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/loop.h>
  43 #include <linux/falloc.h>
  44 #include <linux/fs.h>
  45
  46 #include <xen/events.h>
  47 #include <xen/page.h>
  48 #include <asm/xen/hypervisor.h>
  49 #include <asm/xen/hypercall.h>
  50 #include "common.h"
  51
  52 /*
  53  * These are rather arbitrary. They are fairly large because adjacent requests
  54  * pulled from a communication ring are quite likely to end up being part of
  55  * the same scatter/gather request at the disc.
  56  *
  57  * ** TRY INCREASING 'xen_blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
  58  *
  59  * This will increase the chances of being able to write whole tracks.
  60  * 64 should be enough to keep us competitive with Linux.
  61  */
  62 static int xen_blkif_reqs = 64;
  63 module_param_named(reqs, xen_blkif_reqs, int, 0);
  64 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
  65
  66 /* Run-time switchable: /sys/module/blkback/parameters/ */
  67 static unsigned int log_stats;
  68 module_param(log_stats, int, 0644);
  69
  70 /*
  71  * Each outstanding request that we've passed to the lower device layers has a
  72  * 'pending_req' allocated to it. Each buffer_head that completes decrements
  73  * the pendcnt towards zero. When it hits zero, the specified domain has a
  74  * response queued for it, with the saved 'id' passed back.
  75  */
  76 struct pending_req {
  77         struct xen_blkif        *blkif;
  78         u64                     id;
  79         int                     nr_pages;
  80         atomic_t                pendcnt;
  81         unsigned short          operation;
  82         int                     status;
  83         struct list_head        free_list;
  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  * Little helpful macro to figure out the index and virtual address of the
 105  * pending_pages[..]. For each 'pending_req' we have have up to
 106  * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
 107  * 10 and would index in the pending_pages[..].
 108  */
 109 static inline int vaddr_pagenr(struct pending_req *req, int seg)
 110 {
 111         return (req - blkbk->pending_reqs) *
 112                 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
 113 }
 114
 115 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
 116
 117 static inline unsigned long vaddr(struct pending_req *req, int seg)
 118 {
 119         unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
 120         return (unsigned long)pfn_to_kaddr(pfn);
 121 }
 122
 123 #define pending_handle(_req, _seg) \
 124         (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
 125
 126
 127 static int do_block_io_op(struct xen_blkif *blkif);
 128 static int dispatch_rw_block_io(struct xen_blkif *blkif,
 129                                 struct blkif_request *req,
 130                                 struct pending_req *pending_req);
 131 static void make_response(struct xen_blkif *blkif, u64 id,
 132                           unsigned short op, int st);
 133
 134 /*
 135  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
 136  */
 137 static struct pending_req *alloc_req(void)
 138 {
 139         struct pending_req *req = NULL;
 140         unsigned long flags;
 141
 142         spin_lock_irqsave(&blkbk->pending_free_lock, flags);
 143         if (!list_empty(&blkbk->pending_free)) {
 144                 req = list_entry(blkbk->pending_free.next, struct pending_req,
 145                                  free_list);
 146                 list_del(&req->free_list);
 147         }
 148         spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
 149         return req;
 150 }
 151
 152 /*
 153  * Return the 'pending_req' structure back to the freepool. We also
 154  * wake up the thread if it was waiting for a free page.
 155  */
 156 static void free_req(struct pending_req *req)
 157 {
 158         unsigned long flags;
 159         int was_empty;
 160
 161         spin_lock_irqsave(&blkbk->pending_free_lock, flags);
 162         was_empty = list_empty(&blkbk->pending_free);
 163         list_add(&req->free_list, &blkbk->pending_free);
 164         spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
 165         if (was_empty)
 166                 wake_up(&blkbk->pending_free_wq);
 167 }
 168
 169 /*
 170  * Routines for managing virtual block devices (vbds).
 171  */
 172 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
 173                              int operation)
 174 {
 175         struct xen_vbd *vbd = &blkif->vbd;
 176         int rc = -EACCES;
 177
 178         if ((operation != READ) && vbd->readonly)
 179                 goto out;
 180
 181         if (likely(req->nr_sects)) {
 182                 blkif_sector_t end = req->sector_number + req->nr_sects;
 183
 184                 if (unlikely(end < req->sector_number))
 185                         goto out;
 186                 if (unlikely(end > vbd_sz(vbd)))
 187                         goto out;
 188         }
 189
 190         req->dev  = vbd->pdevice;
 191         req->bdev = vbd->bdev;
 192         rc = 0;
 193
 194  out:
 195         return rc;
 196 }
 197
 198 static void xen_vbd_resize(struct xen_blkif *blkif)
 199 {
 200         struct xen_vbd *vbd = &blkif->vbd;
 201         struct xenbus_transaction xbt;
 202         int err;
 203         struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
 204         unsigned long long new_size = vbd_sz(vbd);
 205
 206         pr_info(DRV_PFX "VBD Resize: Domid: %d, Device: (%d, %d)\n",
 207                 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
 208         pr_info(DRV_PFX "VBD Resize: new size %llu\n", new_size);
 209         vbd->size = new_size;
 210 again:
 211         err = xenbus_transaction_start(&xbt);
 212         if (err) {
 213                 pr_warn(DRV_PFX "Error starting transaction");
 214                 return;
 215         }
 216         err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
 217                             (unsigned long long)vbd_sz(vbd));
 218         if (err) {
 219                 pr_warn(DRV_PFX "Error writing new size");
 220                 goto abort;
 221         }
 222         /*
 223          * Write the current state; we will use this to synchronize
 224          * the front-end. If the current state is "connected" the
 225          * front-end will get the new size information online.
 226          */
 227         err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
 228         if (err) {
 229                 pr_warn(DRV_PFX "Error writing the state");
 230                 goto abort;
 231         }
 232
 233         err = xenbus_transaction_end(xbt, 0);
 234         if (err == -EAGAIN)
 235                 goto again;
 236         if (err)
 237                 pr_warn(DRV_PFX "Error ending transaction");
 238         return;
 239 abort:
 240         xenbus_transaction_end(xbt, 1);
 241 }
 242
 243 /*
 244  * Notification from the guest OS.
 245  */
 246 static void blkif_notify_work(struct xen_blkif *blkif)
 247 {
 248         blkif->waiting_reqs = 1;
 249         wake_up(&blkif->wq);
 250 }
 251
 252 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
 253 {
 254         blkif_notify_work(dev_id);
 255         return IRQ_HANDLED;
 256 }
 257
 258 /*
 259  * SCHEDULER FUNCTIONS
 260  */
 261
 262 static void print_stats(struct xen_blkif *blkif)
 263 {
 264         pr_info("xen-blkback (%s): oo %3d  |  rd %4d  |  wr %4d  |  f %4d"
 265                  "  |  ds %4d\n",
 266                  current->comm, blkif->st_oo_req,
 267                  blkif->st_rd_req, blkif->st_wr_req,
 268                  blkif->st_f_req, blkif->st_ds_req);
 269         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 270         blkif->st_rd_req = 0;
 271         blkif->st_wr_req = 0;
 272         blkif->st_oo_req = 0;
 273         blkif->st_ds_req = 0;
 274 }
 275
 276 int xen_blkif_schedule(void *arg)
 277 {
 278         struct xen_blkif *blkif = arg;
 279         struct xen_vbd *vbd = &blkif->vbd;
 280
 281         xen_blkif_get(blkif);
 282
 283         while (!kthread_should_stop()) {
 284                 if (try_to_freeze())
 285                         continue;
 286                 if (unlikely(vbd->size != vbd_sz(vbd)))
 287                         xen_vbd_resize(blkif);
 288
 289                 wait_event_interruptible(
 290                         blkif->wq,
 291                         blkif->waiting_reqs || kthread_should_stop());
 292                 wait_event_interruptible(
 293                         blkbk->pending_free_wq,
 294                         !list_empty(&blkbk->pending_free) ||
 295                         kthread_should_stop());
 296
 297                 blkif->waiting_reqs = 0;
 298                 smp_mb(); /* clear flag *before* checking for work */
 299
 300                 if (do_block_io_op(blkif))
 301                         blkif->waiting_reqs = 1;
 302
 303                 if (log_stats && time_after(jiffies, blkif->st_print))
 304                         print_stats(blkif);
 305         }
 306
 307         if (log_stats)
 308                 print_stats(blkif);
 309
 310         blkif->xenblkd = NULL;
 311         xen_blkif_put(blkif);
 312
 313         return 0;
 314 }
 315
 316 struct seg_buf {
 317         unsigned long buf;
 318         unsigned int nsec;
 319 };
 320 /*
 321  * Unmap the grant references, and also remove the M2P over-rides
 322  * used in the 'pending_req'.
 323  */
 324 static void xen_blkbk_unmap(struct pending_req *req)
 325 {
 326         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 327         unsigned int i, invcount = 0;
 328         grant_handle_t handle;
 329         int ret;
 330
 331         for (i = 0; i < req->nr_pages; i++) {
 332                 handle = pending_handle(req, i);
 333                 if (handle == BLKBACK_INVALID_HANDLE)
 334                         continue;
 335                 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
 336                                     GNTMAP_host_map, handle);
 337                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
 338                 invcount++;
 339         }
 340
 341         ret = HYPERVISOR_grant_table_op(
 342                 GNTTABOP_unmap_grant_ref, unmap, invcount);
 343         BUG_ON(ret);
 344         /*
 345          * Note, we use invcount, so nr->pages, so we can't index
 346          * using vaddr(req, i).
 347          */
 348         for (i = 0; i < invcount; i++) {
 349                 ret = m2p_remove_override(
 350                         virt_to_page(unmap[i].host_addr), false);
 351                 if (ret) {
 352                         pr_alert(DRV_PFX "Failed to remove M2P override for %lx\n",
 353                                  (unsigned long)unmap[i].host_addr);
 354                         continue;
 355                 }
 356         }
 357 }
 358
 359 static int xen_blkbk_map(struct blkif_request *req,
 360                          struct pending_req *pending_req,
 361                          struct seg_buf seg[])
 362 {
 363         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 364         int i;
 365         int nseg = req->nr_segments;
 366         int ret = 0;
 367
 368         /*
 369          * Fill out preq.nr_sects with proper amount of sectors, and setup
 370          * assign map[..] with the PFN of the page in our domain with the
 371          * corresponding grant reference for each page.
 372          */
 373         for (i = 0; i < nseg; i++) {
 374                 uint32_t flags;
 375
 376                 flags = GNTMAP_host_map;
 377                 if (pending_req->operation != BLKIF_OP_READ)
 378                         flags |= GNTMAP_readonly;
 379                 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
 380                                   req->u.rw.seg[i].gref,
 381                                   pending_req->blkif->domid);
 382         }
 383
 384         ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
 385         BUG_ON(ret);
 386
 387         /*
 388          * Now swizzle the MFN in our domain with the MFN from the other domain
 389          * so that when we access vaddr(pending_req,i) it has the contents of
 390          * the page from the other domain.
 391          */
 392         for (i = 0; i < nseg; i++) {
 393                 if (unlikely(map[i].status != 0)) {
 394                         pr_debug(DRV_PFX "invalid buffer -- could not remap it\n");
 395                         map[i].handle = BLKBACK_INVALID_HANDLE;
 396                         ret |= 1;
 397                 }
 398
 399                 pending_handle(pending_req, i) = map[i].handle;
 400
 401                 if (ret)
 402                         continue;
 403
 404                 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
 405                         blkbk->pending_page(pending_req, i), false);
 406                 if (ret) {
 407                         pr_alert(DRV_PFX "Failed to install M2P override for %lx (ret: %d)\n",
 408                                  (unsigned long)map[i].dev_bus_addr, ret);
 409                         /* We could switch over to GNTTABOP_copy */
 410                         continue;
 411                 }
 412
 413                 seg[i].buf  = map[i].dev_bus_addr |
 414                         (req->u.rw.seg[i].first_sect << 9);
 415         }
 416         return ret;
 417 }
 418
 419 static void xen_blk_discard(struct xen_blkif *blkif, struct blkif_request *req)
 420 {
 421         int err = 0;
 422         int status = BLKIF_RSP_OKAY;
 423         struct block_device *bdev = blkif->vbd.bdev;
 424
 425         if (blkif->blk_backend_type == BLKIF_BACKEND_PHY)
 426                 /* just forward the discard request */
 427                 err = blkdev_issue_discard(bdev,
 428                                 req->u.discard.sector_number,
 429                                 req->u.discard.nr_sectors,
 430                                 GFP_KERNEL, 0);
 431         else if (blkif->blk_backend_type == BLKIF_BACKEND_FILE) {
 432                 /* punch a hole in the backing file */
 433                 struct loop_device *lo = bdev->bd_disk->private_data;
 434                 struct file *file = lo->lo_backing_file;
 435
 436                 if (file->f_op->fallocate)
 437                         err = file->f_op->fallocate(file,
 438                                 FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE,
 439                                 req->u.discard.sector_number << 9,
 440                                 req->u.discard.nr_sectors << 9);
 441                 else
 442                         err = -EOPNOTSUPP;
 443         } else
 444                 err = -EOPNOTSUPP;
 445
 446         if (err == -EOPNOTSUPP) {
 447                 pr_debug(DRV_PFX "discard op failed, not supported\n");
 448                 status = BLKIF_RSP_EOPNOTSUPP;
 449         } else if (err)
 450                 status = BLKIF_RSP_ERROR;
 451
 452         make_response(blkif, req->id, req->operation, status);
 453 }
 454
 455 /*
 456  * Completion callback on the bio's. Called as bh->b_end_io()
 457  */
 458
 459 static void __end_block_io_op(struct pending_req *pending_req, int error)
 460 {
 461         /* An error fails the entire request. */
 462         if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
 463             (error == -EOPNOTSUPP)) {
 464                 pr_debug(DRV_PFX "flush diskcache op failed, not supported\n");
 465                 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
 466                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 467         } else if (error) {
 468                 pr_debug(DRV_PFX "Buffer not up-to-date at end of operation,"
 469                          " error=%d\n", error);
 470                 pending_req->status = BLKIF_RSP_ERROR;
 471         }
 472
 473         /*
 474          * If all of the bio's have completed it is time to unmap
 475          * the grant references associated with 'request' and provide
 476          * the proper response on the ring.
 477          */
 478         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 479                 xen_blkbk_unmap(pending_req);
 480                 make_response(pending_req->blkif, pending_req->id,
 481                               pending_req->operation, pending_req->status);
 482                 xen_blkif_put(pending_req->blkif);
 483                 free_req(pending_req);
 484         }
 485 }
 486
 487 /*
 488  * bio callback.
 489  */
 490 static void end_block_io_op(struct bio *bio, int error)
 491 {
 492         __end_block_io_op(bio->bi_private, error);
 493         bio_put(bio);
 494 }
 495
 496
 497
 498 /*
 499  * Function to copy the from the ring buffer the 'struct blkif_request'
 500  * (which has the sectors we want, number of them, grant references, etc),
 501  * and transmute  it to the block API to hand it over to the proper block disk.
 502  */
 503 static int
 504 __do_block_io_op(struct xen_blkif *blkif)
 505 {
 506         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 507         struct blkif_request req;
 508         struct pending_req *pending_req;
 509         RING_IDX rc, rp;
 510         int more_to_do = 0;
 511
 512         rc = blk_rings->common.req_cons;
 513         rp = blk_rings->common.sring->req_prod;
 514         rmb(); /* Ensure we see queued requests up to 'rp'. */
 515
 516         while (rc != rp) {
 517
 518                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
 519                         break;
 520
 521                 if (kthread_should_stop()) {
 522                         more_to_do = 1;
 523                         break;
 524                 }
 525
 526                 pending_req = alloc_req();
 527                 if (NULL == pending_req) {
 528                         blkif->st_oo_req++;
 529                         more_to_do = 1;
 530                         break;
 531                 }
 532
 533                 switch (blkif->blk_protocol) {
 534                 case BLKIF_PROTOCOL_NATIVE:
 535                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
 536                         break;
 537                 case BLKIF_PROTOCOL_X86_32:
 538                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
 539                         break;
 540                 case BLKIF_PROTOCOL_X86_64:
 541                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
 542                         break;
 543                 default:
 544                         BUG();
 545                 }
 546                 blk_rings->common.req_cons = ++rc; /* before make_response() */
 547
 548                 /* Apply all sanity checks to /private copy/ of request. */
 549                 barrier();
 550
 551                 if (dispatch_rw_block_io(blkif, &req, pending_req))
 552                         break;
 553
 554                 /* Yield point for this unbounded loop. */
 555                 cond_resched();
 556         }
 557
 558         return more_to_do;
 559 }
 560
 561 static int
 562 do_block_io_op(struct xen_blkif *blkif)
 563 {
 564         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 565         int more_to_do;
 566
 567         do {
 568                 more_to_do = __do_block_io_op(blkif);
 569                 if (more_to_do)
 570                         break;
 571
 572                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
 573         } while (more_to_do);
 574
 575         return more_to_do;
 576 }
 577
 578 /*
 579  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
 580  * and call the 'submit_bio' to pass it to the underlying storage.
 581  */
 582 static int dispatch_rw_block_io(struct xen_blkif *blkif,
 583                                 struct blkif_request *req,
 584                                 struct pending_req *pending_req)
 585 {
 586         struct phys_req preq;
 587         struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 588         unsigned int nseg;
 589         struct bio *bio = NULL;
 590         struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 591         int i, nbio = 0;
 592         int operation;
 593         struct blk_plug plug;
 594
 595         switch (req->operation) {
 596         case BLKIF_OP_READ:
 597                 blkif->st_rd_req++;
 598                 operation = READ;
 599                 break;
 600         case BLKIF_OP_WRITE:
 601                 blkif->st_wr_req++;
 602                 operation = WRITE_ODIRECT;
 603                 break;
 604         case BLKIF_OP_FLUSH_DISKCACHE:
 605                 blkif->st_f_req++;
 606                 operation = WRITE_FLUSH;
 607                 break;
 608         case BLKIF_OP_DISCARD:
 609                 blkif->st_ds_req++;
 610                 operation = REQ_DISCARD;
 611                 break;
 612         case BLKIF_OP_WRITE_BARRIER:
 613         default:
 614                 operation = 0; /* make gcc happy */
 615                 goto fail_response;
 616                 break;
 617         }
 618
 619         /* Check that the number of segments is sane. */
 620         nseg = req->nr_segments;
 621         if (unlikely(nseg == 0 && operation != WRITE_FLUSH &&
 622                                 operation != REQ_DISCARD) ||
 623             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
 624                 pr_debug(DRV_PFX "Bad number of segments in request (%d)\n",
 625                          nseg);
 626                 /* Haven't submitted any bio's yet. */
 627                 goto fail_response;
 628         }
 629
 630         preq.dev           = req->handle;
 631         preq.sector_number = req->u.rw.sector_number;
 632         preq.nr_sects      = 0;
 633
 634         pending_req->blkif     = blkif;
 635         pending_req->id        = req->id;
 636         pending_req->operation = req->operation;
 637         pending_req->status    = BLKIF_RSP_OKAY;
 638         pending_req->nr_pages  = nseg;
 639
 640         for (i = 0; i < nseg; i++) {
 641                 seg[i].nsec = req->u.rw.seg[i].last_sect -
 642                         req->u.rw.seg[i].first_sect + 1;
 643                 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
 644                     (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
 645                         goto fail_response;
 646                 preq.nr_sects += seg[i].nsec;
 647
 648         }
 649
 650         if (xen_vbd_translate(&preq, blkif, operation) != 0) {
 651                 pr_debug(DRV_PFX "access denied: %s of [%llu,%llu] on dev=%04x\n",
 652                          operation == READ ? "read" : "write",
 653                          preq.sector_number,
 654                          preq.sector_number + preq.nr_sects, preq.dev);
 655                 goto fail_response;
 656         }
 657
 658         /*
 659          * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
 660          * is set there.
 661          */
 662         for (i = 0; i < nseg; i++) {
 663                 if (((int)preq.sector_number|(int)seg[i].nsec) &
 664                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
 665                         pr_debug(DRV_PFX "Misaligned I/O request from domain %d",
 666                                  blkif->domid);
 667                         goto fail_response;
 668                 }
 669         }
 670
 671         /*
 672          * If we have failed at this point, we need to undo the M2P override,
 673          * set gnttab_set_unmap_op on all of the grant references and perform
 674          * the hypercall to unmap the grants - that is all done in
 675          * xen_blkbk_unmap.
 676          */
 677         if (operation != BLKIF_OP_DISCARD &&
 678                         xen_blkbk_map(req, pending_req, seg))
 679                 goto fail_flush;
 680
 681         /*
 682          * This corresponding xen_blkif_put is done in __end_block_io_op, or
 683          * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
 684          */
 685         xen_blkif_get(blkif);
 686
 687         for (i = 0; i < nseg; i++) {
 688                 while ((bio == NULL) ||
 689                        (bio_add_page(bio,
 690                                      blkbk->pending_page(pending_req, i),
 691                                      seg[i].nsec << 9,
 692                                      seg[i].buf & ~PAGE_MASK) == 0)) {
 693
 694                         bio = bio_alloc(GFP_KERNEL, nseg-i);
 695                         if (unlikely(bio == NULL))
 696                                 goto fail_put_bio;
 697
 698                         biolist[nbio++] = bio;
 699                         bio->bi_bdev    = preq.bdev;
 700                         bio->bi_private = pending_req;
 701                         bio->bi_end_io  = end_block_io_op;
 702                         bio->bi_sector  = preq.sector_number;
 703                 }
 704
 705                 preq.sector_number += seg[i].nsec;
 706         }
 707
 708         /* This will be hit if the operation was a flush or discard. */
 709         if (!bio) {
 710                 BUG_ON(operation != WRITE_FLUSH && operation != REQ_DISCARD);
 711
 712                 if (operation == WRITE_FLUSH) {
 713                         bio = bio_alloc(GFP_KERNEL, 0);
 714                         if (unlikely(bio == NULL))
 715                                 goto fail_put_bio;
 716
 717                         biolist[nbio++] = bio;
 718                         bio->bi_bdev    = preq.bdev;
 719                         bio->bi_private = pending_req;
 720                         bio->bi_end_io  = end_block_io_op;
 721                 } else if (operation == REQ_DISCARD) {
 722                         xen_blk_discard(blkif, req);
 723                         xen_blkif_put(blkif);
 724                         free_req(pending_req);
 725                         return 0;
 726                 }
 727         }
 728
 729         /*
 730          * We set it one so that the last submit_bio does not have to call
 731          * atomic_inc.
 732          */
 733         atomic_set(&pending_req->pendcnt, nbio);
 734
 735         /* Get a reference count for the disk queue and start sending I/O */
 736         blk_start_plug(&plug);
 737
 738         for (i = 0; i < nbio; i++)
 739                 submit_bio(operation, biolist[i]);
 740
 741         /* Let the I/Os go.. */
 742         blk_finish_plug(&plug);
 743
 744         if (operation == READ)
 745                 blkif->st_rd_sect += preq.nr_sects;
 746         else if (operation == WRITE || operation == WRITE_FLUSH)
 747                 blkif->st_wr_sect += preq.nr_sects;
 748
 749         return 0;
 750
 751  fail_flush:
 752         xen_blkbk_unmap(pending_req);
 753  fail_response:
 754         /* Haven't submitted any bio's yet. */
 755         make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
 756         free_req(pending_req);
 757         msleep(1); /* back off a bit */
 758         return -EIO;
 759
 760  fail_put_bio:
 761         for (i = 0; i < nbio; i++)
 762                 bio_put(biolist[i]);
 763         __end_block_io_op(pending_req, -EINVAL);
 764         msleep(1); /* back off a bit */
 765         return -EIO;
 766 }
 767
 768
 769
 770 /*
 771  * Put a response on the ring on how the operation fared.
 772  */
 773 static void make_response(struct xen_blkif *blkif, u64 id,
 774                           unsigned short op, int st)
 775 {
 776         struct blkif_response  resp;
 777         unsigned long     flags;
 778         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 779         int notify;
 780
 781         resp.id        = id;
 782         resp.operation = op;
 783         resp.status    = st;
 784
 785         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
 786         /* Place on the response ring for the relevant domain. */
 787         switch (blkif->blk_protocol) {
 788         case BLKIF_PROTOCOL_NATIVE:
 789                 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
 790                        &resp, sizeof(resp));
 791                 break;
 792         case BLKIF_PROTOCOL_X86_32:
 793                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
 794                        &resp, sizeof(resp));
 795                 break;
 796         case BLKIF_PROTOCOL_X86_64:
 797                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
 798                        &resp, sizeof(resp));
 799                 break;
 800         default:
 801                 BUG();
 802         }
 803         blk_rings->common.rsp_prod_pvt++;
 804         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
 805         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
 806         if (notify)
 807                 notify_remote_via_irq(blkif->irq);
 808 }
 809
 810 static int __init xen_blkif_init(void)
 811 {
 812         int i, mmap_pages;
 813         int rc = 0;
 814
 815         if (!xen_pv_domain())
 816                 return -ENODEV;
 817
 818         blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
 819         if (!blkbk) {
 820                 pr_alert(DRV_PFX "%s: out of memory!\n", __func__);
 821                 return -ENOMEM;
 822         }
 823
 824         mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
 825
 826         blkbk->pending_reqs          = kzalloc(sizeof(blkbk->pending_reqs[0]) *
 827                                         xen_blkif_reqs, GFP_KERNEL);
 828         blkbk->pending_grant_handles = kmalloc(sizeof(blkbk->pending_grant_handles[0]) *
 829                                         mmap_pages, GFP_KERNEL);
 830         blkbk->pending_pages         = kzalloc(sizeof(blkbk->pending_pages[0]) *
 831                                         mmap_pages, GFP_KERNEL);
 832
 833         if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
 834             !blkbk->pending_pages) {
 835                 rc = -ENOMEM;
 836                 goto out_of_memory;
 837         }
 838
 839         for (i = 0; i < mmap_pages; i++) {
 840                 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
 841                 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
 842                 if (blkbk->pending_pages[i] == NULL) {
 843                         rc = -ENOMEM;
 844                         goto out_of_memory;
 845                 }
 846         }
 847         rc = xen_blkif_interface_init();
 848         if (rc)
 849                 goto failed_init;
 850
 851         INIT_LIST_HEAD(&blkbk->pending_free);
 852         spin_lock_init(&blkbk->pending_free_lock);
 853         init_waitqueue_head(&blkbk->pending_free_wq);
 854
 855         for (i = 0; i < xen_blkif_reqs; i++)
 856                 list_add_tail(&blkbk->pending_reqs[i].free_list,
 857                               &blkbk->pending_free);
 858
 859         rc = xen_blkif_xenbus_init();
 860         if (rc)
 861                 goto failed_init;
 862
 863         return 0;
 864
 865  out_of_memory:
 866         pr_alert(DRV_PFX "%s: out of memory\n", __func__);
 867  failed_init:
 868         kfree(blkbk->pending_reqs);
 869         kfree(blkbk->pending_grant_handles);
 870         if (blkbk->pending_pages) {
 871                 for (i = 0; i < mmap_pages; i++) {
 872                         if (blkbk->pending_pages[i])
 873                                 __free_page(blkbk->pending_pages[i]);
 874                 }
 875                 kfree(blkbk->pending_pages);
 876         }
 877         kfree(blkbk);
 878         blkbk = NULL;
 879         return rc;
 880 }
 881
 882 module_init(xen_blkif_init);
 883
 884 MODULE_LICENSE("Dual BSD/GPL");
 885 MODULE_ALIAS("xen-backend:vbd");