When pblk receives a sync, all data up to that point in the write buffer
must be comitted to persistent storage, and as flash memory comes with a
minimal write size there is a significant cost involved both in terms
of time for completing the sync and in terms of write amplification
padded sectors for filling up to the minimal write size.
In order to get a better understanding of the costs involved for syncs,
Add a sysfs attribute to pblk: padded_dist, showing a normalized
distribution of sectors padded. In order to facilitate measurements of
specific workloads during the lifetime of the pblk instance, the
distribution can be reset by writing 0 to the attribute.
Do this by introducing counters for each possible padding:
{0..(minimal write size - 1)} and calculate the normalized distribution
when showing the attribute.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Javier González <javier@cnexlabs.com>
Rearranged total_buckets statement in pblk_sysfs_get_padding_dist
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In a SSD, write amplification, WA, is defined as the average
number of page writes per user page write. Write amplification
negatively affects write performance and decreases the lifetime
of the disk, so it's a useful metric to add to sysfs.
In plkb's case, the number of writes per user sector is the sum of:
(1) number of user writes
(2) number of sectors written by the garbage collector
(3) number of sectors padded (i.e. due to syncs)
This patch adds persistent counters for 1-3 and two sysfs attributes
to export these along with WA calculated with five decimals:
write_amp_mileage: the accumulated write amplification stats
for the lifetime of the pblk instance
write_amp_trip: resetable stats to facilitate delta measurements,
values reset at creation and if 0 is written
to the attribute.
64-bit counters are used as a 32 bit counter would wrap around
already after about 17 TB worth of user data. It will take a
long long time before the 64 bit sector counters wrap around.
The counters are stored after the bad block bitmap in the first
emeta sector of each written line. There is plenty of space in the
first emeta sector, so we don't need to bump the major version of
the line data format.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Unless we protect flush pointer updates with a lock, we risk
resetting new flush points before we've synced all sectors
up to that point.
This patch protects new flush points with the same spin lock
that is being held when advancing the sync pointer and
resetting completed flush points.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move completion of syncs and clearing of flush points to the
write completion path - this ensures that the data has been
comitted to the media before completing bios containing syncs.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Sync point is a really confusing name for keeping track of
the last entry that needs to be flushed so change the name
to to flush_point instead.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Through time, we have generated some redundant helper functions.
Refactor them to eliminate redundant and unnecessary code. Also, reorder
them to improve readability
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We already pass the structure pointer so no need to pass the member.
Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Wait until we know the exact number of ppas to be sent to the device,
before allocating the bio.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a lba either hits the cache or corresponds to an empty entry in the
L2P table, we need to advance the bio according to the position in which
the lba is located. Otherwise, we will copy data in the wrong page, thus
causing data corruption for the application.
In case of a cache hit, we assumed that bio->bi_iter.bi_idx would
contain the correct index, but this is no necessarily true. Instead, use
the local bio advance counter and iterator. This guarantees that lbas
hitting the cache are copied into the right bv_page.
In case of an empty L2P entry, we omitted to advance the bio. In the
cases when the same I/O also contains a cache hit, data corresponding
to this lba will be copied to the wrong bv_page. Fix this by advancing
the bio as we do in the case of a cache hit.
Fixes: a4bd217b43 lightnvm: physical block device (pblk) target
Signed-off-by: Javier González <javier@javigon.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a read is directed to the cache, we risk that the lba has been
updated during the time we made the L2P table lookup and the time we are
actually reading form the cache. We intentionally not hold the L2P lock
not to block other threads.
While strict ordering is not a guarantee at this level (unless REQ_FLUSH
has been previously issued), we have experience that some databases that
have recently implemented direct I/O support, issue metadata reads very
close to the writes, without issuing a fsync in the middle. An easy way
to support them while they is to make an extra effort and check the L2P
map right before reading the cache.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When removing a pblk instance, pad the current line using asynchronous
I/O. This reduces the removal time from ~1 minute in the worst case to a
couple of seconds.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When user threads place data into the write buffer, they reserve space
and do the memory copy out of the lock. As a consequence, when the write
thread starts persisting data, there is a chance that it is not copied
yet. In this case, avoid polling, and schedule before retrying.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Due to user writes being decoupled from media writes because of the need
of an intermediate write buffer, irrecoverable media write errors lead
to pblk stalling; user writes fill up the buffer and end up in an
infinite retry loop.
In order to let user writes fail gracefully, it is necessary for pblk to
keep track of its own internal state and prevent further writes from
being placed into the write buffer.
This patch implements a state machine to keep track of internal errors
and, in case of failure, fail further user writes in an standard way.
Depending on the type of error, pblk will do its best to persist
buffered writes (which are already acknowledged) and close down on a
graceful manner. This way, data might be recovered by re-instantiating
pblk. Such state machine paves out the way for a state-based FTL log.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
At the moment, in order to get enough read parallelism, we have recycled
several lines at the same time. This approach has proven not to work
well when reaching capacity, since we end up mixing valid data from all
lines, thus not maintaining a sustainable free/recycled line ratio.
The new design, relies on a two level workqueue mechanism. In the first
level, we read the metadata for a number of lines based on the GC list
they reside on (this is governed by the number of valid sectors in each
line). In the second level, we recycle a single line at a time. Here, we
issue reads in parallel, while a single GC write thread places data in
the write buffer. This design allows to (i) only move data from one line
at a time, thus maintaining a sane free/recycled ration and (ii)
maintain the GC writer busy with recycled data.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Erase I/Os are scheduled with the following goals in mind: (i) minimize
LUNs collisions with write I/Os, and (ii) even out the price of erasing
on every write, instead of putting all the burden on when garbage
collection runs. This works well on the current design, but is specific
to the default mapping algorithm.
This patch generalizes the erase path so that other mapping algorithms
can select an arbitrary line to be erased instead. It also gets rid of
the erase semaphore since it creates jittering for user writes.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch introduces pblk, a host-side translation layer for
Open-Channel SSDs to expose them like block devices. The translation
layer allows data placement decisions, and I/O scheduling to be
managed by the host, enabling users to optimize the SSD for their
specific workloads.
An open-channel SSD has a set of LUNs (parallel units) and a
collection of blocks. Each block can be read in any order, but
writes must be sequential. Writes may also fail, and if a block
requires it, must also be reset before new writes can be
applied.
To manage the constraints, pblk maintains a logical to
physical address (L2P) table, write cache, garbage
collection logic, recovery scheme, and logic to rate-limit
user I/Os versus garbage collection I/Os.
The L2P table is fully-associative and manages sectors at a
4KB granularity. Pblk stores the L2P table in two places, in
the out-of-band area of the media and on the last page of a
line. In the cause of a power failure, pblk will perform a
scan to recover the L2P table.
The user data is organized into lines. A line is data
striped across blocks and LUNs. The lines enable the host to
reduce the amount of metadata to maintain besides the user
data and makes it easier to implement RAID or erasure coding
in the future.
pblk implements multi-tenant support and can be instantiated
multiple times on the same drive. Each instance owns a
portion of the SSD - both regarding I/O bandwidth and
capacity - providing I/O isolation for each case.
Finally, pblk also exposes a sysfs interface that allows
user-space to peek into the internals of pblk. The interface
is available at /dev/block/*/pblk/ where * is the block
device name exposed.
This work also contains contributions from:
Matias Bjørling <matias@cnexlabs.com>
Simon A. F. Lund <slund@cnexlabs.com>
Young Tack Jin <youngtack.jin@gmail.com>
Huaicheng Li <huaicheng@cs.uchicago.edu>
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Hans Holmberg