| Message ID | 20230607053135.2087354-1-stevensd@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm/khugepaged: fix iteration in collapse_file | expand |
On Wed, Jun 07, 2023 at 02:31:35PM +0900, David Stevens wrote: > From: David Stevens <stevensd@chromium.org> > > Remove an unnecessary call to xas_set(index) when iterating over the > target range in collapse_file. The extra call to xas_set reset the xas > cursor to the top of the tree, causing the xas_next call on the next > iteration to walk the tree to index instead of advancing to index+1. > This returned the same page again, which would cause collapse_file to > fail because the page is already locked. > > This bug was hidden when CONFIG_DEBUG_VM was set. When that config was > used, the xas_load in a subsequent VM_BUG_ON assert would walk xas from > the top of the tree to index, causing the xas_next call on the next loop > iteration to advance the cursor as expected. > > Fixes: a2e17cc2efc7 ("mm/khugepaged: maintain page cache uptodate flag") > Signed-off-by: David Stevens <stevensd@chromium.org> Reviewed-by: Peter Xu <peterx@redhat.com>
On Wed, 7 Jun 2023, David Stevens wrote: > > Remove an unnecessary call to xas_set(index) when iterating over the > target range in collapse_file. The extra call to xas_set reset the xas > cursor to the top of the tree, causing the xas_next call on the next > iteration to walk the tree to index instead of advancing to index+1. > This returned the same page again, which would cause collapse_file to > fail because the page is already locked. > > This bug was hidden when CONFIG_DEBUG_VM was set. When that config was > used, the xas_load in a subsequent VM_BUG_ON assert would walk xas from > the top of the tree to index, causing the xas_next call on the next loop > iteration to advance the cursor as expected. > > Fixes: a2e17cc2efc7 ("mm/khugepaged: maintain page cache uptodate flag") > Signed-off-by: David Stevens <stevensd@chromium.org> This patch seems to be wrong, but I have not investigated why. It's certainly an interesting and worrying observation, if a CONFIG_DEBUG_VM=y kernel goes a significantly different way. I almost always do have CONFIG_DEBUG_VM=y, so you won't be surprised that I never saw the issue. But once I ran an mm-everything with this patch in, I hit that VM_BUG_ON_PAGE(page != xas_load(&xas), page) for the first time (after about 2 hours of huge tmpfs swapping load). As if you have just transferred the problem from DEBUG_VM=n to DEBUG_VM=y. But I then tried a CONFIG_DEBUG_VM off 6.4-rc1 kernel (including the fixee but not this fixer) under similar load, and saw no problem in 14 hours. So I can't even reproduce the bug that is being fixed here: only hit a bug that it introduces. Hugh > --- > mm/khugepaged.c | 1 - > 1 file changed, 1 deletion(-) > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > index 6b9d39d65b73..2d0d58fb4e7f 100644 > --- a/mm/khugepaged.c > +++ b/mm/khugepaged.c > @@ -2070,7 +2070,6 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr, > TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); > > xas_lock_irq(&xas); > - xas_set(&xas, index); > > VM_BUG_ON_PAGE(page != xas_load(&xas), page); > > -- > 2.41.0.rc2.161.g9c6817b8e7-goog
On Sat, Jun 10, 2023 at 5:03 AM Hugh Dickins <hughd@google.com> wrote: > > On Wed, 7 Jun 2023, David Stevens wrote: > > > > Remove an unnecessary call to xas_set(index) when iterating over the > > target range in collapse_file. The extra call to xas_set reset the xas > > cursor to the top of the tree, causing the xas_next call on the next > > iteration to walk the tree to index instead of advancing to index+1. > > This returned the same page again, which would cause collapse_file to > > fail because the page is already locked. > > > > This bug was hidden when CONFIG_DEBUG_VM was set. When that config was > > used, the xas_load in a subsequent VM_BUG_ON assert would walk xas from > > the top of the tree to index, causing the xas_next call on the next loop > > iteration to advance the cursor as expected. > > > > Fixes: a2e17cc2efc7 ("mm/khugepaged: maintain page cache uptodate flag") > > Signed-off-by: David Stevens <stevensd@chromium.org> > > This patch seems to be wrong, but I have not investigated why. > > It's certainly an interesting and worrying observation, > if a CONFIG_DEBUG_VM=y kernel goes a significantly different way. > > I almost always do have CONFIG_DEBUG_VM=y, so you won't be surprised that > I never saw the issue. But once I ran an mm-everything with this patch in, > I hit that VM_BUG_ON_PAGE(page != xas_load(&xas), page) for the first time > (after about 2 hours of huge tmpfs swapping load). Is the particular workload one you can share? I haven't hit that assert so far with my tests. Also, I'm a little surprised that this is the assert which is being hit. My patch series didn't make that many changes to the first loop of the function, and the changes it did make were mostly about missing pages, not present pages. > As if you have just transferred the problem from DEBUG_VM=n to DEBUG_VM=y. > But I then tried a CONFIG_DEBUG_VM off 6.4-rc1 kernel (including the fixee > but not this fixer) under similar load, and saw no problem in 14 hours. > So I can't even reproduce the bug that is being fixed here: only hit a > bug that it introduces. The bug this fixes isn't a crash - it's the fact that khugepage becomes nearly unable to collapse pages. Specifically, it can only collapse if there is exactly one present page, and that page is at index 511. Since MADV_COLLAPSE uses the same code path, it's easy to reproduce the bug I'm trying to fix with this program: #define _GNU_SOURCE #include <assert.h> #include <stdio.h> #include <stdlib.h> #include <sys/mman.h> #include <unistd.h> #define THP_SIZE (2 * 1024 * 1024) #ifndef MADV_COLLAPSE #define MADV_COLLAPSE 25 #endif int main() { int memfd = memfd_create("memfd", MFD_CLOEXEC); assert(memfd >= 0); int ret = ftruncate(memfd, THP_SIZE); assert(ret >= 0); char *addr = mmap(NULL, THP_SIZE, PROT_READ|PROT_WRITE, MAP_SHARED, memfd, 0); assert(addr != MAP_FAILED); addr[0] = 0xff; ret = madvise(addr, THP_SIZE, MADV_COLLAPSE); assert(ret == 0); } If DEBUG_VM isn't set, then the test will trigger the assert. If DEBUG_VM is set or if this fix is included, then the test will pass. -David > Hugh > > > --- > > mm/khugepaged.c | 1 - > > 1 file changed, 1 deletion(-) > > > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > > index 6b9d39d65b73..2d0d58fb4e7f 100644 > > --- a/mm/khugepaged.c > > +++ b/mm/khugepaged.c > > @@ -2070,7 +2070,6 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr, > > TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); > > > > xas_lock_irq(&xas); > > - xas_set(&xas, index); > > > > VM_BUG_ON_PAGE(page != xas_load(&xas), page); > > > > -- > > 2.41.0.rc2.161.g9c6817b8e7-goog
On Sat, 10 Jun 2023, David Stevens wrote: > On Sat, Jun 10, 2023 at 5:03 AM Hugh Dickins <hughd@google.com> wrote: > > On Wed, 7 Jun 2023, David Stevens wrote: > > > > > > Remove an unnecessary call to xas_set(index) when iterating over the > > > target range in collapse_file. The extra call to xas_set reset the xas > > > cursor to the top of the tree, causing the xas_next call on the next > > > iteration to walk the tree to index instead of advancing to index+1. > > > This returned the same page again, which would cause collapse_file to > > > fail because the page is already locked. > > > > > > This bug was hidden when CONFIG_DEBUG_VM was set. When that config was > > > used, the xas_load in a subsequent VM_BUG_ON assert would walk xas from > > > the top of the tree to index, causing the xas_next call on the next loop > > > iteration to advance the cursor as expected. > > > > > > Fixes: a2e17cc2efc7 ("mm/khugepaged: maintain page cache uptodate flag") > > > Signed-off-by: David Stevens <stevensd@chromium.org> > > > > This patch seems to be wrong, I withdraw the accusation! I haven't thought about the patch itself: I'd have to learn a bit more about the xarray rules, so cannot quickly ack it myself; but withdraw my implicit nack. > > but I have not investigated why. > > > > It's certainly an interesting and worrying observation, > > if a CONFIG_DEBUG_VM=y kernel goes a significantly different way. > > > > I almost always do have CONFIG_DEBUG_VM=y, so you won't be surprised that > > I never saw the issue. But once I ran an mm-everything with this patch in, > > I hit that VM_BUG_ON_PAGE(page != xas_load(&xas), page) for the first time > > (after about 2 hours of huge tmpfs swapping load). > > Is the particular workload one you can share? It's a workload familiar to me, and no secret, but which would take much more time to share than I have to spare; with no great likelihood that you could easily reproduce all the conditions at the end of it. > I haven't hit that assert so far with my tests. And nor have I since. I've been trying 6.4-rc5 with and without your patch, mm-everything (of a few days ago) with and without your patch, with and without my pte_offset_map changes. Not once have I seen that VM_BUG_ON_PAGE(page != xas_load(&xas), page) again, even with the exact same kernel. I've no doubt that I did hit it, but I'm now having to assume that it's something very rare (or very rare in my particular testing). And because the first time I saw it coincided with the first time I had in your patch to the last non-blank line above it, I jumped to the conclusion that it was your patch causing it. Not unreasonable at the time, but not justified by later attempts. Perhaps there was something unrelated in that kernel, corrupting an xarray node. If I ever see it again, I hope I'll be able to spend more time investigating; or if I study that code with more leisure, maybe a hypothesis for a rare race will arise. But for now, we'd better just forget about it. > > Also, I'm a little surprised that this is the assert which is being > hit. My patch series didn't make that many changes to the first loop > of the function, and the changes it did make were mostly about missing > pages, not present pages. I was suspicious of that patch series, as I said at the time; and might one day want to revert it; but it has not given me any actual problem - that only came with this fix to it. > > > As if you have just transferred the problem from DEBUG_VM=n to DEBUG_VM=y. > > But I then tried a CONFIG_DEBUG_VM off 6.4-rc1 kernel (including the fixee > > but not this fixer) under similar load, and saw no problem in 14 hours. > > So I can't even reproduce the bug that is being fixed here: only hit a > > bug that it introduces. > > The bug this fixes isn't a crash - it's the fact that khugepage Sorry, yes, I should have re-read your commit message before writing all that. Thanks, Hugh > becomes nearly unable to collapse pages. Specifically, it can only > collapse if there is exactly one present page, and that page is at > index 511. Since MADV_COLLAPSE uses the same code path, it's easy to > reproduce the bug I'm trying to fix with this program: > > #define _GNU_SOURCE > #include <assert.h> > #include <stdio.h> > #include <stdlib.h> > #include <sys/mman.h> > #include <unistd.h> > > #define THP_SIZE (2 * 1024 * 1024) > > #ifndef MADV_COLLAPSE > #define MADV_COLLAPSE 25 > #endif > > int main() { > int memfd = memfd_create("memfd", MFD_CLOEXEC); > assert(memfd >= 0); > > int ret = ftruncate(memfd, THP_SIZE); > assert(ret >= 0); > > char *addr = mmap(NULL, THP_SIZE, PROT_READ|PROT_WRITE, MAP_SHARED, memfd, 0); > assert(addr != MAP_FAILED); > > addr[0] = 0xff; > > ret = madvise(addr, THP_SIZE, MADV_COLLAPSE); > assert(ret == 0); > } > > If DEBUG_VM isn't set, then the test will trigger the assert. If > DEBUG_VM is set or if this fix is included, then the test will pass. > > -David > > > Hugh > > > > > --- > > > mm/khugepaged.c | 1 - > > > 1 file changed, 1 deletion(-) > > > > > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > > > index 6b9d39d65b73..2d0d58fb4e7f 100644 > > > --- a/mm/khugepaged.c > > > +++ b/mm/khugepaged.c > > > @@ -2070,7 +2070,6 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr, > > > TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); > > > > > > xas_lock_irq(&xas); > > > - xas_set(&xas, index); > > > > > > VM_BUG_ON_PAGE(page != xas_load(&xas), page); > > > > > > -- > > > 2.41.0.rc2.161.g9c6817b8e7-goog
diff --git a/mm/khugepaged.c b/mm/khugepaged.c index 6b9d39d65b73..2d0d58fb4e7f 100644 --- a/mm/khugepaged.c +++ b/mm/khugepaged.c @@ -2070,7 +2070,6 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr, TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); xas_lock_irq(&xas); - xas_set(&xas, index); VM_BUG_ON_PAGE(page != xas_load(&xas), page);