At this point the directory's vnode lock is held, so blocking while
waiting for free pages makes the system more susceptible to deadlock in
low memory conditions. This is particularly problematic on NUMA systems
as UMA currently implements a strict first-touch policy.
ufsdirhash_build() already uses M_NOWAIT for other allocations and
already handled failures for the block array allocation, so just convert
to M_NOWAIT.
PR: 253992
Reviewed by: markj, mckusick, vangyzen
(cherry picked from commit f17a590085)
It is unclear without investigating if it can be provided without using
extra memory, so for the time being just don't.
(cherry picked from commit c892d60a1d)
Revert 2d4422e799, Eliminate lock order reversal in UFS ffs_unmount().
After discussion with Chuck Silvers (chs@) we have decided that
there is a better way to resolve this lock order reversal which
will be committed separately.
Sponsored by: Netflix
(cherry picked from commit a63eae65ff)
disk failure.
Each vnode has an embedded lock that controls access to its contents.
However vnodes describing a UFS snapshot all share a single snapshot
lock to coordinate their access and update. As part of mounting a
UFS filesystem with snapshots, each of the vnodes describing a
snapshot has its individual lock replaced with the snapshot lock.
When the filesystem is unmounted the vnode's original lock is
returned replacing the snapshot lock.
When a disk fails while the UFS filesystem it contains is still
mounted (for example when a thumb drive is removed) UFS forcibly
unmounts the filesystem. The loss of the drive causes the GEOM
subsystem to orphan the provider, but the consumer remains until
the filesystem has finished with the unmount. Information describing
the snapshot locks was being prematurely cleared during the orphaning
causing the return of the snapshot vnode's original locks to fail.
The fix is to not clear the needed information prematurely.
Sponsored by: Netflix
with snapshots.
Each vnode has an embedded lock that controls access to its contents.
However vnodes describing a UFS snapshot all share a single snapshot
lock to coordinate their access and update. As part of mounting a
UFS filesystem with snapshots, each of the vnodes describing a
snapshot has its individual lock replaced with the snapshot lock.
When the filesystem is unmounted the vnode's original lock is
returned replacing the snapshot lock.
The lock order reversal happens because vnode locks must be acquired
before snapshot locks. When unmounting we must lock both the snapshot
lock and the vnode lock before swapping them so that the vnode will
be continuously locked during the swap. For each vnode representing
a snapshot, we must first acquire the snapshot lock to ensure
exclusive access to it and its original lock. We then face a lock
order reversal when we try to acquire the original vnode lock. The
problem is eliminated by doing a non-blocking exclusive lock on the
original lock which will always succeed since there are no users
of that lock.
Sponsored by: Netflix
UFS uses a new "mntfs" pseudo file system which provides private
device vnodes for a file system to safely access its disk device.
The original device vnode is saved in um_odevvp to hold the exclusive
lock on the device so that any attempts to open it for writing will
fail. But it is otherwise unused and has its BO_NOBUFS flag set to
enforce that file systems using mntfs vnodes do not accidentally
use the original devfs vnode. When the file system is unmounted,
um_odevvp is no longer needed and is released.
The lock order reversal happens because device vnodes must be locked
before UFS vnodes. During unmount, the root directory vnode lock
is held. When when calling vrele() on um_odevvp, vrele() attempts to
exclusive lock um_odevvp causing the lock order reversal. The problem
is eliminated by doing a non-blocking exclusive lock on um_odevvp
which will always succeed since there are no users of um_odevvp.
With um_odevvp locked, it can be released using vput which does not
attempt to do a blocking exclusive lock request and thus avoids the
lock order reversal.
Sponsored by: Netflix
Respect the new IO_DATASYNC flag when performing synchronous writes.
Compared to O_SYNC, O_DSYNC lets us skip updating the inode in some
cases, matching the behaviour of fdatasync(2).
Reviewed by: kib
Differential Review: https://reviews.freebsd.org/D25160
We use a bitmap to track which cylinder groups have changed between
snapshot creation and filesystem suspension. The "legs" of the bitmap
are four bytes wide (see ACTIVESET()) so we must round up the allocation
size to a multiple of four bytes.
I believe this bug is harmless since UMA/kmem_* will both pad the
allocation and zero the full allocation. Note that malloc() does inline
zeroing when the allocation size is known at compile-time.
Reported by: pho (using KASAN)
Reviewed by: kib, mckusick
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D27731
BA_CLRBUF specifies that existing context of the block will be
completely overwritten by caller, so there is no reason to spend io
fetching existing data. We do the same for indirect blocks.
Reported by: tmunro
Reviewed by: mckusick, tmunro
Tested by: pho, tmunro
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D27353
Replace MAXPHYS by runtime variable maxphys. It is initialized from
MAXPHYS by default, but can be also adjusted with the tunable kern.maxphys.
Make b_pages[] array in struct buf flexible. Size b_pages[] for buffer
cache buffers exactly to atop(maxbcachebuf) (currently it is sized to
atop(MAXPHYS)), and b_pages[] for pbufs is sized to atop(maxphys) + 1.
The +1 for pbufs allow several pbuf consumers, among them vmapbuf(),
to use unaligned buffers still sized to maxphys, esp. when such
buffers come from userspace (*). Overall, we save significant amount
of otherwise wasted memory in b_pages[] for buffer cache buffers,
while bumping MAXPHYS to desired high value.
Eliminate all direct uses of the MAXPHYS constant in kernel and driver
sources, except a place which initialize maxphys. Some random (and
arguably weird) uses of MAXPHYS, e.g. in linuxolator, are converted
straight. Some drivers, which use MAXPHYS to size embeded structures,
get private MAXPHYS-like constant; their convertion is out of scope
for this work.
Changes to cam/, dev/ahci, dev/ata, dev/mpr, dev/mpt, dev/mvs,
dev/siis, where either submitted by, or based on changes by mav.
Suggested by: mav (*)
Reviewed by: imp, mav, imp, mckusick, scottl (intermediate versions)
Tested by: pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D27225
When operating in SU or SU+J mode, ffs_syncvnode() might need to
instantiate other vnode by inode number while owning syncing vnode
lock. Typically this other vnode is the parent of our vnode, but due
to renames occuring right before fsync (or during fsync when we drop
the syncing vnode lock, see below) it might be no longer parent.
More, the called function flush_pagedep_deps() needs to lock other
vnode while owning the lock for vnode which owns the buffer, for which
the dependencies are flushed. This creates another instance of the
same LoR as was fixed in softdep_sync().
Put the generic code for safe relocking into new SU helper
get_parent_vp() and use it in flush_pagedep_deps(). The case for safe
relocking of two vnodes with undefined lock order was extracted into
vn helper vn_lock_pair().
Due to call sequence
ffs_syncvnode()->softdep_sync_buf()->flush_pagedep_deps(),
ffs_syncvnode() indicates with ERELOOKUP that passed vnode was
unlocked in process, and can return ENOENT if the passed vnode
reclaimed. All callers of the function were inspected.
Because UFS namei lookups store auxiliary information about directory
entry in in-memory directory inode, and this information is then used
by UFS code that creates/removed directory entry in the actual
mutating VOPs, it is critical that directory vnode lock is not dropped
between lookup and VOP. For softdep_prelink(), which ensures that
later link/unlink operation can proceed without overflowing the
journal, calls were moved to the place where it is safe to drop
processing VOP because mutations are not yet applied. Then, ERELOOKUP
causes restart of the whole VFS operation (typically VFS syscall) at
top level, including the re-lookup of the involved pathes. [Note that
we already do the same restart for failing calls to vn_start_write(),
so formally this patch does not introduce new behavior.]
Similarly, unsafe calls to fsync in snapshot creation code were
plugged. A possible view on these failures is that it does not make
sense to continue creating snapshot if the snapshot vnode was
reclaimed due to forced unmount.
It is possible that relock/ERELOOKUP situation occurs in
ffs_truncate() called from ufs_inactive(). In this case, dropping the
vnode lock is not safe. Detect the situation with VI_DOINGINACT and
reschedule inactivation by setting VI_OWEINACT. ufs_inactive()
rechecks VI_OWEINACT and avoids reclaiming vnode is truncation failed
this way.
In ffs_truncate(), allocation of the EOF block for partial truncation
is re-done after vnode is synced, since we cannot leave the buffer
locked through ffs_syncvnode().
In collaboration with: pho
Reviewed by: mckusick (previous version), markj
Tested by: markj (syzkaller), pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D26136
