From c6b7a3a26e809c9d2a51ae303764c1d2994f31cf Mon Sep 17 00:00:00 2001 From: Ming Lei Date: Sat, 24 Jun 2023 21:01:05 +0800 Subject: blk-mq: fix two misuses on RQF_USE_SCHED Request allocated from sched tags can't be issued via ->queue_rqs() directly, since driver tag isn't allocated yet. This is the 1st misuse of RQF_USE_SCHED for figuring out plug->has_elevator. Request allocated from sched tags can't be ended by blk_mq_end_request_batch() too, fix the 2nd RQF_USE_SCHED misuse in blk_mq_add_to_batch(). Without this patch, NVMe uring cmd passthrough IO workload can run into hang easily with real io scheduler. Fixes: dd6216bb16e8 ("blk-mq: make sure elevator callbacks aren't called for passthrough request") Reported-by: Guangwu Zhang Closes: https://lore.kernel.org/linux-block/CAGS2=YrBjpLPOKa-gzcKuuOG60AGth5794PNCDwatdnnscB9ug@mail.gmail.com/ Cc: Christoph Hellwig Signed-off-by: Ming Lei Link: https://lore.kernel.org/r/20230624130105.1443879-1-ming.lei@redhat.com Signed-off-by: Jens Axboe --- include/linux/blk-mq.h | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) (limited to 'include') diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h index f401067ac03a..aaed687a454c 100644 --- a/include/linux/blk-mq.h +++ b/include/linux/blk-mq.h @@ -852,7 +852,11 @@ static inline bool blk_mq_add_to_batch(struct request *req, struct io_comp_batch *iob, int ioerror, void (*complete)(struct io_comp_batch *)) { - if (!iob || (req->rq_flags & RQF_USE_SCHED) || ioerror || + /* + * blk_mq_end_request_batch() can't end request allocated from + * sched tags + */ + if (!iob || (req->rq_flags & RQF_SCHED_TAGS) || ioerror || (req->end_io && !blk_rq_is_passthrough(req))) return false; -- cgit From f6c80cffcd47a2d41943e3a41fbe9034d9f6d7b0 Mon Sep 17 00:00:00 2001 From: Keith Busch Date: Mon, 12 Jun 2023 12:03:42 -0700 Subject: block: add request polling helper Provide a direct request polling will for drivers. The interface does not require a bio, and can skip the overhead associated with polling those. The biggest gain from skipping the relatively expensive xarray lookup unnecessary when you already have the request. With this, the simple rq/qc conversion functions have only one caller each, so open code this and remove the helpers. Signed-off-by: Keith Busch Reviewed-by: Kanchan Joshi Reviewed-by: Sagi Grimberg Reviewed-by: Christoph Hellwig Link: https://lore.kernel.org/r/20230612190343.2087040-2-kbusch@meta.com Signed-off-by: Jens Axboe --- block/blk-mq.c | 48 ++++++++++++++++++++++++++++++++---------------- include/linux/blk-mq.h | 2 ++ 2 files changed, 34 insertions(+), 16 deletions(-) (limited to 'include') diff --git a/block/blk-mq.c b/block/blk-mq.c index 98eb31ff914d..5504719b970d 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -49,17 +49,8 @@ static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags); static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, struct list_head *list); - -static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, - blk_qc_t qc) -{ - return xa_load(&q->hctx_table, qc); -} - -static inline blk_qc_t blk_rq_to_qc(struct request *rq) -{ - return rq->mq_hctx->queue_num; -} +static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, + struct io_comp_batch *iob, unsigned int flags); /* * Check if any of the ctx, dispatch list or elevator @@ -1248,7 +1239,7 @@ void blk_mq_start_request(struct request *rq) q->integrity.profile->prepare_fn(rq); #endif if (rq->bio && rq->bio->bi_opf & REQ_POLLED) - WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); + WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); } EXPORT_SYMBOL(blk_mq_start_request); @@ -1354,7 +1345,7 @@ EXPORT_SYMBOL_GPL(blk_rq_is_poll); static void blk_rq_poll_completion(struct request *rq, struct completion *wait) { do { - blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0); + blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0); cond_resched(); } while (!completion_done(wait)); } @@ -4749,10 +4740,9 @@ void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) } EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); -int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, - unsigned int flags) +static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, + struct io_comp_batch *iob, unsigned int flags) { - struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); long state = get_current_state(); int ret; @@ -4777,6 +4767,32 @@ int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch * return 0; } +int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, + struct io_comp_batch *iob, unsigned int flags) +{ + struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie); + + return blk_hctx_poll(q, hctx, iob, flags); +} + +int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, + unsigned int poll_flags) +{ + struct request_queue *q = rq->q; + int ret; + + if (!blk_rq_is_poll(rq)) + return 0; + if (!percpu_ref_tryget(&q->q_usage_counter)) + return 0; + + ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); + blk_queue_exit(q); + + return ret; +} +EXPORT_SYMBOL_GPL(blk_rq_poll); + unsigned int blk_mq_rq_cpu(struct request *rq) { return rq->mq_ctx->cpu; diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h index aaed687a454c..2b7fb8e87793 100644 --- a/include/linux/blk-mq.h +++ b/include/linux/blk-mq.h @@ -715,6 +715,8 @@ int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, void blk_mq_free_tag_set(struct blk_mq_tag_set *set); void blk_mq_free_request(struct request *rq); +int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, + unsigned int poll_flags); bool blk_mq_queue_inflight(struct request_queue *q); -- cgit From 9408d8a37e6cce8803681ab816383450a056c3a9 Mon Sep 17 00:00:00 2001 From: Keith Busch Date: Mon, 12 Jun 2023 12:03:43 -0700 Subject: nvme: improved uring polling Drivers can poll requests directly, so use that. We just need to ensure the driver's request was allocated from a polled hctx, so a special driver flag is added to struct io_uring_cmd. The allows unshared and multipath namespaces to use the same polling callback, and multipath is guaranteed to get the same queue as the command was submitted on. Previously multipath polling might check a different path and poll the wrong info. The other bonus is we don't need a bio payload in order to poll, allowing commands like 'flush' and 'write zeroes' to be submitted on the same high priority queue as read and write commands. Finally, using the request based polling skips the unnecessary bio overhead. Signed-off-by: Keith Busch Reviewed-by: Sagi Grimberg Reviewed-by: Christoph Hellwig Link: https://lore.kernel.org/r/20230612190343.2087040-3-kbusch@meta.com Signed-off-by: Jens Axboe --- drivers/nvme/host/ioctl.c | 70 ++++++++++++------------------------------- drivers/nvme/host/multipath.c | 2 +- drivers/nvme/host/nvme.h | 2 -- include/uapi/linux/io_uring.h | 2 ++ 4 files changed, 22 insertions(+), 54 deletions(-) (limited to 'include') diff --git a/drivers/nvme/host/ioctl.c b/drivers/nvme/host/ioctl.c index 2130ad65b58c..5c3250f36ce7 100644 --- a/drivers/nvme/host/ioctl.c +++ b/drivers/nvme/host/ioctl.c @@ -505,7 +505,6 @@ static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req, { struct io_uring_cmd *ioucmd = req->end_io_data; struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd); - void *cookie = READ_ONCE(ioucmd->cookie); req->bio = pdu->bio; if (nvme_req(req)->flags & NVME_REQ_CANCELLED) @@ -518,10 +517,12 @@ static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req, * For iopoll, complete it directly. * Otherwise, move the completion to task work. */ - if (cookie != NULL && blk_rq_is_poll(req)) + if (blk_rq_is_poll(req)) { + WRITE_ONCE(ioucmd->cookie, NULL); nvme_uring_task_cb(ioucmd, IO_URING_F_UNLOCKED); - else + } else { io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb); + } return RQ_END_IO_FREE; } @@ -531,7 +532,6 @@ static enum rq_end_io_ret nvme_uring_cmd_end_io_meta(struct request *req, { struct io_uring_cmd *ioucmd = req->end_io_data; struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd); - void *cookie = READ_ONCE(ioucmd->cookie); req->bio = pdu->bio; pdu->req = req; @@ -540,10 +540,12 @@ static enum rq_end_io_ret nvme_uring_cmd_end_io_meta(struct request *req, * For iopoll, complete it directly. * Otherwise, move the completion to task work. */ - if (cookie != NULL && blk_rq_is_poll(req)) + if (blk_rq_is_poll(req)) { + WRITE_ONCE(ioucmd->cookie, NULL); nvme_uring_task_meta_cb(ioucmd, IO_URING_F_UNLOCKED); - else + } else { io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_meta_cb); + } return RQ_END_IO_NONE; } @@ -599,7 +601,6 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns, if (issue_flags & IO_URING_F_IOPOLL) rq_flags |= REQ_POLLED; -retry: req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags); if (IS_ERR(req)) return PTR_ERR(req); @@ -613,17 +614,11 @@ retry: return ret; } - if (issue_flags & IO_URING_F_IOPOLL && rq_flags & REQ_POLLED) { - if (unlikely(!req->bio)) { - /* we can't poll this, so alloc regular req instead */ - blk_mq_free_request(req); - rq_flags &= ~REQ_POLLED; - goto retry; - } else { - WRITE_ONCE(ioucmd->cookie, req->bio); - req->bio->bi_opf |= REQ_POLLED; - } + if (blk_rq_is_poll(req)) { + ioucmd->flags |= IORING_URING_CMD_POLLED; + WRITE_ONCE(ioucmd->cookie, req); } + /* to free bio on completion, as req->bio will be null at that time */ pdu->bio = req->bio; pdu->meta_len = d.metadata_len; @@ -785,18 +780,16 @@ int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, struct io_comp_batch *iob, unsigned int poll_flags) { - struct bio *bio; + struct request *req; int ret = 0; - struct nvme_ns *ns; - struct request_queue *q; + + if (!(ioucmd->flags & IORING_URING_CMD_POLLED)) + return 0; rcu_read_lock(); - bio = READ_ONCE(ioucmd->cookie); - ns = container_of(file_inode(ioucmd->file)->i_cdev, - struct nvme_ns, cdev); - q = ns->queue; - if (test_bit(QUEUE_FLAG_POLL, &q->queue_flags) && bio && bio->bi_bdev) - ret = bio_poll(bio, iob, poll_flags); + req = READ_ONCE(ioucmd->cookie); + if (req && blk_rq_is_poll(req)) + ret = blk_rq_poll(req, iob, poll_flags); rcu_read_unlock(); return ret; } @@ -890,31 +883,6 @@ int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd, srcu_read_unlock(&head->srcu, srcu_idx); return ret; } - -int nvme_ns_head_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, - struct io_comp_batch *iob, - unsigned int poll_flags) -{ - struct cdev *cdev = file_inode(ioucmd->file)->i_cdev; - struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev); - int srcu_idx = srcu_read_lock(&head->srcu); - struct nvme_ns *ns = nvme_find_path(head); - struct bio *bio; - int ret = 0; - struct request_queue *q; - - if (ns) { - rcu_read_lock(); - bio = READ_ONCE(ioucmd->cookie); - q = ns->queue; - if (test_bit(QUEUE_FLAG_POLL, &q->queue_flags) && bio - && bio->bi_bdev) - ret = bio_poll(bio, iob, poll_flags); - rcu_read_unlock(); - } - srcu_read_unlock(&head->srcu, srcu_idx); - return ret; -} #endif /* CONFIG_NVME_MULTIPATH */ int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags) diff --git a/drivers/nvme/host/multipath.c b/drivers/nvme/host/multipath.c index 98001eebd275..5aa1592849af 100644 --- a/drivers/nvme/host/multipath.c +++ b/drivers/nvme/host/multipath.c @@ -470,7 +470,7 @@ static const struct file_operations nvme_ns_head_chr_fops = { .unlocked_ioctl = nvme_ns_head_chr_ioctl, .compat_ioctl = compat_ptr_ioctl, .uring_cmd = nvme_ns_head_chr_uring_cmd, - .uring_cmd_iopoll = nvme_ns_head_chr_uring_cmd_iopoll, + .uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll, }; static int nvme_add_ns_head_cdev(struct nvme_ns_head *head) diff --git a/drivers/nvme/host/nvme.h b/drivers/nvme/host/nvme.h index 9a98c14c552a..791cafd9910a 100644 --- a/drivers/nvme/host/nvme.h +++ b/drivers/nvme/host/nvme.h @@ -854,8 +854,6 @@ long nvme_dev_ioctl(struct file *file, unsigned int cmd, unsigned long arg); int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, struct io_comp_batch *iob, unsigned int poll_flags); -int nvme_ns_head_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, - struct io_comp_batch *iob, unsigned int poll_flags); int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags); int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd, diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h index f222d263bc55..08720c7bd92f 100644 --- a/include/uapi/linux/io_uring.h +++ b/include/uapi/linux/io_uring.h @@ -244,8 +244,10 @@ enum io_uring_op { * sqe->uring_cmd_flags * IORING_URING_CMD_FIXED use registered buffer; pass this flag * along with setting sqe->buf_index. + * IORING_URING_CMD_POLLED driver use only */ #define IORING_URING_CMD_FIXED (1U << 0) +#define IORING_URING_CMD_POLLED (1U << 31) /* -- cgit