- Have a stab at clustering the members of vnode_t and vnode_impl_t in a
more cache-conscious way. With that done, go back to adjusting v_usecount
with atomics and keep vi_lock directly in vnode_impl_t (saves KVA).
- Allow VOP_LOCK(LK_NONE) for the benefit of VFS_VGET() and VFS_ROOT().
Make sure LK_UPGRADE always comes with LK_NOWAIT.
- Make cwdinfo use mostly lockless.
- Change the lock on uvm_object, vm_amap and vm_anon to be a RW lock.
- Break v_interlock and vmobjlock apart. v_interlock remains a mutex.
- Do partial PV list locking in the x86 pmap. Others to follow later.
This code path is used both for xfers that are new, and xfers that
are being done piece by piece and are partway done. For the latter
case, skip usbd_xfer_schedule_timeout so we schedule it only once per
xfer.
If you set zfs=YES, presumably you positively want the automatic rc.d
actions, so if there's no /sbin/zfs or if zfs can't find pools with
/etc/zfs/zpool.cache, presumably you would like feedback about that
in rc.log.
Otherwise, we might do
lfs_vflush
-> lfs_seglock
-> lfs_segwait(SEGM_CKP)
-> lfs_writer_enter
which is the reverse of the lfs_writer -> lfs_seglock ordering.
The lock order is lfs_writer -> lfs_seglock. The problem in 52301 is
that lfs_segwrite violates this lock order by sometimes doing
lfs_seglock -> lfs_writer, either (a) when doing a checkpoint or (b),
opportunistically, when there are no dirops pending. Both cases can
deadlock, because dirops sometimes take the seglock (lfs_truncate,
lfs_valloc, lfs_vfree):
(a) There may be dirops pending, and they may be waiting for the
seglock, so we can't wait for them to complete while holding the
seglock.
(b) The test for fs->lfs_dirops == 0 happens unlocked, and the state
may change by the time lfs_writer_enter acquires lfs_lock.
To resolve this in each case:
(a) Do lfs_writer_enter before lfs_seglock, since we will need it
unconditionally anyway. The worst performance impact of this should
be that some dirops get delayed a little bit.
(b) Create a new lfs_writer_tryenter to use at this point so that the
test for fs->lfs_dirops == 0 and the acquisition of lfs_writer happen
atomically under lfs_lock.
Fixes panic:
KASSERT((ip->i_state & IN_ADIROP) == 0) at lfs_vnops.c:1670
lfs_flush_dirops
lfs_check
lfs_setattr
VOP_SETATTR
change_mode
sys_fchmod
syscall
This assertion -- and the assertion that vp->v_uflag has VU_DIROP set
-- is valid only until we release lfs_lock, because we may race with
lfs_unmark_dirop which will remove the nodes and change the flags.
Further, vp itself is valid only as long as it is referenced, which it
is as long as it's on the dchain, but lfs_unmark_dirop drops the
dchain's reference.
