vfs: add VOP_DELAYED_SETSIZE() and related infrastructure The change generalizes code that was initially developed for nfs client to handle filesystems that needs to call vnode_pager_setsize() while only owning the vnode lock shared. Since vnode pager might need to trim or extend the vnode vm_object' page queue, the vnode lock for the call must be owned exclusive. This is typical for filesystems with remote authorative source of file attributes, like nfs/p9/fuse. Handle the conflict by delaying the vnode_pager_setsize() to the next vnode locking to avoid relock. But if the next locking request is in shared mode, lock it exclusively instead, perform the delayed vnode_pager_setsize() call by doing VOP_DEFAULT_SETSIZE(), and then downgrade to shared. Filesystems that opt into the feature must provide the implementation of VOP_DELAYED_SETSIZE() that actually calls vnode_pager_setsize(), and use vn_delay_setsize() helper to mark the vnode as requiring the delay call. nfsclient: convert to use vn_delayed_setsize()
Details
Details
Diff Detail
Diff Detail
- Repository
- rG FreeBSD src repository
- Lint
Lint Skipped - Unit
Tests Skipped
Event Timeline
Comment Actions
20260228 20:28:50 all (1/11): ftruncate3.sh
VNASSERT failed: a->a_gen.a_desc == &vop_lock1_desc not true at vnode_if.c:2095 (VOP_LOCK1_APV)
0xfffffe016a9e6940: type VREG state VSTATE_CONSTRUCTED op 0xffffffff81aaead8
usecount 3, writecount 2, refcount 1 seqc users 0
hold count flags ()
flags (VV_VMSIZEVNLOCK|VI(0x40)|VMP_LAZYLIST)
v_object 0xfffff803dac10c30 ref 2 pages 0 cleanbuf 0 dirtybuf 0
lock type nfs: UNLOCKED
#0 0xffffffff80b990c8 at lockmgr_lock_flags+0x1b8
#1 0xffffffff80cc0f20 at vop_sigdefer+0x30
#2 0xffffffff81213a93 at VOP_LOCK1_APV+0x133
#3 0xffffffff80cfc96b at VOP_LOCK1+0x4b
#4 0xffffffff80cfa260 at _vn_lock+0x140
#5 0xffffffff80cf7902 at vn_truncate+0x82
#6 0xffffffff80c4febe at kern_ftruncate+0xee
#7 0xffffffff8113eaa9 at amd64_syscall+0x169
#8 0xffffffff8110d6db at fast_syscall_common+0xf8
fileid 3 fsid 0x3a3a00ff01
panic: Wrong a_desc in vop_lock1(0xfffffe016a9e6940, 0xfffffe010008fa50)Comment Actions
Accumulated bug fixes after the Peter' testing.
I consider removing the nfs n_mtx and using the vnode interlock instead, as the future change.
Comment Actions
Integrate delayed setsize locking into vn_lock() more tight, now that the code moved out from VOP.
Comment Actions
It all looks good to me, although I won't claim to
understand all the required semantics in
vn_lock_delayed_setsize().
| sys/fs/deadfs/dead_vnops.c | ||
|---|---|---|
| 83 | These VOPs are mostly alphabetically sorted, with the exception of vop_fplookup_*. | |
| sys/kern/vfs_vnops.c | ||
| 2007 | This assertion is redundant. I don't understand why we check v_data == NULL at all. | |
| 2014 | vn_clear_delayed_setsize_locked()? | |
| 2038 | Shouldn't it be __predict_true? This is an unlocked read, so maybe should be protected with atomic_load. | |
| 2045 | Why not pass the arguments normally and let vn_lock_delayed_setsize() build the structure? This calling convention looks odd. | |
| sys/sys/vnode.h | ||
|---|---|---|
| 1256 | It is strange that these helpers are called vn_delay_setsize* but the flag is VI_DELAYEDSSZ. IMO "delayed" is right, and the flag should be VI_DELAYED_SETSIZE or similar, it is not referenced often so does not need to be terse. | |
Comment Actions
markj: sorry, I just committed before your notes come in, your suggested changes went into D55681
| sys/kern/vfs_vnops.c | ||
|---|---|---|
| 2007 | This is similar to the VI_DOOMED check, but with some delicate differences. The vnode being reclaimed might drop the vnode lock in VOP_RECLAIM(), and there might be some pending operations that needs to be done on the vnode. For instance, the buffer flushes. So checking the VI_DOOMED would not allow that operations to finish. | |