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