openwrt/staging/blogic.git
8 years agostm class: dummy_stm: Make nr_dummies parameter read-only
Alexander Shishkin [Fri, 4 Mar 2016 14:22:33 +0000 (16:22 +0200)]
stm class: dummy_stm: Make nr_dummies parameter read-only

Changing nr_dummies after the module has been loaded doesn't actually
change anything, so just make it read-only.

Reported-by: Alan Cox <alan.cox@intel.com>
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reviewed-by: Laurent Fert <laurent.fert@intel.com>
8 years agostm class: Fix integer boundary checks for master range
Chunyan Zhang [Mon, 28 Mar 2016 07:55:42 +0000 (15:55 +0800)]
stm class: Fix integer boundary checks for master range

Master IDs are of unsigned int type, yet in the configfs policy code
we're validating user's input against INT_MAX. This is both pointless
and misleading as the real limits are imposed by the stm device's
[sw_start..sw_end] (which are also limited by the spec to be no larger
than 2^16-1).

Clean this up by getting rid of the redundant comparisons.

Signed-off-by: Chunyan Zhang <zhang.chunyan@linaro.org>
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reviewed-by: Laurent Fert <laurent.fert@intel.com>
8 years agoMAINTAINERS: Add a git tree for the stm class
Alexander Shishkin [Thu, 31 Mar 2016 13:30:15 +0000 (16:30 +0300)]
MAINTAINERS: Add a git tree for the stm class

So that people know where their patches go.

Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reviewed-by: Laurent Fert <laurent.fert@intel.com>
8 years agoLinux 4.6-rc1
Linus Torvalds [Sat, 26 Mar 2016 23:03:24 +0000 (16:03 -0700)]
Linux 4.6-rc1

8 years agoMerge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph...
Linus Torvalds [Sat, 26 Mar 2016 22:53:16 +0000 (15:53 -0700)]
Merge branch 'for-linus' of git://git./linux/kernel/git/sage/ceph-client

Pull Ceph updates from Sage Weil:
 "There is quite a bit here, including some overdue refactoring and
  cleanup on the mon_client and osd_client code from Ilya, scattered
  writeback support for CephFS and a pile of bug fixes from Zheng, and a
  few random cleanups and fixes from others"

[ I already decided not to pull this because of it having been rebased
  recently, but ended up changing my mind after all.  Next time I'll
  really hold people to it.  Oh well.   - Linus ]

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (34 commits)
  libceph: use KMEM_CACHE macro
  ceph: use kmem_cache_zalloc
  rbd: use KMEM_CACHE macro
  ceph: use lookup request to revalidate dentry
  ceph: kill ceph_get_dentry_parent_inode()
  ceph: fix security xattr deadlock
  ceph: don't request vxattrs from MDS
  ceph: fix mounting same fs multiple times
  ceph: remove unnecessary NULL check
  ceph: avoid updating directory inode's i_size accidentally
  ceph: fix race during filling readdir cache
  libceph: use sizeof_footer() more
  ceph: kill ceph_empty_snapc
  ceph: fix a wrong comparison
  ceph: replace CURRENT_TIME by current_fs_time()
  ceph: scattered page writeback
  libceph: add helper that duplicates last extent operation
  libceph: enable large, variable-sized OSD requests
  libceph: osdc->req_mempool should be backed by a slab pool
  libceph: make r_request msg_size calculation clearer
  ...

8 years agoMerge tag 'ofs-pull-tag-1' of git://git.kernel.org/pub/scm/linux/kernel/git/hubcap...
Linus Torvalds [Sat, 26 Mar 2016 19:59:04 +0000 (12:59 -0700)]
Merge tag 'ofs-pull-tag-1' of git://git./linux/kernel/git/hubcap/linux

Pull orangefs filesystem from Mike Marshall.

This finally merges the long-pending orangefs filesystem, which has been
much cleaned up with input from Al Viro over the last six months.  From
the documentation file:

 "OrangeFS is an LGPL userspace scale-out parallel storage system.  It
  is ideal for large storage problems faced by HPC, BigData, Streaming
  Video, Genomics, Bioinformatics.

  Orangefs, originally called PVFS, was first developed in 1993 by Walt
  Ligon and Eric Blumer as a parallel file system for Parallel Virtual
  Machine (PVM) as part of a NASA grant to study the I/O patterns of
  parallel programs.

  Orangefs features include:

    - Distributes file data among multiple file servers
    - Supports simultaneous access by multiple clients
    - Stores file data and metadata on servers using local file system
      and access methods
    - Userspace implementation is easy to install and maintain
    - Direct MPI support
    - Stateless"

see Documentation/filesystems/orangefs.txt for more in-depth details.

* tag 'ofs-pull-tag-1' of git://git.kernel.org/pub/scm/linux/kernel/git/hubcap/linux: (174 commits)
  orangefs: fix orangefs_superblock locking
  orangefs: fix do_readv_writev() handling of error halfway through
  orangefs: have ->kill_sb() evict the VFS side of things first
  orangefs: sanitize ->llseek()
  orangefs-bufmap.h: trim unused junk
  orangefs: saner calling conventions for getting a slot
  orangefs_copy_{to,from}_bufmap(): don't pass bufmap pointer
  orangefs: get rid of readdir_handle_s
  ornagefs: ensure that truncate has an up to date inode size
  orangefs: move code which sets i_link to orangefs_inode_getattr
  orangefs: remove needless wrapper around GFP_KERNEL
  orangefs: remove wrapper around mutex_lock(&inode->i_mutex)
  orangefs: refactor inode type or link_target change detection
  orangefs: use new getattr for revalidate and remove old getattr
  orangefs: use new getattr in inode getattr and permission
  orangefs: use new orangefs_inode_getattr to get size in write and llseek
  orangefs: use new orangefs_inode_getattr to create new inodes
  orangefs: rename orangefs_inode_getattr to orangefs_inode_old_getattr
  orangefs: remove inode->i_lock wrapper
  orangefs: put register_chrdev immediately before register_filesystem
  ...

8 years agoMerge tag 'ntb-4.6' of git://github.com/jonmason/ntb
Linus Torvalds [Sat, 26 Mar 2016 18:37:42 +0000 (11:37 -0700)]
Merge tag 'ntb-4.6' of git://github.com/jonmason/ntb

Pull NTB bug fixes from Jon Mason:
 "NTB bug fixes for tasklet from spinning forever, link errors,
  translation window setup, NULL ptr dereference, and ntb-perf errors.

  Also, a modification to the driver API that makes _addr functions
  optional"

* tag 'ntb-4.6' of git://github.com/jonmason/ntb:
  NTB: Remove _addr functions from ntb_hw_amd
  NTB: Make _addr functions optional in the API
  NTB: Fix incorrect clean up routine in ntb_perf
  NTB: Fix incorrect return check in ntb_perf
  ntb: fix possible NULL dereference
  ntb: add missing setup of translation window
  ntb: stop link work when we do not have memory
  ntb: stop tasklet from spinning forever during shutdown.
  ntb: perf test: fix address space confusion

8 years agoMerge tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
Linus Torvalds [Sat, 26 Mar 2016 18:31:01 +0000 (11:31 -0700)]
Merge tag 'scsi-misc' of git://git./linux/kernel/git/jejb/scsi

Pull more SCSI updates from James Bottomley:
 "The only new stuff which missed the first pull request is an update to
  the UFS driver.

  The rest is an assortment of bug fixes and minor tweaks which appeared
  recently (some are fixes for recent code and some are stuff spotted
  recently by the checkers or the new gcc-6 compiler [most of Arnd's
  stuff])"

* tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi: (32 commits)
  scsi_common: do not clobber fixed sense information
  scsi: ufs: select CONFIG_NLS
  scsi: fc: use get/put_unaligned64 for wwn access
  fnic: move printk()s outside of the critical code section.
  qla2xxx: avoid maybe_uninitialized warning
  megaraid_sas: add missing curly braces in ioctl handler
  lpfc: fix misleading indentation
  scsi_transport_sas: add 'scsi_target_id' sysfs attribute
  scsi_dh_alua: uninitialized variable in alua_check_vpd()
  scsi: ufs-qcom: add printouts of testbus debug registers
  scsi: ufs-qcom: enable/disable the device ref clock
  scsi: ufs-qcom: set PA_Local_TX_LCC_Enable before link startup
  scsi: ufs: add device quirk delay before putting UFS rails in LPM
  scsi: ufs: fix leakage during link off state
  scsi: ufs: tune UniPro parameters to optimize hibern8 exit time
  scsi: ufs: handle non spec compliant bkops behaviour by device
  scsi: ufs: add retry for query descriptors
  scsi: ufs: add error recovery after DL NAC error
  scsi: ufs: make error handling bit faster
  scsi: ufs: disable vccq if it's not needed by UFS device
  ...

8 years agof2fs/crypto: fix xts_tweak initialization
Linus Torvalds [Sat, 26 Mar 2016 17:13:05 +0000 (10:13 -0700)]
f2fs/crypto: fix xts_tweak initialization

Commit 0b81d07790726 ("fs crypto: move per-file encryption from f2fs
tree to fs/crypto") moved the f2fs crypto files to fs/crypto/ and
renamed the symbol prefixes from "f2fs_" to "fscrypt_" (and from "F2FS_"
to just "FS" for preprocessor symbols).

Because of the symbol renaming, it's a bit hard to see it as a file
move: use

    git show -M30 0b81d07790726

to lower the rename detection to just 30% similarity and make git show
the files as renamed (the header file won't be shown as a rename even
then - since all it contains is symbol definitions, it looks almost
completely different).

Even with the renames showing as renames, the diffs are not all that
easy to read, since so much is just the renames.  But Eric Biggers
noticed that it's not just all renames: the initialization of the
xts_tweak had been broken too, using the inode number rather than the
page offset.

That's not right - it makes the xfs_tweak the same for all pages of each
inode.  It _might_ make sense to make the xfs_tweak contain both the
offset _and_ the inode number, but not just the inode number.

Reported-by: Eric Biggers <ebiggers3@gmail.com>
Cc: Jaegeuk Kim <jaegeuk@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoNTB: Remove _addr functions from ntb_hw_amd
Allen Hubbe [Mon, 21 Mar 2016 08:53:14 +0000 (04:53 -0400)]
NTB: Remove _addr functions from ntb_hw_amd

Kernel zero day testing warned about address space confusion.  A virtual
iomem address was used where a physical address is expected.  The
offending functions implement an optional part of the api, so they are
removed.  They can be added later, after testing.

Fixes: a1b3695820aa490e58915d720a1438069813008b
Signed-off-by: Allen Hubbe <Allen.Hubbe@emc.com>
Acked-by: Xiangliang Yu <Xiangliang.Yu@amd.com>
Signed-off-by: Jon Mason <jdmason@kudzu.us>
8 years agoorangefs: fix orangefs_superblock locking
Al Viro [Fri, 25 Mar 2016 23:56:34 +0000 (19:56 -0400)]
orangefs: fix orangefs_superblock locking

* switch orangefs_remount() to taking ORANGEFS_SB(sb) instead of sb
* remove from the list _before_ orangefs_unmount() - request_mutex
in the latter will make sure that nothing observed in the loop in
ORANGEFS_DEV_REMOUNT_ALL handling will get freed until the end
of loop
* on removal, keep the forward pointer and zero the back one.  That
way we can drop and regain the spinlock in the loop body (again,
ORANGEFS_DEV_REMOUNT_ALL one) and still be able to get to the
rest of the list.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs: fix do_readv_writev() handling of error halfway through
Al Viro [Wed, 17 Feb 2016 01:15:43 +0000 (20:15 -0500)]
orangefs: fix do_readv_writev() handling of error halfway through

Error should only be returned if nothing had been read/written.
Otherwise we need to report a short read/write instead.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs: have ->kill_sb() evict the VFS side of things first
Al Viro [Wed, 17 Feb 2016 02:08:29 +0000 (21:08 -0500)]
orangefs: have ->kill_sb() evict the VFS side of things first

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs: sanitize ->llseek()
Al Viro [Wed, 17 Feb 2016 01:25:19 +0000 (20:25 -0500)]
orangefs: sanitize ->llseek()

a) open files can't have NULL inodes
b) it's SEEK_END, not ORANGEFS_SEEK_END; no need to get cute.
c) make_bad_inode() on lseek()?

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs-bufmap.h: trim unused junk
Al Viro [Wed, 17 Feb 2016 01:12:04 +0000 (20:12 -0500)]
orangefs-bufmap.h: trim unused junk

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs: saner calling conventions for getting a slot
Al Viro [Wed, 17 Feb 2016 01:10:26 +0000 (20:10 -0500)]
orangefs: saner calling conventions for getting a slot

just have it return the slot number or -E... - the caller checks
the sign anyway

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs_copy_{to,from}_bufmap(): don't pass bufmap pointer
Al Viro [Wed, 17 Feb 2016 01:06:19 +0000 (20:06 -0500)]
orangefs_copy_{to,from}_bufmap(): don't pass bufmap pointer

it's always __orangefs_bufmap

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoorangefs: get rid of readdir_handle_s
Al Viro [Wed, 17 Feb 2016 00:54:13 +0000 (19:54 -0500)]
orangefs: get rid of readdir_handle_s

no point, really - we couldn't keep those across the calls of
getdents(); it would be too easy to DoS, having all slots exhausted.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
8 years agoMerge branch 'akpm' (patches from Andrew)
Linus Torvalds [Fri, 25 Mar 2016 23:59:11 +0000 (16:59 -0700)]
Merge branch 'akpm' (patches from Andrew)

Merge fourth patch-bomb from Andrew Morton:
 "A lot more stuff than expected, sorry.  A bunch of ocfs2 reviewing was
  finished off.

   - mhocko's oom-reaper out-of-memory-handler changes

   - ocfs2 fixes and features

   - KASAN feature work

   - various fixes"

* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (42 commits)
  thp: fix typo in khugepaged_scan_pmd()
  MAINTAINERS: fill entries for KASAN
  mm/filemap: generic_file_read_iter(): check for zero reads unconditionally
  kasan: test fix: warn if the UAF could not be detected in kmalloc_uaf2
  mm, kasan: stackdepot implementation. Enable stackdepot for SLAB
  arch, ftrace: for KASAN put hard/soft IRQ entries into separate sections
  mm, kasan: add GFP flags to KASAN API
  mm, kasan: SLAB support
  kasan: modify kmalloc_large_oob_right(), add kmalloc_pagealloc_oob_right()
  include/linux/oom.h: remove undefined oom_kills_count()/note_oom_kill()
  mm/page_alloc: prevent merging between isolated and other pageblocks
  drivers/memstick/host/r592.c: avoid gcc-6 warning
  ocfs2: extend enough credits for freeing one truncate record while replaying truncate records
  ocfs2: extend transaction for ocfs2_remove_rightmost_path() and ocfs2_update_edge_lengths() before to avoid inconsistency between inode and et
  ocfs2/dlm: move lock to the tail of grant queue while doing in-place convert
  ocfs2: solve a problem of crossing the boundary in updating backups
  ocfs2: fix occurring deadlock by changing ocfs2_wq from global to local
  ocfs2/dlm: fix BUG in dlm_move_lockres_to_recovery_list
  ocfs2/dlm: fix race between convert and recovery
  ocfs2: fix a deadlock issue in ocfs2_dio_end_io_write()
  ...

8 years agoMerge tag 'pm+acpi-4.6-rc1-3' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael...
Linus Torvalds [Fri, 25 Mar 2016 23:55:37 +0000 (16:55 -0700)]
Merge tag 'pm+acpi-4.6-rc1-3' of git://git./linux/kernel/git/rafael/linux-pm

Pull power management fixlet from Rafael Wysocki:
 "One of commits in my previous pull request changed the permissions of
  drivers/power/avs/rockchip-io-domain.c to executable by mistake"

* tag 'pm+acpi-4.6-rc1-3' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
  Fix permissions of drivers/power/avs/rockchip-io-domain.c

8 years agoMerge tag 'please-pull-preadv2' of git://git.kernel.org/pub/scm/linux/kernel/git...
Linus Torvalds [Fri, 25 Mar 2016 23:48:45 +0000 (16:48 -0700)]
Merge tag 'please-pull-preadv2' of git://git./linux/kernel/git/aegl/linux

Pull ia64 update from Tony Luck:
 "Wire up new system calls p{read,write}v2 for ia64"

* tag 'please-pull-preadv2' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux:
  [IA64] Enable preadv2 and pwritev2 syscalls for ia64

8 years agoMerge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input
Linus Torvalds [Fri, 25 Mar 2016 23:39:05 +0000 (16:39 -0700)]
Merge branch 'for-linus' of git://git./linux/kernel/git/dtor/input

Pull more input updates from Dmitry Torokhov:
 "Second round of updates for the input subsystem.

  The BYD PS/2 protocol driver now uses absolute reporting mode and
  should behave more like other touchpads; Synaptics driver needed to
  extend one of its quirks to a newer firmware version, and a few USB
  drivers got tightened up checks for the contents of their descriptors"

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input:
  Input: sur40 - fix DMA on stack
  Input: ati_remote2 - fix crashes on detecting device with invalid descriptor
  Input: synaptics - handle spurious release of trackstick buttons, again
  Input: synaptics-rmi4 - remove check of Non-NULL array
  Input: byd - enable absolute mode
  Input: ims-pcu - sanity check against missing interfaces
  Input: melfas_mip4 - add hw_version sysfs attribute

8 years agothp: fix typo in khugepaged_scan_pmd()
Kirill A. Shutemov [Fri, 25 Mar 2016 21:22:20 +0000 (14:22 -0700)]
thp: fix typo in khugepaged_scan_pmd()

!PageLRU should lead to SCAN_PAGE_LRU, not SCAN_SCAN_ABORT result.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoMAINTAINERS: fill entries for KASAN
Andrey Ryabinin [Fri, 25 Mar 2016 21:22:17 +0000 (14:22 -0700)]
MAINTAINERS: fill entries for KASAN

Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Acked-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/filemap: generic_file_read_iter(): check for zero reads unconditionally
Nicolai Stange [Fri, 25 Mar 2016 21:22:14 +0000 (14:22 -0700)]
mm/filemap: generic_file_read_iter(): check for zero reads unconditionally

If
 - generic_file_read_iter() gets called with a zero read length,
 - the read offset is at a page boundary,
 - IOCB_DIRECT is not set
-  and the page in question hasn't made it into the page cache yet,
then do_generic_file_read() will trigger a readahead with a req_size hint
of zero.

Since roundup_pow_of_two(0) is undefined, UBSAN reports

  UBSAN: Undefined behaviour in include/linux/log2.h:63:13
  shift exponent 64 is too large for 64-bit type 'long unsigned int'
  CPU: 3 PID: 1017 Comm: sa1 Tainted: G L 4.5.0-next-20160318+ #14
  [...]
  Call Trace:
   [...]
   [<ffffffff813ef61a>] ondemand_readahead+0x3aa/0x3d0
   [<ffffffff813ef61a>] ? ondemand_readahead+0x3aa/0x3d0
   [<ffffffff813c73bd>] ? find_get_entry+0x2d/0x210
   [<ffffffff813ef9c3>] page_cache_sync_readahead+0x63/0xa0
   [<ffffffff813cc04d>] do_generic_file_read+0x80d/0xf90
   [<ffffffff813cc955>] generic_file_read_iter+0x185/0x420
   [...]
   [<ffffffff81510b06>] __vfs_read+0x256/0x3d0
   [...]

when get_init_ra_size() gets called from ondemand_readahead().

The net effect is that the initial readahead size is arch dependent for
requested read lengths of zero: for example, since

  1UL << (sizeof(unsigned long) * 8)

evaluates to 1 on x86 while its result is 0 on ARMv7, the initial readahead
size becomes 4 on the former and 0 on the latter.

What's more, whether or not the file access timestamp is updated for zero
length reads is decided differently for the two cases of IOCB_DIRECT
being set or cleared: in the first case, generic_file_read_iter()
explicitly skips updating that timestamp while in the latter case, it is
always updated through the call to do_generic_file_read().

According to POSIX, zero length reads "do not modify the last data access
timestamp" and thus, the IOCB_DIRECT behaviour is POSIXly correct.

Let generic_file_read_iter() unconditionally check the requested read
length at its entry and return immediately with success if it is zero.

Signed-off-by: Nicolai Stange <nicstange@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agokasan: test fix: warn if the UAF could not be detected in kmalloc_uaf2
Alexander Potapenko [Fri, 25 Mar 2016 21:22:11 +0000 (14:22 -0700)]
kasan: test fix: warn if the UAF could not be detected in kmalloc_uaf2

Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Andrey Konovalov <adech.fo@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, kasan: stackdepot implementation. Enable stackdepot for SLAB
Alexander Potapenko [Fri, 25 Mar 2016 21:22:08 +0000 (14:22 -0700)]
mm, kasan: stackdepot implementation. Enable stackdepot for SLAB

Implement the stack depot and provide CONFIG_STACKDEPOT.  Stack depot
will allow KASAN store allocation/deallocation stack traces for memory
chunks.  The stack traces are stored in a hash table and referenced by
handles which reside in the kasan_alloc_meta and kasan_free_meta
structures in the allocated memory chunks.

IRQ stack traces are cut below the IRQ entry point to avoid unnecessary
duplication.

Right now stackdepot support is only enabled in SLAB allocator.  Once
KASAN features in SLAB are on par with those in SLUB we can switch SLUB
to stackdepot as well, thus removing the dependency on SLUB stack
bookkeeping, which wastes a lot of memory.

This patch is based on the "mm: kasan: stack depots" patch originally
prepared by Dmitry Chernenkov.

Joonsoo has said that he plans to reuse the stackdepot code for the
mm/page_owner.c debugging facility.

[akpm@linux-foundation.org: s/depot_stack_handle/depot_stack_handle_t]
[aryabinin@virtuozzo.com: comment style fixes]
Signed-off-by: Alexander Potapenko <glider@google.com>
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Andrey Konovalov <adech.fo@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoarch, ftrace: for KASAN put hard/soft IRQ entries into separate sections
Alexander Potapenko [Fri, 25 Mar 2016 21:22:05 +0000 (14:22 -0700)]
arch, ftrace: for KASAN put hard/soft IRQ entries into separate sections

KASAN needs to know whether the allocation happens in an IRQ handler.
This lets us strip everything below the IRQ entry point to reduce the
number of unique stack traces needed to be stored.

Move the definition of __irq_entry to <linux/interrupt.h> so that the
users don't need to pull in <linux/ftrace.h>.  Also introduce the
__softirq_entry macro which is similar to __irq_entry, but puts the
corresponding functions to the .softirqentry.text section.

Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Andrey Konovalov <adech.fo@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, kasan: add GFP flags to KASAN API
Alexander Potapenko [Fri, 25 Mar 2016 21:22:02 +0000 (14:22 -0700)]
mm, kasan: add GFP flags to KASAN API

Add GFP flags to KASAN hooks for future patches to use.

This patch is based on the "mm: kasan: unified support for SLUB and SLAB
allocators" patch originally prepared by Dmitry Chernenkov.

Signed-off-by: Alexander Potapenko <glider@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Andrey Konovalov <adech.fo@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, kasan: SLAB support
Alexander Potapenko [Fri, 25 Mar 2016 21:21:59 +0000 (14:21 -0700)]
mm, kasan: SLAB support

Add KASAN hooks to SLAB allocator.

This patch is based on the "mm: kasan: unified support for SLUB and SLAB
allocators" patch originally prepared by Dmitry Chernenkov.

Signed-off-by: Alexander Potapenko <glider@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Andrey Konovalov <adech.fo@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agokasan: modify kmalloc_large_oob_right(), add kmalloc_pagealloc_oob_right()
Alexander Potapenko [Fri, 25 Mar 2016 21:21:56 +0000 (14:21 -0700)]
kasan: modify kmalloc_large_oob_right(), add kmalloc_pagealloc_oob_right()

This patchset implements SLAB support for KASAN

Unlike SLUB, SLAB doesn't store allocation/deallocation stacks for heap
objects, therefore we reimplement this feature in mm/kasan/stackdepot.c.
The intention is to ultimately switch SLUB to use this implementation as
well, which will save a lot of memory (right now SLUB bloats each object
by 256 bytes to store the allocation/deallocation stacks).

Also neither SLUB nor SLAB delay the reuse of freed memory chunks, which
is necessary for better detection of use-after-free errors.  We
introduce memory quarantine (mm/kasan/quarantine.c), which allows
delayed reuse of deallocated memory.

This patch (of 7):

Rename kmalloc_large_oob_right() to kmalloc_pagealloc_oob_right(), as
the test only checks the page allocator functionality.  Also reimplement
kmalloc_large_oob_right() so that the test allocates a large enough
chunk of memory that still does not trigger the page allocator fallback.

Signed-off-by: Alexander Potapenko <glider@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Andrey Konovalov <adech.fo@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoinclude/linux/oom.h: remove undefined oom_kills_count()/note_oom_kill()
Tetsuo Handa [Fri, 25 Mar 2016 21:21:53 +0000 (14:21 -0700)]
include/linux/oom.h: remove undefined oom_kills_count()/note_oom_kill()

A leftover from commit c32b3cbe0d06 ("oom, PM: make OOM detection in the
freezer path raceless").

Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/page_alloc: prevent merging between isolated and other pageblocks
Vlastimil Babka [Fri, 25 Mar 2016 21:21:50 +0000 (14:21 -0700)]
mm/page_alloc: prevent merging between isolated and other pageblocks

Hanjun Guo has reported that a CMA stress test causes broken accounting of
CMA and free pages:

> Before the test, I got:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal:         204800 kB
> CmaFree:          195044 kB
>
>
> After running the test:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal:         204800 kB
> CmaFree:         6602584 kB
>
> So the freed CMA memory is more than total..
>
> Also the the MemFree is more than mem total:
>
> -bash-4.3# cat /proc/meminfo
> MemTotal:       16342016 kB
> MemFree:        22367268 kB
> MemAvailable:   22370528 kB

Laura Abbott has confirmed the issue and suspected the freepage accounting
rewrite around 3.18/4.0 by Joonsoo Kim.  Joonsoo had a theory that this is
caused by unexpected merging between MIGRATE_ISOLATE and MIGRATE_CMA
pageblocks:

> CMA isolates MAX_ORDER aligned blocks, but, during the process,
> partialy isolated block exists. If MAX_ORDER is 11 and
> pageblock_order is 9, two pageblocks make up MAX_ORDER
> aligned block and I can think following scenario because pageblock
> (un)isolation would be done one by one.
>
> (each character means one pageblock. 'C', 'I' means MIGRATE_CMA,
> MIGRATE_ISOLATE, respectively.
>
> CC -> IC -> II (Isolation)
> II -> CI -> CC (Un-isolation)
>
> If some pages are freed at this intermediate state such as IC or CI,
> that page could be merged to the other page that is resident on
> different type of pageblock and it will cause wrong freepage count.

This was supposed to be prevented by CMA operating on MAX_ORDER blocks,
but since it doesn't hold the zone->lock between pageblocks, a race
window does exist.

It's also likely that unexpected merging can occur between
MIGRATE_ISOLATE and non-CMA pageblocks.  This should be prevented in
__free_one_page() since commit 3c605096d315 ("mm/page_alloc: restrict
max order of merging on isolated pageblock").  However, we only check
the migratetype of the pageblock where buddy merging has been initiated,
not the migratetype of the buddy pageblock (or group of pageblocks)
which can be MIGRATE_ISOLATE.

Joonsoo has suggested checking for buddy migratetype as part of
page_is_buddy(), but that would add extra checks in allocator hotpath
and bloat-o-meter has shown significant code bloat (the function is
inline).

This patch reduces the bloat at some expense of more complicated code.
The buddy-merging while-loop in __free_one_page() is initially bounded
to pageblock_border and without any migratetype checks.  The checks are
placed outside, bumping the max_order if merging is allowed, and
returning to the while-loop with a statement which can't be possibly
considered harmful.

This fixes the accounting bug and also removes the arguably weird state
in the original commit 3c605096d315 where buddies could be left
unmerged.

Fixes: 3c605096d315 ("mm/page_alloc: restrict max order of merging on isolated pageblock")
Link: https://lkml.org/lkml/2016/3/2/280
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Hanjun Guo <guohanjun@huawei.com>
Tested-by: Hanjun Guo <guohanjun@huawei.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Debugged-by: Laura Abbott <labbott@redhat.com>
Debugged-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: <stable@vger.kernel.org> [3.18+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agodrivers/memstick/host/r592.c: avoid gcc-6 warning
Arnd Bergmann [Fri, 25 Mar 2016 21:21:47 +0000 (14:21 -0700)]
drivers/memstick/host/r592.c: avoid gcc-6 warning

The r592 driver relies on behavior of the DMA mapping API that is
normally observed but not guaranteed by the API.  Instead it uses a
runtime check to fail transfers if the API ever behaves

When CONFIG_NEED_SG_DMA_LENGTH is not set, one of the checks turns into a
comparison of a variable with itself, which gcc-6.0 now warns about:

drivers/memstick/host/r592.c: In function 'r592_transfer_fifo_dma':
drivers/memstick/host/r592.c:302:31: error: self-comparison always evaluates to false [-Werror=tautological-compare]
    (sg_dma_len(&dev->req->sg) < dev->req->sg.length)) {
                               ^

The check itself is not a problem, so this patch just rephrases the
condition in a way that gcc does not consider an indication of a mistake.
We already know that dev->req->sg.length was initially R592_LFIFO_SIZE, so
we can compare it to that constant again.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Maxim Levitsky <maximlevitsky@gmail.com>
Cc: Quentin Lambert <lambert.quentin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: extend enough credits for freeing one truncate record while replaying truncate...
Xue jiufei [Fri, 25 Mar 2016 21:21:44 +0000 (14:21 -0700)]
ocfs2: extend enough credits for freeing one truncate record while replaying truncate records

Now function ocfs2_replay_truncate_records() first modifies tl_used,
then calls ocfs2_extend_trans() to extend transactions for gd and alloc
inode used for freeing clusters.  jbd2_journal_restart() may be called
and it may happen that tl_used in truncate log is decreased but the
clusters are not freed, which means these clusters are lost.  So we
should avoid extending transactions in these two operations.

Signed-off-by: joyce.xue <xuejiufei@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Acked-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: extend transaction for ocfs2_remove_rightmost_path() and ocfs2_update_edge_len...
Xue jiufei [Fri, 25 Mar 2016 21:21:41 +0000 (14:21 -0700)]
ocfs2: extend transaction for ocfs2_remove_rightmost_path() and ocfs2_update_edge_lengths() before to avoid inconsistency between inode and et

I found that jbd2_journal_restart() is called in some places without
keeping things consistently before.  However, jbd2_journal_restart() may
commit the handle's transaction and restart another one.  If the first
transaction is committed successfully while another not, it may cause
filesystem inconsistency or read only.  This is an effort to fix this
kind of problems.

This patch (of 3):

The following functions will be called while truncating an extent:
ocfs2_remove_btree_range
  -> ocfs2_start_trans
  -> ocfs2_remove_extent
     -> ocfs2_truncate_rec
       -> ocfs2_extend_rotate_transaction
         -> jbd2_journal_restart if jbd2_journal_extend fail
       -> ocfs2_rotate_tree_left
         -> ocfs2_remove_rightmost_path
             -> ocfs2_extend_rotate_transaction
               -> ocfs2_unlink_subtree
                -> ocfs2_update_edge_lengths
                  -> ocfs2_extend_trans
                    -> jbd2_journal_restart if jbd2_journal_extend fail
  -> ocfs2_et_update_clusters
  -> ocfs2_commit_trans

jbd2_journal_restart() may be called and it may happened that the buffers
dirtied in ocfs2_truncate_rec() are committed while buffers dirtied in
ocfs2_et_update_clusters() are not, the total clusters on extent tree and
i_clusters in ocfs2_dinode is inconsistency.  So the clusters got from
ocfs2_dinode is incorrect, and it also cause read-only problem when call
ocfs2_commit_truncate() with the error message: "Inode %llu has empty
extent block at %llu".

We should extend enough credits for function ocfs2_remove_rightmost_path
and ocfs2_update_edge_lengths to avoid this inconsistency.

Signed-off-by: joyce.xue <xuejiufei@huawei.com>
Acked-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2/dlm: move lock to the tail of grant queue while doing in-place convert
xuejiufei [Fri, 25 Mar 2016 21:21:38 +0000 (14:21 -0700)]
ocfs2/dlm: move lock to the tail of grant queue while doing in-place convert

We have found a bug when two nodes doing umount one after another.

1) Node 1 migrate a lockres that has 3 locks in grant queue such as
   N2(PR)<->N3(NL)<->N4(PR) to N2.  After migration, lvb of the lock
   N3(NL) and N4(PR) are empty on node 2 because migration target do not
   copy lvb to these two lock.

2) Node 3 want to convert to PR, it can be granted in
   __dlmconvert_master(), and the order of these locks is unchanged.  The
   lvb of the lock N3(PR) on node 2 is copyed from lockres in function
   dlm_update_lvb() while the lvb of lock N4(PR) is still empty.

3) Node 2 want to leave domain, it will migrate this lockres to node 3.
   Then node 2 will trigger the BUG in dlm_prepare_lvb_for_migration()
   when adding the lock N4(PR) to mres with the following message because
   the lvb of mres is already copied from lock N3(PR), but the lvb of lock
   N4(PR) is empty.

"Mismatched lvb in lock cookie=%u:%llu, name=%.*s, node=%u"

[akpm@linux-foundation.org: tweak comment]
Signed-off-by: xuejiufei <xuejiufei@huawei.com>
Acked-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: solve a problem of crossing the boundary in updating backups
jiangyiwen [Fri, 25 Mar 2016 21:21:35 +0000 (14:21 -0700)]
ocfs2: solve a problem of crossing the boundary in updating backups

In update_backups() there exists a problem of crossing the boundary as
follows:

we assume that lun will be resized to 1TB(cluster_size is 32kb), it will
include 0~33554431 cluster, in update_backups func, it will backup super
block in location of 1TB which is the 33554432th cluster, so the
phenomenon of crossing the boundary happens.

Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Xue jiufei <xuejiufei@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: fix occurring deadlock by changing ocfs2_wq from global to local
jiangyiwen [Fri, 25 Mar 2016 21:21:32 +0000 (14:21 -0700)]
ocfs2: fix occurring deadlock by changing ocfs2_wq from global to local

This patch fixes a deadlock, as follows:

  Node 1                Node 2                  Node 3
1)volume a and b are    only mount vol a        only mount vol b
  mounted

2)                      start to mount b        start to mount a

3)                      check hb of Node 3      check hb of Node 2
                        in vol a, qs_holds++    in vol b, qs_holds++

4) -------------------- all nodes' network down --------------------

5)                      progress of mount b     the same situation as
                        failed, and then call   Node 2
                        ocfs2_dismount_volume.
                        but the process is hung,
                        since there is a work
                        in ocfs2_wq cannot beo
                        completed. This work is
                        about vol a, because
                        ocfs2_wq is global wq.
                        BTW, this work which is
                        scheduled in ocfs2_wq is
                        ocfs2_orphan_scan_work,
                        and the context in this work
                        needs to take inode lock
                        of orphan_dir, because
                        lockres owner are Node 1 and
                        all nodes' nework has been down
                        at the same time, so it can't
                        get the inode lock.

6)                      Why can't this node be fenced
                        when network disconnected?
                        Because the process of
                        mount is hung what caused qs_holds
                        is not equal 0.

Because all works in the ocfs2_wq are relative to the super block.

The solution is to change the ocfs2_wq from global to local.  In other
words, move it into struct ocfs2_super.

Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Xue jiufei <xuejiufei@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2/dlm: fix BUG in dlm_move_lockres_to_recovery_list
Joseph Qi [Fri, 25 Mar 2016 21:21:29 +0000 (14:21 -0700)]
ocfs2/dlm: fix BUG in dlm_move_lockres_to_recovery_list

When master handles convert request, it queues ast first and then
returns status.  This may happen that the ast is sent before the request
status because the above two messages are sent by two threads.  And
right after the ast is sent, if master down, it may trigger BUG in
dlm_move_lockres_to_recovery_list in the requested node because ast
handler moves it to grant list without clear lock->convert_pending.  So
remove BUG_ON statement and check if the ast is processed in
dlmconvert_remote.

Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Reported-by: Yiwen Jiang <jiangyiwen@huawei.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Tariq Saeed <tariq.x.saeed@oracle.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2/dlm: fix race between convert and recovery
Joseph Qi [Fri, 25 Mar 2016 21:21:26 +0000 (14:21 -0700)]
ocfs2/dlm: fix race between convert and recovery

There is a race window between dlmconvert_remote and
dlm_move_lockres_to_recovery_list, which will cause a lock with
OCFS2_LOCK_BUSY in grant list, thus system hangs.

dlmconvert_remote
{
        spin_lock(&res->spinlock);
        list_move_tail(&lock->list, &res->converting);
        lock->convert_pending = 1;
        spin_unlock(&res->spinlock);

        status = dlm_send_remote_convert_request();
        >>>>>> race window, master has queued ast and return DLM_NORMAL,
               and then down before sending ast.
               this node detects master down and calls
               dlm_move_lockres_to_recovery_list, which will revert the
               lock to grant list.
               Then OCFS2_LOCK_BUSY won't be cleared as new master won't
               send ast any more because it thinks already be authorized.

        spin_lock(&res->spinlock);
        lock->convert_pending = 0;
        if (status != DLM_NORMAL)
                dlm_revert_pending_convert(res, lock);
        spin_unlock(&res->spinlock);
}

In this case, check if res->state has DLM_LOCK_RES_RECOVERING bit set
(res is still in recovering) or res master changed (new master has
finished recovery), reset the status to DLM_RECOVERING, then it will
retry convert.

Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Reported-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Tariq Saeed <tariq.x.saeed@oracle.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: fix a deadlock issue in ocfs2_dio_end_io_write()
Ryan Ding [Fri, 25 Mar 2016 21:21:23 +0000 (14:21 -0700)]
ocfs2: fix a deadlock issue in ocfs2_dio_end_io_write()

The code should call ocfs2_free_alloc_context() to free meta_ac &
data_ac before calling ocfs2_run_deallocs().  Because
ocfs2_run_deallocs() will acquire the system inode's i_mutex hold by
meta_ac.  So try to release the lock before ocfs2_run_deallocs().

Fixes: af1310367f41 ("ocfs2: fix sparse file & data ordering issue in direct io.")
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Acked-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: fix disk file size and memory file size mismatch
Ryan Ding [Fri, 25 Mar 2016 21:21:20 +0000 (14:21 -0700)]
ocfs2: fix disk file size and memory file size mismatch

When doing append direct write in an already allocated cluster, and fast
path in ocfs2_dio_get_block() is triggered, function
ocfs2_dio_end_io_write() will be skipped as there is no context
allocated.

As a result, the disk file size will not be changed as it should be.
The solution is to skip fast path when we are about to change file size.

Fixes: af1310367f41 ("ocfs2: fix sparse file & data ordering issue in direct io.")
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Acked-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: take ip_alloc_sem in ocfs2_dio_get_block & ocfs2_dio_end_io_write
Ryan Ding [Fri, 25 Mar 2016 21:21:18 +0000 (14:21 -0700)]
ocfs2: take ip_alloc_sem in ocfs2_dio_get_block & ocfs2_dio_end_io_write

Take ip_alloc_sem to prevent concurrent access to extent tree, which may
cause the extent tree in an unstable state.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: fix ip_unaligned_aio deadlock with dio work queue
Ryan Ding [Fri, 25 Mar 2016 21:21:15 +0000 (14:21 -0700)]
ocfs2: fix ip_unaligned_aio deadlock with dio work queue

In the current implementation of unaligned aio+dio, lock order behave as
follow:

in user process context:
  -> call io_submit()
    -> get i_mutex
<== window1
      -> get ip_unaligned_aio
        -> submit direct io to block device
    -> release i_mutex
  -> io_submit() return

in dio work queue context(the work queue is created in __blockdev_direct_IO):
  -> release ip_unaligned_aio
<== window2
    -> get i_mutex
      -> clear unwritten flag & change i_size
    -> release i_mutex

There is a limitation to the thread number of dio work queue.  256 at
default.  If all 256 thread are in the above 'window2' stage, and there
is a user process in the 'window1' stage, the system will became
deadlock.  Since the user process hold i_mutex to wait ip_unaligned_aio
lock, while there is a direct bio hold ip_unaligned_aio mutex who is
waiting for a dio work queue thread to be schedule.  But all the dio
work queue thread is waiting for i_mutex lock in 'window2'.

This case only happened in a test which send a large number(more than
256) of aio at one io_submit() call.

My design is to remove ip_unaligned_aio lock.  Change it to a sync io
instead.  Just like ip_unaligned_aio lock, serialize the unaligned aio
dio.

[akpm@linux-foundation.org: remove OCFS2_IOCB_UNALIGNED_IO, per Junxiao Bi]
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: code clean up for direct io
Ryan Ding [Fri, 25 Mar 2016 21:21:12 +0000 (14:21 -0700)]
ocfs2: code clean up for direct io

Clean up ocfs2_file_write_iter & ocfs2_prepare_inode_for_write:
 * remove append dio check: it will be checked in ocfs2_direct_IO()
 * remove file hole check: file hole is supported for now
 * remove inline data check: it will be checked in ocfs2_direct_IO()
 * remove the full_coherence check when append dio: we will get the
   inode_lock in ocfs2_dio_get_block, there is no need to fall back to
   buffer io to ensure the coherence semantics.

Now the drop dio procedure is gone.  :)

[akpm@linux-foundation.org: remove unused label]
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: fix sparse file & data ordering issue in direct io
Ryan Ding [Fri, 25 Mar 2016 21:21:09 +0000 (14:21 -0700)]
ocfs2: fix sparse file & data ordering issue in direct io

There are mainly three issues in the direct io code path after commit
24c40b329e03 ("ocfs2: implement ocfs2_direct_IO_write"):

  * Does not support sparse file.
  * Does not support data ordering.  eg: when write to a file hole, it
    will alloc extent first.  If system crashed before io finished, data
    will corrupt.
  * Potential risk when doing aio+dio.  The -EIOCBQUEUED return value is
    likely to be ignored by ocfs2_direct_IO_write().

To resolve above problems, re-design direct io code with following ideas:
  * Use buffer io to fill in holes.  And this will make better
    performance also.
  * Clear unwritten after direct write finished.  So we can make sure
    meta data changes after data write to disk.  (Unwritten extent is
    invisible to user, from user's view, meta data is not changed when
    allocate an unwritten extent.)
  * Clear ocfs2_direct_IO_write().  Do all ending work in end_io.

This patch has passed fs,dio,ltp-aiodio.part1,ltp-aiodio.part2,ltp-aiodio.part4
test cases of ltp.

For performance improvement, see following test result:
ocfs2 cluster size 1MB, ocfs2 volume is mounted on /mnt/.
The original way:
  + rm /mnt/test.img -f
  + dd if=/dev/zero of=/mnt/test.img bs=4K count=1048576 oflag=direct
  1048576+0 records in
  1048576+0 records out
  4294967296 bytes (4.3 GB) copied, 1707.83 s, 2.5 MB/s
  + rm /mnt/test.img -f
  + dd if=/dev/zero of=/mnt/test.img bs=256K count=16384 oflag=direct
  16384+0 records in
  16384+0 records out
  4294967296 bytes (4.3 GB) copied, 582.705 s, 7.4 MB/s

After this patch:
  + rm /mnt/test.img -f
  + dd if=/dev/zero of=/mnt/test.img bs=4K count=1048576 oflag=direct
  1048576+0 records in
  1048576+0 records out
  4294967296 bytes (4.3 GB) copied, 64.6412 s, 66.4 MB/s
  + rm /mnt/test.img -f
  + dd if=/dev/zero of=/mnt/test.img bs=256K count=16384 oflag=direct
  16384+0 records in
  16384+0 records out
  4294967296 bytes (4.3 GB) copied, 34.7611 s, 124 MB/s

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: record UNWRITTEN extents when populate write desc
Ryan Ding [Fri, 25 Mar 2016 21:21:06 +0000 (14:21 -0700)]
ocfs2: record UNWRITTEN extents when populate write desc

To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.

There is still one issue in the direct write procedure.

phase 1: alloc extent with UNWRITTEN flag
phase 2: submit direct data to disk, add zero page to page cache
phase 3: clear UNWRITTEN flag when data has been written to disk

When there are 2 direct write A(0~3KB),B(4~7KB) writing to the same
cluster 0~7KB (cluster size 8KB).  Write request A arrive phase 2 first,
it will zero the region (4~7KB).  Before request A enter to phase 3,
request B arrive phase 2, it will zero region (0~3KB).  This is just like
request B steps request A.

To resolve this issue, we should let request B knows this cluster is already
under zero, to prevent it from steps the previous write request.

This patch will add function ocfs2_unwritten_check() to do this job.  It
will record all clusters that are under direct write(it will be recorded
in the 'ip_unwritten_list' member of inode info), and prevent the later
direct write writing to the same cluster to do the zero work again.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: return the physical address in ocfs2_write_cluster
Ryan Ding [Fri, 25 Mar 2016 21:21:03 +0000 (14:21 -0700)]
ocfs2: return the physical address in ocfs2_write_cluster

To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.

Direct io needs to get the physical address from write_begin, to map the
user page.  This patch is to change the arg 'phys' of
ocfs2_write_cluster to a pointer, so it can be retrieved to write_begin.
And we can retrieve it to the direct io procedure.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: do not change i_size in write_end for direct io
Ryan Ding [Fri, 25 Mar 2016 21:21:01 +0000 (14:21 -0700)]
ocfs2: do not change i_size in write_end for direct io

To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.

Append direct io do not change i_size in get block phase.  It only move
to orphan when starting write.  After data is written to disk, it will
delete itself from orphan and update i_size.  So skip i_size change
section in write_begin for direct io.

And when there is no extents alloc, no meta data changes needed for
direct io (since write_begin start trans for 2 reason: alloc extents &
change i_size.  Now none of them needed).  So we can skip start trans
procedure.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: test target page before change it
Ryan Ding [Fri, 25 Mar 2016 21:20:58 +0000 (14:20 -0700)]
ocfs2: test target page before change it

To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.

Direct io data will not appear in buffer.  The w_target_page member will
not be filled by direct io.  So avoid to use it when it's NULL.  Unlinke
buffer io and mmap, direct io will call write_begin with more than 1
page a time.  So the target_index is not sufficient to describe the
actual data.  change it to a range start at target_index, end in
end_index.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: use c_new to indicate newly allocated extents
Ryan Ding [Fri, 25 Mar 2016 21:20:55 +0000 (14:20 -0700)]
ocfs2: use c_new to indicate newly allocated extents

To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.

There is a problem in ocfs2's direct io implement: if system crashed
after extents allocated, and before data return, we will get a extent
with dirty data on disk.  This problem violate the journal=order
semantics, which means meta changes take effect after data written to
disk.  To resolve this issue, direct write can use the UNWRITTEN flag to
describe a extent during direct data writeback.  The direct write
procedure should act in the following order:

phase 1: alloc extent with UNWRITTEN flag
phase 2: submit direct data to disk, add zero page to page cache
phase 3: clear UNWRITTEN flag when data has been written to disk

This patch is to change the 'c_unwritten' member of
ocfs2_write_cluster_desc to 'c_clear_unwritten'.  Means whether to clear
the unwritten flag.  It do not care if a extent is allocated or not.
And use 'c_new' to specify a newly allocated extent.  So the direct io
procedure can use c_clear_unwritten to control the UNWRITTEN bit on
extent.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: add ocfs2_write_type_t type to identify the caller of write
Ryan Ding [Fri, 25 Mar 2016 21:20:52 +0000 (14:20 -0700)]
ocfs2: add ocfs2_write_type_t type to identify the caller of write

Patchset: fix ocfs2 direct io code patch to support sparse file and data
ordering semantics

The idea is to use buffer io(more precisely use the interface
ocfs2_write_begin_nolock & ocfs2_write_end_nolock) to do the zero work
beyond block size.  And clear UNWRITTEN flag until direct io data has
been written to disk, which can prevent data corruption when system
crashed during direct write.

And we will also archive a better performance: eg.  dd direct write new
file with block size 4KB: before this patchset:
  2.5 MB/s
after this patchset:
  66.4 MB/s

This patch (of 8):

To support direct io in ocfs2_write_begin_nolock &
ocfs2_write_end_nolock.

Remove unused args filp & flags.  Add new arg type.  The type is one of
buffer/direct/mmap.  Indicate 3 way to perform write.  buffer/mmap type
has implemented.  direct type will be implemented later.

Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoocfs2: o2hb: fix double free bug
Junxiao Bi [Fri, 25 Mar 2016 21:20:50 +0000 (14:20 -0700)]
ocfs2: o2hb: fix double free bug

This is a regression issue and caused the following kernel panic when do
ocfs2 multiple test.

  BUG: unable to handle kernel paging request at 00000002000800c0
  IP: [<ffffffff81192978>] kmem_cache_alloc+0x78/0x160
  PGD 7bbe5067 PUD 0
  Oops: 0000 [#1] SMP
  Modules linked in: ocfs2_dlmfs ocfs2_stack_o2cb ocfs2_dlm ocfs2_nodemanager ocfs2_stackglue iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi xen_kbdfront xen_netfront xen_fbfront xen_blkfront
  CPU: 2 PID: 4044 Comm: mpirun Not tainted 4.5.0-rc5-next-20160225 #1
  Hardware name: Xen HVM domU, BIOS 4.3.1OVM 05/14/2014
  task: ffff88007a521a80 ti: ffff88007aed0000 task.ti: ffff88007aed0000
  RIP: 0010:[<ffffffff81192978>]  [<ffffffff81192978>] kmem_cache_alloc+0x78/0x160
  RSP: 0018:ffff88007aed3a48  EFLAGS: 00010282
  RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000001991
  RDX: 0000000000001990 RSI: 00000000024000c0 RDI: 000000000001b330
  RBP: ffff88007aed3a98 R08: ffff88007d29b330 R09: 00000002000800c0
  R10: 0000000c51376d87 R11: ffff8800792cac38 R12: ffff88007cc30f00
  R13: 00000000024000c0 R14: ffffffff811b053f R15: ffff88007aed3ce7
  FS:  0000000000000000(0000) GS:ffff88007d280000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00000002000800c0 CR3: 000000007aeb2000 CR4: 00000000000406e0
  Call Trace:
    __d_alloc+0x2f/0x1a0
    d_alloc+0x17/0x80
    lookup_dcache+0x8a/0xc0
    path_openat+0x3c3/0x1210
    do_filp_open+0x80/0xe0
    do_sys_open+0x110/0x200
    SyS_open+0x19/0x20
    do_syscall_64+0x72/0x230
    entry_SYSCALL64_slow_path+0x25/0x25
  Code: 05 e6 77 e7 7e 4d 8b 08 49 8b 40 10 4d 85 c9 0f 84 dd 00 00 00 48 85 c0 0f 84 d4 00 00 00 49 63 44 24 20 49 8b 3c 24 48 8d 4a 01 <49> 8b 1c 01 4c 89 c8 65 48 0f c7 0f 0f 94 c0 3c 01 75 b6 49 63
  RIP   kmem_cache_alloc+0x78/0x160
  CR2: 00000002000800c0
  ---[ end trace 823969e602e4aaac ]---

Fixes: a4a1dfa4bb8b("ocfs2/cluster: fix memory leak in o2hb_region_release")
Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agodrivers/input: eliminate INPUT_COMPAT_TEST macro
Andrew Morton [Fri, 25 Mar 2016 21:20:47 +0000 (14:20 -0700)]
drivers/input: eliminate INPUT_COMPAT_TEST macro

INPUT_COMPAT_TEST became much simpler after commit f4056b52845283
("input: redefine INPUT_COMPAT_TEST as in_compat_syscall()") so we can
cleanly eliminate it altogether.

Acked-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agooom, oom_reaper: protect oom_reaper_list using simpler way
Tetsuo Handa [Fri, 25 Mar 2016 21:20:44 +0000 (14:20 -0700)]
oom, oom_reaper: protect oom_reaper_list using simpler way

"oom, oom_reaper: disable oom_reaper for oom_kill_allocating_task" tried
to protect oom_reaper_list using MMF_OOM_KILLED flag.  But we can do it
by simply checking tsk->oom_reaper_list != NULL.

Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agooom: make oom_reaper freezable
Michal Hocko [Fri, 25 Mar 2016 21:20:41 +0000 (14:20 -0700)]
oom: make oom_reaper freezable

After "oom: clear TIF_MEMDIE after oom_reaper managed to unmap the
address space" oom_reaper will call exit_oom_victim on the target task
after it is done.  This might however race with the PM freezer:

CPU0 CPU1 CPU2
freeze_processes
  try_to_freeze_tasks
   # Allocation request
out_of_memory
  oom_killer_disable
  wake_oom_reaper(P1)
   __oom_reap_task
  exit_oom_victim(P1)
    wait_event(oom_victims==0)
[...]
     do_exit(P1)
  perform IO/interfere with the freezer

which breaks the oom_killer_disable semantic.  We no longer have a
guarantee that the oom victim won't interfere with the freezer because
it might be anywhere on the way to do_exit while the freezer thinks the
task has already terminated.  It might trigger IO or touch devices which
are frozen already.

In order to close this race, make the oom_reaper thread freezable.  This
will work because
a) already running oom_reaper will block freezer to enter the
   quiescent state
b) wake_oom_reaper will not wake up the reaper after it has been
   frozen
c) the only way to call exit_oom_victim after try_to_freeze_tasks
   is from the oom victim's context when we know the further
   interference shouldn't be possible

Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agooom: make oom_reaper_list single linked
Vladimir Davydov [Fri, 25 Mar 2016 21:20:39 +0000 (14:20 -0700)]
oom: make oom_reaper_list single linked

Entries are only added/removed from oom_reaper_list at head so we can
use a single linked list and hence save a word in task_struct.

Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agooom, oom_reaper: disable oom_reaper for oom_kill_allocating_task
Michal Hocko [Fri, 25 Mar 2016 21:20:36 +0000 (14:20 -0700)]
oom, oom_reaper: disable oom_reaper for oom_kill_allocating_task

Tetsuo has reported that oom_kill_allocating_task=1 will cause
oom_reaper_list corruption because oom_kill_process doesn't follow
standard OOM exclusion (aka ignores TIF_MEMDIE) and allows to enqueue
the same task multiple times - e.g.  by sacrificing the same child
multiple times.

This patch fixes the issue by introducing a new MMF_OOM_KILLED mm flag
which is set in oom_kill_process atomically and oom reaper is disabled
if the flag was already set.

Signed-off-by: Michal Hocko <mhocko@suse.com>
Reported-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, oom_reaper: implement OOM victims queuing
Michal Hocko [Fri, 25 Mar 2016 21:20:33 +0000 (14:20 -0700)]
mm, oom_reaper: implement OOM victims queuing

wake_oom_reaper has allowed only 1 oom victim to be queued.  The main
reason for that was the simplicity as other solutions would require some
way of queuing.  The current approach is racy and that was deemed
sufficient as the oom_reaper is considered a best effort approach to
help with oom handling when the OOM victim cannot terminate in a
reasonable time.  The race could lead to missing an oom victim which can
get stuck

out_of_memory
  wake_oom_reaper
    cmpxchg // OK
     oom_reaper
  oom_reap_task
    __oom_reap_task
oom_victim terminates
      atomic_inc_not_zero // fail
out_of_memory
  wake_oom_reaper
    cmpxchg // fails
  task_to_reap = NULL

This race requires 2 OOM invocations in a short time period which is not
very likely but certainly not impossible.  E.g.  the original victim
might have not released a lot of memory for some reason.

The situation would improve considerably if wake_oom_reaper used a more
robust queuing.  This is what this patch implements.  This means adding
oom_reaper_list list_head into task_struct (eat a hole before embeded
thread_struct for that purpose) and a oom_reaper_lock spinlock for
queuing synchronization.  wake_oom_reaper will then add the task on the
queue and oom_reaper will dequeue it.

Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov@virtuozzo.com>
Cc: Andrea Argangeli <andrea@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, oom_reaper: report success/failure
Michal Hocko [Fri, 25 Mar 2016 21:20:30 +0000 (14:20 -0700)]
mm, oom_reaper: report success/failure

Inform about the successful/failed oom_reaper attempts and dump all the
held locks to tell us more who is blocking the progress.

[akpm@linux-foundation.org: fix CONFIG_MMU=n build]
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Argangeli <andrea@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agooom: clear TIF_MEMDIE after oom_reaper managed to unmap the address space
Michal Hocko [Fri, 25 Mar 2016 21:20:27 +0000 (14:20 -0700)]
oom: clear TIF_MEMDIE after oom_reaper managed to unmap the address space

When oom_reaper manages to unmap all the eligible vmas there shouldn't
be much of the freable memory held by the oom victim left anymore so it
makes sense to clear the TIF_MEMDIE flag for the victim and allow the
OOM killer to select another task.

The lack of TIF_MEMDIE also means that the victim cannot access memory
reserves anymore but that shouldn't be a problem because it would get
the access again if it needs to allocate and hits the OOM killer again
due to the fatal_signal_pending resp.  PF_EXITING check.  We can safely
hide the task from the OOM killer because it is clearly not a good
candidate anymore as everyhing reclaimable has been torn down already.

This patch will allow to cap the time an OOM victim can keep TIF_MEMDIE
and thus hold off further global OOM killer actions granted the oom
reaper is able to take mmap_sem for the associated mm struct.  This is
not guaranteed now but further steps should make sure that mmap_sem for
write should be blocked killable which will help to reduce such a lock
contention.  This is not done by this patch.

Note that exit_oom_victim might be called on a remote task from
__oom_reap_task now so we have to check and clear the flag atomically
otherwise we might race and underflow oom_victims or wake up waiters too
early.

Signed-off-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Argangeli <andrea@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, oom: introduce oom reaper
Michal Hocko [Fri, 25 Mar 2016 21:20:24 +0000 (14:20 -0700)]
mm, oom: introduce oom reaper

This patch (of 5):

This is based on the idea from Mel Gorman discussed during LSFMM 2015
and independently brought up by Oleg Nesterov.

The OOM killer currently allows to kill only a single task in a good
hope that the task will terminate in a reasonable time and frees up its
memory.  Such a task (oom victim) will get an access to memory reserves
via mark_oom_victim to allow a forward progress should there be a need
for additional memory during exit path.

It has been shown (e.g.  by Tetsuo Handa) that it is not that hard to
construct workloads which break the core assumption mentioned above and
the OOM victim might take unbounded amount of time to exit because it
might be blocked in the uninterruptible state waiting for an event (e.g.
lock) which is blocked by another task looping in the page allocator.

This patch reduces the probability of such a lockup by introducing a
specialized kernel thread (oom_reaper) which tries to reclaim additional
memory by preemptively reaping the anonymous or swapped out memory owned
by the oom victim under an assumption that such a memory won't be needed
when its owner is killed and kicked from the userspace anyway.  There is
one notable exception to this, though, if the OOM victim was in the
process of coredumping the result would be incomplete.  This is
considered a reasonable constrain because the overall system health is
more important than debugability of a particular application.

A kernel thread has been chosen because we need a reliable way of
invocation so workqueue context is not appropriate because all the
workers might be busy (e.g.  allocating memory).  Kswapd which sounds
like another good fit is not appropriate as well because it might get
blocked on locks during reclaim as well.

oom_reaper has to take mmap_sem on the target task for reading so the
solution is not 100% because the semaphore might be held or blocked for
write but the probability is reduced considerably wrt.  basically any
lock blocking forward progress as described above.  In order to prevent
from blocking on the lock without any forward progress we are using only
a trylock and retry 10 times with a short sleep in between.  Users of
mmap_sem which need it for write should be carefully reviewed to use
_killable waiting as much as possible and reduce allocations requests
done with the lock held to absolute minimum to reduce the risk even
further.

The API between oom killer and oom reaper is quite trivial.
wake_oom_reaper updates mm_to_reap with cmpxchg to guarantee only
NULL->mm transition and oom_reaper clear this atomically once it is done
with the work.  This means that only a single mm_struct can be reaped at
the time.  As the operation is potentially disruptive we are trying to
limit it to the ncessary minimum and the reaper blocks any updates while
it operates on an mm.  mm_struct is pinned by mm_count to allow parallel
exit_mmap and a race is detected by atomic_inc_not_zero(mm_users).

Signed-off-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Oleg Nesterov <oleg@redhat.com>
Suggested-by: Mel Gorman <mgorman@suse.de>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Argangeli <andrea@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agosched: add schedule_timeout_idle()
Andrew Morton [Fri, 25 Mar 2016 21:20:21 +0000 (14:20 -0700)]
sched: add schedule_timeout_idle()

This will be needed in the patch "mm, oom: introduce oom reaper".

Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years ago[IA64] Enable preadv2 and pwritev2 syscalls for ia64
Tony Luck [Fri, 25 Mar 2016 21:37:32 +0000 (14:37 -0700)]
[IA64] Enable preadv2 and pwritev2 syscalls for ia64

New system calls added in:
      f17d8b35452cab31a70d224964cd583fb2845449
      vfs: vfs: Define new syscalls preadv2,pwritev2

Signed-off-by: Tony Luck <tony.luck@intel.com>
8 years agoFix permissions of drivers/power/avs/rockchip-io-domain.c
Rafael J. Wysocki [Fri, 25 Mar 2016 21:33:10 +0000 (22:33 +0100)]
Fix permissions of drivers/power/avs/rockchip-io-domain.c

The permissions of this file were modified by commit (f447671b9e4f PM /
AVS: rockchip-io: add io selectors and supplies for rk3399) by mistake,
so fix them.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
8 years agolibceph: use KMEM_CACHE macro
Geliang Tang [Sun, 13 Mar 2016 07:18:39 +0000 (15:18 +0800)]
libceph: use KMEM_CACHE macro

Use KMEM_CACHE() instead of kmem_cache_create() to simplify the code.

Signed-off-by: Geliang Tang <geliangtang@163.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agoceph: use kmem_cache_zalloc
Geliang Tang [Sun, 13 Mar 2016 07:26:29 +0000 (15:26 +0800)]
ceph: use kmem_cache_zalloc

Use kmem_cache_zalloc() instead of kmem_cache_alloc() with flag GFP_ZERO.

Signed-off-by: Geliang Tang <geliangtang@163.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agorbd: use KMEM_CACHE macro
Geliang Tang [Sun, 13 Mar 2016 07:17:32 +0000 (15:17 +0800)]
rbd: use KMEM_CACHE macro

Use KMEM_CACHE() instead of kmem_cache_create() to simplify the code.

Signed-off-by: Geliang Tang <geliangtang@163.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agoceph: use lookup request to revalidate dentry
Yan, Zheng [Thu, 17 Mar 2016 06:41:59 +0000 (14:41 +0800)]
ceph: use lookup request to revalidate dentry

If dentry has no lease, ceph_d_revalidate() previously return 0.
This causes VFS to invalidate the dentry and create a new dentry
for later lookup. Invalidating a dentry also detach any underneath
mount points. So mount point inside cephfs can disapear mystically
(even the mount point is not modified by other hosts).

The fix is using lookup request to revalidate dentry without lease.
This can partly solve the mount points disapear issue (as long as
the mount point is not modified by other hosts)

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: kill ceph_get_dentry_parent_inode()
Yan, Zheng [Wed, 16 Mar 2016 08:40:23 +0000 (16:40 +0800)]
ceph: kill ceph_get_dentry_parent_inode()

use vfs helper dget_parent() instead

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: fix security xattr deadlock
Yan, Zheng [Mon, 7 Mar 2016 02:34:50 +0000 (10:34 +0800)]
ceph: fix security xattr deadlock

When security is enabled, security module can call filesystem's
getxattr/setxattr callbacks during d_instantiate(). For cephfs,
d_instantiate() is usually called by MDS' dispatch thread, while
handling MDS reply. If the MDS reply does not include xattrs and
corresponding caps, getxattr/setxattr need to send a new request
to MDS and waits for the reply. This makes MDS' dispatch sleep,
nobody handles later MDS replies.

The fix is make sure lookup/atomic_open reply include xattrs and
corresponding caps. So getxattr can be handled by cached xattrs.
This requires some modification to both MDS and request message.
(Client tells MDS what caps it wants; MDS encodes proper caps in
the reply)

Smack security module may call setxattr during d_instantiate().
Unlike getxattr, we can't force MDS to issue CEPH_CAP_XATTR_EXCL
to us. So just make setxattr return error when called by MDS'
dispatch thread.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: don't request vxattrs from MDS
Yan, Zheng [Sat, 12 Mar 2016 05:32:16 +0000 (13:32 +0800)]
ceph: don't request vxattrs from MDS

It's uselese because MDS reply does not carry any vxattr.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: fix mounting same fs multiple times
Yan, Zheng [Sat, 12 Mar 2016 05:20:48 +0000 (13:20 +0800)]
ceph: fix mounting same fs multiple times

Now __ceph_open_session() only accepts closed client. An opened
client will tigger BUG_ON().

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: remove unnecessary NULL check
Yan, Zheng [Thu, 10 Mar 2016 03:29:34 +0000 (11:29 +0800)]
ceph: remove unnecessary NULL check

If page->mapping is NULL, releasepage() callback does not get called.
Remove the unnecessary NULL check to make static code analysis tool
happy

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: avoid updating directory inode's i_size accidentally
Yan, Zheng [Fri, 26 Feb 2016 09:16:32 +0000 (17:16 +0800)]
ceph: avoid updating directory inode's i_size accidentally

Directory inode's i_size is used by readdir cache.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: fix race during filling readdir cache
Yan, Zheng [Fri, 26 Feb 2016 08:27:13 +0000 (16:27 +0800)]
ceph: fix race during filling readdir cache

Readdir cache uses page cache to save dentry pointers. When adding
dentry pointers to middle of a page, we need to make sure the page
already exists. Otherwise the beginning part of the page will be
invalid pointers.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agolibceph: use sizeof_footer() more
Ilya Dryomov [Sat, 20 Feb 2016 14:56:07 +0000 (15:56 +0100)]
libceph: use sizeof_footer() more

Don't open-code sizeof_footer() in read_partial_message() and
ceph_msg_revoke().  Also, after switching to sizeof_footer(), it's now
possible to use con_out_kvec_add() in prepare_write_message_footer().

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
8 years agoceph: kill ceph_empty_snapc
Ilya Dryomov [Tue, 16 Feb 2016 14:00:24 +0000 (15:00 +0100)]
ceph: kill ceph_empty_snapc

ceph_empty_snapc->num_snaps == 0 at all times.  Passing such a snapc to
ceph_osdc_alloc_request() (possibly through ceph_osdc_new_request()) is
equivalent to passing NULL, as ceph_osdc_alloc_request() uses it only
for sizing the request message.

Further, in all four cases the subsequent ceph_osdc_build_request() is
passed NULL for snapc, meaning that 0 is encoded for seq and num_snaps
and making ceph_empty_snapc entirely useless.  The two cases where it
actually mattered were removed in commits 860560904962 ("ceph: avoid
sending unnessesary FLUSHSNAP message") and 23078637e054 ("ceph: fix
queuing inode to mdsdir's snaprealm").

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: fix a wrong comparison
Anton Protopopov [Wed, 10 Feb 2016 17:38:03 +0000 (12:38 -0500)]
ceph: fix a wrong comparison

A negative value rc compared to the positive value ENOENT in the
finish_read() function.

Signed-off-by: Anton Protopopov <a.s.protopopov@gmail.com>
Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: replace CURRENT_TIME by current_fs_time()
Deepa Dinamani [Wed, 3 Feb 2016 06:07:48 +0000 (22:07 -0800)]
ceph: replace CURRENT_TIME by current_fs_time()

CURRENT_TIME macro is not appropriate for filesystems as it
doesn't use the right granularity for filesystem timestamps.
Use current_fs_time() instead.

Signed-off-by: Deepa Dinamani <deepa.kernel@gmail.com>
Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: scattered page writeback
Yan, Zheng [Thu, 7 Jan 2016 08:00:17 +0000 (16:00 +0800)]
ceph: scattered page writeback

This patch makes ceph_writepages_start() try using single OSD request
to write all dirty pages within a strip unit. When a nonconsecutive
dirty page is found, ceph_writepages_start() tries starting a new write
operation to existing OSD request. If it succeeds, it uses the new
operation to writeback the dirty page.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agolibceph: add helper that duplicates last extent operation
Yan, Zheng [Thu, 7 Jan 2016 09:32:54 +0000 (17:32 +0800)]
libceph: add helper that duplicates last extent operation

This helper duplicates last extent operation in OSD request, then
adjusts the new extent operation's offset and length. The helper
is for scatterd page writeback, which adds nonconsecutive dirty
pages to single OSD request.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: enable large, variable-sized OSD requests
Ilya Dryomov [Tue, 9 Feb 2016 16:50:15 +0000 (17:50 +0100)]
libceph: enable large, variable-sized OSD requests

Turn r_ops into a flexible array member to enable large, consisting of
up to 16 ops, OSD requests.  The use case is scattered writeback in
cephfs and, as far as the kernel client is concerned, 16 is just a made
up number.

r_ops had size 3 for copyup+hint+write, but copyup is really a special
case - it can only happen once.  ceph_osd_request_cache is therefore
stuffed with num_ops=2 requests, anything bigger than that is allocated
with kmalloc().  req_mempool is backed by ceph_osd_request_cache, which
means either num_ops=1 or num_ops=2 for use_mempool=true - all existing
users (ceph_writepages_start(), ceph_osdc_writepages()) are fine with
that.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: osdc->req_mempool should be backed by a slab pool
Ilya Dryomov [Tue, 9 Feb 2016 16:25:31 +0000 (17:25 +0100)]
libceph: osdc->req_mempool should be backed by a slab pool

ceph_osd_request_cache was introduced a long time ago.  Also, osd_req
is about to get a flexible array member, which ceph_osd_request_cache
is going to be aware of.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: make r_request msg_size calculation clearer
Ilya Dryomov [Thu, 11 Feb 2016 12:09:15 +0000 (13:09 +0100)]
libceph: make r_request msg_size calculation clearer

Although msg_size is calculated correctly, the terms are grouped in
a misleading way - snaps appears to not have room for a u32 length.
Move calculation closer to its use and regroup terms.

No functional change.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: move r_reply_op_{len,result} into struct ceph_osd_req_op
Yan, Zheng [Thu, 7 Jan 2016 08:48:57 +0000 (16:48 +0800)]
libceph: move r_reply_op_{len,result} into struct ceph_osd_req_op

This avoids defining large array of r_reply_op_{len,result} in
in struct ceph_osd_request.

Signed-off-by: Yan, Zheng <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: rename ceph_osd_req_op::payload_len to indata_len
Ilya Dryomov [Mon, 8 Feb 2016 12:39:46 +0000 (13:39 +0100)]
libceph: rename ceph_osd_req_op::payload_len to indata_len

Follow userspace nomenclature on this - the next commit adds
outdata_len.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agoceph: remove useless BUG_ON
Yan, Zheng [Wed, 27 Jan 2016 01:30:29 +0000 (09:30 +0800)]
ceph: remove useless BUG_ON

ceph_osdc_start_request() never return -EOLDSNAP

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: don't enable rbytes mount option by default
Yan, Zheng [Mon, 25 Jan 2016 02:44:33 +0000 (10:44 +0800)]
ceph: don't enable rbytes mount option by default

When rbytes mount option is enabled, directory size is recursive
size. Recursive size is not updated instantly. This can cause
directory size to change between successive stat(1)

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agoceph: encode ctime in cap message
Yan, Zheng [Fri, 22 Jan 2016 10:30:12 +0000 (18:30 +0800)]
ceph: encode ctime in cap message

Signed-off-by: Yan, Zheng <zyan@redhat.com>
8 years agolibceph: behave in mon_fault() if cur_mon < 0
Ilya Dryomov [Sat, 23 Jan 2016 14:57:51 +0000 (15:57 +0100)]
libceph: behave in mon_fault() if cur_mon < 0

This can happen if __close_session() in ceph_monc_stop() races with
a connection reset.  We need to ignore such faults, otherwise it's
likely we would take !hunting, call __schedule_delayed() and end up
with delayed_work() executing on invalid memory, among other things.

The (two!) con->private tests are useless, as nothing ever clears
con->private.  Nuke them.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: reschedule tick in mon_fault()
Ilya Dryomov [Fri, 22 Jan 2016 17:42:17 +0000 (18:42 +0100)]
libceph: reschedule tick in mon_fault()

Doing __schedule_delayed() in the hunting branch is pointless, as the
tick will have already been scheduled by then.

What we need to do instead is *reschedule* it in the !hunting branch,
after reopen_session() changes hunt_mult, which affects the delay.
This helps with spacing out connection attempts and avoiding things
like two back-to-back attempts followed by a longer period of waiting
around.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: introduce and switch to reopen_session()
Ilya Dryomov [Thu, 21 Jan 2016 15:33:25 +0000 (16:33 +0100)]
libceph: introduce and switch to reopen_session()

hunting is now set in __open_session() and cleared in finish_hunting(),
instead of all around.  The "session lost" message is printed not only
on connection resets, but also on keepalive timeouts.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: monc hunt rate is 3s with backoff up to 30s
Ilya Dryomov [Thu, 21 Jan 2016 15:33:19 +0000 (16:33 +0100)]
libceph: monc hunt rate is 3s with backoff up to 30s

Unless we are in the process of setting up a client (i.e. connecting to
the monitor cluster for the first time), apply a backoff: every time we
want to reopen a session, increase our timeout by a multiple (currently
2); when we complete the connection, reduce that multipler by 50%.

Mirrors ceph.git commit 794c86fd289bd62a35ed14368fa096c46736e9a2.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: monc ping rate is 10s
Ilya Dryomov [Thu, 21 Jan 2016 15:33:15 +0000 (16:33 +0100)]
libceph: monc ping rate is 10s

Split ping interval and ping timeout: ping interval is 10s; keepalive
timeout is 30s.

Make monc_ping_timeout a constant while at it - it's not actually
exported as a mount option (and the rest of tick-related settings won't
be either), so it's got no place in ceph_options.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: pick a different monitor when reconnecting
Ilya Dryomov [Wed, 20 Jan 2016 16:50:31 +0000 (17:50 +0100)]
libceph: pick a different monitor when reconnecting

Don't try to reconnect to the same monitor when we fail to establish
a session within a timeout or it's lost.

For that, pick_new_mon() needs to see the old value of cur_mon, so
don't clear it in __close_session() - all calls to __close_session()
but one are followed by __open_session() anyway.  __open_session() is
only called when a new session needs to be established, so the "already
open?" branch, which is now in the way, is simply dropped.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: revamp subs code, switch to SUBSCRIBE2 protocol
Ilya Dryomov [Tue, 19 Jan 2016 15:19:06 +0000 (16:19 +0100)]
libceph: revamp subs code, switch to SUBSCRIBE2 protocol

It is currently hard-coded in the mon_client that mdsmap and monmap
subs are continuous, while osdmap sub is always "onetime".  To better
handle full clusters/pools in the osd_client, we need to be able to
issue continuous osdmap subs.  Revamp subs code to allow us to specify
for each sub whether it should be continuous or not.

Although not strictly required for the above, switch to SUBSCRIBE2
protocol while at it, eliminating the ambiguity between a request for
"every map since X" and a request for "just the latest" when we don't
have a map yet (i.e. have epoch 0).  SUBSCRIBE2 feature bit is now
required - it's been supported since pre-argonaut (2010).

Move "got mdsmap" call to the end of ceph_mdsc_handle_map() - calling
in before we validate the epoch and successfully install the new map
can mess up mon_client sub state.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: decouple hunting and subs management
Ilya Dryomov [Fri, 8 Jan 2016 18:17:22 +0000 (21:17 +0300)]
libceph: decouple hunting and subs management

Coupling hunting state with subscribe state is not a good idea.  Clear
hunting when we complete the authentication handshake.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
8 years agolibceph: move debugfs initialization into __ceph_open_session()
Ilya Dryomov [Wed, 6 Jan 2016 09:56:21 +0000 (12:56 +0300)]
libceph: move debugfs initialization into __ceph_open_session()

Our debugfs dir name is a concatenation of cluster fsid and client
unique ID ("global_id").  It used to be the case that we learned
global_id first, nowadays we always learn fsid first - the monmap is
sent before any auth replies are.  ceph_debugfs_client_init() call in
ceph_monc_handle_map() is therefore never executed and can be removed.

Its counterpart in handle_auth_reply() doesn't really belong there
either: having to do monc->client and unlocking early to work around
lockdep is a testament to that.  Move it into __ceph_open_session(),
where it can be called unconditionally.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>