| 1 | /* SPDX-License-Identifier: GPL-2.0-or-later */ |
| 2 | /* |
| 3 | * poll(2) file descriptor monitoring |
| 4 | * |
| 5 | * Uses ppoll(2) when available, g_poll() otherwise. |
| 6 | */ |
| 7 | |
| 8 | #include "qemu/osdep.h" |
| 9 | #include "aio-posix.h" |
| 10 | #include "qemu/rcu_queue.h" |
| 11 | |
| 12 | /* |
| 13 | * These thread-local variables are used only in fdmon_poll_wait() around the |
| 14 | * call to the poll() system call. In particular they are not used while |
| 15 | * aio_poll is performing callbacks, which makes it much easier to think about |
| 16 | * reentrancy! |
| 17 | * |
| 18 | * Stack-allocated arrays would be perfect but they have size limitations; |
| 19 | * heap allocation is expensive enough that we want to reuse arrays across |
| 20 | * calls to aio_poll(). And because poll() has to be called without holding |
| 21 | * any lock, the arrays cannot be stored in AioContext. Thread-local data |
| 22 | * has none of the disadvantages of these three options. |
| 23 | */ |
| 24 | static __thread GPollFD *pollfds; |
| 25 | static __thread AioHandler **nodes; |
| 26 | static __thread unsigned npfd, nalloc; |
| 27 | static __thread Notifier pollfds_cleanup_notifier; |
| 28 | |
| 29 | static void pollfds_cleanup(Notifier *n, void *unused) |
| 30 | { |
| 31 | g_assert(npfd == 0); |
| 32 | g_free(pollfds); |
| 33 | g_free(nodes); |
| 34 | nalloc = 0; |
| 35 | } |
| 36 | |
| 37 | static void add_pollfd(AioHandler *node) |
| 38 | { |
| 39 | if (npfd == nalloc) { |
| 40 | if (nalloc == 0) { |
| 41 | pollfds_cleanup_notifier.notify = pollfds_cleanup; |
| 42 | qemu_thread_atexit_add(&pollfds_cleanup_notifier); |
| 43 | nalloc = 8; |
| 44 | } else { |
| 45 | g_assert(nalloc <= INT_MAX); |
| 46 | nalloc *= 2; |
| 47 | } |
| 48 | pollfds = g_renew(GPollFD, pollfds, nalloc); |
| 49 | nodes = g_renew(AioHandler *, nodes, nalloc); |
| 50 | } |
| 51 | nodes[npfd] = node; |
| 52 | pollfds[npfd] = (GPollFD) { |
| 53 | .fd = node->pfd.fd, |
| 54 | .events = node->pfd.events, |
| 55 | }; |
| 56 | npfd++; |
| 57 | } |
| 58 | |
| 59 | static int fdmon_poll_wait(AioContext *ctx, AioHandlerList *ready_list, |
| 60 | int64_t timeout) |
| 61 | { |
| 62 | AioHandler *node; |
| 63 | int ret; |
| 64 | |
| 65 | assert(npfd == 0); |
| 66 | |
| 67 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 68 | if (!QLIST_IS_INSERTED(node, node_deleted) && node->pfd.events) { |
| 69 | add_pollfd(node); |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | /* epoll(7) is faster above a certain number of fds */ |
| 74 | if (fdmon_epoll_try_upgrade(ctx, npfd)) { |
| 75 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 76 | if (!QLIST_IS_INSERTED(node, node_deleted) && node->pfd.events) { |
| 77 | g_source_remove_poll(&ctx->source, &node->pfd); |
| 78 | } |
| 79 | } |
| 80 | npfd = 0; /* we won't need pollfds[], reset npfd */ |
| 81 | return ctx->fdmon_ops->wait(ctx, ready_list, timeout); |
| 82 | } |
| 83 | |
| 84 | ret = qemu_poll_ns(pollfds, npfd, timeout); |
| 85 | if (ret > 0) { |
| 86 | int i; |
| 87 | |
| 88 | for (i = 0; i < npfd; i++) { |
| 89 | int revents = pollfds[i].revents; |
| 90 | |
| 91 | if (revents) { |
| 92 | aio_add_ready_handler(ready_list, nodes[i], revents); |
| 93 | } |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | npfd = 0; |
| 98 | return ret; |
| 99 | } |
| 100 | |
| 101 | static void fdmon_poll_update(AioContext *ctx, |
| 102 | AioHandler *old_node, |
| 103 | AioHandler *new_node) |
| 104 | { |
| 105 | if (old_node) { |
| 106 | /* |
| 107 | * If the GSource is in the process of being destroyed then |
| 108 | * g_source_remove_poll() causes an assertion failure. Skip removal in |
| 109 | * that case, because glib cleans up its state during destruction |
| 110 | * anyway. |
| 111 | */ |
| 112 | if (!g_source_is_destroyed(&ctx->source)) { |
| 113 | g_source_remove_poll(&ctx->source, &old_node->pfd); |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | if (new_node) { |
| 118 | g_source_add_poll(&ctx->source, &new_node->pfd); |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | static void fdmon_poll_gsource_prepare(AioContext *ctx) |
| 123 | { |
| 124 | /* Do nothing */ |
| 125 | } |
| 126 | |
| 127 | static bool fdmon_poll_gsource_check(AioContext *ctx) |
| 128 | { |
| 129 | AioHandler *node; |
| 130 | bool result = false; |
| 131 | |
| 132 | /* |
| 133 | * We have to walk very carefully in case aio_set_fd_handler is |
| 134 | * called while we're walking. |
| 135 | */ |
| 136 | qemu_lockcnt_inc(&ctx->list_lock); |
| 137 | |
| 138 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 139 | int revents = node->pfd.revents & node->pfd.events; |
| 140 | |
| 141 | if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) { |
| 142 | result = true; |
| 143 | break; |
| 144 | } |
| 145 | if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) { |
| 146 | result = true; |
| 147 | break; |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | qemu_lockcnt_dec(&ctx->list_lock); |
| 152 | |
| 153 | return result; |
| 154 | } |
| 155 | |
| 156 | static void fdmon_poll_gsource_dispatch(AioContext *ctx, |
| 157 | AioHandlerList *ready_list) |
| 158 | { |
| 159 | AioHandler *node; |
| 160 | |
| 161 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 162 | int revents = node->pfd.revents; |
| 163 | |
| 164 | if (revents) { |
| 165 | aio_add_ready_handler(ready_list, node, revents); |
| 166 | } |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | const FDMonOps fdmon_poll_ops = { |
| 171 | .update = fdmon_poll_update, |
| 172 | .wait = fdmon_poll_wait, |
| 173 | .need_wait = aio_poll_disabled, |
| 174 | .gsource_prepare = fdmon_poll_gsource_prepare, |
| 175 | .gsource_check = fdmon_poll_gsource_check, |
| 176 | .gsource_dispatch = fdmon_poll_gsource_dispatch, |
| 177 | }; |
| 178 | |
| 179 | void fdmon_poll_downgrade(AioContext *ctx) |
| 180 | { |
| 181 | AioHandler *node; |
| 182 | |
| 183 | ctx->fdmon_ops = &fdmon_poll_ops; |
| 184 | |
| 185 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 186 | if (!QLIST_IS_INSERTED(node, node_deleted) && node->pfd.events) { |
| 187 | g_source_add_poll(&ctx->source, &node->pfd); |
| 188 | } |
| 189 | } |
| 190 | } |