Switch KM_SLEEP to KM_PUSHPAGE
authorRichard Yao <ryao@cs.stonybrook.edu>
Mon, 7 May 2012 17:49:51 +0000 (13:49 -0400)
committerBrian Behlendorf <behlendorf1@llnl.gov>
Mon, 27 Aug 2012 19:01:37 +0000 (12:01 -0700)
commitb8d06fca089fae4680c3a552fc55c512bfb02202
treedfb5f3d20c5f417110359d39e8af6e8fecb1fcf3
parent991fc1d7ae2589c01a939a9cbd0e866c90fdd03b
Switch KM_SLEEP to KM_PUSHPAGE

Differences between how paging is done on Solaris and Linux can cause
deadlocks if KM_SLEEP is used in any the following contexts.

  * The txg_sync thread
  * The zvol write/discard threads
  * The zpl_putpage() VFS callback

This is because KM_SLEEP will allow for direct reclaim which may result
in the VM calling back in to the filesystem or block layer to write out
pages.  If a lock is held over this operation the potential exists to
deadlock the system.  To ensure forward progress all memory allocations
in these contexts must us KM_PUSHPAGE which disables performing any I/O
to accomplish the memory allocation.

Previously, this behavior was acheived by setting PF_MEMALLOC on the
thread.  However, that resulted in unexpected side effects such as the
exhaustion of pages in ZONE_DMA.  This approach touchs more of the zfs
code, but it is more consistent with the right way to handle these cases
under Linux.

This is patch lays the ground work for being able to safely revert the
following commits which used PF_MEMALLOC:

  21ade34 Disable direct reclaim for z_wr_* threads
  cfc9a5c Fix zpl_writepage() deadlock
  eec8164 Fix ASSERTION(!dsl_pool_sync_context(tx->tx_pool))

Signed-off-by: Richard Yao <ryao@cs.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #726
45 files changed:
include/sys/dbuf.h
include/sys/dsl_dataset.h
include/sys/dsl_dir.h
include/sys/spa.h
module/zcommon/zprop_common.c
module/zfs/arc.c
module/zfs/bplist.c
module/zfs/dbuf.c
module/zfs/ddt.c
module/zfs/dmu.c
module/zfs/dmu_objset.c
module/zfs/dmu_traverse.c
module/zfs/dmu_tx.c
module/zfs/dmu_zfetch.c
module/zfs/dnode.c
module/zfs/dsl_dataset.c
module/zfs/dsl_deadlist.c
module/zfs/dsl_dir.c
module/zfs/dsl_prop.c
module/zfs/lzjb.c
module/zfs/metaslab.c
module/zfs/refcount.c
module/zfs/sa.c
module/zfs/spa.c
module/zfs/spa_config.c
module/zfs/spa_history.c
module/zfs/spa_misc.c
module/zfs/space_map.c
module/zfs/txg.c
module/zfs/unique.c
module/zfs/vdev.c
module/zfs/vdev_cache.c
module/zfs/vdev_disk.c
module/zfs/vdev_file.c
module/zfs/vdev_label.c
module/zfs/vdev_mirror.c
module/zfs/vdev_raidz.c
module/zfs/zap.c
module/zfs/zap_micro.c
module/zfs/zfs_acl.c
module/zfs/zfs_fm.c
module/zfs/zfs_rlock.c
module/zfs/zfs_znode.c
module/zfs/zil.c
module/zfs/zvol.c