drivers/block/xen-blkback/blkback.c

   1 /******************************************************************************
   2  *
   3  * Back-end of the driver for virtual block devices. This portion of the
   4  * driver exports a 'unified' block-device interface that can be accessed
   5  * by any operating system that implements a compatible front end. A
   6  * reference front-end implementation can be found in:
   7  *  drivers/block/xen-blkfront.c
   8  *
   9  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
  10  * Copyright (c) 2005, Christopher Clark
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License version 2
  14  * as published by the Free Software Foundation; or, when distributed
  15  * separately from the Linux kernel or incorporated into other
  16  * software packages, subject to the following license:
  17  *
  18  * Permission is hereby granted, free of charge, to any person obtaining a copy
  19  * of this source file (the "Software"), to deal in the Software without
  20  * restriction, including without limitation the rights to use, copy, modify,
  21  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  22  * and to permit persons to whom the Software is furnished to do so, subject to
  23  * the following conditions:
  24  *
  25  * The above copyright notice and this permission notice shall be included in
  26  * all copies or substantial portions of the Software.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  29  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  30  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  31  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  32  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  33  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  34  * IN THE SOFTWARE.
  35  */
  36
  37 #include <linux/spinlock.h>
  38 #include <linux/kthread.h>
  39 #include <linux/list.h>
  40 #include <linux/delay.h>
  41 #include <linux/freezer.h>
  42 #include <linux/bitmap.h>
  43
  44 #include <xen/events.h>
  45 #include <xen/page.h>
  46 #include <xen/xen.h>
  47 #include <asm/xen/hypervisor.h>
  48 #include <asm/xen/hypercall.h>
  49 #include "common.h"
  50
  51 /*
  52  * These are rather arbitrary. They are fairly large because adjacent requests
  53  * pulled from a communication ring are quite likely to end up being part of
  54  * the same scatter/gather request at the disc.
  55  *
  56  * ** TRY INCREASING 'xen_blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
  57  *
  58  * This will increase the chances of being able to write whole tracks.
  59  * 64 should be enough to keep us competitive with Linux.
  60  */
  61 static int xen_blkif_reqs = 64;
  62 module_param_named(reqs, xen_blkif_reqs, int, 0);
  63 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
  64
  65 /* Run-time switchable: /sys/module/blkback/parameters/ */
  66 static unsigned int log_stats;
  67 module_param(log_stats, int, 0644);
  68
  69 /*
  70  * Each outstanding request that we've passed to the lower device layers has a
  71  * 'pending_req' allocated to it. Each buffer_head that completes decrements
  72  * the pendcnt towards zero. When it hits zero, the specified domain has a
  73  * response queued for it, with the saved 'id' passed back.
  74  */
  75 struct pending_req {
  76         struct xen_blkif        *blkif;
  77         u64                     id;
  78         int                     nr_pages;
  79         atomic_t                pendcnt;
  80         unsigned short          operation;
  81         int                     status;
  82         struct list_head        free_list;
  83         DECLARE_BITMAP(unmap_seg, BLKIF_MAX_SEGMENTS_PER_REQUEST);
  84 };
  85
  86 #define BLKBACK_INVALID_HANDLE (~0)
  87
  88 struct xen_blkbk {
  89         struct pending_req      *pending_reqs;
  90         /* List of all 'pending_req' available */
  91         struct list_head        pending_free;
  92         /* And its spinlock. */
  93         spinlock_t              pending_free_lock;
  94         wait_queue_head_t       pending_free_wq;
  95         /* The list of all pages that are available. */
  96         struct page             **pending_pages;
  97         /* And the grant handles that are available. */
  98         grant_handle_t          *pending_grant_handles;
  99 };
 100
 101 static struct xen_blkbk *blkbk;
 102
 103 /*
 104  * Maximum number of grant pages that can be mapped in blkback.
 105  * BLKIF_MAX_SEGMENTS_PER_REQUEST * RING_SIZE is the maximum number of
 106  * pages that blkback will persistently map.
 107  * Currently, this is:
 108  * RING_SIZE = 32 (for all known ring types)
 109  * BLKIF_MAX_SEGMENTS_PER_REQUEST = 11
 110  * sizeof(struct persistent_gnt) = 48
 111  * So the maximum memory used to store the grants is:
 112  * 32 * 11 * 48 = 16896 bytes
 113  */
 114 static inline unsigned int max_mapped_grant_pages(enum blkif_protocol protocol)
 115 {
 116         switch (protocol) {
 117         case BLKIF_PROTOCOL_NATIVE:
 118                 return __CONST_RING_SIZE(blkif, PAGE_SIZE) *
 119                            BLKIF_MAX_SEGMENTS_PER_REQUEST;
 120         case BLKIF_PROTOCOL_X86_32:
 121                 return __CONST_RING_SIZE(blkif_x86_32, PAGE_SIZE) *
 122                            BLKIF_MAX_SEGMENTS_PER_REQUEST;
 123         case BLKIF_PROTOCOL_X86_64:
 124                 return __CONST_RING_SIZE(blkif_x86_64, PAGE_SIZE) *
 125                            BLKIF_MAX_SEGMENTS_PER_REQUEST;
 126         default:
 127                 BUG();
 128         }
 129         return 0;
 130 }
 131
 132
 133 /*
 134  * Little helpful macro to figure out the index and virtual address of the
 135  * pending_pages[..]. For each 'pending_req' we have have up to
 136  * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
 137  * 10 and would index in the pending_pages[..].
 138  */
 139 static inline int vaddr_pagenr(struct pending_req *req, int seg)
 140 {
 141         return (req - blkbk->pending_reqs) *
 142                 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
 143 }
 144
 145 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
 146
 147 static inline unsigned long vaddr(struct pending_req *req, int seg)
 148 {
 149         unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
 150         return (unsigned long)pfn_to_kaddr(pfn);
 151 }
 152
 153 #define pending_handle(_req, _seg) \
 154         (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
 155
 156
 157 static int do_block_io_op(struct xen_blkif *blkif);
 158 static int dispatch_rw_block_io(struct xen_blkif *blkif,
 159                                 struct blkif_request *req,
 160                                 struct pending_req *pending_req);
 161 static void make_response(struct xen_blkif *blkif, u64 id,
 162                           unsigned short op, int st);
 163
 164 #define foreach_grant_safe(pos, n, rbtree, node) \
 165         for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node), \
 166              (n) = rb_next(&(pos)->node); \
 167              &(pos)->node != NULL; \
 168              (pos) = container_of(n, typeof(*(pos)), node), \
 169              (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL)
 170
 171
 172 static void add_persistent_gnt(struct rb_root *root,
 173                                struct persistent_gnt *persistent_gnt)
 174 {
 175         struct rb_node **new = &(root->rb_node), *parent = NULL;
 176         struct persistent_gnt *this;
 177
 178         /* Figure out where to put new node */
 179         while (*new) {
 180                 this = container_of(*new, struct persistent_gnt, node);
 181
 182                 parent = *new;
 183                 if (persistent_gnt->gnt < this->gnt)
 184                         new = &((*new)->rb_left);
 185                 else if (persistent_gnt->gnt > this->gnt)
 186                         new = &((*new)->rb_right);
 187                 else {
 188                         pr_alert(DRV_PFX " trying to add a gref that's already in the tree\n");
 189                         BUG();
 190                 }
 191         }
 192
 193         /* Add new node and rebalance tree. */
 194         rb_link_node(&(persistent_gnt->node), parent, new);
 195         rb_insert_color(&(persistent_gnt->node), root);
 196 }
 197
 198 static struct persistent_gnt *get_persistent_gnt(struct rb_root *root,
 199                                                  grant_ref_t gref)
 200 {
 201         struct persistent_gnt *data;
 202         struct rb_node *node = root->rb_node;
 203
 204         while (node) {
 205                 data = container_of(node, struct persistent_gnt, node);
 206
 207                 if (gref < data->gnt)
 208                         node = node->rb_left;
 209                 else if (gref > data->gnt)
 210                         node = node->rb_right;
 211                 else
 212                         return data;
 213         }
 214         return NULL;
 215 }
 216
 217 static void free_persistent_gnts(struct rb_root *root, unsigned int num)
 218 {
 219         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 220         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 221         struct persistent_gnt *persistent_gnt;
 222         struct rb_node *n;
 223         int ret = 0;
 224         int segs_to_unmap = 0;
 225
 226         foreach_grant_safe(persistent_gnt, n, root, node) {
 227                 BUG_ON(persistent_gnt->handle ==
 228                         BLKBACK_INVALID_HANDLE);
 229                 gnttab_set_unmap_op(&unmap[segs_to_unmap],
 230                         (unsigned long) pfn_to_kaddr(page_to_pfn(
 231                                 persistent_gnt->page)),
 232                         GNTMAP_host_map,
 233                         persistent_gnt->handle);
 234
 235                 pages[segs_to_unmap] = persistent_gnt->page;
 236
 237                 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
 238                         !rb_next(&persistent_gnt->node)) {
 239                         ret = gnttab_unmap_refs(unmap, NULL, pages,
 240                                 segs_to_unmap);
 241                         BUG_ON(ret);
 242                         segs_to_unmap = 0;
 243                 }
 244
 245                 rb_erase(&persistent_gnt->node, root);
 246                 kfree(persistent_gnt);
 247                 num--;
 248         }
 249         BUG_ON(num != 0);
 250 }
 251
 252 /*
 253  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
 254  */
 255 static struct pending_req *alloc_req(void)
 256 {
 257         struct pending_req *req = NULL;
 258         unsigned long flags;
 259
 260         spin_lock_irqsave(&blkbk->pending_free_lock, flags);
 261         if (!list_empty(&blkbk->pending_free)) {
 262                 req = list_entry(blkbk->pending_free.next, struct pending_req,
 263                                  free_list);
 264                 list_del(&req->free_list);
 265         }
 266         spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
 267         return req;
 268 }
 269
 270 /*
 271  * Return the 'pending_req' structure back to the freepool. We also
 272  * wake up the thread if it was waiting for a free page.
 273  */
 274 static void free_req(struct pending_req *req)
 275 {
 276         unsigned long flags;
 277         int was_empty;
 278
 279         spin_lock_irqsave(&blkbk->pending_free_lock, flags);
 280         was_empty = list_empty(&blkbk->pending_free);
 281         list_add(&req->free_list, &blkbk->pending_free);
 282         spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
 283         if (was_empty)
 284                 wake_up(&blkbk->pending_free_wq);
 285 }
 286
 287 /*
 288  * Routines for managing virtual block devices (vbds).
 289  */
 290 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
 291                              int operation)
 292 {
 293         struct xen_vbd *vbd = &blkif->vbd;
 294         int rc = -EACCES;
 295
 296         if ((operation != READ) && vbd->readonly)
 297                 goto out;
 298
 299         if (likely(req->nr_sects)) {
 300                 blkif_sector_t end = req->sector_number + req->nr_sects;
 301
 302                 if (unlikely(end < req->sector_number))
 303                         goto out;
 304                 if (unlikely(end > vbd_sz(vbd)))
 305                         goto out;
 306         }
 307
 308         req->dev  = vbd->pdevice;
 309         req->bdev = vbd->bdev;
 310         rc = 0;
 311
 312  out:
 313         return rc;
 314 }
 315
 316 static void xen_vbd_resize(struct xen_blkif *blkif)
 317 {
 318         struct xen_vbd *vbd = &blkif->vbd;
 319         struct xenbus_transaction xbt;
 320         int err;
 321         struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
 322         unsigned long long new_size = vbd_sz(vbd);
 323
 324         pr_info(DRV_PFX "VBD Resize: Domid: %d, Device: (%d, %d)\n",
 325                 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
 326         pr_info(DRV_PFX "VBD Resize: new size %llu\n", new_size);
 327         vbd->size = new_size;
 328 again:
 329         err = xenbus_transaction_start(&xbt);
 330         if (err) {
 331                 pr_warn(DRV_PFX "Error starting transaction");
 332                 return;
 333         }
 334         err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
 335                             (unsigned long long)vbd_sz(vbd));
 336         if (err) {
 337                 pr_warn(DRV_PFX "Error writing new size");
 338                 goto abort;
 339         }
 340         /*
 341          * Write the current state; we will use this to synchronize
 342          * the front-end. If the current state is "connected" the
 343          * front-end will get the new size information online.
 344          */
 345         err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
 346         if (err) {
 347                 pr_warn(DRV_PFX "Error writing the state");
 348                 goto abort;
 349         }
 350
 351         err = xenbus_transaction_end(xbt, 0);
 352         if (err == -EAGAIN)
 353                 goto again;
 354         if (err)
 355                 pr_warn(DRV_PFX "Error ending transaction");
 356         return;
 357 abort:
 358         xenbus_transaction_end(xbt, 1);
 359 }
 360
 361 /*
 362  * Notification from the guest OS.
 363  */
 364 static void blkif_notify_work(struct xen_blkif *blkif)
 365 {
 366         blkif->waiting_reqs = 1;
 367         wake_up(&blkif->wq);
 368 }
 369
 370 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
 371 {
 372         blkif_notify_work(dev_id);
 373         return IRQ_HANDLED;
 374 }
 375
 376 /*
 377  * SCHEDULER FUNCTIONS
 378  */
 379
 380 static void print_stats(struct xen_blkif *blkif)
 381 {
 382         pr_info("xen-blkback (%s): oo %3d  |  rd %4d  |  wr %4d  |  f %4d"
 383                  "  |  ds %4d\n",
 384                  current->comm, blkif->st_oo_req,
 385                  blkif->st_rd_req, blkif->st_wr_req,
 386                  blkif->st_f_req, blkif->st_ds_req);
 387         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 388         blkif->st_rd_req = 0;
 389         blkif->st_wr_req = 0;
 390         blkif->st_oo_req = 0;
 391         blkif->st_ds_req = 0;
 392 }
 393
 394 int xen_blkif_schedule(void *arg)
 395 {
 396         struct xen_blkif *blkif = arg;
 397         struct xen_vbd *vbd = &blkif->vbd;
 398
 399         xen_blkif_get(blkif);
 400
 401         while (!kthread_should_stop()) {
 402                 if (try_to_freeze())
 403                         continue;
 404                 if (unlikely(vbd->size != vbd_sz(vbd)))
 405                         xen_vbd_resize(blkif);
 406
 407                 wait_event_interruptible(
 408                         blkif->wq,
 409                         blkif->waiting_reqs || kthread_should_stop());
 410                 wait_event_interruptible(
 411                         blkbk->pending_free_wq,
 412                         !list_empty(&blkbk->pending_free) ||
 413                         kthread_should_stop());
 414
 415                 blkif->waiting_reqs = 0;
 416                 smp_mb(); /* clear flag *before* checking for work */
 417
 418                 if (do_block_io_op(blkif))
 419                         blkif->waiting_reqs = 1;
 420
 421                 if (log_stats && time_after(jiffies, blkif->st_print))
 422                         print_stats(blkif);
 423         }
 424
 425         /* Free all persistent grant pages */
 426         if (!RB_EMPTY_ROOT(&blkif->persistent_gnts))
 427                 free_persistent_gnts(&blkif->persistent_gnts,
 428                         blkif->persistent_gnt_c);
 429
 430         BUG_ON(!RB_EMPTY_ROOT(&blkif->persistent_gnts));
 431         blkif->persistent_gnt_c = 0;
 432
 433         if (log_stats)
 434                 print_stats(blkif);
 435
 436         blkif->xenblkd = NULL;
 437         xen_blkif_put(blkif);
 438
 439         return 0;
 440 }
 441
 442 struct seg_buf {
 443         unsigned long buf;
 444         unsigned int nsec;
 445 };
 446 /*
 447  * Unmap the grant references, and also remove the M2P over-rides
 448  * used in the 'pending_req'.
 449  */
 450 static void xen_blkbk_unmap(struct pending_req *req)
 451 {
 452         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 453         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 454         unsigned int i, invcount = 0;
 455         grant_handle_t handle;
 456         int ret;
 457
 458         for (i = 0; i < req->nr_pages; i++) {
 459                 if (!test_bit(i, req->unmap_seg))
 460                         continue;
 461                 handle = pending_handle(req, i);
 462                 if (handle == BLKBACK_INVALID_HANDLE)
 463                         continue;
 464                 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
 465                                     GNTMAP_host_map, handle);
 466                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
 467                 pages[invcount] = virt_to_page(vaddr(req, i));
 468                 invcount++;
 469         }
 470
 471         ret = gnttab_unmap_refs(unmap, NULL, pages, invcount);
 472         BUG_ON(ret);
 473 }
 474
 475 static int xen_blkbk_map(struct blkif_request *req,
 476                          struct pending_req *pending_req,
 477                          struct seg_buf seg[],
 478                          struct page *pages[])
 479 {
 480         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 481         struct persistent_gnt *persistent_gnts[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 482         struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 483         struct persistent_gnt *persistent_gnt = NULL;
 484         struct xen_blkif *blkif = pending_req->blkif;
 485         phys_addr_t addr = 0;
 486         int i, j;
 487         bool new_map;
 488         int nseg = req->u.rw.nr_segments;
 489         int segs_to_map = 0;
 490         int ret = 0;
 491         int use_persistent_gnts;
 492
 493         use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
 494
 495         BUG_ON(blkif->persistent_gnt_c >
 496                    max_mapped_grant_pages(pending_req->blkif->blk_protocol));
 497
 498         /*
 499          * Fill out preq.nr_sects with proper amount of sectors, and setup
 500          * assign map[..] with the PFN of the page in our domain with the
 501          * corresponding grant reference for each page.
 502          */
 503         for (i = 0; i < nseg; i++) {
 504                 uint32_t flags;
 505
 506                 if (use_persistent_gnts)
 507                         persistent_gnt = get_persistent_gnt(
 508                                 &blkif->persistent_gnts,
 509                                 req->u.rw.seg[i].gref);
 510
 511                 if (persistent_gnt) {
 512                         /*
 513                          * We are using persistent grants and
 514                          * the grant is already mapped
 515                          */
 516                         new_map = false;
 517                 } else if (use_persistent_gnts &&
 518                            blkif->persistent_gnt_c <
 519                            max_mapped_grant_pages(blkif->blk_protocol)) {
 520                         /*
 521                          * We are using persistent grants, the grant is
 522                          * not mapped but we have room for it
 523                          */
 524                         new_map = true;
 525                         persistent_gnt = kmalloc(
 526                                 sizeof(struct persistent_gnt),
 527                                 GFP_KERNEL);
 528                         if (!persistent_gnt)
 529                                 return -ENOMEM;
 530                         persistent_gnt->page = alloc_page(GFP_KERNEL);
 531                         if (!persistent_gnt->page) {
 532                                 kfree(persistent_gnt);
 533                                 return -ENOMEM;
 534                         }
 535                         persistent_gnt->gnt = req->u.rw.seg[i].gref;
 536                         persistent_gnt->handle = BLKBACK_INVALID_HANDLE;
 537
 538                         pages_to_gnt[segs_to_map] =
 539                                 persistent_gnt->page;
 540                         addr = (unsigned long) pfn_to_kaddr(
 541                                 page_to_pfn(persistent_gnt->page));
 542
 543                         add_persistent_gnt(&blkif->persistent_gnts,
 544                                 persistent_gnt);
 545                         blkif->persistent_gnt_c++;
 546                         pr_debug(DRV_PFX " grant %u added to the tree of persistent grants, using %u/%u\n",
 547                                  persistent_gnt->gnt, blkif->persistent_gnt_c,
 548                                  max_mapped_grant_pages(blkif->blk_protocol));
 549                 } else {
 550                         /*
 551                          * We are either using persistent grants and
 552                          * hit the maximum limit of grants mapped,
 553                          * or we are not using persistent grants.
 554                          */
 555                         if (use_persistent_gnts &&
 556                                 !blkif->vbd.overflow_max_grants) {
 557                                 blkif->vbd.overflow_max_grants = 1;
 558                                 pr_alert(DRV_PFX " domain %u, device %#x is using maximum number of persistent grants\n",
 559                                          blkif->domid, blkif->vbd.handle);
 560                         }
 561                         new_map = true;
 562                         pages[i] = blkbk->pending_page(pending_req, i);
 563                         addr = vaddr(pending_req, i);
 564                         pages_to_gnt[segs_to_map] =
 565                                 blkbk->pending_page(pending_req, i);
 566                 }
 567
 568                 if (persistent_gnt) {
 569                         pages[i] = persistent_gnt->page;
 570                         persistent_gnts[i] = persistent_gnt;
 571                 } else {
 572                         persistent_gnts[i] = NULL;
 573                 }
 574
 575                 if (new_map) {
 576                         flags = GNTMAP_host_map;
 577                         if (!persistent_gnt &&
 578                             (pending_req->operation != BLKIF_OP_READ))
 579                                 flags |= GNTMAP_readonly;
 580                         gnttab_set_map_op(&map[segs_to_map++], addr,
 581                                           flags, req->u.rw.seg[i].gref,
 582                                           blkif->domid);
 583                 }
 584         }
 585
 586         if (segs_to_map) {
 587                 ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
 588                 BUG_ON(ret);
 589         }
 590
 591         /*
 592          * Now swizzle the MFN in our domain with the MFN from the other domain
 593          * so that when we access vaddr(pending_req,i) it has the contents of
 594          * the page from the other domain.
 595          */
 596         bitmap_zero(pending_req->unmap_seg, BLKIF_MAX_SEGMENTS_PER_REQUEST);
 597         for (i = 0, j = 0; i < nseg; i++) {
 598                 if (!persistent_gnts[i] ||
 599                     persistent_gnts[i]->handle == BLKBACK_INVALID_HANDLE) {
 600                         /* This is a newly mapped grant */
 601                         BUG_ON(j >= segs_to_map);
 602                         if (unlikely(map[j].status != 0)) {
 603                                 pr_debug(DRV_PFX "invalid buffer -- could not remap it\n");
 604                                 map[j].handle = BLKBACK_INVALID_HANDLE;
 605                                 ret |= 1;
 606                                 if (persistent_gnts[i]) {
 607                                         rb_erase(&persistent_gnts[i]->node,
 608                                                  &blkif->persistent_gnts);
 609                                         blkif->persistent_gnt_c--;
 610                                         kfree(persistent_gnts[i]);
 611                                         persistent_gnts[i] = NULL;
 612                                 }
 613                         }
 614                 }
 615                 if (persistent_gnts[i]) {
 616                         if (persistent_gnts[i]->handle ==
 617                             BLKBACK_INVALID_HANDLE) {
 618                                 /*
 619                                  * If this is a new persistent grant
 620                                  * save the handler
 621                                  */
 622                                 persistent_gnts[i]->handle = map[j].handle;
 623                                 persistent_gnts[i]->dev_bus_addr =
 624                                         map[j++].dev_bus_addr;
 625                         }
 626                         pending_handle(pending_req, i) =
 627                                 persistent_gnts[i]->handle;
 628
 629                         if (ret)
 630                                 continue;
 631
 632                         seg[i].buf = persistent_gnts[i]->dev_bus_addr |
 633                                 (req->u.rw.seg[i].first_sect << 9);
 634                 } else {
 635                         pending_handle(pending_req, i) = map[j].handle;
 636                         bitmap_set(pending_req->unmap_seg, i, 1);
 637
 638                         if (ret) {
 639                                 j++;
 640                                 continue;
 641                         }
 642
 643                         seg[i].buf = map[j++].dev_bus_addr |
 644                                 (req->u.rw.seg[i].first_sect << 9);
 645                 }
 646         }
 647         return ret;
 648 }
 649
 650 static int dispatch_discard_io(struct xen_blkif *blkif,
 651                                 struct blkif_request *req)
 652 {
 653         int err = 0;
 654         int status = BLKIF_RSP_OKAY;
 655         struct block_device *bdev = blkif->vbd.bdev;
 656         unsigned long secure;
 657
 658         blkif->st_ds_req++;
 659
 660         xen_blkif_get(blkif);
 661         secure = (blkif->vbd.discard_secure &&
 662                  (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
 663                  BLKDEV_DISCARD_SECURE : 0;
 664
 665         err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
 666                                    req->u.discard.nr_sectors,
 667                                    GFP_KERNEL, secure);
 668
 669         if (err == -EOPNOTSUPP) {
 670                 pr_debug(DRV_PFX "discard op failed, not supported\n");
 671                 status = BLKIF_RSP_EOPNOTSUPP;
 672         } else if (err)
 673                 status = BLKIF_RSP_ERROR;
 674
 675         make_response(blkif, req->u.discard.id, req->operation, status);
 676         xen_blkif_put(blkif);
 677         return err;
 678 }
 679
 680 static void xen_blk_drain_io(struct xen_blkif *blkif)
 681 {
 682         atomic_set(&blkif->drain, 1);
 683         do {
 684                 /* The initial value is one, and one refcnt taken at the
 685                  * start of the xen_blkif_schedule thread. */
 686                 if (atomic_read(&blkif->refcnt) <= 2)
 687                         break;
 688                 wait_for_completion_interruptible_timeout(
 689                                 &blkif->drain_complete, HZ);
 690
 691                 if (!atomic_read(&blkif->drain))
 692                         break;
 693         } while (!kthread_should_stop());
 694         atomic_set(&blkif->drain, 0);
 695 }
 696
 697 /*
 698  * Completion callback on the bio's. Called as bh->b_end_io()
 699  */
 700
 701 static void __end_block_io_op(struct pending_req *pending_req, int error)
 702 {
 703         /* An error fails the entire request. */
 704         if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
 705             (error == -EOPNOTSUPP)) {
 706                 pr_debug(DRV_PFX "flush diskcache op failed, not supported\n");
 707                 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
 708                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 709         } else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
 710                     (error == -EOPNOTSUPP)) {
 711                 pr_debug(DRV_PFX "write barrier op failed, not supported\n");
 712                 xen_blkbk_barrier(XBT_NIL, pending_req->blkif->be, 0);
 713                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 714         } else if (error) {
 715                 pr_debug(DRV_PFX "Buffer not up-to-date at end of operation,"
 716                          " error=%d\n", error);
 717                 pending_req->status = BLKIF_RSP_ERROR;
 718         }
 719
 720         /*
 721          * If all of the bio's have completed it is time to unmap
 722          * the grant references associated with 'request' and provide
 723          * the proper response on the ring.
 724          */
 725         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 726                 xen_blkbk_unmap(pending_req);
 727                 make_response(pending_req->blkif, pending_req->id,
 728                               pending_req->operation, pending_req->status);
 729                 xen_blkif_put(pending_req->blkif);
 730                 if (atomic_read(&pending_req->blkif->refcnt) <= 2) {
 731                         if (atomic_read(&pending_req->blkif->drain))
 732                                 complete(&pending_req->blkif->drain_complete);
 733                 }
 734                 free_req(pending_req);
 735         }
 736 }
 737
 738 /*
 739  * bio callback.
 740  */
 741 static void end_block_io_op(struct bio *bio, int error)
 742 {
 743         __end_block_io_op(bio->bi_private, error);
 744         bio_put(bio);
 745 }
 746
 747
 748
 749 /*
 750  * Function to copy the from the ring buffer the 'struct blkif_request'
 751  * (which has the sectors we want, number of them, grant references, etc),
 752  * and transmute  it to the block API to hand it over to the proper block disk.
 753  */
 754 static int
 755 __do_block_io_op(struct xen_blkif *blkif)
 756 {
 757         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 758         struct blkif_request req;
 759         struct pending_req *pending_req;
 760         RING_IDX rc, rp;
 761         int more_to_do = 0;
 762
 763         rc = blk_rings->common.req_cons;
 764         rp = blk_rings->common.sring->req_prod;
 765         rmb(); /* Ensure we see queued requests up to 'rp'. */
 766
 767         while (rc != rp) {
 768
 769                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
 770                         break;
 771
 772                 if (kthread_should_stop()) {
 773                         more_to_do = 1;
 774                         break;
 775                 }
 776
 777                 pending_req = alloc_req();
 778                 if (NULL == pending_req) {
 779                         blkif->st_oo_req++;
 780                         more_to_do = 1;
 781                         break;
 782                 }
 783
 784                 switch (blkif->blk_protocol) {
 785                 case BLKIF_PROTOCOL_NATIVE:
 786                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
 787                         break;
 788                 case BLKIF_PROTOCOL_X86_32:
 789                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
 790                         break;
 791                 case BLKIF_PROTOCOL_X86_64:
 792                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
 793                         break;
 794                 default:
 795                         BUG();
 796                 }
 797                 blk_rings->common.req_cons = ++rc; /* before make_response() */
 798
 799                 /* Apply all sanity checks to /private copy/ of request. */
 800                 barrier();
 801                 if (unlikely(req.operation == BLKIF_OP_DISCARD)) {
 802                         free_req(pending_req);
 803                         if (dispatch_discard_io(blkif, &req))
 804                                 break;
 805                 } else if (dispatch_rw_block_io(blkif, &req, pending_req))
 806                         break;
 807
 808                 /* Yield point for this unbounded loop. */
 809                 cond_resched();
 810         }
 811
 812         return more_to_do;
 813 }
 814
 815 static int
 816 do_block_io_op(struct xen_blkif *blkif)
 817 {
 818         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 819         int more_to_do;
 820
 821         do {
 822                 more_to_do = __do_block_io_op(blkif);
 823                 if (more_to_do)
 824                         break;
 825
 826                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
 827         } while (more_to_do);
 828
 829         return more_to_do;
 830 }
 831 /*
 832  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
 833  * and call the 'submit_bio' to pass it to the underlying storage.
 834  */
 835 static int dispatch_rw_block_io(struct xen_blkif *blkif,
 836                                 struct blkif_request *req,
 837                                 struct pending_req *pending_req)
 838 {
 839         struct phys_req preq;
 840         struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 841         unsigned int nseg;
 842         struct bio *bio = NULL;
 843         struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 844         int i, nbio = 0;
 845         int operation;
 846         struct blk_plug plug;
 847         bool drain = false;
 848         struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 849
 850         switch (req->operation) {
 851         case BLKIF_OP_READ:
 852                 blkif->st_rd_req++;
 853                 operation = READ;
 854                 break;
 855         case BLKIF_OP_WRITE:
 856                 blkif->st_wr_req++;
 857                 operation = WRITE_ODIRECT;
 858                 break;
 859         case BLKIF_OP_WRITE_BARRIER:
 860                 drain = true;
 861         case BLKIF_OP_FLUSH_DISKCACHE:
 862                 blkif->st_f_req++;
 863                 operation = WRITE_FLUSH;
 864                 break;
 865         default:
 866                 operation = 0; /* make gcc happy */
 867                 goto fail_response;
 868                 break;
 869         }
 870
 871         /* Check that the number of segments is sane. */
 872         nseg = req->u.rw.nr_segments;
 873
 874         if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
 875             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
 876                 pr_debug(DRV_PFX "Bad number of segments in request (%d)\n",
 877                          nseg);
 878                 /* Haven't submitted any bio's yet. */
 879                 goto fail_response;
 880         }
 881
 882         preq.dev           = req->u.rw.handle;
 883         preq.sector_number = req->u.rw.sector_number;
 884         preq.nr_sects      = 0;
 885
 886         pending_req->blkif     = blkif;
 887         pending_req->id        = req->u.rw.id;
 888         pending_req->operation = req->operation;
 889         pending_req->status    = BLKIF_RSP_OKAY;
 890         pending_req->nr_pages  = nseg;
 891
 892         for (i = 0; i < nseg; i++) {
 893                 seg[i].nsec = req->u.rw.seg[i].last_sect -
 894                         req->u.rw.seg[i].first_sect + 1;
 895                 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
 896                     (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
 897                         goto fail_response;
 898                 preq.nr_sects += seg[i].nsec;
 899
 900         }
 901
 902         if (xen_vbd_translate(&preq, blkif, operation) != 0) {
 903                 pr_debug(DRV_PFX "access denied: %s of [%llu,%llu] on dev=%04x\n",
 904                          operation == READ ? "read" : "write",
 905                          preq.sector_number,
 906                          preq.sector_number + preq.nr_sects, preq.dev);
 907                 goto fail_response;
 908         }
 909
 910         /*
 911          * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
 912          * is set there.
 913          */
 914         for (i = 0; i < nseg; i++) {
 915                 if (((int)preq.sector_number|(int)seg[i].nsec) &
 916                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
 917                         pr_debug(DRV_PFX "Misaligned I/O request from domain %d",
 918                                  blkif->domid);
 919                         goto fail_response;
 920                 }
 921         }
 922
 923         /* Wait on all outstanding I/O's and once that has been completed
 924          * issue the WRITE_FLUSH.
 925          */
 926         if (drain)
 927                 xen_blk_drain_io(pending_req->blkif);
 928
 929         /*
 930          * If we have failed at this point, we need to undo the M2P override,
 931          * set gnttab_set_unmap_op on all of the grant references and perform
 932          * the hypercall to unmap the grants - that is all done in
 933          * xen_blkbk_unmap.
 934          */
 935         if (xen_blkbk_map(req, pending_req, seg, pages))
 936                 goto fail_flush;
 937
 938         /*
 939          * This corresponding xen_blkif_put is done in __end_block_io_op, or
 940          * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
 941          */
 942         xen_blkif_get(blkif);
 943
 944         for (i = 0; i < nseg; i++) {
 945                 while ((bio == NULL) ||
 946                        (bio_add_page(bio,
 947                                      pages[i],
 948                                      seg[i].nsec << 9,
 949                                      seg[i].buf & ~PAGE_MASK) == 0)) {
 950
 951                         bio = bio_alloc(GFP_KERNEL, nseg-i);
 952                         if (unlikely(bio == NULL))
 953                                 goto fail_put_bio;
 954
 955                         biolist[nbio++] = bio;
 956                         bio->bi_bdev    = preq.bdev;
 957                         bio->bi_private = pending_req;
 958                         bio->bi_end_io  = end_block_io_op;
 959                         bio->bi_sector  = preq.sector_number;
 960                 }
 961
 962                 preq.sector_number += seg[i].nsec;
 963         }
 964
 965         /* This will be hit if the operation was a flush or discard. */
 966         if (!bio) {
 967                 BUG_ON(operation != WRITE_FLUSH);
 968
 969                 bio = bio_alloc(GFP_KERNEL, 0);
 970                 if (unlikely(bio == NULL))
 971                         goto fail_put_bio;
 972
 973                 biolist[nbio++] = bio;
 974                 bio->bi_bdev    = preq.bdev;
 975                 bio->bi_private = pending_req;
 976                 bio->bi_end_io  = end_block_io_op;
 977         }
 978
 979         /*
 980          * We set it one so that the last submit_bio does not have to call
 981          * atomic_inc.
 982          */
 983         atomic_set(&pending_req->pendcnt, nbio);
 984
 985         /* Get a reference count for the disk queue and start sending I/O */
 986         blk_start_plug(&plug);
 987
 988         for (i = 0; i < nbio; i++)
 989                 submit_bio(operation, biolist[i]);
 990
 991         /* Let the I/Os go.. */
 992         blk_finish_plug(&plug);
 993
 994         if (operation == READ)
 995                 blkif->st_rd_sect += preq.nr_sects;
 996         else if (operation & WRITE)
 997                 blkif->st_wr_sect += preq.nr_sects;
 998
 999         return 0;
1000
1001  fail_flush:
1002         xen_blkbk_unmap(pending_req);
1003  fail_response:
1004         /* Haven't submitted any bio's yet. */
1005         make_response(blkif, req->u.rw.id, req->operation, BLKIF_RSP_ERROR);
1006         free_req(pending_req);
1007         msleep(1); /* back off a bit */
1008         return -EIO;
1009
1010  fail_put_bio:
1011         for (i = 0; i < nbio; i++)
1012                 bio_put(biolist[i]);
1013         __end_block_io_op(pending_req, -EINVAL);
1014         msleep(1); /* back off a bit */
1015         return -EIO;
1016 }
1017
1018
1019
1020 /*
1021  * Put a response on the ring on how the operation fared.
1022  */
1023 static void make_response(struct xen_blkif *blkif, u64 id,
1024                           unsigned short op, int st)
1025 {
1026         struct blkif_response  resp;
1027         unsigned long     flags;
1028         union blkif_back_rings *blk_rings = &blkif->blk_rings;
1029         int notify;
1030
1031         resp.id        = id;
1032         resp.operation = op;
1033         resp.status    = st;
1034
1035         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
1036         /* Place on the response ring for the relevant domain. */
1037         switch (blkif->blk_protocol) {
1038         case BLKIF_PROTOCOL_NATIVE:
1039                 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
1040                        &resp, sizeof(resp));
1041                 break;
1042         case BLKIF_PROTOCOL_X86_32:
1043                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
1044                        &resp, sizeof(resp));
1045                 break;
1046         case BLKIF_PROTOCOL_X86_64:
1047                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
1048                        &resp, sizeof(resp));
1049                 break;
1050         default:
1051                 BUG();
1052         }
1053         blk_rings->common.rsp_prod_pvt++;
1054         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1055         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
1056         if (notify)
1057                 notify_remote_via_irq(blkif->irq);
1058 }
1059
1060 static int __init xen_blkif_init(void)
1061 {
1062         int i, mmap_pages;
1063         int rc = 0;
1064
1065         if (!xen_domain())
1066                 return -ENODEV;
1067
1068         blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
1069         if (!blkbk) {
1070                 pr_alert(DRV_PFX "%s: out of memory!\n", __func__);
1071                 return -ENOMEM;
1072         }
1073
1074         mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
1075
1076         blkbk->pending_reqs          = kzalloc(sizeof(blkbk->pending_reqs[0]) *
1077                                         xen_blkif_reqs, GFP_KERNEL);
1078         blkbk->pending_grant_handles = kmalloc(sizeof(blkbk->pending_grant_handles[0]) *
1079                                         mmap_pages, GFP_KERNEL);
1080         blkbk->pending_pages         = kzalloc(sizeof(blkbk->pending_pages[0]) *
1081                                         mmap_pages, GFP_KERNEL);
1082
1083         if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
1084             !blkbk->pending_pages) {
1085                 rc = -ENOMEM;
1086                 goto out_of_memory;
1087         }
1088
1089         for (i = 0; i < mmap_pages; i++) {
1090                 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
1091                 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
1092                 if (blkbk->pending_pages[i] == NULL) {
1093                         rc = -ENOMEM;
1094                         goto out_of_memory;
1095                 }
1096         }
1097         rc = xen_blkif_interface_init();
1098         if (rc)
1099                 goto failed_init;
1100
1101         INIT_LIST_HEAD(&blkbk->pending_free);
1102         spin_lock_init(&blkbk->pending_free_lock);
1103         init_waitqueue_head(&blkbk->pending_free_wq);
1104
1105         for (i = 0; i < xen_blkif_reqs; i++)
1106                 list_add_tail(&blkbk->pending_reqs[i].free_list,
1107                               &blkbk->pending_free);
1108
1109         rc = xen_blkif_xenbus_init();
1110         if (rc)
1111                 goto failed_init;
1112
1113         return 0;
1114
1115  out_of_memory:
1116         pr_alert(DRV_PFX "%s: out of memory\n", __func__);
1117  failed_init:
1118         kfree(blkbk->pending_reqs);
1119         kfree(blkbk->pending_grant_handles);
1120         if (blkbk->pending_pages) {
1121                 for (i = 0; i < mmap_pages; i++) {
1122                         if (blkbk->pending_pages[i])
1123                                 __free_page(blkbk->pending_pages[i]);
1124                 }
1125                 kfree(blkbk->pending_pages);
1126         }
1127         kfree(blkbk);
1128         blkbk = NULL;
1129         return rc;
1130 }
1131
1132 module_init(xen_blkif_init);
1133
1134 MODULE_LICENSE("Dual BSD/GPL");
1135 MODULE_ALIAS("xen-backend:vbd");