[Devel] [PATCH vz9 v2 00/65] ploop optimistations and scalling

[Pavel Tikhomirov]

Looks good except one missed hunk.

We can handle excess pio->llist.next = NULL and INIT_LIST_HEADs, and 
list_for_each_ENTRY_safe separately later, as those are not directly 
connected to this rework.

note: Looks like you've accidentally over-wrote original patch authors.

Reviewed-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>

On 2/12/25 17:32, Konstantin Khorenko wrote:
> Ploop processes requsts in a different threads in parallel
> where possible which results in significant improvement in
> performance and makes further optimistations possible.
> 
> v1:
>    - addressed feedback, i've left a few requests to merge changes
>      into bigger patches out, as to keep changes in smaller chunks
>    - patches merged and separated changes - generic cleanup
>    - fix endio for md page writeback
>    - background async allocation of space
>    - move enospc pios dispatch from timer to thread
>    - fixed allocations in atomic context
>    - fixed locking wrt userspace/interrupt context
>    - reworked discard cleanup
>    - removed workqueue, only use kthread code
>    - make metadata writeback works in parallel
>    - end fsync pios in parallel
>    - fixed current flags manipulation
> 
> v2:
>    - commit "dm-ploop: introduce pio.llist" appeared
>      It introduces the the pio.llist along with pio.list, gathered in an
>      union, so during the processes of switching between list users and
>      llist users, we could use appropriate fields without strict pointers
>      casts.
> 
> TODO:
>    - there are multiple INIT_LIST_HEAD() left in the code, most of them
>      are just not needed, but let's looks precisely at each case.
>      Same for "pio->llist.next = NULL;"
> 
>    - md_page.wb_link switched to llist and later - back to list along
>      with introducing ploop.wb_batch_list and ploop.wb_batch_list_prio.
>      Need to understand if we want the latter to be lists or llists and
>      don't push wb_link back and forth.
> 
> RFC series:
> v6 changes:
>   - locking fixes
>   - use READ_ONCE/WRITE_ONCE
>   - properly handle DIO errors (ENOTBLK)
>   - added support for data pios with REQ_FUA
>   - convert more lists to lockless variants
>   - resize/grow is good
> 
> v5 with changes in v4:
>   - BAT writeback delay
>   - Lists/Locking moving to lockless lists to reduce locking
>   - Requests are prepared in the caller to reduce load of worker thread
>   - REQ_FUA and REQ_PREFLUSH support
>   - Convert bit fields to use atomic bitops - reduce locking
>   - remove fsync worker to guarantee order of execution
>   - Map all md pages at creation time - they are only freed on device
>      destroy
> 
> Known bugs:
>    - delayed metadata writeback is not working and is missing error handling
>       - patch to disable it until fixed
>    - fast path is not working - causes rcu lockups - patch to disable it
> 
> Further improvements:
>    - optimize md pages lookups
> 
> Alexander Atanasov (42):
>    dm-ploop: md_pages map all pages at creation time
>    dm-ploop: Use READ_ONCE/WRITE_ONCE to access md page data
>    dm-ploop: fsync after all pios are sent
>    dm-ploop: move md status to use proper bitops
>    dm-ploop: convert wait_list and wb_batch_llist to use lockless lists
>    dm-ploop: convert enospc handling to use lockless lists
>    dm-ploop: convert suspended_pios list to use lockless list
>    dm-ploop: move from wq to kthread
>    dm-ploop: do not use a wrapper for set_bit to make a page writeback
>    dm-ploop: BAT use only one list for writeback
>    dm-ploop: make md writeback timeout to be per page
>    dm-ploop: add interface to disable bat writeback delay
>    dm-ploop: convert wb_batch_list to lockless variant
>    dm-ploop: convert high_prio to status
>    dm-ploop: convert md page rw lock to spin lock
>    dm-ploop: convert bat_rwlock to bat_lock spinlock
>    dm-ploop: prepare bat updates under bat_lock
>    dm-ploop: make ploop_bat_write_complete ready for parallel pio
>      completion
>    dm-ploop: make ploop_submit_metadata_writeback return number of
>      requests sent
>    dm-ploop: introduce pio runner threads
>    dm-ploop: add pio list ids to be used when passing pios to runners
>    dm-ploop: disable metadata writeback delay
>    dm-ploop: use lockless lists for chained cow updates list
>    dm-ploop: use lockless lists for data ready pios
>    dm-ploop: give runner threads better name
>    dm-ploop: resize operation - add holes bitmap locking
>    dm-ploop: remove unnecessary operations
>    dm-ploop: use filp per thread
>    dm-ploop: catch if we try to advance pio past bio end
>    dm-ploop: support REQ_FUA for data pios
>    dm-ploop: proplerly access nr_bat_entries
>    dm-ploop: fix locking and improve error handling when submitting pios
>    dm-ploop: fix how ENOTBLK is handled
>    dm-ploop: sync when suspended or stopping
>    dm-ploop: rework logic in pio processing
>    dm-ploop: end fsync pios in parallel
>    dm-ploop: make filespace preallocations async
>    dm-ploop: resubmit enospc pios from dispatcher thread
>    dm-ploop: dm-ploop: simplify discard completion
>    dm-ploop: use GFP_ATOMIC instead of GFP_NOIO
>    dm-ploop: fix locks used in mixed context
>    dm-ploop: fix how current flags are managed inside threads
> 
> Andrey Zhadchenko (13):
>    dm-ploop: do not flush after metadata writes
>    dm-ploop: set IOCB_DSYNC on all FUA requests
>    dm-ploop: remove extra ploop_cluster_is_in_top_delta()
>    dm-ploop: introduce per-md page locking
>    dm-ploop: reduce BAT accesses on discard completion
>    dm-ploop: simplify llseek
>    dm-ploop: speed up ploop_prepare_bat_update()
>    dm-ploop: make new allocations immediately visible in BAT
>    dm-ploop: drop ploop_cluster_is_in_top_delta()
>    dm-ploop: do not wait for BAT update for non-FUA requests
>    dm-ploop: add delay for metadata writeback
>    dm-ploop: submit all postponed metadata on REQ_OP_FLUSH
>    dm-ploop: handle REQ_PREFLUSH
> 
> Konstantin Khorenko (10):
>    dm-ploop: drop md_page.wb_link initialization
>    dm-ploop: introduce pio.llist
>    dm-ploop: convert the rest of the lists to use llist variant
>    dm-ploop: combine processing of pios thru prepare list and remove
>      fsync worker
>    dm-ploop: move preparations of pios into the caller from worker
>    dm-ploop: fast path execution for reads
>    dm-ploop: split cow processing into two functions
>    dm-ploop: process pios via runners
>    dm-ploop: disable fast path
>    dm-ploop: rework bat completion logic
> 
>   drivers/md/dm-ploop-bat.c    |   74 +-
>   drivers/md/dm-ploop-cmd.c    |  174 ++--
>   drivers/md/dm-ploop-map.c    | 1599 ++++++++++++++++++++++++----------
>   drivers/md/dm-ploop-target.c |  168 +++-
>   drivers/md/dm-ploop.h        |  120 ++-
>   5 files changed, 1484 insertions(+), 651 deletions(-)
> 

-- 
Best regards, Tikhomirov Pavel
Senior Software Developer, Virtuozzo.