Coly Li [Fri, 13 Oct 2017 23:35:29 +0000 (16:35 -0700)]
bcache: check ca->alloc_thread initialized before wake up it
In bcache code, sysfs entries are created before all resources get
allocated, e.g. allocation thread of a cache set.
There is posibility for NULL pointer deference if a resource is accessed
but which is not initialized yet. Indeed Jorg Bornschein catches one on
cache set allocation thread and gets a kernel oops.
The reason for this bug is, when bch_bucket_alloc() is called during
cache set registration and attaching, ca->alloc_thread is not properly
allocated and initialized yet, call wake_up_process() on ca->alloc_thread
triggers NULL pointer deference failure. A simple and fast fix is, before
waking up ca->alloc_thread, checking whether it is allocated, and only
wake up ca->alloc_thread when it is not NULL.
Signed-off-by: Coly Li <colyli@suse.de> Reported-by: Jorg Bornschein <jb@capsec.org> Cc: Kent Overstreet <kent.overstreet@gmail.com> Cc: stable@vger.kernel.org Reviewed-by: Michael Lyle <mlyle@lyle.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Peter Foley [Fri, 13 Oct 2017 23:35:28 +0000 (16:35 -0700)]
bcache: Avoid nested function definition
Fixes below error with clang:
../drivers/md/bcache/sysfs.c:759:3: error: function definition is not allowed here
{ return *((uint16_t *) r) - *((uint16_t *) l); }
^
../drivers/md/bcache/sysfs.c:789:32: error: use of undeclared identifier 'cmp'
sort(p, n, sizeof(uint16_t), cmp, NULL);
^
2 errors generated.
v2:
rename function to __bch_cache_cmp
Signed-off-by: Peter Foley <pefoley2@pefoley.com> Reviewed-by: Coly Li <colyli@suse.de> Reviewed-by: Michael Lyle <mlyle@lyle.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Yafang Shao [Sat, 14 Oct 2017 08:38:27 +0000 (16:38 +0800)]
mm/page-writeback.c: make changes of dirty_writeback_centisecs take effect immediately
This patch is the followup of the prvious patch:
[writeback: schedule periodic writeback with sysctl].
There's another issue to fix.
For example,
- When the tunable was set to one hour and is reset to one second, the
new setting will not take effect for up to one hour.
Kicking the flusher threads immediately fixes it.
Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Fri, 13 Oct 2017 15:36:06 +0000 (09:36 -0600)]
Revert "lightnvm: prevent bd removal if busy"
Christoph correctly points out that this issue is no different
for other block devices, and poking at cross layer internals
is not the right way to solve it.
Javier González [Fri, 13 Oct 2017 12:46:47 +0000 (14:46 +0200)]
lightnvm: implement generic path for sync I/O
Implement a generic path for sending sync I/O on LightNVM. This allows
to reuse the standard synchronous path trough blk_execute_rq(), instead
of implementing a wait_for_completion on the target side (e.g., pblk).
Javier González [Fri, 13 Oct 2017 12:46:45 +0000 (14:46 +0200)]
lightnvm: pblk: avoid being reported as hung on rated GC
The amount of GC I/O on the write buffer is managed by the rate-limiter,
which is calculated as a function of the number of available free
blocks. When reaching the stable point, we risk having scheduled more
I/Os for GC than are allowed on the write buffer. This would result on
the GC semaphore balancing the outstanding read GC I/Os to be reported
as "hung", though the behavior is normal.
Solve this by allowing to schedule when we detect that the read GC path
is not moving forward.
Hans Holmberg [Fri, 13 Oct 2017 12:46:43 +0000 (14:46 +0200)]
lightnvm: pblk: remove spinlock when freeing line metadata
Lockdep complains about being in atomic context while freeing line
metadata - and rightly so as we take a spinlock and end up calling
vfree that might sleep(in pblk_mfree).
There is no need for holding the line manager free_lock while
freeing line metadata as the pipeline as stopped, so remove the lock.
Hans Holmberg [Fri, 13 Oct 2017 12:46:41 +0000 (14:46 +0200)]
lightnvm: pblk: gc all lines in the pipeline before exit
Finish garbage collect of the lines that are in the gc pipeline
before exiting. Ensure that all lines already in in the pipeline
goes through, from read to write.
Do this by keeping track of how many lines are in the pipeline
and waiting for that number to reach zero before exiting the gc
reader task.
Since we're adding a new gc line counter, change the name of
inflight_gc to read_inflight_gc to make the distinction clear.
Hans Holmberg [Fri, 13 Oct 2017 12:46:40 +0000 (14:46 +0200)]
lightnvm: pblk: add l2p crc debug printouts
Print the CRC of the logical-to-physical mapping during exit and
after recovering the L2P table to facilitate detection of meta
data corruption/recovery issues.
The CRC printed after recovery should match the CRC printed during
the previous exit - if it doesn't this indicates that either the meta
data written to the disk is corrupt or recovery failed.
Hans Holmberg [Fri, 13 Oct 2017 12:46:38 +0000 (14:46 +0200)]
lightnvm: pblk: consider bad sectors in emeta during recovery
When recovering lines we need to consider that bad blocks in a line
affect the emeta area size.
Previously it was assumed that the emeta area would grow by the
number of sectors per page * number of bad blocks in the line.
This assumption is not correct - the number of "extra" pages that are
consumed could be both smaller (depending on emeta size) and bigger
(depending on the placement of the bad blocks).
Fix this by calculating the emeta start by iterating backwards
through the line, skipping ppas that map to bad blocks.
Also fix the data types used for ppa indices/counts in
pblk_recov_l2p_from_emeta - we should use u64.
Hans Holmberg [Fri, 13 Oct 2017 12:46:36 +0000 (14:46 +0200)]
lightnvm: pblk: free full lines during recovery
When rebuilding the L2P table, any full lines (lines without any
valid sectors) will be identified. If these lines are not freed,
we risk not being able to allocate the first data line.
This patch refactors the part of GC that frees empty lines
into a separate function and adds a call to this after the
L2P table has been rebuilt.
Hans Holmberg [Fri, 13 Oct 2017 12:46:34 +0000 (14:46 +0200)]
lightnvm: pblk: prevent gc kicks when gc is not operational
GC can be kicked after it has been shut down when closing the last
line during exit, resulting in accesses to freed structures.
Make sure that GC is not triggered while it is not operational.
Also make sure that GC won't be re-activated during exit when
running on another processor by using timer_del_sync.
Rakesh Pandit [Fri, 13 Oct 2017 12:46:28 +0000 (14:46 +0200)]
lightnvm: pblk: fix changing GC group list for a line
pblk_line_gc_list seems to had a bug since the introduction of pblk in
getting GC list for a line. In b20ba1bc7 while redesigning the GC
algorithm, the naming for the GC thresholds was altered, but the
values for high_thrs and mid_thrs were not. The result is that when
moving to the GC lists, the mid threshold is never evaluated.
Javier González [Fri, 13 Oct 2017 12:46:27 +0000 (14:46 +0200)]
lightnvm: pblk: ensure right bad block calculation
Make sure that the variable controlling block threshold for allocating
extra metadata sectors in case of a line with bad blocks does not get a
negative value. Otherwise, the line will be marked as corrupted and
wasted.
Javier González [Fri, 13 Oct 2017 12:46:26 +0000 (14:46 +0200)]
lightnvm: pblk: enable 1 LUN configuration
Metadata I/Os are scheduled to minimize their impact on user data I/Os.
When there are enough LUNs instantiated (i.e., enough bandwidth), it is
easy to interleave metadata and data one after the other so that
metadata I/Os are the ones being blocked and not vice-versa.
We do this by calculating the distance between the I/Os in terms of the
LUNs that are not in used, and selecting a free LUN that satisfies a
the simple heuristic that metadata is scheduled behind. The per-LUN
semaphores guarantee consistency. This works fine on >1 LUN
configuration. However, when a single LUN is instantiated, this design
leads to a deadlock, where metadata waits to be scheduled on a free LUN.
This patch implements the 1 LUN case by simply scheduling the metadada
I/O after the data I/O. In the process, we refactor the way a line is
replaced to ensure that metadata writes are submitted after data writes
in order to guarantee block sequentiality. Note that, since there is
only one LUN, both I/Os will block each other by design. However, such
configuration only pursues tight read latencies, not write bandwidth.
Javier González [Fri, 13 Oct 2017 12:46:25 +0000 (14:46 +0200)]
lightnvm: pblk: remove I/O dependency on write path
pblk schedules user I/O, metadata I/O and erases on the write path in
order to minimize collisions at the media level. Until now, there has
been a dependency between user and metadata I/Os that could lead to a
deadlock as both take the per-LUN semaphore to schedule submission.
This path removes this dependency and guarantees forward progress at a
per I/O granurality.
Javier González [Fri, 13 Oct 2017 12:46:24 +0000 (14:46 +0200)]
lightnvm: pblk: remove redundant check on read path
A partial read I/O in pblk is an I/O where some sectors reside in the
write buffer in main memory and some are persisted on the device. Such
an I/O must at least contain 2 lbas, therefore checking for the case
where a single lba is mapped is not necessary.
Javier González [Fri, 13 Oct 2017 12:46:23 +0000 (14:46 +0200)]
lightnvm: pblk: guarantee line integrity on reads
When a line is recycled during garbage collection, reads can still be
issued to the line. If the line is freed in the middle of this process,
data corruption might occur.
This patch guarantees that lines are not freed in the middle of reads
that target them (lines). Specifically, we use the existing line
reference to decide when a line is eligible for being freed after the
recycle process.
Javier González [Fri, 13 Oct 2017 12:46:22 +0000 (14:46 +0200)]
lightnvm: pblk: check lba sanity on read path
As part of pblk's recovery scheme, we store the lba mapped to each
physical sector on the device's out-of-bound (OOB) area.
On the read path, we can use this information to validate that the data
being delivered to the upper layers corresponds to the lba being
requested. The cost of this check is an extra copy on the DMA region on
the device and an extra comparison in the host, given that (i) the OOB
area is being read together with the data in the media, and (ii) the DMA
region allocated for the ppa list can be reused for the metadata stored
on the OOB area.
Javier González [Fri, 13 Oct 2017 12:46:19 +0000 (14:46 +0200)]
lightnvm: pblk: improve naming for internal req.
Each request type sent to the LightNVM subsystem requires different
metadata. Until now, we have tailored this metadata based on write, read
and erase commands. However, pblk uses different metadata for internal
writes that do not hit the write buffer. Instead of abusing the metadata
for reads, create a new request type - internal write to improve
code readability.
In the process, create internal values for each I/O type instead of
abusing the READ/WRITE macros, as suggested by Christoph.
Javier González [Fri, 13 Oct 2017 12:46:17 +0000 (14:46 +0200)]
lightnvm: pblk: simplify path on REQ_PREFLUSH
On REQ_PREFLUSH, directly tag the I/O context flags to signal a flush in
the write to cache path, instead of finding the correct entry context
and imposing a memory barrier. This simplifies the code and might
potentially prevent race conditions when adding functionality to the
write path.
Javier González [Fri, 13 Oct 2017 12:46:15 +0000 (14:46 +0200)]
lightnvm: pblk: refactor read path on GC
Simplify the part of the garbage collector where data is read from the
line being recycled and moved into an internal queue before being copied
to the memory buffer. This allows to get rid of a dedicated function,
which introduces an unnecessary dependency on the code.
Javier González [Fri, 13 Oct 2017 12:46:14 +0000 (14:46 +0200)]
lightnvm: pblk: simplify data validity check on GC
When a line is selected for recycling by the garbage collector (GC), the
line state changes and the invalid bitmap is frozen, preventing
invalidations from happening. Throughout the GC, the L2P map is checked
to verify that not data being recycled has been updated. The last check
is done before the new map is being stored on the L2P table. Though
this algorithm works, it requires a number of corner cases to be checked
each time the L2P table is being updated. This complicates readability
and is error prone in case that the recycling algorithm is modified.
Instead, this patch makes the invalid bitmap accessible even when the
line is being recycled. When recycled data is being remapped, it is
enough to check the invalid bitmap for the line before updating the L2P
table.
Javier González [Fri, 13 Oct 2017 12:46:09 +0000 (14:46 +0200)]
lightnvm: pblk: do not use a mempool for line bitmaps
pblk holds two sector bitmaps: one to keep track of the mapped sectors
while the line is active and another one to keep track of the invalid
sectors. The latter is kept during the whole live of the line, until it
is recycled. Since we cannot guarantee forward progress for the mempool
in this case, get rid of the mempool and simply allocate memory through
kmalloc.
Javier González [Fri, 13 Oct 2017 12:46:07 +0000 (14:46 +0200)]
lightnvm: pblk: simplify work_queue mempool
In pblk, we have a mempool to allocate a generic structure that we
pass along workqueues. This is heavily used in the GC path in order
to have enough inflight reads and fully utilize the GC bandwidth.
However, the current GC path copies data to the host memory and puts it
back into the write buffer. This requires a vmalloc allocation for the
data and a memory copy. Thus, guaranteeing the allocation by using a
mempool for the structure in itself does not give us much. Until we
implement support for vector copy to avoid moving data through the host,
just allocate the workqueue structure using kmalloc.
Javier González [Fri, 13 Oct 2017 12:46:06 +0000 (14:46 +0200)]
lightnvm: pblk: fix min size for page mempool
pblk uses an internal page mempool for allocating pages on internal
bios. The main two users of this memory pool are partial reads (reads
with some sectors in cache and some on media) and padded writes, which
need to add dummy pages to an existing bio already containing valid
data (and with a large enough bioset allocated). In both cases, the
maximum number of pages per bio is defined by the maximum number of
physical sectors supported by the underlying device.
This patch fixes a bad mempool allocation, where the min_nr of elements
on the pool was fixed (to 16), which is lower than the maximum number
of sectors supported by NVMe (as of the time for this patch). Instead,
use the maximum number of allowed sectors reported by the device.
Javier González [Fri, 13 Oct 2017 12:46:03 +0000 (14:46 +0200)]
lightnvm: pblk: free padded entries in write buffer
When a REQ_FLUSH reaches pblk, the bio cannot be directly completed.
Instead, data on the write buffer is flushed and the bio is completed on
the completion pah. This might require some sectors to be padded in
order to guarantee a successful write.
This patch fixes a memory leak on the padded pages. A consequence of
this bad free was that internal bios not containing data (only a flush)
were not being completed.
Javier González [Fri, 13 Oct 2017 12:46:02 +0000 (14:46 +0200)]
lightnvm: pblk: use right flag for GC allocation
The data buffer for the GC path allocates virtual memory through
vmalloc. When this change was introduced, a flag signaling kmalloc'ed
memory was wrongly introduced. Use the right flag when creating a bio
from this buffer.
Fixes: de54e703a422 ("lightnvm: pblk: use vmalloc for GC data buffer") Signed-off-by: Javier González <javier@cnexlabs.com> Signed-off-by: Matias Bjørling <m@bjorling.me> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rakesh Pandit [Fri, 13 Oct 2017 12:46:00 +0000 (14:46 +0200)]
lightnvm: pblk: reuse pblk_gc_should_kick
This is a trivial change which reuses pblk_gc_should_kick instead of
repeating it again in pblk_rl_free_lines_inc.
Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Made it apply to the common case. Reviewed-by: Javier González <javier@cnexlabs.com> Signed-off-by: Matias Bjørling <m@bjorling.me> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rakesh Pandit [Fri, 13 Oct 2017 12:45:56 +0000 (14:45 +0200)]
lightnvm: pblk: protect line bitmap while submitting meta io
It seems pblk_dealloc_page would race against pblk_alloc_pages for
line bitmap for sector allocation.The chances are very low but might
as well protect the bitmap properly.
Rakesh Pandit [Fri, 13 Oct 2017 12:45:55 +0000 (14:45 +0200)]
lightnvm: include NVM Express driver if OCSSD is selected for build
Because NVM needs BLK_DEV_NVME, select it automatically if we mark NVM
in config file before building kernel. Also append PCI to depends as
select doesn't automatically add dependencies.
Rakesh Pandit [Fri, 13 Oct 2017 12:45:52 +0000 (14:45 +0200)]
lightnvm: protect target type list with correct locks
nvm_tgt_types list was protected by wrong lock for NVM_INFO ioctl call
and can race with addition or removal of target types. Also
unregistering target type was not protected correctly.
Rakesh Pandit [Fri, 13 Oct 2017 12:45:51 +0000 (14:45 +0200)]
lightnvm: prevent bd removal if busy
When a virtual block device is formatted and mounted after creating
with "nvme lnvm create... -t pblk", a removal from "nvm lnvm remove"
would result in this:
446416.309757] bdi-block not registered
[446416.309773] ------------[ cut here ]------------
[446416.309780] WARNING: CPU: 3 PID: 4319 at fs/fs-writeback.c:2159
__mark_inode_dirty+0x268/0x340
Ideally removal should return -EBUSY as block device is mounted after
formatting. This patch tries to address this checking if whole device
or any partition of it already mounted or not before removal.
Whole device is checked using "bd_super" member of block device. This
member is always set once block device has been mounted using a
filesystem. Another member "bd_part_count" takes care of checking any
if any partitions are under use. "bd_part_count" is only updated
under locks when partitions are opened or closed (first open and last
release). This at least does take care sending -EBUSY if removal is
being attempted while whole block device or any partition is mounted.
Rakesh Pandit [Fri, 13 Oct 2017 12:45:50 +0000 (14:45 +0200)]
lightnvm: prevent target type module removal when in use
If target type module e.g. pblk here is unloaded (rmmod) while module
is in use (after creating target) system crashes. We fix this by
using module API refcnt.
Shaohua Li [Sat, 7 Oct 2017 00:56:00 +0000 (17:56 -0700)]
block: set request_list for request
Legacy queue sets request's request_list, mq doesn't. This makes mq does
the same thing, so we can find cgroup of a request. Note, we really
only use blkg field of request_list, it's pointless to allocate mempool
for request_list in mq case.
Signed-off-by: Shaohua Li <shli@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Shaohua Li [Sat, 7 Oct 2017 00:55:59 +0000 (17:55 -0700)]
blk-stat: delete useless code
Fix two issues:
- the per-cpu stat flush is unnecessary, nobody uses per-cpu stat except
sum it to global stat. We can do the calculation there. The flush just
wastes cpu time.
- some fields are signed int/s64. I don't see the point.
Jiufei Xue [Tue, 10 Oct 2017 03:13:32 +0000 (11:13 +0800)]
blk-throttle: fix null pointer dereference while throttling writeback IOs
A null pointer dereference can occur when blkcg is removed manually
with writeback IOs inflight. This is caused by the following case:
Writeback kworker submit the bio and set bio->bi_cg_private to tg
in blk_throtl_assoc_bio.
Then we remove the block cgroup manually, the blkg and tg would be
freed if there is no request inflight.
When the submitted bio come back, blk_throtl_bio_endio() fetch the tg
which was already freed.
Fix this by increasing the refcount of blkg in funcion
blk_throtl_assoc_bio() so that the blkg will not be freed until the
bio_endio called.
Rakesh Pandit [Mon, 9 Oct 2017 10:34:41 +0000 (13:34 +0300)]
writeback: merge try_to_writeback_inodes_sb_nr() into caller
Since commit 925a6efb8ff0c ("Btrfs: stop using
try_to_writeback_inodes_sb_nr to flush delalloc") this function hasn't
been used outside so stop exporting it.
In addition we merge it into try_to_writeback_inodes_sb() which is the
only caller. Also change return type of try_to_writeback_inodes_sb to
void as the only user ext4 doesn't care.
Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Rakesh Pandit <rakesh@tuxera.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Paolo Valente [Mon, 9 Oct 2017 11:11:23 +0000 (13:11 +0200)]
block, bfq: fix unbalanced decrements of burst size
The commit "block, bfq: decrease burst size when queues in burst
exit" introduced the decrement of burst_size on the removal of a
bfq_queue from the burst list. Unfortunately, this decrement can
happen to be performed even when burst size is already equal to 0,
because of unbalanced decrements. A description follows of the cause
of these unbalanced decrements, namely a wrong assumption, and of the
way how this wrong assumption leads to unbalanced decrements.
The wrong assumption is that a bfq_queue can exit only if the process
associated with the bfq_queue has exited. This is false, because a
bfq_queue, say Q, may exit also as a consequence of a merge with
another bfq_queue. In this case, Q exits because the I/O of its
associated process has been redirected to another bfq_queue.
The decrement unbalance occurs because Q may then be re-created after
a split, and added back to the current burst list, *without*
incrementing burst_size. burst_size is not incremented because Q is
not a new bfq_queue added to the burst list, but a bfq_queue only
temporarily removed from the list, and, before the commit "bfq-sq,
bfq-mq: decrease burst size when queues in burst exit", burst_size was
not decremented when Q was removed.
This commit addresses this issue by just checking whether the exiting
bfq_queue is a merged bfq_queue, and, in that case, not decrementing
burst_size. Unfortunately, this still leaves room for unbalanced
decrements, in the following rarer case: on a split, the bfq_queue
happens to be inserted into a different burst list than that it was
removed from when merged. If this happens, the number of elements in
the new burst list becomes higher than burst_size (by one). When the
bfq_queue then exits, it is of course not in a merged state any
longer, thus burst_size is decremented, which results in an unbalanced
decrement. To handle this sporadic, unlucky case in a simple way,
this commit also checks that burst_size is larger than 0 before
decrementing it.
Finally, this commit removes an useless, extra check: the check that
the bfq_queue is sync, performed before checking whether the bfq_queue
is in the burst list. This extra check is redundant, because only sync
bfq_queues can be inserted into the burst list.
Fixes: 7cb04004fa37 ("block, bfq: decrease burst size when queues in burst exit") Reported-by: Philip Müller <philm@manjaro.org> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com> Tested-by: Philip Müller <philm@manjaro.org> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Tested-by: Lee Tibbert <lee.tibbert@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Luca Miccio [Mon, 9 Oct 2017 14:27:21 +0000 (16:27 +0200)]
block,bfq: Disable writeback throttling
Similarly to CFQ, BFQ has its write-throttling heuristics, and it
is better not to combine them with further write-throttling
heuristics of a different nature.
So this commit disables write-back throttling for a device if BFQ
is used as I/O scheduler for that device.
Signed-off-by: Luca Miccio <lucmiccio@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Tested-by: Lee Tibbert <lee.tibbert@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Yafang Shao [Mon, 9 Oct 2017 21:44:13 +0000 (05:44 +0800)]
writeback: schedule periodic writeback with sysctl
After disable periodic writeback by writing 0 to
dirty_writeback_centisecs, the handler wb_workfn() will not be
entered again until the dirty background limit reaches or
sync syscall is executed or no enough free memory available or
vmscan is triggered.
So the periodic writeback can't be enabled by writing a non-zero
value to dirty_writeback_centisecs.
As it can be disabled by sysctl, it should be able to enable by
sysctl as well.
Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
writeback: eliminate work item allocation in bd_start_writeback()
Handle start-all writeback like we do periodic or kupdate
style writeback - by marking the bdi_writeback as needing a full
flush, and simply waking the thread. This eliminates the need to
allocate and queue a specific work item just for this purpose.
After this change, we truly only ever have one of them running at
any point in time. We mark the need to start all flushes, and the
writeback thread will clear it once it has processed the request.
Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Wed, 4 Oct 2017 17:22:24 +0000 (11:22 -0600)]
blk-mq: document the need to have STARTED and COMPLETED share a byte
For memory ordering guarantees on stores, we need to ensure that
these two bits share the same byte of storage in the unsigned
long. Add a comment as to why, and a BUILD_BUG_ON() to ensure that
we don't violate this requirement.
Peter Zijlstra [Wed, 6 Sep 2017 08:00:22 +0000 (10:00 +0200)]
blk-mq: attempt to fix atomic flag memory ordering
Attempt to untangle the ordering in blk-mq. The patch introducing the
single smp_mb__before_atomic() is obviously broken in that it doesn't
clearly specify a pairing barrier and an obtained guarantee.
The comment is further misleading in that it hints that the
deadline store and the COMPLETE store also need to be ordered, but
AFAICT there is no such dependency. However what does appear to be
important is the clear happening _after_ the store, and that worked by
pure accident.
This clarifies blk_mq_start_request() -- we should not get there with
STARTING set -- this simplifies the code and makes the barrier usage
sane (the old code could be read to allow not having _any_ atomic after
the barrier, in which case the barrier hasn't got anything to order). We
then also introduce the missing pairing barrier for it.
Also down-grade the barrier to smp_wmb(), this is cheaper for
PowerPC/ARM and doesn't cost anything extra on x86.
And it documents the STARTING vs COMPLETE ordering. Although I've not
been entirely successful in reverse engineering the blk-mq state
machine so there might still be more funnies around timeout vs
requeue.
If I got anything wrong, feel free to educate me by adding comments to
clarify things ;-)
Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Will Deacon <will.deacon@arm.com> Cc: Ming Lei <tom.leiming@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Andrea Parri <parri.andrea@gmail.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Bart Van Assche <bart.vanassche@wdc.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Fixes: 538b75341835 ("blk-mq: request deadline must be visible before marking rq as started") Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Paolo Valente [Thu, 21 Sep 2017 09:04:03 +0000 (11:04 +0200)]
block, bfq: decrease burst size when queues in burst exit
If many queues belonging to the same group happen to be created
shortly after each other, then the concurrent processes associated
with these queues have typically a common goal, and they get it done
as soon as possible if not hampered by device idling. Examples are
processes spawned by git grep, or by systemd during boot. As for
device idling, this mechanism is currently necessary for weight
raising to succeed in its goal: privileging I/O. In view of these
facts, BFQ does not provide the above queues with either weight
raising or device idling.
On the other hand, a burst of queue creations may be caused also by
the start-up of a complex application. In this case, these queues need
usually to be served one after the other, and as quickly as possible,
to maximise responsiveness. Therefore, in this case the best strategy
is to weight-raise all the queues created during the burst, i.e., the
exact opposite of the strategy for the above case.
To distinguish between the two cases, BFQ uses an empirical burst-size
threshold, found through extensive tests and monitoring of daily
usage. Only large bursts, i.e., burst with a size above this
threshold, are considered as generated by a high number of parallel
processes. In this respect, upstart-based boot proved to be rather
hard to detect as generating a large burst of queue creations, because
with upstart most of the queues created in a burst exit *before* the
next queues in the same burst are created. To address this issue, I
changed the burst-detection mechanism so as to not decrease the size
of the current burst even if one of the queues in the burst is
eliminated.
Unfortunately, this missing decrease causes false positives on very
fast systems: on the start-up of a complex application, such as
libreoffice writer, so many queues are created, served and exited
shortly after each other, that a large burst of queue creations is
wrongly detected as occurring. These false positives just disappear if
the size of a burst is decreased when one of the queues in the burst
exits. This commit restores the missing burst-size decrease, relying
of the fact that upstart is apparently unlikely to be used on systems
running this and future versions of the kernel.
Paolo Valente [Thu, 21 Sep 2017 09:04:02 +0000 (11:04 +0200)]
block, bfq: let early-merged queues be weight-raised on split too
A just-created bfq_queue, say Q, may happen to be merged with another
bfq_queue on the very first invocation of the function
__bfq_insert_request. In such a case, even if Q would clearly deserve
interactive weight raising (as it has just been created), the function
bfq_add_request does not make it to be invoked for Q, and thus to
activate weight raising for Q. As a consequence, when the state of Q
is saved for a possible future restore, after a split of Q from the
other bfq_queue(s), such a state happens to be (unjustly)
non-weight-raised. Then the bfq_queue will not enjoy any weight
raising on the split, even if should still be in an interactive
weight-raising period when the split occurs.
This commit solves this problem as follows, for a just-created
bfq_queue that is being early-merged: it stores directly, in the saved
state of the bfq_queue, the weight-raising state that would have been
assigned to the bfq_queue if not early-merged.
Paolo Valente [Thu, 21 Sep 2017 09:04:01 +0000 (11:04 +0200)]
block, bfq: check and switch back to interactive wr also on queue split
As already explained in the message of commit "block, bfq: fix
wrong init of saved start time for weight raising", if a soft
real-time weight-raising period happens to be nested in a larger
interactive weight-raising period, then BFQ restores the interactive
weight raising at the end of the soft real-time weight raising. In
particular, BFQ checks whether the latter has ended only on request
dispatches.
Unfortunately, the above scheme fails to restore interactive weight
raising in the following corner case: if a bfq_queue, say Q,
1) Is merged with another bfq_queue while it is in a nested soft
real-time weight-raising period. The weight-raising state of Q is
then saved, and not considered any longer until a split occurs.
2) Is split from the other bfq_queue(s) at a time instant when its
soft real-time weight raising is already finished.
On the split, while resuming the previous, soft real-time
weight-raised state of the bfq_queue Q, BFQ checks whether the
current soft real-time weight-raising period is actually over. If so,
BFQ switches weight raising off for Q, *without* checking whether the
soft real-time period was actually nested in a non-yet-finished
interactive weight-raising period.
This commit addresses this issue by adding the above missing check in
bfq_queue splits, and restoring interactive weight raising if needed.
Paolo Valente [Thu, 21 Sep 2017 09:04:00 +0000 (11:04 +0200)]
block, bfq: fix wrong init of saved start time for weight raising
This commit fixes a bug that causes bfq to fail to guarantee a high
responsiveness on some drives, if there is heavy random read+write I/O
in the background. More precisely, such a failure allowed this bug to
be found [1], but the bug may well cause other yet unreported
anomalies.
BFQ raises the weight of the bfq_queues associated with soft real-time
applications, to privilege the I/O, and thus reduce latency, for these
applications. This mechanism is named soft-real-time weight raising in
BFQ. A soft real-time period may happen to be nested into an
interactive weight raising period, i.e., it may happen that, when a
bfq_queue switches to a soft real-time weight-raised state, the
bfq_queue is already being weight-raised because deemed interactive
too. In this case, BFQ saves in a special variable
wr_start_at_switch_to_srt, the time instant when the interactive
weight-raising period started for the bfq_queue, i.e., the time
instant when BFQ started to deem the bfq_queue interactive. This value
is then used to check whether the interactive weight-raising period
would still be in progress when the soft real-time weight-raising
period ends. If so, interactive weight raising is restored for the
bfq_queue. This restore is useful, in particular, because it prevents
bfq_queues from losing their interactive weight raising prematurely,
as a consequence of spurious, short-lived soft real-time
weight-raising periods caused by wrong detections as soft real-time.
If, instead, a bfq_queue switches to soft-real-time weight raising
while it *is not* already in an interactive weight-raising period,
then the variable wr_start_at_switch_to_srt has no meaning during the
following soft real-time weight-raising period. Unfortunately the
handling of this case is wrong in BFQ: not only the variable is not
flagged somehow as meaningless, but it is also set to the time when
the switch to soft real-time weight-raising occurs. This may cause an
interactive weight-raising period to be considered mistakenly as still
in progress, and thus a spurious interactive weight-raising period to
start for the bfq_queue, at the end of the soft-real-time
weight-raising period. In particular the spurious interactive
weight-raising period will be considered as still in progress, if the
soft-real-time weight-raising period does not last very long. The
bfq_queue will then be wrongly privileged and, if I/O bound, will
unjustly steal bandwidth to truly interactive or soft real-time
bfq_queues, harming responsiveness and low latency.
This commit fixes this issue by just setting wr_start_at_switch_to_srt
to minus infinity (farthest past time instant according to jiffies
macros): when the soft-real-time weight-raising period ends, certainly
no interactive weight-raising period will be considered as still in
progress.
[1] Background I/O Type: Random - Background I/O mix: Reads and writes
- Application to start: LibreOffice Writer in
http://www.phoronix.com/scan.php?page=news_item&px=Linux-4.13-IO-Laptop
writeback: only allow one inflight and pending full flush
When someone calls wakeup_flusher_threads() or
wakeup_flusher_threads_bdi(), they schedule writeback of all dirty
pages in the system (or on that bdi). If we are tight on memory, we
can get tons of these queued from kswapd/vmscan. This causes (at
least) two problems:
1) We consume a ton of memory just allocating writeback work items.
We've seen as much as 600 million of these writeback work items
pending. That's a lot of memory to pointlessly hold hostage,
while the box is under memory pressure.
2) We spend so much time processing these work items, that we
introduce a softlockup in writeback processing. This is because
each of the writeback work items don't end up doing any work (it's
hard when you have millions of identical ones coming in to the
flush machinery), so we just sit in a tight loop pulling work
items and deleting/freeing them.
Fix this by adding a 'start_all' bit to the writeback structure, and
set that when someone attempts to flush all dirty pages. The bit is
cleared when we start writeback on that work item. If the bit is
already set when we attempt to queue !nr_pages writeback, then we
simply ignore it.
This provides us one full flush in flight, with one pending as well,
and makes for more efficient handling of this type of writeback.
Acked-by: Johannes Weiner <hannes@cmpxchg.org> Tested-by: Chris Mason <clm@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@kernel.dk>
writeback: move nr_pages == 0 logic to one location
Now that we have no external callers of wb_start_writeback(), we
can shuffle the passing in of 'nr_pages'. Everybody passes in 0
at this point, so just kill the argument and move the dirty
count retrieval to that function.
Acked-by: Johannes Weiner <hannes@cmpxchg.org> Tested-by: Chris Mason <clm@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
We don't have any callers outside of fs-writeback.c anymore,
make it private.
Acked-by: Johannes Weiner <hannes@cmpxchg.org> Tested-by: Chris Mason <clm@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
writeback: pass in '0' for nr_pages writeback in laptop mode
Laptop mode really wants to writeback the number of dirty
pages and inodes. Instead of calculating this in the caller,
just pass in 0 and let wakeup_flusher_threads() handle it.
Use the new wakeup_flusher_threads_bdi() instead of rolling
our own.
Acked-by: Johannes Weiner <hannes@cmpxchg.org> Tested-by: Chris Mason <clm@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Similar to wakeup_flusher_threads(), except that we only wake
up the flusher threads on the specified backing device.
No functional changes in this patch.
Acked-by: Johannes Weiner <hannes@cmpxchg.org> Tested-by: Chris Mason <clm@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
buffer: eliminate the need to call free_more_memory() in __getblk_slow()
Since the previous commit removed any case where grow_buffers()
would return failure due to memory allocations, we can safely
remove the case where we have to call free_more_memory() in
this function.
Since this is also the last user of free_more_memory(), kill
it off completely.
Reviewed-by: Nikolay Borisov <nborisov@suse.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@kernel.dk>
buffer: grow_dev_page() should use __GFP_NOFAIL for all cases
We currently use it for find_or_create_page(), which means that it
cannot fail. Ensure we also pass in 'retry == true' to
alloc_page_buffers(), which also ensure that it cannot fail.
After this, there are no failure cases in grow_dev_page() that
occur because of a failed memory allocation.
Reviewed-by: Nikolay Borisov <nborisov@suse.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@kernel.dk>
buffer: have alloc_page_buffers() use __GFP_NOFAIL
Instead of adding weird retry logic in that function, utilize
__GFP_NOFAIL to ensure that the vm takes care of handling any
potential retries appropriately. This means we don't have to
call free_more_memory() from here.
Reviewed-by: Nikolay Borisov <nborisov@suse.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk-mq: wire up completion notifier for laptop mode
For some reason, the laptop mode IO completion notified was never wired
up for blk-mq. Ensure that we trigger the callback appropriately, to arm
the laptop mode flush timer.
Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>