Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1 | /* |
| 2 | * |
| 3 | * Copyright 2016, Google Inc. |
| 4 | * All rights reserved. |
| 5 | * |
| 6 | * Redistribution and use in source and binary forms, with or without |
| 7 | * modification, are permitted provided that the following conditions are |
| 8 | * met: |
| 9 | * |
| 10 | * * Redistributions of source code must retain the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer. |
| 12 | * * Redistributions in binary form must reproduce the above |
| 13 | * copyright notice, this list of conditions and the following disclaimer |
| 14 | * in the documentation and/or other materials provided with the |
| 15 | * distribution. |
| 16 | * * Neither the name of Google Inc. nor the names of its |
| 17 | * contributors may be used to endorse or promote products derived from |
| 18 | * this software without specific prior written permission. |
| 19 | * |
| 20 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 21 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 22 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 23 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT |
| 24 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
| 25 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT |
| 26 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
| 27 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
| 28 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 29 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 30 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 31 | * |
| 32 | */ |
| 33 | |
murgatroid99 | 5407089 | 2016-08-08 17:01:18 -0700 | [diff] [blame] | 34 | #include "src/core/lib/iomgr/port.h" |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 35 | |
Sree Kuchibhotla | 76a0795 | 2016-06-22 15:09:06 -0700 | [diff] [blame] | 36 | /* This polling engine is only relevant on linux kernels supporting epoll() */ |
murgatroid99 | 623dd4f | 2016-08-08 17:31:27 -0700 | [diff] [blame] | 37 | #ifdef GRPC_LINUX_EPOLL |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 38 | |
Sree Kuchibhotla | 4c11a20 | 2016-06-06 09:23:25 -0700 | [diff] [blame] | 39 | #include "src/core/lib/iomgr/ev_epoll_linux.h" |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 40 | |
| 41 | #include <assert.h> |
| 42 | #include <errno.h> |
| 43 | #include <poll.h> |
Sree Kuchibhotla | 4998e30 | 2016-08-16 07:26:31 -0700 | [diff] [blame] | 44 | #include <pthread.h> |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 45 | #include <signal.h> |
| 46 | #include <string.h> |
| 47 | #include <sys/epoll.h> |
| 48 | #include <sys/socket.h> |
| 49 | #include <unistd.h> |
| 50 | |
| 51 | #include <grpc/support/alloc.h> |
| 52 | #include <grpc/support/log.h> |
| 53 | #include <grpc/support/string_util.h> |
| 54 | #include <grpc/support/tls.h> |
| 55 | #include <grpc/support/useful.h> |
| 56 | |
| 57 | #include "src/core/lib/iomgr/ev_posix.h" |
| 58 | #include "src/core/lib/iomgr/iomgr_internal.h" |
| 59 | #include "src/core/lib/iomgr/wakeup_fd_posix.h" |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 60 | #include "src/core/lib/iomgr/workqueue.h" |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 61 | #include "src/core/lib/profiling/timers.h" |
| 62 | #include "src/core/lib/support/block_annotate.h" |
| 63 | |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 64 | /* TODO: sreek - Move this to init.c and initialize this like other tracers. */ |
| 65 | static int grpc_polling_trace = 0; /* Disabled by default */ |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 66 | #define GRPC_POLLING_TRACE(fmt, ...) \ |
| 67 | if (grpc_polling_trace) { \ |
| 68 | gpr_log(GPR_INFO, (fmt), __VA_ARGS__); \ |
| 69 | } |
| 70 | |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 71 | /* Uncomment the following to enable extra checks on poll_object operations */ |
Sree Kuchibhotla | e6f516e | 2016-12-08 12:20:23 -0800 | [diff] [blame] | 72 | /* #define PO_DEBUG */ |
| 73 | |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 74 | static int grpc_wakeup_signal = -1; |
| 75 | static bool is_grpc_wakeup_signal_initialized = false; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 76 | |
Craig Tiller | b4b8e1e | 2016-11-28 07:33:13 -0800 | [diff] [blame] | 77 | /* TODO: sreek: Right now, this wakes up all pollers. In future we should make |
| 78 | * sure to wake up one polling thread (which can wake up other threads if |
| 79 | * needed) */ |
| 80 | static grpc_wakeup_fd global_wakeup_fd; |
| 81 | |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 82 | /* Implements the function defined in grpc_posix.h. This function might be |
| 83 | * called before even calling grpc_init() to set either a different signal to |
| 84 | * use. If signum == -1, then the use of signals is disabled */ |
| 85 | void grpc_use_signal(int signum) { |
| 86 | grpc_wakeup_signal = signum; |
| 87 | is_grpc_wakeup_signal_initialized = true; |
| 88 | |
| 89 | if (grpc_wakeup_signal < 0) { |
| 90 | gpr_log(GPR_INFO, |
| 91 | "Use of signals is disabled. Epoll engine will not be used"); |
| 92 | } else { |
| 93 | gpr_log(GPR_INFO, "epoll engine will be using signal: %d", |
| 94 | grpc_wakeup_signal); |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | struct polling_island; |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 99 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 100 | typedef enum { |
| 101 | POLL_OBJ_FD, |
| 102 | POLL_OBJ_POLLSET, |
| 103 | POLL_OBJ_POLLSET_SET |
| 104 | } poll_obj_type; |
| 105 | |
| 106 | typedef struct poll_obj { |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 107 | #ifdef PO_DEBUG |
| 108 | poll_obj_type obj_type; |
| 109 | #endif |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 110 | gpr_mu mu; |
| 111 | struct polling_island *pi; |
| 112 | } poll_obj; |
| 113 | |
| 114 | const char *poll_obj_string(poll_obj_type po_type) { |
| 115 | switch (po_type) { |
| 116 | case POLL_OBJ_FD: |
| 117 | return "fd"; |
| 118 | case POLL_OBJ_POLLSET: |
| 119 | return "pollset"; |
| 120 | case POLL_OBJ_POLLSET_SET: |
| 121 | return "pollset_set"; |
| 122 | } |
| 123 | |
| 124 | GPR_UNREACHABLE_CODE(return "UNKNOWN"); |
| 125 | } |
| 126 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 127 | /******************************************************************************* |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 128 | * Fd Declarations |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 129 | */ |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 130 | |
| 131 | #define FD_FROM_PO(po) ((grpc_fd *)(po)) |
| 132 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 133 | struct grpc_fd { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 134 | poll_obj po; |
| 135 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 136 | int fd; |
| 137 | /* refst format: |
Sree Kuchibhotla | 5098f91 | 2016-05-31 10:58:17 -0700 | [diff] [blame] | 138 | bit 0 : 1=Active / 0=Orphaned |
| 139 | bits 1-n : refcount |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 140 | Ref/Unref by two to avoid altering the orphaned bit */ |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 141 | gpr_atm refst; |
| 142 | |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 143 | /* Internally stores data of type (grpc_error *). If the FD is shutdown, this |
| 144 | contains reason for shutdown (i.e a pointer to grpc_error) ORed with |
| 145 | FD_SHUTDOWN_BIT. Since address allocations are word-aligned, the lower bit |
| 146 | of (grpc_error *) addresses is guaranteed to be zero. Even if the |
| 147 | (grpc_error *), is of special types like GRPC_ERROR_NONE, GRPC_ERROR_OOM |
| 148 | etc, the lower bit is guaranteed to be zero. |
| 149 | |
| 150 | Once an fd is shutdown, any pending or future read/write closures on the |
| 151 | fd should fail */ |
Sree Kuchibhotla | 9998338 | 2017-02-12 17:03:27 -0800 | [diff] [blame] | 152 | gpr_atm shutdown_error; |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 153 | |
Sree Kuchibhotla | 9998338 | 2017-02-12 17:03:27 -0800 | [diff] [blame] | 154 | /* The fd is either closed or we relinquished control of it. In either |
| 155 | cases, this indicates that the 'fd' on this structure is no longer |
| 156 | valid */ |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 157 | bool orphaned; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 158 | |
Sree Kuchibhotla | 9998338 | 2017-02-12 17:03:27 -0800 | [diff] [blame] | 159 | /* Closures to call when the fd is readable or writable respectively. These |
| 160 | fields contain one of the following values: |
| 161 | CLOSURE_READY : The fd has an I/O event of interest but there is no |
| 162 | closure yet to execute |
| 163 | |
| 164 | CLOSURE_NOT_READY : The fd has no I/O event of interest |
| 165 | |
Sree Kuchibhotla | a70ccb6 | 2017-02-13 23:16:52 -0800 | [diff] [blame] | 166 | closure ptr : The closure to be executed when the fd has an I/O |
| 167 | event of interest |
| 168 | |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 169 | shutdown_error | FD_SHUTDOWN_BIT : |
| 170 | 'shutdown_error' field ORed with FD_SHUTDOWN_BIT. |
Sree Kuchibhotla | a70ccb6 | 2017-02-13 23:16:52 -0800 | [diff] [blame] | 171 | This indicates that the fd is shutdown. Since all |
| 172 | memory allocations are word-aligned, the lower two |
| 173 | bits of the shutdown_error pointer are always 0. So |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 174 | it is safe to OR these with FD_SHUTDOWN_BIT |
Sree Kuchibhotla | 9998338 | 2017-02-12 17:03:27 -0800 | [diff] [blame] | 175 | |
| 176 | Valid state transitions: |
| 177 | |
| 178 | <closure ptr> <-----3------ CLOSURE_NOT_READY ----1----> CLOSURE_READY |
| 179 | | | ^ | ^ | | |
| 180 | | | | | | | | |
| 181 | | +--------------4----------+ 6 +---------2---------------+ | |
| 182 | | | | |
| 183 | | v | |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 184 | +-----5-------> [shutdown_error | FD_SHUTDOWN_BIT] <----7---------+ |
Sree Kuchibhotla | 9998338 | 2017-02-12 17:03:27 -0800 | [diff] [blame] | 185 | |
| 186 | For 1, 4 : See set_ready() function |
| 187 | For 2, 3 : See notify_on() function |
| 188 | For 5,6,7: See set_shutdown() function */ |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 189 | gpr_atm read_closure; |
| 190 | gpr_atm write_closure; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 191 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 192 | struct grpc_fd *freelist_next; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 193 | grpc_closure *on_done_closure; |
| 194 | |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 195 | /* The pollset that last noticed that the fd is readable. The actual type |
| 196 | * stored in this is (grpc_pollset *) */ |
| 197 | gpr_atm read_notifier_pollset; |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 198 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 199 | grpc_iomgr_object iomgr_object; |
| 200 | }; |
| 201 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 202 | /* Reference counting for fds */ |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 203 | // #define GRPC_FD_REF_COUNT_DEBUG |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 204 | #ifdef GRPC_FD_REF_COUNT_DEBUG |
| 205 | static void fd_ref(grpc_fd *fd, const char *reason, const char *file, int line); |
| 206 | static void fd_unref(grpc_fd *fd, const char *reason, const char *file, |
| 207 | int line); |
| 208 | #define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__) |
| 209 | #define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__) |
| 210 | #else |
| 211 | static void fd_ref(grpc_fd *fd); |
| 212 | static void fd_unref(grpc_fd *fd); |
| 213 | #define GRPC_FD_REF(fd, reason) fd_ref(fd) |
| 214 | #define GRPC_FD_UNREF(fd, reason) fd_unref(fd) |
| 215 | #endif |
| 216 | |
| 217 | static void fd_global_init(void); |
| 218 | static void fd_global_shutdown(void); |
| 219 | |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 220 | #define CLOSURE_NOT_READY ((gpr_atm)0) |
Sree Kuchibhotla | ff4b25d | 2017-02-16 15:07:11 -0800 | [diff] [blame] | 221 | #define CLOSURE_READY ((gpr_atm)2) |
Sree Kuchibhotla | a70ccb6 | 2017-02-13 23:16:52 -0800 | [diff] [blame] | 222 | |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 223 | #define FD_SHUTDOWN_BIT 1 |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 224 | |
| 225 | /******************************************************************************* |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 226 | * Polling island Declarations |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 227 | */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 228 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 229 | #ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 230 | |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 231 | #define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__) |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 232 | #define PI_UNREF(exec_ctx, p, r) \ |
| 233 | pi_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__) |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 234 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 235 | #else /* defined(GRPC_WORKQUEUE_REFCOUNT_DEBUG) */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 236 | |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 237 | #define PI_ADD_REF(p, r) pi_add_ref((p)) |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 238 | #define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p)) |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 239 | |
Yuchen Zeng | 362ac1b | 2016-09-13 16:01:31 -0700 | [diff] [blame] | 240 | #endif /* !defined(GRPC_PI_REF_COUNT_DEBUG) */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 241 | |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 242 | /* This is also used as grpc_workqueue (by directly casing it) */ |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 243 | typedef struct polling_island { |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 244 | grpc_closure_scheduler workqueue_scheduler; |
| 245 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 246 | gpr_mu mu; |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 247 | /* Ref count. Use PI_ADD_REF() and PI_UNREF() macros to increment/decrement |
| 248 | the refcount. |
| 249 | Once the ref count becomes zero, this structure is destroyed which means |
| 250 | we should ensure that there is never a scenario where a PI_ADD_REF() is |
| 251 | racing with a PI_UNREF() that just made the ref_count zero. */ |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 252 | gpr_atm ref_count; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 253 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 254 | /* Pointer to the polling_island this merged into. |
| 255 | * merged_to value is only set once in polling_island's lifetime (and that too |
| 256 | * only if the island is merged with another island). Because of this, we can |
| 257 | * use gpr_atm type here so that we can do atomic access on this and reduce |
| 258 | * lock contention on 'mu' mutex. |
| 259 | * |
| 260 | * Note that if this field is not NULL (i.e not 0), all the remaining fields |
| 261 | * (except mu and ref_count) are invalid and must be ignored. */ |
| 262 | gpr_atm merged_to; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 263 | |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 264 | /* Number of threads currently polling on this island */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 265 | gpr_atm poller_count; |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 266 | /* Mutex guarding the read end of the workqueue (must be held to pop from |
| 267 | * workqueue_items) */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 268 | gpr_mu workqueue_read_mu; |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 269 | /* Queue of closures to be executed */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 270 | gpr_mpscq workqueue_items; |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 271 | /* Count of items in workqueue_items */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 272 | gpr_atm workqueue_item_count; |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 273 | /* Wakeup fd used to wake pollers to check the contents of workqueue_items */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 274 | grpc_wakeup_fd workqueue_wakeup_fd; |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 275 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 276 | /* The fd of the underlying epoll set */ |
| 277 | int epoll_fd; |
| 278 | |
| 279 | /* The file descriptors in the epoll set */ |
| 280 | size_t fd_cnt; |
| 281 | size_t fd_capacity; |
| 282 | grpc_fd **fds; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 283 | } polling_island; |
| 284 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 285 | /******************************************************************************* |
| 286 | * Pollset Declarations |
| 287 | */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 288 | struct grpc_pollset_worker { |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 289 | /* Thread id of this worker */ |
| 290 | pthread_t pt_id; |
| 291 | |
| 292 | /* Used to prevent a worker from getting kicked multiple times */ |
| 293 | gpr_atm is_kicked; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 294 | struct grpc_pollset_worker *next; |
| 295 | struct grpc_pollset_worker *prev; |
| 296 | }; |
| 297 | |
| 298 | struct grpc_pollset { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 299 | poll_obj po; |
| 300 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 301 | grpc_pollset_worker root_worker; |
| 302 | bool kicked_without_pollers; |
| 303 | |
| 304 | bool shutting_down; /* Is the pollset shutting down ? */ |
| 305 | bool finish_shutdown_called; /* Is the 'finish_shutdown_locked()' called ? */ |
| 306 | grpc_closure *shutdown_done; /* Called after after shutdown is complete */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 307 | }; |
| 308 | |
| 309 | /******************************************************************************* |
| 310 | * Pollset-set Declarations |
| 311 | */ |
| 312 | struct grpc_pollset_set { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 313 | poll_obj po; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 314 | }; |
| 315 | |
| 316 | /******************************************************************************* |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 317 | * Common helpers |
| 318 | */ |
| 319 | |
Craig Tiller | f975f74 | 2016-07-01 14:56:27 -0700 | [diff] [blame] | 320 | static bool append_error(grpc_error **composite, grpc_error *error, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 321 | const char *desc) { |
Craig Tiller | f975f74 | 2016-07-01 14:56:27 -0700 | [diff] [blame] | 322 | if (error == GRPC_ERROR_NONE) return true; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 323 | if (*composite == GRPC_ERROR_NONE) { |
| 324 | *composite = GRPC_ERROR_CREATE(desc); |
| 325 | } |
| 326 | *composite = grpc_error_add_child(*composite, error); |
Craig Tiller | f975f74 | 2016-07-01 14:56:27 -0700 | [diff] [blame] | 327 | return false; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 328 | } |
| 329 | |
| 330 | /******************************************************************************* |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 331 | * Polling island Definitions |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 332 | */ |
| 333 | |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 334 | /* The wakeup fd that is used to wake up all threads in a Polling island. This |
| 335 | is useful in the polling island merge operation where we need to wakeup all |
| 336 | the threads currently polling the smaller polling island (so that they can |
| 337 | start polling the new/merged polling island) |
| 338 | |
| 339 | NOTE: This fd is initialized to be readable and MUST NOT be consumed i.e the |
| 340 | threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */ |
| 341 | static grpc_wakeup_fd polling_island_wakeup_fd; |
| 342 | |
Craig Tiller | 2e62013 | 2016-10-10 15:27:44 -0700 | [diff] [blame] | 343 | /* The polling island being polled right now. |
| 344 | See comments in workqueue_maybe_wakeup for why this is tracked. */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 345 | static __thread polling_island *g_current_thread_polling_island; |
| 346 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 347 | /* Forward declaration */ |
Craig Tiller | 2b49ea9 | 2016-07-01 13:21:27 -0700 | [diff] [blame] | 348 | static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi); |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 349 | static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure, |
| 350 | grpc_error *error); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 351 | |
Sree Kuchibhotla | ad2c477 | 2016-06-13 19:06:54 -0700 | [diff] [blame] | 352 | #ifdef GRPC_TSAN |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 353 | /* Currently TSAN may incorrectly flag data races between epoll_ctl and |
| 354 | epoll_wait for any grpc_fd structs that are added to the epoll set via |
| 355 | epoll_ctl and are returned (within a very short window) via epoll_wait(). |
| 356 | |
| 357 | To work-around this race, we establish a happens-before relation between |
| 358 | the code just-before epoll_ctl() and the code after epoll_wait() by using |
| 359 | this atomic */ |
| 360 | gpr_atm g_epoll_sync; |
Sree Kuchibhotla | ad2c477 | 2016-06-13 19:06:54 -0700 | [diff] [blame] | 361 | #endif /* defined(GRPC_TSAN) */ |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 362 | |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 363 | static const grpc_closure_scheduler_vtable workqueue_scheduler_vtable = { |
Craig Tiller | 7c70b6c | 2017-01-23 07:48:42 -0800 | [diff] [blame] | 364 | workqueue_enqueue, workqueue_enqueue, "workqueue"}; |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 365 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 366 | static void pi_add_ref(polling_island *pi); |
| 367 | static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 368 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 369 | #ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG |
Craig Tiller | a10b0b1 | 2016-09-09 16:20:07 -0700 | [diff] [blame] | 370 | static void pi_add_ref_dbg(polling_island *pi, const char *reason, |
| 371 | const char *file, int line) { |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 372 | long old_cnt = gpr_atm_acq_load(&pi->ref_count); |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 373 | pi_add_ref(pi); |
| 374 | gpr_log(GPR_DEBUG, "Add ref pi: %p, old: %ld -> new:%ld (%s) - (%s, %d)", |
| 375 | (void *)pi, old_cnt, old_cnt + 1, reason, file, line); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 376 | } |
| 377 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 378 | static void pi_unref_dbg(grpc_exec_ctx *exec_ctx, polling_island *pi, |
Craig Tiller | a10b0b1 | 2016-09-09 16:20:07 -0700 | [diff] [blame] | 379 | const char *reason, const char *file, int line) { |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 380 | long old_cnt = gpr_atm_acq_load(&pi->ref_count); |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 381 | pi_unref(exec_ctx, pi); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 382 | gpr_log(GPR_DEBUG, "Unref pi: %p, old:%ld -> new:%ld (%s) - (%s, %d)", |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 383 | (void *)pi, old_cnt, (old_cnt - 1), reason, file, line); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 384 | } |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 385 | |
| 386 | static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue, |
| 387 | const char *file, int line, |
| 388 | const char *reason) { |
| 389 | if (workqueue != NULL) { |
Craig Tiller | a10b0b1 | 2016-09-09 16:20:07 -0700 | [diff] [blame] | 390 | pi_add_ref_dbg((polling_island *)workqueue, reason, file, line); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 391 | } |
| 392 | return workqueue; |
| 393 | } |
| 394 | |
| 395 | static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, |
| 396 | const char *file, int line, const char *reason) { |
| 397 | if (workqueue != NULL) { |
Craig Tiller | a10b0b1 | 2016-09-09 16:20:07 -0700 | [diff] [blame] | 398 | pi_unref_dbg(exec_ctx, (polling_island *)workqueue, reason, file, line); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 399 | } |
| 400 | } |
| 401 | #else |
| 402 | static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) { |
| 403 | if (workqueue != NULL) { |
| 404 | pi_add_ref((polling_island *)workqueue); |
| 405 | } |
| 406 | return workqueue; |
| 407 | } |
| 408 | |
| 409 | static void workqueue_unref(grpc_exec_ctx *exec_ctx, |
| 410 | grpc_workqueue *workqueue) { |
| 411 | if (workqueue != NULL) { |
| 412 | pi_unref(exec_ctx, (polling_island *)workqueue); |
| 413 | } |
| 414 | } |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 415 | #endif |
| 416 | |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 417 | static void pi_add_ref(polling_island *pi) { |
| 418 | gpr_atm_no_barrier_fetch_add(&pi->ref_count, 1); |
| 419 | } |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 420 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 421 | static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) { |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 422 | /* If ref count went to zero, delete the polling island. |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 423 | Note that this deletion not be done under a lock. Once the ref count goes |
| 424 | to zero, we are guaranteed that no one else holds a reference to the |
| 425 | polling island (and that there is no racing pi_add_ref() call either). |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 426 | |
| 427 | Also, if we are deleting the polling island and the merged_to field is |
| 428 | non-empty, we should remove a ref to the merged_to polling island |
| 429 | */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 430 | if (1 == gpr_atm_full_fetch_add(&pi->ref_count, -1)) { |
| 431 | polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); |
| 432 | polling_island_delete(exec_ctx, pi); |
| 433 | if (next != NULL) { |
| 434 | PI_UNREF(exec_ctx, next, "pi_delete"); /* Recursive call */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 435 | } |
| 436 | } |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 437 | } |
| 438 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 439 | /* The caller is expected to hold pi->mu lock before calling this function */ |
| 440 | static void polling_island_add_fds_locked(polling_island *pi, grpc_fd **fds, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 441 | size_t fd_count, bool add_fd_refs, |
| 442 | grpc_error **error) { |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 443 | int err; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 444 | size_t i; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 445 | struct epoll_event ev; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 446 | char *err_msg; |
| 447 | const char *err_desc = "polling_island_add_fds"; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 448 | |
Sree Kuchibhotla | ad2c477 | 2016-06-13 19:06:54 -0700 | [diff] [blame] | 449 | #ifdef GRPC_TSAN |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 450 | /* See the definition of g_epoll_sync for more context */ |
Sree Kuchibhotla | 0224dcc | 2016-06-22 18:04:00 -0700 | [diff] [blame] | 451 | gpr_atm_rel_store(&g_epoll_sync, (gpr_atm)0); |
Sree Kuchibhotla | ad2c477 | 2016-06-13 19:06:54 -0700 | [diff] [blame] | 452 | #endif /* defined(GRPC_TSAN) */ |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 453 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 454 | for (i = 0; i < fd_count; i++) { |
| 455 | ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); |
| 456 | ev.data.ptr = fds[i]; |
| 457 | err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, fds[i]->fd, &ev); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 458 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 459 | if (err < 0) { |
| 460 | if (errno != EEXIST) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 461 | gpr_asprintf( |
| 462 | &err_msg, |
| 463 | "epoll_ctl (epoll_fd: %d) add fd: %d failed with error: %d (%s)", |
| 464 | pi->epoll_fd, fds[i]->fd, errno, strerror(errno)); |
| 465 | append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); |
| 466 | gpr_free(err_msg); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 467 | } |
| 468 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 469 | continue; |
| 470 | } |
| 471 | |
| 472 | if (pi->fd_cnt == pi->fd_capacity) { |
| 473 | pi->fd_capacity = GPR_MAX(pi->fd_capacity + 8, pi->fd_cnt * 3 / 2); |
| 474 | pi->fds = gpr_realloc(pi->fds, sizeof(grpc_fd *) * pi->fd_capacity); |
| 475 | } |
| 476 | |
| 477 | pi->fds[pi->fd_cnt++] = fds[i]; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 478 | if (add_fd_refs) { |
| 479 | GRPC_FD_REF(fds[i], "polling_island"); |
| 480 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 481 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 482 | } |
| 483 | |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 484 | /* The caller is expected to hold pi->mu before calling this */ |
| 485 | static void polling_island_add_wakeup_fd_locked(polling_island *pi, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 486 | grpc_wakeup_fd *wakeup_fd, |
| 487 | grpc_error **error) { |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 488 | struct epoll_event ev; |
| 489 | int err; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 490 | char *err_msg; |
| 491 | const char *err_desc = "polling_island_add_wakeup_fd"; |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 492 | |
| 493 | ev.events = (uint32_t)(EPOLLIN | EPOLLET); |
| 494 | ev.data.ptr = wakeup_fd; |
| 495 | err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, |
| 496 | GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), &ev); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 497 | if (err < 0 && errno != EEXIST) { |
| 498 | gpr_asprintf(&err_msg, |
| 499 | "epoll_ctl (epoll_fd: %d) add wakeup fd: %d failed with " |
| 500 | "error: %d (%s)", |
Craig Tiller | 1fa9ddb | 2016-11-28 08:19:37 -0800 | [diff] [blame] | 501 | pi->epoll_fd, GRPC_WAKEUP_FD_GET_READ_FD(&global_wakeup_fd), |
| 502 | errno, strerror(errno)); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 503 | append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); |
| 504 | gpr_free(err_msg); |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 505 | } |
| 506 | } |
| 507 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 508 | /* The caller is expected to hold pi->mu lock before calling this function */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 509 | static void polling_island_remove_all_fds_locked(polling_island *pi, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 510 | bool remove_fd_refs, |
| 511 | grpc_error **error) { |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 512 | int err; |
| 513 | size_t i; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 514 | char *err_msg; |
| 515 | const char *err_desc = "polling_island_remove_fds"; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 516 | |
| 517 | for (i = 0; i < pi->fd_cnt; i++) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 518 | err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, pi->fds[i]->fd, NULL); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 519 | if (err < 0 && errno != ENOENT) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 520 | gpr_asprintf(&err_msg, |
| 521 | "epoll_ctl (epoll_fd: %d) delete fds[%zu]: %d failed with " |
| 522 | "error: %d (%s)", |
| 523 | pi->epoll_fd, i, pi->fds[i]->fd, errno, strerror(errno)); |
| 524 | append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); |
| 525 | gpr_free(err_msg); |
Sree Kuchibhotla | ad162ba | 2016-06-06 16:23:37 -0700 | [diff] [blame] | 526 | } |
| 527 | |
| 528 | if (remove_fd_refs) { |
| 529 | GRPC_FD_UNREF(pi->fds[i], "polling_island"); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 530 | } |
| 531 | } |
| 532 | |
| 533 | pi->fd_cnt = 0; |
| 534 | } |
| 535 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 536 | /* The caller is expected to hold pi->mu lock before calling this function */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 537 | static void polling_island_remove_fd_locked(polling_island *pi, grpc_fd *fd, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 538 | bool is_fd_closed, |
| 539 | grpc_error **error) { |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 540 | int err; |
| 541 | size_t i; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 542 | char *err_msg; |
| 543 | const char *err_desc = "polling_island_remove_fd"; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 544 | |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 545 | /* If fd is already closed, then it would have been automatically been removed |
| 546 | from the epoll set */ |
| 547 | if (!is_fd_closed) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 548 | err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, fd->fd, NULL); |
| 549 | if (err < 0 && errno != ENOENT) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 550 | gpr_asprintf( |
| 551 | &err_msg, |
| 552 | "epoll_ctl (epoll_fd: %d) del fd: %d failed with error: %d (%s)", |
| 553 | pi->epoll_fd, fd->fd, errno, strerror(errno)); |
| 554 | append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); |
| 555 | gpr_free(err_msg); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 556 | } |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 557 | } |
| 558 | |
| 559 | for (i = 0; i < pi->fd_cnt; i++) { |
| 560 | if (pi->fds[i] == fd) { |
| 561 | pi->fds[i] = pi->fds[--pi->fd_cnt]; |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 562 | GRPC_FD_UNREF(fd, "polling_island"); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 563 | break; |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 568 | /* Might return NULL in case of an error */ |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 569 | static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx, |
| 570 | grpc_fd *initial_fd, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 571 | grpc_error **error) { |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 572 | polling_island *pi = NULL; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 573 | const char *err_desc = "polling_island_create"; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 574 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 575 | *error = GRPC_ERROR_NONE; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 576 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 577 | pi = gpr_malloc(sizeof(*pi)); |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 578 | pi->workqueue_scheduler.vtable = &workqueue_scheduler_vtable; |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 579 | gpr_mu_init(&pi->mu); |
| 580 | pi->fd_cnt = 0; |
| 581 | pi->fd_capacity = 0; |
| 582 | pi->fds = NULL; |
| 583 | pi->epoll_fd = -1; |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 584 | |
| 585 | gpr_mu_init(&pi->workqueue_read_mu); |
| 586 | gpr_mpscq_init(&pi->workqueue_items); |
| 587 | gpr_atm_rel_store(&pi->workqueue_item_count, 0); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 588 | |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 589 | gpr_atm_rel_store(&pi->ref_count, 0); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 590 | gpr_atm_rel_store(&pi->poller_count, 0); |
Sree Kuchibhotla | 0224dcc | 2016-06-22 18:04:00 -0700 | [diff] [blame] | 591 | gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 592 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 593 | if (!append_error(error, grpc_wakeup_fd_init(&pi->workqueue_wakeup_fd), |
| 594 | err_desc)) { |
| 595 | goto done; |
| 596 | } |
| 597 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 598 | pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC); |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 599 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 600 | if (pi->epoll_fd < 0) { |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 601 | append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc); |
| 602 | goto done; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 603 | } |
| 604 | |
Craig Tiller | b4b8e1e | 2016-11-28 07:33:13 -0800 | [diff] [blame] | 605 | polling_island_add_wakeup_fd_locked(pi, &global_wakeup_fd, error); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 606 | polling_island_add_wakeup_fd_locked(pi, &pi->workqueue_wakeup_fd, error); |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 607 | |
| 608 | if (initial_fd != NULL) { |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 609 | polling_island_add_fds_locked(pi, &initial_fd, 1, true, error); |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 610 | } |
| 611 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 612 | done: |
| 613 | if (*error != GRPC_ERROR_NONE) { |
Craig Tiller | 0a06cd7 | 2016-07-14 13:21:24 -0700 | [diff] [blame] | 614 | polling_island_delete(exec_ctx, pi); |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 615 | pi = NULL; |
| 616 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 617 | return pi; |
| 618 | } |
| 619 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 620 | static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) { |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 621 | GPR_ASSERT(pi->fd_cnt == 0); |
| 622 | |
Craig Tiller | 0a06cd7 | 2016-07-14 13:21:24 -0700 | [diff] [blame] | 623 | if (pi->epoll_fd >= 0) { |
| 624 | close(pi->epoll_fd); |
| 625 | } |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 626 | GPR_ASSERT(gpr_atm_no_barrier_load(&pi->workqueue_item_count) == 0); |
| 627 | gpr_mu_destroy(&pi->workqueue_read_mu); |
| 628 | gpr_mpscq_destroy(&pi->workqueue_items); |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 629 | gpr_mu_destroy(&pi->mu); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 630 | grpc_wakeup_fd_destroy(&pi->workqueue_wakeup_fd); |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 631 | gpr_free(pi->fds); |
| 632 | gpr_free(pi); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 633 | } |
| 634 | |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 635 | /* Attempts to gets the last polling island in the linked list (liked by the |
| 636 | * 'merged_to' field). Since this does not lock the polling island, there are no |
| 637 | * guarantees that the island returned is the last island */ |
| 638 | static polling_island *polling_island_maybe_get_latest(polling_island *pi) { |
| 639 | polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); |
| 640 | while (next != NULL) { |
| 641 | pi = next; |
| 642 | next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); |
| 643 | } |
| 644 | |
| 645 | return pi; |
| 646 | } |
| 647 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 648 | /* Gets the lock on the *latest* polling island i.e the last polling island in |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 649 | the linked list (linked by the 'merged_to' field). Call gpr_mu_unlock on the |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 650 | returned polling island's mu. |
| 651 | Usage: To lock/unlock polling island "pi", do the following: |
| 652 | polling_island *pi_latest = polling_island_lock(pi); |
| 653 | ... |
| 654 | ... critical section .. |
| 655 | ... |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 656 | gpr_mu_unlock(&pi_latest->mu); // NOTE: use pi_latest->mu. NOT pi->mu */ |
| 657 | static polling_island *polling_island_lock(polling_island *pi) { |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 658 | polling_island *next = NULL; |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 659 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 660 | while (true) { |
| 661 | next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); |
| 662 | if (next == NULL) { |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 663 | /* Looks like 'pi' is the last node in the linked list but unless we check |
| 664 | this by holding the pi->mu lock, we cannot be sure (i.e without the |
| 665 | pi->mu lock, we don't prevent island merges). |
| 666 | To be absolutely sure, check once more by holding the pi->mu lock */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 667 | gpr_mu_lock(&pi->mu); |
| 668 | next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); |
| 669 | if (next == NULL) { |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 670 | /* pi is infact the last node and we have the pi->mu lock. we're done */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 671 | break; |
| 672 | } |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 673 | |
| 674 | /* pi->merged_to is not NULL i.e pi isn't the last node anymore. pi->mu |
| 675 | * isn't the lock we are interested in. Continue traversing the list */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 676 | gpr_mu_unlock(&pi->mu); |
| 677 | } |
| 678 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 679 | pi = next; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 680 | } |
| 681 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 682 | return pi; |
| 683 | } |
| 684 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 685 | /* Gets the lock on the *latest* polling islands in the linked lists pointed by |
| 686 | *p and *q (and also updates *p and *q to point to the latest polling islands) |
| 687 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 688 | This function is needed because calling the following block of code to obtain |
| 689 | locks on polling islands (*p and *q) is prone to deadlocks. |
| 690 | { |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 691 | polling_island_lock(*p, true); |
| 692 | polling_island_lock(*q, true); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 693 | } |
| 694 | |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 695 | Usage/example: |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 696 | polling_island *p1; |
| 697 | polling_island *p2; |
| 698 | .. |
| 699 | polling_island_lock_pair(&p1, &p2); |
| 700 | .. |
| 701 | .. Critical section with both p1 and p2 locked |
| 702 | .. |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 703 | // Release locks: Always call polling_island_unlock_pair() to release locks |
| 704 | polling_island_unlock_pair(p1, p2); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 705 | */ |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 706 | static void polling_island_lock_pair(polling_island **p, polling_island **q) { |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 707 | polling_island *pi_1 = *p; |
| 708 | polling_island *pi_2 = *q; |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 709 | polling_island *next_1 = NULL; |
| 710 | polling_island *next_2 = NULL; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 711 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 712 | /* The algorithm is simple: |
| 713 | - Go to the last polling islands in the linked lists *pi_1 and *pi_2 (and |
| 714 | keep updating pi_1 and pi_2) |
| 715 | - Then obtain locks on the islands by following a lock order rule of |
| 716 | locking polling_island with lower address first |
| 717 | Special case: Before obtaining the locks, check if pi_1 and pi_2 are |
| 718 | pointing to the same island. If that is the case, we can just call |
| 719 | polling_island_lock() |
| 720 | - After obtaining both the locks, double check that the polling islands |
| 721 | are still the last polling islands in their respective linked lists |
| 722 | (this is because there might have been polling island merges before |
| 723 | we got the lock) |
| 724 | - If the polling islands are the last islands, we are done. If not, |
| 725 | release the locks and continue the process from the first step */ |
| 726 | while (true) { |
| 727 | next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to); |
| 728 | while (next_1 != NULL) { |
| 729 | pi_1 = next_1; |
| 730 | next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 731 | } |
| 732 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 733 | next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to); |
| 734 | while (next_2 != NULL) { |
| 735 | pi_2 = next_2; |
| 736 | next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to); |
| 737 | } |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 738 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 739 | if (pi_1 == pi_2) { |
| 740 | pi_1 = pi_2 = polling_island_lock(pi_1); |
| 741 | break; |
| 742 | } |
| 743 | |
| 744 | if (pi_1 < pi_2) { |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 745 | gpr_mu_lock(&pi_1->mu); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 746 | gpr_mu_lock(&pi_2->mu); |
| 747 | } else { |
| 748 | gpr_mu_lock(&pi_2->mu); |
| 749 | gpr_mu_lock(&pi_1->mu); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 750 | } |
| 751 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 752 | next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to); |
| 753 | next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to); |
| 754 | if (next_1 == NULL && next_2 == NULL) { |
| 755 | break; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 756 | } |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 757 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 758 | gpr_mu_unlock(&pi_1->mu); |
| 759 | gpr_mu_unlock(&pi_2->mu); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 760 | } |
| 761 | |
| 762 | *p = pi_1; |
| 763 | *q = pi_2; |
| 764 | } |
| 765 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 766 | static void polling_island_unlock_pair(polling_island *p, polling_island *q) { |
| 767 | if (p == q) { |
| 768 | gpr_mu_unlock(&p->mu); |
| 769 | } else { |
| 770 | gpr_mu_unlock(&p->mu); |
| 771 | gpr_mu_unlock(&q->mu); |
| 772 | } |
| 773 | } |
| 774 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 775 | static void workqueue_maybe_wakeup(polling_island *pi) { |
Craig Tiller | 2e62013 | 2016-10-10 15:27:44 -0700 | [diff] [blame] | 776 | /* If this thread is the current poller, then it may be that it's about to |
| 777 | decrement the current poller count, so we need to look past this thread */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 778 | bool is_current_poller = (g_current_thread_polling_island == pi); |
| 779 | gpr_atm min_current_pollers_for_wakeup = is_current_poller ? 1 : 0; |
| 780 | gpr_atm current_pollers = gpr_atm_no_barrier_load(&pi->poller_count); |
Craig Tiller | 2e62013 | 2016-10-10 15:27:44 -0700 | [diff] [blame] | 781 | /* Only issue a wakeup if it's likely that some poller could come in and take |
| 782 | it right now. Note that since we do an anticipatory mpscq_pop every poll |
| 783 | loop, it's ok if we miss the wakeup here, as we'll get the work item when |
| 784 | the next poller enters anyway. */ |
| 785 | if (current_pollers > min_current_pollers_for_wakeup) { |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 786 | GRPC_LOG_IF_ERROR("workqueue_wakeup_fd", |
| 787 | grpc_wakeup_fd_wakeup(&pi->workqueue_wakeup_fd)); |
| 788 | } |
| 789 | } |
| 790 | |
| 791 | static void workqueue_move_items_to_parent(polling_island *q) { |
| 792 | polling_island *p = (polling_island *)gpr_atm_no_barrier_load(&q->merged_to); |
| 793 | if (p == NULL) { |
| 794 | return; |
| 795 | } |
| 796 | gpr_mu_lock(&q->workqueue_read_mu); |
| 797 | int num_added = 0; |
| 798 | while (gpr_atm_no_barrier_load(&q->workqueue_item_count) > 0) { |
| 799 | gpr_mpscq_node *n = gpr_mpscq_pop(&q->workqueue_items); |
| 800 | if (n != NULL) { |
| 801 | gpr_atm_no_barrier_fetch_add(&q->workqueue_item_count, -1); |
| 802 | gpr_atm_no_barrier_fetch_add(&p->workqueue_item_count, 1); |
| 803 | gpr_mpscq_push(&p->workqueue_items, n); |
| 804 | num_added++; |
| 805 | } |
| 806 | } |
| 807 | gpr_mu_unlock(&q->workqueue_read_mu); |
| 808 | if (num_added > 0) { |
| 809 | workqueue_maybe_wakeup(p); |
| 810 | } |
| 811 | workqueue_move_items_to_parent(p); |
| 812 | } |
| 813 | |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 814 | static polling_island *polling_island_merge(polling_island *p, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 815 | polling_island *q, |
| 816 | grpc_error **error) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 817 | /* Get locks on both the polling islands */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 818 | polling_island_lock_pair(&p, &q); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 819 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 820 | if (p != q) { |
| 821 | /* Make sure that p points to the polling island with fewer fds than q */ |
| 822 | if (p->fd_cnt > q->fd_cnt) { |
| 823 | GPR_SWAP(polling_island *, p, q); |
| 824 | } |
| 825 | |
| 826 | /* Merge p with q i.e move all the fds from p (The one with fewer fds) to q |
| 827 | Note that the refcounts on the fds being moved will not change here. |
| 828 | This is why the last param in the following two functions is 'false') */ |
| 829 | polling_island_add_fds_locked(q, p->fds, p->fd_cnt, false, error); |
| 830 | polling_island_remove_all_fds_locked(p, false, error); |
| 831 | |
| 832 | /* Wakeup all the pollers (if any) on p so that they pickup this change */ |
| 833 | polling_island_add_wakeup_fd_locked(p, &polling_island_wakeup_fd, error); |
| 834 | |
| 835 | /* Add the 'merged_to' link from p --> q */ |
| 836 | gpr_atm_rel_store(&p->merged_to, (gpr_atm)q); |
| 837 | PI_ADD_REF(q, "pi_merge"); /* To account for the new incoming ref from p */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 838 | |
Harvey Tuch | daa9f45 | 2016-11-21 15:42:49 -0500 | [diff] [blame] | 839 | workqueue_move_items_to_parent(p); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 840 | } |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 841 | /* else if p == q, nothing needs to be done */ |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 842 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 843 | polling_island_unlock_pair(p, q); |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 844 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 845 | /* Return the merged polling island (Note that no merge would have happened |
| 846 | if p == q which is ok) */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 847 | return q; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 848 | } |
| 849 | |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 850 | static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure, |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 851 | grpc_error *error) { |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 852 | GPR_TIMER_BEGIN("workqueue.enqueue", 0); |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 853 | grpc_workqueue *workqueue = (grpc_workqueue *)closure->scheduler; |
Craig Tiller | 1dc5dbb | 2016-09-09 17:09:39 -0700 | [diff] [blame] | 854 | /* take a ref to the workqueue: otherwise it can happen that whatever events |
| 855 | * this kicks off ends up destroying the workqueue before this function |
| 856 | * completes */ |
| 857 | GRPC_WORKQUEUE_REF(workqueue, "enqueue"); |
| 858 | polling_island *pi = (polling_island *)workqueue; |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 859 | gpr_atm last = gpr_atm_no_barrier_fetch_add(&pi->workqueue_item_count, 1); |
| 860 | closure->error_data.error = error; |
| 861 | gpr_mpscq_push(&pi->workqueue_items, &closure->next_data.atm_next); |
| 862 | if (last == 0) { |
| 863 | workqueue_maybe_wakeup(pi); |
| 864 | } |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 865 | workqueue_move_items_to_parent(pi); |
Craig Tiller | 1dc5dbb | 2016-09-09 17:09:39 -0700 | [diff] [blame] | 866 | GRPC_WORKQUEUE_UNREF(exec_ctx, workqueue, "enqueue"); |
| 867 | GPR_TIMER_END("workqueue.enqueue", 0); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 868 | } |
| 869 | |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 870 | static grpc_closure_scheduler *workqueue_scheduler(grpc_workqueue *workqueue) { |
| 871 | polling_island *pi = (polling_island *)workqueue; |
Craig Tiller | 801c6cc | 2017-01-03 08:13:13 -0800 | [diff] [blame] | 872 | return workqueue == NULL ? grpc_schedule_on_exec_ctx |
| 873 | : &pi->workqueue_scheduler; |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 874 | } |
| 875 | |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 876 | static grpc_error *polling_island_global_init() { |
| 877 | grpc_error *error = GRPC_ERROR_NONE; |
| 878 | |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 879 | error = grpc_wakeup_fd_init(&polling_island_wakeup_fd); |
| 880 | if (error == GRPC_ERROR_NONE) { |
| 881 | error = grpc_wakeup_fd_wakeup(&polling_island_wakeup_fd); |
| 882 | } |
| 883 | |
| 884 | return error; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 885 | } |
| 886 | |
Sree Kuchibhotla | d627c10 | 2016-06-06 15:49:32 -0700 | [diff] [blame] | 887 | static void polling_island_global_shutdown() { |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 888 | grpc_wakeup_fd_destroy(&polling_island_wakeup_fd); |
Sree Kuchibhotla | d627c10 | 2016-06-06 15:49:32 -0700 | [diff] [blame] | 889 | } |
| 890 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 891 | /******************************************************************************* |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 892 | * Fd Definitions |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 893 | */ |
| 894 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 895 | /* We need to keep a freelist not because of any concerns of malloc performance |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 896 | * but instead so that implementations with multiple threads in (for example) |
| 897 | * epoll_wait deal with the race between pollset removal and incoming poll |
| 898 | * notifications. |
| 899 | * |
| 900 | * The problem is that the poller ultimately holds a reference to this |
| 901 | * object, so it is very difficult to know when is safe to free it, at least |
| 902 | * without some expensive synchronization. |
| 903 | * |
| 904 | * If we keep the object freelisted, in the worst case losing this race just |
| 905 | * becomes a spurious read notification on a reused fd. |
| 906 | */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 907 | |
| 908 | /* The alarm system needs to be able to wakeup 'some poller' sometimes |
| 909 | * (specifically when a new alarm needs to be triggered earlier than the next |
| 910 | * alarm 'epoch'). This wakeup_fd gives us something to alert on when such a |
| 911 | * case occurs. */ |
Sree Kuchibhotla | 9bc3d2d | 2016-06-06 10:27:56 -0700 | [diff] [blame] | 912 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 913 | static grpc_fd *fd_freelist = NULL; |
| 914 | static gpr_mu fd_freelist_mu; |
| 915 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 916 | #ifdef GRPC_FD_REF_COUNT_DEBUG |
| 917 | #define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__) |
| 918 | #define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__) |
| 919 | static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file, |
| 920 | int line) { |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 921 | gpr_log(GPR_DEBUG, "FD %d %p ref %d %ld -> %ld [%s; %s:%d]", fd->fd, |
| 922 | (void *)fd, n, gpr_atm_no_barrier_load(&fd->refst), |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 923 | gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line); |
| 924 | #else |
| 925 | #define REF_BY(fd, n, reason) ref_by(fd, n) |
| 926 | #define UNREF_BY(fd, n, reason) unref_by(fd, n) |
| 927 | static void ref_by(grpc_fd *fd, int n) { |
| 928 | #endif |
| 929 | GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0); |
| 930 | } |
| 931 | |
| 932 | #ifdef GRPC_FD_REF_COUNT_DEBUG |
| 933 | static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file, |
| 934 | int line) { |
| 935 | gpr_atm old; |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 936 | gpr_log(GPR_DEBUG, "FD %d %p unref %d %ld -> %ld [%s; %s:%d]", fd->fd, |
| 937 | (void *)fd, n, gpr_atm_no_barrier_load(&fd->refst), |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 938 | gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line); |
| 939 | #else |
| 940 | static void unref_by(grpc_fd *fd, int n) { |
| 941 | gpr_atm old; |
| 942 | #endif |
| 943 | old = gpr_atm_full_fetch_add(&fd->refst, -n); |
| 944 | if (old == n) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 945 | /* Add the fd to the freelist */ |
| 946 | gpr_mu_lock(&fd_freelist_mu); |
| 947 | fd->freelist_next = fd_freelist; |
| 948 | fd_freelist = fd; |
| 949 | grpc_iomgr_unregister_object(&fd->iomgr_object); |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 950 | |
Sree Kuchibhotla | a70ccb6 | 2017-02-13 23:16:52 -0800 | [diff] [blame] | 951 | grpc_error *err = (grpc_error *)gpr_atm_acq_load(&fd->shutdown_error); |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 952 | /* Clear the least significant bit if it set (in case fd was shutdown) */ |
| 953 | err = (grpc_error *)((intptr_t)err & ~FD_SHUTDOWN_BIT); |
Sree Kuchibhotla | 8b8cbed | 2017-02-09 21:31:27 -0800 | [diff] [blame] | 954 | GRPC_ERROR_UNREF(err); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 955 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 956 | gpr_mu_unlock(&fd_freelist_mu); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 957 | } else { |
| 958 | GPR_ASSERT(old > n); |
| 959 | } |
| 960 | } |
| 961 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 962 | /* Increment refcount by two to avoid changing the orphan bit */ |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 963 | #ifdef GRPC_FD_REF_COUNT_DEBUG |
| 964 | static void fd_ref(grpc_fd *fd, const char *reason, const char *file, |
| 965 | int line) { |
| 966 | ref_by(fd, 2, reason, file, line); |
| 967 | } |
| 968 | |
| 969 | static void fd_unref(grpc_fd *fd, const char *reason, const char *file, |
| 970 | int line) { |
| 971 | unref_by(fd, 2, reason, file, line); |
| 972 | } |
| 973 | #else |
| 974 | static void fd_ref(grpc_fd *fd) { ref_by(fd, 2); } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 975 | static void fd_unref(grpc_fd *fd) { unref_by(fd, 2); } |
| 976 | #endif |
| 977 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 978 | static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } |
| 979 | |
| 980 | static void fd_global_shutdown(void) { |
| 981 | gpr_mu_lock(&fd_freelist_mu); |
| 982 | gpr_mu_unlock(&fd_freelist_mu); |
| 983 | while (fd_freelist != NULL) { |
| 984 | grpc_fd *fd = fd_freelist; |
| 985 | fd_freelist = fd_freelist->freelist_next; |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 986 | gpr_mu_destroy(&fd->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 987 | gpr_free(fd); |
| 988 | } |
| 989 | gpr_mu_destroy(&fd_freelist_mu); |
| 990 | } |
| 991 | |
| 992 | static grpc_fd *fd_create(int fd, const char *name) { |
| 993 | grpc_fd *new_fd = NULL; |
| 994 | |
| 995 | gpr_mu_lock(&fd_freelist_mu); |
| 996 | if (fd_freelist != NULL) { |
| 997 | new_fd = fd_freelist; |
| 998 | fd_freelist = fd_freelist->freelist_next; |
| 999 | } |
| 1000 | gpr_mu_unlock(&fd_freelist_mu); |
| 1001 | |
| 1002 | if (new_fd == NULL) { |
| 1003 | new_fd = gpr_malloc(sizeof(grpc_fd)); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1004 | gpr_mu_init(&new_fd->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1005 | } |
| 1006 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1007 | /* Note: It is not really needed to get the new_fd->po.mu lock here. If this |
| 1008 | * is a newly created fd (or an fd we got from the freelist), no one else |
| 1009 | * would be holding a lock to it anyway. */ |
| 1010 | gpr_mu_lock(&new_fd->po.mu); |
| 1011 | new_fd->po.pi = NULL; |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1012 | #ifdef PO_DEBUG |
| 1013 | new_fd->po.obj_type = POLL_OBJ_FD; |
| 1014 | #endif |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1015 | |
Sree Kuchibhotla | 0224dcc | 2016-06-22 18:04:00 -0700 | [diff] [blame] | 1016 | gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1); |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1017 | new_fd->fd = fd; |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1018 | gpr_atm_no_barrier_store(&new_fd->shutdown_error, (gpr_atm)GRPC_ERROR_NONE); |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1019 | new_fd->orphaned = false; |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1020 | gpr_atm_no_barrier_store(&new_fd->read_closure, CLOSURE_NOT_READY); |
| 1021 | gpr_atm_no_barrier_store(&new_fd->write_closure, CLOSURE_NOT_READY); |
| 1022 | gpr_atm_no_barrier_store(&new_fd->read_notifier_pollset, (gpr_atm)NULL); |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1023 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1024 | new_fd->freelist_next = NULL; |
| 1025 | new_fd->on_done_closure = NULL; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1026 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1027 | gpr_mu_unlock(&new_fd->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1028 | |
| 1029 | char *fd_name; |
| 1030 | gpr_asprintf(&fd_name, "%s fd=%d", name, fd); |
| 1031 | grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1032 | #ifdef GRPC_FD_REF_COUNT_DEBUG |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 1033 | gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, (void *)new_fd, fd_name); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1034 | #endif |
Sree Kuchibhotla | 6a29545 | 2016-06-23 15:53:10 -0700 | [diff] [blame] | 1035 | gpr_free(fd_name); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1036 | return new_fd; |
| 1037 | } |
| 1038 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1039 | static int fd_wrapped_fd(grpc_fd *fd) { |
| 1040 | int ret_fd = -1; |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1041 | gpr_mu_lock(&fd->po.mu); |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1042 | if (!fd->orphaned) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1043 | ret_fd = fd->fd; |
| 1044 | } |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1045 | gpr_mu_unlock(&fd->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1046 | |
| 1047 | return ret_fd; |
| 1048 | } |
| 1049 | |
| 1050 | static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, |
| 1051 | grpc_closure *on_done, int *release_fd, |
| 1052 | const char *reason) { |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1053 | bool is_fd_closed = false; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1054 | grpc_error *error = GRPC_ERROR_NONE; |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 1055 | polling_island *unref_pi = NULL; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1056 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1057 | gpr_mu_lock(&fd->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1058 | fd->on_done_closure = on_done; |
| 1059 | |
| 1060 | /* If release_fd is not NULL, we should be relinquishing control of the file |
| 1061 | descriptor fd->fd (but we still own the grpc_fd structure). */ |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1062 | if (release_fd != NULL) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1063 | *release_fd = fd->fd; |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1064 | } else { |
| 1065 | close(fd->fd); |
| 1066 | is_fd_closed = true; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1067 | } |
| 1068 | |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1069 | fd->orphaned = true; |
| 1070 | |
| 1071 | /* Remove the active status but keep referenced. We want this grpc_fd struct |
| 1072 | to be alive (and not added to freelist) until the end of this function */ |
| 1073 | REF_BY(fd, 1, reason); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1074 | |
| 1075 | /* Remove the fd from the polling island: |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1076 | - Get a lock on the latest polling island (i.e the last island in the |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1077 | linked list pointed by fd->po.pi). This is the island that |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1078 | would actually contain the fd |
| 1079 | - Remove the fd from the latest polling island |
| 1080 | - Unlock the latest polling island |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1081 | - Set fd->po.pi to NULL (but remove the ref on the polling island |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1082 | before doing this.) */ |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1083 | if (fd->po.pi != NULL) { |
| 1084 | polling_island *pi_latest = polling_island_lock(fd->po.pi); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1085 | polling_island_remove_fd_locked(pi_latest, fd, is_fd_closed, &error); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1086 | gpr_mu_unlock(&pi_latest->mu); |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1087 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1088 | unref_pi = fd->po.pi; |
| 1089 | fd->po.pi = NULL; |
Sree Kuchibhotla | 88ee12f | 2016-06-03 19:26:48 -0700 | [diff] [blame] | 1090 | } |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1091 | |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 1092 | grpc_closure_sched(exec_ctx, fd->on_done_closure, GRPC_ERROR_REF(error)); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1093 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1094 | gpr_mu_unlock(&fd->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1095 | UNREF_BY(fd, 2, reason); /* Drop the reference */ |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 1096 | if (unref_pi != NULL) { |
Craig Tiller | 0a06cd7 | 2016-07-14 13:21:24 -0700 | [diff] [blame] | 1097 | /* Unref stale polling island here, outside the fd lock above. |
| 1098 | The polling island owns a workqueue which owns an fd, and unreffing |
| 1099 | inside the lock can cause an eventual lock loop that makes TSAN very |
| 1100 | unhappy. */ |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 1101 | PI_UNREF(exec_ctx, unref_pi, "fd_orphan"); |
| 1102 | } |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1103 | GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error)); |
Yuchen Zeng | a0399f2 | 2016-08-04 17:52:53 -0700 | [diff] [blame] | 1104 | GRPC_ERROR_UNREF(error); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1105 | } |
| 1106 | |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1107 | static void notify_on(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *state, |
| 1108 | grpc_closure *closure) { |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1109 | while (true) { |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1110 | /* Fast-path: CLOSURE_NOT_READY -> <closure> */ |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1111 | /* The 'release' cas here matches the 'acquire' load in set_ready and |
| 1112 | set_shutdown to ensure that the clousure being run happens-after the call |
| 1113 | to notify_on */ |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1114 | if (gpr_atm_rel_cas(state, CLOSURE_NOT_READY, (gpr_atm)closure)) { |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1115 | return; /* Fast-path successful. Return */ |
| 1116 | } |
| 1117 | |
| 1118 | /* Slowpath */ |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1119 | /* The 'acquire' load matches the 'release' cas in set_ready/set_shutdown */ |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1120 | gpr_atm curr = gpr_atm_acq_load(state); |
| 1121 | switch (curr) { |
| 1122 | case CLOSURE_NOT_READY: { |
| 1123 | break; /* retry */ |
| 1124 | } |
| 1125 | |
| 1126 | case CLOSURE_READY: { |
Sree Kuchibhotla | 21f2f06 | 2017-02-16 10:42:37 -0800 | [diff] [blame] | 1127 | /* Change the state to CLOSURE_NOT_READY. If successful: Schedule the |
| 1128 | closure. If not, most likely the state transitioned to shutdown. We |
| 1129 | should retry */ |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1130 | if (gpr_atm_rel_cas(state, CLOSURE_READY, CLOSURE_NOT_READY)) { |
| 1131 | grpc_closure_sched(exec_ctx, closure, GRPC_ERROR_NONE); |
| 1132 | return; /* Slow-path successful. Return */ |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1133 | } |
| 1134 | |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1135 | break; /* retry */ |
| 1136 | } |
| 1137 | |
| 1138 | default: { |
| 1139 | /* 'curr' is either a closure or the fd is shutdown(in which case 'curr' |
| 1140 | contains a pointer to the shutdown-error). If the fd is shutdown, |
| 1141 | schedule the closure with the shutdown error */ |
| 1142 | if ((curr & FD_SHUTDOWN_BIT) > 0) { |
| 1143 | grpc_error *shutdown_err = (grpc_error *)(curr & ~FD_SHUTDOWN_BIT); |
| 1144 | grpc_closure_sched( |
| 1145 | exec_ctx, closure, |
| 1146 | GRPC_ERROR_CREATE_REFERENCING("FD Shutdown", &shutdown_err, 1)); |
| 1147 | return; |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1148 | } |
| 1149 | |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1150 | /* There is already a closure!. This indicates a bug in the code */ |
| 1151 | gpr_log(GPR_ERROR, |
| 1152 | "notify_on called with a previous callback still pending"); |
| 1153 | abort(); |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1154 | } |
| 1155 | } |
| 1156 | } |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1157 | |
| 1158 | GPR_UNREACHABLE_CODE(return ); |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1159 | } |
| 1160 | |
| 1161 | static void set_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *state, |
| 1162 | grpc_error *shutdown_err) { |
| 1163 | /* Try the fast-path first (i.e expect the current value to be |
| 1164 | CLOSURE_NOT_READY */ |
| 1165 | gpr_atm curr = CLOSURE_NOT_READY; |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 1166 | gpr_atm new_state = (gpr_atm)shutdown_err | FD_SHUTDOWN_BIT; |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1167 | |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1168 | while (true) { |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1169 | /* The 'release' cas here matches the 'acquire' load in notify_on and |
| 1170 | set_ready */ |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1171 | if (gpr_atm_rel_cas(state, curr, new_state)) { |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1172 | return; /* Fast-path successful. Return */ |
| 1173 | } |
| 1174 | |
| 1175 | /* Fallback to slowpath */ |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1176 | /* This 'acquire' load matches the 'release' in notify_on and set_ready */ |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1177 | curr = gpr_atm_acq_load(state); |
| 1178 | switch (curr) { |
| 1179 | case CLOSURE_READY: { |
| 1180 | break; /* retry */ |
| 1181 | } |
| 1182 | |
| 1183 | case CLOSURE_NOT_READY: { |
| 1184 | break; /* retry */ |
| 1185 | } |
| 1186 | |
| 1187 | default: { |
| 1188 | /* 'curr' is either a closure or the fd is already shutdown */ |
| 1189 | |
| 1190 | /* If fd is already shutdown, we are done */ |
| 1191 | if ((curr & FD_SHUTDOWN_BIT) > 0) { |
| 1192 | return; |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1193 | } |
| 1194 | |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1195 | /* Fd is not shutdown. Schedule the closure and move the state to |
| 1196 | shutdown state */ |
| 1197 | if (gpr_atm_rel_cas(state, curr, new_state)) { |
| 1198 | grpc_closure_sched( |
| 1199 | exec_ctx, (grpc_closure *)curr, |
| 1200 | GRPC_ERROR_CREATE_REFERENCING("FD Shutdown", &shutdown_err, 1)); |
| 1201 | return; |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1202 | } |
| 1203 | |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1204 | /* 'curr' was a closure but now changed to a different state. We will |
| 1205 | have to retry */ |
| 1206 | break; |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1207 | } |
| 1208 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1209 | } |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1210 | |
| 1211 | GPR_UNREACHABLE_CODE(return ); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1212 | } |
| 1213 | |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1214 | static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *state) { |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1215 | /* Try an optimistic case first (i.e assume current state is |
| 1216 | CLOSURE_NOT_READY) */ |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1217 | /* This 'release' cas matches the 'acquire' load in notify_on ensuring that |
| 1218 | the closure being run happens-after the return from epoll_pwait */ |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1219 | if (gpr_atm_rel_cas(state, CLOSURE_NOT_READY, CLOSURE_READY)) { |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1220 | return; /* early out */ |
| 1221 | } |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1222 | |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1223 | /* The 'acquire' here matches the 'release' in notify_on / set_shutdown */ |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1224 | gpr_atm curr = gpr_atm_acq_load(state); |
| 1225 | switch (curr) { |
| 1226 | case CLOSURE_READY: { |
| 1227 | /* Already ready. We are done here */ |
| 1228 | break; |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1229 | } |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1230 | |
| 1231 | case CLOSURE_NOT_READY: { |
| 1232 | /* The state was not CLOSURE_NOT_READY when we checked initially at the |
| 1233 | beginning of this function but now it is CLOSURE_NOT_READY again. |
| 1234 | This is only possible if the state transitioned out of |
| 1235 | CLOSURE_NOT_READY to either CLOSURE_READY or <some closure> and then |
| 1236 | back to CLOSURE_NOT_READY again (i.e after we entered this function, |
| 1237 | the fd became "ready" and the necessary actions were already done). |
| 1238 | So there is no need to make the state CLOSURE_READY now */ |
| 1239 | break; |
| 1240 | } |
| 1241 | |
| 1242 | default: { |
| 1243 | /* 'curr' is either a closure or the fd is shutdown */ |
| 1244 | if ((curr & FD_SHUTDOWN_BIT) > 0) { |
| 1245 | /* The fd is shutdown. Do nothing */ |
| 1246 | } else if (gpr_atm_rel_cas(state, curr, CLOSURE_NOT_READY)) { |
| 1247 | grpc_closure_sched(exec_ctx, (grpc_closure *)curr, GRPC_ERROR_NONE); |
| 1248 | } |
| 1249 | /* else the state changed again (only possible by either a racing |
| 1250 | set_ready or set_shutdown functions. In both these cases, the closure |
| 1251 | would have been scheduled for execution. So we are done here */ |
| 1252 | break; |
| 1253 | } |
| 1254 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1255 | } |
| 1256 | |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1257 | static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx, |
| 1258 | grpc_fd *fd) { |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1259 | gpr_atm notifier = gpr_atm_acq_load(&fd->read_notifier_pollset); |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1260 | return (grpc_pollset *)notifier; |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1261 | } |
| 1262 | |
Sree Kuchibhotla | 24b6eae | 2016-06-21 18:01:14 -0700 | [diff] [blame] | 1263 | static bool fd_is_shutdown(grpc_fd *fd) { |
Sree Kuchibhotla | 9998338 | 2017-02-12 17:03:27 -0800 | [diff] [blame] | 1264 | grpc_error *err = (grpc_error *)gpr_atm_acq_load(&fd->shutdown_error); |
Sree Kuchibhotla | 61fe094 | 2017-02-14 12:26:56 -0800 | [diff] [blame] | 1265 | return (((intptr_t)err & FD_SHUTDOWN_BIT) > 0); |
Sree Kuchibhotla | 24b6eae | 2016-06-21 18:01:14 -0700 | [diff] [blame] | 1266 | } |
| 1267 | |
Sree Kuchibhotla | 0100b2f | 2016-06-21 17:38:13 -0700 | [diff] [blame] | 1268 | /* Might be called multiple times */ |
Craig Tiller | cda759d | 2017-01-27 11:37:37 -0800 | [diff] [blame] | 1269 | static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_error *why) { |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 1270 | /* Store the shutdown error ORed with FD_SHUTDOWN_BIT in fd->shutdown_error */ |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1271 | if (gpr_atm_rel_cas(&fd->shutdown_error, (gpr_atm)GRPC_ERROR_NONE, |
Sree Kuchibhotla | 2fc2b3e | 2017-02-14 10:05:14 -0800 | [diff] [blame] | 1272 | (gpr_atm)why | FD_SHUTDOWN_BIT)) { |
Sree Kuchibhotla | 0100b2f | 2016-06-21 17:38:13 -0700 | [diff] [blame] | 1273 | shutdown(fd->fd, SHUT_RDWR); |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1274 | |
Sree Kuchibhotla | 91c4da3 | 2017-02-12 14:00:39 -0800 | [diff] [blame] | 1275 | set_shutdown(exec_ctx, fd, &fd->read_closure, why); |
| 1276 | set_shutdown(exec_ctx, fd, &fd->write_closure, why); |
Craig Tiller | cda759d | 2017-01-27 11:37:37 -0800 | [diff] [blame] | 1277 | } else { |
Sree Kuchibhotla | 4db8c82 | 2017-02-12 17:07:31 -0800 | [diff] [blame] | 1278 | /* Shutdown already called */ |
Craig Tiller | cda759d | 2017-01-27 11:37:37 -0800 | [diff] [blame] | 1279 | GRPC_ERROR_UNREF(why); |
Sree Kuchibhotla | 0100b2f | 2016-06-21 17:38:13 -0700 | [diff] [blame] | 1280 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1281 | } |
| 1282 | |
| 1283 | static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, |
| 1284 | grpc_closure *closure) { |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1285 | notify_on(exec_ctx, fd, &fd->read_closure, closure); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1286 | } |
| 1287 | |
| 1288 | static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, |
| 1289 | grpc_closure *closure) { |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1290 | notify_on(exec_ctx, fd, &fd->write_closure, closure); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1291 | } |
| 1292 | |
Craig Tiller | d6ba619 | 2016-06-30 15:42:41 -0700 | [diff] [blame] | 1293 | static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1294 | gpr_mu_lock(&fd->po.mu); |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1295 | grpc_workqueue *workqueue = |
| 1296 | GRPC_WORKQUEUE_REF((grpc_workqueue *)fd->po.pi, "fd_get_workqueue"); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1297 | gpr_mu_unlock(&fd->po.mu); |
Craig Tiller | d6ba619 | 2016-06-30 15:42:41 -0700 | [diff] [blame] | 1298 | return workqueue; |
| 1299 | } |
Craig Tiller | 70bd483 | 2016-06-30 14:20:46 -0700 | [diff] [blame] | 1300 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1301 | /******************************************************************************* |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1302 | * Pollset Definitions |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1303 | */ |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1304 | GPR_TLS_DECL(g_current_thread_pollset); |
| 1305 | GPR_TLS_DECL(g_current_thread_worker); |
Craig Tiller | 1919699 | 2016-06-27 18:45:56 -0700 | [diff] [blame] | 1306 | static __thread bool g_initialized_sigmask; |
| 1307 | static __thread sigset_t g_orig_sigmask; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1308 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1309 | static void sig_handler(int sig_num) { |
Sree Kuchibhotla | d627c10 | 2016-06-06 15:49:32 -0700 | [diff] [blame] | 1310 | #ifdef GRPC_EPOLL_DEBUG |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1311 | gpr_log(GPR_INFO, "Received signal %d", sig_num); |
Sree Kuchibhotla | 9bc3d2d | 2016-06-06 10:27:56 -0700 | [diff] [blame] | 1312 | #endif |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1313 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1314 | |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 1315 | static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); } |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1316 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1317 | /* Global state management */ |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1318 | static grpc_error *pollset_global_init(void) { |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1319 | gpr_tls_init(&g_current_thread_pollset); |
| 1320 | gpr_tls_init(&g_current_thread_worker); |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1321 | poller_kick_init(); |
Craig Tiller | b4b8e1e | 2016-11-28 07:33:13 -0800 | [diff] [blame] | 1322 | return grpc_wakeup_fd_init(&global_wakeup_fd); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1323 | } |
| 1324 | |
| 1325 | static void pollset_global_shutdown(void) { |
Craig Tiller | b4b8e1e | 2016-11-28 07:33:13 -0800 | [diff] [blame] | 1326 | grpc_wakeup_fd_destroy(&global_wakeup_fd); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1327 | gpr_tls_destroy(&g_current_thread_pollset); |
| 1328 | gpr_tls_destroy(&g_current_thread_worker); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1329 | } |
| 1330 | |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1331 | static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { |
| 1332 | grpc_error *err = GRPC_ERROR_NONE; |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 1333 | |
| 1334 | /* Kick the worker only if it was not already kicked */ |
| 1335 | if (gpr_atm_no_barrier_cas(&worker->is_kicked, (gpr_atm)0, (gpr_atm)1)) { |
| 1336 | GRPC_POLLING_TRACE( |
| 1337 | "pollset_worker_kick: Kicking worker: %p (thread id: %ld)", |
| 1338 | (void *)worker, worker->pt_id); |
| 1339 | int err_num = pthread_kill(worker->pt_id, grpc_wakeup_signal); |
| 1340 | if (err_num != 0) { |
| 1341 | err = GRPC_OS_ERROR(err_num, "pthread_kill"); |
| 1342 | } |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1343 | } |
| 1344 | return err; |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1345 | } |
| 1346 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1347 | /* Return 1 if the pollset has active threads in pollset_work (pollset must |
| 1348 | * be locked) */ |
| 1349 | static int pollset_has_workers(grpc_pollset *p) { |
| 1350 | return p->root_worker.next != &p->root_worker; |
| 1351 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1352 | |
| 1353 | static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { |
| 1354 | worker->prev->next = worker->next; |
| 1355 | worker->next->prev = worker->prev; |
| 1356 | } |
| 1357 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1358 | static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) { |
| 1359 | if (pollset_has_workers(p)) { |
| 1360 | grpc_pollset_worker *w = p->root_worker.next; |
| 1361 | remove_worker(p, w); |
| 1362 | return w; |
| 1363 | } else { |
| 1364 | return NULL; |
| 1365 | } |
| 1366 | } |
| 1367 | |
| 1368 | static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) { |
| 1369 | worker->next = &p->root_worker; |
| 1370 | worker->prev = worker->next->prev; |
| 1371 | worker->prev->next = worker->next->prev = worker; |
| 1372 | } |
| 1373 | |
| 1374 | static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) { |
| 1375 | worker->prev = &p->root_worker; |
| 1376 | worker->next = worker->prev->next; |
| 1377 | worker->prev->next = worker->next->prev = worker; |
| 1378 | } |
| 1379 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1380 | /* p->mu must be held before calling this function */ |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1381 | static grpc_error *pollset_kick(grpc_pollset *p, |
| 1382 | grpc_pollset_worker *specific_worker) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1383 | GPR_TIMER_BEGIN("pollset_kick", 0); |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1384 | grpc_error *error = GRPC_ERROR_NONE; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1385 | const char *err_desc = "Kick Failure"; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1386 | grpc_pollset_worker *worker = specific_worker; |
| 1387 | if (worker != NULL) { |
| 1388 | if (worker == GRPC_POLLSET_KICK_BROADCAST) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1389 | if (pollset_has_workers(p)) { |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1390 | GPR_TIMER_BEGIN("pollset_kick.broadcast", 0); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1391 | for (worker = p->root_worker.next; worker != &p->root_worker; |
| 1392 | worker = worker->next) { |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1393 | if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1394 | append_error(&error, pollset_worker_kick(worker), err_desc); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1395 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1396 | } |
Craig Tiller | a218a06 | 2016-06-26 09:58:37 -0700 | [diff] [blame] | 1397 | GPR_TIMER_END("pollset_kick.broadcast", 0); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1398 | } else { |
| 1399 | p->kicked_without_pollers = true; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1400 | } |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1401 | } else { |
| 1402 | GPR_TIMER_MARK("kicked_specifically", 0); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1403 | if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1404 | append_error(&error, pollset_worker_kick(worker), err_desc); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1405 | } |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1406 | } |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1407 | } else if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)p) { |
| 1408 | /* Since worker == NULL, it means that we can kick "any" worker on this |
| 1409 | pollset 'p'. If 'p' happens to be the same pollset this thread is |
| 1410 | currently polling (i.e in pollset_work() function), then there is no need |
| 1411 | to kick any other worker since the current thread can just absorb the |
| 1412 | kick. This is the reason why we enter this case only when |
| 1413 | g_current_thread_pollset is != p */ |
| 1414 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1415 | GPR_TIMER_MARK("kick_anonymous", 0); |
| 1416 | worker = pop_front_worker(p); |
| 1417 | if (worker != NULL) { |
| 1418 | GPR_TIMER_MARK("finally_kick", 0); |
| 1419 | push_back_worker(p, worker); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1420 | append_error(&error, pollset_worker_kick(worker), err_desc); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1421 | } else { |
| 1422 | GPR_TIMER_MARK("kicked_no_pollers", 0); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1423 | p->kicked_without_pollers = true; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1424 | } |
| 1425 | } |
| 1426 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1427 | GPR_TIMER_END("pollset_kick", 0); |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1428 | GRPC_LOG_IF_ERROR("pollset_kick", GRPC_ERROR_REF(error)); |
| 1429 | return error; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1430 | } |
| 1431 | |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1432 | static grpc_error *kick_poller(void) { |
Craig Tiller | b4b8e1e | 2016-11-28 07:33:13 -0800 | [diff] [blame] | 1433 | return grpc_wakeup_fd_wakeup(&global_wakeup_fd); |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1434 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1435 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1436 | static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1437 | gpr_mu_init(&pollset->po.mu); |
| 1438 | *mu = &pollset->po.mu; |
| 1439 | pollset->po.pi = NULL; |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1440 | #ifdef PO_DEBUG |
| 1441 | pollset->po.obj_type = POLL_OBJ_POLLSET; |
| 1442 | #endif |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1443 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1444 | pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1445 | pollset->kicked_without_pollers = false; |
| 1446 | |
| 1447 | pollset->shutting_down = false; |
| 1448 | pollset->finish_shutdown_called = false; |
| 1449 | pollset->shutdown_done = NULL; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1450 | } |
| 1451 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1452 | /* Convert a timespec to milliseconds: |
| 1453 | - Very small or negative poll times are clamped to zero to do a non-blocking |
| 1454 | poll (which becomes spin polling) |
| 1455 | - Other small values are rounded up to one millisecond |
| 1456 | - Longer than a millisecond polls are rounded up to the next nearest |
| 1457 | millisecond to avoid spinning |
| 1458 | - Infinite timeouts are converted to -1 */ |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1459 | static int poll_deadline_to_millis_timeout(gpr_timespec deadline, |
| 1460 | gpr_timespec now) { |
| 1461 | gpr_timespec timeout; |
| 1462 | static const int64_t max_spin_polling_us = 10; |
| 1463 | if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) { |
| 1464 | return -1; |
| 1465 | } |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1466 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1467 | if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros( |
| 1468 | max_spin_polling_us, |
| 1469 | GPR_TIMESPAN))) <= 0) { |
| 1470 | return 0; |
| 1471 | } |
| 1472 | timeout = gpr_time_sub(deadline, now); |
| 1473 | return gpr_time_to_millis(gpr_time_add( |
| 1474 | timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); |
| 1475 | } |
| 1476 | |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1477 | static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd, |
| 1478 | grpc_pollset *notifier) { |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1479 | set_ready(exec_ctx, fd, &fd->read_closure); |
| 1480 | |
Sree Kuchibhotla | 4db8c82 | 2017-02-12 17:07:31 -0800 | [diff] [blame] | 1481 | /* Note, it is possible that fd_become_readable might be called twice with |
Sree Kuchibhotla | 4c60d0d | 2017-02-12 17:09:08 -0800 | [diff] [blame] | 1482 | different 'notifier's when an fd becomes readable and it is in two epoll |
| 1483 | sets (This can happen briefly during polling island merges). In such cases |
| 1484 | it does not really matter which notifer is set as the read_notifier_pollset |
| 1485 | (They would both point to the same polling island anyway) */ |
Sree Kuchibhotla | fb7ced6 | 2017-02-15 18:31:57 -0800 | [diff] [blame] | 1486 | /* Use release store to match with acquire load in fd_get_read_notifier */ |
| 1487 | gpr_atm_rel_store(&fd->read_notifier_pollset, (gpr_atm)notifier); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1488 | } |
| 1489 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1490 | static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { |
Sree Kuchibhotla | 82d7341 | 2017-02-09 18:27:45 -0800 | [diff] [blame] | 1491 | set_ready(exec_ctx, fd, &fd->write_closure); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1492 | } |
| 1493 | |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 1494 | static void pollset_release_polling_island(grpc_exec_ctx *exec_ctx, |
| 1495 | grpc_pollset *ps, char *reason) { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1496 | if (ps->po.pi != NULL) { |
| 1497 | PI_UNREF(exec_ctx, ps->po.pi, reason); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1498 | } |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1499 | ps->po.pi = NULL; |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1500 | } |
| 1501 | |
| 1502 | static void finish_shutdown_locked(grpc_exec_ctx *exec_ctx, |
| 1503 | grpc_pollset *pollset) { |
| 1504 | /* The pollset cannot have any workers if we are at this stage */ |
| 1505 | GPR_ASSERT(!pollset_has_workers(pollset)); |
| 1506 | |
| 1507 | pollset->finish_shutdown_called = true; |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1508 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1509 | /* Release the ref and set pollset->po.pi to NULL */ |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 1510 | pollset_release_polling_island(exec_ctx, pollset, "ps_shutdown"); |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 1511 | grpc_closure_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1512 | } |
| 1513 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1514 | /* pollset->po.mu lock must be held by the caller before calling this */ |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1515 | static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, |
| 1516 | grpc_closure *closure) { |
| 1517 | GPR_TIMER_BEGIN("pollset_shutdown", 0); |
| 1518 | GPR_ASSERT(!pollset->shutting_down); |
| 1519 | pollset->shutting_down = true; |
| 1520 | pollset->shutdown_done = closure; |
| 1521 | pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); |
| 1522 | |
| 1523 | /* If the pollset has any workers, we cannot call finish_shutdown_locked() |
| 1524 | because it would release the underlying polling island. In such a case, we |
| 1525 | let the last worker call finish_shutdown_locked() from pollset_work() */ |
| 1526 | if (!pollset_has_workers(pollset)) { |
| 1527 | GPR_ASSERT(!pollset->finish_shutdown_called); |
| 1528 | GPR_TIMER_MARK("pollset_shutdown.finish_shutdown_locked", 0); |
| 1529 | finish_shutdown_locked(exec_ctx, pollset); |
| 1530 | } |
| 1531 | GPR_TIMER_END("pollset_shutdown", 0); |
| 1532 | } |
| 1533 | |
| 1534 | /* pollset_shutdown is guaranteed to be called before pollset_destroy. So other |
| 1535 | * than destroying the mutexes, there is nothing special that needs to be done |
| 1536 | * here */ |
| 1537 | static void pollset_destroy(grpc_pollset *pollset) { |
| 1538 | GPR_ASSERT(!pollset_has_workers(pollset)); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1539 | gpr_mu_destroy(&pollset->po.mu); |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1540 | } |
| 1541 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1542 | static bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx, |
| 1543 | polling_island *pi) { |
| 1544 | if (gpr_mu_trylock(&pi->workqueue_read_mu)) { |
| 1545 | gpr_mpscq_node *n = gpr_mpscq_pop(&pi->workqueue_items); |
| 1546 | gpr_mu_unlock(&pi->workqueue_read_mu); |
| 1547 | if (n != NULL) { |
| 1548 | if (gpr_atm_full_fetch_add(&pi->workqueue_item_count, -1) > 1) { |
| 1549 | workqueue_maybe_wakeup(pi); |
| 1550 | } |
| 1551 | grpc_closure *c = (grpc_closure *)n; |
Craig Tiller | 061ef74 | 2016-12-29 10:54:09 -0800 | [diff] [blame] | 1552 | grpc_error *error = c->error_data.error; |
| 1553 | c->cb(exec_ctx, c->cb_arg, error); |
| 1554 | GRPC_ERROR_UNREF(error); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1555 | return true; |
| 1556 | } else if (gpr_atm_no_barrier_load(&pi->workqueue_item_count) > 0) { |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 1557 | /* n == NULL might mean there's work but it's not available to be popped |
| 1558 | * yet - try to ensure another workqueue wakes up to check shortly if so |
| 1559 | */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1560 | workqueue_maybe_wakeup(pi); |
| 1561 | } |
| 1562 | } |
| 1563 | return false; |
| 1564 | } |
| 1565 | |
Craig Tiller | 84ea341 | 2016-09-08 14:57:56 -0700 | [diff] [blame] | 1566 | #define GRPC_EPOLL_MAX_EVENTS 100 |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1567 | /* Note: sig_mask contains the signal mask to use *during* epoll_wait() */ |
| 1568 | static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1569 | grpc_pollset *pollset, |
| 1570 | grpc_pollset_worker *worker, int timeout_ms, |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1571 | sigset_t *sig_mask, grpc_error **error) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1572 | struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; |
Sree Kuchibhotla | 88ee12f | 2016-06-03 19:26:48 -0700 | [diff] [blame] | 1573 | int epoll_fd = -1; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1574 | int ep_rv; |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1575 | polling_island *pi = NULL; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1576 | char *err_msg; |
| 1577 | const char *err_desc = "pollset_work_and_unlock"; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1578 | GPR_TIMER_BEGIN("pollset_work_and_unlock", 0); |
| 1579 | |
| 1580 | /* We need to get the epoll_fd to wait on. The epoll_fd is in inside the |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1581 | latest polling island pointed by pollset->po.pi |
Sree Kuchibhotla | 229533b1 | 2016-06-21 20:42:52 -0700 | [diff] [blame] | 1582 | |
| 1583 | Since epoll_fd is immutable, we can read it without obtaining the polling |
| 1584 | island lock. There is however a possibility that the polling island (from |
| 1585 | which we got the epoll_fd) got merged with another island while we are |
| 1586 | in this function. This is still okay because in such a case, we will wakeup |
| 1587 | right-away from epoll_wait() and pick up the latest polling_island the next |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1588 | this function (i.e pollset_work_and_unlock()) is called */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1589 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1590 | if (pollset->po.pi == NULL) { |
| 1591 | pollset->po.pi = polling_island_create(exec_ctx, NULL, error); |
| 1592 | if (pollset->po.pi == NULL) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1593 | GPR_TIMER_END("pollset_work_and_unlock", 0); |
| 1594 | return; /* Fatal error. We cannot continue */ |
| 1595 | } |
| 1596 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1597 | PI_ADD_REF(pollset->po.pi, "ps"); |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1598 | GRPC_POLLING_TRACE("pollset_work: pollset: %p created new pi: %p", |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1599 | (void *)pollset, (void *)pollset->po.pi); |
Sree Kuchibhotla | 88ee12f | 2016-06-03 19:26:48 -0700 | [diff] [blame] | 1600 | } |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1601 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1602 | pi = polling_island_maybe_get_latest(pollset->po.pi); |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1603 | epoll_fd = pi->epoll_fd; |
| 1604 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1605 | /* Update the pollset->po.pi since the island being pointed by |
| 1606 | pollset->po.pi maybe older than the one pointed by pi) */ |
| 1607 | if (pollset->po.pi != pi) { |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1608 | /* Always do PI_ADD_REF before PI_UNREF because PI_UNREF may cause the |
| 1609 | polling island to be deleted */ |
| 1610 | PI_ADD_REF(pi, "ps"); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1611 | PI_UNREF(exec_ctx, pollset->po.pi, "ps"); |
| 1612 | pollset->po.pi = pi; |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1613 | } |
Sree Kuchibhotla | d627c10 | 2016-06-06 15:49:32 -0700 | [diff] [blame] | 1614 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1615 | /* Add an extra ref so that the island does not get destroyed (which means |
| 1616 | the epoll_fd won't be closed) while we are are doing an epoll_wait() on the |
| 1617 | epoll_fd */ |
| 1618 | PI_ADD_REF(pi, "ps_work"); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1619 | gpr_mu_unlock(&pollset->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1620 | |
Craig Tiller | 460502e | 2016-10-13 10:02:08 -0700 | [diff] [blame] | 1621 | /* If we get some workqueue work to do, it might end up completing an item on |
| 1622 | the completion queue, so there's no need to poll... so we skip that and |
| 1623 | redo the complete loop to verify */ |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1624 | if (!maybe_do_workqueue_work(exec_ctx, pi)) { |
| 1625 | gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); |
| 1626 | g_current_thread_polling_island = pi; |
| 1627 | |
Vijay Pai | cef5401 | 2016-08-28 23:05:31 -0700 | [diff] [blame] | 1628 | GRPC_SCHEDULING_START_BLOCKING_REGION; |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1629 | ep_rv = epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, |
| 1630 | sig_mask); |
Vijay Pai | cef5401 | 2016-08-28 23:05:31 -0700 | [diff] [blame] | 1631 | GRPC_SCHEDULING_END_BLOCKING_REGION; |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1632 | if (ep_rv < 0) { |
| 1633 | if (errno != EINTR) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1634 | gpr_asprintf(&err_msg, |
| 1635 | "epoll_wait() epoll fd: %d failed with error: %d (%s)", |
| 1636 | epoll_fd, errno, strerror(errno)); |
| 1637 | append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1638 | } else { |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1639 | /* We were interrupted. Save an interation by doing a zero timeout |
| 1640 | epoll_wait to see if there are any other events of interest */ |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1641 | GRPC_POLLING_TRACE( |
| 1642 | "pollset_work: pollset: %p, worker: %p received kick", |
| 1643 | (void *)pollset, (void *)worker); |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1644 | ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1645 | } |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1646 | } |
Sree Kuchibhotla | 79a6233 | 2016-06-04 14:01:03 -0700 | [diff] [blame] | 1647 | |
Sree Kuchibhotla | ad2c477 | 2016-06-13 19:06:54 -0700 | [diff] [blame] | 1648 | #ifdef GRPC_TSAN |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 1649 | /* See the definition of g_poll_sync for more details */ |
| 1650 | gpr_atm_acq_load(&g_epoll_sync); |
Sree Kuchibhotla | ad2c477 | 2016-06-13 19:06:54 -0700 | [diff] [blame] | 1651 | #endif /* defined(GRPC_TSAN) */ |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 1652 | |
Sree Kuchibhotla | 58e5896 | 2016-06-13 00:52:56 -0700 | [diff] [blame] | 1653 | for (int i = 0; i < ep_rv; ++i) { |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1654 | void *data_ptr = ep_ev[i].data.ptr; |
Craig Tiller | b4b8e1e | 2016-11-28 07:33:13 -0800 | [diff] [blame] | 1655 | if (data_ptr == &global_wakeup_fd) { |
Craig Tiller | 1fa9ddb | 2016-11-28 08:19:37 -0800 | [diff] [blame] | 1656 | append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd), |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1657 | err_desc); |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1658 | } else if (data_ptr == &pi->workqueue_wakeup_fd) { |
Craig Tiller | e49959d | 2017-01-26 08:39:38 -0800 | [diff] [blame] | 1659 | append_error(error, |
| 1660 | grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd), |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1661 | err_desc); |
| 1662 | maybe_do_workqueue_work(exec_ctx, pi); |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1663 | } else if (data_ptr == &polling_island_wakeup_fd) { |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1664 | GRPC_POLLING_TRACE( |
| 1665 | "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " |
| 1666 | "%d) got merged", |
| 1667 | (void *)pollset, (void *)worker, epoll_fd); |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1668 | /* This means that our polling island is merged with a different |
| 1669 | island. We do not have to do anything here since the subsequent call |
| 1670 | to the function pollset_work_and_unlock() will pick up the correct |
| 1671 | epoll_fd */ |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1672 | } else { |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1673 | grpc_fd *fd = data_ptr; |
| 1674 | int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP); |
| 1675 | int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI); |
| 1676 | int write_ev = ep_ev[i].events & EPOLLOUT; |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1677 | if (read_ev || cancel) { |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 1678 | fd_become_readable(exec_ctx, fd, pollset); |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1679 | } |
| 1680 | if (write_ev || cancel) { |
| 1681 | fd_become_writable(exec_ctx, fd); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1682 | } |
| 1683 | } |
Sree Kuchibhotla | e5012ba | 2016-06-06 16:01:45 -0700 | [diff] [blame] | 1684 | } |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 1685 | |
| 1686 | g_current_thread_polling_island = NULL; |
| 1687 | gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); |
| 1688 | } |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1689 | |
| 1690 | GPR_ASSERT(pi != NULL); |
| 1691 | |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1692 | /* Before leaving, release the extra ref we added to the polling island. It |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1693 | is important to use "pi" here (i.e our old copy of pollset->po.pi |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1694 | that we got before releasing the polling island lock). This is because |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1695 | pollset->po.pi pointer might get udpated in other parts of the |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1696 | code when there is an island merge while we are doing epoll_wait() above */ |
Craig Tiller | b39307d | 2016-06-30 15:39:13 -0700 | [diff] [blame] | 1697 | PI_UNREF(exec_ctx, pi, "ps_work"); |
Sree Kuchibhotla | 24b1062 | 2016-06-08 15:20:17 -0700 | [diff] [blame] | 1698 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1699 | GPR_TIMER_END("pollset_work_and_unlock", 0); |
| 1700 | } |
| 1701 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1702 | /* pollset->po.mu lock must be held by the caller before calling this. |
| 1703 | The function pollset_work() may temporarily release the lock (pollset->po.mu) |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1704 | during the course of its execution but it will always re-acquire the lock and |
| 1705 | ensure that it is held by the time the function returns */ |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1706 | static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, |
| 1707 | grpc_pollset_worker **worker_hdl, |
| 1708 | gpr_timespec now, gpr_timespec deadline) { |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1709 | GPR_TIMER_BEGIN("pollset_work", 0); |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1710 | grpc_error *error = GRPC_ERROR_NONE; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1711 | int timeout_ms = poll_deadline_to_millis_timeout(deadline, now); |
| 1712 | |
| 1713 | sigset_t new_mask; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1714 | |
| 1715 | grpc_pollset_worker worker; |
| 1716 | worker.next = worker.prev = NULL; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1717 | worker.pt_id = pthread_self(); |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 1718 | gpr_atm_no_barrier_store(&worker.is_kicked, (gpr_atm)0); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1719 | |
| 1720 | *worker_hdl = &worker; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1721 | |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1722 | gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset); |
| 1723 | gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1724 | |
| 1725 | if (pollset->kicked_without_pollers) { |
| 1726 | /* If the pollset was kicked without pollers, pretend that the current |
| 1727 | worker got the kick and skip polling. A kick indicates that there is some |
| 1728 | work that needs attention like an event on the completion queue or an |
| 1729 | alarm */ |
| 1730 | GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0); |
| 1731 | pollset->kicked_without_pollers = 0; |
| 1732 | } else if (!pollset->shutting_down) { |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1733 | /* We use the posix-signal with number 'grpc_wakeup_signal' for waking up |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 1734 | (i.e 'kicking') a worker in the pollset. A 'kick' is a way to inform the |
| 1735 | worker that there is some pending work that needs immediate attention |
| 1736 | (like an event on the completion queue, or a polling island merge that |
| 1737 | results in a new epoll-fd to wait on) and that the worker should not |
| 1738 | spend time waiting in epoll_pwait(). |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1739 | |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 1740 | A worker can be kicked anytime from the point it is added to the pollset |
| 1741 | via push_front_worker() (or push_back_worker()) to the point it is |
| 1742 | removed via remove_worker(). |
| 1743 | If the worker is kicked before/during it calls epoll_pwait(), it should |
| 1744 | immediately exit from epoll_wait(). If the worker is kicked after it |
| 1745 | returns from epoll_wait(), then nothing really needs to be done. |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1746 | |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 1747 | To accomplish this, we mask 'grpc_wakeup_signal' on this thread at all |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1748 | times *except* when it is in epoll_pwait(). This way, the worker never |
| 1749 | misses acting on a kick */ |
| 1750 | |
Craig Tiller | 1919699 | 2016-06-27 18:45:56 -0700 | [diff] [blame] | 1751 | if (!g_initialized_sigmask) { |
| 1752 | sigemptyset(&new_mask); |
| 1753 | sigaddset(&new_mask, grpc_wakeup_signal); |
| 1754 | pthread_sigmask(SIG_BLOCK, &new_mask, &g_orig_sigmask); |
| 1755 | sigdelset(&g_orig_sigmask, grpc_wakeup_signal); |
| 1756 | g_initialized_sigmask = true; |
| 1757 | /* new_mask: The new thread mask which blocks 'grpc_wakeup_signal'. |
| 1758 | This is the mask used at all times *except during |
| 1759 | epoll_wait()*" |
| 1760 | g_orig_sigmask: The thread mask which allows 'grpc_wakeup_signal' and |
Craig Tiller | 510ff69 | 2016-06-27 20:31:49 -0700 | [diff] [blame] | 1761 | this is the mask to use *during epoll_wait()* |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1762 | |
Craig Tiller | 1919699 | 2016-06-27 18:45:56 -0700 | [diff] [blame] | 1763 | The new_mask is set on the worker before it is added to the pollset |
Craig Tiller | 510ff69 | 2016-06-27 20:31:49 -0700 | [diff] [blame] | 1764 | (i.e before it can be kicked) */ |
Craig Tiller | 1919699 | 2016-06-27 18:45:56 -0700 | [diff] [blame] | 1765 | } |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1766 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1767 | push_front_worker(pollset, &worker); /* Add worker to pollset */ |
| 1768 | |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1769 | pollset_work_and_unlock(exec_ctx, pollset, &worker, timeout_ms, |
| 1770 | &g_orig_sigmask, &error); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1771 | grpc_exec_ctx_flush(exec_ctx); |
| 1772 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1773 | gpr_mu_lock(&pollset->po.mu); |
Sree Kuchibhotla | 3421724 | 2016-06-29 00:19:07 -0700 | [diff] [blame] | 1774 | |
| 1775 | /* Note: There is no need to reset worker.is_kicked to 0 since we are no |
| 1776 | longer going to use this worker */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1777 | remove_worker(pollset, &worker); |
| 1778 | } |
| 1779 | |
| 1780 | /* If we are the last worker on the pollset (i.e pollset_has_workers() is |
| 1781 | false at this point) and the pollset is shutting down, we may have to |
| 1782 | finish the shutdown process by calling finish_shutdown_locked(). |
| 1783 | See pollset_shutdown() for more details. |
| 1784 | |
| 1785 | Note: Continuing to access pollset here is safe; it is the caller's |
| 1786 | responsibility to not destroy a pollset when it has outstanding calls to |
| 1787 | pollset_work() */ |
| 1788 | if (pollset->shutting_down && !pollset_has_workers(pollset) && |
| 1789 | !pollset->finish_shutdown_called) { |
| 1790 | GPR_TIMER_MARK("pollset_work.finish_shutdown_locked", 0); |
| 1791 | finish_shutdown_locked(exec_ctx, pollset); |
| 1792 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1793 | gpr_mu_unlock(&pollset->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1794 | grpc_exec_ctx_flush(exec_ctx); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1795 | gpr_mu_lock(&pollset->po.mu); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1796 | } |
| 1797 | |
| 1798 | *worker_hdl = NULL; |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1799 | |
Sree Kuchibhotla | 8e4926c | 2016-06-08 20:33:19 -0700 | [diff] [blame] | 1800 | gpr_tls_set(&g_current_thread_pollset, (intptr_t)0); |
| 1801 | gpr_tls_set(&g_current_thread_worker, (intptr_t)0); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1802 | |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1803 | GPR_TIMER_END("pollset_work", 0); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1804 | |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 1805 | GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error)); |
| 1806 | return error; |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1807 | } |
| 1808 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1809 | static void add_poll_object(grpc_exec_ctx *exec_ctx, poll_obj *bag, |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1810 | poll_obj_type bag_type, poll_obj *item, |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1811 | poll_obj_type item_type) { |
| 1812 | GPR_TIMER_BEGIN("add_poll_object", 0); |
| 1813 | |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1814 | #ifdef PO_DEBUG |
| 1815 | GPR_ASSERT(item->obj_type == item_type); |
| 1816 | GPR_ASSERT(bag->obj_type == bag_type); |
| 1817 | #endif |
Craig Tiller | 57726ca | 2016-09-12 11:59:45 -0700 | [diff] [blame] | 1818 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1819 | grpc_error *error = GRPC_ERROR_NONE; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 1820 | polling_island *pi_new = NULL; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1821 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1822 | gpr_mu_lock(&bag->mu); |
| 1823 | gpr_mu_lock(&item->mu); |
| 1824 | |
Craig Tiller | 7212c23 | 2016-07-06 13:11:09 -0700 | [diff] [blame] | 1825 | retry: |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1826 | /* |
| 1827 | * 1) If item->pi and bag->pi are both non-NULL and equal, do nothing |
| 1828 | * 2) If item->pi and bag->pi are both NULL, create a new polling island (with |
| 1829 | * a refcount of 2) and point item->pi and bag->pi to the new island |
| 1830 | * 3) If exactly one of item->pi or bag->pi is NULL, update it to point to |
| 1831 | * the other's non-NULL pi |
| 1832 | * 4) Finally if item->pi and bag-pi are non-NULL and not-equal, merge the |
| 1833 | * polling islands and update item->pi and bag->pi to point to the new |
| 1834 | * island |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1835 | */ |
Craig Tiller | 8e8027b | 2016-07-07 10:42:09 -0700 | [diff] [blame] | 1836 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1837 | /* Early out if we are trying to add an 'fd' to a 'bag' but the fd is already |
| 1838 | * orphaned */ |
| 1839 | if (item_type == POLL_OBJ_FD && (FD_FROM_PO(item))->orphaned) { |
| 1840 | gpr_mu_unlock(&item->mu); |
| 1841 | gpr_mu_unlock(&bag->mu); |
Craig Tiller | 42ac6db | 2016-07-06 17:13:56 -0700 | [diff] [blame] | 1842 | return; |
| 1843 | } |
| 1844 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1845 | if (item->pi == bag->pi) { |
| 1846 | pi_new = item->pi; |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 1847 | if (pi_new == NULL) { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1848 | /* GPR_ASSERT(item->pi == bag->pi == NULL) */ |
Sree Kuchibhotla | a129adf | 2016-10-26 16:44:44 -0700 | [diff] [blame] | 1849 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1850 | /* If we are adding an fd to a bag (i.e pollset or pollset_set), then |
| 1851 | * we need to do some extra work to make TSAN happy */ |
| 1852 | if (item_type == POLL_OBJ_FD) { |
| 1853 | /* Unlock before creating a new polling island: the polling island will |
| 1854 | create a workqueue which creates a file descriptor, and holding an fd |
| 1855 | lock here can eventually cause a loop to appear to TSAN (making it |
| 1856 | unhappy). We don't think it's a real loop (there's an epoch point |
| 1857 | where that loop possibility disappears), but the advantages of |
| 1858 | keeping TSAN happy outweigh any performance advantage we might have |
| 1859 | by keeping the lock held. */ |
| 1860 | gpr_mu_unlock(&item->mu); |
| 1861 | pi_new = polling_island_create(exec_ctx, FD_FROM_PO(item), &error); |
| 1862 | gpr_mu_lock(&item->mu); |
Sree Kuchibhotla | a129adf | 2016-10-26 16:44:44 -0700 | [diff] [blame] | 1863 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1864 | /* Need to reverify any assumptions made between the initial lock and |
| 1865 | getting to this branch: if they've changed, we need to throw away our |
| 1866 | work and figure things out again. */ |
| 1867 | if (item->pi != NULL) { |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1868 | GRPC_POLLING_TRACE( |
| 1869 | "add_poll_object: Raced creating new polling island. pi_new: %p " |
| 1870 | "(fd: %d, %s: %p)", |
| 1871 | (void *)pi_new, FD_FROM_PO(item)->fd, poll_obj_string(bag_type), |
| 1872 | (void *)bag); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1873 | /* No need to lock 'pi_new' here since this is a new polling island |
Sree Kuchibhotla | 8214b87 | 2017-02-23 12:49:48 -0800 | [diff] [blame^] | 1874 | and no one has a reference to it yet */ |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1875 | polling_island_remove_all_fds_locked(pi_new, true, &error); |
| 1876 | |
| 1877 | /* Ref and unref so that the polling island gets deleted during unref |
| 1878 | */ |
| 1879 | PI_ADD_REF(pi_new, "dance_of_destruction"); |
| 1880 | PI_UNREF(exec_ctx, pi_new, "dance_of_destruction"); |
| 1881 | goto retry; |
| 1882 | } |
Craig Tiller | 27da642 | 2016-07-06 13:14:46 -0700 | [diff] [blame] | 1883 | } else { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1884 | pi_new = polling_island_create(exec_ctx, NULL, &error); |
Craig Tiller | 7212c23 | 2016-07-06 13:11:09 -0700 | [diff] [blame] | 1885 | } |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1886 | |
| 1887 | GRPC_POLLING_TRACE( |
| 1888 | "add_poll_object: Created new polling island. pi_new: %p (%s: %p, " |
| 1889 | "%s: %p)", |
| 1890 | (void *)pi_new, poll_obj_string(item_type), (void *)item, |
| 1891 | poll_obj_string(bag_type), (void *)bag); |
| 1892 | } else { |
| 1893 | GRPC_POLLING_TRACE( |
| 1894 | "add_poll_object: Same polling island. pi: %p (%s, %s)", |
| 1895 | (void *)pi_new, poll_obj_string(item_type), |
| 1896 | poll_obj_string(bag_type)); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1897 | } |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1898 | } else if (item->pi == NULL) { |
| 1899 | /* GPR_ASSERT(bag->pi != NULL) */ |
| 1900 | /* Make pi_new point to latest pi*/ |
| 1901 | pi_new = polling_island_lock(bag->pi); |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1902 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1903 | if (item_type == POLL_OBJ_FD) { |
| 1904 | grpc_fd *fd = FD_FROM_PO(item); |
| 1905 | polling_island_add_fds_locked(pi_new, &fd, 1, true, &error); |
| 1906 | } |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1907 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1908 | gpr_mu_unlock(&pi_new->mu); |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1909 | GRPC_POLLING_TRACE( |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1910 | "add_poll_obj: item->pi was NULL. pi_new: %p (item(%s): %p, " |
| 1911 | "bag(%s): %p)", |
| 1912 | (void *)pi_new, poll_obj_string(item_type), (void *)item, |
| 1913 | poll_obj_string(bag_type), (void *)bag); |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1914 | } else if (bag->pi == NULL) { |
| 1915 | /* GPR_ASSERT(item->pi != NULL) */ |
| 1916 | /* Make pi_new to point to latest pi */ |
| 1917 | pi_new = polling_island_lock(item->pi); |
| 1918 | gpr_mu_unlock(&pi_new->mu); |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1919 | GRPC_POLLING_TRACE( |
| 1920 | "add_poll_obj: bag->pi was NULL. pi_new: %p (item(%s): %p, " |
| 1921 | "bag(%s): %p)", |
| 1922 | (void *)pi_new, poll_obj_string(item_type), (void *)item, |
| 1923 | poll_obj_string(bag_type), (void *)bag); |
Sree Kuchibhotla | 5098f91 | 2016-05-31 10:58:17 -0700 | [diff] [blame] | 1924 | } else { |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1925 | pi_new = polling_island_merge(item->pi, bag->pi, &error); |
Sree Kuchibhotla | 1e77668 | 2016-06-28 14:09:26 -0700 | [diff] [blame] | 1926 | GRPC_POLLING_TRACE( |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1927 | "add_poll_obj: polling islands merged. pi_new: %p (item(%s): %p, " |
| 1928 | "bag(%s): %p)", |
| 1929 | (void *)pi_new, poll_obj_string(item_type), (void *)item, |
| 1930 | poll_obj_string(bag_type), (void *)bag); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1931 | } |
| 1932 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1933 | /* At this point, pi_new is the polling island that both item->pi and bag->pi |
| 1934 | MUST be pointing to */ |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 1935 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1936 | if (item->pi != pi_new) { |
| 1937 | PI_ADD_REF(pi_new, poll_obj_string(item_type)); |
| 1938 | if (item->pi != NULL) { |
| 1939 | PI_UNREF(exec_ctx, item->pi, poll_obj_string(item_type)); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1940 | } |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1941 | item->pi = pi_new; |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1942 | } |
| 1943 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1944 | if (bag->pi != pi_new) { |
| 1945 | PI_ADD_REF(pi_new, poll_obj_string(bag_type)); |
| 1946 | if (bag->pi != NULL) { |
| 1947 | PI_UNREF(exec_ctx, bag->pi, poll_obj_string(bag_type)); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1948 | } |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1949 | bag->pi = pi_new; |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 1950 | } |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1951 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1952 | gpr_mu_unlock(&item->mu); |
| 1953 | gpr_mu_unlock(&bag->mu); |
Craig Tiller | 1500761 | 2016-07-06 09:36:16 -0700 | [diff] [blame] | 1954 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 1955 | GRPC_LOG_IF_ERROR("add_poll_object", error); |
| 1956 | GPR_TIMER_END("add_poll_object", 0); |
| 1957 | } |
Craig Tiller | 57726ca | 2016-09-12 11:59:45 -0700 | [diff] [blame] | 1958 | |
Sree Kuchibhotla | 9442bab | 2016-05-20 17:54:06 -0700 | [diff] [blame] | 1959 | static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, |
| 1960 | grpc_fd *fd) { |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1961 | add_poll_object(exec_ctx, &pollset->po, POLL_OBJ_POLLSET, &fd->po, |
Sree Kuchibhotla | 499b94b | 2016-11-18 14:35:47 -0800 | [diff] [blame] | 1962 | POLL_OBJ_FD); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1963 | } |
| 1964 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1965 | /******************************************************************************* |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1966 | * Pollset-set Definitions |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1967 | */ |
| 1968 | |
| 1969 | static grpc_pollset_set *pollset_set_create(void) { |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1970 | grpc_pollset_set *pss = gpr_malloc(sizeof(*pss)); |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1971 | gpr_mu_init(&pss->po.mu); |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1972 | pss->po.pi = NULL; |
| 1973 | #ifdef PO_DEBUG |
| 1974 | pss->po.obj_type = POLL_OBJ_POLLSET_SET; |
| 1975 | #endif |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1976 | return pss; |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1977 | } |
| 1978 | |
Craig Tiller | 9e5ac1b | 2017-02-14 22:25:50 -0800 | [diff] [blame] | 1979 | static void pollset_set_destroy(grpc_exec_ctx *exec_ctx, |
| 1980 | grpc_pollset_set *pss) { |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1981 | gpr_mu_destroy(&pss->po.mu); |
| 1982 | |
| 1983 | if (pss->po.pi != NULL) { |
Craig Tiller | 9e5ac1b | 2017-02-14 22:25:50 -0800 | [diff] [blame] | 1984 | PI_UNREF(exec_ctx, pss->po.pi, "pss_destroy"); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1985 | } |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1986 | |
| 1987 | gpr_free(pss); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 1988 | } |
| 1989 | |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1990 | static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss, |
| 1991 | grpc_fd *fd) { |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 1992 | add_poll_object(exec_ctx, &pss->po, POLL_OBJ_POLLSET_SET, &fd->po, |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1993 | POLL_OBJ_FD); |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1994 | } |
| 1995 | |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 1996 | static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss, |
| 1997 | grpc_fd *fd) { |
| 1998 | /* Nothing to do */ |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 1999 | } |
| 2000 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2001 | static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 2002 | grpc_pollset_set *pss, grpc_pollset *ps) { |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 2003 | add_poll_object(exec_ctx, &pss->po, POLL_OBJ_POLLSET_SET, &ps->po, |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 2004 | POLL_OBJ_POLLSET); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2005 | } |
| 2006 | |
| 2007 | static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 2008 | grpc_pollset_set *pss, grpc_pollset *ps) { |
| 2009 | /* Nothing to do */ |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2010 | } |
| 2011 | |
| 2012 | static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, |
| 2013 | grpc_pollset_set *bag, |
| 2014 | grpc_pollset_set *item) { |
Sree Kuchibhotla | a0749a6 | 2016-11-18 20:22:09 -0800 | [diff] [blame] | 2015 | add_poll_object(exec_ctx, &bag->po, POLL_OBJ_POLLSET_SET, &item->po, |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 2016 | POLL_OBJ_POLLSET_SET); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2017 | } |
| 2018 | |
| 2019 | static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, |
| 2020 | grpc_pollset_set *bag, |
| 2021 | grpc_pollset_set *item) { |
Sree Kuchibhotla | 2385fd7 | 2016-11-18 16:30:41 -0800 | [diff] [blame] | 2022 | /* Nothing to do */ |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2023 | } |
| 2024 | |
Sree Kuchibhotla | 2e12db9 | 2016-06-16 16:53:59 -0700 | [diff] [blame] | 2025 | /* Test helper functions |
| 2026 | * */ |
| 2027 | void *grpc_fd_get_polling_island(grpc_fd *fd) { |
| 2028 | polling_island *pi; |
| 2029 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 2030 | gpr_mu_lock(&fd->po.mu); |
| 2031 | pi = fd->po.pi; |
| 2032 | gpr_mu_unlock(&fd->po.mu); |
Sree Kuchibhotla | 2e12db9 | 2016-06-16 16:53:59 -0700 | [diff] [blame] | 2033 | |
| 2034 | return pi; |
| 2035 | } |
| 2036 | |
| 2037 | void *grpc_pollset_get_polling_island(grpc_pollset *ps) { |
| 2038 | polling_island *pi; |
| 2039 | |
Sree Kuchibhotla | f6f33d7 | 2016-11-18 11:38:52 -0800 | [diff] [blame] | 2040 | gpr_mu_lock(&ps->po.mu); |
| 2041 | pi = ps->po.pi; |
| 2042 | gpr_mu_unlock(&ps->po.mu); |
Sree Kuchibhotla | 2e12db9 | 2016-06-16 16:53:59 -0700 | [diff] [blame] | 2043 | |
| 2044 | return pi; |
| 2045 | } |
| 2046 | |
Sree Kuchibhotla | 2e12db9 | 2016-06-16 16:53:59 -0700 | [diff] [blame] | 2047 | bool grpc_are_polling_islands_equal(void *p, void *q) { |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 2048 | polling_island *p1 = p; |
| 2049 | polling_island *p2 = q; |
| 2050 | |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 2051 | /* Note: polling_island_lock_pair() may change p1 and p2 to point to the |
| 2052 | latest polling islands in their respective linked lists */ |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 2053 | polling_island_lock_pair(&p1, &p2); |
Sree Kuchibhotla | 20d0a16 | 2016-06-23 15:14:03 -0700 | [diff] [blame] | 2054 | polling_island_unlock_pair(p1, p2); |
Sree Kuchibhotla | 2f8ade0 | 2016-06-17 13:28:38 -0700 | [diff] [blame] | 2055 | |
| 2056 | return p1 == p2; |
Sree Kuchibhotla | 2e12db9 | 2016-06-16 16:53:59 -0700 | [diff] [blame] | 2057 | } |
| 2058 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2059 | /******************************************************************************* |
Sree Kuchibhotla | 0bcbd79 | 2016-06-01 15:43:03 -0700 | [diff] [blame] | 2060 | * Event engine binding |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2061 | */ |
| 2062 | |
| 2063 | static void shutdown_engine(void) { |
| 2064 | fd_global_shutdown(); |
| 2065 | pollset_global_shutdown(); |
Sree Kuchibhotla | d627c10 | 2016-06-06 15:49:32 -0700 | [diff] [blame] | 2066 | polling_island_global_shutdown(); |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2067 | } |
| 2068 | |
| 2069 | static const grpc_event_engine_vtable vtable = { |
| 2070 | .pollset_size = sizeof(grpc_pollset), |
| 2071 | |
| 2072 | .fd_create = fd_create, |
| 2073 | .fd_wrapped_fd = fd_wrapped_fd, |
| 2074 | .fd_orphan = fd_orphan, |
| 2075 | .fd_shutdown = fd_shutdown, |
Sree Kuchibhotla | 24b6eae | 2016-06-21 18:01:14 -0700 | [diff] [blame] | 2076 | .fd_is_shutdown = fd_is_shutdown, |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2077 | .fd_notify_on_read = fd_notify_on_read, |
| 2078 | .fd_notify_on_write = fd_notify_on_write, |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 2079 | .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset, |
Craig Tiller | 70bd483 | 2016-06-30 14:20:46 -0700 | [diff] [blame] | 2080 | .fd_get_workqueue = fd_get_workqueue, |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2081 | |
| 2082 | .pollset_init = pollset_init, |
| 2083 | .pollset_shutdown = pollset_shutdown, |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2084 | .pollset_destroy = pollset_destroy, |
| 2085 | .pollset_work = pollset_work, |
| 2086 | .pollset_kick = pollset_kick, |
| 2087 | .pollset_add_fd = pollset_add_fd, |
| 2088 | |
| 2089 | .pollset_set_create = pollset_set_create, |
| 2090 | .pollset_set_destroy = pollset_set_destroy, |
| 2091 | .pollset_set_add_pollset = pollset_set_add_pollset, |
| 2092 | .pollset_set_del_pollset = pollset_set_del_pollset, |
| 2093 | .pollset_set_add_pollset_set = pollset_set_add_pollset_set, |
| 2094 | .pollset_set_del_pollset_set = pollset_set_del_pollset_set, |
| 2095 | .pollset_set_add_fd = pollset_set_add_fd, |
| 2096 | .pollset_set_del_fd = pollset_set_del_fd, |
| 2097 | |
| 2098 | .kick_poller = kick_poller, |
| 2099 | |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 2100 | .workqueue_ref = workqueue_ref, |
| 2101 | .workqueue_unref = workqueue_unref, |
Craig Tiller | 91031da | 2016-12-28 15:44:25 -0800 | [diff] [blame] | 2102 | .workqueue_scheduler = workqueue_scheduler, |
Craig Tiller | d8a3c04 | 2016-09-09 12:42:37 -0700 | [diff] [blame] | 2103 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2104 | .shutdown_engine = shutdown_engine, |
| 2105 | }; |
| 2106 | |
Sree Kuchibhotla | 7274402 | 2016-06-09 09:42:06 -0700 | [diff] [blame] | 2107 | /* It is possible that GLIBC has epoll but the underlying kernel doesn't. |
| 2108 | * Create a dummy epoll_fd to make sure epoll support is available */ |
| 2109 | static bool is_epoll_available() { |
| 2110 | int fd = epoll_create1(EPOLL_CLOEXEC); |
| 2111 | if (fd < 0) { |
| 2112 | gpr_log( |
| 2113 | GPR_ERROR, |
| 2114 | "epoll_create1 failed with error: %d. Not using epoll polling engine", |
| 2115 | fd); |
| 2116 | return false; |
| 2117 | } |
| 2118 | close(fd); |
| 2119 | return true; |
| 2120 | } |
| 2121 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2122 | const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 2123 | /* If use of signals is disabled, we cannot use epoll engine*/ |
| 2124 | if (is_grpc_wakeup_signal_initialized && grpc_wakeup_signal < 0) { |
| 2125 | return NULL; |
| 2126 | } |
| 2127 | |
Ken Payson | cd7d047 | 2016-10-11 12:24:20 -0700 | [diff] [blame] | 2128 | if (!grpc_has_wakeup_fd()) { |
Ken Payson | bc544be | 2016-10-06 19:23:47 -0700 | [diff] [blame] | 2129 | return NULL; |
| 2130 | } |
| 2131 | |
Sree Kuchibhotla | 7274402 | 2016-06-09 09:42:06 -0700 | [diff] [blame] | 2132 | if (!is_epoll_available()) { |
| 2133 | return NULL; |
| 2134 | } |
| 2135 | |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 2136 | if (!is_grpc_wakeup_signal_initialized) { |
Sree Kuchibhotla | bd48c91 | 2016-09-27 16:48:25 -0700 | [diff] [blame] | 2137 | grpc_use_signal(SIGRTMIN + 6); |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 2138 | } |
| 2139 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2140 | fd_global_init(); |
Sree Kuchibhotla | 3131c26 | 2016-06-21 17:28:28 -0700 | [diff] [blame] | 2141 | |
| 2142 | if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { |
| 2143 | return NULL; |
| 2144 | } |
| 2145 | |
| 2146 | if (!GRPC_LOG_IF_ERROR("polling_island_global_init", |
| 2147 | polling_island_global_init())) { |
| 2148 | return NULL; |
| 2149 | } |
| 2150 | |
Sree Kuchibhotla | f448c34 | 2016-05-19 10:51:24 -0700 | [diff] [blame] | 2151 | return &vtable; |
| 2152 | } |
| 2153 | |
murgatroid99 | 623dd4f | 2016-08-08 17:31:27 -0700 | [diff] [blame] | 2154 | #else /* defined(GRPC_LINUX_EPOLL) */ |
| 2155 | #if defined(GRPC_POSIX_SOCKET) |
Sree Kuchibhotla | 41622a8 | 2016-06-13 16:43:14 -0700 | [diff] [blame] | 2156 | #include "src/core/lib/iomgr/ev_posix.h" |
murgatroid99 | 623dd4f | 2016-08-08 17:31:27 -0700 | [diff] [blame] | 2157 | /* If GRPC_LINUX_EPOLL is not defined, it means epoll is not available. Return |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 2158 | * NULL */ |
| 2159 | const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { return NULL; } |
murgatroid99 | 623dd4f | 2016-08-08 17:31:27 -0700 | [diff] [blame] | 2160 | #endif /* defined(GRPC_POSIX_SOCKET) */ |
Sree Kuchibhotla | 5855c47 | 2016-06-08 12:56:56 -0700 | [diff] [blame] | 2161 | |
Sree Kuchibhotla | c7be7c6 | 2016-06-09 17:08:50 -0700 | [diff] [blame] | 2162 | void grpc_use_signal(int signum) {} |
murgatroid99 | 623dd4f | 2016-08-08 17:31:27 -0700 | [diff] [blame] | 2163 | #endif /* !defined(GRPC_LINUX_EPOLL) */ |