Skip to content

L
linux
Switch branch/tag
  1. 11 Sep, 2017 5 commits
  3. 28 Aug, 2017 17 commits
  5. 27 Aug, 2017 3 commits
    • Linus Torvalds's avatar
      Avoid page waitqueue race leaving possible page locker waiting · a8b169af
      Linus Torvalds authored 
      The "lock_page_killable()" function waits for exclusive access to the
page lock bit using the WQ_FLAG_EXCLUSIVE bit in the waitqueue entry
set.

That means that if it gets woken up, other waiters may have been
skipped.

That, in turn, means that if it sees the page being unlocked, it *must*
take that lock and return success, even if a lethal signal is also
pending.

So instead of checking for lethal signals first, we need to check for
them after we've checked the actual bit that we were waiting for.  Even
if that might then delay the killing of the process.

This matches the order of the old "wait_on_bit_lock()" infrastructure
that the page locking used to use (and is still used in a few other
areas).

Note that if we still return an error after having unsuccessfully tried
to acquire the page lock, that is ok: that means that some other thread
was able to get ahead of us and lock the page, and when that other
thread then unlocks the page, the wakeup event will be repeated.  So any
other pending waiters will now get properly woken up.

Fixes: 62906027 ("mm: add PageWaiters indicating tasks are waiting for a page bit")
Cc: Nick Piggin <npiggin@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Andi Kleen <ak@linux.intel.com>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      a8b169af
    • Linus Torvalds's avatar
      Minor page waitqueue cleanups · 3510ca20
      Linus Torvalds authored 
      Tim Chen and Kan Liang have been battling a customer load that shows
extremely long page wakeup lists.  The cause seems to be constant NUMA
migration of a hot page that is shared across a lot of threads, but the
actual root cause for the exact behavior has not been found.

Tim has a patch that batches the wait list traversal at wakeup time, so
that we at least don't get long uninterruptible cases where we traverse
and wake up thousands of processes and get nasty latency spikes.  That
is likely 4.14 material, but we're still discussing the page waitqueue
specific parts of it.

In the meantime, I've tried to look at making the page wait queues less
expensive, and failing miserably.  If you have thousands of threads
waiting for the same page, it will be painful.  We'll need to try to
figure out the NUMA balancing issue some day, in addition to avoiding
the excessive spinlock hold times.

That said, having tried to rewrite the page wait queues, I can at least
fix up some of the braindamage in the current situation. In particular:

 (a) we don't want to continue walking the page wait list if the bit
     we're waiting for already got set again (which seems to be one of
     the patterns of the bad load).  That makes no progress and just
     causes pointless cache pollution chasing the pointers.

 (b) we don't want to put the non-locking waiters always on the front of
     the queue, and the locking waiters always on the back.  Not only is
     that unfair, it means that we wake up thousands of reading threads
     that will just end up being blocked by the writer later anyway.

Also add a comment about the layout of 'struct wait_page_key' - there is
an external user of it in the cachefiles code that means that it has to
match the layout of 'struct wait_bit_key' in the two first members.  It
so happens to match, because 'struct page *' and 'unsigned long *' end
up having the same values simply because the page flags are the first
member in struct page.

Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Christopher Lameter <cl@linux.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      3510ca20
    • Linus Torvalds's avatar
      Clarify (and fix) MAX_LFS_FILESIZE macros · 0cc3b0ec
      Linus Torvalds authored 
      We have a MAX_LFS_FILESIZE macro that is meant to be filled in by
filesystems (and other IO targets) that know they are 64-bit clean and
don't have any 32-bit limits in their IO path.

It turns out that our 32-bit value for that limit was bogus.  On 32-bit,
the VM layer is limited by the page cache to only 32-bit index values,
but our logic for that was confusing and actually wrong.  We used to
define that value to

	(((loff_t)PAGE_SIZE << (BITS_PER_LONG-1))-1)

which is actually odd in several ways: it limits the index to 31 bits,
and then it limits files so that they can't have data in that last byte
of a page that has the highest 31-bit index (ie page index 0x7fffffff).

Neither of those limitations make sense.  The index is actually the full
32 bit unsigned value, and we can use that whole full page.  So the
maximum size of the file would logically be "PAGE_SIZE << BITS_PER_LONG".

However, we do wan tto avoid the maximum index, because we have code
that iterates over the page indexes, and we don't want that code to
overflow.  So the maximum size of a file on a 32-bit host should
actually be one page less than the full 32-bit index.

So the actual limit is ULONG_MAX << PAGE_SHIFT.  That means that we will
not actually be using the page of that last index (ULONG_MAX), but we
can grow a file up to that limit.

The wrong value of MAX_LFS_FILESIZE actually caused problems for Doug
Nazar, who was still using a 32-bit host, but with a 9.7TB 2 x RAID5
volume.  It turns out that our old MAX_LFS_FILESIZE was 8TiB (well, one
byte less), but the actual true VM limit is one page less than 16TiB.

This was invisible until commit c2a9737f ("vfs,mm: fix a dead loop
in truncate_inode_pages_range()"), which started applying that
MAX_LFS_FILESIZE limit to block devices too.

NOTE! On 64-bit, the page index isn't a limiter at all, and the limit is
actually just the offset type itself (loff_t), which is signed.  But for
clarity, on 64-bit, just use the maximum signed value, and don't make
people have to count the number of 'f' characters in the hex constant.

So just use LLONG_MAX for the 64-bit case.  That was what the value had
been before too, just written out as a hex constant.

Fixes: c2a9737f ("vfs,mm: fix a dead loop in truncate_inode_pages_range()")
Reported-and-tested-by: default avatarDoug Nazar <nazard@nazar.ca>
Cc: Andreas Dilger <adilger@dilger.ca>
Cc: Mark Fasheh <mfasheh@versity.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Dave Kleikamp <shaggy@kernel.org>
Cc: stable@kernel.org
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      0cc3b0ec
  7. 26 Aug, 2017 13 commits
  9. 25 Aug, 2017 2 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7