source3/modules/vfs_aio_linux.c

   1 /*
   2  * Simulate Posix AIO using Linux kernel AIO.
   3  *
   4  * Copyright (C) Jeremy Allison 2012
   5  * Copyright (C) Volker Lendecke 2012
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 3 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include "includes.h"
  23 #include "system/filesys.h"
  24 #include "smbd/smbd.h"
  25 #include "smbd/globals.h"
  26 #include "lib/util/tevent_unix.h"
  27 #include <sys/eventfd.h>
  28 #include <libaio.h>
  29
  30 static int event_fd = -1;
  31 static io_context_t io_ctx;
  32 static struct tevent_fd *aio_read_event;
  33 static bool used;
  34 static unsigned num_busy;
  35
  36 static void aio_linux_done(struct tevent_context *event_ctx,
  37                            struct tevent_fd *event,
  38                            uint16 flags, void *private_data);
  39
  40 /************************************************************************
  41  Housekeeping. Cleanup if no activity for 30 seconds.
  42 ***********************************************************************/
  43
  44 static void aio_linux_housekeeping(struct tevent_context *event_ctx,
  45                                         struct tevent_timer *te,
  46                                         struct timeval now,
  47                                         void *private_data)
  48 {
  49         /* Remove this timed event handler. */
  50         TALLOC_FREE(te);
  51
  52         if ((num_busy != 0) || used) {
  53                 used = false;
  54
  55                 /* Still busy. Look again in 30 seconds. */
  56                 (void)tevent_add_timer(event_ctx,
  57                                         NULL,
  58                                         timeval_current_ofs(30, 0),
  59                                         aio_linux_housekeeping,
  60                                         NULL);
  61                 return;
  62         }
  63
  64         /* No activity for 30 seconds. Close out kernel resources. */
  65         io_queue_release(io_ctx);
  66         memset(&io_ctx, '\0', sizeof(io_ctx));
  67
  68         if (event_fd != -1) {
  69                 close(event_fd);
  70                 event_fd = -1;
  71         }
  72
  73         TALLOC_FREE(aio_read_event);
  74 }
  75
  76 /************************************************************************
  77  Ensure event fd and aio context are initialized.
  78 ***********************************************************************/
  79
  80 static bool init_aio_linux(struct vfs_handle_struct *handle)
  81 {
  82         struct tevent_timer *te = NULL;
  83
  84         if (event_fd != -1) {
  85                 /* Already initialized. */
  86                 return true;
  87         }
  88
  89         /* Schedule a shutdown event for 30 seconds from now. */
  90         te = tevent_add_timer(handle->conn->sconn->ev_ctx,
  91                                 NULL,
  92                                 timeval_current_ofs(30, 0),
  93                                 aio_linux_housekeeping,
  94                                 NULL);
  95
  96         if (te == NULL) {
  97                 goto fail;
  98         }
  99
 100         event_fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
 101         if (event_fd == -1) {
 102                 goto fail;
 103         }
 104
 105         aio_read_event = tevent_add_fd(server_event_context(),
 106                                 NULL,
 107                                 event_fd,
 108                                 TEVENT_FD_READ,
 109                                 aio_linux_done,
 110                                 NULL);
 111         if (aio_read_event == NULL) {
 112                 goto fail;
 113         }
 114
 115         if (io_queue_init(aio_pending_size, &io_ctx)) {
 116                 goto fail;
 117         }
 118
 119         DEBUG(10,("init_aio_linux: initialized with up to %d events\n",
 120                   aio_pending_size));
 121
 122         return true;
 123
 124   fail:
 125
 126         DEBUG(10,("init_aio_linux: initialization failed\n"));
 127
 128         TALLOC_FREE(te);
 129         TALLOC_FREE(aio_read_event);
 130         if (event_fd != -1) {
 131                 close(event_fd);
 132                 event_fd = -1;
 133         }
 134         memset(&io_ctx, '\0', sizeof(io_ctx));
 135         return false;
 136 }
 137
 138 struct aio_linux_state {
 139         struct iocb event_iocb;
 140         ssize_t ret;
 141         int err;
 142 };
 143
 144 static struct tevent_req *aio_linux_pread_send(
 145         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 146         struct tevent_context *ev, struct files_struct *fsp,
 147         void *data, size_t n, off_t offset)
 148 {
 149         struct tevent_req *req;
 150         struct aio_linux_state *state;
 151         struct iocb *piocb;
 152         int ret;
 153
 154         req = tevent_req_create(mem_ctx, &state, struct aio_linux_state);
 155         if (req == NULL) {
 156                 return NULL;
 157         }
 158         if (!init_aio_linux(handle)) {
 159                 tevent_req_error(req, EIO);
 160                 return tevent_req_post(req, ev);
 161         }
 162
 163         io_prep_pread(&state->event_iocb, fsp->fh->fd, data, n, offset);
 164         io_set_eventfd(&state->event_iocb, event_fd);
 165         state->event_iocb.data = req;
 166
 167         piocb = &state->event_iocb;
 168
 169         ret = io_submit(io_ctx, 1, &piocb);
 170         if (ret < 0) {
 171                 tevent_req_error(req, -ret);
 172                 return tevent_req_post(req, ev);
 173         }
 174         num_busy += 1;
 175         used = true;
 176         return req;
 177 }
 178
 179 static struct tevent_req *aio_linux_pwrite_send(
 180         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 181         struct tevent_context *ev, struct files_struct *fsp,
 182         const void *data, size_t n, off_t offset)
 183 {
 184         struct tevent_req *req;
 185         struct aio_linux_state *state;
 186         struct iocb *piocb;
 187         int ret;
 188
 189         req = tevent_req_create(mem_ctx, &state, struct aio_linux_state);
 190         if (req == NULL) {
 191                 return NULL;
 192         }
 193         if (!init_aio_linux(handle)) {
 194                 tevent_req_error(req, EIO);
 195                 return tevent_req_post(req, ev);
 196         }
 197
 198         io_prep_pwrite(&state->event_iocb, fsp->fh->fd, discard_const(data),
 199                        n, offset);
 200         io_set_eventfd(&state->event_iocb, event_fd);
 201         state->event_iocb.data = req;
 202
 203         piocb = &state->event_iocb;
 204
 205         ret = io_submit(io_ctx, 1, &piocb);
 206         if (ret < 0) {
 207                 tevent_req_error(req, -ret);
 208                 return tevent_req_post(req, ev);
 209         }
 210         num_busy += 1;
 211         used = true;
 212         return req;
 213 }
 214
 215 static struct tevent_req *aio_linux_fsync_send(
 216         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 217         struct tevent_context *ev, struct files_struct *fsp)
 218 {
 219         struct tevent_req *req;
 220         struct aio_linux_state *state;
 221         struct iocb *piocb;
 222         int ret;
 223
 224         req = tevent_req_create(mem_ctx, &state, struct aio_linux_state);
 225         if (req == NULL) {
 226                 return NULL;
 227         }
 228         if (!init_aio_linux(handle)) {
 229                 tevent_req_error(req, EIO);
 230                 return tevent_req_post(req, ev);
 231         }
 232
 233         io_prep_fsync(&state->event_iocb, fsp->fh->fd);
 234         io_set_eventfd(&state->event_iocb, event_fd);
 235         state->event_iocb.data = req;
 236
 237         piocb = &state->event_iocb;
 238
 239         ret = io_submit(io_ctx, 1, &piocb);
 240         if (ret < 0) {
 241                 tevent_req_error(req, -ret);
 242                 return tevent_req_post(req, ev);
 243         }
 244         num_busy += 1;
 245         used = true;
 246         return req;
 247 }
 248
 249 static void aio_linux_done(struct tevent_context *event_ctx,
 250                            struct tevent_fd *event,
 251                            uint16 flags, void *private_data)
 252 {
 253         uint64_t num_events = 0;
 254
 255         DEBUG(10, ("aio_linux_done called with flags=%d\n",
 256                    (int)flags));
 257
 258         /* Read the number of events available. */
 259         if (sys_read(event_fd, &num_events, sizeof(num_events)) !=
 260                         sizeof(num_events)) {
 261                 smb_panic("aio_linux_handle_completion: invalid read");
 262         }
 263
 264         while (num_events > 0) {
 265                 struct timespec ts = { 0, };
 266                 struct io_event finished;
 267                 struct tevent_req *req;
 268                 struct aio_linux_state *state;
 269                 int ret;
 270
 271                 ret = io_getevents(io_ctx, 1, 1, &finished, &ts);
 272                 if (ret < 0) {
 273                         DEBUG(1, ("aio_linux_done: io_getevents returned %s\n",
 274                                   strerror(-ret)));
 275                         return;
 276                 }
 277                 if (ret == 0) {
 278                         DEBUG(10, ("aio_linux_done: io_getvents returned "
 279                                    "0\n"));
 280                         continue;
 281                 }
 282
 283                 num_busy -= 1;
 284
 285                 req = talloc_get_type_abort(finished.data,
 286                                             struct tevent_req);
 287                 state = tevent_req_data(req, struct aio_linux_state);
 288
 289                 if (finished.res < 0) {
 290                         state->ret = -1;
 291                         state->err = -finished.res;
 292                 } else {
 293                         state->ret = finished.res;
 294                         state->err = 0;
 295                 }
 296                 tevent_req_done(req);
 297                 num_events -= 1;
 298         }
 299 }
 300
 301 static ssize_t aio_linux_recv(struct tevent_req *req, int *err)
 302 {
 303         struct aio_linux_state *state = tevent_req_data(
 304                 req, struct aio_linux_state);
 305
 306         if (tevent_req_is_unix_error(req, err)) {
 307                 return -1;
 308         }
 309         if (state->ret == -1) {
 310                 *err = state->err;
 311         }
 312         return state->ret;
 313 }
 314
 315 static int aio_linux_int_recv(struct tevent_req *req, int *err)
 316 {
 317         /*
 318          * Use implicit conversion ssize_t->int
 319          */
 320         return aio_linux_recv(req, err);
 321 }
 322
 323 static int aio_linux_connect(vfs_handle_struct *handle, const char *service,
 324                                const char *user)
 325 {
 326         /*********************************************************************
 327          * How many io_events to initialize ?
 328          * 128 per process seems insane as a default until you realize that
 329          * (a) Throttling is done in SMB2 via the crediting algorithm.
 330          * (b) SMB1 clients are limited to max_mux (50) outstanding
 331          *     requests and Windows clients don't use this anyway.
 332          * Essentially we want this to be unlimited unless smb.conf
 333          * says different.
 334          *********************************************************************/
 335         aio_pending_size = lp_parm_int(
 336                 SNUM(handle->conn), "aio_linux", "aio num events", 128);
 337         return SMB_VFS_NEXT_CONNECT(handle, service, user);
 338 }
 339
 340 static struct vfs_fn_pointers vfs_aio_linux_fns = {
 341         .connect_fn = aio_linux_connect,
 342         .pread_send_fn = aio_linux_pread_send,
 343         .pread_recv_fn = aio_linux_recv,
 344         .pwrite_send_fn = aio_linux_pwrite_send,
 345         .pwrite_recv_fn = aio_linux_recv,
 346         .fsync_send_fn = aio_linux_fsync_send,
 347         .fsync_recv_fn = aio_linux_int_recv,
 348 };
 349
 350 NTSTATUS vfs_aio_linux_init(void)
 351 {
 352         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 353                                 "aio_linux", &vfs_aio_linux_fns);
 354 }