Import the zio nop-write improvement from Illumos. To reduce I/O,
nop-write omits overwriting data if the checksum (cryptographically
secure) of new data matches the checksum of existing data.
It also saves space if snapshots are in use.
It currently works only on datasets with enabled compression, disabled
deduplication and sha256 checksums.
IllumOS 13887:196932ec9e6a and 13888:7204b3392a58
3236 zio nop-write
References:
https://www.illumos.org/issues/3236
MFC after: 2 weeks
* There is no need for the delayed destruction of znodes via taskqueue,
now that we do not need to fear recursion from getnewvnode into
zfs_inactive and zfs_freebsd_reclaim, thus making znode/vnode state
machine a bit simpler.
* More complete porting of zfs_inactive from Solaris VFS model to FreeBSD
vop_inactive and vop_reclaim model. All destructive actions are done
in zfs_freebsd_reclaim.
This allows to simplify zfs_zget logic.
* Allow zfs_zget to return a doomed vnode if the current thread already
has an exclusive lock on the vnode.
* Clean up Solaris-isms like bailing out of reclaim/inactive on certain
values of v_usecount (aka v_count) or directly messing with this counter.
* Do not clear z_vnode while znode is still accessible.
z_vnode should be cleared only after zfs_znode_dmu_fini.
Otherwise zfs_zget may get an effectively half-deconstructed znode.
This allows to simplify zfs_zget logic further.
The above changes fix at least two known/reported problems:
o An indefinite wait in the following code path:
vgone -> VOP_RECLAIM -> zfs_freebsd_reclaim -> vnode_destroy_vobject ->
put_pages -> zfs_write -> zil_commit -> zfs_zget
This happened because vgone marks a vnode as VI_DOOMED before calling
VOP_RECLAIM, but zfs_zget would not return a doomed vnode under any
circumstances.
The fix in this change is not complete as it won't fix a deadlock between
two threads doing VOP_RECLAIM where one thread is in zil_commit trying to
zfs_zget a znode/vnode being reclaimed by the other thread, which would be
blocked trying to enter zil_commit. This type of deadlock has not been
reported as of now.
o An indefinite wait in the unmount path caused by a znode "falling through
the cracks" in inactive+reclaim. This would happen if the znode is unlinked
while its vnode is still active.
To Do: pass locking flags parameter to zfs_zget, so that the zfs-vfs
glue code doesn't have to re-lock a vnode but could ask for proper locking
from the very start. This would also allow for the higher level code to
obtain a doomed vnode when it is expected/requested. Or to avoid blocking
when it is not allowed (see zil_commit example above).
ffs_vgetf seems like a good source of inspiration.
Tested by: Willem Jan Withagen <wjw@digiware.nl>
MFC after: 6 weeks
... otherwise zfs_getpages would mostly be called with one page at a time.
It is expected that ZFS VOP_BMAP is only called from vnode_pager_haspage.
Since ZFS files can have variable block sizes and also because we don't
really know if any given blocks are consecutive, we can not really report
any additional blocks behind or ahead of a given block. Since physical
block numbers do not make sense for ZFS, we do not do any real translation
and thus pass back blk = lblk. The net effect is that vnode_pager_haspage
knows that the block exists and that the pages backed by the block can be
accessed. vnode_pager_haspage may be wrong about the exact count of the
pages backed by the block, because of a variable block size, which
vnode_pager_haspage doesn't really know - it only knows max block size in
a filesystem. So pages from multiple blocks can be passed to zfs_getpages,
but that is expected and correctly handled.
vnode_pager should not call zfs_bmap for any other reason, because ZFS
implements VOP_PUTPAGES and thus vnode_pager_generic_getpages is not used.
vfs_cluster code vfs_bio code should not be called for ZFS, because ZFS does
not use buffer cache layer.
Also, ZFS does not use vn_bmap_seekhole, it has its prviate mechanism for
working with holes.
The above list should cover all the current calls to VOP_BMAP.
Reviewed by: kib
MFC after: 6 weeks
There has not been any complaints about the default behavior, so there
is no need to keep a knob that enables the worse alternative.
Now that the hard-stopping of other CPUs is the only behavior, the panic_cpu
spinlock-like logic can be dropped, because only a single CPU is
supposed to win stop_cpus_hard(other_cpus) race and proceed past that
call.
MFC after: 1 month
Illumos 13886:e3261d03efbf
3349 zpool upgrade -V bumps the on disk version number, but leaves
the in core version
References:
https://www.illumos.org/issues/3349
MFC after: 1 week
... because the latter makes some decision based on the version.
This is especially important for raidz vdevs.
This is similar to what spa_load does.
This is not an issue for upstream because they do not seem to support
using raidz as a root pool.
Reported by: Andrei Lavreniyuk <andy.lavr@gmail.com>
Tested by: Andrei Lavreniyuk <andy.lavr@gmail.com>
MFC after: 6 days
The call is a NOP, because pool version in spa_ubsync.ub_version is not
initialized and thus appears to be zero.
If the version is properly set then the call leads to a NULL pointer
dereference because the spa object is still under-constructed.
The same change was independently made in the upstream as a part of
a larger change (4445fffbbb1ea25fd0e9ea68b9380dd7a6709025).
MFC after: 6 days
Executive code where similar invariant knobs exist.
o) Make the Simple Executive's warning function print "WARNING: " on the same
line as the warning it is displaying, rather than on a separate line.
After chatting with the MAC team, the TSF writes (at least on the 11n
MACs, I don't know about pre-11n MACs) are done as 64 bit writes that
can take some time. So, doing a 32 bit TSF write is definitely not
supported. Leave a comment here which explains that.
Whilst here, add a comment which outlines that after a reset or TSF
write, the TSF write may take a while (up to 50uS) to update.
A write or reset shouldn't be done whilst the previous one is in
flight. Also (and this isn't currently done) a read shouldn't
occur until the SLEEP32_TSF_WRITE_STAT is clear. Right now we're
not doing that, mostly because we haven't been doing lots of TSF
resets/writes until recently.
reducing the number of runtime checks done by the SDK code.
o) Group board/CPU information at early startup by subject matter, so that e.g.
CPU information is adjacent to CPU information and board information is
adjacent to board information.
TSF write.
The TSF_L32 update is fine for the AR5413 (and later, I guess) 11abg NICs
however on the 11n NICs this didn't work. The TSF writes were causing
a much larger time to be skipped, leading to the timing to never
converge.
I've tested this 64 bit TSF read, adjust and write on both the
11n NICs and the AR5413 NIC I've been using for testing. It works
fine on each.
This patch allows the AR5416/AR9280 to be used as a TDMA member.
I don't yet know why the AR9280 is ~7uS accurate rather than ~3uS;
I'll look into it soon.
Tested:
* AR5413, TDMA slave (~ 3us accuracy)
* AR5416, TDMA slave (~ 3us accuracy)
* AR9280, TDMA slave (~ 7us accuracy)
on the 802.11n NICs.
The 802.11n NICs return a TBTT value that continues far past the 16 bit
HAL_BEACON_PERIOD time (in TU.) The code would constrain nextslot to
HAL_BEACON_PERIOD, but it wasn't constraining nexttbtt - the pre-11n
NICs would only return TU values from 0 -> HAL_BEACON_PERIOD. Thus,
when nexttbtt exceeded 64 milliseconds, it would not wrap (but nextslot
did) which lead to a huge tsfdelta.
So until the slot calculation is converted to work in TSF rather than
a mix of TSF and TU, "make" the nexttbtt values match the TU assumptions
for pre-11n NICs.
This fixes the crazy deltatsf calculations but it doesn't fix the
non-convergent tsfdelta issue. That'll be fixed in a subsequent commit.
Rasperry Pi firmware has a set of hardcoded pathes it uses to fill
FDT with system-specific information like display resolution, memory
size, UART and SDHCI clocks, ethernet MAC address. Handle two of them:
- Add placeholder for ethernet MAC address
- Move display node out of "axi" node
