It's pretty obvious that, we reserve qgroup meta rsv in
btrfs_subvolume_reserve_metadata(), but doesn't have corresponding
release/convert calls in btrfs_subvolume_release_metadata().
This leads to the leakage.
[FIX]
To fix this bug, we should follow what we're doing in
btrfs_delalloc_reserve_metadata(), where we reserve qgroup space, and
add it to block_rsv->qgroup_rsv_reserved.
And free the qgroup reserved metadata space when releasing the
block_rsv.
To do this, we need to change the btrfs_subvolume_release_metadata() to
accept btrfs_root, and record the qgroup_to_release number, and call
btrfs_qgroup_convert_reserved_meta() for it.
Was seen in testing. The fix is similar to that of
btrfs: open device without device_list_mutex
where we're holding the device_list_mutex and then grab the bd_mutex,
which pulls in a bunch of dependencies under the bd_mutex. We only ever
call btrfs_close_devices() on mount failure or unmount, so we're save to
not have the device_list_mutex here. We're already holding the
uuid_mutex which keeps us safe from any external modification of the
fs_devices.
Josef Bacik [Thu, 20 Aug 2020 15:18:27 +0000 (11:18 -0400)]
btrfs: move btrfs_rm_dev_replace_free_srcdev outside of all locks
When closing and freeing the source device we could end up doing our
final blkdev_put() on the bdev, which will grab the bd_mutex. As such
we want to be holding as few locks as possible, so move this call
outside of the dev_replace->lock_finishing_cancel_unmount lock. Since
we're modifying the fs_devices we need to make sure we're holding the
uuid_mutex here, so take that as well.
Nikolay Borisov [Wed, 12 Aug 2020 13:26:46 +0000 (16:26 +0300)]
btrfs: remove alloc_list splice in btrfs_prepare_sprout
btrfs_prepare_sprout is called when the first rw device is added to a
seed filesystem. This means the filesystem can't have its alloc_list
be non-empty, since seed filesystems are read only. Simply remove the
code altogether.
Nikolay Borisov [Thu, 16 Jul 2020 07:25:33 +0000 (10:25 +0300)]
btrfs: switch seed device to list api
While this patch touches a bunch of files the conversion is
straighforward. Instead of using the implicit linked list anchored at
btrfs_fs_devices::seed the code is switched to using
list_for_each_entry.
Previous patches in the series already factored out code that processed
both main and seed devices so in those cases the factored out functions
are called on the main fs_devices and then on every seed dev inside
list_for_each_entry.
Using list api also allows to simplify deletion from the seed dev list
performed in btrfs_rm_device and btrfs_rm_dev_replace_free_srcdev by
substituting a while() loop with a simple list_del_init.
Nikolay Borisov [Wed, 15 Jul 2020 10:48:49 +0000 (13:48 +0300)]
btrfs: simplify setting/clearing fs_info to btrfs_fs_devices
It makes no sense to have sysfs-related routines be responsible for
properly initialising the fs_info pointer of struct btrfs_fs_device.
Instead this can be streamlined by making it the responsibility of
btrfs_init_devices_late to initialize it. That function already
initializes fs_info of every individual device in btrfs_fs_devices.
As far as clearing it is concerned it makes sense to move it to
close_fs_devices. That function is only called when struct
btrfs_fs_devices is no longer in use - either for holding seeds or
main devices for a mounted filesystem.
Nikolay Borisov [Wed, 15 Jul 2020 10:48:48 +0000 (13:48 +0300)]
btrfs: make close_fs_devices return void
The return value of this function conveys absolutely no information.
All callers already check the state of fs_devices->opened to decide how
to proceed. So convert the function to returning void. While at it make
btrfs_close_devices also return void.
Nikolay Borisov [Mon, 3 Aug 2020 09:58:46 +0000 (12:58 +0300)]
btrfs: remove err variable from btrfs_get_extent
There's no practical reason too use 'err' as a variable to convey
errors. In fact it's value is either set explicitly in the beginning of
the function or it simply takes the value of 'ret'. Not conforming to
the usual pattern of having ret be the only variable used to convey
errors makes the code more error prone to bugs. In fact one such bug
was introduced by 6bf9e4bd6a27 ("btrfs: inode: Verify inode mode toi
avoid NULL pointer dereference") by assigning the error value to 'ret'
and not 'err'.
Let's fix that issue and make the function less tricky by leaving only
ret to convey error values.
Josef Bacik [Thu, 3 Sep 2020 15:16:51 +0000 (11:16 -0400)]
btrfs: dio iomap DSYNC workaround
iomap dio will run generic_write_sync() for us if the iocb is DSYNC.
This is problematic for us because of 2 reasons:
1. we hold the inode_lock() during this operation, and we take it in
generic_write_sync()
2. we hold a read lock on the dio_sem but take the write lock in fsync
Since we don't want to rip out this code right now, but reworking the
locking is a bit much to do at this point, work around this problem with
this masterpiece of a patch.
First, we clear DSYNC on the iocb so that the iomap stuff doesn't know
that it needs to handle the sync. We save this fact in
current->journal_info, because we need to see do special things once
we're in iomap_begin, and we have no way to pass private information
into iomap_dio_rw().
Next we specify a separate iomap_dio_ops for sync, which implements an
->end_io() callback that gets called when the dio completes. This is
important for AIO, because we really do need to run generic_write_sync()
if we complete asynchronously. However if we're still in the submitting
context when we enter ->end_io() we clear the flag so that the submitter
knows they're the ones that needs to run generic_write_sync().
This is meant to be temporary. We need to work out how to eliminate the
inode_lock() and the dio_sem in our fsync and use another mechanism to
protect these operations.
We're using direct io implementation based on buffer heads. This patch
switches to the new iomap infrastructure.
Switch from __blockdev_direct_IO() to iomap_dio_rw(). Rename
btrfs_get_blocks_direct() to btrfs_dio_iomap_begin() and use it as
iomap_begin() for iomap direct I/O functions. This function allocates
and locks all the blocks required for the I/O. btrfs_submit_direct() is
used as the submit_io() hook for direct I/O ops.
Since we need direct I/O reads to go through iomap_dio_rw(), we change
file_operations.read_iter() to a btrfs_file_read_iter() which calls
btrfs_direct_IO() for direct reads and falls back to
generic_file_buffered_read() for incomplete reads and buffered reads.
We don't need address_space.direct_IO() anymore: set it to noop.
Similarly, we don't need flags used in __blockdev_direct_IO(). iomap is
capable of direct I/O reads from a hole, so we don't need to return
-ENOENT.
Btrfs direct I/O is now done under i_rwsem, shared in case of reads and
exclusive in case of writes. This guards against simultaneous truncates.
Use iomap->iomap_end() to check for failed or incomplete direct I/O:
- for writes, call __endio_write_update_ordered()
- for reads, unlock extents
btrfs_dio_data is now hooked in iomap->private and not
current->journal_info. It carries the reservation variable and the
amount of data submitted, so we can calculate the amount of data to call
__endio_write_update_ordered in case of an error.
This patch removes last use of struct buffer_head from btrfs.
Qu Wenruo [Thu, 20 Aug 2020 07:42:46 +0000 (15:42 +0800)]
btrfs: add owner and fs_info to alloc_state io_tree
Commit 1c11b63eff2a ("btrfs: replace pending/pinned chunks lists with io
tree") introduced btrfs_device::alloc_state extent io tree, but it
doesn't initialize the fs_info and owner member.
This means the following features are not properly supported:
- Fs owner report for insert_state() error
Without fs_info initialized, although btrfs_err() won't panic, it
won't output which fs is causing the error.
- Wrong owner for trace events
alloc_state will get the owner as pinned extents.
Fix this by assiging proper fs_info and owner for
btrfs_device::alloc_state.
Since it's inclusion on 9afc66498a0b ("btrfs: block-group: refactor how
we read one block group item") this function always returned 0, so there
is no need to check for the returned value.
The compilation with W=1 generates the following warnings:
fs/btrfs/sysfs.c:1630:6: warning: variable 'ret' set but not used [-Wunused-but-set-variable]
1630 | int ret;
| ^~~
fs/btrfs/sysfs.c:1629:6: warning: variable 'features' set but not used [-Wunused-but-set-variable]
1629 | u64 features;
| ^~~~~~~~
[ The unused variables are leftover from e410e34fad91 ("Revert "btrfs:
synchronize incompat feature bits with sysfs files""), which needs
to be properly fixed by moving feature bit manipulation from the sysfs
context. Silence the warning to save pepople time, we got several
reports. ]
Filipe Manana [Tue, 11 Aug 2020 11:43:58 +0000 (12:43 +0100)]
btrfs: make fast fsyncs wait only for writeback
Currently regardless of a full or a fast fsync we always wait for ordered
extents to complete, and then start logging the inode after that. However
for fast fsyncs we can just wait for the writeback to complete, we don't
need to wait for the ordered extents to complete since we use the list of
modified extents maps to figure out which extents we must log and we can
get their checksums directly from the ordered extents that are still in
flight, otherwise look them up from the checksums tree.
Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at
fsync time"), for fast fsyncs, we used to start logging without even
waiting for the writeback to complete first, we would wait for it to
complete after logging, while holding a transaction open, which lead to
performance issues when using cgroups and probably for other cases too,
as wait for IO while holding a transaction handle should be avoided as
much as possible. After that, for fast fsyncs, we started to wait for
ordered extents to complete before starting to log, which adds some
latency to fsyncs and we even got at least one report about a performance
drop which bisected to that particular change:
This change makes fast fsyncs only wait for writeback to finish before
starting to log the inode, instead of waiting for both the writeback to
finish and for the ordered extents to complete. This brings back part of
the logic we had that extracts checksums from in flight ordered extents,
which are not yet in the checksums tree, and making sure transaction
commits wait for the completion of ordered extents previously logged
(by far most of the time they have already completed by the time a
transaction commit starts, resulting in no wait at all), to avoid any
data loss if an ordered extent completes after the transaction used to
log an inode is committed, followed by a power failure.
When there are no other tasks accessing the checksums and the subvolume
btrees, the ordered extent completion is pretty fast, typically taking
100 to 200 microseconds only in my observations. However when there are
other tasks accessing these btrees, ordered extent completion can take a
lot more time due to lock contention on nodes and leaves of these btrees.
I've seen cases over 2 milliseconds, which starts to be significant. In
particular when we do have concurrent fsyncs against different files there
is a lot of contention on the checksums btree, since we have many tasks
writing the checksums into the btree and other tasks that already started
the logging phase are doing lookups for checksums in the btree.
This change also turns all ranged fsyncs into full ranged fsyncs, which
is something we already did when not using the NO_HOLES features or when
doing a full fsync. This is to guarantee we never miss checksums due to
writeback having been triggered only for a part of an extent, and we end
up logging the full extent but only checksums for the written range, which
results in missing checksums after log replay. Allowing ranged fsyncs to
operate again only in the original range, when using the NO_HOLES feature
and doing a fast fsync is doable but requires some non trivial changes to
the writeback path, which can always be worked on later if needed, but I
don't think they are a very common use case.
Several tests were performed using fio for different numbers of concurrent
jobs, each writing and fsyncing its own file, for both sequential and
random file writes. The tests were run on bare metal, no virtualization,
on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device,
with a kernel configuration that is the default of typical distributions
(debian in this case), without debug options enabled (kasan, kmemleak,
slub debug, debug of page allocations, lock debugging, etc).
The following script that calls fio was used:
$ cat test-fsync.sh
#!/bin/bash
DEV=/dev/nvme0n1
MNT=/mnt/btrfs
MOUNT_OPTIONS="-o ssd -o space_cache=v2"
MKFS_OPTIONS="-d single -m single"
if [ $# -ne 5 ]; then
echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]"
exit 1
fi
Filipe Manana [Tue, 11 Aug 2020 11:43:48 +0000 (12:43 +0100)]
btrfs: do not commit logs and transactions during link and rename operations
Since commit d4682ba03ef618 ("Btrfs: sync log after logging new name") we
started to commit logs, and fallback to transaction commits when we failed
to log the new names or commit the logs, after link and rename operations
when the target inodes (or their parents) were previously logged in the
current transaction. This was to avoid losing directories despite an
explicit fsync on them when they are ancestors of some inode that got a
new named logged, due to a link or rename operation. However that adds the
cost of starting IO and waiting for it to complete, which can cause higher
latencies for applications.
Instead of doing that, just make sure that when we log a new name for an
inode we don't mark any of its ancestors as logged, so that if any one
does an fsync against any of them, without doing any other change on them,
the fsync commits the log. This way we only pay the cost of a log commit
(or a transaction commit if something goes wrong or a new block group was
created) if the application explicitly asks to fsync any of the parent
directories.
Using dbench, which mixes several filesystems operations including renames,
revealed some significant latency gains. The following script that uses
dbench was used to test this:
echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
mkfs.btrfs -f $MKFS_OPTIONS $DEV
mount $MOUNT_OPTIONS $DEV $MNT
dbench -t 300 -D $MNT $THREADS
umount $MNT
The test was run on bare metal, no virtualization, on a box with 12 cores
(Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel
configuration that is the default of typical distributions (debian in this
case), without debug options enabled (kasan, kmemleak, slub debug, debug
of page allocations, lock debugging, etc).
Filipe Manana [Tue, 11 Aug 2020 11:43:37 +0000 (12:43 +0100)]
btrfs: do not take the log_mutex of the subvolume when pinning the log
During a rename we pin the log to make sure no one commits a log that
reflects an ongoing rename operation, as it might result in a committed
log where it recorded the unlink of the old name without having recorded
the new name. However we are taking the subvolume's log_mutex before
incrementing the log_writers counter, which is not necessary since that
counter is atomic and we only remove the old name from the log and add
the new name to the log after we have incremented log_writers, ensuring
that no one can commit the log after we have removed the old name from
the log and before we added the new name to the log.
By taking the log_mutex lock we are just adding unnecessary contention on
the lock, which can become visible for workloads that mix renames with
fsyncs, writes for files opened with O_SYNC and unlink operations (if the
inode or its parent were fsynced before in the current transaction).
So just remove the lock and unlock of the subvolume's log_mutex at
btrfs_pin_log_trans().
Using dbench, which mixes different types of operations that end up taking
that mutex (fsyncs, renames, unlinks and writes into files opened with
O_SYNC) revealed some small gains. The following script that calls dbench
was used:
echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
mkfs.btrfs -f $MKFS_OPTIONS $DEV
mount $MOUNT_OPTIONS $DEV $MNT
dbench -s -t 600 -D $MNT $THREADS
umount $MNT
The test was run on bare metal, no virtualization, on a box with 12 cores
(Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel
configuration that is the default of typical distributions (debian in this
case), without debug options enabled (kasan, kmemleak, slub debug, debug
of page allocations, lock debugging, etc).
David Sterba [Mon, 17 Aug 2020 10:16:57 +0000 (12:16 +0200)]
btrfs: send: remove indirect callback parameter for changed_cb
There's a custom callback passed to btrfs_compare_trees which happens to
be named exactly same as the existing function implementing it. This is
confusing and the indirection is not necessary for our needs. Compiler
is clever enough to call it directly so there's effectively no change.
David Sterba [Mon, 17 Aug 2020 10:12:38 +0000 (12:12 +0200)]
btrfs: scrub: rename ratelimit state varaible to avoid shadowing
There's already defined _rs within ctree.h:btrfs_printk_ratelimited,
local variables should not use _ to avoid such name clashes with
macro-local variables.
David Sterba [Mon, 17 Aug 2020 08:58:38 +0000 (10:58 +0200)]
btrfs: compression: move declarations to header
The declarations of compression algorithm callbacks are defined in the
.c file as they're used from there. Compiler warns that there are no
declarations for public functions when compiling lzo.c/zlib.c/zstd.c.
Fix that by moving the declarations to the header as it's the common
place for all of them.
David Sterba [Mon, 17 Aug 2020 08:56:00 +0000 (10:56 +0200)]
btrfs: remove const from btrfs_feature_set_name
The function btrfs_feature_set_name returns a const char pointer, the
second const is not necessary and reported as a warning:
In file included from fs/btrfs/space-info.c:6:
fs/btrfs/sysfs.h:16:1: warning: type qualifiers ignored on function return type [-Wignored-qualifiers]
16 | const char * const btrfs_feature_set_name(enum btrfs_feature_set set);
| ^~~~~
Qu Wenruo [Thu, 13 Aug 2020 06:33:52 +0000 (14:33 +0800)]
btrfs: cleanup calculation of lockend in lock_and_cleanup_extent_if_need()
We're just doing rounding up to sectorsize to calculate the lockend.
There is no need to do the unnecessary length calculation, just direct
round_up() is enough.
Josef Bacik [Tue, 25 Aug 2020 20:56:59 +0000 (16:56 -0400)]
btrfs: fix possible infinite loop in data async reclaim
Dave reported an issue where generic/102 would sometimes hang. This
turned out to be because we'd get into this spot where we were no longer
making progress on data reservations because our exit condition was not
met. The log is basically
while (!space_info->full && !list_empty(&space_info->tickets))
flush_space(space_info, flush_state);
where flush state is our various flush states, but doesn't include
ALLOC_CHUNK_FORCE. This is because we actually lead with allocating
chunks, and so the assumption was that once you got to the actual
flushing states you could no longer allocate chunks. This was a stupid
assumption, because you could have deleted block groups that would be
reclaimed by a transaction commit, thus unsetting space_info->full.
This is essentially what happens with generic/102, and so sometimes
you'd get stuck in the flushing loop because we weren't allocating
chunks, but flushing space wasn't giving us what we needed to make
progress.
Fix this by adding ALLOC_CHUNK_FORCE to the end of our flushing states,
that way we will eventually bail out because we did end up with
space_info->full if we free'd a chunk previously. Otherwise, as is the
case for this test, we'll allocate our chunk and continue on our happy
merry way.
Josef Bacik [Tue, 21 Jul 2020 14:22:34 +0000 (10:22 -0400)]
btrfs: add a comment explaining the data flush steps
The data flushing steps are not obvious to people other than myself and
Chris. Write a giant comment explaining the reasoning behind each flush
step for data as well as why it is in that particular order.
Josef Bacik [Tue, 21 Jul 2020 14:22:33 +0000 (10:22 -0400)]
btrfs: do async reclaim for data reservations
Now that we have the data ticketing stuff in place, move normal data
reservations to use an async reclaim helper to satisfy tickets. Before
we could have multiple tasks race in and both allocate chunks, resulting
in more data chunks than we would necessarily need. Serializing these
allocations and making a single thread responsible for flushing will
only allocate chunks as needed, as well as cut down on transaction
commits and other flush related activities.
Priority reservations will still work as they have before, simply
trying to allocate a chunk until they can make their reservation.
Josef Bacik [Tue, 21 Jul 2020 14:22:32 +0000 (10:22 -0400)]
btrfs: flush delayed refs when trying to reserve data space
We can end up with freed extents in the delayed refs, and thus
may_commit_transaction() may not think we have enough pinned space to
commit the transaction and we'll ENOSPC early. Handle this by running
the delayed refs in order to make sure pinned is uptodate before we try
to commit the transaction.
Josef Bacik [Tue, 21 Jul 2020 14:22:31 +0000 (10:22 -0400)]
btrfs: run delayed iputs before committing the transaction for data
Before we were waiting on iputs after we committed the transaction, but
this doesn't really make much sense. We want to reclaim any space we
may have in order to be more likely to commit the transaction, due to
pinned space being added by running the delayed iputs. Fix this by
making delayed iputs run before committing the transaction.
Josef Bacik [Tue, 21 Jul 2020 14:22:30 +0000 (10:22 -0400)]
btrfs: don't force commit if we are data
We used to unconditionally commit the transaction at least 2 times and
then on the 3rd try check against pinned space to make sure committing
the transaction was worth the effort. This is overkill, we know nobody
is going to steal our reservation, and if we can't make our reservation
with the pinned amount simply bail out.
This also cleans up the passing of bytes_needed to
may_commit_transaction, as that was the thing we added into place in
order to accomplish this behavior. We no longer need it so remove that
mess.
Josef Bacik [Tue, 21 Jul 2020 14:22:29 +0000 (10:22 -0400)]
btrfs: drop the commit_cycles stuff for data reservations
This was an old wart left over from how we previously did data
reservations. Before we could have people race in and take a
reservation while we were flushing space, so we needed to make sure we
looped a few times before giving up. Now that we're using the ticketing
infrastructure we don't have to worry about this and can drop the logic
altogether.
Josef Bacik [Tue, 21 Jul 2020 14:22:28 +0000 (10:22 -0400)]
btrfs: use the same helper for data and metadata reservations
Now that data reservations follow the same pattern as metadata
reservations we can simply rename __reserve_metadata_bytes to
__reserve_bytes and use that helper for data reservations.
Things to keep in mind, btrfs_can_overcommit() returns 0 for data,
because we can never overcommit. We also will never pass in FLUSH_ALL
for data, so we'll simply be added to the priority list and go straight
into handle_reserve_ticket.
Josef Bacik [Tue, 21 Jul 2020 14:22:27 +0000 (10:22 -0400)]
btrfs: serialize data reservations if we are flushing
Nikolay reported a problem where generic/371 would fail sometimes with a
slow drive. The gist of the test is that we fallocate a file in
parallel with a pwrite of a different file. These two files combined
are smaller than the file system, but sometimes the pwrite would ENOSPC.
A fair bit of investigation uncovered the fact that the fallocate
workload was racing in and grabbing the free space that the pwrite
workload was trying to free up so it could make its own reservation.
After a few loops of this eventually the pwrite workload would error out
with an ENOSPC.
We've had the same problem with metadata as well, and we serialized all
metadata allocations to satisfy this problem. This wasn't usually a
problem with data because data reservations are more straightforward,
but obviously could still happen.
Fix this by not allowing reservations to occur if there are any pending
tickets waiting to be satisfied on the space info.
Josef Bacik [Tue, 21 Jul 2020 14:22:26 +0000 (10:22 -0400)]
btrfs: use ticketing for data space reservations
Now that we have all the infrastructure in place, use the ticketing
infrastructure to make data allocations. This still maintains the exact
same flushing behavior, but now we're using tickets to get our
reservations satisfied.
Josef Bacik [Tue, 21 Jul 2020 14:22:25 +0000 (10:22 -0400)]
btrfs: add btrfs_reserve_data_bytes and use it
Create a new function btrfs_reserve_data_bytes() in order to handle data
reservations. This uses the new flush types and flush states to handle
making data reservations.
This patch specifically does not change any functionality, and is
purposefully not cleaned up in order to make bisection easier for the
future patches. The new helper is identical to the old helper in how it
handles data reservations. We first try to force a chunk allocation,
and then we run through the flush states all at once and in the same
order that they were done with the old helper.
Subsequent patches will clean this up and change the behavior of the
flushing, and it is important to keep those changes separate so we can
easily bisect down to the patch that caused the regression, rather than
the patch that made us start using the new infrastructure.
Josef Bacik [Tue, 21 Jul 2020 14:22:24 +0000 (10:22 -0400)]
btrfs: add the data transaction commit logic into may_commit_transaction
Data space flushing currently unconditionally commits the transaction
twice in a row, and the last time it checks if there's enough pinned
extents to satisfy its reservation before deciding to commit the
transaction for the 3rd and final time.
Encode this logic into may_commit_transaction(). In the next patch we
will pass in U64_MAX for bytes_needed the first two times, and the final
time we will pass in the actual bytes we need so the normal logic will
apply.
This patch exists solely to make the logical changes I will make to the
flushing state machine separate to make it easier to bisect any
performance related regressions.
Josef Bacik [Tue, 21 Jul 2020 14:22:23 +0000 (10:22 -0400)]
btrfs: add flushing states for handling data reservations
Currently the way we do data reservations is by seeing if we have enough
space in our space_info. If we do not and we're a normal inode we'll
1) Attempt to force a chunk allocation until we can't anymore.
2) If that fails we'll flush delalloc, then commit the transaction, then
run the delayed iputs.
If we are a free space inode we're only allowed to force a chunk
allocation. In order to use the normal flushing mechanism we need to
encode this into a flush state array for normal inodes. Since both will
start with allocating chunks until the space info is full there is no
need to add this as a flush state, this will be handled specially.
Josef Bacik [Tue, 21 Jul 2020 14:22:22 +0000 (10:22 -0400)]
btrfs: check tickets after waiting on ordered extents
Right now if the space is freed up after the ordered extents complete
(which is likely since the reservations are held until they complete),
we would do extra delalloc flushing before we'd notice that we didn't
have any more tickets. Fix this by moving the tickets check after our
wait_ordered_extents check.
Josef Bacik [Tue, 21 Jul 2020 14:22:21 +0000 (10:22 -0400)]
btrfs: use btrfs_start_delalloc_roots in shrink_delalloc
The original iteration of flushing had us flushing delalloc and then
checking to see if we could make our reservation, thus we were very
careful about how many pages we would flush at once.
But now that everything is async and we satisfy tickets as the space
becomes available we don't have to keep track of any of this, simply
try and flush the number of dirty inodes we may have in order to
reclaim space to make our reservation. This cleans up our delalloc
flushing significantly.
The async_pages stuff is dropped because btrfs_start_delalloc_roots()
handles the case that we generate async extents for us, so we no longer
require this extra logic.
Josef Bacik [Tue, 21 Jul 2020 14:22:20 +0000 (10:22 -0400)]
btrfs: use the btrfs_space_info_free_bytes_may_use helper for delalloc
We are going to use the ticket infrastructure for data, so use the
btrfs_space_info_free_bytes_may_use() helper in
btrfs_free_reserved_data_space_noquota() so we get the
btrfs_try_granting_tickets call when we free our reservation.
Josef Bacik [Tue, 21 Jul 2020 14:22:18 +0000 (10:22 -0400)]
btrfs: call btrfs_try_granting_tickets when unpinning anything
When unpinning we were only calling btrfs_try_granting_tickets() if
global_rsv->space_info == space_info, which is problematic because we
use ticketing for SYSTEM chunks, and want to use it for DATA as well.
Fix this by moving this call outside of that if statement.
Josef Bacik [Tue, 21 Jul 2020 14:22:17 +0000 (10:22 -0400)]
btrfs: call btrfs_try_granting_tickets when freeing reserved bytes
We were missing a call to btrfs_try_granting_tickets in
btrfs_free_reserved_bytes, so add it to handle the case where we're able
to satisfy an allocation because we've freed a pending reservation.
Josef Bacik [Tue, 21 Jul 2020 14:22:16 +0000 (10:22 -0400)]
btrfs: make ALLOC_CHUNK use the space info flags
We have traditionally used flush_space() to flush metadata space, so
we've been unconditionally using btrfs_metadata_alloc_profile() for our
profile to allocate a chunk. However if we're going to use this for
data we need to use btrfs_get_alloc_profile() on the space_info we pass
in.
Josef Bacik [Tue, 21 Jul 2020 14:22:15 +0000 (10:22 -0400)]
btrfs: make shrink_delalloc take space_info as an arg
Currently shrink_delalloc just looks up the metadata space info, but
this won't work if we're trying to reclaim space for data chunks. We
get the right space_info we want passed into flush_space, so simply pass
that along to shrink_delalloc.
Josef Bacik [Tue, 21 Jul 2020 14:22:12 +0000 (10:22 -0400)]
btrfs: change nr to u64 in btrfs_start_delalloc_roots
We have btrfs_wait_ordered_roots() which takes a u64 for nr, but
btrfs_start_delalloc_roots() that takes an int for nr, which makes using
them in conjunction, especially for something like (u64)-1, annoying and
inconsistent. Fix btrfs_start_delalloc_roots() to take a u64 for nr and
adjust start_delalloc_inodes() and it's callers appropriately.
This means we've adjusted start_delalloc_inodes() to take a pointer of
nr since we want to preserve the ability for start-delalloc_inodes() to
return an error, so simply make it do the nr adjusting as necessary.
Part of adjusting the callers to this means changing
btrfs_writeback_inodes_sb_nr() to take a u64 for items. This may be
confusing because it seems unrelated, but the caller of
btrfs_writeback_inodes_sb_nr() already passes in a u64, it's just the
function variable that needs to be changed.
Nikolay Borisov [Wed, 12 Aug 2020 13:18:51 +0000 (16:18 +0300)]
btrfs: remove fsid argument from btrfs_sysfs_update_sprout_fsid
It can be accessed from 'fs_devices' as it's identical to
fs_info->fs_devices. Also add a comment about why we are calling the
function. No semantic changes.
Marc Zyngier [Mon, 5 Oct 2020 14:02:17 +0000 (15:02 +0100)]
gpio: pca953x: Survive spurious interrupts
The pca953x driver never checks the result of irq_find_mapping(),
which returns 0 when no mapping is found. When a spurious interrupt
is delivered (which can happen under obscure circumstances), the
kernel explodes as it still tries to handle the error code as
a real interrupt.
Handle this particular case and warn on spurious interrupts.
Tony Luck [Tue, 6 Oct 2020 21:09:10 +0000 (14:09 -0700)]
x86/mce: Decode a kernel instruction to determine if it is copying from user
All instructions copying data between kernel and user memory
are tagged with either _ASM_EXTABLE_UA or _ASM_EXTABLE_CPY
entries in the exception table. ex_fault_handler_type() returns
EX_HANDLER_UACCESS for both of these.
Recovery is only possible when the machine check was triggered
on a read from user memory. In this case the same strategy for
recovery applies as if the user had made the access in ring3. If
the fault was in kernel memory while copying to user there is no
current recovery plan.
For MOV and MOVZ instructions a full decode of the instruction
is done to find the source address. For MOVS instructions
the source address is in the %rsi register. The function
fault_in_kernel_space() determines whether the source address is
kernel or user, upgrade it from "static" so it can be used here.
Tony Luck [Tue, 6 Oct 2020 21:09:09 +0000 (14:09 -0700)]
x86/mce: Recover from poison found while copying from user space
Existing kernel code can only recover from a machine check on code that
is tagged in the exception table with a fault handling recovery path.
Add two new fields in the task structure to pass information from
machine check handler to the "task_work" that is queued to run before
the task returns to user mode:
+ mce_vaddr: will be initialized to the user virtual address of the fault
in the case where the fault occurred in the kernel copying data from
a user address. This is so that kill_me_maybe() can provide that
information to the user SIGBUS handler.
+ mce_kflags: copy of the struct mce.kflags needed by kill_me_maybe()
to determine if mce_vaddr is applicable to this error.
Add code to recover from a machine check while copying data from user
space to the kernel. Action for this case is the same as if the user
touched the poison directly; unmap the page and send a SIGBUS to the task.
Use a new helper function to share common code between the "fault
in user mode" case and the "fault while copying from user" case.
New code paths will be activated by the next patch which sets
MCE_IN_KERNEL_COPYIN.
Tony Luck [Tue, 6 Oct 2020 21:09:08 +0000 (14:09 -0700)]
x86/mce: Avoid tail copy when machine check terminated a copy from user
In the page fault case it is ok to see if a few more unaligned bytes
can be copied from the source address. Worst case is that the page fault
will be triggered again.
Machine checks are more serious. Just give up at the point where the
main copy loop triggered the #MC and return from the copy code as if
the copy succeeded. The machine check handler will use task_work_add() to
make sure that the task is sent a SIGBUS.
Ben Chuang [Mon, 5 Oct 2020 10:55:09 +0000 (18:55 +0800)]
mmc: sdhci-pci-gli: Add CQHCI Support for GL9763E
Add CQHCI initialization and implement CQHCI operations for GL9763E.
Use bit19 of the register (0x888) to decide whether to disable command
queuing. If the bit is set, the command queuing will be disabled.
Youquan Song [Tue, 6 Oct 2020 21:09:07 +0000 (14:09 -0700)]
x86/mce: Add _ASM_EXTABLE_CPY for copy user access
_ASM_EXTABLE_UA is a general exception entry to record the exception fixup
for all exception spots between kernel and user space access.
To enable recovery from machine checks while coping data from user
addresses it is necessary to be able to distinguish the places that are
looping copying data from those that copy a single byte/word/etc.
Add a new macro _ASM_EXTABLE_CPY and use it in place of _ASM_EXTABLE_UA
in the copy functions.
Record the exception reason number to regs->ax at
ex_handler_uaccess which is used to check MCE triggered.
The new fixup routine ex_handler_copy() is almost an exact copy of
ex_handler_uaccess() The difference is that it sets regs->ax to the trap
number. Following patches use this to avoid trying to copy remaining
bytes from the tail of the copy and possibly hitting the poison again.
New mce.kflags bit MCE_IN_KERNEL_COPYIN will be used by mce_severity()
calculation to indicate that a machine check is recoverable because the
kernel was copying from user space.
Youquan Song [Tue, 6 Oct 2020 21:09:05 +0000 (14:09 -0700)]
x86/mce: Pass pointer to saved pt_regs to severity calculation routines
New recovery features require additional information about processor
state when a machine check occurred. Pass pt_regs down to the routines
that need it.
Mike Travis [Mon, 5 Oct 2020 20:39:28 +0000 (15:39 -0500)]
x86/platform/uv: Update for UV5 NMI MMR changes
The UV NMI MMR addresses and fields moved between UV4 and UV5
necessitating a rewrite of the UV NMI handler. Adjust references
to accommodate those changes.
Mike Travis [Mon, 5 Oct 2020 20:39:26 +0000 (15:39 -0500)]
x86/platform/uv: Update node present counting
The changes in the UV5 arch shrunk the NODE PRESENT table to just 2x64
entries (128 total) so are in to 64 bit MMRs instead of a depth of 64
bits in an array. Adjust references when counting up the nodes present.
Mike Travis [Mon, 5 Oct 2020 20:39:22 +0000 (15:39 -0500)]
x86/platform/uv: Add and decode Arch Type in UVsystab
When the UV BIOS starts the kernel it passes the UVsystab info struct to
the kernel which contains information elements more specific than ACPI,
and generally pertinent only to the MMRs. These are read only fields
so information is passed one way only. A new field starting with UV5 is
the UV architecture type so the ACPI OEM_ID field can be used for other
purposes going forward. The UV Arch Type selects the entirety of the
MMRs available, with their addresses and fields defined in uv_mmrs.h.
Mike Travis [Mon, 5 Oct 2020 20:39:21 +0000 (15:39 -0500)]
x86/platform/uv: Add UV5 direct references
Add new references to UV5 (and UVY class) system MMR addresses and
fields primarily caused by the expansion from 46 to 52 bits of physical
memory address.
Mike Travis [Mon, 5 Oct 2020 20:39:20 +0000 (15:39 -0500)]
x86/platform/uv: Update UV MMRs for UV5
Update UV MMRs in uv_mmrs.h for UV5 based on Verilog output from the
UV Hub hardware design files. This is the next UV architecture with
a new class (UVY) being defined for 52 bit physical address masks.
Uses a bitmask for UV arch identification so a single test can cover
multiple versions. Includes other adjustments to match the uv_mmrs.h
file to keep from encountering compile errors. New UV5 functionality
is added in the patches that follow.
Mike Travis [Tue, 6 Oct 2020 21:34:27 +0000 (16:34 -0500)]
drivers/misc/sgi-xp: Adjust references in UV kernel modules
Remove the define is_uv() is_uv_system and just use the latter as is.
This removes a conflict with a new symbol in the generated uv_mmrs.h
file (is_uv()).
Mike Travis [Mon, 5 Oct 2020 20:39:18 +0000 (15:39 -0500)]
x86/platform/uv: Remove SCIR MMR references for UV systems
UV class systems no longer use System Controller for monitoring of CPU
activity provided by this driver. Other methods have been developed for
BIOS and the management controller (BMC). Remove that supporting code.
Mike Travis [Mon, 5 Oct 2020 20:39:17 +0000 (15:39 -0500)]
x86/platform/uv: Remove UV BAU TLB Shootdown Handler
The Broadcast Assist Unit (BAU) TLB shootdown handler is being rewritten
to become the UV BAU APIC driver. It is designed to speed up sending
IPIs to selective CPUs within the system. Remove the current TLB
shutdown handler (tlb_uv.c) file and a couple of kernel hooks in the
interim.
In nvme_set_queue_limits() we initialize vwc to false and later add
a condition to set vwc true. The value of the vwc can be declare
initialized which makes all the blk_queue_XXX() calls uniform.
In nvme_validate_ns() the exra variable ctrl is used only twice.
Using ns->ctrl directly still maintains the redability and original
length of the lines in the code. Get rid of the extra variable ctrl &
use ns->ctrl directly.
Move the logic to revalidate the block_device size or remove the
namespace from the caller into nvme_validate_ns. This removes
the return value and thus the status code translation. Additionally
it also catches non-permanent errors from nvme_update_ns_info using
the existing logic.
nvme: query namespace identifiers before adding the namespace
Check the namespace identifier list first thing when scanning namespaces.
This keeps the code to query the CSI common between the alloc and validate
path, and helps to structure the code better for multiple command set
support.
Now that the queue is frozen before updating ->lba_shift we can't hit the
invalid references mentioned in the comment any more. More importantly
this code would not have helped us if the format was changed by another
controller or through implementation defined back channels.
Ensure that there can't be any I/O in flight went we change the disk
geometry in nvme_update_ns_info, most notable the LBA size by lifting
the queue free from nvme_update_disk_info into the caller
Factor out a helper from nvme_update_ns_info that configures the
per-namespaces metadata and PI settings. Also make sure the helpers
clear the flags explicitly instead of all of ->features to allow for
potentially reusing ->features for future non-metadata flags.
nvme: call nvme_identify_ns as the first thing in nvme_alloc_ns_block
Check if the namespace actually exists as the very first thing and don't
bother with any extra work if not. This should speed up and simplify
the sequential scanning for NVMe 1.0 devices.
Rename _nvme_revalidate_disk to nvme_validate_ns to better describe
what the function does, and pass the struct nvme_ns instead of the
gendisk to better match the call chain.
The removal of the ->revalidate_disk method broke the initialization of
the zone bitmaps, as nvme_revalidate_disk now never gets called during
initialization.
Move the zone related code from nvme_revalidate_disk into a new helper in
zns.c, and call it from nvme_alloc_ns in addition to nvme_validate_ns to
ensure the zone bitmaps are initialized during probe.
block: optimize blk_queue_zoned_model for !CONFIG_BLK_DEV_ZONED
Always return BLK_ZONED_NONE if zoned device support is not enabled.
This allows various compiler optimizations including the dead code
elimination that we so like for avoiding ifdefs.
The function nvme_init_ctrl() gets the ctrl reference & when it fails it
does put the ctrl reference in the error unwind code.
When creating loop ctrl in nvme_loop_create_ctrl() if nvme_init_ctrl()
returns non zero (i.e. error) value it jumps to the "out_put_ctrl" label
which calls nvme_put_ctrl(), that will lead to douple ctrl put in error
unwind path.
Update nvme_loop_create_ctrl() such that this patch removes the
"out_put_ctrl" label, add a new "out" label after nvme_put_ctrl() in
error unwind path and jump to newly added label when nvme_init_ctrl()
call retuns an error.
projects / linux.git / log