Merge tag 'xfs-for-linus-4.3' of git://git./linux/kernel/git/dgc/linux-xfs

Pull xfs updates from Dave Chinner:
 "There isn't a whole lot to this update - it's mostly bug fixes and
  they are spread pretty much all over XFS.  There are some corruption
  fixes, some fixes for log recovery, some fixes that prevent unount
  from hanging, a lockdep annotation rework for inode locking to prevent
  false positives and the usual random bunch of cleanups and minor
  improvements.

  Deatils:

   - large rework of EFI/EFD lifecycle handling to fix log recovery
     corruption issues, crashes and unmount hangs

   - separate metadata UUID on disk to enable changing boot label UUID
     for v5 filesystems

   - fixes for gcc miscompilation on certain platforms and optimisation
     levels

   - remote attribute allocation and recovery corruption fixes

   - inode lockdep annotation rework to fix bugs with too many
     subclasses

   - directory inode locking changes to prevent lockdep false positives

   - a handful of minor corruption fixes

   - various other small cleanups and bug fixes"

* tag 'xfs-for-linus-4.3' of git://git.kernel.org/pub/scm/linux/kernel/git/dgc/linux-xfs: (42 commits)
  xfs: fix error gotos in xfs_setattr_nonsize
  xfs: add mssing inode cache attempts counter increment
  xfs: return errors from partial I/O failures to files
  libxfs: bad magic number should set da block buffer error
  xfs: fix non-debug build warnings
  xfs: collapse allocsize and biosize mount option handling
  xfs: Fix file type directory corruption for btree directories
  xfs: lockdep annotations throw warnings on non-debug builds
  xfs: Fix uninitialized return value in xfs_alloc_fix_freelist()
  xfs: inode lockdep annotations broke non-lockdep build
  xfs: flush entire file on dio read/write to cached file
  xfs: Fix xfs_attr_leafblock definition
  libxfs: readahead of dir3 data blocks should use the read verifier
  xfs: stop holding ILOCK over filldir callbacks
  xfs: clean up inode lockdep annotations
  xfs: swap leaf buffer into path struct atomically during path shift
  xfs: relocate sparse inode mount warning
  xfs: dquots should be stamped with sb_meta_uuid
  xfs: log recovery needs to validate against sb_meta_uuid
  xfs: growfs not aware of sb_meta_uuid
  ...

Merge branch 'for-linus' of git://git./linux/kernel/git/viro/vfs

Pull vfs updates from Al Viro:
 "In this one:

   - d_move fixes (Eric Biederman)

   - UFS fixes (me; locking is mostly sane now, a bunch of bugs in error
     handling ought to be fixed)

   - switch of sb_writers to percpu rwsem (Oleg Nesterov)

   - superblock scalability (Josef Bacik and Dave Chinner)

   - swapon(2) race fix (Hugh Dickins)"

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (65 commits)
  vfs: Test for and handle paths that are unreachable from their mnt_root
  dcache: Reduce the scope of i_lock in d_splice_alias
  dcache: Handle escaped paths in prepend_path
  mm: fix potential data race in SyS_swapon
  inode: don't softlockup when evicting inodes
  inode: rename i_wb_list to i_io_list
  sync: serialise per-superblock sync operations
  inode: convert inode_sb_list_lock to per-sb
  inode: add hlist_fake to avoid the inode hash lock in evict
  writeback: plug writeback at a high level
  change sb_writers to use percpu_rw_semaphore
  shift percpu_counter_destroy() into destroy_super_work()
  percpu-rwsem: kill CONFIG_PERCPU_RWSEM
  percpu-rwsem: introduce percpu_rwsem_release() and percpu_rwsem_acquire()
  percpu-rwsem: introduce percpu_down_read_trylock()
  document rwsem_release() in sb_wait_write()
  fix the broken lockdep logic in __sb_start_write()
  introduce __sb_writers_{acquired,release}() helpers
  ufs_inode_get{frag,block}(): get rid of 'phys' argument
  ufs_getfrag_block(): tidy up a bit
  ...

Merge tag 'for-linus-4.3-merge-window-part-1' of git://git./linux/kernel/git/ericvh/v9fs

Pull 9p updates from Eric Van Hensbergen:
 "Just a few cleanups for 4.3 merge window for the 9p file system.  I've
  gotten several more over the past week, but this group has been in
  for-next for at least a couple of weeks so I figured I'd push them
  first while I test the rest.

  Most of the ones not in this set are bug-fixes anyways so I could hold
  them for rc1"

* tag 'for-linus-4.3-merge-window-part-1' of git://git.kernel.org/pub/scm/linux/kernel/git/ericvh/v9fs:
  9p: fix return code of read() when count is 0
  9p: remove unused option Opt_trans

Merge tag 'media/v4.3-1' of git://git./linux/kernel/git/mchehab/linux-media

Pull media updates from Mauro Carvalho Chehab:
 - new DVB frontend drivers: ascot2e, cxd2841er, horus3a, lnbh25
 - new HDMI capture driver: tc358743
 - new driver for NetUP DVB new boards (netup_unidvb)
 - IR support for DVBSky cards (smipcie-ir)
 - Coda driver has gain macroblock tiling support
 - Renesas R-Car gains JPEG codec driver
 - new DVB platform driver for STi boards: c8sectpfe
 - added documentation for the media core kABI to device-drivers DocBook
 - lots of driver fixups, cleanups and improvements

* tag 'media/v4.3-1' of git://git.kernel.org/pub/scm/linux/kernel/git/mchehab/linux-media: (297 commits)
  [media] c8sectpfe: Remove select on undefined LIBELF_32
  [media] i2c: fix platform_no_drv_owner.cocci warnings
  [media] cx231xx: Use wake_up_interruptible() instead of wake_up_interruptible_nr()
  [media] tc358743: only queue subdev notifications if devnode is set
  [media] tc358743: add missing Kconfig dependency/select
  [media] c8sectpfe: Use %pad to print 'dma_addr_t'
  [media] DocBook media: Fix typo "the the" in xml files
  [media] tc358743: make reset gpio optional
  [media] tc358743: set direction of reset gpio using devm_gpiod_get
  [media] dvbdev: document most of the functions/data structs
  [media] dvb_frontend.h: document the struct dvb_frontend
  [media] dvb-frontend.h: document struct dtv_frontend_properties
  [media] dvb-frontend.h: document struct dvb_frontend_ops
  [media] dvb: Use DVBFE_ALGO_HW where applicable
  [media] dvb_frontend.h: document struct analog_demod_ops
  [media] dvb_frontend.h: Document struct dvb_tuner_ops
  [media] Docbook: Document struct analog_parameters
  [media] dvb_frontend.h: get rid of dvbfe_modcod
  [media] add documentation for struct dvb_tuner_info
  [media] dvb_frontend: document dvb_frontend_tune_settings
  ...

Merge branch 'mailbox-for-next' of git://git.linaro.org/landing-teams/working/fujitsu/integration

Pull mailbox updates from Jassi Brar:
 "Mainly we move from jiffy based timer to HRTIMER for finer control
  over polling.  Then a controller reduces its polling period from 10 to
  1ms"

* 'mailbox-for-next' of git://git.linaro.org/landing-teams/working/fujitsu/integration:
  mailbox: arm_mhu: reduce txpoll_period from 10ms to 1 ms
  mailbox: switch to hrtimer for tx_complete polling
  mailbox: Drop owner assignment from platform_driver

Merge tag 'md/4.3' of git://neil.brown.name/md

Pull md updates from Neil Brown:

 - an assortment of little fixes, several for minor races only likely to
   be hit during testing

 - further cluster-md-raid1 development, not ready for real use yet.

 - new RAID6 syndrome code for ARM NEON

 - fix a race where a write can return before failure of one device is
   properly recorded in metadata, so an immediate crash might result in
   that write being lost.

* tag 'md/4.3' of git://neil.brown.name/md: (33 commits)
  md/raid5: ensure device failure recorded before write request returns.
  md/raid5: use bio_list for the list of bios to return.
  md/raid10: ensure device failure recorded before write request returns.
  md/raid1: ensure device failure recorded before write request returns.
  md-cluster: remove inappropriate try_module_get from join()
  md: extend spinlock protection in register_md_cluster_operations
  md-cluster: Read the disk bitmap sb and check if it needs recovery
  md-cluster: only call complete(&cinfo->completion) when node join cluster
  md-cluster: add missed lockres_free
  md-cluster: remove the unused sb_lock
  md-cluster: init suspend_list and suspend_lock early in join
  md-cluster: add the error check if failed to get dlm lock
  md-cluster: init completion within lockres_init
  md-cluster: fix deadlock issue on message lock
  md-cluster: transfer the resync ownership to another node
  md-cluster: split recover_slot for future code reuse
  md-cluster: use %pU to print UUIDs
  md: setup safemode_timer before it's being used
  md/raid5: handle possible race as reshape completes.
  md: sync sync_completed has correct value as recovery finishes.
  ...

Merge tag 'nfsd-4.3' of git://linux-nfs.org/~bfields/linux

Pull nfsd updates from Bruce Fields:
 "Nothing major, but:

   - Add Jeff Layton as an nfsd co-maintainer: no change to existing
     practice, just an acknowledgement of the status quo.

   - Two patches ("nfsd: ensure that...") for a race overlooked by the
     state locking rewrite, causing a crash noticed by multiple users.

   - Lots of smaller bugfixes all over from Kinglong Mee.

   - From Jeff, some cleanup of server rpc code in preparation for
     possible shift of nfsd threads to workqueues"

* tag 'nfsd-4.3' of git://linux-nfs.org/~bfields/linux: (52 commits)
  nfsd: deal with DELEGRETURN racing with CB_RECALL
  nfsd: return CLID_INUSE for unexpected SETCLIENTID_CONFIRM case
  nfsd: ensure that delegation stateid hash references are only put once
  nfsd: ensure that the ol stateid hash reference is only put once
  net: sunrpc: fix tracepoint Warning: unknown op '->'
  nfsd: allow more than one laundry job to run at a time
  nfsd: don't WARN/backtrace for invalid container deployment.
  fs: fix fs/locks.c kernel-doc warning
  nfsd: Add Jeff Layton as co-maintainer
  NFSD: Return word2 bitmask if setting security label in OPEN/CREATE
  NFSD: Set the attributes used to store the verifier for EXCLUSIVE4_1
  nfsd: SUPPATTR_EXCLCREAT must be encoded before SECURITY_LABEL.
  nfsd: Fix an FS_LAYOUT_TYPES/LAYOUT_TYPES encode bug
  NFSD: Store parent's stat in a separate value
  nfsd: Fix two typos in comments
  lockd: NLM grace period shouldn't block NFSv4 opens
  nfsd: include linux/nfs4.h in export.h
  sunrpc: Switch to using hash list instead single list
  sunrpc/nfsd: Remove redundant code by exports seq_operations functions
  sunrpc: Store cache_detail in seq_file's private directly
  ...

Merge branch 'for-linus-4.3' of git://git./linux/kernel/git/mason/linux-btrfs

Pull btrfs updates from Chris Mason:
 "This has Jeff Mahoney's long standing trim patch that fixes corners
  where trims were missing.  Omar has some raid5/6 fixes, especially for
  using scrub and device replace when devices are missing.

  Zhao Lie continues cleaning and fixing things, this series fixes some
  really hard to hit corners in xfstests.  I had to pull it last merge
  window due to some deadlocks, but those are now resolved.

  I added support for Tejun's new blkio controllers.  It seems to work
  well for single devices, we'll expand to multi-device as well"

* 'for-linus-4.3' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (47 commits)
  btrfs: fix compile when block cgroups are not enabled
  Btrfs: fix file read corruption after extent cloning and fsync
  Btrfs: check if previous transaction aborted to avoid fs corruption
  btrfs: use __GFP_NOFAIL in alloc_btrfs_bio
  btrfs: Prevent from early transaction abort
  btrfs: Remove unused arguments in tree-log.c
  btrfs: Remove useless condition in start_log_trans()
  Btrfs: add support for blkio controllers
  Btrfs: remove unused mutex from struct 'btrfs_fs_info'
  Btrfs: fix parity scrub of RAID 5/6 with missing device
  Btrfs: fix device replace of a missing RAID 5/6 device
  Btrfs: add RAID 5/6 BTRFS_RBIO_REBUILD_MISSING operation
  Btrfs: count devices correctly in readahead during RAID 5/6 replace
  Btrfs: remove misleading handling of missing device scrub
  btrfs: fix clone / extent-same deadlocks
  Btrfs: fix defrag to merge tail file extent
  Btrfs: fix warning in backref walking
  btrfs: Add WARN_ON() for double lock in btrfs_tree_lock()
  btrfs: Remove root argument in extent_data_ref_count()
  btrfs: Fix wrong comment of btrfs_alloc_tree_block()
  ...

Merge branch 'akpm' (patches from Andrew)

Merge patch-bomb from Andrew Morton:

 - a few misc things

 - Andy's "ambient capabilities"

 - fs/nofity updates

 - the ocfs2 queue

 - kernel/watchdog.c updates and feature work.

 - some of MM.  Includes Andrea's userfaultfd feature.

[ Hadn't noticed that userfaultfd was 'default y' when applying the
  patches, so that got fixed in this merge instead.  We do _not_ mark
  new features that nobody uses yet 'default y'   - Linus ]

* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (118 commits)
  mm/hugetlb.c: make vma_has_reserves() return bool
  mm/madvise.c: make madvise_behaviour_valid() return bool
  mm/memory.c: make tlb_next_batch() return bool
  mm/dmapool.c: change is_page_busy() return from int to bool
  mm: remove struct node_active_region
  mremap: simplify the "overlap" check in mremap_to()
  mremap: don't do uneccesary checks if new_len == old_len
  mremap: don't do mm_populate(new_addr) on failure
  mm: move ->mremap() from file_operations to vm_operations_struct
  mremap: don't leak new_vma if f_op->mremap() fails
  mm/hugetlb.c: make vma_shareable() return bool
  mm: make GUP handle pfn mapping unless FOLL_GET is requested
  mm: fix status code which move_pages() returns for zero page
  mm: memcontrol: bring back the VM_BUG_ON() in mem_cgroup_swapout()
  genalloc: add support of multiple gen_pools per device
  genalloc: add name arg to gen_pool_get() and devm_gen_pool_create()
  mm/memblock: WARN_ON when nid differs from overlap region
  Documentation/features/vm: add feature description and arch support status for batched TLB flush after unmap
  mm: defer flush of writable TLB entries
  mm: send one IPI per CPU to TLB flush all entries after unmapping pages
  ...

task_work: remove fifo ordering guarantee

In commit f341861fb0b ("task_work: add a scheduling point in
task_work_run()") I fixed a latency problem adding a cond_resched()
call.

Later, commit ac3d0da8f329 added yet another loop to reverse a list,
bringing back the latency spike :

I've seen in some cases this loop taking 275 ms, if for example a
process with 2,000,000 files is killed.

We could add yet another cond_resched() in the reverse loop, or we
can simply remove the reversal, as I do not think anything
would depend on order of task_work_add() submitted works.

Fixes: ac3d0da8f329 ("task_work: Make task_work_add() lockless")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Maciej Żenczykowski <maze@google.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Merge linux-block/for-4.3/core into md/for-linux

There were a few conflicts that are fairly easy to resolve.

Signed-off-by: NeilBrown <neilb@suse.com>

mm/hugetlb.c: make vma_has_reserves() return bool

This makes vma_has_reserves() return bool due to this particular function
only returning either one or zero as its return value.

Signed-off-by: Nicholas Krause <xerofoify@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/madvise.c: make madvise_behaviour_valid() return bool

This makes the madvise_bahaviour_valid() function return bool due to
this particular function always returning the value of either one or
zero as its return value.

Signed-off-by: Nicholas Krause <xerofoify@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memory.c: make tlb_next_batch() return bool

This makes the tlb_next_batch() bool due to this particular function only
ever returning either one or zero as its return value.

Signed-off-by: Nicholas Krause <xerofoify@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/dmapool.c: change is_page_busy() return from int to bool

This makes the function is_page_busy() return bool rather then an int now
due to this particular function's single return statement only ever
evaulating to either one or zero.

Signed-off-by: Nicholas Krause <xerofoify@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remove struct node_active_region

struct node_active_region is not used anymore.  Remove it.

Signed-off-by: minkyung88.kim <minkyung88.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mremap: simplify the "overlap" check in mremap_to()

Minor, but this check is overcomplicated.  Two half-intervals do NOT
overlap if END1 <= START2 || END2 <= START1, mremap_to() just needs to
negate this check.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mremap: don't do uneccesary checks if new_len == old_len

The "new_len > old_len" branch in vma_to_resize() looks very confusing.
It only covers the VM_DONTEXPAND/pgoff checks but everything below is
equally unneeded if new_len == old_len.

Change this code to return if "new_len == old_len", new_len < old_len is
not possible, otherwise the code below is wrong anyway.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mremap: don't do mm_populate(new_addr) on failure

move_vma() sets *locked even if move_page_tables() or ->mremap() fails,
change sys_mremap() to check "ret & ~PAGE_MASK".

I think we should simply remove the VM_LOCKED code in move_vma(), that is
why this patch doesn't change move_vma().  But this needs more cleanups.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: move ->mremap() from file_operations to vm_operations_struct

vma->vm_ops->mremap() looks more natural and clean in move_vma(), and this
way ->mremap() can have more users.  Say, vdso.

While at it, s/aio_ring_remap/aio_ring_mremap/.

Note: this is the minimal change before ->mremap() finds another user in
file_operations; this method should have more arguments, and it can be
used to kill arch_remap().

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mremap: don't leak new_vma if f_op->mremap() fails

move_vma() can't just return if f_op->mremap() fails, we should unmap the
new vma like we do if move_page_tables() fails.  To avoid the code
duplication this patch moves the "move entries back" under the new "if
(err)" branch.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/hugetlb.c: make vma_shareable() return bool

This makes vma_shareable() return bool now due to this particular function
only ever returning either one or zero as its return value.

Signed-off-by: Nicholas Krause <xerofoify@gmail.com>
Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: make GUP handle pfn mapping unless FOLL_GET is requested

With DAX, pfn mapping becoming more common.  The patch adjusts GUP code to
cover pfn mapping for cases when we don't need struct page to proceed.

To make it possible, let's change follow_page() code to return -EEXIST
error code if proper page table entry exists, but no corresponding struct
page.  __get_user_page() would ignore the error code and move to the next
page frame.

The immediate effect of the change is working MAP_POPULATE and mlock() on
DAX mappings.

[akpm@linux-foundation.org: fix arm64 build]
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reviewed-by: Toshi Kani <toshi.kani@hp.com>
Acked-by: Matthew Wilcox <willy@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fix status code which move_pages() returns for zero page

The manpage for move_pages(2) specifies that status code for zero page is
supposed to be -EFAULT.  Currently kernel return -ENOENT in this case.

follow_page() can do it for us, if we would ask for FOLL_DUMP.  The use of
FOLL_DUMP also means that the upper layer page tables pages are no longer
allocated.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: memcontrol: bring back the VM_BUG_ON() in mem_cgroup_swapout()

Clark stumbled over a VM_BUG_ON() in -RT which was then was removed by
Johannes in commit f371763a79d ("mm: memcontrol: fix false-positive
VM_BUG_ON() on -rt").  The comment before that patch was a tiny bit better
than it is now.  While the patch claimed to fix a false-postive on -RT
this was not the case.  None of the -RT folks ACKed it and it was not a
false positive report.  That was a *real* problem.

This patch updates the comment that is improper because it refers to
"disabled preemption" as a consequence of that lock being taken.  A
spin_lock() disables preemption, true, but in this case the code relies on
the fact that the lock _also_ disables interrupts once it is acquired.
And this is the important detail (which was checked the VM_BUG_ON()) which
needs to be pointed out.  This is the hint one needs while looking at the
code.  It was explained by Johannes on the list that the per-CPU variables
are protected by local_irq_save().  The BUG_ON() was helpful.  This code
has been workarounded in -RT in the meantime.  I wouldn't mind running
into more of those if the code in question uses *special* kind of locking
since now there is no verification (in terms of lockdep or BUG_ON()) and
therefore I bring the VM_BUG_ON() check back in.

The two functions after the comment could also have a "local_irq_save()"
dance around them in order to serialize access to the per-CPU variables.
This has been avoided because the interrupts should be off.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Clark Williams <williams@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

genalloc: add support of multiple gen_pools per device

This change fills devm_gen_pool_create()/gen_pool_get() "name" argument
stub with contents and extends of_gen_pool_get() functionality on this
basis.

If there is no associated platform device with a device node passed to
of_gen_pool_get(), the function attempts to get a label property or device
node name (= repeats MTD OF partition standard) and seeks for a named
gen_pool registered by device of the parent device node.

The main idea of the change is to allow registration of independent
gen_pools under the same umbrella device, say "partitions" on "storage
device", the original functionality of one "partition" per "storage
device" is untouched.

[akpm@linux-foundation.org: fix constness in devres_find()]
[dan.carpenter@oracle.com: freeing const data pointers]
Signed-off-by: Vladimir Zapolskiy <vladimir_zapolskiy@mentor.com>
Cc: Philipp Zabel <p.zabel@pengutronix.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Nicolas Ferre <nicolas.ferre@atmel.com>
Cc: Alexandre Belloni <alexandre.belloni@free-electrons.com>
Cc: Jean-Christophe Plagniol-Villard <plagnioj@jcrosoft.com>
Cc: Shawn Guo <shawnguo@kernel.org>
Cc: Sascha Hauer <kernel@pengutronix.de>
Cc: Mauro Carvalho Chehab <mchehab@osg.samsung.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

genalloc: add name arg to gen_pool_get() and devm_gen_pool_create()

This change modifies gen_pool_get() and devm_gen_pool_create() client
interfaces adding one more argument "name" of a gen_pool object.

Due to implementation gen_pool_get() is capable to retrieve only one
gen_pool associated with a device even if multiple gen_pools are created,
fortunately right at the moment it is sufficient for the clients, hence
provide NULL as a valid argument on both producer devm_gen_pool_create()
and consumer gen_pool_get() sides.

Because only one created gen_pool per device is addressable, explicitly
add a restriction to devm_gen_pool_create() to create only one gen_pool
per device, this implies two possible error codes returned by the
function, account it on client side (only misc/sram).  This completes
client side changes related to genalloc updates.

[akpm@linux-foundation.org: gen_pool_get() cleanup]
Signed-off-by: Vladimir Zapolskiy <vladimir_zapolskiy@mentor.com>
Cc: Philipp Zabel <p.zabel@pengutronix.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Nicolas Ferre <nicolas.ferre@atmel.com>
Cc: Alexandre Belloni <alexandre.belloni@free-electrons.com>
Cc: Jean-Christophe Plagniol-Villard <plagnioj@jcrosoft.com>
Cc: Shawn Guo <shawnguo@kernel.org>
Cc: Sascha Hauer <kernel@pengutronix.de>
Cc: Mauro Carvalho Chehab <mchehab@osg.samsung.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memblock: WARN_ON when nid differs from overlap region

Each memblock_region has nid to indicates the Node ID of this range.  For
the overlap case, memblock_add_range() inserts the lower part and leave
the upper part as indicated in the overlapped region.

If the nid of the new range differs from the overlapped region, the
information recorded is not correct.

This patch adds a WARN_ON when the nid of the new range differs from the
overlapped region.

Signed-off-by: Wei Yang <weiyang@linux.vnet.ibm.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Documentation/features/vm: add feature description and arch support status for batched TLB flush after unmap

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: defer flush of writable TLB entries

If a PTE is unmapped and it's dirty then it was writable recently.  Due to
deferred TLB flushing, it's best to assume a writable TLB cache entry
exists.  With that assumption, the TLB must be flushed before any IO can
start or the page is freed to avoid lost writes or data corruption.  This
patch defers flushing of potentially writable TLBs as long as possible.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: send one IPI per CPU to TLB flush all entries after unmapping pages

An IPI is sent to flush remote TLBs when a page is unmapped that was
potentially accesssed by other CPUs.  There are many circumstances where
this happens but the obvious one is kswapd reclaiming pages belonging to a
running process as kswapd and the task are likely running on separate
CPUs.

On small machines, this is not a significant problem but as machine gets
larger with more cores and more memory, the cost of these IPIs can be
high.  This patch uses a simple structure that tracks CPUs that
potentially have TLB entries for pages being unmapped.  When the unmapping
is complete, the full TLB is flushed on the assumption that a refill cost
is lower than flushing individual entries.

Architectures wishing to do this must give the following guarantee.

        If a clean page is unmapped and not immediately flushed, the
        architecture must guarantee that a write to that linear address
        from a CPU with a cached TLB entry will trap a page fault.

This is essentially what the kernel already depends on but the window is
much larger with this patch applied and is worth highlighting.  The
architecture should consider whether the cost of the full TLB flush is
higher than sending an IPI to flush each individual entry.  An additional
architecture helper called flush_tlb_local is required.  It's a trivial
wrapper with some accounting in the x86 case.

The impact of this patch depends on the workload as measuring any benefit
requires both mapped pages co-located on the LRU and memory pressure.  The
case with the biggest impact is multiple processes reading mapped pages
taken from the vm-scalability test suite.  The test case uses NR_CPU
readers of mapped files that consume 10*RAM.

Linear mapped reader on a 4-node machine with 64G RAM and 48 CPUs

                                           4.2.0-rc1          4.2.0-rc1
                                             vanilla       flushfull-v7
Ops lru-file-mmap-read-elapsed      159.62 (  0.00%)   120.68 ( 24.40%)
Ops lru-file-mmap-read-time_range    30.59 (  0.00%)     2.80 ( 90.85%)
Ops lru-file-mmap-read-time_stddv     6.70 (  0.00%)     0.64 ( 90.38%)

           4.2.0-rc1    4.2.0-rc1
             vanilla flushfull-v7
User          581.00       611.43
System       5804.93      4111.76
Elapsed       161.03       122.12

This is showing that the readers completed 24.40% faster with 29% less
system CPU time.  From vmstats, it is known that the vanilla kernel was
interrupted roughly 900K times per second during the steady phase of the
test and the patched kernel was interrupts 180K times per second.

The impact is lower on a single socket machine.

                                           4.2.0-rc1          4.2.0-rc1
                                             vanilla       flushfull-v7
Ops lru-file-mmap-read-elapsed       25.33 (  0.00%)    20.38 ( 19.54%)
Ops lru-file-mmap-read-time_range     0.91 (  0.00%)     1.44 (-58.24%)
Ops lru-file-mmap-read-time_stddv     0.28 (  0.00%)     0.47 (-65.34%)

           4.2.0-rc1    4.2.0-rc1
             vanilla flushfull-v7
User           58.09        57.64
System        111.82        76.56
Elapsed        27.29        22.55

It's still a noticeable improvement with vmstat showing interrupts went
from roughly 500K per second to 45K per second.

The patch will have no impact on workloads with no memory pressure or have
relatively few mapped pages.  It will have an unpredictable impact on the
workload running on the CPU being flushed as it'll depend on how many TLB
entries need to be refilled and how long that takes.  Worst case, the TLB
will be completely cleared of active entries when the target PFNs were not
resident at all.

[sasha.levin@oracle.com: trace tlb flush after disabling preemption in try_to_unmap_flush]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

x86, mm: trace when an IPI is about to be sent

When unmapping pages it is necessary to flush the TLB.  If that page was
accessed by another CPU then an IPI is used to flush the remote CPU.  That
is a lot of IPIs if kswapd is scanning and unmapping >100K pages per
second.

There already is a window between when a page is unmapped and when it is
TLB flushed.  This series increases the window so multiple pages can be
flushed using a single IPI.  This should be safe or the kernel is hosed
already.

Patch 1 simply made the rest of the series easier to write as ftrace
        could identify all the senders of TLB flush IPIS.

Patch 2 tracks what CPUs potentially map a PFN and then sends an IPI
        to flush the entire TLB.

Patch 3 tracks when there potentially are writable TLB entries that
        need to be batched differently

Patch 4 increases SWAP_CLUSTER_MAX to further batch flushes

The performance impact is documented in the changelogs but in the optimistic
case on a 4-socket machine the full series reduces interrupts from 900K
interrupts/second to 60K interrupts/second.

This patch (of 4):

It is easy to trace when an IPI is received to flush a TLB but harder to
detect what event sent it.  This patch makes it easy to identify the
source of IPIs being transmitted for TLB flushes on x86.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Dave Hansen <dave.hansen@intel.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: selftest

This test allocates two virtual areas and bounces the physical memory
across the two virtual areas using only userfaultfd.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Shuah Khan <shuah.kh@samsung.com>
Cc: Shuah Khan <shuahkh@osg.samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: avoid missing wakeups during refile in userfaultfd_read

During the refile in userfaultfd_read both waitqueues could look empty to
the lockless wake_userfault().  Use a seqcount to prevent this false
negative that could leave an userfault blocked.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: propagate the full address in THP faults

The THP faults were not propagating the original fault address.  The
latest version of the API with uffd.arg.pagefault.address is supposed to
propagate the full address through THP faults.

This was not a kernel crashing bug and it wouldn't risk to corrupt user
memory, but it would cause a SIGBUS failure because the wrong page was
being copied.

For various reasons this wasn't easily reproducible in the qemu workload,
but the strestest exposed the problem immediately.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: allow signals to interrupt a userfault

This is only simple to achieve if the userfault is going to return to
userland (not to the kernel) because we can avoid returning VM_FAULT_RETRY
despite we temporarily released the mmap_sem.  The fault would just be
retried by userland then.  This is safe at least on x86 and powerpc (the
two archs with the syscall implemented so far).

Hint to verify for which archs this is safe: after handle_mm_fault
returns, no access to data structures protected by the mmap_sem must be
done by the fault code in arch/*/mm/fault.c until up_read(&mm->mmap_sem)
is called.

This has two main benefits: signals can run with lower latency in
production (signals aren't blocked by userfaults and userfaults are
immediately repeated after signal processing) and gdb can then trivially
debug the threads blocked in this kind of userfaults coming directly from
userland.

On a side note: while gdb has a need to get signal processed, coredumps
always worked perfectly with userfaults, no matter if the userfault is
triggered by GUP a kernel copy_user or directly from userland.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: require UFFDIO_API before other ioctls

UFFDIO_API was already forced before read/poll could work.  This makes the
code more strict to force it also for all other ioctls.

All users would already have been required to call UFFDIO_API before
invoking other ioctls but this makes it more explicit.

This will ensure we can change all ioctls (all but UFFDIO_API/struct
uffdio_api) with a bump of uffdio_api.api.

There's no actual plan or need to change the API or the ioctl, the current
API already should cover fine even the non cooperative usage, but this is
just for the longer term future just in case.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: UFFDIO_COPY and UFFDIO_ZEROPAGE

These two ioctl allows to either atomically copy or to map zeropages
into the virtual address space. This is used by the thread that opened
the userfaultfd to resolve the userfaults.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: avoid mmap_sem read recursion in mcopy_atomic

If the rwsem starves writers it wasn't strictly a bug but lockdep
doesn't like it and this avoids depending on lowlevel implementation
details of the lock.

[akpm@linux-foundation.org: delete weird BUILD_BUG_ON()]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: mcopy_atomic|mfill_zeropage: UFFDIO_COPY|UFFDIO_ZEROPAGE preparation

This implements mcopy_atomic and mfill_zeropage that are the lowlevel
VM methods that are invoked respectively by the UFFDIO_COPY and
UFFDIO_ZEROPAGE userfaultfd commands.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: UFFDIO_COPY|UFFDIO_ZEROPAGE uAPI

This implements the uABI of UFFDIO_COPY and UFFDIO_ZEROPAGE.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: activate syscall

This activates the userfaultfd syscall.

[sfr@canb.auug.org.au: activate syscall fix]
[akpm@linux-foundation.org: don't enable userfaultfd on powerpc]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: buildsystem activation

This allows to select the userfaultfd during configuration to build it.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: solve the race between UFFDIO_COPY|ZEROPAGE and read

Solve in-kernel the race between UFFDIO_COPY|ZEROPAGE and
userfaultfd_read if they are run on different threads simultaneously.

Until now qemu solved the race in userland: the race was explicitly
and intentionally left for userland to solve. However we can also
solve it in kernel.

Requiring all users to solve this race if they use two threads (one
for the background transfer and one for the userfault reads) isn't
very attractive from an API prospective, furthermore this allows to
remove a whole bunch of mutex and bitmap code from qemu, making it
faster. The cost of __get_user_pages_fast should be insignificant
considering it scales perfectly and the pagetables are already hot in
the CPU cache, compared to the overhead in userland to maintain those
structures.

Applying this patch is backwards compatible with respect to the
userfaultfd userland API, however reverting this change wouldn't be
backwards compatible anymore.

Without this patch qemu in the background transfer thread, has to read
the old state, and do UFFDIO_WAKE if old_state is missing but it
become REQUESTED by the time it tries to set it to RECEIVED (signaling
the other side received an userfault).

    vcpu                background_thr userfault_thr
    -----               -----          -----
    vcpu0 handle_mm_fault()

                        postcopy_place_page
                        read old_state -> MISSING
                        UFFDIO_COPY 0x7fb76a139000 (no wakeup, still pending)

    vcpu0 fault at 0x7fb76a139000 enters handle_userfault
    poll() is kicked

                                        poll() -> POLLIN
                                        read() -> 0x7fb76a139000
                                        postcopy_pmi_change_state(MISSING, REQUESTED) -> REQUESTED

                        tmp_state = postcopy_pmi_change_state(old_state, RECEIVED) -> REQUESTED
                        /* check that no userfault raced with UFFDIO_COPY */
                        if (old_state == MISSING && tmp_state == REQUESTED)
                                UFFDIO_WAKE from background thread

And a second case where a UFFDIO_WAKE would be needed is in the userfault thread:

    vcpu                background_thr userfault_thr
    -----               -----          -----
    vcpu0 handle_mm_fault()

                        postcopy_place_page
                        read old_state -> MISSING
                        UFFDIO_COPY 0x7fb76a139000 (no wakeup, still pending)
                        tmp_state = postcopy_pmi_change_state(old_state, RECEIVED) -> RECEIVED

    vcpu0 fault at 0x7fb76a139000 enters handle_userfault
    poll() is kicked

                                        poll() -> POLLIN
                                        read() -> 0x7fb76a139000

                                        if (postcopy_pmi_change_state(MISSING, REQUESTED) == RECEIVED)
                                                UFFDIO_WAKE from userfault thread

This patch removes the need of both UFFDIO_WAKE and of the associated
per-page tristate as well.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: allocate the userfaultfd_ctx cacheline aligned

Use proper slab to guarantee alignment.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: optimize read() and poll() to be O(1)

This makes read O(1) and poll that was already O(1) becomes lockless.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: wake pending userfaults

This is an optimization but it's a userland visible one and it affects
the API.

The downside of this optimization is that if you call poll() and you
get POLLIN, read(ufd) may still return -EAGAIN. The blocked userfault
may be waken by a different thread, before read(ufd) comes
around. This in short means that poll() isn't really usable if the
userfaultfd is opened in blocking mode.

userfaults won't wait in "pending" state to be read anymore and any
UFFDIO_WAKE or similar operations that has the objective of waking
userfaults after their resolution, will wake all blocked userfaults
for the resolved range, including those that haven't been read() by
userland yet.

The behavior of poll() becomes not standard, but this obviates the
need of "spurious" UFFDIO_WAKE and it lets the userland threads to
restart immediately without requiring an UFFDIO_WAKE. This is even
more significant in case of repeated faults on the same address from
multiple threads.

This optimization is justified by the measurement that the number of
spurious UFFDIO_WAKE accounts for 5% and 10% of the total
userfaults for heavy workloads, so it's worth optimizing those away.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: change the read API to return a uffd_msg

I had requests to return the full address (not the page aligned one) to
userland.

It's not entirely clear how the page offset could be relevant because
userfaults aren't like SIGBUS that can sigjump to a different place and it
actually skip resolving the fault depending on a page offset.  There's
currently no real way to skip the fault especially because after a
UFFDIO_COPY|ZEROPAGE, the fault is optimized to be retried within the
kernel without having to return to userland first (not even self modifying
code replacing the .text that touched the faulting address would prevent
the fault to be repeated).  Userland cannot skip repeating the fault even
more so if the fault was triggered by a KVM secondary page fault or any
get_user_pages or any copy-user inside some syscall which will return to
kernel code.  The second time FAULT_FLAG_RETRY_NOWAIT won't be set leading
to a SIGBUS being raised because the userfault can't wait if it cannot
release the mmap_map first (and FAULT_FLAG_RETRY_NOWAIT is required for
that).

Still returning userland a proper structure during the read() on the uffd,
can allow to use the current UFFD_API for the future non-cooperative
extensions too and it looks cleaner as well.  Once we get additional
fields there's no point to return the fault address page aligned anymore
to reuse the bits below PAGE_SHIFT.

The only downside is that the read() syscall will read 32bytes instead of
8bytes but that's not going to be measurable overhead.

The total number of new events that can be extended or of new future bits
for already shipped events, is limited to 64 by the features field of the
uffdio_api structure.  If more will be needed a bump of UFFD_API will be
required.

[akpm@linux-foundation.org: use __packed]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: Rename uffd_api.bits into .features

This is (seems to be) the minimal thing that is required to unblock
standard uffd usage from the non-cooperative one.  Now more bits can be
added to the features field indicating e.g.  UFFD_FEATURE_FORK and others
needed for the latter use-case.

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: add new syscall to provide memory externalization

Once an userfaultfd has been created and certain region of the process
virtual address space have been registered into it, the thread responsible
for doing the memory externalization can manage the page faults in
userland by talking to the kernel using the userfaultfd protocol.

poll() can be used to know when there are new pending userfaults to be
read (POLLIN).

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: prevent khugepaged to merge if userfaultfd is armed

If userfaultfd is armed on a certain vma we can't "fill" the holes with
zeroes or we'll break the userland on demand paging.  The holes if the
userfault is armed, are really missing information (not zeroes) that the
userland has to load from network or elsewhere.

The same issue happens for wrprotected ptes that we can't just convert
into a single writable pmd_trans_huge.

We could however in theory still merge across zeropages if only
VM_UFFD_MISSING is set (so if VM_UFFD_WP is not set)...  that could be
slightly improved but it'd be much more complex code for a tiny corner
case.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: teach vma_merge to merge across vma->vm_userfaultfd_ctx

vma->vm_userfaultfd_ctx is yet another vma parameter that vma_merge
must be aware about so that we can merge vmas back like they were
originally before arming the userfaultfd on some memory range.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: call handle_userfault() for userfaultfd_missing() faults

This is where the page faults must be modified to call
handle_userfault() if userfaultfd_missing() is true (so if the
vma->vm_flags had VM_UFFD_MISSING set).

handle_userfault() then takes care of blocking the page fault and
delivering it to userland.

The fault flags must also be passed as parameter so the "read|write"
kind of fault can be passed to userland.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: add VM_UFFD_MISSING and VM_UFFD_WP

These two flags gets set in vma->vm_flags to tell the VM common code
if the userfaultfd is armed and in which mode (only tracking missing
faults, only tracking wrprotect faults or both). If neither flags is
set it means the userfaultfd is not armed on the vma.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: add vm_userfaultfd_ctx to the vm_area_struct

This adds the vm_userfaultfd_ctx to the vm_area_struct.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: linux/userfaultfd_k.h

Kernel header defining the methods needed by the VM common code to
interact with the userfaultfd.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: uAPI

Defines the uAPI of the userfaultfd, notably the ioctl numbers and protocol.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: waitqueue: add nr wake parameter to __wake_up_locked_key

userfaultfd needs to wake all waitqueues (pass 0 as nr parameter), instead
of the current hardcoded 1 (that would wake just the first waitqueue in
the head list).

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

userfaultfd: linux/Documentation/vm/userfaultfd.txt

This is the latest userfaultfd patchset.  The postcopy live migration
feature on the qemu side is mostly ready to be merged and it entirely
depends on the userfaultfd syscall to be merged as well.  So it'd be great
if this patchset could be reviewed for merging in -mm.

Userfaults allow to implement on demand paging from userland and more
generally they allow userland to more efficiently take control of the
behavior of page faults than what was available before (PROT_NONE +
SIGSEGV trap).

The use cases are:

1) KVM postcopy live migration (one form of cloud memory
   externalization).

   KVM postcopy live migration is the primary driver of this work:

    http://blog.zhaw.ch/icclab/setting-up-post-copy-live-migration-in-openstack/
    http://lists.gnu.org/archive/html/qemu-devel/2015-02/msg04873.html

2) postcopy live migration of binaries inside linux containers:

    http://thread.gmane.org/gmane.linux.kernel.mm/132662

3) KVM postcopy live snapshotting (allowing to limit/throttle the
   memory usage, unlike fork would, plus the avoidance of fork
   overhead in the first place).

   While the wrprotect tracking is not implemented yet, the syscall API is
   already contemplating the wrprotect fault tracking and it's generic enough
   to allow its later implementation in a backwards compatible fashion.

4) KVM userfaults on shared memory. The UFFDIO_COPY lowlevel method
   should be extended to work also on tmpfs and then the
   uffdio_register.ioctls will notify userland that UFFDIO_COPY is
   available even when the registered virtual memory range is tmpfs
   backed.

5) alternate mechanism to notify web browsers or apps on embedded
   devices that volatile pages have been reclaimed. This basically
   avoids the need to run a syscall before the app can access with the
   CPU the virtual regions marked volatile. This depends on point 4)
   to be fulfilled first, as volatile pages happily apply to tmpfs.

Even though there wasn't a real use case requesting it yet, it also
allows to implement distributed shared memory in a way that readonly
shared mappings can exist simultaneously in different hosts and they
can be become exclusive at the first wrprotect fault.

This patch (of 22):

Add documentation.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Cc: zhang.zhanghailiang@huawei.com
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Huangpeng (Peter)" <peter.huangpeng@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/slab.h: fix argument order in cache_from_obj's error message

While debugging a networking issue, I hit a condition that triggered an
object to be freed into the wrong kmem cache, and thus triggered the
warning in cache_from_obj().

The arguments in the error message are in wrong order: the location
of the object's kmem cache is in cachep, not s.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/slub: don't wait for high-order page allocation

Description is almost copied from commit fb05e7a89f50 ("net: don't wait
for order-3 page allocation").

I saw excessive direct memory reclaim/compaction triggered by slub.  This
causes performance issues and add latency.  Slub uses high-order
allocation to reduce internal fragmentation and management overhead.  But,
direct memory reclaim/compaction has high overhead and the benefit of
high-order allocation can't compensate the overhead of both work.

This patch makes auxiliary high-order allocation atomic.  If there is no
memory pressure and memory isn't fragmented, the alloction will still
success, so we don't sacrifice high-order allocation's benefit here.  If
the atomic allocation fails, direct memory reclaim/compaction will not be
triggered, allocation fallback to low-order immediately, hence the direct
memory reclaim/compaction overhead is avoided.  In the allocation failure
case, kswapd is waken up and trying to make high-order freepages, so
allocation could success next time.

Following is the test to measure effect of this patch.

System: QEMU, CPU 8, 512 MB
Mem: 25% memory is allocated at random position to make fragmentation.
 Memory-hogger occupies 150 MB memory.
Workload: hackbench -g 20 -l 1000

Average result by 10 runs (Base va Patched)

elapsed_time(s): 4.3468 vs 2.9838
compact_stall: 461.7 vs 73.6
pgmigrate_success: 28315.9 vs 7256.1

Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/slub: fix slab double-free in case of duplicate sysfs filename

sysfs_slab_add() shouldn't call kobject_put at error path: this puts last
reference of kmem-cache kobject and frees it.  Kmem cache will be freed
second time at error path in kmem_cache_create().

For example this happens when slub debug was enabled in runtime and
somebody creates new kmem cache:

# echo 1 | tee /sys/kernel/slab/*/sanity_checks
# modprobe configfs

"configfs_dir_cache" cannot be merged because existing slab have debug and
cannot create new slab because unique name ":t-0000096" already taken.

Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/slub: move slab initialization into irq enabled region

Initializing a new slab can introduce rather large latencies because most
of the initialization runs always with interrupts disabled.

There is no point in doing so.  The newly allocated slab is not visible
yet, so there is no reason to protect it against concurrent alloc/free.

Move the expensive parts of the initialization into allocate_slab(), so
for all allocations with GFP_WAIT set, interrupts are enabled.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

slub: add support for kmem_cache_debug in bulk calls

Per request of Joonsoo Kim adding kmem debug support.

I've tested that when debugging is disabled, then there is almost no
performance impact as this code basically gets removed by the compiler.

Need some guidance in enabling and testing this.

bulk- PREVIOUS                  - THIS-PATCH
  1 -  43 cycles(tsc) 10.811 ns -  44 cycles(tsc) 11.236 ns  improved  -2.3%
  2 -  27 cycles(tsc)  6.867 ns -  28 cycles(tsc)  7.019 ns  improved  -3.7%
  3 -  21 cycles(tsc)  5.496 ns -  22 cycles(tsc)  5.526 ns  improved  -4.8%
  4 -  24 cycles(tsc)  6.038 ns -  19 cycles(tsc)  4.786 ns  improved  20.8%
  8 -  17 cycles(tsc)  4.280 ns -  18 cycles(tsc)  4.572 ns  improved  -5.9%
 16 -  17 cycles(tsc)  4.483 ns -  18 cycles(tsc)  4.658 ns  improved  -5.9%
 30 -  18 cycles(tsc)  4.531 ns -  18 cycles(tsc)  4.568 ns  improved   0.0%
 32 -  58 cycles(tsc) 14.586 ns -  65 cycles(tsc) 16.454 ns  improved -12.1%
 34 -  53 cycles(tsc) 13.391 ns -  63 cycles(tsc) 15.932 ns  improved -18.9%
 48 -  65 cycles(tsc) 16.268 ns -  50 cycles(tsc) 12.506 ns  improved  23.1%
 64 -  53 cycles(tsc) 13.440 ns -  63 cycles(tsc) 15.929 ns  improved -18.9%
128 -  79 cycles(tsc) 19.899 ns -  86 cycles(tsc) 21.583 ns  improved  -8.9%
158 -  90 cycles(tsc) 22.732 ns -  90 cycles(tsc) 22.552 ns  improved   0.0%
250 -  95 cycles(tsc) 23.916 ns -  98 cycles(tsc) 24.589 ns  improved  -3.2%

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

slub: initial bulk free implementation

This implements SLUB specific kmem_cache_free_bulk().  SLUB allocator now
both have bulk alloc and free implemented.

Choose to reenable local IRQs while calling slowpath __slab_free().  In
worst case, where all objects hit slowpath call, the performance should
still be faster than fallback function __kmem_cache_free_bulk(), because
local_irq_{disable+enable} is very fast (7-cycles), while the fallback
invokes this_cpu_cmpxchg() which is slightly slower (9-cycles).
Nitpicking, this should be faster for N>=4, due to the entry cost of
local_irq_{disable+enable}.

Do notice that the save+restore variant is very expensive, this is key to
why this optimization works.

CPU: i7-4790K CPU @ 4.00GHz
 * local_irq_{disable,enable}:  7 cycles(tsc) - 1.821 ns
 * local_irq_{save,restore}  : 37 cycles(tsc) - 9.443 ns

Measurements on CPU CPU i7-4790K @ 4.00GHz
Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns

Bulk- fallback                   - this-patch
  1 -  58 cycles(tsc) 14.542 ns  -  43 cycles(tsc) 10.811 ns  improved 25.9%
  2 -  50 cycles(tsc) 12.659 ns  -  27 cycles(tsc)  6.867 ns  improved 46.0%
  3 -  48 cycles(tsc) 12.168 ns  -  21 cycles(tsc)  5.496 ns  improved 56.2%
  4 -  47 cycles(tsc) 11.987 ns  -  24 cycles(tsc)  6.038 ns  improved 48.9%
  8 -  46 cycles(tsc) 11.518 ns  -  17 cycles(tsc)  4.280 ns  improved 63.0%
 16 -  45 cycles(tsc) 11.366 ns  -  17 cycles(tsc)  4.483 ns  improved 62.2%
 30 -  45 cycles(tsc) 11.433 ns  -  18 cycles(tsc)  4.531 ns  improved 60.0%
 32 -  75 cycles(tsc) 18.983 ns  -  58 cycles(tsc) 14.586 ns  improved 22.7%
 34 -  71 cycles(tsc) 17.940 ns  -  53 cycles(tsc) 13.391 ns  improved 25.4%
 48 -  80 cycles(tsc) 20.077 ns  -  65 cycles(tsc) 16.268 ns  improved 18.8%
 64 -  71 cycles(tsc) 17.799 ns  -  53 cycles(tsc) 13.440 ns  improved 25.4%
128 -  91 cycles(tsc) 22.980 ns  -  79 cycles(tsc) 19.899 ns  improved 13.2%
158 - 100 cycles(tsc) 25.241 ns  -  90 cycles(tsc) 22.732 ns  improved 10.0%
250 - 102 cycles(tsc) 25.583 ns  -  95 cycles(tsc) 23.916 ns  improved  6.9%

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

slub: improve bulk alloc strategy

Call slowpath __slab_alloc() from within the bulk loop, as the side-effect
of this call likely repopulates c->freelist.

Choose to reenable local IRQs while calling slowpath.

Saving some optimizations for later.  E.g.  it is possible to extract
parts of __slab_alloc() and avoid the unnecessary and expensive (37
cycles) local_irq_{save,restore}.  For now, be happy calling
__slab_alloc() this lower icache impact of this func and I don't have to
worry about correctness.

Measurements on CPU CPU i7-4790K @ 4.00GHz
Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns

Bulk- fallback                   - this-patch
  1 -  58 cycles(tsc) 14.516 ns  -  49 cycles(tsc) 12.459 ns  improved 15.5%
  2 -  51 cycles(tsc) 12.930 ns  -  38 cycles(tsc)  9.605 ns  improved 25.5%
  3 -  49 cycles(tsc) 12.274 ns  -  34 cycles(tsc)  8.525 ns  improved 30.6%
  4 -  48 cycles(tsc) 12.058 ns  -  32 cycles(tsc)  8.036 ns  improved 33.3%
  8 -  46 cycles(tsc) 11.609 ns  -  31 cycles(tsc)  7.756 ns  improved 32.6%
 16 -  45 cycles(tsc) 11.451 ns  -  32 cycles(tsc)  8.148 ns  improved 28.9%
 30 -  79 cycles(tsc) 19.865 ns  -  68 cycles(tsc) 17.164 ns  improved 13.9%
 32 -  76 cycles(tsc) 19.212 ns  -  66 cycles(tsc) 16.584 ns  improved 13.2%
 34 -  74 cycles(tsc) 18.600 ns  -  63 cycles(tsc) 15.954 ns  improved 14.9%
 48 -  88 cycles(tsc) 22.092 ns  -  77 cycles(tsc) 19.373 ns  improved 12.5%
 64 -  80 cycles(tsc) 20.043 ns  -  68 cycles(tsc) 17.188 ns  improved 15.0%
128 -  99 cycles(tsc) 24.818 ns  -  89 cycles(tsc) 22.404 ns  improved 10.1%
158 -  99 cycles(tsc) 24.977 ns  -  92 cycles(tsc) 23.089 ns  improved  7.1%
250 - 106 cycles(tsc) 26.552 ns  -  99 cycles(tsc) 24.785 ns  improved  6.6%

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

slub bulk alloc: extract objects from the per cpu slab

First piece: acceleration of retrieval of per cpu objects

If we are allocating lots of objects then it is advantageous to disable
interrupts and avoid the this_cpu_cmpxchg() operation to get these objects
faster.

Note that we cannot do the fast operation if debugging is enabled, because
we would have to add extra code to do all the debugging checks.  And it
would not be fast anyway.

Note also that the requirement of having interrupts disabled avoids having
to do processor flag operations.

Allocate as many objects as possible in the fast way and then fall back to
the generic implementation for the rest of the objects.

Measurements on CPU CPU i7-4790K @ 4.00GHz
Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns

Bulk- fallback                   - this-patch
  1 -  57 cycles(tsc) 14.432 ns  -  48 cycles(tsc) 12.155 ns  improved 15.8%
  2 -  50 cycles(tsc) 12.746 ns  -  37 cycles(tsc)  9.390 ns  improved 26.0%
  3 -  48 cycles(tsc) 12.180 ns  -  33 cycles(tsc)  8.417 ns  improved 31.2%
  4 -  48 cycles(tsc) 12.015 ns  -  32 cycles(tsc)  8.045 ns  improved 33.3%
  8 -  46 cycles(tsc) 11.526 ns  -  30 cycles(tsc)  7.699 ns  improved 34.8%
 16 -  45 cycles(tsc) 11.418 ns  -  32 cycles(tsc)  8.205 ns  improved 28.9%
 30 -  80 cycles(tsc) 20.246 ns  -  73 cycles(tsc) 18.328 ns  improved  8.8%
 32 -  79 cycles(tsc) 19.946 ns  -  72 cycles(tsc) 18.208 ns  improved  8.9%
 34 -  78 cycles(tsc) 19.659 ns  -  71 cycles(tsc) 17.987 ns  improved  9.0%
 48 -  86 cycles(tsc) 21.516 ns  -  82 cycles(tsc) 20.566 ns  improved  4.7%
 64 -  93 cycles(tsc) 23.423 ns  -  89 cycles(tsc) 22.480 ns  improved  4.3%
128 - 100 cycles(tsc) 25.170 ns  -  99 cycles(tsc) 24.871 ns  improved  1.0%
158 - 102 cycles(tsc) 25.549 ns  - 101 cycles(tsc) 25.375 ns  improved  1.0%
250 - 101 cycles(tsc) 25.344 ns  - 100 cycles(tsc) 25.182 ns  improved  1.0%

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

slab: infrastructure for bulk object allocation and freeing

Add the basic infrastructure for alloc/free operations on pointer arrays.
It includes a generic function in the common slab code that is used in
this infrastructure patch to create the unoptimized functionality for slab
bulk operations.

Allocators can then provide optimized allocation functions for situations
in which large numbers of objects are needed.  These optimization may
avoid taking locks repeatedly and bypass metadata creation if all objects
in slab pages can be used to provide the objects required.

Allocators can extend the skeletons provided and add their own code to the
bulk alloc and free functions.  They can keep the generic allocation and
freeing and just fall back to those if optimizations would not work (like
for example when debugging is on).

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

slub: fix spelling succedd to succeed

With this patchset the SLUB allocator now has both bulk alloc and free
implemented.

This patchset mostly optimizes the "fastpath" where objects are available
on the per CPU fastpath page.  This mostly amortize the less-heavy
none-locked cmpxchg_double used on fastpath.

The "fallback" bulking (e.g __kmem_cache_free_bulk) provides a good basis
for comparison.  Measurements[1] of the fallback functions
__kmem_cache_{free,alloc}_bulk have been copied from slab_common.c and
forced "noinline" to force a function call like slab_common.c.

Measurements on CPU CPU i7-4790K @ 4.00GHz
Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns

Measurements last-patch with disabled debugging:

Bulk- fallback                   - this-patch
  1 -  57 cycles(tsc) 14.448 ns  -  44 cycles(tsc) 11.236 ns  improved 22.8%
  2 -  51 cycles(tsc) 12.768 ns  -  28 cycles(tsc)  7.019 ns  improved 45.1%
  3 -  48 cycles(tsc) 12.232 ns  -  22 cycles(tsc)  5.526 ns  improved 54.2%
  4 -  48 cycles(tsc) 12.025 ns  -  19 cycles(tsc)  4.786 ns  improved 60.4%
  8 -  46 cycles(tsc) 11.558 ns  -  18 cycles(tsc)  4.572 ns  improved 60.9%
 16 -  45 cycles(tsc) 11.458 ns  -  18 cycles(tsc)  4.658 ns  improved 60.0%
 30 -  45 cycles(tsc) 11.499 ns  -  18 cycles(tsc)  4.568 ns  improved 60.0%
 32 -  79 cycles(tsc) 19.917 ns  -  65 cycles(tsc) 16.454 ns  improved 17.7%
 34 -  78 cycles(tsc) 19.655 ns  -  63 cycles(tsc) 15.932 ns  improved 19.2%
 48 -  68 cycles(tsc) 17.049 ns  -  50 cycles(tsc) 12.506 ns  improved 26.5%
 64 -  80 cycles(tsc) 20.009 ns  -  63 cycles(tsc) 15.929 ns  improved 21.3%
128 -  94 cycles(tsc) 23.749 ns  -  86 cycles(tsc) 21.583 ns  improved  8.5%
158 -  97 cycles(tsc) 24.299 ns  -  90 cycles(tsc) 22.552 ns  improved  7.2%
250 - 102 cycles(tsc) 25.681 ns  -  98 cycles(tsc) 24.589 ns  improved  3.9%

Benchmarking shows impressive improvements in the "fastpath" with a small
number of objects in the working set.  Once the working set increases,
resulting in activating the "slowpath" (that contains the heavier locked
cmpxchg_double) the improvement decreases.

I'm currently working on also optimizing the "slowpath" (as network stack
use-case hits this), but this patchset should provide a good foundation
for further improvements.  Rest of my patch queue in this area needs some
more work, but preliminary results are good.  I'm attending Netfilter
Workshop[2] next week, and I'll hopefully return working on further
improvements in this area.

This patch (of 6):

s/succedd/succeed/

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

watchdog: rename watchdog_suspend() and watchdog_resume()

Rename watchdog_suspend() to lockup_detector_suspend() and
watchdog_resume() to lockup_detector_resume() to avoid confusion with the
watchdog subsystem and to be consistent with the existing name
lockup_detector_init().

Also provide comment blocks to explain the watchdog_running and
watchdog_suspended variables and their relationship.

Signed-off-by: Ulrich Obergfell <uobergfe@redhat.com>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Stephane Eranian <eranian@google.com>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

watchdog: use suspend/resume interface in fixup_ht_bug()

Remove watchdog_nmi_disable_all() and watchdog_nmi_enable_all() since
these functions are no longer needed.  If a subsystem has a need to
deactivate the watchdog temporarily, it should utilize the
watchdog_suspend() and watchdog_resume() functions.

[akpm@linux-foundation.org: fix build with CONFIG_LOCKUP_DETECTOR=m]
Signed-off-by: Ulrich Obergfell <uobergfe@redhat.com>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Stephane Eranian <eranian@google.com>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

watchdog: use park/unpark functions in update_watchdog_all_cpus()

Remove update_watchdog() and restart_watchdog_hrtimer() since these
functions are no longer needed.  Changes of parameters such as the sample
period are honored at the time when the watchdog threads are being
unparked.

Signed-off-by: Ulrich Obergfell <uobergfe@redhat.com>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Stephane Eranian <eranian@google.com>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

watchdog: introduce watchdog_suspend() and watchdog_resume()

This interface can be utilized to deactivate the hard and soft lockup
detector temporarily.  Callers are expected to minimize the duration of
deactivation.  Multiple deactivations are allowed to occur in parallel but
should be rare in practice.

[akpm@linux-foundation.org: remove unneeded static initialization]
Signed-off-by: Ulrich Obergfell <uobergfe@redhat.com>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Stephane Eranian <eranian@google.com>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

watchdog: introduce watchdog_park_threads() and watchdog_unpark_threads()

Originally watchdog_nmi_enable(cpu) and watchdog_nmi_disable(cpu) were
only called in watchdog thread context.  However, the following commits
utilize these functions outside of watchdog thread context too.

  commit 9809b18fcf6b8d8ec4d3643677345907e6b50eca
  Author: Michal Hocko <mhocko@suse.cz>
  Date:   Tue Sep 24 15:27:30 2013 -0700

      watchdog: update watchdog_thresh properly

  commit b3738d29323344da3017a91010530cf3a58590fc
  Author: Stephane Eranian <eranian@google.com>
  Date:   Mon Nov 17 20:07:03 2014 +0100

      watchdog: Add watchdog enable/disable all functions

Hence, it is now possible that these functions execute concurrently with
the same 'cpu' argument.  This concurrency is problematic because per-cpu
'watchdog_ev' can be accessed/modified without adequate synchronization.

The patch series aims to address the above problem.  However, instead of
introducing locks to protect per-cpu 'watchdog_ev' a different approach is
taken: Invoke these functions by parking and unparking the watchdog
threads (to ensure they are always called in watchdog thread context).

  static struct smp_hotplug_thread watchdog_threads = {
           ...
          .park   = watchdog_disable, // calls watchdog_nmi_disable()
          .unpark = watchdog_enable,  // calls watchdog_nmi_enable()
  };

Both previously mentioned commits call these functions in a similar way
and thus in principle contain some duplicate code.  The patch series also
avoids this duplication by providing a commonly usable mechanism.

- Patch 1/4 introduces the watchdog_{park|unpark}_threads functions that
  park/unpark all watchdog threads specified in 'watchdog_cpumask'. They
  are intended to be called inside of kernel/watchdog.c only.

- Patch 2/4 introduces the watchdog_{suspend|resume} functions which can
  be utilized by external callers to deactivate the hard and soft lockup
  detector temporarily.

- Patch 3/4 utilizes watchdog_{park|unpark}_threads to replace some code
  that was introduced by commit 9809b18fcf6b8d8ec4d3643677345907e6b50eca.

- Patch 4/4 utilizes watchdog_{suspend|resume} to replace some code that
  was introduced by commit b3738d29323344da3017a91010530cf3a58590fc.

A few corner cases should be mentioned here for completeness.

- kthread_park() of watchdog/N could hang if cpu N is already locked up.
  However, if watchdog is enabled the lockup will be detected anyway.

- kthread_unpark() of watchdog/N could hang if cpu N got locked up after
  kthread_park(). The occurrence of this scenario should be _very_ rare
  in practice, in particular because it is not expected that temporary
  deactivation will happen frequently, and if it happens at all it is
  expected that the duration of deactivation will be short.

This patch (of 4): introduce watchdog_park_threads() and watchdog_unpark_threads()

These functions are intended to be used only from inside kernel/watchdog.c
to park/unpark all watchdog threads that are specified in
watchdog_cpumask.

Signed-off-by: Ulrich Obergfell <uobergfe@redhat.com>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Stephane Eranian <eranian@google.com>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

kernel/watchdog: move NMI function header declarations from watchdog.h to nmi.h

The kernel's NMI watchdog has nothing to do with the watchdog subsystem.
Its header declarations should be in linux/nmi.h, not linux/watchdog.h.

The code provided two sets of dummy functions if HARDLOCKUP_DETECTOR is
not configured, one in the include file and one in kernel/watchdog.c.
Remove the dummy functions from kernel/watchdog.c and use those from the
include file.

Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Cc: Stephane Eranian <eranian@google.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Don Zickus <dzickus@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

watchdog: simplify housekeeping affinity with the appropriate mask

housekeeping_mask gathers all the CPUs that aren't part of the nohz_full
set.  This is exactly what we want the watchdog to be affine to without
the need to use complicated cpumask operations.

Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

smpboot: allow passing the cpumask on per-cpu thread registration

It makes the registration cheaper and simpler for the smpboot per-cpu
kthread users that don't need to always update the cpumask after threads
creation.

[sfr@canb.auug.org.au: fix for allow passing the cpumask on per-cpu thread registration]
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

smpboot: make cleanup to mirror setup

The per-cpu kthread cleanup() callback is the mirror of the setup()
callback.  When the per-cpu kthread is started, it first calls setup()
to initialize the resources which are then released by cleanup() when
the kthread exits.

Now since the introduction of a per-cpu kthread cpumask, the kthreads
excluded by the cpumask on boot may happen to be parked immediately
after their creation without taking the setup() stage, waiting to be
asked to unpark to do so.  Then when smpboot_unregister_percpu_thread()
is later called, the kthread is stopped without having ever called
setup().

But this triggers a bug as the kthread unconditionally calls cleanup()
on exit but this doesn't mirror any setup().  Thus the kernel crashes
because we try to free resources that haven't been initialized, as in
the watchdog case:

    WATCHDOG disable 0
    WATCHDOG disable 1
    WATCHDOG disable 2
    BUG: unable to handle kernel NULL pointer dereference at           (null)
    IP: hrtimer_active+0x26/0x60
    [...]
    Call Trace:
      hrtimer_try_to_cancel+0x1c/0x280
      hrtimer_cancel+0x1d/0x30
      watchdog_disable+0x56/0x70
      watchdog_cleanup+0xe/0x10
      smpboot_thread_fn+0x23c/0x2c0
      kthread+0xf8/0x110
      ret_from_fork+0x3f/0x70

This bug is currently masked with explicit kthread unparking before
kthread_stop() on smpboot_destroy_threads(). This forces a call to
setup() and then unpark().

We could fix this by unconditionally calling setup() on kthread entry.
But setup() isn't always cheap.  In the case of watchdog it launches
hrtimer, perf events, etc...  So we may as well like to skip it if there
are chances the kthread will never be used, as in a reduced cpumask value.

So let's simply do a state machine check before calling cleanup() that
makes sure setup() has been called before mirroring it.

And remove the nasty hack workaround.

Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

smpboot: fix memory leak on error handling

The cpumask is allocated before threads get created. If the latter step
fails, we need to free the cpumask.

Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs: create and use seq_show_option for escaping

Many file systems that implement the show_options hook fail to correctly
escape their output which could lead to unescaped characters (e.g.  new
lines) leaking into /proc/mounts and /proc/[pid]/mountinfo files.  This
could lead to confusion, spoofed entries (resulting in things like
systemd issuing false d-bus "mount" notifications), and who knows what
else.  This looks like it would only be the root user stepping on
themselves, but it's possible weird things could happen in containers or
in other situations with delegated mount privileges.

Here's an example using overlay with setuid fusermount trusting the
contents of /proc/mounts (via the /etc/mtab symlink).  Imagine the use
of "sudo" is something more sneaky:

  $ BASE="ovl"
  $ MNT="$BASE/mnt"
  $ LOW="$BASE/lower"
  $ UP="$BASE/upper"
  $ WORK="$BASE/work/ 0 0
  none /proc fuse.pwn user_id=1000"
  $ mkdir -p "$LOW" "$UP" "$WORK"
  $ sudo mount -t overlay -o "lowerdir=$LOW,upperdir=$UP,workdir=$WORK" none /mnt
  $ cat /proc/mounts
  none /root/ovl/mnt overlay rw,relatime,lowerdir=ovl/lower,upperdir=ovl/upper,workdir=ovl/work/ 0 0
  none /proc fuse.pwn user_id=1000 0 0
  $ fusermount -u /proc
  $ cat /proc/mounts
  cat: /proc/mounts: No such file or directory

This fixes the problem by adding new seq_show_option and
seq_show_option_n helpers, and updating the vulnerable show_option
handlers to use them as needed.  Some, like SELinux, need to be open
coded due to unusual existing escape mechanisms.

[akpm@linux-foundation.org: add lost chunk, per Kees]
[keescook@chromium.org: seq_show_option should be using const parameters]
Signed-off-by: Kees Cook <keescook@chromium.org>
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Acked-by: Jan Kara <jack@suse.com>
Acked-by: Paul Moore <paul@paul-moore.com>
Cc: J. R. Okajima <hooanon05g@gmail.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: clean up redundant NULL checks before kfree

NULL check before kfree is redundant and so clean them up.

Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: neaten do_error, ocfs2_error and ocfs2_abort

These uses sometimes do and sometimes don't have '\n' terminations.  Make
the uses consistently use '\n' terminations and remove the newline from
the functions.

Miscellanea:

o Coalesce formats
o Realign arguments

Signed-off-by: Joe Perches <joe@perches.com>
Reviewed-by: 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>

ocfs2: do not set fs read-only if rec[0] is empty while committing truncate

While appending an extent to a file, it will call these functions:
ocfs2_insert_extent

  -> call ocfs2_grow_tree() if there's no free rec
     -> ocfs2_add_branch add a new branch to extent tree,
        now rec[0] in the leaf of rightmost path is empty
  -> ocfs2_do_insert_extent
     -> ocfs2_rotate_tree_right
       -> ocfs2_extend_rotate_transaction
          -> jbd2_journal_restart if jbd2_journal_extend fail
     -> ocfs2_insert_path
        -> ocfs2_extend_trans
          -> jbd2_journal_restart if jbd2_journal_extend fail
        -> ocfs2_insert_at_leaf
     -> ocfs2_et_update_clusters
Function jbd2_journal_restart() may be called and it may happened that
buffers dirtied in ocfs2_add_branch() are committed
while buffers dirtied in ocfs2_insert_at_leaf() and
ocfs2_et_update_clusters() are not.
So an empty rec[0] is left in rightmost path which will cause
read-only filesystem when call ocfs2_commit_truncate()
with the error message: "Inode %lu has an empty extent record".

This is not a serious problem, so remove the rightmost path when call
ocfs2_commit_truncate().

Signed-off-by: joyce.xue <xuejiufei@huawei.com>
Reviewed-by: 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>

ocfs2: call ocfs2_journal_access_di() before ocfs2_journal_dirty() in ocfs2_write_end_nolock()

1: After we call ocfs2_journal_access_di() in ocfs2_write_begin(),
   jbd2_journal_restart() may also be called, in this function transaction
   A's t_updates-- and obtains a new transaction B.  If
   jbd2_journal_commit_transaction() is happened to commit transaction A,
   when t_updates==0, it will continue to complete commit and unfile
   buffer.

   So when jbd2_journal_dirty_metadata(), the handle is pointed a new
   transaction B, and the buffer head's journal head is already freed,
   jh->b_transaction == NULL, jh->b_next_transaction == NULL, it returns
   EINVAL, So it triggers the BUG_ON(status).

thread 1                                          jbd2
ocfs2_write_begin                     jbd2_journal_commit_transaction
ocfs2_write_begin_nolock
  ocfs2_start_trans
    jbd2__journal_start(t_updates+1,
                       transaction A)
    ocfs2_journal_access_di
    ocfs2_write_cluster_by_desc
      ocfs2_mark_extent_written
        ocfs2_change_extent_flag
          ocfs2_split_extent
            ocfs2_extend_rotate_transaction
              jbd2_journal_restart
              (t_updates-1,transaction B) t_updates==0
                                        __jbd2_journal_refile_buffer
                                        (jh->b_transaction = NULL)
ocfs2_write_end
ocfs2_write_end_nolock
    ocfs2_journal_dirty
        jbd2_journal_dirty_metadata(bug)
   ocfs2_commit_trans

2.  In ext4, I found that: jbd2_journal_get_write_access() called by
   ext4_write_end.

ext4_write_begin
    ext4_journal_start
        __ext4_journal_start_sb
            ext4_journal_check_start
            jbd2__journal_start

ext4_write_end
    ext4_mark_inode_dirty
        ext4_reserve_inode_write
            ext4_journal_get_write_access
                jbd2_journal_get_write_access
        ext4_mark_iloc_dirty
            ext4_do_update_inode
                ext4_handle_dirty_metadata
                    jbd2_journal_dirty_metadata

3. So I think we should put ocfs2_journal_access_di before
   ocfs2_journal_dirty in the ocfs2_write_end.  and it works well after my
   modification.

Signed-off-by: vicky <vicky.yangwenfang@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Zhangguanghui <zhang.guanghui@h3c.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: use 64bit variables to track heartbeat time

o2hb_elapsed_msecs computes the time taken for a disk heartbeat.
'struct timeval' variables are used to store start and end times.  On
32-bit systems, the 'tv_sec' component of 'struct timeval' will overflow
in year 2038 and beyond.

This patch solves the overflow with the following:

1. Replace o2hb_elapsed_msecs using 'ktime_t' values to measure start
   and end time, and built-in function 'ktime_ms_delta' to compute the
   elapsed time.  ktime_get_real() is used since the code prints out the
   wallclock time.

2. Changes format string to print time as a single 64-bit nanoseconds
   value ("%lld") instead of seconds and microseconds.  This simplifies
   the code since converting ktime_t to that format would need expensive
   computation.  However, the debug log string is less readable than the
   previous format.

Signed-off-by: Tina Ruchandani <ruchandani.tina@gmail.com>
Suggested by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: 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>

ocfs2: fix race between crashed dio and rm

There is a race case between crashed dio and rm, which will lead to
OCFS2_VALID_FL not set read-only.

  N1                              N2
  ------------------------------------------------------------------------
  dd with direct flag
                                  rm file
  crashed with an dio entry left
  in orphan dir
                                  clear OCFS2_VALID_FL in
                                  ocfs2_remove_inode
                                  recover N1 and read the corrupted inode,
                                  and set filesystem read-only

So we skip the inode deletion this time and wait for dio entry recovered
first.

Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Reviewed-by: 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>

ocfs2: avoid access invalid address when read o2dlm debug messages

The following case will lead to a lockres is freed but is still in use.

cat /sys/kernel/debug/o2dlm/locking_state dlm_thread
lockres_seq_start
    -> lock dlm->track_lock
    -> get resA
                                                resA->refs decrease to 0,
                                                call dlm_lockres_release,
                                                and wait for "cat" unlock.
Although resA->refs is already set to 0,
increase resA->refs, and then unlock
                                                lock dlm->track_lock
                                                    -> list_del_init()
                                                    -> unlock
                                                    -> free resA

In such a race case, invalid address access may occurs.  So we should
delete list res->tracking before resA->refs decrease to 0.

Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: take inode lock in ocfs2_iop_set/get_acl()

This bug in mainline code is pointed out by Mark Fasheh.  When
ocfs2_iop_set_acl() and ocfs2_iop_get_acl() are entered from VFS layer,
inode lock is not held.  This seems to be regression from older kernels.
The patch is to fix that.

Orabug: 20189959
Signed-off-by: Tariq Saeed <tariq.x.saeed@oracle.com>
Reviewed-by: 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>

ocfs2: fix BUG_ON() in ocfs2_ci_checkpointed()

PID: 614    TASK: ffff882a739da580  CPU: 3   COMMAND: "ocfs2dc"
  #0 [ffff882ecc3759b0] machine_kexec at ffffffff8103b35d
  #1 [ffff882ecc375a20] crash_kexec at ffffffff810b95b5
  #2 [ffff882ecc375af0] oops_end at ffffffff815091d8
  #3 [ffff882ecc375b20] die at ffffffff8101868b
  #4 [ffff882ecc375b50] do_trap at ffffffff81508bb0
  #5 [ffff882ecc375ba0] do_invalid_op at ffffffff810165e5
  #6 [ffff882ecc375c40] invalid_op at ffffffff815116fb
     [exception RIP: ocfs2_ci_checkpointed+208]
     RIP: ffffffffa0a7e940  RSP: ffff882ecc375cf0  RFLAGS: 00010002
     RAX: 0000000000000001  RBX: 000000000000654b  RCX: ffff8812dc83f1f8
     RDX: 00000000000017d9  RSI: ffff8812dc83f1f8  RDI: ffffffffa0b2c318
     RBP: ffff882ecc375d20   R8: ffff882ef6ecfa60   R9: ffff88301f272200
     R10: 0000000000000000  R11: 0000000000000000  R12: ffffffffffffffff
     R13: ffff8812dc83f4f0  R14: 0000000000000000  R15: ffff8812dc83f1f8
     ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018
  #7 [ffff882ecc375d28] ocfs2_check_meta_downconvert at ffffffffa0a7edbd [ocfs2]
  #8 [ffff882ecc375d38] ocfs2_unblock_lock at ffffffffa0a84af8 [ocfs2]
  #9 [ffff882ecc375dc8] ocfs2_process_blocked_lock at ffffffffa0a85285 [ocfs2]
#10 [ffff882ecc375e18] ocfs2_downconvert_thread_do_work at ffffffffa0a85445 [ocfs2]
#11 [ffff882ecc375e68] ocfs2_downconvert_thread at ffffffffa0a854de [ocfs2]
#12 [ffff882ecc375ee8] kthread at ffffffff81090da7
#13 [ffff882ecc375f48] kernel_thread_helper at ffffffff81511884
assert is tripped because the tran is not checkpointed and the lock level is PR.

Some time ago, chmod command had been executed. As result, the following call
chain left the inode cluster lock in PR state, latter on causing the assert.
system_call_fastpath
  -> my_chmod
   -> sys_chmod
    -> sys_fchmodat
     -> notify_change
      -> ocfs2_setattr
       -> posix_acl_chmod
        -> ocfs2_iop_set_acl
         -> ocfs2_set_acl
          -> ocfs2_acl_set_mode
Here is how.
1119 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1120 {
1247         ocfs2_inode_unlock(inode, 1); <<< WRONG thing to do.
..
1258         if (!status && attr->ia_valid & ATTR_MODE) {
1259                 status =  posix_acl_chmod(inode, inode->i_mode);

519 posix_acl_chmod(struct inode *inode, umode_t mode)
520 {
..
539         ret = inode->i_op->set_acl(inode, acl, ACL_TYPE_ACCESS);

287 int ocfs2_iop_set_acl(struct inode *inode, struct posix_acl *acl, ...
288 {
289         return ocfs2_set_acl(NULL, inode, NULL, type, acl, NULL, NULL);

224 int ocfs2_set_acl(handle_t *handle,
225                          struct inode *inode, ...
231 {
..
252                                 ret = ocfs2_acl_set_mode(inode, di_bh,
253                                                          handle, mode);

168 static int ocfs2_acl_set_mode(struct inode *inode, struct buffer_head ...
170 {
183         if (handle == NULL) {
                    >>> BUG: inode lock not held in ex at this point <<<
184                 handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
185                                            OCFS2_INODE_UPDATE_CREDITS);

ocfs2_setattr.#1247 we unlock and at #1259 call posix_acl_chmod. When we reach
ocfs2_acl_set_mode.#181 and do trans, the inode cluster lock is not held in EX
mode (it should be). How this could have happended?

We are the lock master, were holding lock EX and have released it in
ocfs2_setattr.#1247.  Note that there are no holders of this lock at
this point.  Another node needs the lock in PR, and we downconvert from
EX to PR.  So the inode lock is PR when do the trans in
ocfs2_acl_set_mode.#184.  The trans stays in core (not flushed to disc).
Now another node want the lock in EX, downconvert thread gets kicked
(the one that tripped assert abovt), finds an unflushed trans but the
lock is not EX (it is PR).  If the lock was at EX, it would have flushed
the trans ocfs2_ci_checkpointed -> ocfs2_start_checkpoint before
downconverting (to NULL) for the request.

ocfs2_setattr must not drop inode lock ex in this code path.  If it
does, takes it again before the trans, say in ocfs2_set_acl, another
cluster node can get in between, execute another setattr, overwriting
the one in progress on this node, resulting in a mode acl size combo
that is a mix of the two.

Orabug: 20189959
Signed-off-by: Tariq Saeed <tariq.x.saeed@oracle.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Joseph Qi <joseph.qi@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: optimize error handling in dlm_request_join

Currently error handling in dlm_request_join is a little obscure, so
optimize it to promote readability.

If packet.code is invalid, reset it to JOIN_DISALLOW to keep it
meaningful.  It only influences the log printing.

Signed-off-by: Norton.Zhu <norton.zhu@huawei.com>
Cc: Srinivas Eeda <srinivas.eeda@oracle.com>
Reviewed-by: 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>

ocfs2: fix a tiny case that inode can not removed

When running dirop_fileop_racer we found a case that inode
can not removed.

Two nodes, say Node A and Node B, mount the same ocfs2 volume.  Create
two dirs /race/1/ and /race/2/ in the filesystem.

  Node A                            Node B
  rm -r /race/2/
                                    mv /race/1/ /race/2/
  call ocfs2_unlink(), get
  the EX mode of /race/2/
                                    wait for B unlock /race/2/
  decrease i_nlink of /race/2/ to 0,
  and add inode of /race/2/ into
  orphan dir, unlock /race/2/
                                    got EX mode of /race/2/. because
                                    /race/1/ is dir, so inc i_nlink
                                    of /race/2/ and update into disk,
                                    unlock /race/2/
  because i_nlink of /race/2/
  is not zero, this inode will
  always remain in orphan dir

This patch fixes this case by test whether i_nlink of new dir is zero.

Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Xue jiufei <xuejiufei@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: add ip_alloc_sem in direct IO to protect allocation changes

In ocfs2, ip_alloc_sem is used to protect allocation changes on the
node.  In direct IO, we add ip_alloc_sem to protect date consistent
between direct-io and ocfs2_truncate_file race (buffer io use
ip_alloc_sem already).  Although inode->i_mutex lock is used to avoid
concurrency of above situation, i think ip_alloc_sem is still needed
because protect allocation changes is significant.

Other filesystem like ext4 also uses rw_semaphore to protect data
consistent between get_block-vs-truncate race by other means, So
ip_alloc_sem in ocfs2 direct io is needed.

Signed-off-by: Weiwei Wang <wangww631@huawei.com>
Signed-off-by: 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>

ocfs2: clear the rest of the buffers on error

In case a validation fails, clear the rest of the buffers and return the
error to the calling function.

This also facilitates bubbling up the error originating from ocfs2_error
to calling functions.

Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: 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>

ocfs2: acknowledge return value of ocfs2_error()

Caveat: This may return -EROFS for a read case, which seems wrong.  This
is happening even without this patch series though.  Should we convert
EROFS to EIO?

Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.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>

ocfs2: add errors=continue

OCFS2 is often used in high-availaibility systems.  However, ocfs2
converts the filesystem to read-only at the drop of the hat.  This may
not be necessary, since turning the filesystem read-only would affect
other running processes as well, decreasing availability.

This attempt is to add errors=continue, which would return the EIO to
the calling process and terminate furhter processing so that the
filesystem is not corrupted further.  However, the filesystem is not
converted to read-only.

As a future plan, I intend to create a small utility or extend
fsck.ocfs2 to fix small errors such as in the inode.  The input to the
utility such as the inode can come from the kernel logs so we don't have
to schedule a downtime for fixing small-enough errors.

The patch changes the ocfs2_error to return an error.  The error
returned depends on the mount option set.  If none is set, the default
is to turn the filesystem read-only.

Perhaps errors=continue is not the best option name.  Historically it is
used for making an attempt to progress in the current process itself.
Should we call it errors=eio? or errors=killproc? Suggestions/Comments
welcome.

Sources are available at:
  https://github.com/goldwynr/linux/tree/error-cont

Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: 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>

ocfs2: flush inode data to disk and free inode when i_count becomes zero

Disk inode deletion may be heavily delayed when one node unlink a file
after the same dentry is freed on another node(say N1) because of memory
shrink but inode is left in memory.  This inode can only be freed while
N1 doing the orphan scan work.

However, N1 may skip orphan scan for several times because other nodes
may do the work earlier.  In our tests, it may take 1 hour on 4 nodes
cluster and it hurts the user experience.  So we think the inode should
be freed after the data flushed to disk when i_count becomes zero to
avoid such circumstances.

Signed-off-by: Joyce.xue <xuejiufei@huawei.com>
Cc: Joel Becker <jlbec@evilplan.org>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: trusted xattr missing CAP_SYS_ADMIN check

The trusted extended attributes are only visible to the process which
hvae CAP_SYS_ADMIN capability but the check is missing in ocfs2
xattr_handler trusted list.  The check is important because this will be
used for implementing mechanisms in the userspace for which other
ordinary processes should not have access to.

Signed-off-by: Sanidhya Kashyap <sanidhya.gatech@gmail.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Taesoo kim <taesoo@gatech.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2: set filesytem read-only when ocfs2_delete_entry failed.

In ocfs2_rename, it will lead to an inode with two entried(old and new) if
ocfs2_delete_entry(old) failed.  Thus, filesystem will be inconsistent.

The case is described below:

ocfs2_rename
    -> ocfs2_start_trans
    -> ocfs2_add_entry(new)
    -> ocfs2_delete_entry(old)
        -> __ocfs2_journal_access *failed* because of -ENOMEM
    -> ocfs2_commit_trans

So filesystem should be set to read-only at the moment.

Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Joel Becker <jlbec@evilplan.org>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ocfs2/dlm: use list_for_each_entry instead of list_for_each

Use list_for_each_entry instead of list_for_each to simplify code.

Signed-off-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>

ocfs2: remove unneeded code in dlm_register_domain_handlers

The last goto statement is unneeded, so remove it.

Signed-off-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>