[lkml]   [2011]   [May]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[PATCHSET v3.1 0/7] data integrity: Stabilize pages during writeback for various fses
Hi all,

This is v3.1 of the stable-page-writes patchset for ext4/3/2, xfs, and fat.
The purpose of this patchset is to prohibit processes from writing on memory
pages that are currently being written to disk because certain storage setups
(e.g. SCSI disks with DIF integrity checksums) will fail a write if the page
contents don't match the checksum. btrfs already guarantees page stability, so
it does not use these changes.

The technique used is fairly simple -- whenever a page is about to become
writable (either because of a write fault to a mapped page, or a buffered write
is in progress), wait for the page writeback flag to be clear, indicating that
the page is not being written to disk. This means that it is necessary (1) to
add wait for writeback code to grab_cache_page_write_begin to take care of
buffered writes, and (2) all filesystems must have a page_mkwrite that locks a
page, waits for writeback, and returns the locked page. For filesystems that
piggyback on the generic block_page_mkwrite, the patchset adds the writeback
wait to that function; for filesystems that do not use the page_mkwrite hook at
all, the patchset provides a stub page_mkwrite.

I ran my write-after-checksum ("wac") reproducer program to try to create the
DIF checksum errors by madly rewriting the same memory pages. In fact, I tried
the following combinations against ext2/3/4, xfs, btrfs, and vfat:

a. 64 write() threads + sync_file_range
b. 64 mmap write threads + msync
c. 32 write() threads + sync_file_range + 32 mmap write threads + msync
d. Same as C, but with all threads in directio mode
e. Same as A, but with all threads in directio mode
f. Same as B, but with all threads in directio mode

After running profiles A-F for 30 minutes each on 6 different machines, ext2,
ext4, xfs, and vfat reported no errors. ext3 still has a lingering failure
case (which I will touch on briefly later) and btrfs eventually reports -ENOSPC
and fails the test, though it does that even without any of the patches applied.

To assess the performance impact of stable page writes, I moved to a disk that
doesn't have DIF support so that I could measure just the impact of waiting for
writeback. I first ran wac with 64 threads madly scribbling on a 64k file and
saw about a 12 percent performance decrease. I then reran the wac program with
64 threads and a 64MB file and saw about the same performance numbers. As I
suspected, the patchset only seems to impact workloads that rewrite the same
memory page frequently.

I am still chasing down what exactly is broken in ext3. data=writeback mode
passes with no failures. data=ordered, however, does not pass; my current
suspicion is that jbd is calling submit_bh on data buffers but doesn't call
page_mkclean to kick the userspace programs off the page before writing it.

Per various comments regarding v3 of this patchset, I've integrated his
suggestions, reworked the patch descriptions to make it clearer which ones
touch all the filesystems and which ones are to fix remaining holes in specific
filesystems, and expanded the scope of filesystems that got fixed.

As always, questions and comments are welcome; and thank you to all the
previous reviewers of this patchset. I am also soliciting people's opinions on
whether or not these patches could go upstream for .40.


 \ /
  Last update: 2011-05-10 01:07    [W:0.138 / U:3.900 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site