lkml.org 
[lkml]   [2013]   [Jan]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH 0/6 RFC] Mapping range lock
Date

Hello,

As I promised in my LSF/MM summit proposal here are initial patches
implementing mapping range lock. There's ext3 converted to fully use the
range locks, converting other filesystems shouldn't be difficult but I want
to spend time on it only after we are sure what we want. The following part
is copied from the LSF/MM proposal, below it are some performance numbers.

There are several different motivations for implementing mapping range
locking:

a) Punch hole is currently racy wrt mmap (page can be faulted in in the
punched range after page cache has been invalidated) leading to nasty
results as fs corruption (we can end up writing to already freed block),
user exposure of uninitialized data, etc. To fix this we need some new
mechanism of serializing hole punching and page faults.

b) There is an uncomfortable number of mechanisms serializing various paths
manipulating pagecache and data underlying it. We have i_mutex, page lock,
checks for page beyond EOF in pagefault code, i_dio_count for direct IO.
Different pairs of operations are serialized by different mechanisms and
not all the cases are covered. Case (a) above is likely the worst but DIO
vs buffered IO isn't ideal either (we provide only limited consistency).
The range locking should somewhat simplify serialization of pagecache
operations. So i_dio_count can be removed completely, i_mutex to certain
extent (we still need something for things like timestamp updates,
possibly for i_size changes although those can be dealt with I think).

c) i_mutex doesn't allow any paralellism of operations using it and some
filesystems workaround this for specific cases (e.g. DIO reads). Using
range locking allows for concurrent operations (e.g. writes, DIO) on
different parts of the file. Of course, range locking itself isn't
enough to make the parallelism possible. Filesystems still have to
somehow deal with the concurrency when manipulating inode allocation
data. But the range locking at least provides a common VFS mechanism for
serialization VFS itself needs and it's upto each filesystem to
serialize more if it needs to.

How it works:
------------

General idea is that range lock for range x-y prevents creation of pages in
that range.

In practice this means:
----------------------

All read paths adding page to page cache and grab_cache_page_write_begin()
first take range lock for the index, then insert locked page, and finally
unlock the range. See below on why buffered IO uses range locks on per-page
basis.

DIO gets range lock at the moment it submits bio for the range covering
pages in the bio. Then pagecache is truncated and bio submitted. Range lock
is unlocked once bio is completed.

Punch hole for range x-y takes range lock for the range before truncating
page cache and the lock is released after filesystem blocks for the range
are freed.

Truncate to size x is equivalent to punch hole for the range x - ~0UL.

The reason why we take the range lock for buffered IO on per-page basis and
for DIO for each bio separately is lock ordering with mmap_sem. Page faults
need to instantiate page under mmap_sem. That establishes mmap_sem > range
lock. Buffered IO takes mmap_sem when prefaulting pages so we cannot hold
range lock at that moment. Similarly get_user_pages() in DIO code takes
mmap_sem so we have be sure not to hold range lock when calling that.

How much does it cost:
---------------------

There's a memory cost - an extra pointer and spinlock in struct
address_space, 64 bytes on stack for buffered IO, truncate, punch hole, and
dynamically allocated 72-byte structure per each BIO submitted by direct IO.

And there's a cpu cost. I measured it on an 8 CPU machine with 4 GB of memory
with ext2 (yes, I added support also for ext2 and used it for measurements as
especially write results are much less noisy) over 1G ramdisk. The workloads
were generated by FIO and were 1) read 800 MB file, 2) overwrite 800 MB file,
3) mmap read 800 MB file. Each test was run 30 times.

The results are here (times to complete in ms):
Vanilla Range Locks
Avg Stddev Avg Stddev
READ 1133.566667 11.954590 1137.06666 7.827019
WRITE 1069.300000 7.996458 1101.200000 8.607748
MMAP 1416.733333 28.459250 1421.900000 30.636960

So we see READ and MMAP time changes are in the noise (although for reads
there seem to be about 1% cost if I compare more tests), for WRITE the cost
barely stands out of the noise at ~3% (and here I verified with perf what's
going on and indeed the range_lock() and range_unlock() calls cost in total
close to 3% of CPU time).

So the cost is noticeable. Is it a problem? Maybe, not sure... We could
likely optimize the lock-single-page range case but I wanted to start
simple and get some feedback first.

Honza


\
 
 \ /
  Last update: 2013-01-31 23:42    [W:0.264 / U:0.324 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site