| Message ID | 20250410-ublk_task_per_io-v3-2-b811e8f4554a@purestorage.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | ublk: decouple server threads from hctxs | expand |
On Thu, Apr 10, 2025 at 06:17:51PM -0600, Uday Shankar wrote: > Currently, ublk_drv associates to each hardware queue (hctx) a unique > task (called the queue's ubq_daemon) which is allowed to issue > COMMIT_AND_FETCH commands against the hctx. If any other task attempts > to do so, the command fails immediately with EINVAL. When considered > together with the block layer architecture, the result is that for each > CPU C on the system, there is a unique ublk server thread which is > allowed to handle I/O submitted on CPU C. This can lead to suboptimal > performance under imbalanced load generation. For an extreme example, > suppose all the load is generated on CPUs mapping to a single ublk > server thread. Then that thread may be fully utilized and become the > bottleneck in the system, while other ublk server threads are totally > idle. > > This issue can also be addressed directly in the ublk server without > kernel support by having threads dequeue I/Os and pass them around to > ensure even load. But this solution requires inter-thread communication > at least twice for each I/O (submission and completion), which is > generally a bad pattern for performance. The problem gets even worse > with zero copy, as more inter-thread communication would be required to > have the buffer register/unregister calls to come from the correct > thread. Agree. The limit is actually originated from current implementation, both REGISTER_IO_BUF and UNREGISTER_IO_BUF should be fine to run from other pthread because the request buffer 'meta' is actually read-only. > > Therefore, address this issue in ublk_drv by requiring a unique task per > I/O instead of per queue/hctx. Imbalanced load can then be balanced > across all ublk server threads by having threads issue FETCH_REQs in a > round-robin manner. As a small toy example, consider a system with a > single ublk device having 2 queues, each of queue depth 4. A ublk server > having 4 threads could issue its FETCH_REQs against this device as > follows (where each entry is the qid,tag pair that the FETCH_REQ > targets): > > poller thread: T0 T1 T2 T3 > 0,0 0,1 0,2 0,3 > 1,3 1,0 1,1 1,2 > > Since tags appear to be allocated in sequential chunks, this setup > provides a rough approximation to distributing I/Os round-robin across > all ublk server threads, while letting I/Os stay fully thread-local. BLK_MQ_F_TAG_RR can be set for this way, so is it possible to make this as one feature? And set BLK_MQ_F_TAG_RR for this feature. Also can you share what the preferred implementation is for ublk server? I think per-io pthread may not be good, maybe partition tags space into fixed range/pthread? `ublk_queue' reference is basically read-only in IO code path, I think it need to be declared explicitly as 'const' pointer in IO code/uring code path first. Otherwise, it is easy to trigger data race with per-io task since it is lockless. Thanks, Ming
On Fri, Apr 11, 2025 at 04:53:19PM +0800, Ming Lei wrote: > On Thu, Apr 10, 2025 at 06:17:51PM -0600, Uday Shankar wrote: > > Currently, ublk_drv associates to each hardware queue (hctx) a unique > > task (called the queue's ubq_daemon) which is allowed to issue > > COMMIT_AND_FETCH commands against the hctx. If any other task attempts > > to do so, the command fails immediately with EINVAL. When considered > > together with the block layer architecture, the result is that for each > > CPU C on the system, there is a unique ublk server thread which is > > allowed to handle I/O submitted on CPU C. This can lead to suboptimal > > performance under imbalanced load generation. For an extreme example, > > suppose all the load is generated on CPUs mapping to a single ublk > > server thread. Then that thread may be fully utilized and become the > > bottleneck in the system, while other ublk server threads are totally > > idle. > > > > This issue can also be addressed directly in the ublk server without > > kernel support by having threads dequeue I/Os and pass them around to > > ensure even load. But this solution requires inter-thread communication > > at least twice for each I/O (submission and completion), which is > > generally a bad pattern for performance. The problem gets even worse > > with zero copy, as more inter-thread communication would be required to > > have the buffer register/unregister calls to come from the correct > > thread. > > Agree. > > The limit is actually originated from current implementation, both > REGISTER_IO_BUF and UNREGISTER_IO_BUF should be fine to run from other > pthread because the request buffer 'meta' is actually read-only. > > > > > Therefore, address this issue in ublk_drv by requiring a unique task per > > I/O instead of per queue/hctx. Imbalanced load can then be balanced > > across all ublk server threads by having threads issue FETCH_REQs in a > > round-robin manner. As a small toy example, consider a system with a > > single ublk device having 2 queues, each of queue depth 4. A ublk server > > having 4 threads could issue its FETCH_REQs against this device as > > follows (where each entry is the qid,tag pair that the FETCH_REQ > > targets): > > > > poller thread: T0 T1 T2 T3 > > 0,0 0,1 0,2 0,3 > > 1,3 1,0 1,1 1,2 > > > > Since tags appear to be allocated in sequential chunks, this setup > > provides a rough approximation to distributing I/Os round-robin across > > all ublk server threads, while letting I/Os stay fully thread-local. > > BLK_MQ_F_TAG_RR can be set for this way, so is it possible to make this > as one feature? And set BLK_MQ_F_TAG_RR for this feature. Yes, it would be easy enough to add. However we have been testing with the v1 patch [1] for a while now, and have seen pretty even load balancing even without BLK_MQ_F_TAG_RR. So I am not sure if it is worth it/if we will use the flag, especially considering that it is documented as reducing performance. [1] https://lore.kernel.org/all/20241002224437.3088981-1-ushankar@purestorage.com/ > Also can you share what the preferred implementation is for ublk server? > > I think per-io pthread may not be good, maybe partition tags space into > fixed range/pthread? By "unique task per io" I mean that each io can have its own task (including two ios in the same queue can have different tasks), but two ios can have the same task. That's roughly what we're doing, we have a handful of threads (around 8-16) and we split up the I/Os between them. With this patch we lift the restriction that each thread corresponds 1:1 with a ublk_queue/hctx. > `ublk_queue' reference is basically read-only in IO code path, I think > it need to be declared explicitly as 'const' pointer in IO code/uring code > path first. Otherwise, it is easy to trigger data race with per-io task > since it is lockless. That is a good suggestion.
On Tue, Apr 15, 2025 at 06:12:09PM -0600, Uday Shankar wrote: > On Fri, Apr 11, 2025 at 04:53:19PM +0800, Ming Lei wrote: > > On Thu, Apr 10, 2025 at 06:17:51PM -0600, Uday Shankar wrote: > > > Currently, ublk_drv associates to each hardware queue (hctx) a unique > > > task (called the queue's ubq_daemon) which is allowed to issue > > > COMMIT_AND_FETCH commands against the hctx. If any other task attempts > > > to do so, the command fails immediately with EINVAL. When considered > > > together with the block layer architecture, the result is that for each > > > CPU C on the system, there is a unique ublk server thread which is > > > allowed to handle I/O submitted on CPU C. This can lead to suboptimal > > > performance under imbalanced load generation. For an extreme example, > > > suppose all the load is generated on CPUs mapping to a single ublk > > > server thread. Then that thread may be fully utilized and become the > > > bottleneck in the system, while other ublk server threads are totally > > > idle. > > > > > > This issue can also be addressed directly in the ublk server without > > > kernel support by having threads dequeue I/Os and pass them around to > > > ensure even load. But this solution requires inter-thread communication > > > at least twice for each I/O (submission and completion), which is > > > generally a bad pattern for performance. The problem gets even worse > > > with zero copy, as more inter-thread communication would be required to > > > have the buffer register/unregister calls to come from the correct > > > thread. > > > > Agree. > > > > The limit is actually originated from current implementation, both > > REGISTER_IO_BUF and UNREGISTER_IO_BUF should be fine to run from other > > pthread because the request buffer 'meta' is actually read-only. > > > > > > > > Therefore, address this issue in ublk_drv by requiring a unique task per > > > I/O instead of per queue/hctx. Imbalanced load can then be balanced > > > across all ublk server threads by having threads issue FETCH_REQs in a > > > round-robin manner. As a small toy example, consider a system with a > > > single ublk device having 2 queues, each of queue depth 4. A ublk server > > > having 4 threads could issue its FETCH_REQs against this device as > > > follows (where each entry is the qid,tag pair that the FETCH_REQ > > > targets): > > > > > > poller thread: T0 T1 T2 T3 > > > 0,0 0,1 0,2 0,3 > > > 1,3 1,0 1,1 1,2 > > > > > > Since tags appear to be allocated in sequential chunks, this setup > > > provides a rough approximation to distributing I/Os round-robin across > > > all ublk server threads, while letting I/Os stay fully thread-local. > > > > BLK_MQ_F_TAG_RR can be set for this way, so is it possible to make this > > as one feature? And set BLK_MQ_F_TAG_RR for this feature. > > Yes, it would be easy enough to add. However we have been testing with That is why I suggest to add it as one feature, such as, PER_IO_TASK, then you can run any optimization on this feature only in future. There are other differences for this feature, such as, how to set each io task's affinity, how to partition tag space in optimized way, ... BTW, recently I found it is helpful to get good perf by only selecting one cpu as the queue thread's sched affinity. One feature flag also has document benefit. Also `Documentation/block/ublk.rst` need to be updated with this change/feature. Fortunately the cancel code patch has been generic enough to cover PER_IO_TASK already. > the v1 patch [1] for a while now, and have seen pretty even load > balancing even without BLK_MQ_F_TAG_RR. So I am not sure if it is worth > it/if we will use the flag, especially considering that it is documented > as reducing performance. per-io task actually depends on IO balance over each partitioned tag space, which highly relies on tag allocation algorithm. > > [1] https://lore.kernel.org/all/20241002224437.3088981-1-ushankar@purestorage.com/ > > > Also can you share what the preferred implementation is for ublk server? > > > > I think per-io pthread may not be good, maybe partition tags space into > > fixed range/pthread? > > By "unique task per io" I mean that each io can have its own task > (including two ios in the same queue can have different tasks), but two > ios can have the same task. > > That's roughly what we're doing, we have a handful of threads (around > 8-16) and we split up the I/Os between them. With this patch we lift the > restriction that each thread corresponds 1:1 with a ublk_queue/hctx. OK, care to add one command line(such as queue_tasks) to enable it in ublk kernel selftest? Then it can serve: - the added code can be covered in selftest - avoid to break this feature by future change - example for showing how to use this feature - run performance evaluation with different target setting The main change should be in ublk_io_handler_fn() & ublk_queue_init() by allocating one io_uring array for each queue. For target code, we already have ublk_queue_alloc_sqes(), in which the ring selection can be done centrally & transparently. > > > `ublk_queue' reference is basically read-only in IO code path, I think > > it need to be declared explicitly as 'const' pointer in IO code/uring code > > path first. Otherwise, it is easy to trigger data race with per-io task > > since it is lockless. > > That is a good suggestion. Great to see you have started it. Maybe it can be the prepare patches, which can be merged first. Thanks, Ming
diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c index 812789f58704cece9b661713cd0107807c789531..567cd314480be6273876ff7918250efc1e1b242f 100644 --- a/drivers/block/ublk_drv.c +++ b/drivers/block/ublk_drv.c @@ -150,6 +150,7 @@ struct ublk_io { int res; struct io_uring_cmd *cmd; + struct task_struct *task; }; struct ublk_queue { @@ -157,11 +158,9 @@ struct ublk_queue { int q_depth; unsigned long flags; - struct task_struct *ubq_daemon; struct ublksrv_io_desc *io_cmd_buf; bool force_abort; - bool timeout; bool canceling; bool fail_io; /* copy of dev->state == UBLK_S_DEV_FAIL_IO */ unsigned short nr_io_ready; /* how many ios setup */ @@ -1072,11 +1071,6 @@ static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu( return io_uring_cmd_to_pdu(ioucmd, struct ublk_uring_cmd_pdu); } -static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq) -{ - return ubq->ubq_daemon->flags & PF_EXITING; -} - /* todo: handle partial completion */ static inline void __ublk_complete_rq(struct request *req) { @@ -1202,13 +1196,13 @@ static void ublk_dispatch_req(struct ublk_queue *ubq, /* * Task is exiting if either: * - * (1) current != ubq_daemon. + * (1) current != io->task. * io_uring_cmd_complete_in_task() tries to run task_work - * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING. + * in a workqueue if cmd's task is PF_EXITING. * * (2) current->flags & PF_EXITING. */ - if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) { + if (unlikely(current != io->task || current->flags & PF_EXITING)) { __ublk_abort_rq(ubq, req); return; } @@ -1314,23 +1308,20 @@ static void ublk_queue_cmd_list(struct ublk_queue *ubq, struct rq_list *l) static enum blk_eh_timer_return ublk_timeout(struct request *rq) { struct ublk_queue *ubq = rq->mq_hctx->driver_data; + struct ublk_io *io = &ubq->ios[rq->tag]; unsigned int nr_inflight = 0; int i; if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) { - if (!ubq->timeout) { - send_sig(SIGKILL, ubq->ubq_daemon, 0); - ubq->timeout = true; - } - + send_sig(SIGKILL, io->task, 0); return BLK_EH_DONE; } - if (!ubq_daemon_is_dying(ubq)) + if (!(io->task->flags & PF_EXITING)) return BLK_EH_RESET_TIMER; for (i = 0; i < ubq->q_depth; i++) { - struct ublk_io *io = &ubq->ios[i]; + io = &ubq->ios[i]; if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) nr_inflight++; @@ -1529,8 +1520,8 @@ static void ublk_commit_completion(struct ublk_device *ub, } /* - * Called from ubq_daemon context via cancel fn, meantime quiesce ublk - * blk-mq queue, so we are called exclusively with blk-mq and ubq_daemon + * Called from io task context via cancel fn, meantime quiesce ublk + * blk-mq queue, so we are called exclusively with blk-mq and io task * context, so everything is serialized. */ static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq) @@ -1646,13 +1637,13 @@ static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd, return; task = io_uring_cmd_get_task(cmd); - if (WARN_ON_ONCE(task && task != ubq->ubq_daemon)) + io = &ubq->ios[pdu->tag]; + if (WARN_ON_ONCE(task && task != io->task)) return; ub = ubq->dev; need_schedule = ublk_abort_requests(ub, ubq); - io = &ubq->ios[pdu->tag]; WARN_ON_ONCE(io->cmd != cmd); ublk_cancel_cmd(ubq, io, issue_flags); @@ -1813,8 +1804,6 @@ static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq) { ubq->nr_io_ready++; if (ublk_queue_ready(ubq)) { - ubq->ubq_daemon = current; - get_task_struct(ubq->ubq_daemon); ub->nr_queues_ready++; if (capable(CAP_SYS_ADMIN)) @@ -1928,14 +1917,14 @@ static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd, if (!ubq || ub_cmd->q_id != ubq->q_id) goto out; - if (ubq->ubq_daemon && ubq->ubq_daemon != current) - goto out; - if (tag >= ubq->q_depth) goto out; io = &ubq->ios[tag]; + if (io->task && io->task != current) + goto out; + /* there is pending io cmd, something must be wrong */ if (io->flags & UBLK_IO_FLAG_ACTIVE) { ret = -EBUSY; @@ -1996,6 +1985,7 @@ static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd, ublk_fill_io_cmd(io, cmd, ub_cmd->addr); ublk_mark_io_ready(ub, ubq); + io->task = get_task_struct(current); mutex_unlock(&ub->mutex); break; case UBLK_IO_COMMIT_AND_FETCH_REQ: @@ -2235,9 +2225,15 @@ static void ublk_deinit_queue(struct ublk_device *ub, int q_id) { int size = ublk_queue_cmd_buf_size(ub, q_id); struct ublk_queue *ubq = ublk_get_queue(ub, q_id); + struct ublk_io *io; + int i; + + for (i = 0; i < ubq->q_depth; i++) { + io = &ubq->ios[i]; + if (io->task) + put_task_struct(io->task); + } - if (ubq->ubq_daemon) - put_task_struct(ubq->ubq_daemon); if (ubq->io_cmd_buf) free_pages((unsigned long)ubq->io_cmd_buf, get_order(size)); } @@ -2928,15 +2924,8 @@ static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq) { int i; - WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq))); - /* All old ioucmds have to be completed */ ubq->nr_io_ready = 0; - /* old daemon is PF_EXITING, put it now */ - put_task_struct(ubq->ubq_daemon); - /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */ - ubq->ubq_daemon = NULL; - ubq->timeout = false; ubq->canceling = false; for (i = 0; i < ubq->q_depth; i++) { @@ -2946,6 +2935,10 @@ static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq) io->flags = 0; io->cmd = NULL; io->addr = 0; + + WARN_ON_ONCE(!(io->task && (io->task->flags & PF_EXITING))); + put_task_struct(io->task); + io->task = NULL; } } @@ -2986,7 +2979,7 @@ static int ublk_ctrl_start_recovery(struct ublk_device *ub, pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id); for (i = 0; i < ub->dev_info.nr_hw_queues; i++) ublk_queue_reinit(ub, ublk_get_queue(ub, i)); - /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */ + /* set to NULL, otherwise new tasks cannot mmap the io_cmd_buf */ ub->mm = NULL; ub->nr_queues_ready = 0; ub->nr_privileged_daemon = 0; @@ -3005,14 +2998,14 @@ static int ublk_ctrl_end_recovery(struct ublk_device *ub, int ret = -EINVAL; int i; - pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n", - __func__, ub->dev_info.nr_hw_queues, header->dev_id); - /* wait until new ubq_daemon sending all FETCH_REQ */ + pr_devel("%s: Waiting for all FETCH_REQs, dev id %d...\n", __func__, + header->dev_id); + if (wait_for_completion_interruptible(&ub->completion)) return -EINTR; - pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n", - __func__, ub->dev_info.nr_hw_queues, header->dev_id); + pr_devel("%s: All FETCH_REQs received, dev id %d\n", __func__, + header->dev_id); mutex_lock(&ub->mutex); if (ublk_nosrv_should_stop_dev(ub))