptable. For one it will mess up the page fault
> > > > statistics.
> > > > In my mind, having an explicit loop for CPU2 waiting for the PTE to
> > > > show up is still better than this page fault loop. You can have more
> > > > CPU power friendly loops.
> > >
> > > I assume you mean something like
> > >
> > > while(!pte_same())
> > > cpu_relax();
> > >
> > > then we still have a chance to miss the change of B.
> > >
> > > For example, another thread is changing pte to A->B->A, our loop can
> > > miss B. Thus we will trap into an infinite loop. this is even worse.
> >
> > Yes. You are right, it is worse. Thanks for catching that. That is why
> > I say this needs more discussion, I haven't fully thought it through
> > :-)
>
> Hi Chris and Barry,
>
> Thanks for the comments!
>
> The worst thing I know of returning in do_swap_page without handling
> the swap, is an increase of some statistic counters, note it will not
> cause major page fault counters to grow, only things like perf counter
> and vma lock statistic are affected.
>
> And actually there are multiple already existing return points in
> do_swap_page that will return without handling it, which may
> re-trigger the page fault.
Thanks for pointing that out. I take a look at those, which seems
different than the case here. In those cases, it truely can not make
forward progress.
Here we actually have all the data it needs to complete the page
fault. Just a data synchronization issue preventing making forward
progress.
Ideally we can have some clever data structure to solve the
synchronization issue and make forward progress.
> When do_swap_page is called, many pre-checks have been applied, and
> they could all be invalidated if something raced, simply looping
> inside here could miss a lot of corner cases, so we have to go through
> that again.
Actually, I think about it. Looping it here seems worse in the sense
that it is already holding some locks. Return and retry the page fault
at least release those locks and let others have a chance to make
progress.
>
> This patch did increase the chance of false positive increase of some
> counters, maybe something like returning a VM_FAULT_RETRY could make
> it better, but code is more complex and will cause other counters to
> grow.
This is certainly not ideal. It might upset the feedback loop that
uses the swap fault statistic as input to adjust the swapping
behavior.
Chris