The superblock is also metadata of the filesystem so the relevant IO
should be tagged as such. We also tag it as high priority, as it's the
last block committed for metadata from a given transaction. Any delays
would effectively block the whole transaction, also blocking any other
operation holding the device_list_mutex.
Reviewed-by: NJosef Bacik <jbacik@fb.com>
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

    This patch provides a band aid to improve the 'out of the box'
behaviour of btrfs for disks that are detected as being an ssd.  In a
general purpose mixed workload scenario, the current ssd mode causes
overallocation of available raw disk space for data, while leaving
behind increasing amounts of unused fragmented free space. This
situation leads to early ENOSPC problems which are harming user
experience and adoption of btrfs as a general purpose filesystem.

This patch modifies the data extent allocation behaviour of the ssd mode
to make it behave identical to nossd mode.  The metadata behaviour and
additional ssd_spread option stay untouched so far.

Recommendations for future development are to reconsider the current
oversimplified nossd / ssd distinction and the broken detection
mechanism based on the rotational attribute in sysfs and provide
experienced users with a more flexible way to choose allocator behaviour
for data and metadata, optimized for certain use cases, while keeping
sane 'out of the box' default settings.  The internals of the current
btrfs code have more potential than what currently gets exposed to the
user to choose from.

    The SSD story...

    In the first year of btrfs development, around early 2008, btrfs
gained a mount option which enables specific functionality for
filesystems on solid state devices. The first occurance of this
functionality is in commit e18e4809, labeled "Add mount -o ssd, which
includes optimizations for seek free storage".

The effect on allocating free space for doing (data) writes is to
'cluster' writes together, writing them out in contiguous space, as
opposed to a 'tetris' way of putting all separate writes into any free
space fragment that fits (which is what the -o nossd behaviour does).

A somewhat simplified explanation of what happens is that, when for
example, the 'cluster' size is set to 2MiB, when we do some writes, the
data allocator will search for a free space block that is 2MiB big, and
put the writes in there. The ssd mode itself might allow a 2MiB cluster
to be composed of multiple free space extents with some existing data in
between, while the additional ssd_spread mount option kills off this
option and requires fully free space.

The idea behind this is (commit 536ac8ae): "The [...] clusters make it
more likely a given IO will completely overwrite the ssd block, so it
doesn't have to do an internal rwm cycle."; ssd block meaning nand erase
block. So, effectively this means applying a "locality based algorithm"
and trying to outsmart the actual ssd.

Since then, various changes have been made to the involved code, but the
basic idea is still present, and gets activated whenever the ssd mount
option is active. This also happens by default, when the rotational flag
as seen at /sys/block/<device>/queue/rotational is set to 0.

    However, there's a number of problems with this approach.

    First, what the optimization is trying to do is outsmart the ssd by
assuming there is a relation between the physical address space of the
block device as seen by btrfs and the actual physical storage of the
ssd, and then adjusting data placement. However, since the introduction
of the Flash Translation Layer (FTL) which is a part of the internal
controller of an ssd, these attempts are futile. The use of good quality
FTL in consumer ssd products might have been limited in 2008, but this
situation has changed drastically soon after that time. Today, even the
flash memory in your automatic cat feeding machine or your grandma's
wheelchair has a full featured one.

Second, the behaviour as described above results in the filesystem being
filled up with badly fragmented free space extents because of relatively
small pieces of space that are freed up by deletes, but not selected
again as part of a 'cluster'. Since the algorithm prefers allocating a
new chunk over going back to tetris mode, the end result is a filesystem
in which all raw space is allocated, but which is composed of
underutilized chunks with a 'shotgun blast' pattern of fragmented free
space. Usually, the next problematic thing that happens is the
filesystem wanting to allocate new space for metadata, which causes the
filesystem to fail in spectacular ways.

Third, the default mount options you get for an ssd ('ssd' mode enabled,
'discard' not enabled), in combination with spreading out writes over
the full address space and ignoring freed up space leads to worst case
behaviour in providing information to the ssd itself, since it will
never learn that all the free space left behind is actually free.  There
are two ways to let an ssd know previously written data does not have to
be preserved, which are sending explicit signals using discard or
fstrim, or by simply overwriting the space with new data.  The worst
case behaviour is the btrfs ssd_spread mount option in combination with
not having discard enabled. It has a side effect of minimizing the reuse
of free space previously written in.

Fourth, the rotational flag in /sys/ does not reliably indicate if the
device is a locally attached ssd. For example, iSCSI or NBD displays as
non-rotational, while a loop device on an ssd shows up as rotational.

The combination of the second and third problem effectively means that
despite all the good intentions, the btrfs ssd mode reliably causes the
ssd hardware and the filesystem structures and performance to be choked
to death. The clickbait version of the title of this story would have
been "Btrfs ssd optimizations considered harmful for ssds".

The current nossd 'tetris' mode (even still without discard) allows a
pattern of overwriting much more previously used space, causing many
more implicit discards to happen because of the overwrite information
the ssd gets. The actual location in the physical address space, as seen
from the point of view of btrfs is irrelevant, because the actual writes
to the low level flash are reordered anyway thanks to the FTL.

    Changes made in the code

1. Make ssd mode data allocation identical to tetris mode, like nossd.
2. Adjust and clean up filesystem mount messages so that we can easily
identify if a kernel has this patch applied or not, when providing
support to end users. Also, make better use of the *_and_info helpers to
only trigger messages on actual state changes.

    Backporting notes

Notes for whoever wants to backport this patch to their 4.9 LTS kernel:
* First apply commit 951e7966 "btrfs: drop the nossd flag when
  remounting with -o ssd", or fixup the differences manually.
* The rest of the conflicts are because of the fs_info refactoring. So,
  for example, instead of using fs_info, it's root->fs_info in
  extent-tree.c
Signed-off-by: NHans van Kranenburg <hans.van.kranenburg@mendix.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Although this bio has no data attached, it will reach this condition
(bio->bi_opf & REQ_PREFLUSH) and then update the flush_gen of dev_state
in __btrfsic_submit_bio. So we should still submit it through integrity
checker. Otherwise, the integrity checker will throw the following warning
when I mount a newly created btrfs filesystem.

[10264.755497] btrfs: attempt to write superblock which references block M @29523968 (sdb1/1111654400/0) which is not flushed out of disk's write cache (block flush_gen=1, dev->flush_gen=0)!
[10264.755498] btrfs: attempt to write superblock which references block M @29523968 (sdb1/37912576/0) which is not flushed out of disk's write cache (block flush_gen=1, dev->flush_gen=0)!
Signed-off-by: NLu Fengqi <lufq.fnst@cn.fujitsu.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Though BTRFS_FSID_SIZE and BTRFS_UUID_SIZE are of the same size, we
should use the matching constant for the fsid buffer.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The pinned chunks might be left over so we clean them but at this point
of close_ctree, there's noone to race with, the locking can be removed.
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Signed-off-by: NDavid Sterba <dsterba@suse.com>

Superblock is read and written using buffer heads, we need to set the
bdev blocksize. The magic constant has been hardcoded in several places,
so replace it with a named constant.
Signed-off-by: NDavid Sterba <dsterba@suse.com>

There are two independent parts, one that writes the superblocks and
another that waits for completion. No functional changes, but cleanups,
reformatting and comment updates.
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Signed-off-by: NDavid Sterba <dsterba@suse.com>

As we use per-chunk degradable check, the global
num_tolerated_disk_barrier_failures is of no use.

We can now remove it.
Signed-off-by: NQu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The last user of num_tolerated_disk_barrier_failures is
barrier_all_devices().
But it can be easily changed to the new per-chunk degradable check
framework.
Signed-off-by: NQu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Now use the btrfs_check_rw_degradable() to check if we can mount in the
degraded mode.

With this patch, we can mount in the following case:
 # mkfs.btrfs -f -m raid1 -d single /dev/sdb /dev/sdc
 # wipefs -a /dev/sdc
 # mount /dev/sdb /mnt/btrfs -o degraded
 As the single data chunk is only on sdb, so it's OK to mount as
 degraded, as missing one device is OK for RAID1.

But still fail in the following case as expected:
 # mkfs.btrfs -f -m raid1 -d single /dev/sdb /dev/sdc
 # wipefs -a /dev/sdb
 # mount /dev/sdc /mnt/btrfs -o degraded
 As the data chunk is only in sdb, so it's not OK to mount it as
 degraded.
Reported-by: NZhao Lei <zhaolei@cn.fujitsu.com>
Reported-by: NAnand Jain <anand.jain@oracle.com>
Signed-off-by: NQu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

We've started using cloned bios more in 4.13, there are some specifics
regarding the iteration.  Filipe found [1] that the raid56 iterated a
cloned bio using bio_for_each_segment_all, which is incorrect. The
cloned bios have wrong bi_vcnt and this could lead to silent
corruptions.  This patch adds assertions to all remaining
bio_for_each_segment_all cases.

[1] https://patchwork.kernel.org/patch/9838535/Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

We should really just wait in wait_dev_flush and let the caller decide
what to do with the error value.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Similar to what submit_bio_wait does, we should account for IO while
waiting for a bio completion. This has marginal visible effects, flush
bio is short-lived.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

For devices that support flushing, we allocate a bio, submit, wait for
it and then free it. The bio allocation does not fail so ENOMEM is not a
problem but we still may unnecessarily stress the allocation subsystem.

Instead, we can allocate the bio at the same time we allocate the device
and reuse it each time we need to flush the barriers. The bio is reset
before each use. Reference counting is simplified to just device
allocation (get) and freeing (put).

The bio used to be submitted through the integrity checker which will
find out that bio has no data attached and call submit_bio.

Status of the bio in flight needs to be tracked separately in case the
device caches get switched off between write and wait.
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Currently, percpu_counter_add is a wrapper around __percpu_counter_add
which is preempt safe due to explicit calls to preempt_disable.  Given
how __ prefix is used in percpu related interfaces, the naming
unfortunately creates the false sense that __percpu_counter_add is
less safe than percpu_counter_add.  In terms of context-safety,
they're equivalent.  The only difference is that the __ version takes
a batch parameter.

Make this a bit more explicit by just renaming __percpu_counter_add to
percpu_counter_add_batch.

This patch doesn't cause any functional changes.

tj: Minor updates to patch description for clarity.  Cosmetic
    indentation updates.
Signed-off-by: NNikolay Borisov <nborisov@suse.com>
Signed-off-by: NTejun Heo <tj@kernel.org>
Cc: Chris Mason <clm@fb.com>
Cc: Josef Bacik <jbacik@fb.com>
Cc: David Sterba <dsterba@suse.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: Jan Kara <jack@suse.com>
Cc: Jens Axboe <axboe@fb.com>
Cc: linux-mm@kvack.org
Cc: "David S. Miller" <davem@davemloft.net>

We can keep the state among the other fs_info flags, there's no reason
why fs_frozen would need to be separate.
Reviewed-by: NNikolay Borisov <nborisov@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Submit and wait parts of write_dev_flush() can be split into two
separate functions for better readability.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

There is no extra benefit to count null bdev during the submit loop,
as these null devices will be anyway checked during command
completion device loop just after the submit loop. We are holding the
device_list_mutex, the device->bdev status won't change in between.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Since commit "btrfs: btrfs_io_bio_alloc never fails, skip error handling"
write_dev_flush will not return ENOMEM in the sending part. We do not
need to check for it in the callers.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
[ updated changelog ]
Signed-off-by: NDavid Sterba <dsterba@suse.com>

We can hardcode GFP_NOFS to btrfs_io_bio_alloc, although it means we
change it back from GFP_KERNEL in scrub. I'd rather save a few stack
bytes from not passing the gfp flags in the remaining, more imporatant,
contexts and the bio allocating API now looks more consistent.
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Update direct callers of btrfs_io_bio_alloc that do error handling, that
we can now remove.
Reviewed-by: NAnand Jain <anand.jain@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Observing the number of slab objects of btrfs_transaction, there's just
one active on an almost quiescent filesystem, and the number of objects
goes to about ten when sync is in progress. Then the nubmer goes down to
1.  This matches the expectations of the transaction lifetime.

For such use the separate slab cache is not justified, as we do not
reuse objects frequently. For the shortlived transaction, the generic
slab (size 512) should be ok. We can optimistically expect that the 512
slabs are not all used (fragmentation) and there are free slots to take
when we do the allocation, compared to potentially allocating a whole new
page for the separate slab.

We'll lose the stats about the object use, which could be added later if
we really need them.
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Nothing checks its return value.

Is it safe to skip checking return value of btrfs_wait_tree_block_writeback?

Liu Bo: I think yes, it's used in walk_log_tree which is called in two
places, free_log_tree and log replay.  For free_log_tree, it waits for
any running writeback of the extent buffer under freeing to finish in
case we need to access the eb pointer from page->private, and it's OK to
not check the return value, while for log replay, it's doesn't wait
because wc->wait is not set. So neither cares about the writeback error.
Signed-off-by: NJeff Layton <jlayton@redhat.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
[ added more explanation to changelog, from Liu Bo ]
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The end io work queue items have been tracked by the work queues since
"Btrfs: Add async worker threads for pre and post IO checksumming"
(8b712842) (2008).
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The list used to track checksums in the early version (2.6.29), but I
was able not pinpoint the commit that stopped using it. Everything
apparently works without it for a long time.
Signed-off-by: NDavid Sterba <dsterba@suse.com>

For extent_io tree's we have carried the address_mapping of the inode
around in the io tree in order to pull the inode back out for calling
into various tree ops hooks.  This works fine when everything that has
an extent_io_tree has an inode.  But we are going to remove the
btree_inode, so we need to change this.  Instead just have a generic
void * for private data that we can initialize with, and have all the
tree ops use that instead.  This had a lot of cascading changes but
should be relatively straightforward.
Signed-off-by: NJosef Bacik <jbacik@fb.com>
Reviewed-by: NChandan Rajendra <chandan@linux.vnet.ibm.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
[ minor reordering of the callback prototypes ]
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The ->free_chunk_space variable is used to track the unallocated space
and access to it is protected by a spinlock, which is not used for
anything else.  Make the code a bit self-explanatory by switching the
variable to an atomic64_t type and kill the spinlock.
Signed-off-by: NNikolay Borisov <nborisov@suse.com>
[ not a performance critical code, use of atomic type is ok ]
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

This adds comments to the flush error handling part of the code, and
hopes to maintain the same logic with a framework which can be used to
handle the errors at the volume level.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Replace bi_error with a new bi_status to allow for a clear conversion.
Note that device mapper overloaded bi_error with a private value, which
we'll have to keep arround at least for now and thus propagate to a
proper blk_status_t value.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Commit b685d3d6 "block: treat REQ_FUA and REQ_PREFLUSH as
synchronous" removed REQ_SYNC flag from WRITE_{FUA|PREFLUSH|...}
definitions.  generic_make_request_checks() however strips REQ_FUA and
REQ_PREFLUSH flags from a bio when the storage doesn't report volatile
write cache and thus write effectively becomes asynchronous which can
lead to performance regressions

Fix the problem by making sure all bios which are synchronous are
properly marked with REQ_SYNC.

CC: David Sterba <dsterba@suse.com>
CC: linux-btrfs@vger.kernel.org
Fixes: b685d3d6Signed-off-by: NJan Kara <jack@suse.cz>
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

Commits cc8385b5 and 7ef70b4d added preallocation for the
reada radix trees and also switched them over to GFP_KERNEL for the
default gfp mask.

Since we're doing radix tree insertions under spinlocks, we need
to make sure the mask doesn't allow sleeping.  This fix keeps
the radix preallocation but switches back to the original gfp_mask.
Reported-by: NFilipe Manana <fdmanana@suse.com>
Signed-off-by: NChris Mason <clm@fb.com>

Allocate struct backing_dev_info separately instead of embedding it
inside superblock. This unifies handling of bdi among users.

CC: Chris Mason <clm@fb.com>
CC: Josef Bacik <jbacik@fb.com>
CC: David Sterba <dsterba@suse.com>
CC: linux-btrfs@vger.kernel.org
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NJens Axboe <axboe@fb.com>

The block layer call chain from submit_bio will check if the write cache
is enabled for the given queue before submitting the flush. This will
add a code to fail fast if its not.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
[ updated changelog to reflect current code stat, blkdev_issue_flush is
  not used yet ]
Signed-off-by: NDavid Sterba <dsterba@suse.com>

The last consumer of nobarriers is removed by the commit [1] and sync
won't fail with EOPNOTSUPP anymore. Thus, now when write cache is write
through it just return success without actually transpiring such a
request to the block device/lun.

[1]
commit b25de9d6
block: remove BIO_EOPNOTSUPP

And, as the device/lun write cache state may change dynamically saving
such as state won't help either. So deleting the member nobarriers.
Signed-off-by: NAnand Jain <anand.jain@oracle.com>
Reviewed-by: NDavid Sterba <dsterba@suse.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

We can read fs_info from eb.
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

We can preallocate the node so insertion does not have to do that under
the lock. The GFP flags for the global radix tree are initialized to
 GFP_NOFS & ~__GFP_DIRECT_RECLAIM
but we can use GFP_KERNEL, because readahead is optional and not on any
critical writeout path.
Reviewed-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

refcount_t type and corresponding API should be
used instead of atomic_t when the variable is used as
a reference counter. This allows to avoid accidental
refcounter overflows that might lead to use-after-free
situations.
Signed-off-by: NElena Reshetova <elena.reshetova@intel.com>
Signed-off-by: NHans Liljestrand <ishkamiel@gmail.com>
Signed-off-by: NKees Cook <keescook@chromium.org>
Signed-off-by: NDavid Windsor <dwindsor@gmail.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>

refcount_t type and corresponding API should be
used instead of atomic_t when the variable is used as
a reference counter. This allows to avoid accidental
refcounter overflows that might lead to use-after-free
situations.
Signed-off-by: NElena Reshetova <elena.reshetova@intel.com>
Signed-off-by: NHans Liljestrand <ishkamiel@gmail.com>
Signed-off-by: NKees Cook <keescook@chromium.org>
Signed-off-by: NDavid Windsor <dwindsor@gmail.com>
Signed-off-by: NDavid Sterba <dsterba@suse.com>