- 10 Nov, 2014 28 commits
-
-
Joe Thornber authored
Safely allow the discard blocksize to be larger than the cache blocksize by using the bio prison's range locking support. This also improves discard performance considerly because larger discards are issued to the dm-cache device. The discard blocksize was always intended to be greater than the cache blocksize. But until now it wasn't implemented safely. Also, by safely restoring the ability to have discard blocksize larger than cache blocksize we're able to significantly reduce the memory used for the cache's discard bitset. Before, with a small discard blocksize, the discard bitset could get quite large because its size is a function of the discard blocksize and the origin device's size. For example, previously, using a 32KB cache blocksize with a 40TB origin resulted in 1280MB of incore memory use for the discard bitset! Now, the discard blocksize is scaled up accordingly to ensure the discard bitset is capped at 2**14 bits, or 16KB. Signed-off-by:
Joe Thornber <ejt@redhat.com> Signed-off-by:
Mike Snitzer <snitzer@redhat.com>
-
Joe Thornber authored
This reverts commit d132cc6d because we actually do want to allow the discard blocksize to be larger than the cache blocksize. Further dm-cache discard changes will make this possible. Signed-off-by:
-
Joe Thornber authored
This reverts commit 64ab346a because we actually do want to allow the discard blocksize to be larger than the cache blocksize. Further dm-cache discard changes will make this possible. Signed-off-by:
-
Joe Thornber authored
Ranges will be placed in the same cell if they overlap. Range locking is a prerequisite for more efficient multi-block discard support in both the cache and thin-provisioning targets. Signed-off-by:
-
Mike Snitzer authored
Before, if the user wanted sequential IO to be promoted to the cache they'd have to set sequential_threshold to some nebulous large value. Now, the user may easily disable sequential IO detection (and sequential IO's implicit bypass of the cache) by setting sequential_threshold to 0. Signed-off-by: -
Joe Thornber authored
Rather than maintaining a separate promote_threshold variable that we periodically update we now use the hit count of the oldest clean block. Also add a fudge factor to discourage demoting dirty blocks. With some tests this has a sizeable difference, because the old code was too eager to demote blocks. For example, device-mapper-test-suite's git_extract_cache_quick test goes from taking 190 seconds, to 142 (linear on spindle takes 250). Signed-off-by:
-
Hannes Reinecke authored
When creating new devices dm_sync_table() calls synchronize_rcu_expedited(), causing _all_ pending RCU pointers to be flushed. This causes a latency overhead that is especially noticeable when creating lots of devices. And all of this is pointless as there are no old maps to be disconnected, and hence no stale pointers which would need to be cleared up. Signed-off-by:
Hannes Reinecke <hare@suse.de> Reviewed-by:
Mikulas Patocka <mpatocka@redhat.com> Signed-off-by:
-
Pranith Kumar authored
Annotate the map field with __rcu since this is a rcu pointer which is checked by sparse. Signed-off-by:
Pranith Kumar <bobby.prani@gmail.com> Signed-off-by:
Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by:
-
Pranith Kumar authored
The map field in 'struct mapped_device' is an rcu pointer. Use rcu_dereference() while accessing it. Signed-off-by:
-
Mike Snitzer authored
Also refactor some other bio_list erroring helpers. Signed-off-by: -
Mike Snitzer authored
Eliminate redundant should_error_unserviceable_bio check and error loop. Signed-off-by: -
Joe Thornber authored
Sort the cells in logical block order before processing each cell in process_thin_deferred_cells(). This significantly improves the ondisk layout on rotational storage, whereby improving read performance. Signed-off-by:
-
Joe Thornber authored
This use of direct submission in process_shared_bio() reduces latency for submitting bios in the shared cell by avoiding adding those bios to the deferred list and waiting for the next iteration of the worker. Signed-off-by:
-
Joe Thornber authored
This use of direct submission in process_prepared_mapping() reduces latency for submitting bios in a cell by avoiding adding those bios to the deferred list and waiting for the next iteration of the worker. But this direct submission exposes the potential for a race between releasing a cell and incrementing deferred set. Fix this by introducing dm_cell_visit_release() and refactoring inc_remap_and_issue_cell() accordingly. Signed-off-by:
-
Joe Thornber authored
This avoids dropping the cell, so increases the probability that other bios will collect within the cell, rather than being passed individually to the worker. Also add required process_cell and process_discard_cell error handling wrappers and set associated pool-mode function pointers accordingly. Signed-off-by:
-
Mike Snitzer authored
Purely cleanup of duplicated code, no functional change. Signed-off-by: -
Joe Thornber authored
When processing a discard bio, if the block is already quiesced do the discard immediately rather than adding the mapping to a list for the next iteration of the worker thread. Discarding a fully provisioned 100G thin volume with 64k block size goes from 860s to 95s with this change. Clearly there's something wrong with the worker architecture, more investigation needed. Signed-off-by:
-
Mike Snitzer authored
Introduce thin_merge so that any additional constraints from the data volume may be taken into account when determing the maximum number of sectors that can be issued relative to the specified logical offset. This is particularly important if/when the data volume is layered ontop of a more sophisticated device (e.g. dm-raid or some other DM target). Reviewed-by:
Heinz Mauelshagen <heinzm@redhat.com> Signed-off-by:
-
Mike Snitzer authored
These code changes do not introduce a functional change. But bio_add_page() will never attempt to build up a bio larger than queue_max_sectors(). Similarly, bio_get_nr_vecs() is also bound by queue_max_sectors(). Therefore, there is no point in allowing dm_merge_bvec() to answer "how many sectors can a bio have at this offset?" with anything larger than queue_max_sectors(). Using queue_max_sectors() rather than BIO_MAX_SECTORS serves to more accurately convey the limits that are being imposed. Also, use unlikely() to clarify the fact that the defensive code in dm_merge_bvec() relative to max_size going negative shouldn't ever happen -- if it does happen there is a bug in the block layer for requesting larger than dm_merge_bvec()'s initial response for a given offset. Also, update a comment in dm_merge_bvec() relative to max_hw_sectors_kb. And fix empty newline whitespace. Signed-off-by: -
Mike Snitzer authored
Allows for filesystems to submit bios that are a factor of the thinp blocksize, improving dm-thinp efficiency (particularly when the data volume is RAID). Also set io_min to max_sectors_kb if it is a factor of the thinp blocksize. Signed-off-by: -
Joe Thornber authored
Throttle IO based on the time it's taking the worker to do one loop. There were reports of hung task timeouts occuring and it was observed that the excessively long avgqu-sz (as reported by iostat) was contributing to these hung tasks. Throttling definitely helps dm-thinp perform better under heavy IO load (without being detremental by being overzealous). It reduces avgqu-sz drastically, e.g.: from 60K to ~6K, and even as low as 150 once metadata is cached by bufio, when dirty_ratio=5, dirty_background_ratio=2. And avgqu-sz stays at or below 30K even with dirty_ratio=20, dirty_background_ratio=10. Signed-off-by:
-
Joe Thornber authored
Prefetch metadata at the start of the worker thread and then again every 128th bio processed from the deferred list. Signed-off-by:
-
Joe Thornber authored
Introduce the dm_tm_issue_prefetches interface. If you're using a non-blocking clone the tm will build up a list of requested blocks that weren't in core. dm_tm_issue_prefetches will request those blocks to be prefetched. Signed-off-by:
-
Joe Thornber authored
This change is a prerequisite for allowing metadata to be prefetched. Signed-off-by:
-
Joe Thornber authored
Previously it was using a fixed sized hash table. There are times when very many concurrent cells are held (such as when processing a very large discard). When this happens the hash table performance becomes very poor. Signed-off-by:
-
Joe Thornber authored
These changes help keep metadata backed by dm-bufio in-core longer which fixes reports of metadata churn in the face of heavy random IO workloads. Before, bufio evicted all buffers older than DM_BUFIO_DEFAULT_AGE_SECS. Having a device (e.g. dm-thinp or dm-cache) lose all metadata just because associated buffers had been idle for some time is unfriendly. Now, the user may now configure the number of bytes that bufio retains using the 'retain_bytes' module parameter. The default is 256K. Also, the DM_BUFIO_WORK_TIMER_SECS and DM_BUFIO_DEFAULT_AGE_SECS defaults were quite low so increase them (to 30 and 300 respectively). Signed-off-by:
-
Joe Thornber authored
Converting over to using an rbtree eliminates a fixed 8MB allocation from vmalloc space for the hash table. Signed-off-by:
-
Joe Thornber authored
The walk code was using a 'ro_spine' to hold it's locked btree nodes. But this data structure is designed for the rolling lock scheme, and as such automatically unlocks blocks that are two steps up the call chain. This is not suitable for the simple recursive walk algorithm, which retraces its steps. This code is only used by the persistent array code, which in turn is only used by dm-cache. In order to trigger it you need to have a mapping tree that is more than 2 levels deep; which equates to 8-16 million cache blocks. For instance a 4T ssd with a very small block size of 32k only just triggers this bug. The fix just places the locked blocks on the stack, and stops using the ro_spine altogether. Signed-off-by:
-
- 04 Nov, 2014 1 commit
-
-
Joe Thornber authored
Avoids normal IO racing with discard. Signed-off-by:
-
- 29 Oct, 2014 1 commit
-
-
Heinz Mauelshagen authored
Commit 48cf06bc ("dm raid: add discard support for RAID levels 4, 5 and 6") did not properly handle missing metadata device(s). A failing read of the superblock causes the metadata and data devices to be removed from the dev array in struct raid_set, setting references to both devices to NULL. configure_discard_support() nonetheless tries to access the data dev unconditionally causing an oops. Signed-off-by:
-
- 21 Oct, 2014 1 commit
-
-
Heinz Mauelshagen authored
The dm-raid superblock (struct dm_raid_superblock) is padded to 512 bytes and that size is being used to read it in from the metadata device into one preallocated page. Reading or writing this on a 512-byte sector device works fine but on a 4096-byte sector device this fails. Set the dm-raid superblock's size to the logical block size of the metadata device, because IO at that size is guaranteed too work. Also add a size check to avoid silent partial metadata loss in case the superblock should ever grow past the logical block size or PAGE_SIZE. [includes pointer math fix from Dan Carpenter] Reported-by:
"Liuhua Wang" <lwang@suse.com> Signed-off-by:
Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by:
-
- 17 Oct, 2014 1 commit
-
-
Mikulas Patocka authored
The shrinker uses gfp flags to indicate what kind of operation can the driver wait for. If __GFP_IO flag is present, the driver can wait for block I/O operations, if __GFP_FS flag is present, the driver can wait on operations involving the filesystem. dm-bufio tested for __GFP_IO. However, dm-bufio can run on a loop block device that makes calls into the filesystem. If __GFP_IO is present and __GFP_FS isn't, dm-bufio could still block on filesystem operations if it runs on a loop block device. The change from __GFP_IO to __GFP_FS supposedly fixes one observed (though unreproducible) deadlock involving dm-bufio and loop device. Signed-off-by:
-
- 11 Oct, 2014 1 commit
-
-
Pavitra Kumar authored
Fix a potential struct stripe_c leak that would occur if the chunk_size exceeded the maximum allowed by dm_set_target_max_io_len (UINT_MAX). However, in practice there is no possibility of this occuring given that chunk_size is of type uint32_t. But it is good to fix this to future-proof in case dm_set_target_max_io_len's implementation were to change. Signed-off-by:
Pavitra Kumar <pavitrak@nvidia.com> Signed-off-by:
-
- 06 Oct, 2014 7 commits
-
-
Alexey Khoroshilov authored
If cn_add_callback() fails in dm_ulog_tfr_init(), it does not deallocate prealloced memory but calls cn_del_callback(). Found by Linux Driver Verification project (linuxtesting.org). Signed-off-by:
Alexey Khoroshilov <khoroshilov@ispras.ru> Reviewed-by:
Jonathan Brassow <jbrassow@redhat.com> Signed-off-by:
-
Mikulas Patocka authored
When __scan frees the required number of buffer entries that the shrinker requested (nr_to_scan becomes zero) it must return. Before this fix the __scan code exited only the inner loop and continued in the outer loop -- which could result in reduced performance due to extra buffers being freed (e.g. unnecessarily evicted thinp metadata needing to be synchronously re-read into bufio's cache). Also, move dm_bufio_cond_resched to __scan's inner loop, so that iterating the bufio client's lru lists doesn't result in scheduling latency. Reported-by:
Joe Thornber <thornber@redhat.com> Signed-off-by:
-
Joe Thornber authored
The 'last_accessed' member of the dm_buffer structure was only set when the the buffer was created. This led to each buffer being discarded after dm_bufio_max_age time even if it was used recently. In practice this resulted in all thinp metadata being evicted soon after being read -- this is particularly problematic for metadata intensive workloads like multithreaded small random IO. 'last_accessed' is now updated each time the buffer is moved to the head of the LRU list, so the buffer is now properly discarded if it was not used in dm_bufio_max_age time. Signed-off-by:
-
Heinz Mauelshagen authored
In case of RAID levels 4, 5 and 6 we have to verify each RAID members' ability to zero data on discards to avoid stripe data corruption -- if discard_zeroes_data is not set for each RAID member discard support must be disabled. But given the uncertainty of whether or not a RAID member properly supports zeroing data on discard we require the user to explicitly allow discard support on RAID levels 4, 5, and 6 by setting a dm-raid module paramter, e.g.: dm-raid.devices_handle_discard_safely=Y Otherwise, discards could cause data corruption on RAID4/5/6. Signed-off-by:
-
Heinz Mauelshagen authored
Discard support is not enabled for RAID levels 4, 5, and 6 at this time due to concerns about unreliable discard_zeroes_data support on some hardware. Otherwise, discards could cause stripe data corruption (classic example of bad apples spoiling the bunch). Signed-off-by:
-
Benjamin Marzinski authored
Until this change, when loading a new DM table, DM core would re-open all of the devices in the DM table. Now, DM core will avoid redundant device opens (and closes when destroying the old table) if the old table already has a device open using the same mode. This is achieved by managing reference counts on the table_devices that DM core now stores in the mapped_device structure (rather than in the dm_table structure). So a mapped_device's active and inactive dm_tables' dm_dev lists now just point to the dm_devs stored in the mapped_device's table_devices list. This improvement in DM core's device reference counting has the side-effect of fixing a long-standing limitation of the multipath target: a DM multipath table couldn't include any paths that were unusable (failed). For example: if all paths have failed and you add a new, working, path to the table; you can't use it since the table load would fail due to it still containing failed paths. Now a re-load of a multipath table can include failed devices and when those devices become active again they can be used instantly. The device list code in dm.c isn't a straight copy/paste from the code in dm-table.c, but it's very close (aside from some variable renames). One subtle difference is that find_table_device for the tables_devices list will only match devices with the same name and mode. This is because we don't want to upgrade a device's mode in the active table when an inactive table is loaded. Access to the mapped_device structure's tables_devices list requires a mutex (tables_devices_lock), so that tables cannot be created and destroyed concurrently. Signed-off-by:
Benjamin Marzinski <bmarzins@redhat.com> Signed-off-by:
-
Benjamin Marzinski authored
'queue_io' is set so that IO is queued while paths are being initialized. Clear queue_io in __choose_pgpath if there are no valid paths, since there are obviously no paths that can be initialized. Otherwise IOs to the device will back up. Signed-off-by:
-
