drivers/block/xen-blkback/blkback.c

   1 /******************************************************************************
   2  *
   3  * Back-end of the driver for virtual block devices. This portion of the
   4  * driver exports a 'unified' block-device interface that can be accessed
   5  * by any operating system that implements a compatible front end. A
   6  * reference front-end implementation can be found in:
   7  *  drivers/block/xen-blkfront.c
   8  *
   9  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
  10  * Copyright (c) 2005, Christopher Clark
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License version 2
  14  * as published by the Free Software Foundation; or, when distributed
  15  * separately from the Linux kernel or incorporated into other
  16  * software packages, subject to the following license:
  17  *
  18  * Permission is hereby granted, free of charge, to any person obtaining a copy
  19  * of this source file (the "Software"), to deal in the Software without
  20  * restriction, including without limitation the rights to use, copy, modify,
  21  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  22  * and to permit persons to whom the Software is furnished to do so, subject to
  23  * the following conditions:
  24  *
  25  * The above copyright notice and this permission notice shall be included in
  26  * all copies or substantial portions of the Software.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  29  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  30  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  31  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  32  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  33  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  34  * IN THE SOFTWARE.
  35  */
  36
  37 #include <linux/spinlock.h>
  38 #include <linux/kthread.h>
  39 #include <linux/list.h>
  40 #include <linux/delay.h>
  41 #include <linux/freezer.h>
  42 #include <linux/bitmap.h>
  43
  44 #include <xen/events.h>
  45 #include <xen/page.h>
  46 #include <xen/xen.h>
  47 #include <asm/xen/hypervisor.h>
  48 #include <asm/xen/hypercall.h>
  49 #include <xen/balloon.h>
  50 #include "common.h"
  51
  52 /*
  53  * Maximum number of unused free pages to keep in the internal buffer.
  54  * Setting this to a value too low will reduce memory used in each backend,
  55  * but can have a performance penalty.
  56  *
  57  * A sane value is xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST, but can
  58  * be set to a lower value that might degrade performance on some intensive
  59  * IO workloads.
  60  */
  61
  62 static int xen_blkif_max_buffer_pages = 1024;
  63 module_param_named(max_buffer_pages, xen_blkif_max_buffer_pages, int, 0644);
  64 MODULE_PARM_DESC(max_buffer_pages,
  65 "Maximum number of free pages to keep in each block backend buffer");
  66
  67 /*
  68  * Maximum number of grants to map persistently in blkback. For maximum
  69  * performance this should be the total numbers of grants that can be used
  70  * to fill the ring, but since this might become too high, specially with
  71  * the use of indirect descriptors, we set it to a value that provides good
  72  * performance without using too much memory.
  73  *
  74  * When the list of persistent grants is full we clean it up using a LRU
  75  * algorithm.
  76  */
  77
  78 static int xen_blkif_max_pgrants = 1056;
  79 module_param_named(max_persistent_grants, xen_blkif_max_pgrants, int, 0644);
  80 MODULE_PARM_DESC(max_persistent_grants,
  81                  "Maximum number of grants to map persistently");
  82
  83 /*
  84  * The LRU mechanism to clean the lists of persistent grants needs to
  85  * be executed periodically. The time interval between consecutive executions
  86  * of the purge mechanism is set in ms.
  87  */
  88 #define LRU_INTERVAL 100
  89
  90 /*
  91  * When the persistent grants list is full we will remove unused grants
  92  * from the list. The percent number of grants to be removed at each LRU
  93  * execution.
  94  */
  95 #define LRU_PERCENT_CLEAN 5
  96
  97 /* Run-time switchable: /sys/module/blkback/parameters/ */
  98 static unsigned int log_stats;
  99 module_param(log_stats, int, 0644);
 100
 101 #define BLKBACK_INVALID_HANDLE (~0)
 102
 103 /* Number of free pages to remove on each call to free_xenballooned_pages */
 104 #define NUM_BATCH_FREE_PAGES 10
 105
 106 static inline int get_free_page(struct xen_blkif *blkif, struct page **page)
 107 {
 108         unsigned long flags;
 109
 110         spin_lock_irqsave(&blkif->free_pages_lock, flags);
 111         if (list_empty(&blkif->free_pages)) {
 112                 BUG_ON(blkif->free_pages_num != 0);
 113                 spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
 114                 return alloc_xenballooned_pages(1, page, false);
 115         }
 116         BUG_ON(blkif->free_pages_num == 0);
 117         page[0] = list_first_entry(&blkif->free_pages, struct page, lru);
 118         list_del(&page[0]->lru);
 119         blkif->free_pages_num--;
 120         spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
 121
 122         return 0;
 123 }
 124
 125 static inline void put_free_pages(struct xen_blkif *blkif, struct page **page,
 126                                   int num)
 127 {
 128         unsigned long flags;
 129         int i;
 130
 131         spin_lock_irqsave(&blkif->free_pages_lock, flags);
 132         for (i = 0; i < num; i++)
 133                 list_add(&page[i]->lru, &blkif->free_pages);
 134         blkif->free_pages_num += num;
 135         spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
 136 }
 137
 138 static inline void shrink_free_pagepool(struct xen_blkif *blkif, int num)
 139 {
 140         /* Remove requested pages in batches of NUM_BATCH_FREE_PAGES */
 141         struct page *page[NUM_BATCH_FREE_PAGES];
 142         unsigned int num_pages = 0;
 143         unsigned long flags;
 144
 145         spin_lock_irqsave(&blkif->free_pages_lock, flags);
 146         while (blkif->free_pages_num > num) {
 147                 BUG_ON(list_empty(&blkif->free_pages));
 148                 page[num_pages] = list_first_entry(&blkif->free_pages,
 149                                                    struct page, lru);
 150                 list_del(&page[num_pages]->lru);
 151                 blkif->free_pages_num--;
 152                 if (++num_pages == NUM_BATCH_FREE_PAGES) {
 153                         spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
 154                         free_xenballooned_pages(num_pages, page);
 155                         spin_lock_irqsave(&blkif->free_pages_lock, flags);
 156                         num_pages = 0;
 157                 }
 158         }
 159         spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
 160         if (num_pages != 0)
 161                 free_xenballooned_pages(num_pages, page);
 162 }
 163
 164 #define vaddr(page) ((unsigned long)pfn_to_kaddr(page_to_pfn(page)))
 165
 166 static int do_block_io_op(struct xen_blkif *blkif);
 167 static int dispatch_rw_block_io(struct xen_blkif *blkif,
 168                                 struct blkif_request *req,
 169                                 struct pending_req *pending_req);
 170 static void make_response(struct xen_blkif *blkif, u64 id,
 171                           unsigned short op, int st);
 172
 173 #define foreach_grant_safe(pos, n, rbtree, node) \
 174         for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node), \
 175              (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL; \
 176              &(pos)->node != NULL; \
 177              (pos) = container_of(n, typeof(*(pos)), node), \
 178              (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL)
 179
 180
 181 /*
 182  * We don't need locking around the persistent grant helpers
 183  * because blkback uses a single-thread for each backed, so we
 184  * can be sure that this functions will never be called recursively.
 185  *
 186  * The only exception to that is put_persistent_grant, that can be called
 187  * from interrupt context (by xen_blkbk_unmap), so we have to use atomic
 188  * bit operations to modify the flags of a persistent grant and to count
 189  * the number of used grants.
 190  */
 191 static int add_persistent_gnt(struct xen_blkif *blkif,
 192                                struct persistent_gnt *persistent_gnt)
 193 {
 194         struct rb_node **new = NULL, *parent = NULL;
 195         struct persistent_gnt *this;
 196
 197         if (blkif->persistent_gnt_c >= xen_blkif_max_pgrants) {
 198                 if (!blkif->vbd.overflow_max_grants)
 199                         blkif->vbd.overflow_max_grants = 1;
 200                 return -EBUSY;
 201         }
 202         /* Figure out where to put new node */
 203         new = &blkif->persistent_gnts.rb_node;
 204         while (*new) {
 205                 this = container_of(*new, struct persistent_gnt, node);
 206
 207                 parent = *new;
 208                 if (persistent_gnt->gnt < this->gnt)
 209                         new = &((*new)->rb_left);
 210                 else if (persistent_gnt->gnt > this->gnt)
 211                         new = &((*new)->rb_right);
 212                 else {
 213                         pr_alert_ratelimited(DRV_PFX " trying to add a gref that's already in the tree\n");
 214                         return -EINVAL;
 215                 }
 216         }
 217
 218         bitmap_zero(persistent_gnt->flags, PERSISTENT_GNT_FLAGS_SIZE);
 219         set_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
 220         /* Add new node and rebalance tree. */
 221         rb_link_node(&(persistent_gnt->node), parent, new);
 222         rb_insert_color(&(persistent_gnt->node), &blkif->persistent_gnts);
 223         blkif->persistent_gnt_c++;
 224         atomic_inc(&blkif->persistent_gnt_in_use);
 225         return 0;
 226 }
 227
 228 static struct persistent_gnt *get_persistent_gnt(struct xen_blkif *blkif,
 229                                                  grant_ref_t gref)
 230 {
 231         struct persistent_gnt *data;
 232         struct rb_node *node = NULL;
 233
 234         node = blkif->persistent_gnts.rb_node;
 235         while (node) {
 236                 data = container_of(node, struct persistent_gnt, node);
 237
 238                 if (gref < data->gnt)
 239                         node = node->rb_left;
 240                 else if (gref > data->gnt)
 241                         node = node->rb_right;
 242                 else {
 243                         if(test_bit(PERSISTENT_GNT_ACTIVE, data->flags)) {
 244                                 pr_alert_ratelimited(DRV_PFX " requesting a grant already in use\n");
 245                                 return NULL;
 246                         }
 247                         set_bit(PERSISTENT_GNT_ACTIVE, data->flags);
 248                         atomic_inc(&blkif->persistent_gnt_in_use);
 249                         return data;
 250                 }
 251         }
 252         return NULL;
 253 }
 254
 255 static void put_persistent_gnt(struct xen_blkif *blkif,
 256                                struct persistent_gnt *persistent_gnt)
 257 {
 258         if(!test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
 259                   pr_alert_ratelimited(DRV_PFX " freeing a grant already unused");
 260         set_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
 261         clear_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
 262         atomic_dec(&blkif->persistent_gnt_in_use);
 263 }
 264
 265 static void free_persistent_gnts(struct xen_blkif *blkif, struct rb_root *root,
 266                                  unsigned int num)
 267 {
 268         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 269         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 270         struct persistent_gnt *persistent_gnt;
 271         struct rb_node *n;
 272         int ret = 0;
 273         int segs_to_unmap = 0;
 274
 275         foreach_grant_safe(persistent_gnt, n, root, node) {
 276                 BUG_ON(persistent_gnt->handle ==
 277                         BLKBACK_INVALID_HANDLE);
 278                 gnttab_set_unmap_op(&unmap[segs_to_unmap],
 279                         (unsigned long) pfn_to_kaddr(page_to_pfn(
 280                                 persistent_gnt->page)),
 281                         GNTMAP_host_map,
 282                         persistent_gnt->handle);
 283
 284                 pages[segs_to_unmap] = persistent_gnt->page;
 285
 286                 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
 287                         !rb_next(&persistent_gnt->node)) {
 288                         ret = gnttab_unmap_refs(unmap, NULL, pages,
 289                                 segs_to_unmap);
 290                         BUG_ON(ret);
 291                         put_free_pages(blkif, pages, segs_to_unmap);
 292                         segs_to_unmap = 0;
 293                 }
 294
 295                 rb_erase(&persistent_gnt->node, root);
 296                 kfree(persistent_gnt);
 297                 num--;
 298         }
 299         BUG_ON(num != 0);
 300 }
 301
 302 void xen_blkbk_unmap_purged_grants(struct work_struct *work)
 303 {
 304         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 305         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 306         struct persistent_gnt *persistent_gnt;
 307         int ret, segs_to_unmap = 0;
 308         struct xen_blkif *blkif = container_of(work, typeof(*blkif), persistent_purge_work);
 309
 310         while(!list_empty(&blkif->persistent_purge_list)) {
 311                 persistent_gnt = list_first_entry(&blkif->persistent_purge_list,
 312                                                   struct persistent_gnt,
 313                                                   remove_node);
 314                 list_del(&persistent_gnt->remove_node);
 315
 316                 gnttab_set_unmap_op(&unmap[segs_to_unmap],
 317                         vaddr(persistent_gnt->page),
 318                         GNTMAP_host_map,
 319                         persistent_gnt->handle);
 320
 321                 pages[segs_to_unmap] = persistent_gnt->page;
 322
 323                 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 324                         ret = gnttab_unmap_refs(unmap, NULL, pages,
 325                                 segs_to_unmap);
 326                         BUG_ON(ret);
 327                         put_free_pages(blkif, pages, segs_to_unmap);
 328                         segs_to_unmap = 0;
 329                 }
 330                 kfree(persistent_gnt);
 331         }
 332         if (segs_to_unmap > 0) {
 333                 ret = gnttab_unmap_refs(unmap, NULL, pages, segs_to_unmap);
 334                 BUG_ON(ret);
 335                 put_free_pages(blkif, pages, segs_to_unmap);
 336         }
 337 }
 338
 339 static void purge_persistent_gnt(struct xen_blkif *blkif)
 340 {
 341         struct persistent_gnt *persistent_gnt;
 342         struct rb_node *n;
 343         unsigned int num_clean, total;
 344         bool scan_used = false, clean_used = false;
 345         struct rb_root *root;
 346
 347         if (blkif->persistent_gnt_c < xen_blkif_max_pgrants ||
 348             (blkif->persistent_gnt_c == xen_blkif_max_pgrants &&
 349             !blkif->vbd.overflow_max_grants)) {
 350                 return;
 351         }
 352
 353         if (work_pending(&blkif->persistent_purge_work)) {
 354                 pr_alert_ratelimited(DRV_PFX "Scheduled work from previous purge is still pending, cannot purge list\n");
 355                 return;
 356         }
 357
 358         num_clean = (xen_blkif_max_pgrants / 100) * LRU_PERCENT_CLEAN;
 359         num_clean = blkif->persistent_gnt_c - xen_blkif_max_pgrants + num_clean;
 360         num_clean = min(blkif->persistent_gnt_c, num_clean);
 361         if ((num_clean == 0) ||
 362             (num_clean > (blkif->persistent_gnt_c - atomic_read(&blkif->persistent_gnt_in_use))))
 363                 return;
 364
 365         /*
 366          * At this point, we can assure that there will be no calls
 367          * to get_persistent_grant (because we are executing this code from
 368          * xen_blkif_schedule), there can only be calls to put_persistent_gnt,
 369          * which means that the number of currently used grants will go down,
 370          * but never up, so we will always be able to remove the requested
 371          * number of grants.
 372          */
 373
 374         total = num_clean;
 375
 376         pr_debug(DRV_PFX "Going to purge %u persistent grants\n", num_clean);
 377
 378         BUG_ON(!list_empty(&blkif->persistent_purge_list));
 379         root = &blkif->persistent_gnts;
 380 purge_list:
 381         foreach_grant_safe(persistent_gnt, n, root, node) {
 382                 BUG_ON(persistent_gnt->handle ==
 383                         BLKBACK_INVALID_HANDLE);
 384
 385                 if (clean_used) {
 386                         clear_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
 387                         continue;
 388                 }
 389
 390                 if (test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
 391                         continue;
 392                 if (!scan_used &&
 393                     (test_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags)))
 394                         continue;
 395
 396                 rb_erase(&persistent_gnt->node, root);
 397                 list_add(&persistent_gnt->remove_node,
 398                          &blkif->persistent_purge_list);
 399                 if (--num_clean == 0)
 400                         goto finished;
 401         }
 402         /*
 403          * If we get here it means we also need to start cleaning
 404          * grants that were used since last purge in order to cope
 405          * with the requested num
 406          */
 407         if (!scan_used && !clean_used) {
 408                 pr_debug(DRV_PFX "Still missing %u purged frames\n", num_clean);
 409                 scan_used = true;
 410                 goto purge_list;
 411         }
 412 finished:
 413         if (!clean_used) {
 414                 pr_debug(DRV_PFX "Finished scanning for grants to clean, removing used flag\n");
 415                 clean_used = true;
 416                 goto purge_list;
 417         }
 418
 419         blkif->persistent_gnt_c -= (total - num_clean);
 420         blkif->vbd.overflow_max_grants = 0;
 421
 422         /* We can defer this work */
 423         schedule_work(&blkif->persistent_purge_work);
 424         pr_debug(DRV_PFX "Purged %u/%u\n", (total - num_clean), total);
 425         return;
 426 }
 427
 428 /*
 429  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
 430  */
 431 static struct pending_req *alloc_req(struct xen_blkif *blkif)
 432 {
 433         struct pending_req *req = NULL;
 434         unsigned long flags;
 435
 436         spin_lock_irqsave(&blkif->pending_free_lock, flags);
 437         if (!list_empty(&blkif->pending_free)) {
 438                 req = list_entry(blkif->pending_free.next, struct pending_req,
 439                                  free_list);
 440                 list_del(&req->free_list);
 441         }
 442         spin_unlock_irqrestore(&blkif->pending_free_lock, flags);
 443         return req;
 444 }
 445
 446 /*
 447  * Return the 'pending_req' structure back to the freepool. We also
 448  * wake up the thread if it was waiting for a free page.
 449  */
 450 static void free_req(struct xen_blkif *blkif, struct pending_req *req)
 451 {
 452         unsigned long flags;
 453         int was_empty;
 454
 455         spin_lock_irqsave(&blkif->pending_free_lock, flags);
 456         was_empty = list_empty(&blkif->pending_free);
 457         list_add(&req->free_list, &blkif->pending_free);
 458         spin_unlock_irqrestore(&blkif->pending_free_lock, flags);
 459         if (was_empty)
 460                 wake_up(&blkif->pending_free_wq);
 461 }
 462
 463 /*
 464  * Routines for managing virtual block devices (vbds).
 465  */
 466 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
 467                              int operation)
 468 {
 469         struct xen_vbd *vbd = &blkif->vbd;
 470         int rc = -EACCES;
 471
 472         if ((operation != READ) && vbd->readonly)
 473                 goto out;
 474
 475         if (likely(req->nr_sects)) {
 476                 blkif_sector_t end = req->sector_number + req->nr_sects;
 477
 478                 if (unlikely(end < req->sector_number))
 479                         goto out;
 480                 if (unlikely(end > vbd_sz(vbd)))
 481                         goto out;
 482         }
 483
 484         req->dev  = vbd->pdevice;
 485         req->bdev = vbd->bdev;
 486         rc = 0;
 487
 488  out:
 489         return rc;
 490 }
 491
 492 static void xen_vbd_resize(struct xen_blkif *blkif)
 493 {
 494         struct xen_vbd *vbd = &blkif->vbd;
 495         struct xenbus_transaction xbt;
 496         int err;
 497         struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
 498         unsigned long long new_size = vbd_sz(vbd);
 499
 500         pr_info(DRV_PFX "VBD Resize: Domid: %d, Device: (%d, %d)\n",
 501                 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
 502         pr_info(DRV_PFX "VBD Resize: new size %llu\n", new_size);
 503         vbd->size = new_size;
 504 again:
 505         err = xenbus_transaction_start(&xbt);
 506         if (err) {
 507                 pr_warn(DRV_PFX "Error starting transaction");
 508                 return;
 509         }
 510         err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
 511                             (unsigned long long)vbd_sz(vbd));
 512         if (err) {
 513                 pr_warn(DRV_PFX "Error writing new size");
 514                 goto abort;
 515         }
 516         /*
 517          * Write the current state; we will use this to synchronize
 518          * the front-end. If the current state is "connected" the
 519          * front-end will get the new size information online.
 520          */
 521         err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
 522         if (err) {
 523                 pr_warn(DRV_PFX "Error writing the state");
 524                 goto abort;
 525         }
 526
 527         err = xenbus_transaction_end(xbt, 0);
 528         if (err == -EAGAIN)
 529                 goto again;
 530         if (err)
 531                 pr_warn(DRV_PFX "Error ending transaction");
 532         return;
 533 abort:
 534         xenbus_transaction_end(xbt, 1);
 535 }
 536
 537 /*
 538  * Notification from the guest OS.
 539  */
 540 static void blkif_notify_work(struct xen_blkif *blkif)
 541 {
 542         blkif->waiting_reqs = 1;
 543         wake_up(&blkif->wq);
 544 }
 545
 546 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
 547 {
 548         blkif_notify_work(dev_id);
 549         return IRQ_HANDLED;
 550 }
 551
 552 /*
 553  * SCHEDULER FUNCTIONS
 554  */
 555
 556 static void print_stats(struct xen_blkif *blkif)
 557 {
 558         pr_info("xen-blkback (%s): oo %3llu  |  rd %4llu  |  wr %4llu  |  f %4llu"
 559                  "  |  ds %4llu | pg: %4u/%4d\n",
 560                  current->comm, blkif->st_oo_req,
 561                  blkif->st_rd_req, blkif->st_wr_req,
 562                  blkif->st_f_req, blkif->st_ds_req,
 563                  blkif->persistent_gnt_c,
 564                  xen_blkif_max_pgrants);
 565         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 566         blkif->st_rd_req = 0;
 567         blkif->st_wr_req = 0;
 568         blkif->st_oo_req = 0;
 569         blkif->st_ds_req = 0;
 570 }
 571
 572 int xen_blkif_schedule(void *arg)
 573 {
 574         struct xen_blkif *blkif = arg;
 575         struct xen_vbd *vbd = &blkif->vbd;
 576         unsigned long timeout;
 577         int ret;
 578
 579         xen_blkif_get(blkif);
 580
 581         while (!kthread_should_stop()) {
 582                 if (try_to_freeze())
 583                         continue;
 584                 if (unlikely(vbd->size != vbd_sz(vbd)))
 585                         xen_vbd_resize(blkif);
 586
 587                 timeout = msecs_to_jiffies(LRU_INTERVAL);
 588
 589                 timeout = wait_event_interruptible_timeout(
 590                         blkif->wq,
 591                         blkif->waiting_reqs || kthread_should_stop(),
 592                         timeout);
 593                 if (timeout == 0)
 594                         goto purge_gnt_list;
 595                 timeout = wait_event_interruptible_timeout(
 596                         blkif->pending_free_wq,
 597                         !list_empty(&blkif->pending_free) ||
 598                         kthread_should_stop(),
 599                         timeout);
 600                 if (timeout == 0)
 601                         goto purge_gnt_list;
 602
 603                 blkif->waiting_reqs = 0;
 604                 smp_mb(); /* clear flag *before* checking for work */
 605
 606                 ret = do_block_io_op(blkif);
 607                 if (ret > 0)
 608                         blkif->waiting_reqs = 1;
 609                 if (ret == -EACCES)
 610                         wait_event_interruptible(blkif->shutdown_wq,
 611                                                  kthread_should_stop());
 612
 613 purge_gnt_list:
 614                 if (blkif->vbd.feature_gnt_persistent &&
 615                     time_after(jiffies, blkif->next_lru)) {
 616                         purge_persistent_gnt(blkif);
 617                         blkif->next_lru = jiffies + msecs_to_jiffies(LRU_INTERVAL);
 618                 }
 619
 620                 /* Shrink if we have more than xen_blkif_max_buffer_pages */
 621                 shrink_free_pagepool(blkif, xen_blkif_max_buffer_pages);
 622
 623                 if (log_stats && time_after(jiffies, blkif->st_print))
 624                         print_stats(blkif);
 625         }
 626
 627         /* Drain pending purge work */
 628         flush_work(&blkif->persistent_purge_work);
 629
 630         if (log_stats)
 631                 print_stats(blkif);
 632
 633         blkif->xenblkd = NULL;
 634         xen_blkif_put(blkif);
 635
 636         return 0;
 637 }
 638
 639 /*
 640  * Remove persistent grants and empty the pool of free pages
 641  */
 642 void xen_blkbk_free_caches(struct xen_blkif *blkif)
 643 {
 644         /* Free all persistent grant pages */
 645         if (!RB_EMPTY_ROOT(&blkif->persistent_gnts))
 646                 free_persistent_gnts(blkif, &blkif->persistent_gnts,
 647                         blkif->persistent_gnt_c);
 648
 649         BUG_ON(!RB_EMPTY_ROOT(&blkif->persistent_gnts));
 650         blkif->persistent_gnt_c = 0;
 651
 652         /* Since we are shutting down remove all pages from the buffer */
 653         shrink_free_pagepool(blkif, 0 /* All */);
 654 }
 655
 656 /*
 657  * Unmap the grant references, and also remove the M2P over-rides
 658  * used in the 'pending_req'.
 659  */
 660 static void xen_blkbk_unmap(struct xen_blkif *blkif,
 661                             struct grant_page *pages[],
 662                             int num)
 663 {
 664         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 665         struct page *unmap_pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 666         unsigned int i, invcount = 0;
 667         int ret;
 668
 669         for (i = 0; i < num; i++) {
 670                 if (pages[i]->persistent_gnt != NULL) {
 671                         put_persistent_gnt(blkif, pages[i]->persistent_gnt);
 672                         continue;
 673                 }
 674                 if (pages[i]->handle == BLKBACK_INVALID_HANDLE)
 675                         continue;
 676                 unmap_pages[invcount] = pages[i]->page;
 677                 gnttab_set_unmap_op(&unmap[invcount], vaddr(pages[i]->page),
 678                                     GNTMAP_host_map, pages[i]->handle);
 679                 pages[i]->handle = BLKBACK_INVALID_HANDLE;
 680                 if (++invcount == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 681                         ret = gnttab_unmap_refs(unmap, NULL, unmap_pages,
 682                                                 invcount);
 683                         BUG_ON(ret);
 684                         put_free_pages(blkif, unmap_pages, invcount);
 685                         invcount = 0;
 686                 }
 687         }
 688         if (invcount) {
 689                 ret = gnttab_unmap_refs(unmap, NULL, unmap_pages, invcount);
 690                 BUG_ON(ret);
 691                 put_free_pages(blkif, unmap_pages, invcount);
 692         }
 693 }
 694
 695 static int xen_blkbk_map(struct xen_blkif *blkif,
 696                          struct grant_page *pages[],
 697                          int num, bool ro)
 698 {
 699         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 700         struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 701         struct persistent_gnt *persistent_gnt = NULL;
 702         phys_addr_t addr = 0;
 703         int i, seg_idx, new_map_idx;
 704         int segs_to_map = 0;
 705         int ret = 0;
 706         int last_map = 0, map_until = 0;
 707         int use_persistent_gnts;
 708
 709         use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
 710
 711         /*
 712          * Fill out preq.nr_sects with proper amount of sectors, and setup
 713          * assign map[..] with the PFN of the page in our domain with the
 714          * corresponding grant reference for each page.
 715          */
 716 again:
 717         for (i = map_until; i < num; i++) {
 718                 uint32_t flags;
 719
 720                 if (use_persistent_gnts)
 721                         persistent_gnt = get_persistent_gnt(
 722                                 blkif,
 723                                 pages[i]->gref);
 724
 725                 if (persistent_gnt) {
 726                         /*
 727                          * We are using persistent grants and
 728                          * the grant is already mapped
 729                          */
 730                         pages[i]->page = persistent_gnt->page;
 731                         pages[i]->persistent_gnt = persistent_gnt;
 732                 } else {
 733                         if (get_free_page(blkif, &pages[i]->page))
 734                                 goto out_of_memory;
 735                         addr = vaddr(pages[i]->page);
 736                         pages_to_gnt[segs_to_map] = pages[i]->page;
 737                         pages[i]->persistent_gnt = NULL;
 738                         flags = GNTMAP_host_map;
 739                         if (!use_persistent_gnts && ro)
 740                                 flags |= GNTMAP_readonly;
 741                         gnttab_set_map_op(&map[segs_to_map++], addr,
 742                                           flags, pages[i]->gref,
 743                                           blkif->domid);
 744                 }
 745                 map_until = i + 1;
 746                 if (segs_to_map == BLKIF_MAX_SEGMENTS_PER_REQUEST)
 747                         break;
 748         }
 749
 750         if (segs_to_map) {
 751                 ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
 752                 BUG_ON(ret);
 753         }
 754
 755         /*
 756          * Now swizzle the MFN in our domain with the MFN from the other domain
 757          * so that when we access vaddr(pending_req,i) it has the contents of
 758          * the page from the other domain.
 759          */
 760         for (seg_idx = last_map, new_map_idx = 0; seg_idx < map_until; seg_idx++) {
 761                 if (!pages[seg_idx]->persistent_gnt) {
 762                         /* This is a newly mapped grant */
 763                         BUG_ON(new_map_idx >= segs_to_map);
 764                         if (unlikely(map[new_map_idx].status != 0)) {
 765                                 pr_debug(DRV_PFX "invalid buffer -- could not remap it\n");
 766                                 put_free_pages(blkif, &pages[seg_idx]->page, 1);
 767                                 pages[seg_idx]->handle = BLKBACK_INVALID_HANDLE;
 768                                 ret |= 1;
 769                                 goto next;
 770                         }
 771                         pages[seg_idx]->handle = map[new_map_idx].handle;
 772                 } else {
 773                         continue;
 774                 }
 775                 if (use_persistent_gnts &&
 776                     blkif->persistent_gnt_c < xen_blkif_max_pgrants) {
 777                         /*
 778                          * We are using persistent grants, the grant is
 779                          * not mapped but we might have room for it.
 780                          */
 781                         persistent_gnt = kmalloc(sizeof(struct persistent_gnt),
 782                                                  GFP_KERNEL);
 783                         if (!persistent_gnt) {
 784                                 /*
 785                                  * If we don't have enough memory to
 786                                  * allocate the persistent_gnt struct
 787                                  * map this grant non-persistenly
 788                                  */
 789                                 goto next;
 790                         }
 791                         persistent_gnt->gnt = map[new_map_idx].ref;
 792                         persistent_gnt->handle = map[new_map_idx].handle;
 793                         persistent_gnt->page = pages[seg_idx]->page;
 794                         if (add_persistent_gnt(blkif,
 795                                                persistent_gnt)) {
 796                                 kfree(persistent_gnt);
 797                                 persistent_gnt = NULL;
 798                                 goto next;
 799                         }
 800                         pages[seg_idx]->persistent_gnt = persistent_gnt;
 801                         pr_debug(DRV_PFX " grant %u added to the tree of persistent grants, using %u/%u\n",
 802                                  persistent_gnt->gnt, blkif->persistent_gnt_c,
 803                                  xen_blkif_max_pgrants);
 804                         goto next;
 805                 }
 806                 if (use_persistent_gnts && !blkif->vbd.overflow_max_grants) {
 807                         blkif->vbd.overflow_max_grants = 1;
 808                         pr_debug(DRV_PFX " domain %u, device %#x is using maximum number of persistent grants\n",
 809                                  blkif->domid, blkif->vbd.handle);
 810                 }
 811                 /*
 812                  * We could not map this grant persistently, so use it as
 813                  * a non-persistent grant.
 814                  */
 815 next:
 816                 new_map_idx++;
 817         }
 818         segs_to_map = 0;
 819         last_map = map_until;
 820         if (map_until != num)
 821                 goto again;
 822
 823         return ret;
 824
 825 out_of_memory:
 826         pr_alert(DRV_PFX "%s: out of memory\n", __func__);
 827         put_free_pages(blkif, pages_to_gnt, segs_to_map);
 828         return -ENOMEM;
 829 }
 830
 831 static int xen_blkbk_map_seg(struct pending_req *pending_req)
 832 {
 833         int rc;
 834
 835         rc = xen_blkbk_map(pending_req->blkif, pending_req->segments,
 836                            pending_req->nr_pages,
 837                            (pending_req->operation != BLKIF_OP_READ));
 838
 839         return rc;
 840 }
 841
 842 static int xen_blkbk_parse_indirect(struct blkif_request *req,
 843                                     struct pending_req *pending_req,
 844                                     struct seg_buf seg[],
 845                                     struct phys_req *preq)
 846 {
 847         struct grant_page **pages = pending_req->indirect_pages;
 848         struct xen_blkif *blkif = pending_req->blkif;
 849         int indirect_grefs, rc, n, nseg, i;
 850         struct blkif_request_segment *segments = NULL;
 851
 852         nseg = pending_req->nr_pages;
 853         indirect_grefs = INDIRECT_PAGES(nseg);
 854         BUG_ON(indirect_grefs > BLKIF_MAX_INDIRECT_PAGES_PER_REQUEST);
 855
 856         for (i = 0; i < indirect_grefs; i++)
 857                 pages[i]->gref = req->u.indirect.indirect_grefs[i];
 858
 859         rc = xen_blkbk_map(blkif, pages, indirect_grefs, true);
 860         if (rc)
 861                 goto unmap;
 862
 863         for (n = 0, i = 0; n < nseg; n++) {
 864                 if ((n % SEGS_PER_INDIRECT_FRAME) == 0) {
 865                         /* Map indirect segments */
 866                         if (segments)
 867                                 kunmap_atomic(segments);
 868                         segments = kmap_atomic(pages[n/SEGS_PER_INDIRECT_FRAME]->page);
 869                 }
 870                 i = n % SEGS_PER_INDIRECT_FRAME;
 871                 pending_req->segments[n]->gref = segments[i].gref;
 872                 seg[n].nsec = segments[i].last_sect -
 873                         segments[i].first_sect + 1;
 874                 seg[n].offset = (segments[i].first_sect << 9);
 875                 if ((segments[i].last_sect >= (PAGE_SIZE >> 9)) ||
 876                     (segments[i].last_sect < segments[i].first_sect)) {
 877                         rc = -EINVAL;
 878                         goto unmap;
 879                 }
 880                 preq->nr_sects += seg[n].nsec;
 881         }
 882
 883 unmap:
 884         if (segments)
 885                 kunmap_atomic(segments);
 886         xen_blkbk_unmap(blkif, pages, indirect_grefs);
 887         return rc;
 888 }
 889
 890 static int dispatch_discard_io(struct xen_blkif *blkif,
 891                                 struct blkif_request *req)
 892 {
 893         int err = 0;
 894         int status = BLKIF_RSP_OKAY;
 895         struct block_device *bdev = blkif->vbd.bdev;
 896         unsigned long secure;
 897         struct phys_req preq;
 898
 899         xen_blkif_get(blkif);
 900
 901         preq.sector_number = req->u.discard.sector_number;
 902         preq.nr_sects      = req->u.discard.nr_sectors;
 903
 904         err = xen_vbd_translate(&preq, blkif, WRITE);
 905         if (err) {
 906                 pr_warn(DRV_PFX "access denied: DISCARD [%llu->%llu] on dev=%04x\n",
 907                         preq.sector_number,
 908                         preq.sector_number + preq.nr_sects, blkif->vbd.pdevice);
 909                 goto fail_response;
 910         }
 911         blkif->st_ds_req++;
 912
 913         secure = (blkif->vbd.discard_secure &&
 914                  (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
 915                  BLKDEV_DISCARD_SECURE : 0;
 916
 917         err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
 918                                    req->u.discard.nr_sectors,
 919                                    GFP_KERNEL, secure);
 920 fail_response:
 921         if (err == -EOPNOTSUPP) {
 922                 pr_debug(DRV_PFX "discard op failed, not supported\n");
 923                 status = BLKIF_RSP_EOPNOTSUPP;
 924         } else if (err)
 925                 status = BLKIF_RSP_ERROR;
 926
 927         make_response(blkif, req->u.discard.id, req->operation, status);
 928         xen_blkif_put(blkif);
 929         return err;
 930 }
 931
 932 static int dispatch_other_io(struct xen_blkif *blkif,
 933                              struct blkif_request *req,
 934                              struct pending_req *pending_req)
 935 {
 936         free_req(blkif, pending_req);
 937         make_response(blkif, req->u.other.id, req->operation,
 938                       BLKIF_RSP_EOPNOTSUPP);
 939         return -EIO;
 940 }
 941
 942 static void xen_blk_drain_io(struct xen_blkif *blkif)
 943 {
 944         atomic_set(&blkif->drain, 1);
 945         do {
 946                 if (atomic_read(&blkif->inflight) == 0)
 947                         break;
 948                 wait_for_completion_interruptible_timeout(
 949                                 &blkif->drain_complete, HZ);
 950
 951                 if (!atomic_read(&blkif->drain))
 952                         break;
 953         } while (!kthread_should_stop());
 954         atomic_set(&blkif->drain, 0);
 955 }
 956
 957 /*
 958  * Completion callback on the bio's. Called as bh->b_end_io()
 959  */
 960
 961 static void __end_block_io_op(struct pending_req *pending_req, int error)
 962 {
 963         /* An error fails the entire request. */
 964         if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
 965             (error == -EOPNOTSUPP)) {
 966                 pr_debug(DRV_PFX "flush diskcache op failed, not supported\n");
 967                 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
 968                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 969         } else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
 970                     (error == -EOPNOTSUPP)) {
 971                 pr_debug(DRV_PFX "write barrier op failed, not supported\n");
 972                 xen_blkbk_barrier(XBT_NIL, pending_req->blkif->be, 0);
 973                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 974         } else if (error) {
 975                 pr_debug(DRV_PFX "Buffer not up-to-date at end of operation,"
 976                          " error=%d\n", error);
 977                 pending_req->status = BLKIF_RSP_ERROR;
 978         }
 979
 980         /*
 981          * If all of the bio's have completed it is time to unmap
 982          * the grant references associated with 'request' and provide
 983          * the proper response on the ring.
 984          */
 985         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 986                 struct xen_blkif *blkif = pending_req->blkif;
 987
 988                 xen_blkbk_unmap(blkif,
 989                                 pending_req->segments,
 990                                 pending_req->nr_pages);
 991                 make_response(blkif, pending_req->id,
 992                               pending_req->operation, pending_req->status);
 993                 free_req(blkif, pending_req);
 994                 /*
 995                  * Make sure the request is freed before releasing blkif,
 996                  * or there could be a race between free_req and the
 997                  * cleanup done in xen_blkif_free during shutdown.
 998                  *
 999                  * NB: The fact that we might try to wake up pending_free_wq
1000                  * before drain_complete (in case there's a drain going on)
1001                  * it's not a problem with our current implementation
1002                  * because we can assure there's no thread waiting on
1003                  * pending_free_wq if there's a drain going on, but it has
1004                  * to be taken into account if the current model is changed.
1005                  */
1006                 if (atomic_dec_and_test(&blkif->inflight) && atomic_read(&blkif->drain)) {
1007                         complete(&blkif->drain_complete);
1008                 }
1009                 xen_blkif_put(blkif);
1010         }
1011 }
1012
1013 /*
1014  * bio callback.
1015  */
1016 static void end_block_io_op(struct bio *bio, int error)
1017 {
1018         __end_block_io_op(bio->bi_private, error);
1019         bio_put(bio);
1020 }
1021
1022
1023
1024 /*
1025  * Function to copy the from the ring buffer the 'struct blkif_request'
1026  * (which has the sectors we want, number of them, grant references, etc),
1027  * and transmute  it to the block API to hand it over to the proper block disk.
1028  */
1029 static int
1030 __do_block_io_op(struct xen_blkif *blkif)
1031 {
1032         union blkif_back_rings *blk_rings = &blkif->blk_rings;
1033         struct blkif_request req;
1034         struct pending_req *pending_req;
1035         RING_IDX rc, rp;
1036         int more_to_do = 0;
1037
1038         rc = blk_rings->common.req_cons;
1039         rp = blk_rings->common.sring->req_prod;
1040         rmb(); /* Ensure we see queued requests up to 'rp'. */
1041
1042         if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
1043                 rc = blk_rings->common.rsp_prod_pvt;
1044                 pr_warn(DRV_PFX "Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
1045                         rp, rc, rp - rc, blkif->vbd.pdevice);
1046                 return -EACCES;
1047         }
1048         while (rc != rp) {
1049
1050                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
1051                         break;
1052
1053                 if (kthread_should_stop()) {
1054                         more_to_do = 1;
1055                         break;
1056                 }
1057
1058                 pending_req = alloc_req(blkif);
1059                 if (NULL == pending_req) {
1060                         blkif->st_oo_req++;
1061                         more_to_do = 1;
1062                         break;
1063                 }
1064
1065                 switch (blkif->blk_protocol) {
1066                 case BLKIF_PROTOCOL_NATIVE:
1067                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
1068                         break;
1069                 case BLKIF_PROTOCOL_X86_32:
1070                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
1071                         break;
1072                 case BLKIF_PROTOCOL_X86_64:
1073                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
1074                         break;
1075                 default:
1076                         BUG();
1077                 }
1078                 blk_rings->common.req_cons = ++rc; /* before make_response() */
1079
1080                 /* Apply all sanity checks to /private copy/ of request. */
1081                 barrier();
1082
1083                 switch (req.operation) {
1084                 case BLKIF_OP_READ:
1085                 case BLKIF_OP_WRITE:
1086                 case BLKIF_OP_WRITE_BARRIER:
1087                 case BLKIF_OP_FLUSH_DISKCACHE:
1088                 case BLKIF_OP_INDIRECT:
1089                         if (dispatch_rw_block_io(blkif, &req, pending_req))
1090                                 goto done;
1091                         break;
1092                 case BLKIF_OP_DISCARD:
1093                         free_req(blkif, pending_req);
1094                         if (dispatch_discard_io(blkif, &req))
1095                                 goto done;
1096                         break;
1097                 default:
1098                         if (dispatch_other_io(blkif, &req, pending_req))
1099                                 goto done;
1100                         break;
1101                 }
1102
1103                 /* Yield point for this unbounded loop. */
1104                 cond_resched();
1105         }
1106 done:
1107         return more_to_do;
1108 }
1109
1110 static int
1111 do_block_io_op(struct xen_blkif *blkif)
1112 {
1113         union blkif_back_rings *blk_rings = &blkif->blk_rings;
1114         int more_to_do;
1115
1116         do {
1117                 more_to_do = __do_block_io_op(blkif);
1118                 if (more_to_do)
1119                         break;
1120
1121                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
1122         } while (more_to_do);
1123
1124         return more_to_do;
1125 }
1126 /*
1127  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
1128  * and call the 'submit_bio' to pass it to the underlying storage.
1129  */
1130 static int dispatch_rw_block_io(struct xen_blkif *blkif,
1131                                 struct blkif_request *req,
1132                                 struct pending_req *pending_req)
1133 {
1134         struct phys_req preq;
1135         struct seg_buf *seg = pending_req->seg;
1136         unsigned int nseg;
1137         struct bio *bio = NULL;
1138         struct bio **biolist = pending_req->biolist;
1139         int i, nbio = 0;
1140         int operation;
1141         struct blk_plug plug;
1142         bool drain = false;
1143         struct grant_page **pages = pending_req->segments;
1144         unsigned short req_operation;
1145
1146         req_operation = req->operation == BLKIF_OP_INDIRECT ?
1147                         req->u.indirect.indirect_op : req->operation;
1148         if ((req->operation == BLKIF_OP_INDIRECT) &&
1149             (req_operation != BLKIF_OP_READ) &&
1150             (req_operation != BLKIF_OP_WRITE)) {
1151                 pr_debug(DRV_PFX "Invalid indirect operation (%u)\n",
1152                          req_operation);
1153                 goto fail_response;
1154         }
1155
1156         switch (req_operation) {
1157         case BLKIF_OP_READ:
1158                 blkif->st_rd_req++;
1159                 operation = READ;
1160                 break;
1161         case BLKIF_OP_WRITE:
1162                 blkif->st_wr_req++;
1163                 operation = WRITE_ODIRECT;
1164                 break;
1165         case BLKIF_OP_WRITE_BARRIER:
1166                 drain = true;
1167         case BLKIF_OP_FLUSH_DISKCACHE:
1168                 blkif->st_f_req++;
1169                 operation = WRITE_FLUSH;
1170                 break;
1171         default:
1172                 operation = 0; /* make gcc happy */
1173                 goto fail_response;
1174                 break;
1175         }
1176
1177         /* Check that the number of segments is sane. */
1178         nseg = req->operation == BLKIF_OP_INDIRECT ?
1179                req->u.indirect.nr_segments : req->u.rw.nr_segments;
1180
1181         if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
1182             unlikely((req->operation != BLKIF_OP_INDIRECT) &&
1183                      (nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) ||
1184             unlikely((req->operation == BLKIF_OP_INDIRECT) &&
1185                      (nseg > MAX_INDIRECT_SEGMENTS))) {
1186                 pr_debug(DRV_PFX "Bad number of segments in request (%d)\n",
1187                          nseg);
1188                 /* Haven't submitted any bio's yet. */
1189                 goto fail_response;
1190         }
1191
1192         preq.nr_sects      = 0;
1193
1194         pending_req->blkif     = blkif;
1195         pending_req->id        = req->u.rw.id;
1196         pending_req->operation = req_operation;
1197         pending_req->status    = BLKIF_RSP_OKAY;
1198         pending_req->nr_pages  = nseg;
1199
1200         if (req->operation != BLKIF_OP_INDIRECT) {
1201                 preq.dev               = req->u.rw.handle;
1202                 preq.sector_number     = req->u.rw.sector_number;
1203                 for (i = 0; i < nseg; i++) {
1204                         pages[i]->gref = req->u.rw.seg[i].gref;
1205                         seg[i].nsec = req->u.rw.seg[i].last_sect -
1206                                 req->u.rw.seg[i].first_sect + 1;
1207                         seg[i].offset = (req->u.rw.seg[i].first_sect << 9);
1208                         if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
1209                             (req->u.rw.seg[i].last_sect <
1210                              req->u.rw.seg[i].first_sect))
1211                                 goto fail_response;
1212                         preq.nr_sects += seg[i].nsec;
1213                 }
1214         } else {
1215                 preq.dev               = req->u.indirect.handle;
1216                 preq.sector_number     = req->u.indirect.sector_number;
1217                 if (xen_blkbk_parse_indirect(req, pending_req, seg, &preq))
1218                         goto fail_response;
1219         }
1220
1221         if (xen_vbd_translate(&preq, blkif, operation) != 0) {
1222                 pr_debug(DRV_PFX "access denied: %s of [%llu,%llu] on dev=%04x\n",
1223                          operation == READ ? "read" : "write",
1224                          preq.sector_number,
1225                          preq.sector_number + preq.nr_sects,
1226                          blkif->vbd.pdevice);
1227                 goto fail_response;
1228         }
1229
1230         /*
1231          * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
1232          * is set there.
1233          */
1234         for (i = 0; i < nseg; i++) {
1235                 if (((int)preq.sector_number|(int)seg[i].nsec) &
1236                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
1237                         pr_debug(DRV_PFX "Misaligned I/O request from domain %d",
1238                                  blkif->domid);
1239                         goto fail_response;
1240                 }
1241         }
1242
1243         /* Wait on all outstanding I/O's and once that has been completed
1244          * issue the WRITE_FLUSH.
1245          */
1246         if (drain)
1247                 xen_blk_drain_io(pending_req->blkif);
1248
1249         /*
1250          * If we have failed at this point, we need to undo the M2P override,
1251          * set gnttab_set_unmap_op on all of the grant references and perform
1252          * the hypercall to unmap the grants - that is all done in
1253          * xen_blkbk_unmap.
1254          */
1255         if (xen_blkbk_map_seg(pending_req))
1256                 goto fail_flush;
1257
1258         /*
1259          * This corresponding xen_blkif_put is done in __end_block_io_op, or
1260          * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
1261          */
1262         xen_blkif_get(blkif);
1263         atomic_inc(&blkif->inflight);
1264
1265         for (i = 0; i < nseg; i++) {
1266                 while ((bio == NULL) ||
1267                        (bio_add_page(bio,
1268                                      pages[i]->page,
1269                                      seg[i].nsec << 9,
1270                                      seg[i].offset) == 0)) {
1271
1272                         int nr_iovecs = min_t(int, (nseg-i), BIO_MAX_PAGES);
1273                         bio = bio_alloc(GFP_KERNEL, nr_iovecs);
1274                         if (unlikely(bio == NULL))
1275                                 goto fail_put_bio;
1276
1277                         biolist[nbio++] = bio;
1278                         bio->bi_bdev    = preq.bdev;
1279                         bio->bi_private = pending_req;
1280                         bio->bi_end_io  = end_block_io_op;
1281                         bio->bi_iter.bi_sector  = preq.sector_number;
1282                 }
1283
1284                 preq.sector_number += seg[i].nsec;
1285         }
1286
1287         /* This will be hit if the operation was a flush or discard. */
1288         if (!bio) {
1289                 BUG_ON(operation != WRITE_FLUSH);
1290
1291                 bio = bio_alloc(GFP_KERNEL, 0);
1292                 if (unlikely(bio == NULL))
1293                         goto fail_put_bio;
1294
1295                 biolist[nbio++] = bio;
1296                 bio->bi_bdev    = preq.bdev;
1297                 bio->bi_private = pending_req;
1298                 bio->bi_end_io  = end_block_io_op;
1299         }
1300
1301         atomic_set(&pending_req->pendcnt, nbio);
1302         blk_start_plug(&plug);
1303
1304         for (i = 0; i < nbio; i++)
1305                 submit_bio(operation, biolist[i]);
1306
1307         /* Let the I/Os go.. */
1308         blk_finish_plug(&plug);
1309
1310         if (operation == READ)
1311                 blkif->st_rd_sect += preq.nr_sects;
1312         else if (operation & WRITE)
1313                 blkif->st_wr_sect += preq.nr_sects;
1314
1315         return 0;
1316
1317  fail_flush:
1318         xen_blkbk_unmap(blkif, pending_req->segments,
1319                         pending_req->nr_pages);
1320  fail_response:
1321         /* Haven't submitted any bio's yet. */
1322         make_response(blkif, req->u.rw.id, req_operation, BLKIF_RSP_ERROR);
1323         free_req(blkif, pending_req);
1324         msleep(1); /* back off a bit */
1325         return -EIO;
1326
1327  fail_put_bio:
1328         for (i = 0; i < nbio; i++)
1329                 bio_put(biolist[i]);
1330         atomic_set(&pending_req->pendcnt, 1);
1331         __end_block_io_op(pending_req, -EINVAL);
1332         msleep(1); /* back off a bit */
1333         return -EIO;
1334 }
1335
1336
1337
1338 /*
1339  * Put a response on the ring on how the operation fared.
1340  */
1341 static void make_response(struct xen_blkif *blkif, u64 id,
1342                           unsigned short op, int st)
1343 {
1344         struct blkif_response  resp;
1345         unsigned long     flags;
1346         union blkif_back_rings *blk_rings = &blkif->blk_rings;
1347         int notify;
1348
1349         resp.id        = id;
1350         resp.operation = op;
1351         resp.status    = st;
1352
1353         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
1354         /* Place on the response ring for the relevant domain. */
1355         switch (blkif->blk_protocol) {
1356         case BLKIF_PROTOCOL_NATIVE:
1357                 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
1358                        &resp, sizeof(resp));
1359                 break;
1360         case BLKIF_PROTOCOL_X86_32:
1361                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
1362                        &resp, sizeof(resp));
1363                 break;
1364         case BLKIF_PROTOCOL_X86_64:
1365                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
1366                        &resp, sizeof(resp));
1367                 break;
1368         default:
1369                 BUG();
1370         }
1371         blk_rings->common.rsp_prod_pvt++;
1372         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1373         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
1374         if (notify)
1375                 notify_remote_via_irq(blkif->irq);
1376 }
1377
1378 static int __init xen_blkif_init(void)
1379 {
1380         int rc = 0;
1381
1382         if (!xen_domain())
1383                 return -ENODEV;
1384
1385         rc = xen_blkif_interface_init();
1386         if (rc)
1387                 goto failed_init;
1388
1389         rc = xen_blkif_xenbus_init();
1390         if (rc)
1391                 goto failed_init;
1392
1393  failed_init:
1394         return rc;
1395 }
1396
1397 module_init(xen_blkif_init);
1398
1399 MODULE_LICENSE("Dual BSD/GPL");
1400 MODULE_ALIAS("xen-backend:vbd");