Paolo Bonzini | f42b220 | 2012-06-09 04:01:51 +0200 | [diff] [blame^] | 1 | /* |
| 2 | * QEMU aio implementation |
| 3 | * |
| 4 | * Copyright IBM Corp., 2008 |
| 5 | * Copyright Red Hat Inc., 2012 |
| 6 | * |
| 7 | * Authors: |
| 8 | * Anthony Liguori <aliguori@us.ibm.com> |
| 9 | * Paolo Bonzini <pbonzini@redhat.com> |
| 10 | * |
| 11 | * This work is licensed under the terms of the GNU GPL, version 2. See |
| 12 | * the COPYING file in the top-level directory. |
| 13 | * |
| 14 | * Contributions after 2012-01-13 are licensed under the terms of the |
| 15 | * GNU GPL, version 2 or (at your option) any later version. |
| 16 | */ |
| 17 | |
| 18 | #include "qemu-common.h" |
| 19 | #include "block.h" |
| 20 | #include "qemu-queue.h" |
| 21 | #include "qemu_socket.h" |
| 22 | |
| 23 | struct AioHandler { |
| 24 | EventNotifier *e; |
| 25 | EventNotifierHandler *io_notify; |
| 26 | AioFlushEventNotifierHandler *io_flush; |
| 27 | GPollFD pfd; |
| 28 | int deleted; |
| 29 | QLIST_ENTRY(AioHandler) node; |
| 30 | }; |
| 31 | |
| 32 | void aio_set_event_notifier(AioContext *ctx, |
| 33 | EventNotifier *e, |
| 34 | EventNotifierHandler *io_notify, |
| 35 | AioFlushEventNotifierHandler *io_flush) |
| 36 | { |
| 37 | AioHandler *node; |
| 38 | |
| 39 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { |
| 40 | if (node->e == e && !node->deleted) { |
| 41 | break; |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | /* Are we deleting the fd handler? */ |
| 46 | if (!io_notify) { |
| 47 | if (node) { |
| 48 | /* If the lock is held, just mark the node as deleted */ |
| 49 | if (ctx->walking_handlers) { |
| 50 | node->deleted = 1; |
| 51 | node->pfd.revents = 0; |
| 52 | } else { |
| 53 | /* Otherwise, delete it for real. We can't just mark it as |
| 54 | * deleted because deleted nodes are only cleaned up after |
| 55 | * releasing the walking_handlers lock. |
| 56 | */ |
| 57 | QLIST_REMOVE(node, node); |
| 58 | g_free(node); |
| 59 | } |
| 60 | } |
| 61 | } else { |
| 62 | if (node == NULL) { |
| 63 | /* Alloc and insert if it's not already there */ |
| 64 | node = g_malloc0(sizeof(AioHandler)); |
| 65 | node->e = e; |
| 66 | node->pfd.fd = (uintptr_t)event_notifier_get_handle(e); |
| 67 | node->pfd.events = G_IO_IN; |
| 68 | QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node); |
| 69 | } |
| 70 | /* Update handler with latest information */ |
| 71 | node->io_notify = io_notify; |
| 72 | node->io_flush = io_flush; |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | bool aio_pending(AioContext *ctx) |
| 77 | { |
| 78 | AioHandler *node; |
| 79 | |
| 80 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { |
| 81 | if (node->pfd.revents && node->io_notify) { |
| 82 | return true; |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | return false; |
| 87 | } |
| 88 | |
| 89 | bool aio_poll(AioContext *ctx, bool blocking) |
| 90 | { |
| 91 | AioHandler *node; |
| 92 | HANDLE events[MAXIMUM_WAIT_OBJECTS + 1]; |
| 93 | bool busy, progress; |
| 94 | int count; |
| 95 | |
| 96 | progress = false; |
| 97 | |
| 98 | /* |
| 99 | * If there are callbacks left that have been queued, we need to call then. |
| 100 | * Do not call select in this case, because it is possible that the caller |
| 101 | * does not need a complete flush (as is the case for qemu_aio_wait loops). |
| 102 | */ |
| 103 | if (aio_bh_poll(ctx)) { |
| 104 | blocking = false; |
| 105 | progress = true; |
| 106 | } |
| 107 | |
| 108 | /* |
| 109 | * Then dispatch any pending callbacks from the GSource. |
| 110 | * |
| 111 | * We have to walk very carefully in case qemu_aio_set_fd_handler is |
| 112 | * called while we're walking. |
| 113 | */ |
| 114 | node = QLIST_FIRST(&ctx->aio_handlers); |
| 115 | while (node) { |
| 116 | AioHandler *tmp; |
| 117 | |
| 118 | ctx->walking_handlers++; |
| 119 | |
| 120 | if (node->pfd.revents && node->io_notify) { |
| 121 | node->pfd.revents = 0; |
| 122 | node->io_notify(node->e); |
| 123 | progress = true; |
| 124 | } |
| 125 | |
| 126 | tmp = node; |
| 127 | node = QLIST_NEXT(node, node); |
| 128 | |
| 129 | ctx->walking_handlers--; |
| 130 | |
| 131 | if (!ctx->walking_handlers && tmp->deleted) { |
| 132 | QLIST_REMOVE(tmp, node); |
| 133 | g_free(tmp); |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | if (progress && !blocking) { |
| 138 | return true; |
| 139 | } |
| 140 | |
| 141 | ctx->walking_handlers++; |
| 142 | |
| 143 | /* fill fd sets */ |
| 144 | busy = false; |
| 145 | count = 0; |
| 146 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { |
| 147 | /* If there aren't pending AIO operations, don't invoke callbacks. |
| 148 | * Otherwise, if there are no AIO requests, qemu_aio_wait() would |
| 149 | * wait indefinitely. |
| 150 | */ |
| 151 | if (!node->deleted && node->io_flush) { |
| 152 | if (node->io_flush(node->e) == 0) { |
| 153 | continue; |
| 154 | } |
| 155 | busy = true; |
| 156 | } |
| 157 | if (!node->deleted && node->io_notify) { |
| 158 | events[count++] = event_notifier_get_handle(node->e); |
| 159 | } |
| 160 | } |
| 161 | |
| 162 | ctx->walking_handlers--; |
| 163 | |
| 164 | /* No AIO operations? Get us out of here */ |
| 165 | if (!busy) { |
| 166 | return progress; |
| 167 | } |
| 168 | |
| 169 | /* wait until next event */ |
| 170 | for (;;) { |
| 171 | int timeout = blocking ? INFINITE : 0; |
| 172 | int ret = WaitForMultipleObjects(count, events, FALSE, timeout); |
| 173 | |
| 174 | /* if we have any signaled events, dispatch event */ |
| 175 | if ((DWORD) (ret - WAIT_OBJECT_0) >= count) { |
| 176 | break; |
| 177 | } |
| 178 | |
| 179 | blocking = false; |
| 180 | |
| 181 | /* we have to walk very carefully in case |
| 182 | * qemu_aio_set_fd_handler is called while we're walking */ |
| 183 | node = QLIST_FIRST(&ctx->aio_handlers); |
| 184 | while (node) { |
| 185 | AioHandler *tmp; |
| 186 | |
| 187 | ctx->walking_handlers++; |
| 188 | |
| 189 | if (!node->deleted && |
| 190 | event_notifier_get_handle(node->e) == events[ret - WAIT_OBJECT_0] && |
| 191 | node->io_notify) { |
| 192 | node->io_notify(node->e); |
| 193 | progress = true; |
| 194 | } |
| 195 | |
| 196 | tmp = node; |
| 197 | node = QLIST_NEXT(node, node); |
| 198 | |
| 199 | ctx->walking_handlers--; |
| 200 | |
| 201 | if (!ctx->walking_handlers && tmp->deleted) { |
| 202 | QLIST_REMOVE(tmp, node); |
| 203 | g_free(tmp); |
| 204 | } |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | return progress; |
| 209 | } |