diff options
| author | Linus Torvalds <torvalds@linux-foundation.org> | 2010-08-04 01:33:38 +0400 |
|---|---|---|
| committer | Linus Torvalds <torvalds@linux-foundation.org> | 2010-08-04 01:33:38 +0400 |
| commit | 51102ee5b8853d230e534cbcd0d888f0134738a3 (patch) | |
| tree | 5a34844b3040dbc099dd0d3ebc2e6fca0bc65855 /fs/xfs/xfs_filestream.c | |
| parent | 54161df1fb1469d66bce3a3b14d8281adbb69263 (diff) | |
| parent | 6b0a2996a0c023d84bc27ec7528a6e54cb5ea264 (diff) | |
| download | linux-51102ee5b8853d230e534cbcd0d888f0134738a3.tar.xz | |
Merge branch 'for-linus' of git://oss.sgi.com/xfs/xfs
* 'for-linus' of git://oss.sgi.com/xfs/xfs: (49 commits)
xfs simplify and speed up direct I/O completions
xfs: move aio completion after unwritten extent conversion
direct-io: move aio_complete into ->end_io
xfs: fix big endian build
xfs: clean up xfs_bmap_get_bp
xfs: simplify xfs_truncate_file
xfs: kill the b_strat callback in xfs_buf
xfs: remove obsolete osyncisosync mount option
xfs: clean up filestreams helpers
xfs: fix gcc 4.6 set but not read and unused statement warnings
xfs: Fix build when CONFIG_XFS_POSIX_ACL=n
xfs: fix unsigned underflow in xfs_free_eofblocks
xfs: use GFP_NOFS for page cache allocation
xfs: fix memory reclaim recursion deadlock on locked inode buffer
xfs: fix xfs_trans_add_item() lockdep warnings
xfs: simplify and remove xfs_ireclaim
xfs: don't block on buffer read errors
xfs: move inode shrinker unregister even earlier
xfs: remove a dmapi leftover
xfs: writepage always has buffers
...
Diffstat (limited to 'fs/xfs/xfs_filestream.c')
| -rw-r--r-- | fs/xfs/xfs_filestream.c | 84 |
1 files changed, 77 insertions, 7 deletions
diff --git a/fs/xfs/xfs_filestream.c b/fs/xfs/xfs_filestream.c index 390850ee6603..9b715dce5699 100644 --- a/fs/xfs/xfs_filestream.c +++ b/fs/xfs/xfs_filestream.c @@ -18,13 +18,9 @@ #include "xfs.h" #include "xfs_bmap_btree.h" #include "xfs_inum.h" -#include "xfs_dir2.h" -#include "xfs_dir2_sf.h" -#include "xfs_attr_sf.h" #include "xfs_dinode.h" #include "xfs_inode.h" #include "xfs_ag.h" -#include "xfs_dmapi.h" #include "xfs_log.h" #include "xfs_trans.h" #include "xfs_sb.h" @@ -127,6 +123,82 @@ typedef struct fstrm_item xfs_inode_t *pip; /* Parent directory inode pointer. */ } fstrm_item_t; +/* + * Allocation group filestream associations are tracked with per-ag atomic + * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a + * particular AG already has active filestreams associated with it. The mount + * point's m_peraglock is used to protect these counters from per-ag array + * re-allocation during a growfs operation. When xfs_growfs_data_private() is + * about to reallocate the array, it calls xfs_filestream_flush() with the + * m_peraglock held in write mode. + * + * Since xfs_mru_cache_flush() guarantees that all the free functions for all + * the cache elements have finished executing before it returns, it's safe for + * the free functions to use the atomic counters without m_peraglock protection. + * This allows the implementation of xfs_fstrm_free_func() to be agnostic about + * whether it was called with the m_peraglock held in read mode, write mode or + * not held at all. The race condition this addresses is the following: + * + * - The work queue scheduler fires and pulls a filestream directory cache + * element off the LRU end of the cache for deletion, then gets pre-empted. + * - A growfs operation grabs the m_peraglock in write mode, flushes all the + * remaining items from the cache and reallocates the mount point's per-ag + * array, resetting all the counters to zero. + * - The work queue thread resumes and calls the free function for the element + * it started cleaning up earlier. In the process it decrements the + * filestreams counter for an AG that now has no references. + * + * With a shrinkfs feature, the above scenario could panic the system. + * + * All other uses of the following macros should be protected by either the + * m_peraglock held in read mode, or the cache's internal locking exposed by the + * interval between a call to xfs_mru_cache_lookup() and a call to + * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode + * when new elements are added to the cache. + * + * Combined, these locking rules ensure that no associations will ever exist in + * the cache that reference per-ag array elements that have since been + * reallocated. + */ +static int +xfs_filestream_peek_ag( + xfs_mount_t *mp, + xfs_agnumber_t agno) +{ + struct xfs_perag *pag; + int ret; + + pag = xfs_perag_get(mp, agno); + ret = atomic_read(&pag->pagf_fstrms); + xfs_perag_put(pag); + return ret; +} + +static int +xfs_filestream_get_ag( + xfs_mount_t *mp, + xfs_agnumber_t agno) +{ + struct xfs_perag *pag; + int ret; + + pag = xfs_perag_get(mp, agno); + ret = atomic_inc_return(&pag->pagf_fstrms); + xfs_perag_put(pag); + return ret; +} + +static void +xfs_filestream_put_ag( + xfs_mount_t *mp, + xfs_agnumber_t agno) +{ + struct xfs_perag *pag; + + pag = xfs_perag_get(mp, agno); + atomic_dec(&pag->pagf_fstrms); + xfs_perag_put(pag); +} /* * Scan the AGs starting at startag looking for an AG that isn't in use and has @@ -355,16 +427,14 @@ xfs_fstrm_free_func( { fstrm_item_t *item = (fstrm_item_t *)data; xfs_inode_t *ip = item->ip; - int ref; ASSERT(ip->i_ino == ino); xfs_iflags_clear(ip, XFS_IFILESTREAM); /* Drop the reference taken on the AG when the item was added. */ - ref = xfs_filestream_put_ag(ip->i_mount, item->ag); + xfs_filestream_put_ag(ip->i_mount, item->ag); - ASSERT(ref >= 0); TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, xfs_filestream_peek_ag(ip->i_mount, item->ag)); |