On 2026/10/10 18:16, illofspeed wrote: > v2: no code changes. The only change is my author address: v1 went out on > 2026-10-10 from my previous address. Please apply this version instead of v1. > > These patches fix two problems in dm-zoned's reclaim worker, found while > running md RAID5 on three dm-zoned targets on 27 TB host-managed SMR drives > (WD Ultrastar DC HC680), each target with a regular cache device. > > 1. On a target with cache zones and no free sequential zone, an idle > target keeps copying random zones into other random zones, forever > (patch 2; patch 1 makes the resulting -ENOSPC back off instead of > spinning). After a full md rebuild every chunk is mapped, so this > happens right after a rebuild: the idle member read and wrote about > 72 MB/s nonstop for hours. > > 2. After a reclaim pass that ends while the target is busy, the idle poll > is never re-armed, so zones filled under load are not reclaimed when the > target goes idle (patch 3). Users notice it as "the buffer never drains"; > the workaround has been a timer sending "dmsetup message <dev> 0 reclaim". > > Testing: the three patches were built as an out-of-tree dm-zoned module for > 7.2.6 (the changed code is identical in current master) and tested > - on emulated host-managed drives (tcmu-runner ZBC handler) in a VM: the > idle copy loop reproduced with the stock module (340 MB/s on an idle > target, zone counters unchanged), 0 MB/s with the patches; > - on the three real drives since 2026-10-05: crash test, three 1.5 TiB > write benchmarks, and eight conversions between single-device and > cache-device layouts; emptied zones were reclaimed within minutes of the > targets going idle without the reclaim timer. > - with a reproducer that needs no special hardware (scsi_debug zbc=managed, > 64 MiB zones, plus a 1 GiB loop cache device): stock module ~600 MB/s of > read+write on the idle target with unchanged zone counters, patched > module 0 MB/s: > https://github.com/illofspeed/synology-host-managed-smr/tree/main/repro > Not done: a test on current master itself (only 7.2.6), and a rebuild onto a > cached member with the patches (the first problem was seen after a rebuild > with the stock module). > > Related, not addressed here: the double-free in dmz_load_sb() reported on > dm-devel on 2026-05-30. > > The problems were found and the patches written with the help of an AI > coding assistant (hence Assisted-by); the analysis, the tests on real > hardware and this description were reviewed by me.
What changes from v1? > > illofspeed (3): > dm zoned: back off when reclaim finds no destination zone > dm zoned: do not reclaim a random zone into another random zone > dm zoned: keep polling for idle reclaim after a reclaim pass > > drivers/md/dm-zoned-reclaim.c | 28 ++++++++++++++++++++++++++-- > 1 file changed, 26 insertions(+), 2 deletions(-) > -- Damien Le Moal Western Digital Research
