Jens Axboe | 21b4aa5 | 2019-03-06 09:03:50 -0700 | [diff] [blame] | 1 | #include <sys/types.h> |
| 2 | #include <sys/stat.h> |
| 3 | #include <sys/mman.h> |
| 4 | #include <unistd.h> |
| 5 | #include <errno.h> |
| 6 | #include <string.h> |
| 7 | |
| 8 | #include "liburing.h" |
| 9 | #include "barrier.h" |
| 10 | |
| 11 | static int __io_uring_get_completion(struct io_uring *ring, |
| 12 | struct io_uring_cqe **cqe_ptr, int wait) |
| 13 | { |
| 14 | struct io_uring_cq *cq = &ring->cq; |
| 15 | const unsigned mask = *cq->kring_mask; |
| 16 | unsigned head; |
| 17 | int ret; |
| 18 | |
| 19 | *cqe_ptr = NULL; |
| 20 | head = *cq->khead; |
| 21 | do { |
| 22 | /* |
| 23 | * It's necessary to use a read_barrier() before reading |
| 24 | * the CQ tail, since the kernel updates it locklessly. The |
| 25 | * kernel has the matching store barrier for the update. The |
| 26 | * kernel also ensures that previous stores to CQEs are ordered |
| 27 | * with the tail update. |
| 28 | */ |
| 29 | read_barrier(); |
| 30 | if (head != *cq->ktail) { |
| 31 | *cqe_ptr = &cq->cqes[head & mask]; |
| 32 | break; |
| 33 | } |
| 34 | if (!wait) |
| 35 | break; |
| 36 | ret = io_uring_enter(ring->ring_fd, 0, 1, |
| 37 | IORING_ENTER_GETEVENTS, NULL); |
| 38 | if (ret < 0) |
| 39 | return -errno; |
| 40 | } while (1); |
| 41 | |
| 42 | if (*cqe_ptr) { |
| 43 | *cq->khead = head + 1; |
| 44 | /* |
| 45 | * Ensure that the kernel sees our new head, the kernel has |
| 46 | * the matching read barrier. |
| 47 | */ |
| 48 | write_barrier(); |
| 49 | } |
| 50 | |
| 51 | return 0; |
| 52 | } |
| 53 | |
| 54 | /* |
| 55 | * Return an IO completion, if one is readily available |
| 56 | */ |
| 57 | int io_uring_get_completion(struct io_uring *ring, |
| 58 | struct io_uring_cqe **cqe_ptr) |
| 59 | { |
| 60 | return __io_uring_get_completion(ring, cqe_ptr, 0); |
| 61 | } |
| 62 | |
| 63 | /* |
| 64 | * Return an IO completion, waiting for it if necessary |
| 65 | */ |
| 66 | int io_uring_wait_completion(struct io_uring *ring, |
| 67 | struct io_uring_cqe **cqe_ptr) |
| 68 | { |
| 69 | return __io_uring_get_completion(ring, cqe_ptr, 1); |
| 70 | } |
| 71 | |
| 72 | /* |
| 73 | * Submit sqes acquired from io_uring_get_sqe() to the kernel. |
| 74 | * |
| 75 | * Returns number of sqes submitted |
| 76 | */ |
| 77 | int io_uring_submit(struct io_uring *ring) |
| 78 | { |
| 79 | struct io_uring_sq *sq = &ring->sq; |
| 80 | const unsigned mask = *sq->kring_mask; |
| 81 | unsigned ktail, ktail_next, submitted; |
| 82 | int ret; |
| 83 | |
| 84 | /* |
| 85 | * If we have pending IO in the kring, submit it first. We need a |
| 86 | * read barrier here to match the kernels store barrier when updating |
| 87 | * the SQ head. |
| 88 | */ |
| 89 | read_barrier(); |
| 90 | if (*sq->khead != *sq->ktail) { |
| 91 | submitted = *sq->kring_entries; |
| 92 | goto submit; |
| 93 | } |
| 94 | |
| 95 | if (sq->sqe_head == sq->sqe_tail) |
| 96 | return 0; |
| 97 | |
| 98 | /* |
| 99 | * Fill in sqes that we have queued up, adding them to the kernel ring |
| 100 | */ |
| 101 | submitted = 0; |
| 102 | ktail = ktail_next = *sq->ktail; |
| 103 | while (sq->sqe_head < sq->sqe_tail) { |
| 104 | ktail_next++; |
| 105 | read_barrier(); |
| 106 | |
| 107 | sq->array[ktail & mask] = sq->sqe_head & mask; |
| 108 | ktail = ktail_next; |
| 109 | |
| 110 | sq->sqe_head++; |
| 111 | submitted++; |
| 112 | } |
| 113 | |
| 114 | if (!submitted) |
| 115 | return 0; |
| 116 | |
| 117 | if (*sq->ktail != ktail) { |
| 118 | /* |
| 119 | * First write barrier ensures that the SQE stores are updated |
| 120 | * with the tail update. This is needed so that the kernel |
| 121 | * will never see a tail update without the preceeding sQE |
| 122 | * stores being done. |
| 123 | */ |
| 124 | write_barrier(); |
| 125 | *sq->ktail = ktail; |
| 126 | /* |
| 127 | * The kernel has the matching read barrier for reading the |
| 128 | * SQ tail. |
| 129 | */ |
| 130 | write_barrier(); |
| 131 | } |
| 132 | |
| 133 | submit: |
| 134 | ret = io_uring_enter(ring->ring_fd, submitted, 0, |
| 135 | IORING_ENTER_GETEVENTS, NULL); |
| 136 | if (ret < 0) |
| 137 | return -errno; |
| 138 | |
| 139 | return 0; |
| 140 | } |
| 141 | |
| 142 | /* |
| 143 | * Return an sqe to fill. Application must later call io_uring_submit() |
| 144 | * when it's ready to tell the kernel about it. The caller may call this |
| 145 | * function multiple times before calling io_uring_submit(). |
| 146 | * |
| 147 | * Returns a vacant sqe, or NULL if we're full. |
| 148 | */ |
| 149 | struct io_uring_sqe *io_uring_get_sqe(struct io_uring *ring) |
| 150 | { |
| 151 | struct io_uring_sq *sq = &ring->sq; |
| 152 | unsigned next = sq->sqe_tail + 1; |
| 153 | struct io_uring_sqe *sqe; |
| 154 | |
| 155 | /* |
| 156 | * All sqes are used |
| 157 | */ |
| 158 | if (next - sq->sqe_head > *sq->kring_entries) |
| 159 | return NULL; |
| 160 | |
| 161 | sqe = &sq->sqes[sq->sqe_tail & *sq->kring_mask]; |
| 162 | sq->sqe_tail = next; |
| 163 | return sqe; |
| 164 | } |