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.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Cleanup up unused and static functions across the whole codebase.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Shut down the GC workqueues and tasks in the right order.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Start GC if needed, directly after init, as we might
need to garbage collect in order to make room for user writes.
Create a helper function that allows to kick GC without exposing the
internals of the GC/rate-limiter interaction.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
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>
Use a constant to set the maximum number of inflight GC requests
allowed.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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.
This allows us to have a much smaller mempool.
Reported-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For now, we allocate a per I/O buffer for GC data. Since the potential
size of the buffer is 256KB and GC is not in the fast path, do this
allocation with vmalloc. This puts lets pressure on the memory
allocator at no performance cost.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use the right types and conversions on le64 variables. Reported by
sparse.
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>
At the moment, we separate the closed lines on three different list
based on their number of valid sectors. GC recycles lines from each list
based on capacity. Lines from each list are taken in a FIFO fashion.
Since the number of lines is limited (it corresponds to the number of
blocks in a LUN, which is somewhere between 1000-2000), we can afford
scanning the lists to choose the optimal line to be recycled. This helps
specially in lines with a high number of valid sectors.
If the number of blocks per LUN increases, we will consider a more
efficient policy.
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, line metadata is persisted on a separate work queue, that
is kicked each time that a line is closed. The assumption when designing
this was that freeing the write thread from creating a new write request
was better than the potential impact of writes colliding on the media
(user I/O and metadata I/O). Experimentation has proven that this
assumption is wrong; collision can cause up to 25% of bandwidth and
introduce long tail latencies on the write thread, which potentially
cause user write threads to spend more time spinning to get a free entry
on the write buffer.
This patch moves the metadata logic to the write thread. When a line is
closed, remaining metadata is written in memory and is placed on a
metadata queue. The write thread then takes the metadata corresponding
to the previous line, creates the write request and schedules it to
minimize collisions on the media. Using this approach, we see that we
can saturate the media's bandwidth, which helps reducing both write
latencies and the spinning time for user writer threads.
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 block erases fail, these blocks are marked bad. The number of valid
blocks in the line was not updated, which could cause an infinite loop
on the erase path.
Fix this atomic counter and, in order to avoid taking an irq lock on the
interrupt context, make the erase counters atomic too.
Also, in the case that a significant number of blocks become bad in a
line, the result is the double shared metadata buffer (emeta) to stop
the pipeline until all metadata is flushed to the media. Increase the
number of metadata lines from 2 to 4 to avoid this case.
Fixes: a4bd217b43 "lightnvm: physical block device (pblk) target"
Signed-off-by: Javier González <javier@cnexlabs.com>
Reviewed-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
These labels are reversed so we could end up dereferencing an error
pointer or leaking.
Fixes: 7f347ba6bb3a ("lightnvm: physical block device (pblk) target")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
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>
