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 #define pr_fmt(fmt) "xen-blkback: " fmt
  38
  39 #include <linux/spinlock.h>
  40 #include <linux/kthread.h>
  41 #include <linux/list.h>
  42 #include <linux/delay.h>
  43 #include <linux/freezer.h>
  44 #include <linux/bitmap.h>
  45
  46 #include <xen/events.h>
  47 #include <xen/page.h>
  48 #include <xen/xen.h>
  49 #include <asm/xen/hypervisor.h>
  50 #include <asm/xen/hypercall.h>
  51 #include <xen/balloon.h>
  52 #include <xen/grant_table.h>
  53 #include "common.h"
  54
  55 /*
  56  * Maximum number of unused free pages to keep in the internal buffer.
  57  * Setting this to a value too low will reduce memory used in each backend,
  58  * but can have a performance penalty.
  59  *
  60  * A sane value is xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST, but can
  61  * be set to a lower value that might degrade performance on some intensive
  62  * IO workloads.
  63  */
  64
  65 static int xen_blkif_max_buffer_pages = 1024;
  66 module_param_named(max_buffer_pages, xen_blkif_max_buffer_pages, int, 0644);
  67 MODULE_PARM_DESC(max_buffer_pages,
  68 "Maximum number of free pages to keep in each block backend buffer");
  69
  70 /*
  71  * Maximum number of grants to map persistently in blkback. For maximum
  72  * performance this should be the total numbers of grants that can be used
  73  * to fill the ring, but since this might become too high, specially with
  74  * the use of indirect descriptors, we set it to a value that provides good
  75  * performance without using too much memory.
  76  *
  77  * When the list of persistent grants is full we clean it up using a LRU
  78  * algorithm.
  79  */
  80
  81 static int xen_blkif_max_pgrants = 1056;
  82 module_param_named(max_persistent_grants, xen_blkif_max_pgrants, int, 0644);
  83 MODULE_PARM_DESC(max_persistent_grants,
  84                  "Maximum number of grants to map persistently");
  85
  86 /*
  87  * How long a persistent grant is allowed to remain allocated without being in
  88  * use. The time is in seconds, 0 means indefinitely long.
  89  */
  90
  91 static unsigned int xen_blkif_pgrant_timeout = 60;
  92 module_param_named(persistent_grant_unused_seconds, xen_blkif_pgrant_timeout,
  93                    uint, 0644);
  94 MODULE_PARM_DESC(persistent_grant_unused_seconds,
  95                  "Time in seconds an unused persistent grant is allowed to "
  96                  "remain allocated. Default is 60, 0 means unlimited.");
  97
  98 /*
  99  * Maximum number of rings/queues blkback supports, allow as many queues as there
 100  * are CPUs if user has not specified a value.
 101  */
 102 unsigned int xenblk_max_queues;
 103 module_param_named(max_queues, xenblk_max_queues, uint, 0644);
 104 MODULE_PARM_DESC(max_queues,
 105                  "Maximum number of hardware queues per virtual disk." \
 106                  "By default it is the number of online CPUs.");
 107
 108 /*
 109  * Maximum order of pages to be used for the shared ring between front and
 110  * backend, 4KB page granularity is used.
 111  */
 112 unsigned int xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
 113 module_param_named(max_ring_page_order, xen_blkif_max_ring_order, int, 0444);
 114 MODULE_PARM_DESC(max_ring_page_order, "Maximum order of pages to be used for the shared ring");
 115 /*
 116  * The LRU mechanism to clean the lists of persistent grants needs to
 117  * be executed periodically. The time interval between consecutive executions
 118  * of the purge mechanism is set in ms.
 119  */
 120 #define LRU_INTERVAL 100
 121
 122 /*
 123  * When the persistent grants list is full we will remove unused grants
 124  * from the list. The percent number of grants to be removed at each LRU
 125  * execution.
 126  */
 127 #define LRU_PERCENT_CLEAN 5
 128
 129 /* Run-time switchable: /sys/module/blkback/parameters/ */
 130 static unsigned int log_stats;
 131 module_param(log_stats, int, 0644);
 132
 133 #define BLKBACK_INVALID_HANDLE (~0)
 134
 135 /* Number of free pages to remove on each call to gnttab_free_pages */
 136 #define NUM_BATCH_FREE_PAGES 10
 137
 138 static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt)
 139 {
 140         return xen_blkif_pgrant_timeout &&
 141                (jiffies - persistent_gnt->last_used >=
 142                 HZ * xen_blkif_pgrant_timeout);
 143 }
 144
 145 static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page)
 146 {
 147         unsigned long flags;
 148
 149         spin_lock_irqsave(&ring->free_pages_lock, flags);
 150         if (list_empty(&ring->free_pages)) {
 151                 BUG_ON(ring->free_pages_num != 0);
 152                 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
 153                 return gnttab_alloc_pages(1, page);
 154         }
 155         BUG_ON(ring->free_pages_num == 0);
 156         page[0] = list_first_entry(&ring->free_pages, struct page, lru);
 157         list_del(&page[0]->lru);
 158         ring->free_pages_num--;
 159         spin_unlock_irqrestore(&ring->free_pages_lock, flags);
 160
 161         return 0;
 162 }
 163
 164 static inline void put_free_pages(struct xen_blkif_ring *ring, struct page **page,
 165                                   int num)
 166 {
 167         unsigned long flags;
 168         int i;
 169
 170         spin_lock_irqsave(&ring->free_pages_lock, flags);
 171         for (i = 0; i < num; i++)
 172                 list_add(&page[i]->lru, &ring->free_pages);
 173         ring->free_pages_num += num;
 174         spin_unlock_irqrestore(&ring->free_pages_lock, flags);
 175 }
 176
 177 static inline void shrink_free_pagepool(struct xen_blkif_ring *ring, int num)
 178 {
 179         /* Remove requested pages in batches of NUM_BATCH_FREE_PAGES */
 180         struct page *page[NUM_BATCH_FREE_PAGES];
 181         unsigned int num_pages = 0;
 182         unsigned long flags;
 183
 184         spin_lock_irqsave(&ring->free_pages_lock, flags);
 185         while (ring->free_pages_num > num) {
 186                 BUG_ON(list_empty(&ring->free_pages));
 187                 page[num_pages] = list_first_entry(&ring->free_pages,
 188                                                    struct page, lru);
 189                 list_del(&page[num_pages]->lru);
 190                 ring->free_pages_num--;
 191                 if (++num_pages == NUM_BATCH_FREE_PAGES) {
 192                         spin_unlock_irqrestore(&ring->free_pages_lock, flags);
 193                         gnttab_free_pages(num_pages, page);
 194                         spin_lock_irqsave(&ring->free_pages_lock, flags);
 195                         num_pages = 0;
 196                 }
 197         }
 198         spin_unlock_irqrestore(&ring->free_pages_lock, flags);
 199         if (num_pages != 0)
 200                 gnttab_free_pages(num_pages, page);
 201 }
 202
 203 #define vaddr(page) ((unsigned long)pfn_to_kaddr(page_to_pfn(page)))
 204
 205 static int do_block_io_op(struct xen_blkif_ring *ring);
 206 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
 207                                 struct blkif_request *req,
 208                                 struct pending_req *pending_req);
 209 static void make_response(struct xen_blkif_ring *ring, u64 id,
 210                           unsigned short op, int st);
 211
 212 #define foreach_grant_safe(pos, n, rbtree, node) \
 213         for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node), \
 214              (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL; \
 215              &(pos)->node != NULL; \
 216              (pos) = container_of(n, typeof(*(pos)), node), \
 217              (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL)
 218
 219
 220 /*
 221  * We don't need locking around the persistent grant helpers
 222  * because blkback uses a single-thread for each backend, so we
 223  * can be sure that this functions will never be called recursively.
 224  *
 225  * The only exception to that is put_persistent_grant, that can be called
 226  * from interrupt context (by xen_blkbk_unmap), so we have to use atomic
 227  * bit operations to modify the flags of a persistent grant and to count
 228  * the number of used grants.
 229  */
 230 static int add_persistent_gnt(struct xen_blkif_ring *ring,
 231                                struct persistent_gnt *persistent_gnt)
 232 {
 233         struct rb_node **new = NULL, *parent = NULL;
 234         struct persistent_gnt *this;
 235         struct xen_blkif *blkif = ring->blkif;
 236
 237         if (ring->persistent_gnt_c >= xen_blkif_max_pgrants) {
 238                 if (!blkif->vbd.overflow_max_grants)
 239                         blkif->vbd.overflow_max_grants = 1;
 240                 return -EBUSY;
 241         }
 242         /* Figure out where to put new node */
 243         new = &ring->persistent_gnts.rb_node;
 244         while (*new) {
 245                 this = container_of(*new, struct persistent_gnt, node);
 246
 247                 parent = *new;
 248                 if (persistent_gnt->gnt < this->gnt)
 249                         new = &((*new)->rb_left);
 250                 else if (persistent_gnt->gnt > this->gnt)
 251                         new = &((*new)->rb_right);
 252                 else {
 253                         pr_alert_ratelimited("trying to add a gref that's already in the tree\n");
 254                         return -EINVAL;
 255                 }
 256         }
 257
 258         bitmap_zero(persistent_gnt->flags, PERSISTENT_GNT_FLAGS_SIZE);
 259         set_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
 260         /* Add new node and rebalance tree. */
 261         rb_link_node(&(persistent_gnt->node), parent, new);
 262         rb_insert_color(&(persistent_gnt->node), &ring->persistent_gnts);
 263         ring->persistent_gnt_c++;
 264         atomic_inc(&ring->persistent_gnt_in_use);
 265         return 0;
 266 }
 267
 268 static struct persistent_gnt *get_persistent_gnt(struct xen_blkif_ring *ring,
 269                                                  grant_ref_t gref)
 270 {
 271         struct persistent_gnt *data;
 272         struct rb_node *node = NULL;
 273
 274         node = ring->persistent_gnts.rb_node;
 275         while (node) {
 276                 data = container_of(node, struct persistent_gnt, node);
 277
 278                 if (gref < data->gnt)
 279                         node = node->rb_left;
 280                 else if (gref > data->gnt)
 281                         node = node->rb_right;
 282                 else {
 283                         if(test_bit(PERSISTENT_GNT_ACTIVE, data->flags)) {
 284                                 pr_alert_ratelimited("requesting a grant already in use\n");
 285                                 return NULL;
 286                         }
 287                         set_bit(PERSISTENT_GNT_ACTIVE, data->flags);
 288                         atomic_inc(&ring->persistent_gnt_in_use);
 289                         return data;
 290                 }
 291         }
 292         return NULL;
 293 }
 294
 295 static void put_persistent_gnt(struct xen_blkif_ring *ring,
 296                                struct persistent_gnt *persistent_gnt)
 297 {
 298         if(!test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
 299                 pr_alert_ratelimited("freeing a grant already unused\n");
 300         persistent_gnt->last_used = jiffies;
 301         clear_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
 302         atomic_dec(&ring->persistent_gnt_in_use);
 303 }
 304
 305 static void free_persistent_gnts(struct xen_blkif_ring *ring, struct rb_root *root,
 306                                  unsigned int num)
 307 {
 308         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 309         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 310         struct persistent_gnt *persistent_gnt;
 311         struct rb_node *n;
 312         int segs_to_unmap = 0;
 313         struct gntab_unmap_queue_data unmap_data;
 314
 315         unmap_data.pages = pages;
 316         unmap_data.unmap_ops = unmap;
 317         unmap_data.kunmap_ops = NULL;
 318
 319         foreach_grant_safe(persistent_gnt, n, root, node) {
 320                 BUG_ON(persistent_gnt->handle ==
 321                         BLKBACK_INVALID_HANDLE);
 322                 gnttab_set_unmap_op(&unmap[segs_to_unmap],
 323                         (unsigned long) pfn_to_kaddr(page_to_pfn(
 324                                 persistent_gnt->page)),
 325                         GNTMAP_host_map,
 326                         persistent_gnt->handle);
 327
 328                 pages[segs_to_unmap] = persistent_gnt->page;
 329
 330                 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
 331                         !rb_next(&persistent_gnt->node)) {
 332
 333                         unmap_data.count = segs_to_unmap;
 334                         BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
 335
 336                         put_free_pages(ring, pages, segs_to_unmap);
 337                         segs_to_unmap = 0;
 338                 }
 339
 340                 rb_erase(&persistent_gnt->node, root);
 341                 kfree(persistent_gnt);
 342                 num--;
 343         }
 344         BUG_ON(num != 0);
 345 }
 346
 347 void xen_blkbk_unmap_purged_grants(struct work_struct *work)
 348 {
 349         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 350         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 351         struct persistent_gnt *persistent_gnt;
 352         int segs_to_unmap = 0;
 353         struct xen_blkif_ring *ring = container_of(work, typeof(*ring), persistent_purge_work);
 354         struct gntab_unmap_queue_data unmap_data;
 355
 356         unmap_data.pages = pages;
 357         unmap_data.unmap_ops = unmap;
 358         unmap_data.kunmap_ops = NULL;
 359
 360         while(!list_empty(&ring->persistent_purge_list)) {
 361                 persistent_gnt = list_first_entry(&ring->persistent_purge_list,
 362                                                   struct persistent_gnt,
 363                                                   remove_node);
 364                 list_del(&persistent_gnt->remove_node);
 365
 366                 gnttab_set_unmap_op(&unmap[segs_to_unmap],
 367                         vaddr(persistent_gnt->page),
 368                         GNTMAP_host_map,
 369                         persistent_gnt->handle);
 370
 371                 pages[segs_to_unmap] = persistent_gnt->page;
 372
 373                 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 374                         unmap_data.count = segs_to_unmap;
 375                         BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
 376                         put_free_pages(ring, pages, segs_to_unmap);
 377                         segs_to_unmap = 0;
 378                 }
 379                 kfree(persistent_gnt);
 380         }
 381         if (segs_to_unmap > 0) {
 382                 unmap_data.count = segs_to_unmap;
 383                 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
 384                 put_free_pages(ring, pages, segs_to_unmap);
 385         }
 386 }
 387
 388 static void purge_persistent_gnt(struct xen_blkif_ring *ring)
 389 {
 390         struct persistent_gnt *persistent_gnt;
 391         struct rb_node *n;
 392         unsigned int num_clean, total;
 393         bool scan_used = false;
 394         struct rb_root *root;
 395
 396         if (work_busy(&ring->persistent_purge_work)) {
 397                 pr_alert_ratelimited("Scheduled work from previous purge is still busy, cannot purge list\n");
 398                 goto out;
 399         }
 400
 401         if (ring->persistent_gnt_c < xen_blkif_max_pgrants ||
 402             (ring->persistent_gnt_c == xen_blkif_max_pgrants &&
 403             !ring->blkif->vbd.overflow_max_grants)) {
 404                 num_clean = 0;
 405         } else {
 406                 num_clean = (xen_blkif_max_pgrants / 100) * LRU_PERCENT_CLEAN;
 407                 num_clean = ring->persistent_gnt_c - xen_blkif_max_pgrants +
 408                             num_clean;
 409                 num_clean = min(ring->persistent_gnt_c, num_clean);
 410                 pr_debug("Going to purge at least %u persistent grants\n",
 411                          num_clean);
 412         }
 413
 414         /*
 415          * At this point, we can assure that there will be no calls
 416          * to get_persistent_grant (because we are executing this code from
 417          * xen_blkif_schedule), there can only be calls to put_persistent_gnt,
 418          * which means that the number of currently used grants will go down,
 419          * but never up, so we will always be able to remove the requested
 420          * number of grants.
 421          */
 422
 423         total = 0;
 424
 425         BUG_ON(!list_empty(&ring->persistent_purge_list));
 426         root = &ring->persistent_gnts;
 427 purge_list:
 428         foreach_grant_safe(persistent_gnt, n, root, node) {
 429                 BUG_ON(persistent_gnt->handle ==
 430                         BLKBACK_INVALID_HANDLE);
 431
 432                 if (test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
 433                         continue;
 434                 if (!scan_used && !persistent_gnt_timeout(persistent_gnt))
 435                         continue;
 436                 if (scan_used && total >= num_clean)
 437                         continue;
 438
 439                 rb_erase(&persistent_gnt->node, root);
 440                 list_add(&persistent_gnt->remove_node,
 441                          &ring->persistent_purge_list);
 442                 total++;
 443         }
 444         /*
 445          * Check whether we also need to start cleaning
 446          * grants that were used since last purge in order to cope
 447          * with the requested num
 448          */
 449         if (!scan_used && total < num_clean) {
 450                 pr_debug("Still missing %u purged frames\n", num_clean - total);
 451                 scan_used = true;
 452                 goto purge_list;
 453         }
 454
 455         if (total) {
 456                 ring->persistent_gnt_c -= total;
 457                 ring->blkif->vbd.overflow_max_grants = 0;
 458
 459                 /* We can defer this work */
 460                 schedule_work(&ring->persistent_purge_work);
 461                 pr_debug("Purged %u/%u\n", num_clean, total);
 462         }
 463
 464 out:
 465         return;
 466 }
 467
 468 /*
 469  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
 470  */
 471 static struct pending_req *alloc_req(struct xen_blkif_ring *ring)
 472 {
 473         struct pending_req *req = NULL;
 474         unsigned long flags;
 475
 476         spin_lock_irqsave(&ring->pending_free_lock, flags);
 477         if (!list_empty(&ring->pending_free)) {
 478                 req = list_entry(ring->pending_free.next, struct pending_req,
 479                                  free_list);
 480                 list_del(&req->free_list);
 481         }
 482         spin_unlock_irqrestore(&ring->pending_free_lock, flags);
 483         return req;
 484 }
 485
 486 /*
 487  * Return the 'pending_req' structure back to the freepool. We also
 488  * wake up the thread if it was waiting for a free page.
 489  */
 490 static void free_req(struct xen_blkif_ring *ring, struct pending_req *req)
 491 {
 492         unsigned long flags;
 493         int was_empty;
 494
 495         spin_lock_irqsave(&ring->pending_free_lock, flags);
 496         was_empty = list_empty(&ring->pending_free);
 497         list_add(&req->free_list, &ring->pending_free);
 498         spin_unlock_irqrestore(&ring->pending_free_lock, flags);
 499         if (was_empty)
 500                 wake_up(&ring->pending_free_wq);
 501 }
 502
 503 /*
 504  * Routines for managing virtual block devices (vbds).
 505  */
 506 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
 507                              int operation)
 508 {
 509         struct xen_vbd *vbd = &blkif->vbd;
 510         int rc = -EACCES;
 511
 512         if ((operation != REQ_OP_READ) && vbd->readonly)
 513                 goto out;
 514
 515         if (likely(req->nr_sects)) {
 516                 blkif_sector_t end = req->sector_number + req->nr_sects;
 517
 518                 if (unlikely(end < req->sector_number))
 519                         goto out;
 520                 if (unlikely(end > vbd_sz(vbd)))
 521                         goto out;
 522         }
 523
 524         req->dev  = vbd->pdevice;
 525         req->bdev = vbd->bdev;
 526         rc = 0;
 527
 528  out:
 529         return rc;
 530 }
 531
 532 static void xen_vbd_resize(struct xen_blkif *blkif)
 533 {
 534         struct xen_vbd *vbd = &blkif->vbd;
 535         struct xenbus_transaction xbt;
 536         int err;
 537         struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
 538         unsigned long long new_size = vbd_sz(vbd);
 539
 540         pr_info("VBD Resize: Domid: %d, Device: (%d, %d)\n",
 541                 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
 542         pr_info("VBD Resize: new size %llu\n", new_size);
 543         vbd->size = new_size;
 544 again:
 545         err = xenbus_transaction_start(&xbt);
 546         if (err) {
 547                 pr_warn("Error starting transaction\n");
 548                 return;
 549         }
 550         err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
 551                             (unsigned long long)vbd_sz(vbd));
 552         if (err) {
 553                 pr_warn("Error writing new size\n");
 554                 goto abort;
 555         }
 556         /*
 557          * Write the current state; we will use this to synchronize
 558          * the front-end. If the current state is "connected" the
 559          * front-end will get the new size information online.
 560          */
 561         err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
 562         if (err) {
 563                 pr_warn("Error writing the state\n");
 564                 goto abort;
 565         }
 566
 567         err = xenbus_transaction_end(xbt, 0);
 568         if (err == -EAGAIN)
 569                 goto again;
 570         if (err)
 571                 pr_warn("Error ending transaction\n");
 572         return;
 573 abort:
 574         xenbus_transaction_end(xbt, 1);
 575 }
 576
 577 /*
 578  * Notification from the guest OS.
 579  */
 580 static void blkif_notify_work(struct xen_blkif_ring *ring)
 581 {
 582         ring->waiting_reqs = 1;
 583         wake_up(&ring->wq);
 584 }
 585
 586 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
 587 {
 588         blkif_notify_work(dev_id);
 589         return IRQ_HANDLED;
 590 }
 591
 592 /*
 593  * SCHEDULER FUNCTIONS
 594  */
 595
 596 static void print_stats(struct xen_blkif_ring *ring)
 597 {
 598         pr_info("(%s): oo %3llu  |  rd %4llu  |  wr %4llu  |  f %4llu"
 599                  "  |  ds %4llu | pg: %4u/%4d\n",
 600                  current->comm, ring->st_oo_req,
 601                  ring->st_rd_req, ring->st_wr_req,
 602                  ring->st_f_req, ring->st_ds_req,
 603                  ring->persistent_gnt_c,
 604                  xen_blkif_max_pgrants);
 605         ring->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 606         ring->st_rd_req = 0;
 607         ring->st_wr_req = 0;
 608         ring->st_oo_req = 0;
 609         ring->st_ds_req = 0;
 610 }
 611
 612 int xen_blkif_schedule(void *arg)
 613 {
 614         struct xen_blkif_ring *ring = arg;
 615         struct xen_blkif *blkif = ring->blkif;
 616         struct xen_vbd *vbd = &blkif->vbd;
 617         unsigned long timeout;
 618         int ret;
 619
 620         set_freezable();
 621         while (!kthread_should_stop()) {
 622                 if (try_to_freeze())
 623                         continue;
 624                 if (unlikely(vbd->size != vbd_sz(vbd)))
 625                         xen_vbd_resize(blkif);
 626
 627                 timeout = msecs_to_jiffies(LRU_INTERVAL);
 628
 629                 timeout = wait_event_interruptible_timeout(
 630                         ring->wq,
 631                         ring->waiting_reqs || kthread_should_stop(),
 632                         timeout);
 633                 if (timeout == 0)
 634                         goto purge_gnt_list;
 635                 timeout = wait_event_interruptible_timeout(
 636                         ring->pending_free_wq,
 637                         !list_empty(&ring->pending_free) ||
 638                         kthread_should_stop(),
 639                         timeout);
 640                 if (timeout == 0)
 641                         goto purge_gnt_list;
 642
 643                 ring->waiting_reqs = 0;
 644                 smp_mb(); /* clear flag *before* checking for work */
 645
 646                 ret = do_block_io_op(ring);
 647                 if (ret > 0)
 648                         ring->waiting_reqs = 1;
 649                 if (ret == -EACCES)
 650                         wait_event_interruptible(ring->shutdown_wq,
 651                                                  kthread_should_stop());
 652
 653 purge_gnt_list:
 654                 if (blkif->vbd.feature_gnt_persistent &&
 655                     time_after(jiffies, ring->next_lru)) {
 656                         purge_persistent_gnt(ring);
 657                         ring->next_lru = jiffies + msecs_to_jiffies(LRU_INTERVAL);
 658                 }
 659
 660                 /* Shrink if we have more than xen_blkif_max_buffer_pages */
 661                 shrink_free_pagepool(ring, xen_blkif_max_buffer_pages);
 662
 663                 if (log_stats && time_after(jiffies, ring->st_print))
 664                         print_stats(ring);
 665         }
 666
 667         /* Drain pending purge work */
 668         flush_work(&ring->persistent_purge_work);
 669
 670         if (log_stats)
 671                 print_stats(ring);
 672
 673         ring->xenblkd = NULL;
 674
 675         return 0;
 676 }
 677
 678 /*
 679  * Remove persistent grants and empty the pool of free pages
 680  */
 681 void xen_blkbk_free_caches(struct xen_blkif_ring *ring)
 682 {
 683         /* Free all persistent grant pages */
 684         if (!RB_EMPTY_ROOT(&ring->persistent_gnts))
 685                 free_persistent_gnts(ring, &ring->persistent_gnts,
 686                         ring->persistent_gnt_c);
 687
 688         BUG_ON(!RB_EMPTY_ROOT(&ring->persistent_gnts));
 689         ring->persistent_gnt_c = 0;
 690
 691         /* Since we are shutting down remove all pages from the buffer */
 692         shrink_free_pagepool(ring, 0 /* All */);
 693 }
 694
 695 static unsigned int xen_blkbk_unmap_prepare(
 696         struct xen_blkif_ring *ring,
 697         struct grant_page **pages,
 698         unsigned int num,
 699         struct gnttab_unmap_grant_ref *unmap_ops,
 700         struct page **unmap_pages)
 701 {
 702         unsigned int i, invcount = 0;
 703
 704         for (i = 0; i < num; i++) {
 705                 if (pages[i]->persistent_gnt != NULL) {
 706                         put_persistent_gnt(ring, pages[i]->persistent_gnt);
 707                         continue;
 708                 }
 709                 if (pages[i]->handle == BLKBACK_INVALID_HANDLE)
 710                         continue;
 711                 unmap_pages[invcount] = pages[i]->page;
 712                 gnttab_set_unmap_op(&unmap_ops[invcount], vaddr(pages[i]->page),
 713                                     GNTMAP_host_map, pages[i]->handle);
 714                 pages[i]->handle = BLKBACK_INVALID_HANDLE;
 715                 invcount++;
 716         }
 717
 718         return invcount;
 719 }
 720
 721 static void xen_blkbk_unmap_and_respond_callback(int result, struct gntab_unmap_queue_data *data)
 722 {
 723         struct pending_req *pending_req = (struct pending_req *)(data->data);
 724         struct xen_blkif_ring *ring = pending_req->ring;
 725         struct xen_blkif *blkif = ring->blkif;
 726
 727         /* BUG_ON used to reproduce existing behaviour,
 728            but is this the best way to deal with this? */
 729         BUG_ON(result);
 730
 731         put_free_pages(ring, data->pages, data->count);
 732         make_response(ring, pending_req->id,
 733                       pending_req->operation, pending_req->status);
 734         free_req(ring, pending_req);
 735         /*
 736          * Make sure the request is freed before releasing blkif,
 737          * or there could be a race between free_req and the
 738          * cleanup done in xen_blkif_free during shutdown.
 739          *
 740          * NB: The fact that we might try to wake up pending_free_wq
 741          * before drain_complete (in case there's a drain going on)
 742          * it's not a problem with our current implementation
 743          * because we can assure there's no thread waiting on
 744          * pending_free_wq if there's a drain going on, but it has
 745          * to be taken into account if the current model is changed.
 746          */
 747         if (atomic_dec_and_test(&ring->inflight) && atomic_read(&blkif->drain)) {
 748                 complete(&blkif->drain_complete);
 749         }
 750         xen_blkif_put(blkif);
 751 }
 752
 753 static void xen_blkbk_unmap_and_respond(struct pending_req *req)
 754 {
 755         struct gntab_unmap_queue_data* work = &req->gnttab_unmap_data;
 756         struct xen_blkif_ring *ring = req->ring;
 757         struct grant_page **pages = req->segments;
 758         unsigned int invcount;
 759
 760         invcount = xen_blkbk_unmap_prepare(ring, pages, req->nr_segs,
 761                                            req->unmap, req->unmap_pages);
 762
 763         work->data = req;
 764         work->done = xen_blkbk_unmap_and_respond_callback;
 765         work->unmap_ops = req->unmap;
 766         work->kunmap_ops = NULL;
 767         work->pages = req->unmap_pages;
 768         work->count = invcount;
 769
 770         gnttab_unmap_refs_async(&req->gnttab_unmap_data);
 771 }
 772
 773
 774 /*
 775  * Unmap the grant references.
 776  *
 777  * This could accumulate ops up to the batch size to reduce the number
 778  * of hypercalls, but since this is only used in error paths there's
 779  * no real need.
 780  */
 781 static void xen_blkbk_unmap(struct xen_blkif_ring *ring,
 782                             struct grant_page *pages[],
 783                             int num)
 784 {
 785         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 786         struct page *unmap_pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 787         unsigned int invcount = 0;
 788         int ret;
 789
 790         while (num) {
 791                 unsigned int batch = min(num, BLKIF_MAX_SEGMENTS_PER_REQUEST);
 792
 793                 invcount = xen_blkbk_unmap_prepare(ring, pages, batch,
 794                                                    unmap, unmap_pages);
 795                 if (invcount) {
 796                         ret = gnttab_unmap_refs(unmap, NULL, unmap_pages, invcount);
 797                         BUG_ON(ret);
 798                         put_free_pages(ring, unmap_pages, invcount);
 799                 }
 800                 pages += batch;
 801                 num -= batch;
 802         }
 803 }
 804
 805 static int xen_blkbk_map(struct xen_blkif_ring *ring,
 806                          struct grant_page *pages[],
 807                          int num, bool ro)
 808 {
 809         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 810         struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 811         struct persistent_gnt *persistent_gnt = NULL;
 812         phys_addr_t addr = 0;
 813         int i, seg_idx, new_map_idx;
 814         int segs_to_map = 0;
 815         int ret = 0;
 816         int last_map = 0, map_until = 0;
 817         int use_persistent_gnts;
 818         struct xen_blkif *blkif = ring->blkif;
 819
 820         use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
 821
 822         /*
 823          * Fill out preq.nr_sects with proper amount of sectors, and setup
 824          * assign map[..] with the PFN of the page in our domain with the
 825          * corresponding grant reference for each page.
 826          */
 827 again:
 828         for (i = map_until; i < num; i++) {
 829                 uint32_t flags;
 830
 831                 if (use_persistent_gnts) {
 832                         persistent_gnt = get_persistent_gnt(
 833                                 ring,
 834                                 pages[i]->gref);
 835                 }
 836
 837                 if (persistent_gnt) {
 838                         /*
 839                          * We are using persistent grants and
 840                          * the grant is already mapped
 841                          */
 842                         pages[i]->page = persistent_gnt->page;
 843                         pages[i]->persistent_gnt = persistent_gnt;
 844                 } else {
 845                         if (get_free_page(ring, &pages[i]->page))
 846                                 goto out_of_memory;
 847                         addr = vaddr(pages[i]->page);
 848                         pages_to_gnt[segs_to_map] = pages[i]->page;
 849                         pages[i]->persistent_gnt = NULL;
 850                         flags = GNTMAP_host_map;
 851                         if (!use_persistent_gnts && ro)
 852                                 flags |= GNTMAP_readonly;
 853                         gnttab_set_map_op(&map[segs_to_map++], addr,
 854                                           flags, pages[i]->gref,
 855                                           blkif->domid);
 856                 }
 857                 map_until = i + 1;
 858                 if (segs_to_map == BLKIF_MAX_SEGMENTS_PER_REQUEST)
 859                         break;
 860         }
 861
 862         if (segs_to_map) {
 863                 ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
 864                 BUG_ON(ret);
 865         }
 866
 867         /*
 868          * Now swizzle the MFN in our domain with the MFN from the other domain
 869          * so that when we access vaddr(pending_req,i) it has the contents of
 870          * the page from the other domain.
 871          */
 872         for (seg_idx = last_map, new_map_idx = 0; seg_idx < map_until; seg_idx++) {
 873                 if (!pages[seg_idx]->persistent_gnt) {
 874                         /* This is a newly mapped grant */
 875                         BUG_ON(new_map_idx >= segs_to_map);
 876                         if (unlikely(map[new_map_idx].status != 0)) {
 877                                 pr_debug("invalid buffer -- could not remap it\n");
 878                                 put_free_pages(ring, &pages[seg_idx]->page, 1);
 879                                 pages[seg_idx]->handle = BLKBACK_INVALID_HANDLE;
 880                                 ret |= 1;
 881                                 goto next;
 882                         }
 883                         pages[seg_idx]->handle = map[new_map_idx].handle;
 884                 } else {
 885                         continue;
 886                 }
 887                 if (use_persistent_gnts &&
 888                     ring->persistent_gnt_c < xen_blkif_max_pgrants) {
 889                         /*
 890                          * We are using persistent grants, the grant is
 891                          * not mapped but we might have room for it.
 892                          */
 893                         persistent_gnt = kmalloc(sizeof(struct persistent_gnt),
 894                                                  GFP_KERNEL);
 895                         if (!persistent_gnt) {
 896                                 /*
 897                                  * If we don't have enough memory to
 898                                  * allocate the persistent_gnt struct
 899                                  * map this grant non-persistenly
 900                                  */
 901                                 goto next;
 902                         }
 903                         persistent_gnt->gnt = map[new_map_idx].ref;
 904                         persistent_gnt->handle = map[new_map_idx].handle;
 905                         persistent_gnt->page = pages[seg_idx]->page;
 906                         if (add_persistent_gnt(ring,
 907                                                persistent_gnt)) {
 908                                 kfree(persistent_gnt);
 909                                 persistent_gnt = NULL;
 910                                 goto next;
 911                         }
 912                         pages[seg_idx]->persistent_gnt = persistent_gnt;
 913                         pr_debug("grant %u added to the tree of persistent grants, using %u/%u\n",
 914                                  persistent_gnt->gnt, ring->persistent_gnt_c,
 915                                  xen_blkif_max_pgrants);
 916                         goto next;
 917                 }
 918                 if (use_persistent_gnts && !blkif->vbd.overflow_max_grants) {
 919                         blkif->vbd.overflow_max_grants = 1;
 920                         pr_debug("domain %u, device %#x is using maximum number of persistent grants\n",
 921                                  blkif->domid, blkif->vbd.handle);
 922                 }
 923                 /*
 924                  * We could not map this grant persistently, so use it as
 925                  * a non-persistent grant.
 926                  */
 927 next:
 928                 new_map_idx++;
 929         }
 930         segs_to_map = 0;
 931         last_map = map_until;
 932         if (map_until != num)
 933                 goto again;
 934
 935         return ret;
 936
 937 out_of_memory:
 938         pr_alert("%s: out of memory\n", __func__);
 939         put_free_pages(ring, pages_to_gnt, segs_to_map);
 940         return -ENOMEM;
 941 }
 942
 943 static int xen_blkbk_map_seg(struct pending_req *pending_req)
 944 {
 945         int rc;
 946
 947         rc = xen_blkbk_map(pending_req->ring, pending_req->segments,
 948                            pending_req->nr_segs,
 949                            (pending_req->operation != BLKIF_OP_READ));
 950
 951         return rc;
 952 }
 953
 954 static int xen_blkbk_parse_indirect(struct blkif_request *req,
 955                                     struct pending_req *pending_req,
 956                                     struct seg_buf seg[],
 957                                     struct phys_req *preq)
 958 {
 959         struct grant_page **pages = pending_req->indirect_pages;
 960         struct xen_blkif_ring *ring = pending_req->ring;
 961         int indirect_grefs, rc, n, nseg, i;
 962         struct blkif_request_segment *segments = NULL;
 963
 964         nseg = pending_req->nr_segs;
 965         indirect_grefs = INDIRECT_PAGES(nseg);
 966         BUG_ON(indirect_grefs > BLKIF_MAX_INDIRECT_PAGES_PER_REQUEST);
 967
 968         for (i = 0; i < indirect_grefs; i++)
 969                 pages[i]->gref = req->u.indirect.indirect_grefs[i];
 970
 971         rc = xen_blkbk_map(ring, pages, indirect_grefs, true);
 972         if (rc)
 973                 goto unmap;
 974
 975         for (n = 0, i = 0; n < nseg; n++) {
 976                 uint8_t first_sect, last_sect;
 977
 978                 if ((n % SEGS_PER_INDIRECT_FRAME) == 0) {
 979                         /* Map indirect segments */
 980                         if (segments)
 981                                 kunmap_atomic(segments);
 982                         segments = kmap_atomic(pages[n/SEGS_PER_INDIRECT_FRAME]->page);
 983                 }
 984                 i = n % SEGS_PER_INDIRECT_FRAME;
 985
 986                 pending_req->segments[n]->gref = segments[i].gref;
 987
 988                 first_sect = READ_ONCE(segments[i].first_sect);
 989                 last_sect = READ_ONCE(segments[i].last_sect);
 990                 if (last_sect >= (XEN_PAGE_SIZE >> 9) || last_sect < first_sect) {
 991                         rc = -EINVAL;
 992                         goto unmap;
 993                 }
 994
 995                 seg[n].nsec = last_sect - first_sect + 1;
 996                 seg[n].offset = first_sect << 9;
 997                 preq->nr_sects += seg[n].nsec;
 998         }
 999
1000 unmap:
1001         if (segments)
1002                 kunmap_atomic(segments);
1003         xen_blkbk_unmap(ring, pages, indirect_grefs);
1004         return rc;
1005 }
1006
1007 static int dispatch_discard_io(struct xen_blkif_ring *ring,
1008                                 struct blkif_request *req)
1009 {
1010         int err = 0;
1011         int status = BLKIF_RSP_OKAY;
1012         struct xen_blkif *blkif = ring->blkif;
1013         struct block_device *bdev = blkif->vbd.bdev;
1014         unsigned long secure;
1015         struct phys_req preq;
1016
1017         xen_blkif_get(blkif);
1018
1019         preq.sector_number = req->u.discard.sector_number;
1020         preq.nr_sects      = req->u.discard.nr_sectors;
1021
1022         err = xen_vbd_translate(&preq, blkif, REQ_OP_WRITE);
1023         if (err) {
1024                 pr_warn("access denied: DISCARD [%llu->%llu] on dev=%04x\n",
1025                         preq.sector_number,
1026                         preq.sector_number + preq.nr_sects, blkif->vbd.pdevice);
1027                 goto fail_response;
1028         }
1029         ring->st_ds_req++;
1030
1031         secure = (blkif->vbd.discard_secure &&
1032                  (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
1033                  BLKDEV_DISCARD_SECURE : 0;
1034
1035         err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
1036                                    req->u.discard.nr_sectors,
1037                                    GFP_KERNEL, secure);
1038 fail_response:
1039         if (err == -EOPNOTSUPP) {
1040                 pr_debug("discard op failed, not supported\n");
1041                 status = BLKIF_RSP_EOPNOTSUPP;
1042         } else if (err)
1043                 status = BLKIF_RSP_ERROR;
1044
1045         make_response(ring, req->u.discard.id, req->operation, status);
1046         xen_blkif_put(blkif);
1047         return err;
1048 }
1049
1050 static int dispatch_other_io(struct xen_blkif_ring *ring,
1051                              struct blkif_request *req,
1052                              struct pending_req *pending_req)
1053 {
1054         free_req(ring, pending_req);
1055         make_response(ring, req->u.other.id, req->operation,
1056                       BLKIF_RSP_EOPNOTSUPP);
1057         return -EIO;
1058 }
1059
1060 static void xen_blk_drain_io(struct xen_blkif_ring *ring)
1061 {
1062         struct xen_blkif *blkif = ring->blkif;
1063
1064         atomic_set(&blkif->drain, 1);
1065         do {
1066                 if (atomic_read(&ring->inflight) == 0)
1067                         break;
1068                 wait_for_completion_interruptible_timeout(
1069                                 &blkif->drain_complete, HZ);
1070
1071                 if (!atomic_read(&blkif->drain))
1072                         break;
1073         } while (!kthread_should_stop());
1074         atomic_set(&blkif->drain, 0);
1075 }
1076
1077 static void __end_block_io_op(struct pending_req *pending_req,
1078                 blk_status_t error)
1079 {
1080         /* An error fails the entire request. */
1081         if (pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE &&
1082             error == BLK_STS_NOTSUPP) {
1083                 pr_debug("flush diskcache op failed, not supported\n");
1084                 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->ring->blkif->be, 0);
1085                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1086         } else if (pending_req->operation == BLKIF_OP_WRITE_BARRIER &&
1087                    error == BLK_STS_NOTSUPP) {
1088                 pr_debug("write barrier op failed, not supported\n");
1089                 xen_blkbk_barrier(XBT_NIL, pending_req->ring->blkif->be, 0);
1090                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1091         } else if (error) {
1092                 pr_debug("Buffer not up-to-date at end of operation,"
1093                          " error=%d\n", error);
1094                 pending_req->status = BLKIF_RSP_ERROR;
1095         }
1096
1097         /*
1098          * If all of the bio's have completed it is time to unmap
1099          * the grant references associated with 'request' and provide
1100          * the proper response on the ring.
1101          */
1102         if (atomic_dec_and_test(&pending_req->pendcnt))
1103                 xen_blkbk_unmap_and_respond(pending_req);
1104 }
1105
1106 /*
1107  * bio callback.
1108  */
1109 static void end_block_io_op(struct bio *bio)
1110 {
1111         __end_block_io_op(bio->bi_private, bio->bi_status);
1112         bio_put(bio);
1113 }
1114
1115
1116
1117 /*
1118  * Function to copy the from the ring buffer the 'struct blkif_request'
1119  * (which has the sectors we want, number of them, grant references, etc),
1120  * and transmute  it to the block API to hand it over to the proper block disk.
1121  */
1122 static int
1123 __do_block_io_op(struct xen_blkif_ring *ring)
1124 {
1125         union blkif_back_rings *blk_rings = &ring->blk_rings;
1126         struct blkif_request req;
1127         struct pending_req *pending_req;
1128         RING_IDX rc, rp;
1129         int more_to_do = 0;
1130
1131         rc = blk_rings->common.req_cons;
1132         rp = blk_rings->common.sring->req_prod;
1133         rmb(); /* Ensure we see queued requests up to 'rp'. */
1134
1135         if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
1136                 rc = blk_rings->common.rsp_prod_pvt;
1137                 pr_warn("Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
1138                         rp, rc, rp - rc, ring->blkif->vbd.pdevice);
1139                 return -EACCES;
1140         }
1141         while (rc != rp) {
1142
1143                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
1144                         break;
1145
1146                 if (kthread_should_stop()) {
1147                         more_to_do = 1;
1148                         break;
1149                 }
1150
1151                 pending_req = alloc_req(ring);
1152                 if (NULL == pending_req) {
1153                         ring->st_oo_req++;
1154                         more_to_do = 1;
1155                         break;
1156                 }
1157
1158                 switch (ring->blkif->blk_protocol) {
1159                 case BLKIF_PROTOCOL_NATIVE:
1160                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
1161                         break;
1162                 case BLKIF_PROTOCOL_X86_32:
1163                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
1164                         break;
1165                 case BLKIF_PROTOCOL_X86_64:
1166                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
1167                         break;
1168                 default:
1169                         BUG();
1170                 }
1171                 blk_rings->common.req_cons = ++rc; /* before make_response() */
1172
1173                 /* Apply all sanity checks to /private copy/ of request. */
1174                 barrier();
1175
1176                 switch (req.operation) {
1177                 case BLKIF_OP_READ:
1178                 case BLKIF_OP_WRITE:
1179                 case BLKIF_OP_WRITE_BARRIER:
1180                 case BLKIF_OP_FLUSH_DISKCACHE:
1181                 case BLKIF_OP_INDIRECT:
1182                         if (dispatch_rw_block_io(ring, &req, pending_req))
1183                                 goto done;
1184                         break;
1185                 case BLKIF_OP_DISCARD:
1186                         free_req(ring, pending_req);
1187                         if (dispatch_discard_io(ring, &req))
1188                                 goto done;
1189                         break;
1190                 default:
1191                         if (dispatch_other_io(ring, &req, pending_req))
1192                                 goto done;
1193                         break;
1194                 }
1195
1196                 /* Yield point for this unbounded loop. */
1197                 cond_resched();
1198         }
1199 done:
1200         return more_to_do;
1201 }
1202
1203 static int
1204 do_block_io_op(struct xen_blkif_ring *ring)
1205 {
1206         union blkif_back_rings *blk_rings = &ring->blk_rings;
1207         int more_to_do;
1208
1209         do {
1210                 more_to_do = __do_block_io_op(ring);
1211                 if (more_to_do)
1212                         break;
1213
1214                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
1215         } while (more_to_do);
1216
1217         return more_to_do;
1218 }
1219 /*
1220  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
1221  * and call the 'submit_bio' to pass it to the underlying storage.
1222  */
1223 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
1224                                 struct blkif_request *req,
1225                                 struct pending_req *pending_req)
1226 {
1227         struct phys_req preq;
1228         struct seg_buf *seg = pending_req->seg;
1229         unsigned int nseg;
1230         struct bio *bio = NULL;
1231         struct bio **biolist = pending_req->biolist;
1232         int i, nbio = 0;
1233         int operation;
1234         int operation_flags = 0;
1235         struct blk_plug plug;
1236         bool drain = false;
1237         struct grant_page **pages = pending_req->segments;
1238         unsigned short req_operation;
1239
1240         req_operation = req->operation == BLKIF_OP_INDIRECT ?
1241                         req->u.indirect.indirect_op : req->operation;
1242
1243         if ((req->operation == BLKIF_OP_INDIRECT) &&
1244             (req_operation != BLKIF_OP_READ) &&
1245             (req_operation != BLKIF_OP_WRITE)) {
1246                 pr_debug("Invalid indirect operation (%u)\n", req_operation);
1247                 goto fail_response;
1248         }
1249
1250         switch (req_operation) {
1251         case BLKIF_OP_READ:
1252                 ring->st_rd_req++;
1253                 operation = REQ_OP_READ;
1254                 break;
1255         case BLKIF_OP_WRITE:
1256                 ring->st_wr_req++;
1257                 operation = REQ_OP_WRITE;
1258                 operation_flags = REQ_SYNC | REQ_IDLE;
1259                 break;
1260         case BLKIF_OP_WRITE_BARRIER:
1261                 drain = true;
1262                 /* fall through */
1263         case BLKIF_OP_FLUSH_DISKCACHE:
1264                 ring->st_f_req++;
1265                 operation = REQ_OP_WRITE;
1266                 operation_flags = REQ_PREFLUSH;
1267                 break;
1268         default:
1269                 operation = 0; /* make gcc happy */
1270                 goto fail_response;
1271                 break;
1272         }
1273
1274         /* Check that the number of segments is sane. */
1275         nseg = req->operation == BLKIF_OP_INDIRECT ?
1276                req->u.indirect.nr_segments : req->u.rw.nr_segments;
1277
1278         if (unlikely(nseg == 0 && operation_flags != REQ_PREFLUSH) ||
1279             unlikely((req->operation != BLKIF_OP_INDIRECT) &&
1280                      (nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) ||
1281             unlikely((req->operation == BLKIF_OP_INDIRECT) &&
1282                      (nseg > MAX_INDIRECT_SEGMENTS))) {
1283                 pr_debug("Bad number of segments in request (%d)\n", nseg);
1284                 /* Haven't submitted any bio's yet. */
1285                 goto fail_response;
1286         }
1287
1288         preq.nr_sects      = 0;
1289
1290         pending_req->ring      = ring;
1291         pending_req->id        = req->u.rw.id;
1292         pending_req->operation = req_operation;
1293         pending_req->status    = BLKIF_RSP_OKAY;
1294         pending_req->nr_segs   = nseg;
1295
1296         if (req->operation != BLKIF_OP_INDIRECT) {
1297                 preq.dev               = req->u.rw.handle;
1298                 preq.sector_number     = req->u.rw.sector_number;
1299                 for (i = 0; i < nseg; i++) {
1300                         pages[i]->gref = req->u.rw.seg[i].gref;
1301                         seg[i].nsec = req->u.rw.seg[i].last_sect -
1302                                 req->u.rw.seg[i].first_sect + 1;
1303                         seg[i].offset = (req->u.rw.seg[i].first_sect << 9);
1304                         if ((req->u.rw.seg[i].last_sect >= (XEN_PAGE_SIZE >> 9)) ||
1305                             (req->u.rw.seg[i].last_sect <
1306                              req->u.rw.seg[i].first_sect))
1307                                 goto fail_response;
1308                         preq.nr_sects += seg[i].nsec;
1309                 }
1310         } else {
1311                 preq.dev               = req->u.indirect.handle;
1312                 preq.sector_number     = req->u.indirect.sector_number;
1313                 if (xen_blkbk_parse_indirect(req, pending_req, seg, &preq))
1314                         goto fail_response;
1315         }
1316
1317         if (xen_vbd_translate(&preq, ring->blkif, operation) != 0) {
1318                 pr_debug("access denied: %s of [%llu,%llu] on dev=%04x\n",
1319                          operation == REQ_OP_READ ? "read" : "write",
1320                          preq.sector_number,
1321                          preq.sector_number + preq.nr_sects,
1322                          ring->blkif->vbd.pdevice);
1323                 goto fail_response;
1324         }
1325
1326         /*
1327          * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
1328          * is set there.
1329          */
1330         for (i = 0; i < nseg; i++) {
1331                 if (((int)preq.sector_number|(int)seg[i].nsec) &
1332                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
1333                         pr_debug("Misaligned I/O request from domain %d\n",
1334                                  ring->blkif->domid);
1335                         goto fail_response;
1336                 }
1337         }
1338
1339         /* Wait on all outstanding I/O's and once that has been completed
1340          * issue the flush.
1341          */
1342         if (drain)
1343                 xen_blk_drain_io(pending_req->ring);
1344
1345         /*
1346          * If we have failed at this point, we need to undo the M2P override,
1347          * set gnttab_set_unmap_op on all of the grant references and perform
1348          * the hypercall to unmap the grants - that is all done in
1349          * xen_blkbk_unmap.
1350          */
1351         if (xen_blkbk_map_seg(pending_req))
1352                 goto fail_flush;
1353
1354         /*
1355          * This corresponding xen_blkif_put is done in __end_block_io_op, or
1356          * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
1357          */
1358         xen_blkif_get(ring->blkif);
1359         atomic_inc(&ring->inflight);
1360
1361         for (i = 0; i < nseg; i++) {
1362                 while ((bio == NULL) ||
1363                        (bio_add_page(bio,
1364                                      pages[i]->page,
1365                                      seg[i].nsec << 9,
1366                                      seg[i].offset) == 0)) {
1367
1368                         int nr_iovecs = min_t(int, (nseg-i), BIO_MAX_PAGES);
1369                         bio = bio_alloc(GFP_KERNEL, nr_iovecs);
1370                         if (unlikely(bio == NULL))
1371                                 goto fail_put_bio;
1372
1373                         biolist[nbio++] = bio;
1374                         bio_set_dev(bio, preq.bdev);
1375                         bio->bi_private = pending_req;
1376                         bio->bi_end_io  = end_block_io_op;
1377                         bio->bi_iter.bi_sector  = preq.sector_number;
1378                         bio_set_op_attrs(bio, operation, operation_flags);
1379                 }
1380
1381                 preq.sector_number += seg[i].nsec;
1382         }
1383
1384         /* This will be hit if the operation was a flush or discard. */
1385         if (!bio) {
1386                 BUG_ON(operation_flags != REQ_PREFLUSH);
1387
1388                 bio = bio_alloc(GFP_KERNEL, 0);
1389                 if (unlikely(bio == NULL))
1390                         goto fail_put_bio;
1391
1392                 biolist[nbio++] = bio;
1393                 bio_set_dev(bio, preq.bdev);
1394                 bio->bi_private = pending_req;
1395                 bio->bi_end_io  = end_block_io_op;
1396                 bio_set_op_attrs(bio, operation, operation_flags);
1397         }
1398
1399         atomic_set(&pending_req->pendcnt, nbio);
1400         blk_start_plug(&plug);
1401
1402         for (i = 0; i < nbio; i++)
1403                 submit_bio(biolist[i]);
1404
1405         /* Let the I/Os go.. */
1406         blk_finish_plug(&plug);
1407
1408         if (operation == REQ_OP_READ)
1409                 ring->st_rd_sect += preq.nr_sects;
1410         else if (operation == REQ_OP_WRITE)
1411                 ring->st_wr_sect += preq.nr_sects;
1412
1413         return 0;
1414
1415  fail_flush:
1416         xen_blkbk_unmap(ring, pending_req->segments,
1417                         pending_req->nr_segs);
1418  fail_response:
1419         /* Haven't submitted any bio's yet. */
1420         make_response(ring, req->u.rw.id, req_operation, BLKIF_RSP_ERROR);
1421         free_req(ring, pending_req);
1422         msleep(1); /* back off a bit */
1423         return -EIO;
1424
1425  fail_put_bio:
1426         for (i = 0; i < nbio; i++)
1427                 bio_put(biolist[i]);
1428         atomic_set(&pending_req->pendcnt, 1);
1429         __end_block_io_op(pending_req, BLK_STS_RESOURCE);
1430         msleep(1); /* back off a bit */
1431         return -EIO;
1432 }
1433
1434
1435
1436 /*
1437  * Put a response on the ring on how the operation fared.
1438  */
1439 static void make_response(struct xen_blkif_ring *ring, u64 id,
1440                           unsigned short op, int st)
1441 {
1442         struct blkif_response *resp;
1443         unsigned long     flags;
1444         union blkif_back_rings *blk_rings;
1445         int notify;
1446
1447         spin_lock_irqsave(&ring->blk_ring_lock, flags);
1448         blk_rings = &ring->blk_rings;
1449         /* Place on the response ring for the relevant domain. */
1450         switch (ring->blkif->blk_protocol) {
1451         case BLKIF_PROTOCOL_NATIVE:
1452                 resp = RING_GET_RESPONSE(&blk_rings->native,
1453                                          blk_rings->native.rsp_prod_pvt);
1454                 break;
1455         case BLKIF_PROTOCOL_X86_32:
1456                 resp = RING_GET_RESPONSE(&blk_rings->x86_32,
1457                                          blk_rings->x86_32.rsp_prod_pvt);
1458                 break;
1459         case BLKIF_PROTOCOL_X86_64:
1460                 resp = RING_GET_RESPONSE(&blk_rings->x86_64,
1461                                          blk_rings->x86_64.rsp_prod_pvt);
1462                 break;
1463         default:
1464                 BUG();
1465         }
1466
1467         resp->id        = id;
1468         resp->operation = op;
1469         resp->status    = st;
1470
1471         blk_rings->common.rsp_prod_pvt++;
1472         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1473         spin_unlock_irqrestore(&ring->blk_ring_lock, flags);
1474         if (notify)
1475                 notify_remote_via_irq(ring->irq);
1476 }
1477
1478 static int __init xen_blkif_init(void)
1479 {
1480         int rc = 0;
1481
1482         if (!xen_domain())
1483                 return -ENODEV;
1484
1485         if (xen_blkif_max_ring_order > XENBUS_MAX_RING_GRANT_ORDER) {
1486                 pr_info("Invalid max_ring_order (%d), will use default max: %d.\n",
1487                         xen_blkif_max_ring_order, XENBUS_MAX_RING_GRANT_ORDER);
1488                 xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
1489         }
1490
1491         if (xenblk_max_queues == 0)
1492                 xenblk_max_queues = num_online_cpus();
1493
1494         rc = xen_blkif_interface_init();
1495         if (rc)
1496                 goto failed_init;
1497
1498         rc = xen_blkif_xenbus_init();
1499         if (rc)
1500                 goto failed_init;
1501
1502  failed_init:
1503         return rc;
1504 }
1505
1506 module_init(xen_blkif_init);
1507
1508 MODULE_LICENSE("Dual BSD/GPL");
1509 MODULE_ALIAS("xen-backend:vbd");