On Mon, Feb 5, 2024 at 8:25 PM Barry Song <21cnbao@xxxxxxxxx> wrote:
>
Hi, Barry
Thanks for the comments.
> On Mon, Feb 5, 2024 at 7:10 PM Kairui Song <ryncsn@xxxxxxxxx> wrote:
> >
> > From: Kairui Song <kasong@xxxxxxxxxxx>
> >
> > In the direct swapin path, when two or more threads swapin the same entry
> > at the same time, they get different pages (A, B) because swap cache is
> > skipped. 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 modify and swap-out the page again, using the
>
> Even if T0's swap_read_folio is later than T1, problems can still happen.
> after T1 swaps in and sets ptes, then frees the swap entry. T0 reads zRAM
> later. it will get zero as zRAM will fill zero for freed slot,
>
> static int zram_read_from_zspool(struct zram *zram, struct page *page,
> u32 index)
> {
> ...
>
>
> value = handle ? zram_get_element(zram, index) : 0;
> mem = kmap_local_page(page);
> zram_fill_page(mem, PAGE_SIZE, value);
> kunmap_local(mem);
> return 0;
> }
> }
>
> Even though nobody modifies the data before the page is swapped out to the
> same swap offset as before tT0's orig_pte, T0's pte_same check is still true
> and T0 will map filled zeroed page to pte.
>
> so there is more than one risk besides modified data losses.
Thanks for the complement, I think this is true, and it shares the
same problem of the entry reuse, so this patch also covered this
potential race. I can add more words later to cover this case as well.
>
> > same entry. It break the pte_same check because PTE value is unchanged,
> > causing ABA problem. Then thread (T0) will then install the stalled page
> > (A) into the PTE so new data in page (B) is lost, 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_readpage() <- read to page A swap_readpage() <- read to page B
> > <slow on later locks or interrupt> <finished swapin first>
> > .. set_pte_at()
> > swap_free() <- Now the entry is freed.
> > <write to page B, now page A stalled>
> > <swap out page B using 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!
> >
> > 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 the
> > cached swapin path, two swapin path being used at the same time
> > 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!
>
> i am also reading these codes recently. It is quite unbelievable this
> is really happening
> now. as freeing swaps is returning slot to slots_ret, but allocating
> swap is from slots.
> so if swapfile is large, the chance that the newly allocated swap was
> a recently freed swap
> is close to 0%. but yes, the code does have the risk.
Indeed, for reproducing I used a 32M swap device, and the data being
swapped in/out is large enough to make full use of it. So the
reproduce rate is increased by a lot. It's not a completely fictional
test as some low end device do have smaller swaps, and real world race
could happen in many strange ways.
> >
> > 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")
> > Link: https://github.com/ryncsn/emm-test-project/tree/master/swap-stress-race [1]
> > Signed-off-by: Kairui Song <kasong@xxxxxxxxxxx>
>
> I will also run your patch on my problem I reported today[1]. will update
> the result to you this week.
>
> [1] https://lore.kernel.org/linux-mm/d4f602db-403b-4b1f-a3de-affeb40bc499@xxxxxxx/T/#m41701d0c0e127cdae636e97a13ab521364a810f4
>
Thanks!
