drivers/xen/blkback/xenbus.c

   1 /*  Xenbus code for blkif backend
   2     Copyright (C) 2005 Rusty Russell <rusty@rustcorp.com.au>
   3     Copyright (C) 2005 XenSource Ltd
   4
   5     This program is free software; you can redistribute it and/or modify
   6     it under the terms of the GNU General Public License as published by
   7     the Free Software Foundation; either version 2 of the License, or
   8     (at your option) any later version.
   9
  10     This program is distributed in the hope that it will be useful,
  11     but WITHOUT ANY WARRANTY; without even the implied warranty of
  12     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13     GNU General Public License for more details.
  14
  15     You should have received a copy of the GNU General Public License
  16     along with this program; if not, write to the Free Software
  17     Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18 */
  19
  20 #include <stdarg.h>
  21 #include <linux/kthread.h>
  22 #include "common.h"
  23 #include "../core/domctl.h"
  24
  25 #undef DPRINTK
  26 #define DPRINTK(fmt, args...)                           \
  27         pr_debug("blkback/xenbus (%s:%d) " fmt ".\n",   \
  28                  __FUNCTION__, __LINE__, ##args)
  29
  30 static void connect(struct backend_info *);
  31 static int connect_ring(struct backend_info *);
  32 static void backend_changed(struct xenbus_watch *, const char **,
  33                             unsigned int);
  34
  35 static int blkback_name(blkif_t *blkif, char *buf)
  36 {
  37         char *devpath, *devname;
  38         struct xenbus_device *dev = blkif->be->dev;
  39
  40         devpath = xenbus_read(XBT_NIL, dev->nodename, "dev", NULL);
  41         if (IS_ERR(devpath))
  42                 return PTR_ERR(devpath);
  43
  44         if ((devname = strstr(devpath, "/dev/")) != NULL)
  45                 devname += strlen("/dev/");
  46         else
  47                 devname  = devpath;
  48
  49         snprintf(buf, TASK_COMM_LEN, "blkback.%d.%s", blkif->domid, devname);
  50         kfree(devpath);
  51
  52         return 0;
  53 }
  54
  55 static void update_blkif_status(blkif_t *blkif)
  56 {
  57         int err;
  58         char name[TASK_COMM_LEN];
  59
  60         /* Not ready to connect? */
  61         if (!blkif->irq)
  62                 return;
  63
  64         /* Already connected? */
  65         if (blkif->be->dev->state == XenbusStateConnected)
  66                 return;
  67
  68         /* Attempt to connect: exit if we fail to. */
  69         connect(blkif->be);
  70         if (blkif->be->dev->state != XenbusStateConnected)
  71                 return;
  72
  73         err = blkback_name(blkif, name);
  74         if (err) {
  75                 xenbus_dev_error(blkif->be->dev, err, "get blkback dev name");
  76                 return;
  77         }
  78
  79         if (blkif->vbd.bdev) {
  80                 struct address_space *mapping
  81                         = blkif->vbd.bdev->bd_inode->i_mapping;
  82
  83                 err = filemap_write_and_wait(mapping);
  84                 if (err) {
  85                         xenbus_dev_error(blkif->be->dev, err, "block flush");
  86                         return;
  87                 }
  88                 invalidate_inode_pages2(mapping);
  89         }
  90
  91         blkif->xenblkd = kthread_run(blkif_schedule, blkif, name);
  92         if (IS_ERR(blkif->xenblkd)) {
  93                 err = PTR_ERR(blkif->xenblkd);
  94                 blkif->xenblkd = NULL;
  95                 xenbus_dev_error(blkif->be->dev, err, "start xenblkd");
  96         }
  97 }
  98
  99
 100 /****************************************************************
 101  *  sysfs interface for VBD I/O requests
 102  */
 103
 104 #define VBD_SHOW(name, format, args...)                                 \
 105         static ssize_t show_##name(struct device *_dev,                 \
 106                                    struct device_attribute *attr,       \
 107                                    char *buf)                           \
 108         {                                                               \
 109                 ssize_t ret = -ENODEV;                                  \
 110                 struct xenbus_device *dev;                              \
 111                 struct backend_info *be;                                \
 112                                                                         \
 113                 if (!get_device(_dev))                                  \
 114                         return ret;                                     \
 115                 dev = to_xenbus_device(_dev);                           \
 116                 if ((be = dev_get_drvdata(&dev->dev)) != NULL)          \
 117                         ret = sprintf(buf, format, ##args);             \
 118                 put_device(_dev);                                       \
 119                 return ret;                                             \
 120         }                                                               \
 121         static DEVICE_ATTR(name, S_IRUGO, show_##name, NULL)
 122
 123 VBD_SHOW(oo_req,  "%d\n", be->blkif->st_oo_req);
 124 VBD_SHOW(rd_req,  "%d\n", be->blkif->st_rd_req);
 125 VBD_SHOW(wr_req,  "%d\n", be->blkif->st_wr_req);
 126 VBD_SHOW(br_req,  "%d\n", be->blkif->st_br_req);
 127 VBD_SHOW(fl_req,  "%d\n", be->blkif->st_fl_req);
 128 VBD_SHOW(ds_req,  "%d\n", be->blkif->st_ds_req);
 129 VBD_SHOW(rd_sect, "%d\n", be->blkif->st_rd_sect);
 130 VBD_SHOW(wr_sect, "%d\n", be->blkif->st_wr_sect);
 131
 132 static struct attribute *vbdstat_attrs[] = {
 133         &dev_attr_oo_req.attr,
 134         &dev_attr_rd_req.attr,
 135         &dev_attr_wr_req.attr,
 136         &dev_attr_br_req.attr,
 137         &dev_attr_fl_req.attr,
 138         &dev_attr_ds_req.attr,
 139         &dev_attr_rd_sect.attr,
 140         &dev_attr_wr_sect.attr,
 141         NULL
 142 };
 143
 144 static const struct attribute_group vbdstat_group = {
 145         .name = "statistics",
 146         .attrs = vbdstat_attrs,
 147 };
 148
 149 VBD_SHOW(physical_device, "%x:%x\n", be->major, be->minor);
 150 VBD_SHOW(mode, "%s\n", be->mode);
 151
 152 int xenvbd_sysfs_addif(struct xenbus_device *dev)
 153 {
 154         int error;
 155
 156         error = device_create_file(&dev->dev, &dev_attr_physical_device);
 157         if (error)
 158                 goto fail1;
 159
 160         error = device_create_file(&dev->dev, &dev_attr_mode);
 161         if (error)
 162                 goto fail2;
 163
 164         error = sysfs_create_group(&dev->dev.kobj, &vbdstat_group);
 165         if (error)
 166                 goto fail3;
 167
 168         return 0;
 169
 170 fail3:  sysfs_remove_group(&dev->dev.kobj, &vbdstat_group);
 171 fail2:  device_remove_file(&dev->dev, &dev_attr_mode);
 172 fail1:  device_remove_file(&dev->dev, &dev_attr_physical_device);
 173         return error;
 174 }
 175
 176 void xenvbd_sysfs_delif(struct xenbus_device *dev)
 177 {
 178         sysfs_remove_group(&dev->dev.kobj, &vbdstat_group);
 179         device_remove_file(&dev->dev, &dev_attr_mode);
 180         device_remove_file(&dev->dev, &dev_attr_physical_device);
 181 }
 182
 183 static int blkback_remove(struct xenbus_device *dev)
 184 {
 185         struct backend_info *be = dev_get_drvdata(&dev->dev);
 186
 187         DPRINTK("");
 188
 189         if (be->major || be->minor)
 190                 xenvbd_sysfs_delif(dev);
 191
 192         if (be->backend_watch.node) {
 193                 unregister_xenbus_watch(&be->backend_watch);
 194                 kfree(be->backend_watch.node);
 195                 be->backend_watch.node = NULL;
 196         }
 197
 198         if (be->cdrom_watch.node) {
 199                 unregister_xenbus_watch(&be->cdrom_watch);
 200                 kfree(be->cdrom_watch.node);
 201                 be->cdrom_watch.node = NULL;
 202         }
 203
 204         if (be->blkif) {
 205                 blkif_disconnect(be->blkif);
 206                 vbd_free(&be->blkif->vbd);
 207                 blkif_free(be->blkif);
 208                 be->blkif = NULL;
 209         }
 210
 211         kfree(be);
 212         dev_set_drvdata(&dev->dev, NULL);
 213         return 0;
 214 }
 215
 216 int blkback_barrier(struct xenbus_transaction xbt,
 217                     struct backend_info *be, int state)
 218 {
 219         struct xenbus_device *dev = be->dev;
 220         int err;
 221
 222         err = xenbus_printf(xbt, dev->nodename, "feature-barrier",
 223                             "%d", state);
 224         if (err)
 225                 xenbus_dev_fatal(dev, err, "writing feature-barrier");
 226
 227         return err;
 228 }
 229
 230 int blkback_flush_diskcache(struct xenbus_transaction xbt,
 231                             struct backend_info *be, int state)
 232 {
 233         struct xenbus_device *dev = be->dev;
 234         int err;
 235
 236         err = xenbus_printf(xbt, dev->nodename, "feature-flush-cache",
 237                             "%d", state);
 238         if (err)
 239                 xenbus_dev_fatal(dev, err, "writing feature-flush-cache");
 240
 241         return err;
 242 }
 243
 244 static int xen_blkbk_discard(struct xenbus_transaction xbt,
 245                              struct backend_info *be)
 246 {
 247         struct xenbus_device *dev = be->dev;
 248         blkif_t *blkif = be->blkif;
 249         char *type;
 250         int err;
 251         int state = 0;
 252
 253         type = xenbus_read(XBT_NIL, dev->nodename, "type", NULL);
 254         if (!IS_ERR(type)) {
 255                 if (strncmp(type, "file", 4) == 0) {
 256                         state = 1;
 257                         blkif->blk_backend_type = BLKIF_BACKEND_FILE;
 258                 }
 259                 if (strncmp(type, "phy", 3) == 0) {
 260                         struct request_queue *q;
 261
 262                         q = bdev_get_queue(blkif->vbd.bdev);
 263                         if (blk_queue_discard(q)) {
 264                                 err = xenbus_printf(xbt, dev->nodename,
 265                                         "discard-granularity", "%u",
 266                                         q->limits.discard_granularity);
 267                                 if (err) {
 268                                         xenbus_dev_fatal(dev, err,
 269                                                 "writing discard-granularity");
 270                                         goto kfree;
 271                                 }
 272                                 err = xenbus_printf(xbt, dev->nodename,
 273                                         "discard-alignment", "%u",
 274                                         q->limits.discard_alignment);
 275                                 if (err) {
 276                                         xenbus_dev_fatal(dev, err,
 277                                                 "writing discard-alignment");
 278                                         goto kfree;
 279                                 }
 280                                 state = 1;
 281                                 blkif->blk_backend_type = BLKIF_BACKEND_PHY;
 282                         }
 283                         /* Optional. */
 284                         err = xenbus_printf(xbt, dev->nodename,
 285                                             "discard-secure", "%d",
 286                                             blkif->vbd.discard_secure);
 287                         if (err) {
 288                                 xenbus_dev_fatal(dev, err,
 289                                                  "writing discard-secure");
 290                                 goto kfree;
 291                         }
 292                 }
 293         } else {
 294                 err = PTR_ERR(type);
 295                 xenbus_dev_fatal(dev, err, "reading type");
 296                 goto out;
 297         }
 298
 299         err = xenbus_printf(xbt, dev->nodename, "feature-discard",
 300                             "%d", state);
 301         if (err)
 302                 xenbus_dev_fatal(dev, err, "writing feature-discard");
 303 kfree:
 304         kfree(type);
 305 out:
 306         return err;
 307 }
 308
 309 /**
 310  * Entry point to this code when a new device is created.  Allocate the basic
 311  * structures, and watch the store waiting for the hotplug scripts to tell us
 312  * the device's physical major and minor numbers.  Switch to InitWait.
 313  */
 314 static int blkback_probe(struct xenbus_device *dev,
 315                          const struct xenbus_device_id *id)
 316 {
 317         int err;
 318         struct backend_info *be = kzalloc(sizeof(struct backend_info),
 319                                           GFP_KERNEL);
 320         if (!be) {
 321                 xenbus_dev_fatal(dev, -ENOMEM,
 322                                  "allocating backend structure");
 323                 return -ENOMEM;
 324         }
 325         be->dev = dev;
 326         dev_set_drvdata(&dev->dev, be);
 327
 328         be->blkif = blkif_alloc(dev->otherend_id);
 329         if (IS_ERR(be->blkif)) {
 330                 err = PTR_ERR(be->blkif);
 331                 be->blkif = NULL;
 332                 xenbus_dev_fatal(dev, err, "creating block interface");
 333                 goto fail;
 334         }
 335
 336         /* setup back pointer */
 337         be->blkif->be = be;
 338
 339         err = xenbus_watch_path2(dev, dev->nodename, "physical-device",
 340                                  &be->backend_watch, backend_changed);
 341         if (err)
 342                 goto fail;
 343
 344         err = xenbus_switch_state(dev, XenbusStateInitWait);
 345         if (err)
 346                 goto fail;
 347
 348         return 0;
 349
 350 fail:
 351         DPRINTK("failed");
 352         blkback_remove(dev);
 353         return err;
 354 }
 355
 356
 357 /**
 358  * Callback received when the hotplug scripts have placed the physical-device
 359  * node.  Read it and the mode node, and create a vbd.  If the frontend is
 360  * ready, connect.
 361  */
 362 static void backend_changed(struct xenbus_watch *watch,
 363                             const char **vec, unsigned int len)
 364 {
 365         int err;
 366         unsigned major;
 367         unsigned minor;
 368         struct backend_info *be
 369                 = container_of(watch, struct backend_info, backend_watch);
 370         struct xenbus_device *dev = be->dev;
 371         int cdrom = 0;
 372         char *device_type;
 373
 374         DPRINTK("");
 375
 376         err = xenbus_scanf(XBT_NIL, dev->nodename, "physical-device", "%x:%x",
 377                            &major, &minor);
 378         if (XENBUS_EXIST_ERR(err)) {
 379                 /* Since this watch will fire once immediately after it is
 380                    registered, we expect this.  Ignore it, and wait for the
 381                    hotplug scripts. */
 382                 return;
 383         }
 384         if (err != 2) {
 385                 xenbus_dev_fatal(dev, err, "reading physical-device");
 386                 return;
 387         }
 388
 389         if ((be->major || be->minor) &&
 390             ((be->major != major) || (be->minor != minor))) {
 391                 pr_warning("blkback: changing physical device (from %x:%x to"
 392                            " %x:%x) not supported\n", be->major, be->minor,
 393                            major, minor);
 394                 return;
 395         }
 396
 397         be->mode = xenbus_read(XBT_NIL, dev->nodename, "mode", NULL);
 398         if (IS_ERR(be->mode)) {
 399                 err = PTR_ERR(be->mode);
 400                 be->mode = NULL;
 401                 xenbus_dev_fatal(dev, err, "reading mode");
 402                 return;
 403         }
 404
 405         device_type = xenbus_read(XBT_NIL, dev->otherend, "device-type", NULL);
 406         if (!IS_ERR(device_type)) {
 407                 cdrom = strcmp(device_type, "cdrom") == 0;
 408                 kfree(device_type);
 409         }
 410
 411         if (be->major == 0 && be->minor == 0) {
 412                 /* Front end dir is a number, which is used as the handle. */
 413
 414                 char *p = strrchr(dev->otherend, '/') + 1;
 415                 long handle = simple_strtoul(p, NULL, 0);
 416
 417                 be->major = major;
 418                 be->minor = minor;
 419
 420                 err = vbd_create(be->blkif, handle, major, minor,
 421                                  FMODE_READ
 422                                  | (strchr(be->mode, 'w') ? FMODE_WRITE : 0)
 423                                  | (strchr(be->mode, '!') ? 0 : FMODE_EXCL),
 424                                  cdrom);
 425                 switch (err) {
 426                 case -ENOMEDIUM:
 427                         if (be->blkif->vbd.type
 428                             & (VDISK_CDROM | VDISK_REMOVABLE))
 429                 case 0:
 430                                 break;
 431                 default:
 432                         be->major = be->minor = 0;
 433                         xenbus_dev_fatal(dev, err, "creating vbd structure");
 434                         return;
 435                 }
 436
 437                 err = xenvbd_sysfs_addif(dev);
 438                 if (err) {
 439                         vbd_free(&be->blkif->vbd);
 440                         be->major = be->minor = 0;
 441                         xenbus_dev_fatal(dev, err, "creating sysfs entries");
 442                         return;
 443                 }
 444
 445                 /* We're potentially connected now */
 446                 update_blkif_status(be->blkif);
 447
 448                 /* Add watch for cdrom media status if necessay */
 449                 cdrom_add_media_watch(be);
 450         }
 451 }
 452
 453
 454 /**
 455  * Callback received when the frontend's state changes.
 456  */
 457 static void frontend_changed(struct xenbus_device *dev,
 458                              enum xenbus_state frontend_state)
 459 {
 460         struct backend_info *be = dev_get_drvdata(&dev->dev);
 461         int err;
 462
 463         DPRINTK("%s", xenbus_strstate(frontend_state));
 464
 465         switch (frontend_state) {
 466         case XenbusStateInitialising:
 467                 if (dev->state == XenbusStateClosed) {
 468                         pr_info("%s: %s: prepare for reconnect\n",
 469                                 __FUNCTION__, dev->nodename);
 470                         xenbus_switch_state(dev, XenbusStateInitWait);
 471                 }
 472                 break;
 473
 474         case XenbusStateInitialised:
 475         case XenbusStateConnected:
 476                 /* Ensure we connect even when two watches fire in
 477                    close successsion and we miss the intermediate value
 478                    of frontend_state. */
 479                 if (dev->state == XenbusStateConnected)
 480                         break;
 481
 482                 /* Enforce precondition before potential leak point.
 483                  * blkif_disconnect() is idempotent.
 484                  */
 485                 blkif_disconnect(be->blkif);
 486
 487                 err = connect_ring(be);
 488                 if (err)
 489                         break;
 490                 if (be->blkif->vbd.bdev)
 491                         update_blkif_status(be->blkif);
 492                 break;
 493
 494         case XenbusStateClosing:
 495                 blkif_disconnect(be->blkif);
 496                 xenbus_switch_state(dev, XenbusStateClosing);
 497                 break;
 498
 499         case XenbusStateClosed:
 500                 xenbus_switch_state(dev, XenbusStateClosed);
 501                 if (xenbus_dev_is_online(dev))
 502                         break;
 503                 /* fall through if not online */
 504         case XenbusStateUnknown:
 505                 /* implies blkif_disconnect() via blkback_remove() */
 506                 device_unregister(&dev->dev);
 507                 break;
 508
 509         default:
 510                 xenbus_dev_fatal(dev, -EINVAL, "saw state %d at frontend",
 511                                  frontend_state);
 512                 break;
 513         }
 514 }
 515
 516
 517 /* ** Connection ** */
 518
 519
 520 /**
 521  * Write the physical details regarding the block device to the store, and
 522  * switch to Connected state.
 523  */
 524 static void connect(struct backend_info *be)
 525 {
 526         struct xenbus_transaction xbt;
 527         int err;
 528         struct xenbus_device *dev = be->dev;
 529
 530         DPRINTK("%s", dev->otherend);
 531
 532         /* Supply the information about the device the frontend needs */
 533 again:
 534         err = xenbus_transaction_start(&xbt);
 535         if (err) {
 536                 xenbus_dev_fatal(dev, err, "starting transaction");
 537                 return;
 538         }
 539
 540         err = blkback_flush_diskcache(xbt, be, be->blkif->vbd.flush_support);
 541         if (err)
 542                 goto abort;
 543
 544         err = blkback_barrier(xbt, be, be->blkif->vbd.flush_support);
 545         if (err)
 546                 goto abort;
 547
 548         err = xen_blkbk_discard(xbt, be);
 549
 550         err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
 551                             vbd_size(&be->blkif->vbd));
 552         if (err) {
 553                 xenbus_dev_fatal(dev, err, "writing %s/sectors",
 554                                  dev->nodename);
 555                 goto abort;
 556         }
 557
 558         /* FIXME: use a typename instead */
 559         err = xenbus_printf(xbt, dev->nodename, "info", "%u",
 560                             be->blkif->vbd.type);
 561         if (err) {
 562                 xenbus_dev_fatal(dev, err, "writing %s/info",
 563                                  dev->nodename);
 564                 goto abort;
 565         }
 566         err = xenbus_printf(xbt, dev->nodename, "sector-size", "%lu",
 567                             vbd_secsize(&be->blkif->vbd));
 568         if (err) {
 569                 xenbus_dev_fatal(dev, err, "writing %s/sector-size",
 570                                  dev->nodename);
 571                 goto abort;
 572         }
 573
 574         err = xenbus_transaction_end(xbt, 0);
 575         if (err == -EAGAIN)
 576                 goto again;
 577         if (err)
 578                 xenbus_dev_fatal(dev, err, "ending transaction");
 579
 580         err = xenbus_switch_state(dev, XenbusStateConnected);
 581         if (err)
 582                 xenbus_dev_fatal(dev, err, "%s: switching to Connected state",
 583                                  dev->nodename);
 584
 585         return;
 586  abort:
 587         xenbus_transaction_end(xbt, 1);
 588 }
 589
 590
 591 static int connect_ring(struct backend_info *be)
 592 {
 593         struct xenbus_device *dev = be->dev;
 594         unsigned int ring_ref, evtchn;
 595         char *protocol;
 596         int err;
 597
 598         DPRINTK("%s", dev->otherend);
 599
 600         err = xenbus_gather(XBT_NIL, dev->otherend, "ring-ref", "%u", &ring_ref,
 601                             "event-channel", "%u", &evtchn, NULL);
 602         if (err) {
 603                 xenbus_dev_fatal(dev, err,
 604                                  "reading %s/ring-ref and event-channel",
 605                                  dev->otherend);
 606                 return err;
 607         }
 608
 609         be->blkif->blk_protocol = BLKIF_PROTOCOL_NATIVE;
 610         err = xenbus_gather(XBT_NIL, dev->otherend, "protocol",
 611                             NULL, &protocol, NULL);
 612         if (err) {
 613                 protocol = NULL;
 614                 be->blkif->blk_protocol = xen_guest_blkif_protocol(be->blkif->domid);
 615         }
 616         else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_NATIVE))
 617                 be->blkif->blk_protocol = BLKIF_PROTOCOL_NATIVE;
 618         else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_32))
 619                 be->blkif->blk_protocol = BLKIF_PROTOCOL_X86_32;
 620         else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_64))
 621                 be->blkif->blk_protocol = BLKIF_PROTOCOL_X86_64;
 622         else {
 623                 xenbus_dev_fatal(dev, err, "unknown fe protocol %s", protocol);
 624                 kfree(protocol);
 625                 return -1;
 626         }
 627         pr_info("blkback: ring-ref %u, event-channel %u, protocol %d (%s)\n",
 628                 ring_ref, evtchn, be->blkif->blk_protocol,
 629                 protocol ?: "unspecified");
 630         kfree(protocol);
 631
 632         /* Map the shared frame, irq etc. */
 633         err = blkif_map(be->blkif, ring_ref, evtchn);
 634         if (err) {
 635                 xenbus_dev_fatal(dev, err, "mapping ring-ref %u port %u",
 636                                  ring_ref, evtchn);
 637                 return err;
 638         }
 639
 640         return 0;
 641 }
 642
 643
 644 /* ** Driver Registration ** */
 645
 646
 647 static const struct xenbus_device_id blkback_ids[] = {
 648         { "vbd" },
 649         { "" }
 650 };
 651
 652 static DEFINE_XENBUS_DRIVER(blkback, ,
 653         .probe = blkback_probe,
 654         .remove = blkback_remove,
 655         .otherend_changed = frontend_changed
 656 );
 657
 658
 659 void blkif_xenbus_init(void)
 660 {
 661         WARN_ON(xenbus_register_backend(&blkback_driver));
 662 }