Pull io_uring updates from Jens Axboe: - Add supported for more directly managed task_work running. This is beneficial for real world applications that end up issuing lots of system calls as part of handling work. Normal task_work will always execute as we transition in and out of the kernel, even for "unrelated" system calls. It's more efficient to defer the handling of io_uring's deferred work until the application wants it to be run, generally in batches. As part of ongoing work to write an io_uring network backend for Thrift, this has been shown to greatly improve performance. (Dylan) - Add IOPOLL support for passthrough (Kanchan) - Improvements and fixes to the send zero-copy support (Pavel) - Partial IO handling fixes (Pavel) - CQE ordering fixes around CQ ring overflow (Pavel) - Support sendto() for non-zc as well (Pavel) - Support sendmsg for zerocopy (Pavel) - Networking iov_iter fix (Stefan) - Misc fixes and cleanups (Pavel, me) * tag 'for-6.1/io_uring-2022-10-03' of git://git.kernel.dk/linux: (56 commits) io_uring/net: fix notif cqe reordering io_uring/net: don't update msg_name if not provided io_uring: don't gate task_work run on TIF_NOTIFY_SIGNAL io_uring/rw: defer fsnotify calls to task context io_uring/net: fix fast_iov assignment in io_setup_async_msg() io_uring/net: fix non-zc send with address io_uring/net: don't skip notifs for failed requests io_uring/rw: don't lose short results on io_setup_async_rw() io_uring/rw: fix unexpected link breakage io_uring/net: fix cleanup double free free_iov init io_uring: fix CQE reordering io_uring/net: fix UAF in io_sendrecv_fail() selftest/net: adjust io_uring sendzc notif handling io_uring: ensure local task_work marks task as running io_uring/net: zerocopy sendmsg io_uring/net: combine fail handlers io_uring/net: rename io_sendzc() io_uring/net: support non-zerocopy sendto io_uring/net: refactor io_setup_async_addr io_uring/net: don't lose partial send_zc on fail ...
Unified · Split
| @@ -1233,7 +1233,7 @@ static void blk_end_sync_rq(struct request *rq, blk_status_t ret) · 1232 unmodified lines | ||
| 1233 | 1233 | complete(&wait->done); |
| 1234 | 1234 | } |
| 1235 | 1235 | |
| 1236 | −static bool blk_rq_is_poll(struct request *rq) | |
| 1236 | +bool blk_rq_is_poll(struct request *rq) | |
| 1237 | 1237 | { |
| 1238 | 1238 | if (!rq->mq_hctx) |
| 1239 | 1239 | return false; |
| @@ -1243,6 +1243,7 @@ static bool blk_rq_is_poll(struct request *rq) · 3 unmodified lines | ||
| 1243 | 1243 | return false; |
| 1244 | 1244 | return true; |
| 1245 | 1245 | } |
| 1246 | +EXPORT_SYMBOL_GPL(blk_rq_is_poll); | |
| 1246 | 1247 | |
| 1247 | 1248 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) |
| 1248 | 1249 | { |
| @@ -3976,6 +3976,7 @@ static const struct file_operations nvme_ns_chr_fops = { · 3975 unmodified lines | ||
| 3976 | 3976 | .unlocked_ioctl = nvme_ns_chr_ioctl, |
| 3977 | 3977 | .compat_ioctl = compat_ptr_ioctl, |
| 3978 | 3978 | .uring_cmd = nvme_ns_chr_uring_cmd, |
| 3979 | + .uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll, | |
| 3979 | 3980 | }; |
| 3980 | 3981 | |
| 3981 | 3982 | static int nvme_add_ns_cdev(struct nvme_ns *ns) |
| @@ -391,11 +391,19 @@ static void nvme_uring_cmd_end_io(struct request *req, blk_status_t err) · 390 unmodified lines | ||
| 391 | 391 | struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd); |
| 392 | 392 | /* extract bio before reusing the same field for request */ |
| 393 | 393 | struct bio *bio = pdu->bio; |
| 394 | + void *cookie = READ_ONCE(ioucmd->cookie); | |
| 394 | 395 | |
| 395 | 396 | pdu->req = req; |
| 396 | 397 | req->bio = bio; |
| 397 | − /* this takes care of moving rest of completion-work to task context */ | |
| 398 | − io_uring_cmd_complete_in_task(ioucmd, nvme_uring_task_cb); | |
| 398 | + | |
| 399 | + /* | |
| 400 | + * For iopoll, complete it directly. | |
| 401 | + * Otherwise, move the completion to task work. | |
| 402 | + */ | |
| 403 | + if (cookie != NULL && blk_rq_is_poll(req)) | |
| 404 | + nvme_uring_task_cb(ioucmd); | |
| 405 | + else | |
| 406 | + io_uring_cmd_complete_in_task(ioucmd, nvme_uring_task_cb); | |
| 399 | 407 | } |
| 400 | 408 | |
| 401 | 409 | static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns, |
| @@ -445,7 +453,10 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns, · 43 unmodified lines | ||
| 445 | 453 | rq_flags = REQ_NOWAIT; |
| 446 | 454 | blk_flags = BLK_MQ_REQ_NOWAIT; |
| 447 | 455 | } |
| 456 | + if (issue_flags & IO_URING_F_IOPOLL) | |
| 457 | + rq_flags |= REQ_POLLED; | |
| 448 | 458 | |
| 459 | +retry: | |
| 449 | 460 | req = nvme_alloc_user_request(q, &c, nvme_to_user_ptr(d.addr), |
| 450 | 461 | d.data_len, nvme_to_user_ptr(d.metadata), |
| 451 | 462 | d.metadata_len, 0, &meta, d.timeout_ms ? |
| @@ -456,6 +467,17 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns, · 4 unmodified lines | ||
| 456 | 467 | req->end_io = nvme_uring_cmd_end_io; |
| 457 | 468 | req->end_io_data = ioucmd; |
| 458 | 469 | |
| 470 | + if (issue_flags & IO_URING_F_IOPOLL && rq_flags & REQ_POLLED) { | |
| 471 | + if (unlikely(!req->bio)) { | |
| 472 | + /* we can't poll this, so alloc regular req instead */ | |
| 473 | + blk_mq_free_request(req); | |
| 474 | + rq_flags &= ~REQ_POLLED; | |
| 475 | + goto retry; | |
| 476 | + } else { | |
| 477 | + WRITE_ONCE(ioucmd->cookie, req->bio); | |
| 478 | + req->bio->bi_opf |= REQ_POLLED; | |
| 479 | + } | |
| 480 | + } | |
| 459 | 481 | /* to free bio on completion, as req->bio will be null at that time */ |
| 460 | 482 | pdu->bio = req->bio; |
| 461 | 483 | pdu->meta = meta; |
| @@ -559,9 +581,6 @@ long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg) · 97 unmodified lines | ||
| 559 | 581 | |
| 560 | 582 | static int nvme_uring_cmd_checks(unsigned int issue_flags) |
| 561 | 583 | { |
| 562 | − /* IOPOLL not supported yet */ | |
| 563 | − if (issue_flags & IO_URING_F_IOPOLL) | |
| 564 | − return -EOPNOTSUPP; | |
| 565 | 584 | |
| 566 | 585 | /* NVMe passthrough requires big SQE/CQE support */ |
| 567 | 586 | if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) != |
| @@ -604,6 +623,25 @@ int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags) · 36 unmodified lines | ||
| 604 | 623 | return nvme_ns_uring_cmd(ns, ioucmd, issue_flags); |
| 605 | 624 | } |
| 606 | 625 | |
| 626 | +int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, | |
| 627 | + struct io_comp_batch *iob, | |
| 628 | + unsigned int poll_flags) | |
| 629 | +{ | |
| 630 | + struct bio *bio; | |
| 631 | + int ret = 0; | |
| 632 | + struct nvme_ns *ns; | |
| 633 | + struct request_queue *q; | |
| 634 | + | |
| 635 | + rcu_read_lock(); | |
| 636 | + bio = READ_ONCE(ioucmd->cookie); | |
| 637 | + ns = container_of(file_inode(ioucmd->file)->i_cdev, | |
| 638 | + struct nvme_ns, cdev); | |
| 639 | + q = ns->queue; | |
| 640 | + if (test_bit(QUEUE_FLAG_POLL, &q->queue_flags) && bio && bio->bi_bdev) | |
| 641 | + ret = bio_poll(bio, iob, poll_flags); | |
| 642 | + rcu_read_unlock(); | |
| 643 | + return ret; | |
| 644 | +} | |
| 607 | 645 | #ifdef CONFIG_NVME_MULTIPATH |
| 608 | 646 | static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd, |
| 609 | 647 | void __user *argp, struct nvme_ns_head *head, int srcu_idx) |
| @@ -685,6 +723,31 @@ int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd, · 75 unmodified lines | ||
| 685 | 723 | srcu_read_unlock(&head->srcu, srcu_idx); |
| 686 | 724 | return ret; |
| 687 | 725 | } |
| 726 | + | |
| 727 | +int nvme_ns_head_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, | |
| 728 | + struct io_comp_batch *iob, | |
| 729 | + unsigned int poll_flags) | |
| 730 | +{ | |
| 731 | + struct cdev *cdev = file_inode(ioucmd->file)->i_cdev; | |
| 732 | + struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev); | |
| 733 | + int srcu_idx = srcu_read_lock(&head->srcu); | |
| 734 | + struct nvme_ns *ns = nvme_find_path(head); | |
| 735 | + struct bio *bio; | |
| 736 | + int ret = 0; | |
| 737 | + struct request_queue *q; | |
| 738 | + | |
| 739 | + if (ns) { | |
| 740 | + rcu_read_lock(); | |
| 741 | + bio = READ_ONCE(ioucmd->cookie); | |
| 742 | + q = ns->queue; | |
| 743 | + if (test_bit(QUEUE_FLAG_POLL, &q->queue_flags) && bio | |
| 744 | + && bio->bi_bdev) | |
| 745 | + ret = bio_poll(bio, iob, poll_flags); | |
| 746 | + rcu_read_unlock(); | |
| 747 | + } | |
| 748 | + srcu_read_unlock(&head->srcu, srcu_idx); | |
| 749 | + return ret; | |
| 750 | +} | |
| 688 | 751 | #endif /* CONFIG_NVME_MULTIPATH */ |
| 689 | 752 | |
| 690 | 753 | int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags) |
| @@ -692,6 +755,10 @@ int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags) · 1 unmodified line | ||
| 692 | 755 | struct nvme_ctrl *ctrl = ioucmd->file->private_data; |
| 693 | 756 | int ret; |
| 694 | 757 | |
| 758 | + /* IOPOLL not supported yet */ | |
| 759 | + if (issue_flags & IO_URING_F_IOPOLL) | |
| 760 | + return -EOPNOTSUPP; | |
| 761 | + | |
| 695 | 762 | ret = nvme_uring_cmd_checks(issue_flags); |
| 696 | 763 | if (ret) |
| 697 | 764 | return ret; |
| @@ -439,6 +439,7 @@ static const struct file_operations nvme_ns_head_chr_fops = { · 438 unmodified lines | ||
| 439 | 439 | .unlocked_ioctl = nvme_ns_head_chr_ioctl, |
| 440 | 440 | .compat_ioctl = compat_ptr_ioctl, |
| 441 | 441 | .uring_cmd = nvme_ns_head_chr_uring_cmd, |
| 442 | + .uring_cmd_iopoll = nvme_ns_head_chr_uring_cmd_iopoll, | |
| 442 | 443 | }; |
| 443 | 444 | |
| 444 | 445 | static int nvme_add_ns_head_cdev(struct nvme_ns_head *head) |
| @@ -821,6 +821,10 @@ long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd, · 820 unmodified lines | ||
| 821 | 821 | unsigned long arg); |
| 822 | 822 | long nvme_dev_ioctl(struct file *file, unsigned int cmd, |
| 823 | 823 | unsigned long arg); |
| 824 | +int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, | |
| 825 | + struct io_comp_batch *iob, unsigned int poll_flags); | |
| 826 | +int nvme_ns_head_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd, | |
| 827 | + struct io_comp_batch *iob, unsigned int poll_flags); | |
| 824 | 828 | int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, |
| 825 | 829 | unsigned int issue_flags); |
| 826 | 830 | int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd, |
| @@ -69,17 +69,17 @@ __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n) · 68 unmodified lines | ||
| 69 | 69 | * it returns false, the eventfd_signal() call should be deferred to a |
| 70 | 70 | * safe context. |
| 71 | 71 | */ |
| 72 | − if (WARN_ON_ONCE(current->in_eventfd_signal)) | |
| 72 | + if (WARN_ON_ONCE(current->in_eventfd)) | |
| 73 | 73 | return 0; |
| 74 | 74 | |
| 75 | 75 | spin_lock_irqsave(&ctx->wqh.lock, flags); |
| 76 | − current->in_eventfd_signal = 1; | |
| 76 | + current->in_eventfd = 1; | |
| 77 | 77 | if (ULLONG_MAX - ctx->count < n) |
| 78 | 78 | n = ULLONG_MAX - ctx->count; |
| 79 | 79 | ctx->count += n; |
| 80 | 80 | if (waitqueue_active(&ctx->wqh)) |
| 81 | 81 | wake_up_locked_poll(&ctx->wqh, EPOLLIN); |
| 82 | − current->in_eventfd_signal = 0; | |
| 82 | + current->in_eventfd = 0; | |
| 83 | 83 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); |
| 84 | 84 | |
| 85 | 85 | return n; |
| @@ -253,8 +253,10 @@ static ssize_t eventfd_read(struct kiocb *iocb, struct iov_iter *to) · 167 unmodified lines | ||
| 253 | 253 | __set_current_state(TASK_RUNNING); |
| 254 | 254 | } |
| 255 | 255 | eventfd_ctx_do_read(ctx, &ucnt); |
| 256 | + current->in_eventfd = 1; | |
| 256 | 257 | if (waitqueue_active(&ctx->wqh)) |
| 257 | 258 | wake_up_locked_poll(&ctx->wqh, EPOLLOUT); |
| 259 | + current->in_eventfd = 0; | |
| 258 | 260 | spin_unlock_irq(&ctx->wqh.lock); |
| 259 | 261 | if (unlikely(copy_to_iter(&ucnt, sizeof(ucnt), to) != sizeof(ucnt))) |
| 260 | 262 | return -EFAULT; |
| @@ -301,8 +303,10 @@ static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t c · 40 unmodified lines | ||
| 301 | 303 | } |
| 302 | 304 | if (likely(res > 0)) { |
| 303 | 305 | ctx->count += ucnt; |
| 306 | + current->in_eventfd = 1; | |
| 304 | 307 | if (waitqueue_active(&ctx->wqh)) |
| 305 | 308 | wake_up_locked_poll(&ctx->wqh, EPOLLIN); |
| 309 | + current->in_eventfd = 0; | |
| 306 | 310 | } |
| 307 | 311 | spin_unlock_irq(&ctx->wqh.lock); |
| 308 | 312 | |
| @@ -980,6 +980,7 @@ int blk_rq_map_kern(struct request_queue *, struct request *, void *, · 979 unmodified lines | ||
| 980 | 980 | int blk_rq_append_bio(struct request *rq, struct bio *bio); |
| 981 | 981 | void blk_execute_rq_nowait(struct request *rq, bool at_head); |
| 982 | 982 | blk_status_t blk_execute_rq(struct request *rq, bool at_head); |
| 983 | +bool blk_rq_is_poll(struct request *rq); | |
| 983 | 984 | |
| 984 | 985 | struct req_iterator { |
| 985 | 986 | struct bvec_iter iter; |
| @@ -46,7 +46,7 @@ void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt); · 45 unmodified lines | ||
| 46 | 46 | |
| 47 | 47 | static inline bool eventfd_signal_allowed(void) |
| 48 | 48 | { |
| 49 | − return !current->in_eventfd_signal; | |
| 49 | + return !current->in_eventfd; | |
| 50 | 50 | } |
| 51 | 51 | |
| 52 | 52 | #else /* CONFIG_EVENTFD */ |
| @@ -2133,6 +2133,8 @@ struct file_operations { · 2132 unmodified lines | ||
| 2133 | 2133 | loff_t len, unsigned int remap_flags); |
| 2134 | 2134 | int (*fadvise)(struct file *, loff_t, loff_t, int); |
| 2135 | 2135 | int (*uring_cmd)(struct io_uring_cmd *ioucmd, unsigned int issue_flags); |
| 2136 | + int (*uring_cmd_iopoll)(struct io_uring_cmd *, struct io_comp_batch *, | |
| 2137 | + unsigned int poll_flags); | |
| 2136 | 2138 | } __randomize_layout; |
| 2137 | 2139 | |
| 2138 | 2140 | struct inode_operations { |
| @@ -20,8 +20,12 @@ enum io_uring_cmd_flags { · 19 unmodified lines | ||
| 20 | 20 | struct io_uring_cmd { |
| 21 | 21 | struct file *file; |
| 22 | 22 | const void *cmd; |
| 23 | − /* callback to defer completions to task context */ | |
| 24 | − void (*task_work_cb)(struct io_uring_cmd *cmd); | |
| 23 | + union { | |
| 24 | + /* callback to defer completions to task context */ | |
| 25 | + void (*task_work_cb)(struct io_uring_cmd *cmd); | |
| 26 | + /* used for polled completion */ | |
| 27 | + void *cookie; | |
| 28 | + }; | |
| 25 | 29 | u32 cmd_op; |
| 26 | 30 | u32 pad; |
| 27 | 31 | u8 pdu[32]; /* available inline for free use */ |
| @@ -184,6 +184,8 @@ struct io_ev_fd { · 183 unmodified lines | ||
| 184 | 184 | struct eventfd_ctx *cq_ev_fd; |
| 185 | 185 | unsigned int eventfd_async: 1; |
| 186 | 186 | struct rcu_head rcu; |
| 187 | + atomic_t refs; | |
| 188 | + atomic_t ops; | |
| 187 | 189 | }; |
| 188 | 190 | |
| 189 | 191 | struct io_alloc_cache { |
| @@ -301,6 +303,8 @@ struct io_ring_ctx { · 111 unmodified lines | ||
| 301 | 303 | struct io_hash_table cancel_table; |
| 302 | 304 | bool poll_multi_queue; |
| 303 | 305 | |
| 306 | + struct llist_head work_llist; | |
| 307 | + | |
| 304 | 308 | struct list_head io_buffers_comp; |
| 305 | 309 | } ____cacheline_aligned_in_smp; |
| 306 | 310 | |
| @@ -936,7 +936,7 @@ struct task_struct { · 935 unmodified lines | ||
| 936 | 936 | #endif |
| 937 | 937 | #ifdef CONFIG_EVENTFD |
| 938 | 938 | /* Recursion prevention for eventfd_signal() */ |
| 939 | − unsigned in_eventfd_signal:1; | |
| 939 | + unsigned in_eventfd:1; | |
| 940 | 940 | #endif |
| 941 | 941 | #ifdef CONFIG_IOMMU_SVA |
| 942 | 942 | unsigned pasid_activated:1; |
| @@ -655,6 +655,35 @@ TRACE_EVENT(io_uring_short_write, · 654 unmodified lines | ||
| 655 | 655 | __entry->wanted, __entry->got) |
| 656 | 656 | ); |
| 657 | 657 | |
| 658 | +/* | |
| 659 | + * io_uring_local_work_run - ran ring local task work | |
| 660 | + * | |
| 661 | + * @tctx: pointer to a io_uring_ctx | |
| 662 | + * @count: how many functions it ran | |
| 663 | + * @loops: how many loops it ran | |
| 664 | + * | |
| 665 | + */ | |
| 666 | +TRACE_EVENT(io_uring_local_work_run, | |
| 667 | + | |
| 668 | + TP_PROTO(void *ctx, int count, unsigned int loops), | |
| 669 | + | |
| 670 | + TP_ARGS(ctx, count, loops), | |
| 671 | + | |
| 672 | + TP_STRUCT__entry ( | |
| 673 | + __field(void *, ctx ) | |
| 674 | + __field(int, count ) | |
| 675 | + __field(unsigned int, loops ) | |
| 676 | + ), | |
| 677 | + | |
| 678 | + TP_fast_assign( | |
| 679 | + __entry->ctx = ctx; | |
| 680 | + __entry->count = count; | |
| 681 | + __entry->loops = loops; | |
| 682 | + ), | |
| 683 | + | |
| 684 | + TP_printk("ring %p, count %d, loops %u", __entry->ctx, __entry->count, __entry->loops) | |
| 685 | +); | |
| 686 | + | |
| 658 | 687 | #endif /* _TRACE_IO_URING_H */ |
| 659 | 688 | |
| 660 | 689 | /* This part must be outside protection */ |
| @@ -157,6 +157,13 @@ enum { · 156 unmodified lines | ||
| 157 | 157 | */ |
| 158 | 158 | #define IORING_SETUP_SINGLE_ISSUER (1U << 12) |
| 159 | 159 | |
| 160 | +/* | |
| 161 | + * Defer running task work to get events. | |
| 162 | + * Rather than running bits of task work whenever the task transitions | |
| 163 | + * try to do it just before it is needed. | |
| 164 | + */ | |
| 165 | +#define IORING_SETUP_DEFER_TASKRUN (1U << 13) | |
| 166 | + | |
| 160 | 167 | enum io_uring_op { |
| 161 | 168 | IORING_OP_NOP, |
| 162 | 169 | IORING_OP_READV, |
| @@ -206,6 +213,7 @@ enum io_uring_op { · 43 unmodified lines | ||
| 206 | 213 | IORING_OP_SOCKET, |
| 207 | 214 | IORING_OP_URING_CMD, |
| 208 | 215 | IORING_OP_SEND_ZC, |
| 216 | + IORING_OP_SENDMSG_ZC, | |
| 209 | 217 | |
| 210 | 218 | /* this goes last, obviously */ |
| 211 | 219 | IORING_OP_LAST, |
| @@ -292,7 +292,7 @@ int io_sync_cancel(struct io_ring_ctx *ctx, void __user *arg) · 291 unmodified lines | ||
| 292 | 292 | break; |
| 293 | 293 | |
| 294 | 294 | mutex_unlock(&ctx->uring_lock); |
| 295 | − ret = io_run_task_work_sig(); | |
| 295 | + ret = io_run_task_work_sig(ctx); | |
| 296 | 296 | if (ret < 0) { |
| 297 | 297 | mutex_lock(&ctx->uring_lock); |
| 298 | 298 | break; |
| @@ -60,13 +60,15 @@ static __cold void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, · 59 unmodified lines | ||
| 60 | 60 | unsigned int cq_head = READ_ONCE(r->cq.head); |
| 61 | 61 | unsigned int cq_tail = READ_ONCE(r->cq.tail); |
| 62 | 62 | unsigned int cq_shift = 0; |
| 63 | + unsigned int sq_shift = 0; | |
| 63 | 64 | unsigned int sq_entries, cq_entries; |
| 64 | 65 | bool has_lock; |
| 65 | − bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
| 66 | 66 | unsigned int i; |
| 67 | 67 | |
| 68 | − if (is_cqe32) | |
| 68 | + if (ctx->flags & IORING_SETUP_CQE32) | |
| 69 | 69 | cq_shift = 1; |
| 70 | + if (ctx->flags & IORING_SETUP_SQE128) | |
| 71 | + sq_shift = 1; | |
| 70 | 72 | |
| 71 | 73 | /* |
| 72 | 74 | * we may get imprecise sqe and cqe info if uring is actively running |
| @@ -82,19 +84,36 @@ static __cold void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, · 9 unmodified lines | ||
| 82 | 84 | seq_printf(m, "CqHead:\t%u\n", cq_head); |
| 83 | 85 | seq_printf(m, "CqTail:\t%u\n", cq_tail); |
| 84 | 86 | seq_printf(m, "CachedCqTail:\t%u\n", ctx->cached_cq_tail); |
| 85 | − seq_printf(m, "SQEs:\t%u\n", sq_tail - ctx->cached_sq_head); | |
| 87 | + seq_printf(m, "SQEs:\t%u\n", sq_tail - sq_head); | |
| 86 | 88 | sq_entries = min(sq_tail - sq_head, ctx->sq_entries); |
| 87 | 89 | for (i = 0; i < sq_entries; i++) { |
| 88 | 90 | unsigned int entry = i + sq_head; |
| 89 | − unsigned int sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]); | |
| 90 | 91 | struct io_uring_sqe *sqe; |
| 92 | + unsigned int sq_idx; | |
| 91 | 93 | |
| 94 | + sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]); | |
| 92 | 95 | if (sq_idx > sq_mask) |
| 93 | 96 | continue; |
| 94 | − sqe = &ctx->sq_sqes[sq_idx]; | |
| 95 | − seq_printf(m, "%5u: opcode:%d, fd:%d, flags:%x, user_data:%llu\n", | |
| 96 | − sq_idx, sqe->opcode, sqe->fd, sqe->flags, | |
| 97 | − sqe->user_data); | |
| 97 | + sqe = &ctx->sq_sqes[sq_idx << 1]; | |
| 98 | + seq_printf(m, "%5u: opcode:%s, fd:%d, flags:%x, off:%llu, " | |
| 99 | + "addr:0x%llx, rw_flags:0x%x, buf_index:%d " | |
| 100 | + "user_data:%llu", | |
| 101 | + sq_idx, io_uring_get_opcode(sqe->opcode), sqe->fd, | |
| 102 | + sqe->flags, (unsigned long long) sqe->off, | |
| 103 | + (unsigned long long) sqe->addr, sqe->rw_flags, | |
| 104 | + sqe->buf_index, sqe->user_data); | |
| 105 | + if (sq_shift) { | |
| 106 | + u64 *sqeb = (void *) (sqe + 1); | |
| 107 | + int size = sizeof(struct io_uring_sqe) / sizeof(u64); | |
| 108 | + int j; | |
| 109 | + | |
| 110 | + for (j = 0; j < size; j++) { | |
| 111 | + seq_printf(m, ", e%d:0x%llx", j, | |
| 112 | + (unsigned long long) *sqeb); | |
| 113 | + sqeb++; | |
| 114 | + } | |
| 115 | + } | |
| 116 | + seq_printf(m, "\n"); | |
| 98 | 117 | } |
| 99 | 118 | seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head); |
| 100 | 119 | cq_entries = min(cq_tail - cq_head, ctx->cq_entries); |
| @@ -102,16 +121,13 @@ static __cold void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, · 1 unmodified line | ||
| 102 | 121 | unsigned int entry = i + cq_head; |
| 103 | 122 | struct io_uring_cqe *cqe = &r->cqes[(entry & cq_mask) << cq_shift]; |
| 104 | 123 | |
| 105 | − if (!is_cqe32) { | |
| 106 | − seq_printf(m, "%5u: user_data:%llu, res:%d, flag:%x\n", | |
| 124 | + seq_printf(m, "%5u: user_data:%llu, res:%d, flag:%x", | |
| 107 | 125 | entry & cq_mask, cqe->user_data, cqe->res, |
| 108 | 126 | cqe->flags); |
| 109 | − } else { | |
| 110 | − seq_printf(m, "%5u: user_data:%llu, res:%d, flag:%x, " | |
| 111 | − "extra1:%llu, extra2:%llu\n", | |
| 112 | − entry & cq_mask, cqe->user_data, cqe->res, | |
| 113 | − cqe->flags, cqe->big_cqe[0], cqe->big_cqe[1]); | |
| 114 | − } | |
| 127 | + if (cq_shift) | |
| 128 | + seq_printf(m, ", extra1:%llu, extra2:%llu\n", | |
| 129 | + cqe->big_cqe[0], cqe->big_cqe[1]); | |
| 130 | + seq_printf(m, "\n"); | |
| 115 | 131 | } |
| 116 | 132 | |
| 117 | 133 | /* |
| @@ -125,6 +125,11 @@ enum { · 124 unmodified lines | ||
| 125 | 125 | IO_CHECK_CQ_DROPPED_BIT, |
| 126 | 126 | }; |
| 127 | 127 | |
| 128 | +enum { | |
| 129 | + IO_EVENTFD_OP_SIGNAL_BIT, | |
| 130 | + IO_EVENTFD_OP_FREE_BIT, | |
| 131 | +}; | |
| 132 | + | |
| 128 | 133 | struct io_defer_entry { |
| 129 | 134 | struct list_head list; |
| 130 | 135 | struct io_kiocb *req; |
| @@ -142,7 +147,7 @@ static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, · 11 unmodified lines | ||
| 142 | 147 | static void io_dismantle_req(struct io_kiocb *req); |
| 143 | 148 | static void io_clean_op(struct io_kiocb *req); |
| 144 | 149 | static void io_queue_sqe(struct io_kiocb *req); |
| 145 | − | |
| 150 | +static void io_move_task_work_from_local(struct io_ring_ctx *ctx); | |
| 146 | 151 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
| 147 | 152 | |
| 148 | 153 | static struct kmem_cache *req_cachep; |
| @@ -316,6 +321,7 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) · 167 unmodified lines | ||
| 316 | 321 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
| 317 | 322 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
| 318 | 323 | init_llist_head(&ctx->rsrc_put_llist); |
| 324 | + init_llist_head(&ctx->work_llist); | |
| 319 | 325 | INIT_LIST_HEAD(&ctx->tctx_list); |
| 320 | 326 | ctx->submit_state.free_list.next = NULL; |
| 321 | 327 | INIT_WQ_LIST(&ctx->locked_free_list); |
| @@ -477,25 +483,28 @@ static __cold void io_queue_deferred(struct io_ring_ctx *ctx) · 155 unmodified lines | ||
| 477 | 483 | } |
| 478 | 484 | } |
| 479 | 485 | |
| 480 | −static void io_eventfd_signal(struct io_ring_ctx *ctx) | |
| 486 | + | |
| 487 | +static void io_eventfd_ops(struct rcu_head *rcu) | |
| 481 | 488 | { |
| 482 | − struct io_ev_fd *ev_fd; | |
| 483 | − bool skip; | |
| 489 | + struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); | |
| 490 | + int ops = atomic_xchg(&ev_fd->ops, 0); | |
| 484 | 491 | |
| 485 | − spin_lock(&ctx->completion_lock); | |
| 486 | − /* | |
| 487 | − * Eventfd should only get triggered when at least one event has been | |
| 488 | − * posted. Some applications rely on the eventfd notification count only | |
| 489 | − * changing IFF a new CQE has been added to the CQ ring. There's no | |
| 490 | − * depedency on 1:1 relationship between how many times this function is | |
| 491 | − * called (and hence the eventfd count) and number of CQEs posted to the | |
| 492 | − * CQ ring. | |
| 492 | + if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) | |
| 493 | + eventfd_signal(ev_fd->cq_ev_fd, 1); | |
| 494 | + | |
| 495 | + /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback | |
| 496 | + * ordering in a race but if references are 0 we know we have to free | |
| 497 | + * it regardless. | |
| 493 | 498 | */ |
| 494 | − skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
| 495 | − ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
| 496 | − spin_unlock(&ctx->completion_lock); | |
| 497 | − if (skip) | |
| 498 | − return; | |
| 499 | + if (atomic_dec_and_test(&ev_fd->refs)) { | |
| 500 | + eventfd_ctx_put(ev_fd->cq_ev_fd); | |
| 501 | + kfree(ev_fd); | |
| 502 | + } | |
| 503 | +} | |
| 504 | + | |
| 505 | +static void io_eventfd_signal(struct io_ring_ctx *ctx) | |
| 506 | +{ | |
| 507 | + struct io_ev_fd *ev_fd = NULL; | |
| 499 | 508 | |
| 500 | 509 | rcu_read_lock(); |
| 501 | 510 | /* |
| @@ -513,13 +522,46 @@ static void io_eventfd_signal(struct io_ring_ctx *ctx) · 11 unmodified lines | ||
| 513 | 522 | goto out; |
| 514 | 523 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
| 515 | 524 | goto out; |
| 525 | + if (ev_fd->eventfd_async && !io_wq_current_is_worker()) | |
| 526 | + goto out; | |
| 516 | 527 | |
| 517 | − if (!ev_fd->eventfd_async || io_wq_current_is_worker()) | |
| 528 | + if (likely(eventfd_signal_allowed())) { | |
| 518 | 529 | eventfd_signal(ev_fd->cq_ev_fd, 1); |
| 530 | + } else { | |
| 531 | + atomic_inc(&ev_fd->refs); | |
| 532 | + if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
| 533 | + call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
| 534 | + else | |
| 535 | + atomic_dec(&ev_fd->refs); | |
| 536 | + } | |
| 537 | + | |
| 519 | 538 | out: |
| 520 | 539 | rcu_read_unlock(); |
| 521 | 540 | } |
| 522 | 541 | |
| 542 | +static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) | |
| 543 | +{ | |
| 544 | + bool skip; | |
| 545 | + | |
| 546 | + spin_lock(&ctx->completion_lock); | |
| 547 | + | |
| 548 | + /* | |
| 549 | + * Eventfd should only get triggered when at least one event has been | |
| 550 | + * posted. Some applications rely on the eventfd notification count | |
| 551 | + * only changing IFF a new CQE has been added to the CQ ring. There's | |
| 552 | + * no depedency on 1:1 relationship between how many times this | |
| 553 | + * function is called (and hence the eventfd count) and number of CQEs | |
| 554 | + * posted to the CQ ring. | |
| 555 | + */ | |
| 556 | + skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
| 557 | + ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
| 558 | + spin_unlock(&ctx->completion_lock); | |
| 559 | + if (skip) | |
| 560 | + return; | |
| 561 | + | |
| 562 | + io_eventfd_signal(ctx); | |
| 563 | +} | |
| 564 | + | |
| 523 | 565 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
| 524 | 566 | { |
| 525 | 567 | if (ctx->off_timeout_used || ctx->drain_active) { |
| @@ -531,7 +573,7 @@ void __io_commit_cqring_flush(struct io_ring_ctx *ctx) · 5 unmodified lines | ||
| 531 | 573 | spin_unlock(&ctx->completion_lock); |
| 532 | 574 | } |
| 533 | 575 | if (ctx->has_evfd) |
| 534 | − io_eventfd_signal(ctx); | |
| 576 | + io_eventfd_flush_signal(ctx); | |
| 535 | 577 | } |
| 536 | 578 | |
| 537 | 579 | static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
| @@ -567,7 +609,7 @@ static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) · 29 unmodified lines | ||
| 567 | 609 | |
| 568 | 610 | io_cq_lock(ctx); |
| 569 | 611 | while (!list_empty(&ctx->cq_overflow_list)) { |
| 570 | − struct io_uring_cqe *cqe = io_get_cqe(ctx); | |
| 612 | + struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); | |
| 571 | 613 | struct io_overflow_cqe *ocqe; |
| 572 | 614 | |
| 573 | 615 | if (!cqe && !force) |
| @@ -694,12 +736,19 @@ bool io_req_cqe_overflow(struct io_kiocb *req) · 120 unmodified lines | ||
| 694 | 736 | * control dependency is enough as we're using WRITE_ONCE to |
| 695 | 737 | * fill the cq entry |
| 696 | 738 | */ |
| 697 | −struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx) | |
| 739 | +struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) | |
| 698 | 740 | { |
| 699 | 741 | struct io_rings *rings = ctx->rings; |
| 700 | 742 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); |
| 701 | 743 | unsigned int free, queued, len; |
| 702 | 744 | |
| 745 | + /* | |
| 746 | + * Posting into the CQ when there are pending overflowed CQEs may break | |
| 747 | + * ordering guarantees, which will affect links, F_MORE users and more. | |
| 748 | + * Force overflow the completion. | |
| 749 | + */ | |
| 750 | + if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
| 751 | + return NULL; | |
| 703 | 752 | |
| 704 | 753 | /* userspace may cheat modifying the tail, be safe and do min */ |
| 705 | 754 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); |
| @@ -823,8 +872,12 @@ inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags) · 117 unmodified lines | ||
| 823 | 872 | |
| 824 | 873 | void io_req_complete_failed(struct io_kiocb *req, s32 res) |
| 825 | 874 | { |
| 875 | + const struct io_op_def *def = &io_op_defs[req->opcode]; | |
| 876 | + | |
| 826 | 877 | req_set_fail(req); |
| 827 | 878 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
| 879 | + if (def->fail) | |
| 880 | + def->fail(req); | |
| 828 | 881 | io_req_complete_post(req); |
| 829 | 882 | } |
| 830 | 883 | |
| @@ -1047,17 +1100,40 @@ void tctx_task_work(struct callback_head *cb) · 216 unmodified lines | ||
| 1047 | 1100 | trace_io_uring_task_work_run(tctx, count, loops); |
| 1048 | 1101 | } |
| 1049 | 1102 | |
| 1050 | −void io_req_task_work_add(struct io_kiocb *req) | |
| 1103 | +static void io_req_local_work_add(struct io_kiocb *req) | |
| 1104 | +{ | |
| 1105 | + struct io_ring_ctx *ctx = req->ctx; | |
| 1106 | + | |
| 1107 | + if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) | |
| 1108 | + return; | |
| 1109 | + | |
| 1110 | + if (unlikely(atomic_read(&req->task->io_uring->in_idle))) { | |
| 1111 | + io_move_task_work_from_local(ctx); | |
| 1112 | + return; | |
| 1113 | + } | |
| 1114 | + | |
| 1115 | + if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
| 1116 | + atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
| 1117 | + | |
| 1118 | + if (ctx->has_evfd) | |
| 1119 | + io_eventfd_signal(ctx); | |
| 1120 | + io_cqring_wake(ctx); | |
| 1121 | + | |
| 1122 | +} | |
| 1123 | + | |
| 1124 | +static inline void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) | |
| 1051 | 1125 | { |
| 1052 | 1126 | struct io_uring_task *tctx = req->task->io_uring; |
| 1053 | 1127 | struct io_ring_ctx *ctx = req->ctx; |
| 1054 | 1128 | struct llist_node *node; |
| 1055 | − bool running; | |
| 1056 | 1129 | |
| 1057 | − running = !llist_add(&req->io_task_work.node, &tctx->task_list); | |
| 1130 | + if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
| 1131 | + io_req_local_work_add(req); | |
| 1132 | + return; | |
| 1133 | + } | |
| 1058 | 1134 | |
| 1059 | 1135 | /* task_work already pending, we're done */ |
| 1060 | − if (running) | |
| 1136 | + if (!llist_add(&req->io_task_work.node, &tctx->task_list)) | |
| 1061 | 1137 | return; |
| 1062 | 1138 | |
| 1063 | 1139 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
| @@ -1077,6 +1153,84 @@ void io_req_task_work_add(struct io_kiocb *req) · 13 unmodified lines | ||
| 1077 | 1153 | } |
| 1078 | 1154 | } |
| 1079 | 1155 | |
| 1156 | +void io_req_task_work_add(struct io_kiocb *req) | |
| 1157 | +{ | |
| 1158 | + __io_req_task_work_add(req, true); | |
| 1159 | +} | |
| 1160 | + | |
| 1161 | +static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) | |
| 1162 | +{ | |
| 1163 | + struct llist_node *node; | |
| 1164 | + | |
| 1165 | + node = llist_del_all(&ctx->work_llist); | |
| 1166 | + while (node) { | |
| 1167 | + struct io_kiocb *req = container_of(node, struct io_kiocb, | |
| 1168 | + io_task_work.node); | |
| 1169 | + | |
| 1170 | + node = node->next; | |
| 1171 | + __io_req_task_work_add(req, false); | |
| 1172 | + } | |
| 1173 | +} | |
| 1174 | + | |
| 1175 | +int __io_run_local_work(struct io_ring_ctx *ctx, bool locked) | |
| 1176 | +{ | |
| 1177 | + struct llist_node *node; | |
| 1178 | + struct llist_node fake; | |
| 1179 | + struct llist_node *current_final = NULL; | |
| 1180 | + int ret; | |
| 1181 | + unsigned int loops = 1; | |
| 1182 | + | |
| 1183 | + if (unlikely(ctx->submitter_task != current)) | |
| 1184 | + return -EEXIST; | |
| 1185 | + | |
| 1186 | + node = io_llist_xchg(&ctx->work_llist, &fake); | |
| 1187 | + ret = 0; | |
| 1188 | +again: | |
| 1189 | + while (node != current_final) { | |
| 1190 | + struct llist_node *next = node->next; | |
| 1191 | + struct io_kiocb *req = container_of(node, struct io_kiocb, | |
| 1192 | + io_task_work.node); | |
| 1193 | + prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
| 1194 | + req->io_task_work.func(req, &locked); | |
| 1195 | + ret++; | |
| 1196 | + node = next; | |
| 1197 | + } | |
| 1198 | + | |
| 1199 | + if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
| 1200 | + atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
| 1201 | + | |
| 1202 | + node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL); | |
| 1203 | + if (node != &fake) { | |
| 1204 | + loops++; | |
| 1205 | + current_final = &fake; | |
| 1206 | + node = io_llist_xchg(&ctx->work_llist, &fake); | |
| 1207 | + goto again; | |
| 1208 | + } | |
| 1209 | + | |
| 1210 | + if (locked) | |
| 1211 | + io_submit_flush_completions(ctx); | |
| 1212 | + trace_io_uring_local_work_run(ctx, ret, loops); | |
| 1213 | + return ret; | |
| 1214 | + | |
| 1215 | +} | |
| 1216 | + | |
| 1217 | +int io_run_local_work(struct io_ring_ctx *ctx) | |
| 1218 | +{ | |
| 1219 | + bool locked; | |
| 1220 | + int ret; | |
| 1221 | + | |
| 1222 | + if (llist_empty(&ctx->work_llist)) | |
| 1223 | + return 0; | |
| 1224 | + | |
| 1225 | + __set_current_state(TASK_RUNNING); | |
| 1226 | + locked = mutex_trylock(&ctx->uring_lock); | |
| 1227 | + ret = __io_run_local_work(ctx, locked); | |
| 1228 | + if (locked) | |
| 1229 | + mutex_unlock(&ctx->uring_lock); | |
| 1230 | + | |
| 1231 | + return ret; | |
| 1232 | +} | |
| 1233 | + | |
| 1080 | 1234 | static void io_req_tw_post(struct io_kiocb *req, bool *locked) |
| 1081 | 1235 | { |
| 1082 | 1236 | io_req_complete_post(req); |
| @@ -1183,7 +1337,7 @@ static void __io_submit_flush_completions(struct io_ring_ctx *ctx) · 100 unmodified lines | ||
| 1183 | 1337 | struct io_wq_work_node *node, *prev; |
| 1184 | 1338 | struct io_submit_state *state = &ctx->submit_state; |
| 1185 | 1339 | |
| 1186 | − spin_lock(&ctx->completion_lock); | |
| 1340 | + io_cq_lock(ctx); | |
| 1187 | 1341 | wq_list_for_each(node, prev, &state->compl_reqs) { |
| 1188 | 1342 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
| 1189 | 1343 | comp_list); |
| @@ -1254,6 +1408,9 @@ static int io_iopoll_check(struct io_ring_ctx *ctx, long min) · 64 unmodified lines | ||
| 1254 | 1408 | int ret = 0; |
| 1255 | 1409 | unsigned long check_cq; |
| 1256 | 1410 | |
| 1411 | + if (!io_allowed_run_tw(ctx)) | |
| 1412 | + return -EEXIST; | |
| 1413 | + | |
| 1257 | 1414 | check_cq = READ_ONCE(ctx->check_cq); |
| 1258 | 1415 | if (unlikely(check_cq)) { |
| 1259 | 1416 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
| @@ -1284,13 +1441,19 @@ static int io_iopoll_check(struct io_ring_ctx *ctx, long min) · 24 unmodified lines | ||
| 1284 | 1441 | * forever, while the workqueue is stuck trying to acquire the |
| 1285 | 1442 | * very same mutex. |
| 1286 | 1443 | */ |
| 1287 | − if (wq_list_empty(&ctx->iopoll_list)) { | |
| 1444 | + if (wq_list_empty(&ctx->iopoll_list) || | |
| 1445 | + io_task_work_pending(ctx)) { | |
| 1288 | 1446 | u32 tail = ctx->cached_cq_tail; |
| 1289 | 1447 | |
| 1290 | − mutex_unlock(&ctx->uring_lock); | |
| 1291 | − io_run_task_work(); | |
| 1292 | − mutex_lock(&ctx->uring_lock); | |
| 1448 | + if (!llist_empty(&ctx->work_llist)) | |
| 1449 | + __io_run_local_work(ctx, true); | |
| 1293 | 1450 | |
| 1451 | + if (task_work_pending(current) || | |
| 1452 | + wq_list_empty(&ctx->iopoll_list)) { | |
| 1453 | + mutex_unlock(&ctx->uring_lock); | |
| 1454 | + io_run_task_work(); | |
| 1455 | + mutex_lock(&ctx->uring_lock); | |
| 1456 | + } | |
| 1294 | 1457 | /* some requests don't go through iopoll_list */ |
| 1295 | 1458 | if (tail != ctx->cached_cq_tail || |
| 1296 | 1459 | wq_list_empty(&ctx->iopoll_list)) |
| @@ -1732,10 +1895,6 @@ static void io_queue_async(struct io_kiocb *req, int ret) · 435 unmodified lines | ||
| 1732 | 1895 | io_req_task_queue(req); |
| 1733 | 1896 | break; |
| 1734 | 1897 | case IO_APOLL_ABORTED: |
| 1735 | − /* | |
| 1736 | − * Queued up for async execution, worker will release | |
| 1737 | − * submit reference when the iocb is actually submitted. | |
| 1738 | − */ | |
| 1739 | 1898 | io_kbuf_recycle(req, 0); |
| 1740 | 1899 | io_queue_iowq(req, NULL); |
| 1741 | 1900 | break; |
| @@ -2149,6 +2308,13 @@ struct io_wait_queue { · 407 unmodified lines | ||
| 2149 | 2308 | unsigned nr_timeouts; |
| 2150 | 2309 | }; |
| 2151 | 2310 | |
| 2311 | +static inline bool io_has_work(struct io_ring_ctx *ctx) | |
| 2312 | +{ | |
| 2313 | + return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || | |
| 2314 | + ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && | |
| 2315 | + !llist_empty(&ctx->work_llist)); | |
| 2316 | +} | |
| 2317 | + | |
| 2152 | 2318 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
| 2153 | 2319 | { |
| 2154 | 2320 | struct io_ring_ctx *ctx = iowq->ctx; |
| @@ -2167,20 +2333,20 @@ static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, · 12 unmodified lines | ||
| 2167 | 2333 | { |
| 2168 | 2334 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, |
| 2169 | 2335 | wq); |
| 2336 | + struct io_ring_ctx *ctx = iowq->ctx; | |
| 2170 | 2337 | |
| 2171 | 2338 | /* |
| 2172 | 2339 | * Cannot safely flush overflowed CQEs from here, ensure we wake up |
| 2173 | 2340 | * the task, and the next invocation will do it. |
| 2174 | 2341 | */ |
| 2175 | − if (io_should_wake(iowq) || | |
| 2176 | − test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq)) | |
| 2342 | + if (io_should_wake(iowq) || io_has_work(ctx)) | |
| 2177 | 2343 | return autoremove_wake_function(curr, mode, wake_flags, key); |
| 2178 | 2344 | return -1; |
| 2179 | 2345 | } |
| 2180 | 2346 | |
| 2181 | −int io_run_task_work_sig(void) | |
| 2347 | +int io_run_task_work_sig(struct io_ring_ctx *ctx) | |
| 2182 | 2348 | { |
| 2183 | − if (io_run_task_work()) | |
| 2349 | + if (io_run_task_work_ctx(ctx) > 0) | |
| 2184 | 2350 | return 1; |
| 2185 | 2351 | if (task_sigpending(current)) |
| 2186 | 2352 | return -EINTR; |
| @@ -2196,7 +2362,7 @@ static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, · 9 unmodified lines | ||
| 2196 | 2362 | unsigned long check_cq; |
| 2197 | 2363 | |
| 2198 | 2364 | /* make sure we run task_work before checking for signals */ |
| 2199 | − ret = io_run_task_work_sig(); | |
| 2365 | + ret = io_run_task_work_sig(ctx); | |
| 2200 | 2366 | if (ret || io_should_wake(iowq)) |
| 2201 | 2367 | return ret; |
| 2202 | 2368 | |
| @@ -2226,13 +2392,19 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, · 23 unmodified lines | ||
| 2226 | 2392 | ktime_t timeout = KTIME_MAX; |
| 2227 | 2393 | int ret; |
| 2228 | 2394 | |
| 2395 | + if (!io_allowed_run_tw(ctx)) | |
| 2396 | + return -EEXIST; | |
| 2397 | + | |
| 2229 | 2398 | do { |
| 2399 | + /* always run at least 1 task work to process local work */ | |
| 2400 | + ret = io_run_task_work_ctx(ctx); | |
| 2401 | + if (ret < 0) | |
| 2402 | + return ret; | |
| 2230 | 2403 | io_cqring_overflow_flush(ctx); |
| 2404 | + | |
| 2231 | 2405 | if (io_cqring_events(ctx) >= min_events) |
| 2232 | 2406 | return 0; |
| 2233 | − if (!io_run_task_work()) | |
| 2234 | − break; | |
| 2235 | − } while (1); | |
| 2407 | + } while (ret > 0); | |
| 2236 | 2408 | |
| 2237 | 2409 | if (sig) { |
| 2238 | 2410 | #ifdef CONFIG_COMPAT |
| @@ -2366,17 +2538,11 @@ static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, · 127 unmodified lines | ||
| 2366 | 2538 | ev_fd->eventfd_async = eventfd_async; |
| 2367 | 2539 | ctx->has_evfd = true; |
| 2368 | 2540 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); |
| 2541 | + atomic_set(&ev_fd->refs, 1); | |
| 2542 | + atomic_set(&ev_fd->ops, 0); | |
| 2369 | 2543 | return 0; |
| 2370 | 2544 | } |
| 2371 | 2545 | |
| 2372 | −static void io_eventfd_put(struct rcu_head *rcu) | |
| 2373 | −{ | |
| 2374 | − struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); | |
| 2375 | − | |
| 2376 | − eventfd_ctx_put(ev_fd->cq_ev_fd); | |
| 2377 | − kfree(ev_fd); | |
| 2378 | −} | |
| 2379 | − | |
| 2380 | 2546 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
| 2381 | 2547 | { |
| 2382 | 2548 | struct io_ev_fd *ev_fd; |
| @@ -2386,7 +2552,8 @@ static int io_eventfd_unregister(struct io_ring_ctx *ctx) · 3 unmodified lines | ||
| 2386 | 2552 | if (ev_fd) { |
| 2387 | 2553 | ctx->has_evfd = false; |
| 2388 | 2554 | rcu_assign_pointer(ctx->io_ev_fd, NULL); |
| 2389 | − call_rcu(&ev_fd->rcu, io_eventfd_put); | |
| 2555 | + if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) | |
| 2556 | + call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
| 2390 | 2557 | return 0; |
| 2391 | 2558 | } |
| 2392 | 2559 | |
| @@ -2509,8 +2676,8 @@ static __poll_t io_uring_poll(struct file *file, poll_table *wait) · 116 unmodified lines | ||
| 2509 | 2676 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this |
| 2510 | 2677 | * pushs them to do the flush. |
| 2511 | 2678 | */ |
| 2512 | − if (io_cqring_events(ctx) || | |
| 2513 | − test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
| 2679 | + | |
| 2680 | + if (io_cqring_events(ctx) || io_has_work(ctx)) | |
| 2514 | 2681 | mask |= EPOLLIN | EPOLLRDNORM; |
| 2515 | 2682 | |
| 2516 | 2683 | return mask; |
| @@ -2573,6 +2740,9 @@ static __cold void io_ring_exit_work(struct work_struct *work) · 56 unmodified lines | ||
| 2573 | 2740 | * as nobody else will be looking for them. |
| 2574 | 2741 | */ |
| 2575 | 2742 | do { |
| 2743 | + if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
| 2744 | + io_move_task_work_from_local(ctx); | |
| 2745 | + | |
| 2576 | 2746 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
| 2577 | 2747 | cond_resched(); |
| 2578 | 2748 | |
| @@ -2770,13 +2940,15 @@ static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, · 191 unmodified lines | ||
| 2770 | 2940 | } |
| 2771 | 2941 | } |
| 2772 | 2942 | |
| 2943 | + if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
| 2944 | + ret |= io_run_local_work(ctx) > 0; | |
| 2773 | 2945 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
| 2774 | 2946 | mutex_lock(&ctx->uring_lock); |
| 2775 | 2947 | ret |= io_poll_remove_all(ctx, task, cancel_all); |
| 2776 | 2948 | mutex_unlock(&ctx->uring_lock); |
| 2777 | 2949 | ret |= io_kill_timeouts(ctx, task, cancel_all); |
| 2778 | 2950 | if (task) |
| 2779 | − ret |= io_run_task_work(); | |
| 2951 | + ret |= io_run_task_work() > 0; | |
| 2780 | 2952 | return ret; |
| 2781 | 2953 | } |
| 2782 | 2954 | |
| @@ -2992,8 +3164,6 @@ SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, · 209 unmodified lines | ||
| 2992 | 3164 | struct fd f; |
| 2993 | 3165 | long ret; |
| 2994 | 3166 | |
| 2995 | − io_run_task_work(); | |
| 2996 | − | |
| 2997 | 3167 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
| 2998 | 3168 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
| 2999 | 3169 | IORING_ENTER_REGISTERED_RING))) |
| @@ -3063,8 +3233,10 @@ SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, · 63 unmodified lines | ||
| 3063 | 3233 | goto iopoll_locked; |
| 3064 | 3234 | mutex_unlock(&ctx->uring_lock); |
| 3065 | 3235 | } |
| 3236 | + | |
| 3066 | 3237 | if (flags & IORING_ENTER_GETEVENTS) { |
| 3067 | 3238 | int ret2; |
| 3239 | + | |
| 3068 | 3240 | if (ctx->syscall_iopoll) { |
| 3069 | 3241 | /* |
| 3070 | 3242 | * We disallow the app entering submit/complete with |
| @@ -3293,17 +3465,29 @@ static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, · 222 unmodified lines | ||
| 3293 | 3465 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
| 3294 | 3466 | /* IPI related flags don't make sense with SQPOLL */ |
| 3295 | 3467 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
| 3296 | − IORING_SETUP_TASKRUN_FLAG)) | |
| 3468 | + IORING_SETUP_TASKRUN_FLAG | | |
| 3469 | + IORING_SETUP_DEFER_TASKRUN)) | |
| 3297 | 3470 | goto err; |
| 3298 | 3471 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
| 3299 | 3472 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
| 3300 | 3473 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
| 3301 | 3474 | } else { |
| 3302 | − if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
| 3475 | + if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && | |
| 3476 | + !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
| 3303 | 3477 | goto err; |
| 3304 | 3478 | ctx->notify_method = TWA_SIGNAL; |
| 3305 | 3479 | } |
| 3306 | 3480 | |
| 3481 | + /* | |
| 3482 | + * For DEFER_TASKRUN we require the completion task to be the same as the | |
| 3483 | + * submission task. This implies that there is only one submitter, so enforce | |
| 3484 | + * that. | |
| 3485 | + */ | |
| 3486 | + if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
| 3487 | + !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
| 3488 | + goto err; | |
| 3489 | + } | |
| 3490 | + | |
| 3307 | 3491 | /* |
| 3308 | 3492 | * This is just grabbed for accounting purposes. When a process exits, |
| 3309 | 3493 | * the mm is exited and dropped before the files, hence we need to hang |
| @@ -3408,7 +3592,7 @@ static long io_uring_setup(u32 entries, struct io_uring_params __user *params) · 98 unmodified lines | ||
| 3408 | 3592 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
| 3409 | 3593 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
| 3410 | 3594 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
| 3411 | − IORING_SETUP_SINGLE_ISSUER)) | |
| 3595 | + IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) | |
| 3412 | 3596 | return -EINVAL; |
| 3413 | 3597 | |
| 3414 | 3598 | return io_uring_create(entries, &p, params); |
| @@ -3874,7 +4058,7 @@ SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, · 459 unmodified lines | ||
| 3874 | 4058 | |
| 3875 | 4059 | ctx = f.file->private_data; |
| 3876 | 4060 | |
| 3877 | − io_run_task_work(); | |
| 4061 | + io_run_task_work_ctx(ctx); | |
| 3878 | 4062 | |
| 3879 | 4063 | mutex_lock(&ctx->uring_lock); |
| 3880 | 4064 | ret = __io_uring_register(ctx, opcode, arg, nr_args); |
| @@ -24,9 +24,11 @@ enum { · 23 unmodified lines | ||
| 24 | 24 | IOU_STOP_MULTISHOT = -ECANCELED, |
| 25 | 25 | }; |
| 26 | 26 | |
| 27 | −struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx); | |
| 27 | +struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow); | |
| 28 | 28 | bool io_req_cqe_overflow(struct io_kiocb *req); |
| 29 | −int io_run_task_work_sig(void); | |
| 29 | +int io_run_task_work_sig(struct io_ring_ctx *ctx); | |
| 30 | +int __io_run_local_work(struct io_ring_ctx *ctx, bool locked); | |
| 31 | +int io_run_local_work(struct io_ring_ctx *ctx); | |
| 30 | 32 | void io_req_complete_failed(struct io_kiocb *req, s32 res); |
| 31 | 33 | void __io_req_complete(struct io_kiocb *req, unsigned issue_flags); |
| 32 | 34 | void io_req_complete_post(struct io_kiocb *req); |
| @@ -91,7 +93,8 @@ static inline void io_cq_lock(struct io_ring_ctx *ctx) · 58 unmodified lines | ||
| 91 | 93 | |
| 92 | 94 | void io_cq_unlock_post(struct io_ring_ctx *ctx); |
| 93 | 95 | |
| 94 | −static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) | |
| 96 | +static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx, | |
| 97 | + bool overflow) | |
| 95 | 98 | { |
| 96 | 99 | if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) { |
| 97 | 100 | struct io_uring_cqe *cqe = ctx->cqe_cached; |
| @@ -103,7 +106,12 @@ static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) · 5 unmodified lines | ||
| 103 | 106 | return cqe; |
| 104 | 107 | } |
| 105 | 108 | |
| 106 | − return __io_get_cqe(ctx); | |
| 109 | + return __io_get_cqe(ctx, overflow); | |
| 110 | +} | |
| 111 | + | |
| 112 | +static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) | |
| 113 | +{ | |
| 114 | + return io_get_cqe_overflow(ctx, false); | |
| 107 | 115 | } |
| 108 | 116 | |
| 109 | 117 | static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx, |
| @@ -221,17 +229,43 @@ static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) · 111 unmodified lines | ||
| 221 | 229 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; |
| 222 | 230 | } |
| 223 | 231 | |
| 224 | −static inline bool io_run_task_work(void) | |
| 232 | +static inline int io_run_task_work(void) | |
| 225 | 233 | { |
| 226 | − if (test_thread_flag(TIF_NOTIFY_SIGNAL)) { | |
| 234 | + if (task_work_pending(current)) { | |
| 235 | + if (test_thread_flag(TIF_NOTIFY_SIGNAL)) | |
| 236 | + clear_notify_signal(); | |
| 227 | 237 | __set_current_state(TASK_RUNNING); |
| 228 | − clear_notify_signal(); | |
| 229 | − if (task_work_pending(current)) | |
| 230 | − task_work_run(); | |
| 231 | − return true; | |
| 238 | + task_work_run(); | |
| 239 | + return 1; | |
| 232 | 240 | } |
| 233 | 241 | |
| 234 | − return false; | |
| 242 | + return 0; | |
| 243 | +} | |
| 244 | + | |
| 245 | +static inline bool io_task_work_pending(struct io_ring_ctx *ctx) | |
| 246 | +{ | |
| 247 | + return test_thread_flag(TIF_NOTIFY_SIGNAL) || | |
| 248 | + !wq_list_empty(&ctx->work_llist); | |
| 249 | +} | |
| 250 | + | |
| 251 | +static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx) | |
| 252 | +{ | |
| 253 | + int ret = 0; | |
| 254 | + int ret2; | |
| 255 | + | |
| 256 | + if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
| 257 | + ret = io_run_local_work(ctx); | |
| 258 | + | |
| 259 | + /* want to run this after in case more is added */ | |
| 260 | + ret2 = io_run_task_work(); | |
| 261 | + | |
| 262 | + /* Try propagate error in favour of if tasks were run, | |
| 263 | + * but still make sure to run them if requested | |
| 264 | + */ | |
| 265 | + if (ret >= 0) | |
| 266 | + ret += ret2; | |
| 267 | + | |
| 268 | + return ret; | |
| 235 | 269 | } |
| 236 | 270 | |
| 237 | 271 | static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked) |
| @@ -301,4 +335,10 @@ static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) · 63 unmodified lines | ||
| 301 | 335 | return container_of(node, struct io_kiocb, comp_list); |
| 302 | 336 | } |
| 303 | 337 | |
| 338 | +static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx) | |
| 339 | +{ | |
| 340 | + return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) || | |
| 341 | + ctx->submitter_task == current); | |
| 342 | +} | |
| 343 | + | |
| 304 | 344 | #endif |
| @@ -86,18 +86,6 @@ static inline bool io_do_buffer_select(struct io_kiocb *req) · 85 unmodified lines | ||
| 86 | 86 | |
| 87 | 87 | static inline void io_kbuf_recycle(struct io_kiocb *req, unsigned issue_flags) |
| 88 | 88 | { |
| 89 | − /* | |
| 90 | − * READV uses fields in `struct io_rw` (len/addr) to stash the selected | |
| 91 | − * buffer data. However if that buffer is recycled the original request | |
| 92 | − * data stored in addr is lost. Therefore forbid recycling for now. | |
| 93 | − */ | |
| 94 | − if (req->opcode == IORING_OP_READV) { | |
| 95 | − if ((req->flags & REQ_F_BUFFER_RING) && req->buf_list) { | |
| 96 | − req->buf_list->head++; | |
| 97 | − req->buf_list = NULL; | |
| 98 | − } | |
| 99 | − return; | |
| 100 | − } | |
| 101 | 89 | if (req->flags & REQ_F_BUFFER_SELECTED) |
| 102 | 90 | io_kbuf_recycle_legacy(req, issue_flags); |
| 103 | 91 | if (req->flags & REQ_F_BUFFER_RING) |
| @@ -55,21 +55,14 @@ struct io_sr_msg { · 54 unmodified lines | ||
| 55 | 55 | struct user_msghdr __user *umsg; |
| 56 | 56 | void __user *buf; |
| 57 | 57 | }; |
| 58 | + unsigned len; | |
| 59 | + unsigned done_io; | |
| 58 | 60 | unsigned msg_flags; |
| 59 | − unsigned flags; | |
| 60 | − size_t len; | |
| 61 | − size_t done_io; | |
| 62 | −}; | |
| 63 | − | |
| 64 | −struct io_sendzc { | |
| 65 | − struct file *file; | |
| 66 | − void __user *buf; | |
| 67 | − size_t len; | |
| 68 | − unsigned msg_flags; | |
| 69 | − unsigned flags; | |
| 70 | − unsigned addr_len; | |
| 61 | + u16 flags; | |
| 62 | + /* initialised and used only by !msg send variants */ | |
| 63 | + u16 addr_len; | |
| 71 | 64 | void __user *addr; |
| 72 | − size_t done_io; | |
| 65 | + /* used only for send zerocopy */ | |
| 73 | 66 | struct io_kiocb *notif; |
| 74 | 67 | }; |
| 75 | 68 | |
| @@ -126,28 +119,36 @@ static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags) · 50 unmodified lines | ||
| 126 | 119 | } |
| 127 | 120 | } |
| 128 | 121 | |
| 129 | −static struct io_async_msghdr *io_recvmsg_alloc_async(struct io_kiocb *req, | |
| 130 | − unsigned int issue_flags) | |
| 122 | +static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req, | |
| 123 | + unsigned int issue_flags) | |
| 131 | 124 | { |
| 132 | 125 | struct io_ring_ctx *ctx = req->ctx; |
| 133 | 126 | struct io_cache_entry *entry; |
| 127 | + struct io_async_msghdr *hdr; | |
| 134 | 128 | |
| 135 | 129 | if (!(issue_flags & IO_URING_F_UNLOCKED) && |
| 136 | 130 | (entry = io_alloc_cache_get(&ctx->netmsg_cache)) != NULL) { |
| 137 | − struct io_async_msghdr *hdr; | |
| 138 | − | |
| 139 | 131 | hdr = container_of(entry, struct io_async_msghdr, cache); |
| 132 | + hdr->free_iov = NULL; | |
| 140 | 133 | req->flags |= REQ_F_ASYNC_DATA; |
| 141 | 134 | req->async_data = hdr; |
| 142 | 135 | return hdr; |
| 143 | 136 | } |
| 144 | 137 | |
| 145 | − if (!io_alloc_async_data(req)) | |
| 146 | − return req->async_data; | |
| 147 | − | |
| 138 | + if (!io_alloc_async_data(req)) { | |
| 139 | + hdr = req->async_data; | |
| 140 | + hdr->free_iov = NULL; | |
| 141 | + return hdr; | |
| 142 | + } | |
| 148 | 143 | return NULL; |
| 149 | 144 | } |
| 150 | 145 | |
| 146 | +static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req) | |
| 147 | +{ | |
| 148 | + /* ->prep_async is always called from the submission context */ | |
| 149 | + return io_msg_alloc_async(req, 0); | |
| 150 | +} | |
| 151 | + | |
| 151 | 152 | static int io_setup_async_msg(struct io_kiocb *req, |
| 152 | 153 | struct io_async_msghdr *kmsg, |
| 153 | 154 | unsigned int issue_flags) |
| @@ -156,17 +157,20 @@ static int io_setup_async_msg(struct io_kiocb *req, · 2 unmodified lines | ||
| 156 | 157 | |
| 157 | 158 | if (req_has_async_data(req)) |
| 158 | 159 | return -EAGAIN; |
| 159 | − async_msg = io_recvmsg_alloc_async(req, issue_flags); | |
| 160 | + async_msg = io_msg_alloc_async(req, issue_flags); | |
| 160 | 161 | if (!async_msg) { |
| 161 | 162 | kfree(kmsg->free_iov); |
| 162 | 163 | return -ENOMEM; |
| 163 | 164 | } |
| 164 | 165 | req->flags |= REQ_F_NEED_CLEANUP; |
| 165 | 166 | memcpy(async_msg, kmsg, sizeof(*kmsg)); |
| 166 | − async_msg->msg.msg_name = &async_msg->addr; | |
| 167 | + if (async_msg->msg.msg_name) | |
| 168 | + async_msg->msg.msg_name = &async_msg->addr; | |
| 167 | 169 | /* if were using fast_iov, set it to the new one */ |
| 168 | − if (!async_msg->free_iov) | |
| 169 | − async_msg->msg.msg_iter.iov = async_msg->fast_iov; | |
| 170 | + if (!kmsg->free_iov) { | |
| 171 | + size_t fast_idx = kmsg->msg.msg_iter.iov - kmsg->fast_iov; | |
| 172 | + async_msg->msg.msg_iter.iov = &async_msg->fast_iov[fast_idx]; | |
| 173 | + } | |
| 170 | 174 | |
| 171 | 175 | return -EAGAIN; |
| 172 | 176 | } |
| @@ -182,34 +186,34 @@ static int io_sendmsg_copy_hdr(struct io_kiocb *req, · 9 unmodified lines | ||
| 182 | 186 | &iomsg->free_iov); |
| 183 | 187 | } |
| 184 | 188 | |
| 185 | −int io_sendzc_prep_async(struct io_kiocb *req) | |
| 189 | +int io_send_prep_async(struct io_kiocb *req) | |
| 186 | 190 | { |
| 187 | − struct io_sendzc *zc = io_kiocb_to_cmd(req, struct io_sendzc); | |
| 191 | + struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 188 | 192 | struct io_async_msghdr *io; |
| 189 | 193 | int ret; |
| 190 | 194 | |
| 191 | 195 | if (!zc->addr || req_has_async_data(req)) |
| 192 | 196 | return 0; |
| 193 | − if (io_alloc_async_data(req)) | |
| 197 | + io = io_msg_alloc_async_prep(req); | |
| 198 | + if (!io) | |
| 194 | 199 | return -ENOMEM; |
| 195 | − | |
| 196 | − io = req->async_data; | |
| 197 | 200 | ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr); |
| 198 | 201 | return ret; |
| 199 | 202 | } |
| 200 | 203 | |
| 201 | 204 | static int io_setup_async_addr(struct io_kiocb *req, |
| 202 | − struct sockaddr_storage *addr, | |
| 205 | + struct sockaddr_storage *addr_storage, | |
| 203 | 206 | unsigned int issue_flags) |
| 204 | 207 | { |
| 208 | + struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 205 | 209 | struct io_async_msghdr *io; |
| 206 | 210 | |
| 207 | − if (!addr || req_has_async_data(req)) | |
| 211 | + if (!sr->addr || req_has_async_data(req)) | |
| 208 | 212 | return -EAGAIN; |
| 209 | − if (io_alloc_async_data(req)) | |
| 213 | + io = io_msg_alloc_async(req, issue_flags); | |
| 214 | + if (!io) | |
| 210 | 215 | return -ENOMEM; |
| 211 | − io = req->async_data; | |
| 212 | − memcpy(&io->addr, addr, sizeof(io->addr)); | |
| 216 | + memcpy(&io->addr, addr_storage, sizeof(io->addr)); | |
| 213 | 217 | return -EAGAIN; |
| 214 | 218 | } |
| 215 | 219 | |
| @@ -217,6 +221,8 @@ int io_sendmsg_prep_async(struct io_kiocb *req) · 1 unmodified line | ||
| 217 | 221 | { |
| 218 | 222 | int ret; |
| 219 | 223 | |
| 224 | + if (!io_msg_alloc_async_prep(req)) | |
| 225 | + return -ENOMEM; | |
| 220 | 226 | ret = io_sendmsg_copy_hdr(req, req->async_data); |
| 221 | 227 | if (!ret) |
| 222 | 228 | req->flags |= REQ_F_NEED_CLEANUP; |
| @@ -234,8 +240,14 @@ int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · 11 unmodified lines | ||
| 234 | 240 | { |
| 235 | 241 | struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg); |
| 236 | 242 | |
| 237 | − if (unlikely(sqe->file_index || sqe->addr2)) | |
| 243 | + if (req->opcode == IORING_OP_SEND) { | |
| 244 | + if (READ_ONCE(sqe->__pad3[0])) | |
| 245 | + return -EINVAL; | |
| 246 | + sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
| 247 | + sr->addr_len = READ_ONCE(sqe->addr_len); | |
| 248 | + } else if (sqe->addr2 || sqe->file_index) { | |
| 238 | 249 | return -EINVAL; |
| 250 | + } | |
| 239 | 251 | |
| 240 | 252 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
| 241 | 253 | sr->len = READ_ONCE(sqe->len); |
| @@ -291,13 +303,13 @@ int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags) · 49 unmodified lines | ||
| 291 | 303 | if (ret < min_ret) { |
| 292 | 304 | if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK)) |
| 293 | 305 | return io_setup_async_msg(req, kmsg, issue_flags); |
| 294 | − if (ret == -ERESTARTSYS) | |
| 295 | − ret = -EINTR; | |
| 296 | 306 | if (ret > 0 && io_net_retry(sock, flags)) { |
| 297 | 307 | sr->done_io += ret; |
| 298 | 308 | req->flags |= REQ_F_PARTIAL_IO; |
| 299 | 309 | return io_setup_async_msg(req, kmsg, issue_flags); |
| 300 | 310 | } |
| 311 | + if (ret == -ERESTARTSYS) | |
| 312 | + ret = -EINTR; | |
| 301 | 313 | req_set_fail(req); |
| 302 | 314 | } |
| 303 | 315 | /* fast path, check for non-NULL to avoid function call */ |
| @@ -315,6 +327,7 @@ int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags) · 11 unmodified lines | ||
| 315 | 327 | |
| 316 | 328 | int io_send(struct io_kiocb *req, unsigned int issue_flags) |
| 317 | 329 | { |
| 330 | + struct sockaddr_storage __address; | |
| 318 | 331 | struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg); |
| 319 | 332 | struct msghdr msg; |
| 320 | 333 | struct iovec iov; |
| @@ -323,9 +336,29 @@ int io_send(struct io_kiocb *req, unsigned int issue_flags) · 2 unmodified lines | ||
| 323 | 336 | int min_ret = 0; |
| 324 | 337 | int ret; |
| 325 | 338 | |
| 339 | + msg.msg_name = NULL; | |
| 340 | + msg.msg_control = NULL; | |
| 341 | + msg.msg_controllen = 0; | |
| 342 | + msg.msg_namelen = 0; | |
| 343 | + msg.msg_ubuf = NULL; | |
| 344 | + | |
| 345 | + if (sr->addr) { | |
| 346 | + if (req_has_async_data(req)) { | |
| 347 | + struct io_async_msghdr *io = req->async_data; | |
| 348 | + | |
| 349 | + msg.msg_name = &io->addr; | |
| 350 | + } else { | |
| 351 | + ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address); | |
| 352 | + if (unlikely(ret < 0)) | |
| 353 | + return ret; | |
| 354 | + msg.msg_name = (struct sockaddr *)&__address; | |
| 355 | + } | |
| 356 | + msg.msg_namelen = sr->addr_len; | |
| 357 | + } | |
| 358 | + | |
| 326 | 359 | if (!(req->flags & REQ_F_POLLED) && |
| 327 | 360 | (sr->flags & IORING_RECVSEND_POLL_FIRST)) |
| 328 | − return -EAGAIN; | |
| 361 | + return io_setup_async_addr(req, &__address, issue_flags); | |
| 329 | 362 | |
| 330 | 363 | sock = sock_from_file(req->file); |
| 331 | 364 | if (unlikely(!sock)) |
| @@ -335,12 +368,6 @@ int io_send(struct io_kiocb *req, unsigned int issue_flags) · 3 unmodified lines | ||
| 335 | 368 | if (unlikely(ret)) |
| 336 | 369 | return ret; |
| 337 | 370 | |
| 338 | − msg.msg_name = NULL; | |
| 339 | − msg.msg_control = NULL; | |
| 340 | − msg.msg_controllen = 0; | |
| 341 | − msg.msg_namelen = 0; | |
| 342 | − msg.msg_ubuf = NULL; | |
| 343 | − | |
| 344 | 371 | flags = sr->msg_flags; |
| 345 | 372 | if (issue_flags & IO_URING_F_NONBLOCK) |
| 346 | 373 | flags |= MSG_DONTWAIT; |
| @@ -351,16 +378,17 @@ int io_send(struct io_kiocb *req, unsigned int issue_flags) · 4 unmodified lines | ||
| 351 | 378 | ret = sock_sendmsg(sock, &msg); |
| 352 | 379 | if (ret < min_ret) { |
| 353 | 380 | if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK)) |
| 354 | − return -EAGAIN; | |
| 355 | − if (ret == -ERESTARTSYS) | |
| 356 | − ret = -EINTR; | |
| 381 | + return io_setup_async_addr(req, &__address, issue_flags); | |
| 382 | + | |
| 357 | 383 | if (ret > 0 && io_net_retry(sock, flags)) { |
| 358 | 384 | sr->len -= ret; |
| 359 | 385 | sr->buf += ret; |
| 360 | 386 | sr->done_io += ret; |
| 361 | 387 | req->flags |= REQ_F_PARTIAL_IO; |
| 362 | − return -EAGAIN; | |
| 388 | + return io_setup_async_addr(req, &__address, issue_flags); | |
| 363 | 389 | } |
| 390 | + if (ret == -ERESTARTSYS) | |
| 391 | + ret = -EINTR; | |
| 364 | 392 | req_set_fail(req); |
| 365 | 393 | } |
| 366 | 394 | if (ret >= 0) |
| @@ -454,7 +482,6 @@ static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, · 87 unmodified lines | ||
| 454 | 482 | |
| 455 | 483 | if (msg.msg_iovlen == 0) { |
| 456 | 484 | sr->len = 0; |
| 457 | − iomsg->free_iov = NULL; | |
| 458 | 485 | } else if (msg.msg_iovlen > 1) { |
| 459 | 486 | return -EINVAL; |
| 460 | 487 | } else { |
| @@ -465,7 +492,6 @@ static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, · 4 unmodified lines | ||
| 465 | 492 | if (clen < 0) |
| 466 | 493 | return -EINVAL; |
| 467 | 494 | sr->len = clen; |
| 468 | − iomsg->free_iov = NULL; | |
| 469 | 495 | } |
| 470 | 496 | |
| 471 | 497 | if (req->flags & REQ_F_APOLL_MULTISHOT) { |
| @@ -504,6 +530,8 @@ int io_recvmsg_prep_async(struct io_kiocb *req) · 32 unmodified lines | ||
| 504 | 530 | { |
| 505 | 531 | int ret; |
| 506 | 532 | |
| 533 | + if (!io_msg_alloc_async_prep(req)) | |
| 534 | + return -ENOMEM; | |
| 507 | 535 | ret = io_recvmsg_copy_hdr(req, req->async_data); |
| 508 | 536 | if (!ret) |
| 509 | 537 | req->flags |= REQ_F_NEED_CLEANUP; |
| @@ -751,13 +779,13 @@ retry_multishot: · 241 unmodified lines | ||
| 751 | 779 | } |
| 752 | 780 | return ret; |
| 753 | 781 | } |
| 754 | − if (ret == -ERESTARTSYS) | |
| 755 | − ret = -EINTR; | |
| 756 | 782 | if (ret > 0 && io_net_retry(sock, flags)) { |
| 757 | 783 | sr->done_io += ret; |
| 758 | 784 | req->flags |= REQ_F_PARTIAL_IO; |
| 759 | 785 | return io_setup_async_msg(req, kmsg, issue_flags); |
| 760 | 786 | } |
| 787 | + if (ret == -ERESTARTSYS) | |
| 788 | + ret = -EINTR; | |
| 761 | 789 | req_set_fail(req); |
| 762 | 790 | } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) { |
| 763 | 791 | req_set_fail(req); |
| @@ -847,8 +875,6 @@ retry_multishot: · 83 unmodified lines | ||
| 847 | 875 | |
| 848 | 876 | return -EAGAIN; |
| 849 | 877 | } |
| 850 | − if (ret == -ERESTARTSYS) | |
| 851 | − ret = -EINTR; | |
| 852 | 878 | if (ret > 0 && io_net_retry(sock, flags)) { |
| 853 | 879 | sr->len -= ret; |
| 854 | 880 | sr->buf += ret; |
| @@ -856,6 +882,8 @@ retry_multishot: · 1 unmodified line | ||
| 856 | 882 | req->flags |= REQ_F_PARTIAL_IO; |
| 857 | 883 | return -EAGAIN; |
| 858 | 884 | } |
| 885 | + if (ret == -ERESTARTSYS) | |
| 886 | + ret = -EINTR; | |
| 859 | 887 | req_set_fail(req); |
| 860 | 888 | } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) { |
| 861 | 889 | out_free: |
| @@ -879,23 +907,30 @@ out_free: · 17 unmodified lines | ||
| 879 | 907 | return ret; |
| 880 | 908 | } |
| 881 | 909 | |
| 882 | −void io_sendzc_cleanup(struct io_kiocb *req) | |
| 910 | +void io_send_zc_cleanup(struct io_kiocb *req) | |
| 883 | 911 | { |
| 884 | − struct io_sendzc *zc = io_kiocb_to_cmd(req, struct io_sendzc); | |
| 912 | + struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 913 | + struct io_async_msghdr *io; | |
| 885 | 914 | |
| 886 | − zc->notif->flags |= REQ_F_CQE_SKIP; | |
| 887 | − io_notif_flush(zc->notif); | |
| 888 | − zc->notif = NULL; | |
| 915 | + if (req_has_async_data(req)) { | |
| 916 | + io = req->async_data; | |
| 917 | + /* might be ->fast_iov if *msg_copy_hdr failed */ | |
| 918 | + if (io->free_iov != io->fast_iov) | |
| 919 | + kfree(io->free_iov); | |
| 920 | + } | |
| 921 | + if (zc->notif) { | |
| 922 | + io_notif_flush(zc->notif); | |
| 923 | + zc->notif = NULL; | |
| 924 | + } | |
| 889 | 925 | } |
| 890 | 926 | |
| 891 | −int io_sendzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
| 927 | +int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
| 892 | 928 | { |
| 893 | − struct io_sendzc *zc = io_kiocb_to_cmd(req, struct io_sendzc); | |
| 929 | + struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 894 | 930 | struct io_ring_ctx *ctx = req->ctx; |
| 895 | 931 | struct io_kiocb *notif; |
| 896 | 932 | |
| 897 | − if (READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3) || | |
| 898 | − READ_ONCE(sqe->__pad3[0])) | |
| 933 | + if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3))) | |
| 899 | 934 | return -EINVAL; |
| 900 | 935 | /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */ |
| 901 | 936 | if (req->flags & REQ_F_CQE_SKIP) |
| @@ -922,14 +957,24 @@ int io_sendzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · 20 unmodified lines | ||
| 922 | 957 | io_req_set_rsrc_node(notif, ctx, 0); |
| 923 | 958 | } |
| 924 | 959 | |
| 960 | + if (req->opcode == IORING_OP_SEND_ZC) { | |
| 961 | + if (READ_ONCE(sqe->__pad3[0])) | |
| 962 | + return -EINVAL; | |
| 963 | + zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
| 964 | + zc->addr_len = READ_ONCE(sqe->addr_len); | |
| 965 | + } else { | |
| 966 | + if (unlikely(sqe->addr2 || sqe->file_index)) | |
| 967 | + return -EINVAL; | |
| 968 | + if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF)) | |
| 969 | + return -EINVAL; | |
| 970 | + } | |
| 971 | + | |
| 925 | 972 | zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
| 926 | 973 | zc->len = READ_ONCE(sqe->len); |
| 927 | 974 | zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL; |
| 928 | 975 | if (zc->msg_flags & MSG_DONTWAIT) |
| 929 | 976 | req->flags |= REQ_F_NOWAIT; |
| 930 | 977 | |
| 931 | − zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
| 932 | − zc->addr_len = READ_ONCE(sqe->addr_len); | |
| 933 | 978 | zc->done_io = 0; |
| 934 | 979 | |
| 935 | 980 | #ifdef CONFIG_COMPAT |
| @@ -939,6 +984,13 @@ int io_sendzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · 3 unmodified lines | ||
| 939 | 984 | return 0; |
| 940 | 985 | } |
| 941 | 986 | |
| 987 | +static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb, | |
| 988 | + struct iov_iter *from, size_t length) | |
| 989 | +{ | |
| 990 | + skb_zcopy_downgrade_managed(skb); | |
| 991 | + return __zerocopy_sg_from_iter(NULL, sk, skb, from, length); | |
| 992 | +} | |
| 993 | + | |
| 942 | 994 | static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb, |
| 943 | 995 | struct iov_iter *from, size_t length) |
| 944 | 996 | { |
| @@ -949,13 +1001,10 @@ static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb, · 4 unmodified lines | ||
| 949 | 1001 | ssize_t copied = 0; |
| 950 | 1002 | unsigned long truesize = 0; |
| 951 | 1003 | |
| 952 | − if (!shinfo->nr_frags) | |
| 1004 | + if (!frag) | |
| 953 | 1005 | shinfo->flags |= SKBFL_MANAGED_FRAG_REFS; |
| 954 | − | |
| 955 | − if (!skb_zcopy_managed(skb) || !iov_iter_is_bvec(from)) { | |
| 956 | − skb_zcopy_downgrade_managed(skb); | |
| 1006 | + else if (unlikely(!skb_zcopy_managed(skb))) | |
| 957 | 1007 | return __zerocopy_sg_from_iter(NULL, sk, skb, from, length); |
| 958 | − } | |
| 959 | 1008 | |
| 960 | 1009 | bi.bi_size = min(from->count, length); |
| 961 | 1010 | bi.bi_bvec_done = from->iov_offset; |
| @@ -993,14 +1042,14 @@ static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb, · 31 unmodified lines | ||
| 993 | 1042 | return ret; |
| 994 | 1043 | } |
| 995 | 1044 | |
| 996 | −int io_sendzc(struct io_kiocb *req, unsigned int issue_flags) | |
| 1045 | +int io_send_zc(struct io_kiocb *req, unsigned int issue_flags) | |
| 997 | 1046 | { |
| 998 | − struct sockaddr_storage __address, *addr = NULL; | |
| 999 | − struct io_sendzc *zc = io_kiocb_to_cmd(req, struct io_sendzc); | |
| 1047 | + struct sockaddr_storage __address; | |
| 1048 | + struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 1000 | 1049 | struct msghdr msg; |
| 1001 | 1050 | struct iovec iov; |
| 1002 | 1051 | struct socket *sock; |
| 1003 | − unsigned msg_flags, cflags; | |
| 1052 | + unsigned msg_flags; | |
| 1004 | 1053 | int ret, min_ret = 0; |
| 1005 | 1054 | |
| 1006 | 1055 | sock = sock_from_file(req->file); |
| @@ -1016,26 +1065,26 @@ int io_sendzc(struct io_kiocb *req, unsigned int issue_flags) · 9 unmodified lines | ||
| 1016 | 1065 | if (req_has_async_data(req)) { |
| 1017 | 1066 | struct io_async_msghdr *io = req->async_data; |
| 1018 | 1067 | |
| 1019 | − msg.msg_name = addr = &io->addr; | |
| 1068 | + msg.msg_name = &io->addr; | |
| 1020 | 1069 | } else { |
| 1021 | 1070 | ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address); |
| 1022 | 1071 | if (unlikely(ret < 0)) |
| 1023 | 1072 | return ret; |
| 1024 | 1073 | msg.msg_name = (struct sockaddr *)&__address; |
| 1025 | − addr = &__address; | |
| 1026 | 1074 | } |
| 1027 | 1075 | msg.msg_namelen = zc->addr_len; |
| 1028 | 1076 | } |
| 1029 | 1077 | |
| 1030 | 1078 | if (!(req->flags & REQ_F_POLLED) && |
| 1031 | 1079 | (zc->flags & IORING_RECVSEND_POLL_FIRST)) |
| 1032 | − return io_setup_async_addr(req, addr, issue_flags); | |
| 1080 | + return io_setup_async_addr(req, &__address, issue_flags); | |
| 1033 | 1081 | |
| 1034 | 1082 | if (zc->flags & IORING_RECVSEND_FIXED_BUF) { |
| 1035 | 1083 | ret = io_import_fixed(WRITE, &msg.msg_iter, req->imu, |
| 1036 | 1084 | (u64)(uintptr_t)zc->buf, zc->len); |
| 1037 | 1085 | if (unlikely(ret)) |
| 1038 | 1086 | return ret; |
| 1087 | + msg.sg_from_iter = io_sg_from_iter; | |
| 1039 | 1088 | } else { |
| 1040 | 1089 | ret = import_single_range(WRITE, zc->buf, zc->len, &iov, |
| 1041 | 1090 | &msg.msg_iter); |
| @@ -1044,6 +1093,7 @@ int io_sendzc(struct io_kiocb *req, unsigned int issue_flags) · 2 unmodified lines | ||
| 1044 | 1093 | ret = io_notif_account_mem(zc->notif, zc->len); |
| 1045 | 1094 | if (unlikely(ret)) |
| 1046 | 1095 | return ret; |
| 1096 | + msg.sg_from_iter = io_sg_from_iter_iovec; | |
| 1047 | 1097 | } |
| 1048 | 1098 | |
| 1049 | 1099 | msg_flags = zc->msg_flags | MSG_ZEROCOPY; |
| @@ -1054,22 +1104,19 @@ int io_sendzc(struct io_kiocb *req, unsigned int issue_flags) · 4 unmodified lines | ||
| 1054 | 1104 | |
| 1055 | 1105 | msg.msg_flags = msg_flags; |
| 1056 | 1106 | msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg; |
| 1057 | − msg.sg_from_iter = io_sg_from_iter; | |
| 1058 | 1107 | ret = sock_sendmsg(sock, &msg); |
| 1059 | 1108 | |
| 1060 | 1109 | if (unlikely(ret < min_ret)) { |
| 1061 | 1110 | if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK)) |
| 1062 | − return io_setup_async_addr(req, addr, issue_flags); | |
| 1111 | + return io_setup_async_addr(req, &__address, issue_flags); | |
| 1063 | 1112 | |
| 1064 | 1113 | if (ret > 0 && io_net_retry(sock, msg.msg_flags)) { |
| 1065 | 1114 | zc->len -= ret; |
| 1066 | 1115 | zc->buf += ret; |
| 1067 | 1116 | zc->done_io += ret; |
| 1068 | 1117 | req->flags |= REQ_F_PARTIAL_IO; |
| 1069 | − return io_setup_async_addr(req, addr, issue_flags); | |
| 1118 | + return io_setup_async_addr(req, &__address, issue_flags); | |
| 1070 | 1119 | } |
| 1071 | − if (ret < 0 && !zc->done_io) | |
| 1072 | − zc->notif->flags |= REQ_F_CQE_SKIP; | |
| 1073 | 1120 | if (ret == -ERESTARTSYS) |
| 1074 | 1121 | ret = -EINTR; |
| 1075 | 1122 | req_set_fail(req); |
| @@ -1080,13 +1127,102 @@ int io_sendzc(struct io_kiocb *req, unsigned int issue_flags) · 4 unmodified lines | ||
| 1080 | 1127 | else if (zc->done_io) |
| 1081 | 1128 | ret = zc->done_io; |
| 1082 | 1129 | |
| 1083 | − io_notif_flush(zc->notif); | |
| 1084 | − req->flags &= ~REQ_F_NEED_CLEANUP; | |
| 1085 | − cflags = ret >= 0 ? IORING_CQE_F_MORE : 0; | |
| 1086 | − io_req_set_res(req, ret, cflags); | |
| 1130 | + /* | |
| 1131 | + * If we're in io-wq we can't rely on tw ordering guarantees, defer | |
| 1132 | + * flushing notif to io_send_zc_cleanup() | |
| 1133 | + */ | |
| 1134 | + if (!(issue_flags & IO_URING_F_UNLOCKED)) { | |
| 1135 | + io_notif_flush(zc->notif); | |
| 1136 | + req->flags &= ~REQ_F_NEED_CLEANUP; | |
| 1137 | + } | |
| 1138 | + io_req_set_res(req, ret, IORING_CQE_F_MORE); | |
| 1087 | 1139 | return IOU_OK; |
| 1088 | 1140 | } |
| 1089 | 1141 | |
| 1142 | +int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags) | |
| 1143 | +{ | |
| 1144 | + struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 1145 | + struct io_async_msghdr iomsg, *kmsg; | |
| 1146 | + struct socket *sock; | |
| 1147 | + unsigned flags; | |
| 1148 | + int ret, min_ret = 0; | |
| 1149 | + | |
| 1150 | + sock = sock_from_file(req->file); | |
| 1151 | + if (unlikely(!sock)) | |
| 1152 | + return -ENOTSOCK; | |
| 1153 | + | |
| 1154 | + if (req_has_async_data(req)) { | |
| 1155 | + kmsg = req->async_data; | |
| 1156 | + } else { | |
| 1157 | + ret = io_sendmsg_copy_hdr(req, &iomsg); | |
| 1158 | + if (ret) | |
| 1159 | + return ret; | |
| 1160 | + kmsg = &iomsg; | |
| 1161 | + } | |
| 1162 | + | |
| 1163 | + if (!(req->flags & REQ_F_POLLED) && | |
| 1164 | + (sr->flags & IORING_RECVSEND_POLL_FIRST)) | |
| 1165 | + return io_setup_async_msg(req, kmsg, issue_flags); | |
| 1166 | + | |
| 1167 | + flags = sr->msg_flags | MSG_ZEROCOPY; | |
| 1168 | + if (issue_flags & IO_URING_F_NONBLOCK) | |
| 1169 | + flags |= MSG_DONTWAIT; | |
| 1170 | + if (flags & MSG_WAITALL) | |
| 1171 | + min_ret = iov_iter_count(&kmsg->msg.msg_iter); | |
| 1172 | + | |
| 1173 | + kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg; | |
| 1174 | + kmsg->msg.sg_from_iter = io_sg_from_iter_iovec; | |
| 1175 | + ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); | |
| 1176 | + | |
| 1177 | + if (unlikely(ret < min_ret)) { | |
| 1178 | + if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK)) | |
| 1179 | + return io_setup_async_msg(req, kmsg, issue_flags); | |
| 1180 | + | |
| 1181 | + if (ret > 0 && io_net_retry(sock, flags)) { | |
| 1182 | + sr->done_io += ret; | |
| 1183 | + req->flags |= REQ_F_PARTIAL_IO; | |
| 1184 | + return io_setup_async_msg(req, kmsg, issue_flags); | |
| 1185 | + } | |
| 1186 | + if (ret == -ERESTARTSYS) | |
| 1187 | + ret = -EINTR; | |
| 1188 | + req_set_fail(req); | |
| 1189 | + } | |
| 1190 | + /* fast path, check for non-NULL to avoid function call */ | |
| 1191 | + if (kmsg->free_iov) { | |
| 1192 | + kfree(kmsg->free_iov); | |
| 1193 | + kmsg->free_iov = NULL; | |
| 1194 | + } | |
| 1195 | + | |
| 1196 | + io_netmsg_recycle(req, issue_flags); | |
| 1197 | + if (ret >= 0) | |
| 1198 | + ret += sr->done_io; | |
| 1199 | + else if (sr->done_io) | |
| 1200 | + ret = sr->done_io; | |
| 1201 | + | |
| 1202 | + /* | |
| 1203 | + * If we're in io-wq we can't rely on tw ordering guarantees, defer | |
| 1204 | + * flushing notif to io_send_zc_cleanup() | |
| 1205 | + */ | |
| 1206 | + if (!(issue_flags & IO_URING_F_UNLOCKED)) { | |
| 1207 | + io_notif_flush(sr->notif); | |
| 1208 | + req->flags &= ~REQ_F_NEED_CLEANUP; | |
| 1209 | + } | |
| 1210 | + io_req_set_res(req, ret, IORING_CQE_F_MORE); | |
| 1211 | + return IOU_OK; | |
| 1212 | +} | |
| 1213 | + | |
| 1214 | +void io_sendrecv_fail(struct io_kiocb *req) | |
| 1215 | +{ | |
| 1216 | + struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg); | |
| 1217 | + | |
| 1218 | + if (req->flags & REQ_F_PARTIAL_IO) | |
| 1219 | + req->cqe.res = sr->done_io; | |
| 1220 | + | |
| 1221 | + if ((req->flags & REQ_F_NEED_CLEANUP) && | |
| 1222 | + (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC)) | |
| 1223 | + req->cqe.flags |= IORING_CQE_F_MORE; | |
| 1224 | +} | |
| 1225 | + | |
| 1090 | 1226 | int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
| 1091 | 1227 | { |
| 1092 | 1228 | struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept); |
| @@ -31,18 +31,21 @@ struct io_async_connect { · 30 unmodified lines | ||
| 31 | 31 | int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| 32 | 32 | int io_shutdown(struct io_kiocb *req, unsigned int issue_flags); |
| 33 | 33 | |
| 34 | −int io_sendzc_prep_async(struct io_kiocb *req); | |
| 35 | 34 | int io_sendmsg_prep_async(struct io_kiocb *req); |
| 36 | 35 | void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req); |
| 37 | 36 | int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| 38 | 37 | int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags); |
| 38 | + | |
| 39 | 39 | int io_send(struct io_kiocb *req, unsigned int issue_flags); |
| 40 | +int io_send_prep_async(struct io_kiocb *req); | |
| 40 | 41 | |
| 41 | 42 | int io_recvmsg_prep_async(struct io_kiocb *req); |
| 42 | 43 | int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| 43 | 44 | int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags); |
| 44 | 45 | int io_recv(struct io_kiocb *req, unsigned int issue_flags); |
| 45 | 46 | |
| 47 | +void io_sendrecv_fail(struct io_kiocb *req); | |
| 48 | + | |
| 46 | 49 | int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| 47 | 50 | int io_accept(struct io_kiocb *req, unsigned int issue_flags); |
| 48 | 51 | |
| @@ -53,9 +56,10 @@ int io_connect_prep_async(struct io_kiocb *req); · 4 unmodified lines | ||
| 53 | 56 | int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| 54 | 57 | int io_connect(struct io_kiocb *req, unsigned int issue_flags); |
| 55 | 58 | |
| 56 | −int io_sendzc(struct io_kiocb *req, unsigned int issue_flags); | |
| 57 | −int io_sendzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); | |
| 58 | −void io_sendzc_cleanup(struct io_kiocb *req); | |
| 59 | +int io_send_zc(struct io_kiocb *req, unsigned int issue_flags); | |
| 60 | +int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags); | |
| 61 | +int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); | |
| 62 | +void io_send_zc_cleanup(struct io_kiocb *req); | |
| 59 | 63 | |
| 60 | 64 | void io_netmsg_cache_free(struct io_cache_entry *entry); |
| 61 | 65 | #else |
| @@ -69,6 +69,7 @@ const struct io_op_def io_op_defs[] = { · 68 unmodified lines | ||
| 69 | 69 | .issue = io_read, |
| 70 | 70 | .prep_async = io_readv_prep_async, |
| 71 | 71 | .cleanup = io_readv_writev_cleanup, |
| 72 | + .fail = io_rw_fail, | |
| 72 | 73 | }, |
| 73 | 74 | [IORING_OP_WRITEV] = { |
| 74 | 75 | .needs_file = 1, |
| @@ -85,6 +86,7 @@ const struct io_op_def io_op_defs[] = { · 10 unmodified lines | ||
| 85 | 86 | .issue = io_write, |
| 86 | 87 | .prep_async = io_writev_prep_async, |
| 87 | 88 | .cleanup = io_readv_writev_cleanup, |
| 89 | + .fail = io_rw_fail, | |
| 88 | 90 | }, |
| 89 | 91 | [IORING_OP_FSYNC] = { |
| 90 | 92 | .needs_file = 1, |
| @@ -105,6 +107,7 @@ const struct io_op_def io_op_defs[] = { · 14 unmodified lines | ||
| 105 | 107 | .name = "READ_FIXED", |
| 106 | 108 | .prep = io_prep_rw, |
| 107 | 109 | .issue = io_read, |
| 110 | + .fail = io_rw_fail, | |
| 108 | 111 | }, |
| 109 | 112 | [IORING_OP_WRITE_FIXED] = { |
| 110 | 113 | .needs_file = 1, |
| @@ -119,6 +122,7 @@ const struct io_op_def io_op_defs[] = { · 8 unmodified lines | ||
| 119 | 122 | .name = "WRITE_FIXED", |
| 120 | 123 | .prep = io_prep_rw, |
| 121 | 124 | .issue = io_write, |
| 125 | + .fail = io_rw_fail, | |
| 122 | 126 | }, |
| 123 | 127 | [IORING_OP_POLL_ADD] = { |
| 124 | 128 | .needs_file = 1, |
| @@ -146,6 +150,7 @@ const struct io_op_def io_op_defs[] = { · 21 unmodified lines | ||
| 146 | 150 | .unbound_nonreg_file = 1, |
| 147 | 151 | .pollout = 1, |
| 148 | 152 | .ioprio = 1, |
| 153 | + .manual_alloc = 1, | |
| 149 | 154 | .name = "SENDMSG", |
| 150 | 155 | #if defined(CONFIG_NET) |
| 151 | 156 | .async_size = sizeof(struct io_async_msghdr), |
| @@ -153,6 +158,7 @@ const struct io_op_def io_op_defs[] = { · 1 unmodified line | ||
| 153 | 158 | .issue = io_sendmsg, |
| 154 | 159 | .prep_async = io_sendmsg_prep_async, |
| 155 | 160 | .cleanup = io_sendmsg_recvmsg_cleanup, |
| 161 | + .fail = io_sendrecv_fail, | |
| 156 | 162 | #else |
| 157 | 163 | .prep = io_eopnotsupp_prep, |
| 158 | 164 | #endif |
| @@ -163,6 +169,7 @@ const struct io_op_def io_op_defs[] = { · 4 unmodified lines | ||
| 163 | 169 | .pollin = 1, |
| 164 | 170 | .buffer_select = 1, |
| 165 | 171 | .ioprio = 1, |
| 172 | + .manual_alloc = 1, | |
| 166 | 173 | .name = "RECVMSG", |
| 167 | 174 | #if defined(CONFIG_NET) |
| 168 | 175 | .async_size = sizeof(struct io_async_msghdr), |
| @@ -170,6 +177,7 @@ const struct io_op_def io_op_defs[] = { · 1 unmodified line | ||
| 170 | 177 | .issue = io_recvmsg, |
| 171 | 178 | .prep_async = io_recvmsg_prep_async, |
| 172 | 179 | .cleanup = io_sendmsg_recvmsg_cleanup, |
| 180 | + .fail = io_sendrecv_fail, | |
| 173 | 181 | #else |
| 174 | 182 | .prep = io_eopnotsupp_prep, |
| 175 | 183 | #endif |
| @@ -273,6 +281,7 @@ const struct io_op_def io_op_defs[] = { · 97 unmodified lines | ||
| 273 | 281 | .name = "READ", |
| 274 | 282 | .prep = io_prep_rw, |
| 275 | 283 | .issue = io_read, |
| 284 | + .fail = io_rw_fail, | |
| 276 | 285 | }, |
| 277 | 286 | [IORING_OP_WRITE] = { |
| 278 | 287 | .needs_file = 1, |
| @@ -287,6 +296,7 @@ const struct io_op_def io_op_defs[] = { · 8 unmodified lines | ||
| 287 | 296 | .name = "WRITE", |
| 288 | 297 | .prep = io_prep_rw, |
| 289 | 298 | .issue = io_write, |
| 299 | + .fail = io_rw_fail, | |
| 290 | 300 | }, |
| 291 | 301 | [IORING_OP_FADVISE] = { |
| 292 | 302 | .needs_file = 1, |
| @@ -306,10 +316,14 @@ const struct io_op_def io_op_defs[] = { · 13 unmodified lines | ||
| 306 | 316 | .pollout = 1, |
| 307 | 317 | .audit_skip = 1, |
| 308 | 318 | .ioprio = 1, |
| 319 | + .manual_alloc = 1, | |
| 309 | 320 | .name = "SEND", |
| 310 | 321 | #if defined(CONFIG_NET) |
| 322 | + .async_size = sizeof(struct io_async_msghdr), | |
| 311 | 323 | .prep = io_sendmsg_prep, |
| 312 | 324 | .issue = io_send, |
| 325 | + .fail = io_sendrecv_fail, | |
| 326 | + .prep_async = io_send_prep_async, | |
| 313 | 327 | #else |
| 314 | 328 | .prep = io_eopnotsupp_prep, |
| 315 | 329 | #endif |
| @@ -325,6 +339,7 @@ const struct io_op_def io_op_defs[] = { · 9 unmodified lines | ||
| 325 | 339 | #if defined(CONFIG_NET) |
| 326 | 340 | .prep = io_recvmsg_prep, |
| 327 | 341 | .issue = io_recv, |
| 342 | + .fail = io_sendrecv_fail, | |
| 328 | 343 | #else |
| 329 | 344 | .prep = io_eopnotsupp_prep, |
| 330 | 345 | #endif |
| @@ -465,6 +480,7 @@ const struct io_op_def io_op_defs[] = { · 134 unmodified lines | ||
| 465 | 480 | .needs_file = 1, |
| 466 | 481 | .plug = 1, |
| 467 | 482 | .name = "URING_CMD", |
| 483 | + .iopoll = 1, | |
| 468 | 484 | .async_size = uring_cmd_pdu_size(1), |
| 469 | 485 | .prep = io_uring_cmd_prep, |
| 470 | 486 | .issue = io_uring_cmd, |
| @@ -480,10 +496,30 @@ const struct io_op_def io_op_defs[] = { · 9 unmodified lines | ||
| 480 | 496 | .manual_alloc = 1, |
| 481 | 497 | #if defined(CONFIG_NET) |
| 482 | 498 | .async_size = sizeof(struct io_async_msghdr), |
| 483 | − .prep = io_sendzc_prep, | |
| 484 | − .issue = io_sendzc, | |
| 485 | − .prep_async = io_sendzc_prep_async, | |
| 486 | − .cleanup = io_sendzc_cleanup, | |
| 499 | + .prep = io_send_zc_prep, | |
| 500 | + .issue = io_send_zc, | |
| 501 | + .prep_async = io_send_prep_async, | |
| 502 | + .cleanup = io_send_zc_cleanup, | |
| 503 | + .fail = io_sendrecv_fail, | |
| 504 | +#else | |
| 505 | + .prep = io_eopnotsupp_prep, | |
| 506 | +#endif | |
| 507 | + }, | |
| 508 | + [IORING_OP_SENDMSG_ZC] = { | |
| 509 | + .name = "SENDMSG_ZC", | |
| 510 | + .needs_file = 1, | |
| 511 | + .unbound_nonreg_file = 1, | |
| 512 | + .pollout = 1, | |
| 513 | + .audit_skip = 1, | |
| 514 | + .ioprio = 1, | |
| 515 | + .manual_alloc = 1, | |
| 516 | +#if defined(CONFIG_NET) | |
| 517 | + .async_size = sizeof(struct io_async_msghdr), | |
| 518 | + .prep = io_send_zc_prep, | |
| 519 | + .issue = io_sendmsg_zc, | |
| 520 | + .prep_async = io_sendmsg_prep_async, | |
| 521 | + .cleanup = io_send_zc_cleanup, | |
| 522 | + .fail = io_sendrecv_fail, | |
| 487 | 523 | #else |
| 488 | 524 | .prep = io_eopnotsupp_prep, |
| 489 | 525 | #endif |
| @@ -36,6 +36,7 @@ struct io_op_def { · 35 unmodified lines | ||
| 36 | 36 | int (*issue)(struct io_kiocb *, unsigned int); |
| 37 | 37 | int (*prep_async)(struct io_kiocb *); |
| 38 | 38 | void (*cleanup)(struct io_kiocb *); |
| 39 | + void (*fail)(struct io_kiocb *); | |
| 39 | 40 | }; |
| 40 | 41 | |
| 41 | 42 | extern const struct io_op_def io_op_defs[]; |
| @@ -341,7 +341,7 @@ __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, · 340 unmodified lines | ||
| 341 | 341 | flush_delayed_work(&ctx->rsrc_put_work); |
| 342 | 342 | reinit_completion(&data->done); |
| 343 | 343 | |
| 344 | − ret = io_run_task_work_sig(); | |
| 344 | + ret = io_run_task_work_sig(ctx); | |
| 345 | 345 | mutex_lock(&ctx->uring_lock); |
| 346 | 346 | } while (ret >= 0); |
| 347 | 347 | data->quiesce = false; |
| @@ -33,6 +33,46 @@ static inline bool io_file_supports_nowait(struct io_kiocb *req) · 32 unmodified lines | ||
| 33 | 33 | return req->flags & REQ_F_SUPPORT_NOWAIT; |
| 34 | 34 | } |
| 35 | 35 | |
| 36 | +#ifdef CONFIG_COMPAT | |
| 37 | +static int io_iov_compat_buffer_select_prep(struct io_rw *rw) | |
| 38 | +{ | |
| 39 | + struct compat_iovec __user *uiov; | |
| 40 | + compat_ssize_t clen; | |
| 41 | + | |
| 42 | + uiov = u64_to_user_ptr(rw->addr); | |
| 43 | + if (!access_ok(uiov, sizeof(*uiov))) | |
| 44 | + return -EFAULT; | |
| 45 | + if (__get_user(clen, &uiov->iov_len)) | |
| 46 | + return -EFAULT; | |
| 47 | + if (clen < 0) | |
| 48 | + return -EINVAL; | |
| 49 | + | |
| 50 | + rw->len = clen; | |
| 51 | + return 0; | |
| 52 | +} | |
| 53 | +#endif | |
| 54 | + | |
| 55 | +static int io_iov_buffer_select_prep(struct io_kiocb *req) | |
| 56 | +{ | |
| 57 | + struct iovec __user *uiov; | |
| 58 | + struct iovec iov; | |
| 59 | + struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 60 | + | |
| 61 | + if (rw->len != 1) | |
| 62 | + return -EINVAL; | |
| 63 | + | |
| 64 | +#ifdef CONFIG_COMPAT | |
| 65 | + if (req->ctx->compat) | |
| 66 | + return io_iov_compat_buffer_select_prep(rw); | |
| 67 | +#endif | |
| 68 | + | |
| 69 | + uiov = u64_to_user_ptr(rw->addr); | |
| 70 | + if (copy_from_user(&iov, uiov, sizeof(*uiov))) | |
| 71 | + return -EFAULT; | |
| 72 | + rw->len = iov.iov_len; | |
| 73 | + return 0; | |
| 74 | +} | |
| 75 | + | |
| 36 | 76 | int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
| 37 | 77 | { |
| 38 | 78 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
| @@ -69,6 +109,16 @@ int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) · 30 unmodified lines | ||
| 69 | 109 | rw->addr = READ_ONCE(sqe->addr); |
| 70 | 110 | rw->len = READ_ONCE(sqe->len); |
| 71 | 111 | rw->flags = READ_ONCE(sqe->rw_flags); |
| 112 | + | |
| 113 | + /* Have to do this validation here, as this is in io_read() rw->len might | |
| 114 | + * have chanaged due to buffer selection | |
| 115 | + */ | |
| 116 | + if (req->opcode == IORING_OP_READV && req->flags & REQ_F_BUFFER_SELECT) { | |
| 117 | + ret = io_iov_buffer_select_prep(req); | |
| 118 | + if (ret) | |
| 119 | + return ret; | |
| 120 | + } | |
| 121 | + | |
| 72 | 122 | return 0; |
| 73 | 123 | } |
| 74 | 124 | |
| @@ -186,14 +236,6 @@ static void kiocb_end_write(struct io_kiocb *req) · 111 unmodified lines | ||
| 186 | 236 | |
| 187 | 237 | static bool __io_complete_rw_common(struct io_kiocb *req, long res) |
| 188 | 238 | { |
| 189 | − struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 190 | − | |
| 191 | − if (rw->kiocb.ki_flags & IOCB_WRITE) { | |
| 192 | − kiocb_end_write(req); | |
| 193 | − fsnotify_modify(req->file); | |
| 194 | − } else { | |
| 195 | − fsnotify_access(req->file); | |
| 196 | − } | |
| 197 | 239 | if (unlikely(res != req->cqe.res)) { |
| 198 | 240 | if ((res == -EAGAIN || res == -EOPNOTSUPP) && |
| 199 | 241 | io_rw_should_reissue(req)) { |
| @@ -220,6 +262,20 @@ static inline int io_fixup_rw_res(struct io_kiocb *req, long res) · 20 unmodified lines | ||
| 220 | 262 | return res; |
| 221 | 263 | } |
| 222 | 264 | |
| 265 | +static void io_req_rw_complete(struct io_kiocb *req, bool *locked) | |
| 266 | +{ | |
| 267 | + struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 268 | + | |
| 269 | + if (rw->kiocb.ki_flags & IOCB_WRITE) { | |
| 270 | + kiocb_end_write(req); | |
| 271 | + fsnotify_modify(req->file); | |
| 272 | + } else { | |
| 273 | + fsnotify_access(req->file); | |
| 274 | + } | |
| 275 | + | |
| 276 | + io_req_task_complete(req, locked); | |
| 277 | +} | |
| 278 | + | |
| 223 | 279 | static void io_complete_rw(struct kiocb *kiocb, long res) |
| 224 | 280 | { |
| 225 | 281 | struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb); |
| @@ -228,7 +284,7 @@ static void io_complete_rw(struct kiocb *kiocb, long res) · 2 unmodified lines | ||
| 228 | 284 | if (__io_complete_rw_common(req, res)) |
| 229 | 285 | return; |
| 230 | 286 | io_req_set_res(req, io_fixup_rw_res(req, res), 0); |
| 231 | − req->io_task_work.func = io_req_task_complete; | |
| 287 | + req->io_task_work.func = io_req_rw_complete; | |
| 232 | 288 | io_req_task_work_add(req); |
| 233 | 289 | } |
| 234 | 290 | |
| @@ -279,79 +335,6 @@ static int kiocb_done(struct io_kiocb *req, ssize_t ret, · 44 unmodified lines | ||
| 279 | 335 | return IOU_ISSUE_SKIP_COMPLETE; |
| 280 | 336 | } |
| 281 | 337 | |
| 282 | −#ifdef CONFIG_COMPAT | |
| 283 | −static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
| 284 | − unsigned int issue_flags) | |
| 285 | −{ | |
| 286 | − struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 287 | − struct compat_iovec __user *uiov; | |
| 288 | − compat_ssize_t clen; | |
| 289 | − void __user *buf; | |
| 290 | − size_t len; | |
| 291 | − | |
| 292 | − uiov = u64_to_user_ptr(rw->addr); | |
| 293 | − if (!access_ok(uiov, sizeof(*uiov))) | |
| 294 | − return -EFAULT; | |
| 295 | − if (__get_user(clen, &uiov->iov_len)) | |
| 296 | − return -EFAULT; | |
| 297 | − if (clen < 0) | |
| 298 | − return -EINVAL; | |
| 299 | − | |
| 300 | − len = clen; | |
| 301 | − buf = io_buffer_select(req, &len, issue_flags); | |
| 302 | − if (!buf) | |
| 303 | − return -ENOBUFS; | |
| 304 | − rw->addr = (unsigned long) buf; | |
| 305 | − iov[0].iov_base = buf; | |
| 306 | − rw->len = iov[0].iov_len = (compat_size_t) len; | |
| 307 | − return 0; | |
| 308 | −} | |
| 309 | −#endif | |
| 310 | − | |
| 311 | −static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
| 312 | − unsigned int issue_flags) | |
| 313 | −{ | |
| 314 | − struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 315 | − struct iovec __user *uiov = u64_to_user_ptr(rw->addr); | |
| 316 | − void __user *buf; | |
| 317 | − ssize_t len; | |
| 318 | − | |
| 319 | − if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
| 320 | − return -EFAULT; | |
| 321 | − | |
| 322 | − len = iov[0].iov_len; | |
| 323 | − if (len < 0) | |
| 324 | − return -EINVAL; | |
| 325 | − buf = io_buffer_select(req, &len, issue_flags); | |
| 326 | − if (!buf) | |
| 327 | − return -ENOBUFS; | |
| 328 | − rw->addr = (unsigned long) buf; | |
| 329 | − iov[0].iov_base = buf; | |
| 330 | − rw->len = iov[0].iov_len = len; | |
| 331 | − return 0; | |
| 332 | −} | |
| 333 | − | |
| 334 | −static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
| 335 | − unsigned int issue_flags) | |
| 336 | −{ | |
| 337 | − struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 338 | − | |
| 339 | − if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) { | |
| 340 | − iov[0].iov_base = u64_to_user_ptr(rw->addr); | |
| 341 | − iov[0].iov_len = rw->len; | |
| 342 | − return 0; | |
| 343 | − } | |
| 344 | − if (rw->len != 1) | |
| 345 | − return -EINVAL; | |
| 346 | − | |
| 347 | −#ifdef CONFIG_COMPAT | |
| 348 | − if (req->ctx->compat) | |
| 349 | − return io_compat_import(req, iov, issue_flags); | |
| 350 | −#endif | |
| 351 | − | |
| 352 | − return __io_iov_buffer_select(req, iov, issue_flags); | |
| 353 | −} | |
| 354 | − | |
| 355 | 338 | static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req, |
| 356 | 339 | struct io_rw_state *s, |
| 357 | 340 | unsigned int issue_flags) |
| @@ -374,7 +357,8 @@ static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req, · 16 unmodified lines | ||
| 374 | 357 | buf = u64_to_user_ptr(rw->addr); |
| 375 | 358 | sqe_len = rw->len; |
| 376 | 359 | |
| 377 | − if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { | |
| 360 | + if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE || | |
| 361 | + (req->flags & REQ_F_BUFFER_SELECT)) { | |
| 378 | 362 | if (io_do_buffer_select(req)) { |
| 379 | 363 | buf = io_buffer_select(req, &sqe_len, issue_flags); |
| 380 | 364 | if (!buf) |
| @@ -390,14 +374,6 @@ static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req, · 9 unmodified lines | ||
| 390 | 374 | } |
| 391 | 375 | |
| 392 | 376 | iovec = s->fast_iov; |
| 393 | − if (req->flags & REQ_F_BUFFER_SELECT) { | |
| 394 | − ret = io_iov_buffer_select(req, iovec, issue_flags); | |
| 395 | − if (ret) | |
| 396 | − return ERR_PTR(ret); | |
| 397 | − iov_iter_init(iter, ddir, iovec, 1, iovec->iov_len); | |
| 398 | − return NULL; | |
| 399 | − } | |
| 400 | − | |
| 401 | 377 | ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter, |
| 402 | 378 | req->ctx->compat); |
| 403 | 379 | if (unlikely(ret < 0)) |
| @@ -794,10 +770,12 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags) · 390 unmodified lines | ||
| 794 | 770 | iov_iter_restore(&s->iter, &s->iter_state); |
| 795 | 771 | |
| 796 | 772 | ret2 = io_setup_async_rw(req, iovec, s, true); |
| 797 | − if (ret2) | |
| 798 | − return ret2; | |
| 799 | − | |
| 800 | 773 | iovec = NULL; |
| 774 | + if (ret2) { | |
| 775 | + ret = ret > 0 ? ret : ret2; | |
| 776 | + goto done; | |
| 777 | + } | |
| 778 | + | |
| 801 | 779 | io = req->async_data; |
| 802 | 780 | s = &io->s; |
| 803 | 781 | /* |
| @@ -823,6 +801,7 @@ int io_read(struct io_kiocb *req, unsigned int issue_flags) · 19 unmodified lines | ||
| 823 | 801 | return -EAGAIN; |
| 824 | 802 | } |
| 825 | 803 | |
| 804 | + req->cqe.res = iov_iter_count(&s->iter); | |
| 826 | 805 | /* |
| 827 | 806 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If |
| 828 | 807 | * we get -EIOCBQUEUED, then we'll get a notification when the |
| @@ -984,6 +963,14 @@ static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx) · 155 unmodified lines | ||
| 984 | 963 | io_cqring_wake(ctx); |
| 985 | 964 | } |
| 986 | 965 | |
| 966 | +void io_rw_fail(struct io_kiocb *req) | |
| 967 | +{ | |
| 968 | + int res; | |
| 969 | + | |
| 970 | + res = io_fixup_rw_res(req, req->cqe.res); | |
| 971 | + io_req_set_res(req, res, req->cqe.flags); | |
| 972 | +} | |
| 973 | + | |
| 987 | 974 | int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin) |
| 988 | 975 | { |
| 989 | 976 | struct io_wq_work_node *pos, *start, *prev; |
| @@ -1000,7 +987,7 @@ int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin) · 10 unmodified lines | ||
| 1000 | 987 | |
| 1001 | 988 | wq_list_for_each(pos, start, &ctx->iopoll_list) { |
| 1002 | 989 | struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list); |
| 1003 | − struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 990 | + struct file *file = req->file; | |
| 1004 | 991 | int ret; |
| 1005 | 992 | |
| 1006 | 993 | /* |
| @@ -1011,7 +998,17 @@ int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin) · 4 unmodified lines | ||
| 1011 | 998 | if (READ_ONCE(req->iopoll_completed)) |
| 1012 | 999 | break; |
| 1013 | 1000 | |
| 1014 | − ret = rw->kiocb.ki_filp->f_op->iopoll(&rw->kiocb, &iob, poll_flags); | |
| 1001 | + if (req->opcode == IORING_OP_URING_CMD) { | |
| 1002 | + struct io_uring_cmd *ioucmd; | |
| 1003 | + | |
| 1004 | + ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd); | |
| 1005 | + ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob, | |
| 1006 | + poll_flags); | |
| 1007 | + } else { | |
| 1008 | + struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
| 1009 | + | |
| 1010 | + ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags); | |
| 1011 | + } | |
| 1015 | 1012 | if (unlikely(ret < 0)) |
| 1016 | 1013 | return ret; |
| 1017 | 1014 | else if (ret) |
| @@ -21,3 +21,4 @@ int io_readv_prep_async(struct io_kiocb *req); · 20 unmodified lines | ||
| 21 | 21 | int io_write(struct io_kiocb *req, unsigned int issue_flags); |
| 22 | 22 | int io_writev_prep_async(struct io_kiocb *req); |
| 23 | 23 | void io_readv_writev_cleanup(struct io_kiocb *req); |
| 24 | +void io_rw_fail(struct io_kiocb *req); | |
| @@ -149,11 +149,10 @@ static inline void io_remove_next_linked(struct io_kiocb *req) · 148 unmodified lines | ||
| 149 | 149 | nxt->link = NULL; |
| 150 | 150 | } |
| 151 | 151 | |
| 152 | −bool io_disarm_next(struct io_kiocb *req) | |
| 152 | +void io_disarm_next(struct io_kiocb *req) | |
| 153 | 153 | __must_hold(&req->ctx->completion_lock) |
| 154 | 154 | { |
| 155 | 155 | struct io_kiocb *link = NULL; |
| 156 | − bool posted = false; | |
| 157 | 156 | |
| 158 | 157 | if (req->flags & REQ_F_ARM_LTIMEOUT) { |
| 159 | 158 | link = req->link; |
| @@ -161,7 +160,6 @@ bool io_disarm_next(struct io_kiocb *req) · 1 unmodified line | ||
| 161 | 160 | if (link && link->opcode == IORING_OP_LINK_TIMEOUT) { |
| 162 | 161 | io_remove_next_linked(req); |
| 163 | 162 | io_req_tw_post_queue(link, -ECANCELED, 0); |
| 164 | − posted = true; | |
| 165 | 163 | } |
| 166 | 164 | } else if (req->flags & REQ_F_LINK_TIMEOUT) { |
| 167 | 165 | struct io_ring_ctx *ctx = req->ctx; |
| @@ -169,17 +167,12 @@ bool io_disarm_next(struct io_kiocb *req) · 1 unmodified line | ||
| 169 | 167 | spin_lock_irq(&ctx->timeout_lock); |
| 170 | 168 | link = io_disarm_linked_timeout(req); |
| 171 | 169 | spin_unlock_irq(&ctx->timeout_lock); |
| 172 | − if (link) { | |
| 173 | − posted = true; | |
| 170 | + if (link) | |
| 174 | 171 | io_req_tw_post_queue(link, -ECANCELED, 0); |
| 175 | − } | |
| 176 | 172 | } |
| 177 | 173 | if (unlikely((req->flags & REQ_F_FAIL) && |
| 178 | − !(req->flags & REQ_F_HARDLINK))) { | |
| 179 | − posted |= (req->link != NULL); | |
| 174 | + !(req->flags & REQ_F_HARDLINK))) | |
| 180 | 175 | io_fail_links(req); |
| 181 | − } | |
| 182 | − return posted; | |
| 183 | 176 | } |
| 184 | 177 | |
| 185 | 178 | struct io_kiocb *__io_disarm_linked_timeout(struct io_kiocb *req, |
| @@ -27,7 +27,7 @@ int io_timeout_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd); · 26 unmodified lines | ||
| 27 | 27 | __cold bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk, |
| 28 | 28 | bool cancel_all); |
| 29 | 29 | void io_queue_linked_timeout(struct io_kiocb *req); |
| 30 | −bool io_disarm_next(struct io_kiocb *req); | |
| 30 | +void io_disarm_next(struct io_kiocb *req); | |
| 31 | 31 | |
| 32 | 32 | int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| 33 | 33 | int io_link_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe); |
| @@ -50,7 +50,11 @@ void io_uring_cmd_done(struct io_uring_cmd *ioucmd, ssize_t ret, ssize_t res2) · 49 unmodified lines | ||
| 50 | 50 | io_req_set_res(req, ret, 0); |
| 51 | 51 | if (req->ctx->flags & IORING_SETUP_CQE32) |
| 52 | 52 | io_req_set_cqe32_extra(req, res2, 0); |
| 53 | − __io_req_complete(req, 0); | |
| 53 | + if (req->ctx->flags & IORING_SETUP_IOPOLL) | |
| 54 | + /* order with io_iopoll_req_issued() checking ->iopoll_complete */ | |
| 55 | + smp_store_release(&req->iopoll_completed, 1); | |
| 56 | + else | |
| 57 | + __io_req_complete(req, 0); | |
| 54 | 58 | } |
| 55 | 59 | EXPORT_SYMBOL_GPL(io_uring_cmd_done); |
| 56 | 60 | |
| @@ -97,8 +101,11 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags) · 40 unmodified lines | ||
| 97 | 101 | issue_flags |= IO_URING_F_SQE128; |
| 98 | 102 | if (ctx->flags & IORING_SETUP_CQE32) |
| 99 | 103 | issue_flags |= IO_URING_F_CQE32; |
| 100 | − if (ctx->flags & IORING_SETUP_IOPOLL) | |
| 104 | + if (ctx->flags & IORING_SETUP_IOPOLL) { | |
| 101 | 105 | issue_flags |= IO_URING_F_IOPOLL; |
| 106 | + req->iopoll_completed = 0; | |
| 107 | + WRITE_ONCE(ioucmd->cookie, NULL); | |
| 108 | + } | |
| 102 | 109 | |
| 103 | 110 | if (req_has_async_data(req)) |
| 104 | 111 | ioucmd->cmd = req->async_data; |
| @@ -400,7 +400,6 @@ static void do_tx(int domain, int type, int protocol) · 399 unmodified lines | ||
| 400 | 400 | cfg_payload_len, msg_flags); |
| 401 | 401 | sqe->user_data = NONZC_TAG; |
| 402 | 402 | } else { |
| 403 | − compl_cqes++; | |
| 404 | 403 | io_uring_prep_sendzc(sqe, fd, payload, |
| 405 | 404 | cfg_payload_len, |
| 406 | 405 | msg_flags, zc_flags); |
| @@ -430,18 +429,23 @@ static void do_tx(int domain, int type, int protocol) · 23 unmodified lines | ||
| 430 | 429 | if (cqe->flags & IORING_CQE_F_NOTIF) { |
| 431 | 430 | if (cqe->flags & IORING_CQE_F_MORE) |
| 432 | 431 | error(1, -EINVAL, "invalid notif flags"); |
| 432 | + if (compl_cqes <= 0) | |
| 433 | + error(1, -EINVAL, "notification mismatch"); | |
| 433 | 434 | compl_cqes--; |
| 434 | 435 | i--; |
| 435 | − } else if (cqe->res <= 0) { | |
| 436 | − if (cqe->flags & IORING_CQE_F_MORE) | |
| 437 | − error(1, cqe->res, "more with a failed send"); | |
| 438 | − error(1, cqe->res, "send failed"); | |
| 439 | − } else { | |
| 440 | − if (cqe->user_data == ZC_TAG && | |
| 441 | − !(cqe->flags & IORING_CQE_F_MORE)) | |
| 442 | − error(1, cqe->res, "missing more flag"); | |
| 436 | + io_uring_cqe_seen(&ring); | |
| 437 | + continue; | |
| 438 | + } | |
| 439 | + if (cqe->flags & IORING_CQE_F_MORE) { | |
| 440 | + if (cqe->user_data != ZC_TAG) | |
| 441 | + error(1, cqe->res, "unexpected F_MORE"); | |
| 442 | + compl_cqes++; | |
| 443 | + } | |
| 444 | + if (cqe->res >= 0) { | |
| 443 | 445 | packets++; |
| 444 | 446 | bytes += cqe->res; |
| 447 | + } else if (cqe->res != -EAGAIN) { | |
| 448 | + error(1, cqe->res, "send failed"); | |
| 445 | 449 | } |
| 446 | 450 | io_uring_cqe_seen(&ring); |
| 447 | 451 | } |