- 09 Oct, 2018 35 commits
-
-
Javier González authored
In the OCSSD 2.0 spec, each chunk reports its write pointer. This means that pblk does not need to scan open lines to find the write pointer, but instead, it can retrieve it directly (and verify it). This patch uses the write pointer on open lines to (i) recover the line up until the last written lba and (ii) reconstruct the map bitmap and rest of line metadata so that the line can be used for new data. Since the 1.2 path in lightnvm core has been re-implemented to populate the chunk structure and thus recover the write pointer on initialization, this patch removes 1.2 specific recovery, as the 2.0 path can be reused. Signed-off-by:
Javier González <javier@cnexlabs.com> Signed-off-by:
Matias Bjørling <mb@lightnvm.io> Signed-off-by:
Jens Axboe <axboe@kernel.dk>
-
Javier González authored
pblk guarantees write ordering at a chunk level through a per open chunk semaphore. At this point, since we only have an open I/O stream for both user and GC data, the semaphore is per parallel unit. For the metadata I/O that is synchronous, the semaphore is not needed as ordering is guaranteed. However, if the metadata scheme changes or multiple streams are open, this guarantee might not be preserved. This patch makes sure that all writes go through the semaphore, even for synchronous I/O. This is consistent with pblk's write I/O model. It also simplifies maintenance since changes in the metadata scheme could cause ordering issues. Signed-off-by:
-
Javier González authored
pblk maintains two different metadata paths for smeta and emeta, which store metadata at the start of the line and at the end of the line, respectively. Until now, these path has been common for writing and retrieving metadata, however, as these paths diverge, the common code becomes less clear and unnecessary complicated. In preparation for further changes to the metadata write path, this patch separates the write and read paths for smeta and emeta and removes the synchronous emeta path as it not used anymore (emeta is scheduled asynchronously to prevent jittering due to internal I/Os). Signed-off-by:
-
Javier González authored
dma allocations for ppa_list and meta_list in rqd are replicated in several places across the pblk codebase. Make helpers to encapsulate creation and deletion to simplify the code. Signed-off-by:
-
Javier González authored
The lightnvm subsystem provides helpers to retrieve chunk metadata, where the target needs to provide a buffer to store the metadata. An implicit assumption is that this buffer is contiguous and can be used to retrieve the data from the device. If the device exposes too many chunks, then kmalloc might fail, thus failing instance creation. This patch removes this assumption by implementing an internal buffer in the lightnvm subsystem to retrieve chunk metadata. Targets can then use virtual memory allocations. Since this is a target API change, adapt pblk accordingly. Signed-off-by:
Hans Holmberg <hans.holmberg@cnexlabs.com> Signed-off-by:
-
Jia-Ju Bai authored
The driver may sleep with holding a spinlock. The function call paths (from bottom to top) in Linux-4.16 are: [FUNC] nvm_dev_dma_alloc(GFP_KERNEL) drivers/lightnvm/pblk-core.c, 754: nvm_dev_dma_alloc in pblk_line_submit_smeta_io drivers/lightnvm/pblk-core.c, 1048: pblk_line_submit_smeta_io in pblk_line_init_bb drivers/lightnvm/pblk-core.c, 1434: pblk_line_init_bb in pblk_line_replace_data drivers/lightnvm/pblk-recovery.c, 980: pblk_line_replace_data in pblk_recov_l2p drivers/lightnvm/pblk-recovery.c, 976: spin_lock in pblk_recov_l2p [FUNC] bio_map_kern(GFP_KERNEL) drivers/lightnvm/pblk-core.c, 762: bio_map_kern in pblk_line_submit_smeta_io drivers/lightnvm/pblk-core.c, 1048: pblk_line_submit_smeta_io in pblk_line_init_bb drivers/lightnvm/pblk-core.c, 1434: pblk_line_init_bb in pblk_line_replace_data drivers/lightnvm/pblk-recovery.c, 980: pblk_line_replace_data in pblk_recov_l2p drivers/lightnvm/pblk-recovery.c, 976: spin_lock in pblk_recov_l2p To fix these bugs, the call to pblk_line_replace_data() is moved out of the spinlock protection. These bugs are found by my static analysis tool DSAC. Signed-off-by:
Jia-Ju Bai <baijiaju1990@gmail.com> Reviewed-by:
-
Hans Holmberg authored
On 1.2-devices, the mapping-out of remaning sectors in the failed-write's block can result in an infinite loop, stalling the write pipeline, fix this. Fixes: 6a3abf5b ("lightnvm: pblk: rework write error recovery path") Signed-off-by:
-
Hans Holmberg authored
Pblk should not create a set of global caches every time a pblk instance is created. The global caches should be made available only when there is one or more pblk instances. This patch bundles the global caches together with a kref keeping track of whether the caches should be available or not. Also, turn the global pblk lock into a mutex that explicitly protects the caches (as this was the only purpose of the lock). Signed-off-by:
-
Javier González authored
If a line is padded, calculate the pad distance directly on the helper being used for this purpose. Signed-off-by:
-
Javier González authored
Continuing the effort of moving 1.2 and 2.0 specific code to core, move 64_to_32 and 32_to_64 ppa helpers from pblk to core. Signed-off-by:
-
Hans Holmberg authored
Trace state of chunk resets. Signed-off-by:
-
Hans Holmberg authored
Add trace events for tracking pblk state changes. Signed-off-by:
-
Hans Holmberg authored
Add trace events for logging for line state changes. Signed-off-by:
-
Hans Holmberg authored
Introduce trace points for tracking chunk states in pblk - this is useful for inspection of the entire state of the drive, and real handy for both fw and pblk debugging. Signed-off-by:
-
Matias Bjørling authored
Remove the debug only iteration within __pblk_down_page, which then allows us to reduce the number of arguments down to pblk and the parallel unit from the functions that calls it. Simplifying the callers logic considerably. Also, rename the functions pblk_[down/up]_page to pblk_[down/up]_chunk, to communicate that it manages the write pointer of the chunk. Note that it also protects the parallel unit such that at most one chunk is active per parallel unit. Signed-off-by:
-
Hans Holmberg authored
When the user data counter exceeds 32 bits, the write amplification calculation does not provide the right value. Fix this by using div64_u64 in stead of div64. Fixes: 76758390 ("lightnvm: pblk: export write amplification counters to sysfs") Signed-off-by:
-
Hans Holmberg authored
The prefix when printing ppas in pblk_read_check_rand should be "rnd" not "seq", so fix this so we can differentiate between lba missmatches in random and sequential reads. Also change the print order so we align with pblk_read_check_seq, printing read lba first. Signed-off-by:
-
Hans Holmberg authored
The parameters nr_ppas and ppa_list are not used, so remove them. Signed-off-by:
-
Hans Holmberg authored
Line map bitmap allocations are fairly large and can fail. Allocation failures are fatal to pblk, stopping the write pipeline. To avoid this, allocate the bitmaps using a mempool instead. Mempool allocations never fail if called from a process context, and pblk *should* only allocate map bitmaps in process context, but keep the failure handling for robustness sake. Signed-off-by:
-
Hans Holmberg authored
There is a number of places in the lightnvm subsystem where the user iterates over the ppa list. Before iterating, the user must know if it is a single or multiple LBAs due to vector commands using either the nvm_rq ->ppa_addr or ->ppa_list fields on command submission, which leads to open-coding the if/else statement. Instead of having multiple if/else's, move it into a function that can be called by its users. A nice side effect of this cleanup is that this patch fixes up a bunch of cases where we don't consider the single-ppa case in pblk. Signed-off-by:
-
Javier González authored
If a line is recovered from open chunks, the memory structures for emeta have not necessarily been properly set on line initialization. When closing a line, make sure that emeta is consistent so that the line can be recovered on the fast path on next reboot. Also, remove a couple of empty lines at the end of the function. Signed-off-by:
-
Javier González authored
Removed unused struct ppa_addr variable. Signed-off-by:
-
Javier González authored
Fix comment typo Decrese -> Decrease Signed-off-by:
-
Javier González authored
The current helper to obtain a line from a ppa returns the line id, which requires its users to explicitly retrieve the pointer to the line with the id. Make 2 different helpers: one returning the line id and one returning the line directly. Signed-off-by:
-
Javier González authored
Implement helpers to go from ppas to a chunk within a line and an address within a chunk. These helpers will be used on the patches adding trace support in pblk, which will be sent in this window. Signed-off-by:
-
Matias Bjørling authored
The read completion path uses the put_line variable to decide whether the reference on a line should be released. The function name used for that is pblk_read_put_rqd_kref, which could lead one to believe that it is the rqd that is releasing the reference, while it is the line reference that is put. Rename and also split the function in two to account for either rqd or single ppa callers and move it to core, such that it later can be used in the write path as well. Signed-off-by:
Heiner Litz <hlitz@ucsc.edu> Signed-off-by:
-
Matias Bjørling authored
The I/O size and capacity checks are already done by the block layer. Signed-off-by:
-
Matias Bjørling authored
The calculation of pblk->min_write_pgs should only use the optimal write size attribute provided by the drive, it does not correlate to the memory page size of the system, which can be smaller or larger than the LBA size reported. Signed-off-by:
-
Matias Bjørling authored
Both NVM_MAX_VLBA and PBLK_MAX_REQ_ADDRS define how many LBAs that are available in a vector command. pblk uses them interchangeably in its implementation. Use NVM_MAX_VLBA as the main one and remove usages of PBLK_MAX_REQ_ADDRS. Also remove the power representation that only has one user, and instead calculate it at runtime. Signed-off-by:
-
Matias Bjørling authored
pblk implements two data paths for recovery line state. One for 1.2 and another for 2.0, instead of having pblk implement these, combine them in the core to reduce complexity and make available to other targets. The new interface will adhere to the 2.0 chunk definition, including managing open chunks with an active write pointer. To provide this interface, a 1.2 device recovers the state of the chunks by manually detecting if a chunk is either free/open/close/offline, and if open, scanning the flash pages sequentially to find the next writeable page. This process takes on average ~10 seconds on a device with 64 dies, 1024 blocks and 60us read access time. The process can be parallelized but is left out for maintenance simplicity, as the 1.2 specification is deprecated. For 2.0 devices, the logic is maintained internally in the drive and retrieved through the 2.0 interface. Signed-off-by:
-
Javier González authored
In pblk, when a new line is allocated, metadata for the previously written line is scheduled. This is done through a fixed memory region that is shared through time and contexts across different lines and therefore protected by a lock. Unfortunately, this lock is not properly covering all the metadata used for sharing this memory regions, resulting in a race condition. This patch fixes this race condition by protecting this metadata properly. Fixes: dd2a4343 ("lightnvm: pblk: sched. metadata on write thread") Signed-off-by:
-
Matias Bjørling authored
A 1.2 device is able to manage the logical to physical mapping table internally or leave it to the host. A target only supports one of those approaches, and therefore must check on initialization. Move this check to core to avoid each target implement the check. Signed-off-by:
-
Matias Bjørling authored
rqd.error is masked by the return value of pblk_submit_io_sync. The rqd structure is then passed on to the end_io function, which assumes that any error should lead to a chunk being marked offline/bad. Since the pblk_submit_io_sync can fail before the command is issued to the device, the error value maybe not correspond to a media failure, leading to chunks being immaturely retired. Also, the pblk_blk_erase_sync function prints an error message in case the erase fails. Since the caller prints an error message by itself, remove the error message in this function. Signed-off-by:
-
Matias Bjørling authored
Add nvm_set_flags helper to enable core to appropriately set the command flags for read/write/erase depending on which version a drive supports. The flags arguments can be distilled into the access hint, scrambling, and program/erase suspend. Replace the access hint with a "is_seq" parameter. The rest of the flags are dependent on the command opcode, which is trivial to detect and set. Signed-off-by:
-
Matias Bjørling authored
No need to force NVMe device driver to be compiled in if the lightnvm subsystem is selected. Also no need for PCI to be selected as well, as it would be selected by the device driver that hooks into the subsystem. Signed-off-by:
-
- 08 Oct, 2018 5 commits
-
-
Ming Lei authored
Lot of controllers may have only one irq vector for completing IO request. And usually affinity of the only irq vector is all possible CPUs, however, on most of ARCH, there may be only one specific CPU for handling this interrupt. So if all IOs are completed in hardirq context, it is inevitable to degrade IO performance because of increased irq latency. This patch tries to address this issue by allowing to complete request in softirq context, like the legacy IO path. IOPS is observed as ~13%+ in the following randread test on raid0 over virtio-scsi. mdadm --create --verbose /dev/md0 --level=0 --chunk=1024 --raid-devices=8 /dev/sdb /dev/sdc /dev/sdd /dev/sde /dev/sdf /dev/sdg /dev/sdh /dev/sdi fio --time_based --name=benchmark --runtime=30 --filename=/dev/md0 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=32 --rw=randread --blocksize=4k Cc: Dongli Zhang <dongli.zhang@oracle.com> Cc: Zach Marano <zmarano@google.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <bvanassche@acm.org> Cc: Jianchao Wang <jianchao.w.wang@oracle.com> Signed-off-by:
Ming Lei <ming.lei@redhat.com> Signed-off-by:
-
Dongbo Cao authored
when the nbuckets of cache device is smaller than 1024, making cache device will trigger BUG_ON in kernel, add a condition to avoid this. Reported-by:
nitroxis <n@nxs.re> Signed-off-by:
Dongbo Cao <cdbdyx@163.com> Signed-off-by:
Coly Li <colyli@suse.de> Signed-off-by:
-
Dongbo Cao authored
Split the combined '||' statements in if() check, to make the code easier for debug. Signed-off-by:
-
Shenghui Wang authored
Current cache_set has MAX_CACHES_PER_SET caches most, and the macro is used for " struct cache *cache_by_alloc[MAX_CACHES_PER_SET]; " in the define of struct cache_set. Use MAX_CACHES_PER_SET instead of magic number 8 in __bch_bucket_alloc_set. Signed-off-by:
Shenghui Wang <shhuiw@foxmail.com> Signed-off-by:
-
Coly Li authored
In extents.c:bch_extent_bad(), number 96 is used as parameter to call btree_bug_on(). The purpose is to check whether stale gen value exceeds BUCKET_GC_GEN_MAX, so it is better to use macro BUCKET_GC_GEN_MAX to make the code more understandable. Signed-off-by:
-
