On Fri, Oct 11, 2019 at 2:05 PM Josef Bacik josef@toxicpanda.com wrote:

We hit a regression while rolling out 5.2 internally where we were hitting the following panic

kernel BUG at mm/page-writeback.c:2659! RIP: 0010:clear_page_dirty_for_io+0xe6/0x1f0 Call Trace: __process_pages_contig+0x25a/0x350 ? extent_clear_unlock_delalloc+0x43/0x70 submit_compressed_extents+0x359/0x4d0 normal_work_helper+0x15a/0x330 process_one_work+0x1f5/0x3f0 worker_thread+0x2d/0x3d0 ? rescuer_thread+0x340/0x340 kthread+0x111/0x130 ? kthread_create_on_node+0x60/0x60 ret_from_fork+0x1f/0x30

this is happening because the page is not locked when doing clear_page_dirty_for_io. Looking at the core dump it was because our async_extent had a ram_size of 24576 but our async_chunk range only spanned 20480, so we had a whole extra page in our ram_size for our async_extent.

This happened because we try not to compress pages outside of our i_size, however a cleanup patch changed us to do

actual_end = min_t(u64, i_size_read(inode), end + 1);

which is problematic because i_size_read() can evaluate to different values in between checking and assigning. So either a expanding truncate or a fallocate could increase our i_size while we're doing

Well, not just those operations but a write starting at or beyond eof could also do it, since at writeback time we don't hold the inode's lock and the buffered write path doesn't lock the range if it starts at or past current i_size.

writeout and actual_end would end up being past the range we have locked.

I confirmed this was what was happening by installing a debug kernel that had

actual_end = min_t(u64, i_size_read(inode), end + 1); if (actual_end > end + 1) { printk(KERN_ERR "WE GOT FUCKED\n");

I think we coud be more polite on changelogs and mailing lists, or we could add a tag in the subjects warning about improper content like video games and music records :)

    actual_end = end + 1;

}

and installing it onto 500 boxes of the tier that had been seeing the problem regularly. Last night I got my debug message and no panic, confirming what I expected.

Fixes: 62b37622718c ("btrfs: Remove isize local variable in compress_file_range") cc: stable@vger.kernel.org Signed-off-by: Josef Bacik josef@toxicpanda.com

Anyway, looks good to me.

Reviewed-by: Filipe Manana fdmanana@suse.com

---

fs/btrfs/inode.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c index 2eb1d7249f83..9a483d1f61f8 100644 --- a/fs/btrfs/inode.c +++ b/fs/btrfs/inode.c @@ -474,6 +474,7 @@ static noinline int compress_file_range(struct async_chunk *async_chunk) u64 start = async_chunk->start; u64 end = async_chunk->end; u64 actual_end;

•   loff_t i_size = i_size_read(inode);
    int ret = 0;
    struct page **pages = NULL;
    unsigned long nr_pages;
  

@@ -488,7 +489,13 @@ static noinline int compress_file_range(struct async_chunk *async_chunk) inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1, SZ_16K);

•   actual_end = min_t(u64, i_size_read(inode), end + 1);
  

•   /*
  

•    * We need to save i_size before now because it could change in between
  

•    * us evaluating the size and assigning it.  This is because we lock and
  

•    * unlock the page in truncate and fallocate, and then modify the i_size
  

•    * later on.
  

•    */
  

•   actual_end = min_t(u64, i_size, end + 1);
  

again: will_compress = 0; nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1; -- 2.21.0

On Fri, Oct 11, 2019 at 09:03:54AM -0400, Josef Bacik wrote:

this is happening because the page is not locked when doing clear_page_dirty_for_io. Looking at the core dump it was because our async_extent had a ram_size of 24576 but our async_chunk range only spanned 20480, so we had a whole extra page in our ram_size for our async_extent.

This happened because we try not to compress pages outside of our i_size, however a cleanup patch changed us to do

actual_end = min_t(u64, i_size_read(inode), end + 1);

which is problematic because i_size_read() can evaluate to different values in between checking and assigning.

The other patch adding READ_ONCE to i_size_read actually has the interesting bits why this happens, so this changelog should have them too.

So either a expanding truncate or a fallocate could increase our i_size while we're doing writeout and actual_end would end up being past the range we have locked.

Yeah, it's there, the problematic part is the double load of inode size

mov 0x20(%rsp),%rax cmp %rax,0x48(%r15) movl $0x0,0x18(%rsp) mov %rax,%r12 mov %r14,%rax cmovbe 0x48(%r15),%r12

Where r15 is inode and 0x48 is offset of i_size.

I confirmed this was what was happening by installing a debug kernel that had

actual_end = min_t(u64, i_size_read(inode), end + 1); if (actual_end > end + 1) { printk(KERN_ERR "WE GOT FUCKED\n"); actual_end = end + 1; }

and installing it onto 500 boxes of the tier that had been seeing the problem regularly. Last night I got my debug message and no panic, confirming what I expected.

Fixes: 62b37622718c ("btrfs: Remove isize local variable in compress_file_range") cc: stable@vger.kernel.org Signed-off-by: Josef Bacik josef@toxicpanda.com

---

fs/btrfs/inode.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c index 2eb1d7249f83..9a483d1f61f8 100644 --- a/fs/btrfs/inode.c +++ b/fs/btrfs/inode.c @@ -474,6 +474,7 @@ static noinline int compress_file_range(struct async_chunk *async_chunk) u64 start = async_chunk->start; u64 end = async_chunk->end; u64 actual_end;

• loff_t i_size = i_size_read(inode); int ret = 0; struct page **pages = NULL; unsigned long nr_pages;

@@ -488,7 +489,13 @@ static noinline int compress_file_range(struct async_chunk *async_chunk) inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1, SZ_16K);

• actual_end = min_t(u64, i_size_read(inode), end + 1);

• /*

• * We need to save i_size before now because it could change in between
  

• * us evaluating the size and assigning it.  This is because we lock and
  

• * unlock the page in truncate and fallocate, and then modify the i_size
  

• * later on.
  

• */
  

• actual_end = min_t(u64, i_size, end + 1);

Unfortunatelly this is not sufficient and does not prevent the compiler to do the same optimization, though current compilers seem to do just one load it's not prevented in principle. The only cure is READ_ONCE, either globally or locally done in compress_file_range.

The other patch addding READ_ONCE to i_size_read would be IMHO ideal.

I've experimented with various ways to do the 'once' semantics, the following worked:

* force READ_ONCE on the isize varaible, as READ_ONCE requires lvalue, we can't do READ_ONCE(i_size_read())

u64 isize; ... isize = i_read_size(inode); isize = READ_ONCE(isize);

* insert barrier

u64 isize; ... isize = i_size_read(inode); barrier();

* i_size_read returns READ_ONCE(inode->i_size)

same as current code, ie. actual_end = min_t(..., i_size_read(inode), ...)

All verified on the assembly level that this does not do the double load.