Segment info indexing also used sizeof(struct) instead of the
4K metadata stride, so info_index could point between slots and
subsequent writes would advance incorrectly. Derive info_index
from the pointer returned by the segment meta search using
PCACHE_SEG_INFO_SIZE and advance to the next slot for future
updates.
Signed-off-by: Li Chen <chenl311@chinatelecom.cn>
Signed-off-by: Dongsheng Yang <dongsheng.yang@linux.dev>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Reviewed-by: Zheng Gu <cengku@gmail.com>
Cc: stable@vger.kernel.org # 6.18
In dm-pcache, in order to ensure crash-consistency, a dual-copy scheme
is used to alternately update metadata, and there is a slot index that
records the current slot. However, in the write path the current
implementation writes directly to the current slot indexed by slot
index, and then advances the slot — which ends up overwriting the
existing slot, violating the crash-consistency guarantee.
This patch fixes that behavior, preventing metadata from being
overwritten incorrectly.
In addition, this patch add a missing pmem_wmb() after memcpy_flushcache().
Signed-off-by: Dongsheng Yang <dongsheng.yang@linux.dev>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Reviewed-by: Zheng Gu <cengku@gmail.com>
Cc: stable@vger.kernel.org # 6.18
The smatch checker reports a “scheduler in atomic context” problem in
the following call chain:
miss_read_end_req()
-> cache_seg_put()
-> cache_seg_invalidate()
-> cache_seg_gen_increase()
-> mutex_lock(&cache_seg->ctrl_lock);
In practice, this `mutex_lock` will not actually schedule, because it is
only called when `cache_seg_put()` drops the last reference, which is
single-threaded. That is also why the issue never shows up during real
testing.
However, the code is still buggy. The original purpose of `ctrl_lock`
was to prevent read/write conflicts on the cache segment control
information. Looking at the current usage, all control information
accesses are single-threaded: reads only occur during the init phase,
where no conflicts are possible, and writes happen once in the init
phase (also single-threaded) and once when `cache_seg_put()` drops the
last reference (again single-threaded).
Therefore, this patch removes `ctrl_lock` entirely and adds comments in
the appropriate places to document this logic.
Signed-off-by: Dongsheng Yang <dongsheng.yang@linux.dev>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
This patch introduces dm-pcache, a new DM target that places a DAX-
capable persistent-memory device in front of any slower block device and
uses it as a high-throughput, low-latency cache.
Design highlights
-----------------
- DAX data path – data is copied directly between DRAM and the pmem
mapping, bypassing the block layer’s overhead.
- Segmented, crash-consistent layout
- all layout metadata are dual-replicated CRC-protected.
- atomic kset flushes; key replay on mount guarantees cache integrity
even after power loss.
- Striped multi-tree index
- Multi‑tree indexing for high parallelism.
- overlap-resolution logic ensures non-intersecting cached extents.
- Background services
- write-back worker flushes dirty keys in order, preserving backing-device
crash consistency. This is important for checkpoint in cloud storage.
- garbage collector reclaims clean segments when utilisation exceeds a
tunable threshold.
- Data integrity – optional CRC32 on cached payload; metadata always protected.
Comparison with existing block-level caches
---------------------------------------------------------------------------------------------------------------------------------
| Feature | pcache (this patch) | bcache | dm-writecache |
|----------------------------------|---------------------------------|------------------------------|---------------------------|
| pmem access method | DAX | bio (block I/O) | DAX |
| Write latency (4 K rand-write) | ~5 µs | ~20 µs | ~5 µs |
| Concurrency | multi subtree index | global index tree | single tree + wc_lock |
| IOPS (4K randwrite, 32 numjobs) | 2.1 M | 352 K | 283 K |
| Read-cache support | YES | YES | NO |
| Deployment | no re-format of backend | backend devices must be | no re-format of backend |
| | | reformatted | |
| Write-back ordering | log-structured; | no ordering guarantee | no ordering guarantee |
| | preserves app-IO-order | | |
| Data integrity checks | metadata + data CRC(optional) | metadata CRC only | none |
---------------------------------------------------------------------------------------------------------------------------------
Signed-off-by: Dongsheng Yang <dongsheng.yang@linux.dev>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
