drivers/xen/blkback/blkback.c

   1 /******************************************************************************
   2  * arch/xen/drivers/blkif/backend/main.c
   3  *
   4  * Back-end of the driver for virtual block devices. This portion of the
   5  * driver exports a 'unified' block-device interface that can be accessed
   6  * by any operating system that implements a compatible front end. A
   7  * reference front-end implementation can be found in:
   8  *  arch/xen/drivers/blkif/frontend
   9  *
  10  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
  11  * Copyright (c) 2005, Christopher Clark
  12  *
  13  * This program is free software; you can redistribute it and/or
  14  * modify it under the terms of the GNU General Public License version 2
  15  * as published by the Free Software Foundation; or, when distributed
  16  * separately from the Linux kernel or incorporated into other
  17  * software packages, subject to the following license:
  18  *
  19  * Permission is hereby granted, free of charge, to any person obtaining a copy
  20  * of this source file (the "Software"), to deal in the Software without
  21  * restriction, including without limitation the rights to use, copy, modify,
  22  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  23  * and to permit persons to whom the Software is furnished to do so, subject to
  24  * the following conditions:
  25  *
  26  * The above copyright notice and this permission notice shall be included in
  27  * all copies or substantial portions of the Software.
  28  *
  29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  30  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  31  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  32  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  33  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  34  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  35  * IN THE SOFTWARE.
  36  */
  37
  38 #include <linux/spinlock.h>
  39 #include <linux/kthread.h>
  40 #include <linux/list.h>
  41 #include <xen/balloon.h>
  42 #include <asm/hypervisor.h>
  43 #include <asm/hypercall.h>
  44 #include "common.h"
  45
  46 /*
  47  * These are rather arbitrary. They are fairly large because adjacent requests
  48  * pulled from a communication ring are quite likely to end up being part of
  49  * the same scatter/gather request at the disc.
  50  *
  51  * ** TRY INCREASING 'blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
  52  *
  53  * This will increase the chances of being able to write whole tracks.
  54  * 64 should be enough to keep us competitive with Linux.
  55  */
  56 static int blkif_reqs = 64;
  57 module_param_named(reqs, blkif_reqs, int, 0);
  58 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
  59
  60 /* Run-time switchable: /sys/module/blkback/parameters/ */
  61 static unsigned int log_stats = 0;
  62 static unsigned int debug_lvl = 0;
  63 module_param(log_stats, int, 0644);
  64 module_param(debug_lvl, int, 0644);
  65
  66 /*
  67  * Each outstanding request that we've passed to the lower device layers has a
  68  * 'pending_req' allocated to it. Each buffer_head that completes decrements
  69  * the pendcnt towards zero. When it hits zero, the specified domain has a
  70  * response queued for it, with the saved 'id' passed back.
  71  */
  72 typedef struct {
  73         blkif_t       *blkif;
  74         unsigned long  id;
  75         int            nr_pages;
  76         atomic_t       pendcnt;
  77         unsigned short operation;
  78         int            status;
  79         struct list_head free_list;
  80 } pending_req_t;
  81
  82 static pending_req_t *pending_reqs;
  83 static struct list_head pending_free;
  84 static DEFINE_SPINLOCK(pending_free_lock);
  85 static DECLARE_WAIT_QUEUE_HEAD(pending_free_wq);
  86
  87 #define BLKBACK_INVALID_HANDLE (~0)
  88
  89 static struct page **pending_pages;
  90 static grant_handle_t *pending_grant_handles;
  91
  92 static inline int vaddr_pagenr(pending_req_t *req, int seg)
  93 {
  94         return (req - pending_reqs) * BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
  95 }
  96
  97 static inline unsigned long vaddr(pending_req_t *req, int seg)
  98 {
  99         unsigned long pfn = page_to_pfn(pending_pages[vaddr_pagenr(req, seg)]);
 100         return (unsigned long)pfn_to_kaddr(pfn);
 101 }
 102
 103 #define pending_handle(_req, _seg) \
 104         (pending_grant_handles[vaddr_pagenr(_req, _seg)])
 105
 106
 107 static int do_block_io_op(blkif_t *blkif);
 108 static void dispatch_rw_block_io(blkif_t *blkif,
 109                                  blkif_request_t *req,
 110                                  pending_req_t *pending_req);
 111 static void make_response(blkif_t *blkif, unsigned long id,
 112                           unsigned short op, int st);
 113
 114 /******************************************************************
 115  * misc small helpers
 116  */
 117 static pending_req_t* alloc_req(void)
 118 {
 119         pending_req_t *req = NULL;
 120         unsigned long flags;
 121
 122         spin_lock_irqsave(&pending_free_lock, flags);
 123         if (!list_empty(&pending_free)) {
 124                 req = list_entry(pending_free.next, pending_req_t, free_list);
 125                 list_del(&req->free_list);
 126         }
 127         spin_unlock_irqrestore(&pending_free_lock, flags);
 128         return req;
 129 }
 130
 131 static void free_req(pending_req_t *req)
 132 {
 133         unsigned long flags;
 134         int was_empty;
 135
 136         spin_lock_irqsave(&pending_free_lock, flags);
 137         was_empty = list_empty(&pending_free);
 138         list_add(&req->free_list, &pending_free);
 139         spin_unlock_irqrestore(&pending_free_lock, flags);
 140         if (was_empty)
 141                 wake_up(&pending_free_wq);
 142 }
 143
 144 static void unplug_queue(blkif_t *blkif)
 145 {
 146         if (blkif->plug == NULL)
 147                 return;
 148         if (blkif->plug->unplug_fn)
 149                 blkif->plug->unplug_fn(blkif->plug);
 150         blk_put_queue(blkif->plug);
 151         blkif->plug = NULL;
 152 }
 153
 154 static void plug_queue(blkif_t *blkif, struct bio *bio)
 155 {
 156         request_queue_t *q = bdev_get_queue(bio->bi_bdev);
 157
 158         if (q == blkif->plug)
 159                 return;
 160         unplug_queue(blkif);
 161         blk_get_queue(q);
 162         blkif->plug = q;
 163 }
 164
 165 static void fast_flush_area(pending_req_t *req)
 166 {
 167         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 168         unsigned int i, invcount = 0;
 169         grant_handle_t handle;
 170         int ret;
 171
 172         for (i = 0; i < req->nr_pages; i++) {
 173                 handle = pending_handle(req, i);
 174                 if (handle == BLKBACK_INVALID_HANDLE)
 175                         continue;
 176                 gnttab_set_unmap_op(&unmap[i], vaddr(req, i), GNTMAP_host_map,
 177                                     handle);
 178                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
 179                 invcount++;
 180         }
 181
 182         ret = HYPERVISOR_grant_table_op(
 183                 GNTTABOP_unmap_grant_ref, unmap, invcount);
 184         BUG_ON(ret);
 185 }
 186
 187 /******************************************************************
 188  * SCHEDULER FUNCTIONS
 189  */
 190
 191 static void print_stats(blkif_t *blkif)
 192 {
 193         printk(KERN_DEBUG "%s: oo %3d  |  rd %4d  |  wr %4d\n",
 194                current->comm, blkif->st_oo_req,
 195                blkif->st_rd_req, blkif->st_wr_req);
 196         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 197         blkif->st_rd_req = 0;
 198         blkif->st_wr_req = 0;
 199         blkif->st_oo_req = 0;
 200 }
 201
 202 int blkif_schedule(void *arg)
 203 {
 204         blkif_t *blkif = arg;
 205
 206         blkif_get(blkif);
 207
 208         if (debug_lvl)
 209                 printk(KERN_DEBUG "%s: started\n", current->comm);
 210
 211         while (!kthread_should_stop()) {
 212                 wait_event_interruptible(
 213                         blkif->wq,
 214                         blkif->waiting_reqs || kthread_should_stop());
 215                 wait_event_interruptible(
 216                         pending_free_wq,
 217                         !list_empty(&pending_free) || kthread_should_stop());
 218
 219                 blkif->waiting_reqs = 0;
 220                 smp_mb(); /* clear flag *before* checking for work */
 221
 222                 if (do_block_io_op(blkif))
 223                         blkif->waiting_reqs = 1;
 224                 unplug_queue(blkif);
 225
 226                 if (log_stats && time_after(jiffies, blkif->st_print))
 227                         print_stats(blkif);
 228         }
 229
 230         if (log_stats)
 231                 print_stats(blkif);
 232         if (debug_lvl)
 233                 printk(KERN_DEBUG "%s: exiting\n", current->comm);
 234
 235         blkif->xenblkd = NULL;
 236         blkif_put(blkif);
 237
 238         return 0;
 239 }
 240
 241 /******************************************************************
 242  * COMPLETION CALLBACK -- Called as bh->b_end_io()
 243  */
 244
 245 static void __end_block_io_op(pending_req_t *pending_req, int uptodate)
 246 {
 247         /* An error fails the entire request. */
 248         if (!uptodate) {
 249                 DPRINTK("Buffer not up-to-date at end of operation\n");
 250                 pending_req->status = BLKIF_RSP_ERROR;
 251         }
 252
 253         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 254                 fast_flush_area(pending_req);
 255                 make_response(pending_req->blkif, pending_req->id,
 256                               pending_req->operation, pending_req->status);
 257                 blkif_put(pending_req->blkif);
 258                 free_req(pending_req);
 259         }
 260 }
 261
 262 static int end_block_io_op(struct bio *bio, unsigned int done, int error)
 263 {
 264         if (bio->bi_size != 0)
 265                 return 1;
 266         __end_block_io_op(bio->bi_private, !error);
 267         bio_put(bio);
 268         return error;
 269 }
 270
 271
 272 /******************************************************************************
 273  * NOTIFICATION FROM GUEST OS.
 274  */
 275
 276 static void blkif_notify_work(blkif_t *blkif)
 277 {
 278         blkif->waiting_reqs = 1;
 279         wake_up(&blkif->wq);
 280 }
 281
 282 irqreturn_t blkif_be_int(int irq, void *dev_id)
 283 {
 284         blkif_notify_work(dev_id);
 285         return IRQ_HANDLED;
 286 }
 287
 288
 289
 290 /******************************************************************
 291  * DOWNWARD CALLS -- These interface with the block-device layer proper.
 292  */
 293
 294 static int do_block_io_op(blkif_t *blkif)
 295 {
 296         blkif_back_ring_t *blk_ring = &blkif->blk_ring;
 297         blkif_request_t req;
 298         pending_req_t *pending_req;
 299         RING_IDX rc, rp;
 300         int more_to_do = 0;
 301
 302         rc = blk_ring->req_cons;
 303         rp = blk_ring->sring->req_prod;
 304         rmb(); /* Ensure we see queued requests up to 'rp'. */
 305
 306         while ((rc != rp) && !RING_REQUEST_CONS_OVERFLOW(blk_ring, rc)) {
 307
 308                 pending_req = alloc_req();
 309                 if (NULL == pending_req) {
 310                         blkif->st_oo_req++;
 311                         more_to_do = 1;
 312                         break;
 313                 }
 314
 315                 memcpy(&req, RING_GET_REQUEST(blk_ring, rc), sizeof(req));
 316                 blk_ring->req_cons = ++rc; /* before make_response() */
 317
 318                 switch (req.operation) {
 319                 case BLKIF_OP_READ:
 320                         blkif->st_rd_req++;
 321                         dispatch_rw_block_io(blkif, &req, pending_req);
 322                         break;
 323                 case BLKIF_OP_WRITE:
 324                         blkif->st_wr_req++;
 325                         dispatch_rw_block_io(blkif, &req, pending_req);
 326                         break;
 327                 default:
 328                         DPRINTK("error: unknown block io operation [%d]\n",
 329                                 req.operation);
 330                         make_response(blkif, req.id, req.operation,
 331                                       BLKIF_RSP_ERROR);
 332                         free_req(pending_req);
 333                         break;
 334                 }
 335         }
 336         return more_to_do;
 337 }
 338
 339 static void dispatch_rw_block_io(blkif_t *blkif,
 340                                  blkif_request_t *req,
 341                                  pending_req_t *pending_req)
 342 {
 343         extern void ll_rw_block(int rw, int nr, struct buffer_head * bhs[]);
 344         int operation = (req->operation == BLKIF_OP_WRITE) ? WRITE : READ;
 345         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 346         struct phys_req preq;
 347         struct {
 348                 unsigned long buf; unsigned int nsec;
 349         } seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 350         unsigned int nseg;
 351         struct bio *bio = NULL, *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 352         int ret, i, nbio = 0;
 353
 354         /* Check that number of segments is sane. */
 355         nseg = req->nr_segments;
 356         if (unlikely(nseg == 0) ||
 357             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
 358                 DPRINTK("Bad number of segments in request (%d)\n", nseg);
 359                 goto fail_response;
 360         }
 361
 362         preq.dev           = req->handle;
 363         preq.sector_number = req->sector_number;
 364         preq.nr_sects      = 0;
 365
 366         pending_req->blkif     = blkif;
 367         pending_req->id        = req->id;
 368         pending_req->operation = operation;
 369         pending_req->status    = BLKIF_RSP_OKAY;
 370         pending_req->nr_pages  = nseg;
 371
 372         for (i = 0; i < nseg; i++) {
 373                 uint32_t flags;
 374
 375                 seg[i].nsec = req->seg[i].last_sect -
 376                         req->seg[i].first_sect + 1;
 377
 378                 if ((req->seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
 379                     (req->seg[i].last_sect < req->seg[i].first_sect))
 380                         goto fail_response;
 381                 preq.nr_sects += seg[i].nsec;
 382
 383                 flags = GNTMAP_host_map;
 384                 if ( operation == WRITE )
 385                         flags |= GNTMAP_readonly;
 386                 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
 387                                   req->seg[i].gref, blkif->domid);
 388         }
 389
 390         ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
 391         BUG_ON(ret);
 392
 393         for (i = 0; i < nseg; i++) {
 394                 if (unlikely(map[i].status != 0)) {
 395                         DPRINTK("invalid buffer -- could not remap it\n");
 396                         map[i].handle = BLKBACK_INVALID_HANDLE;
 397                         ret |= 1;
 398                 }
 399
 400                 pending_handle(pending_req, i) = map[i].handle;
 401
 402                 if (ret)
 403                         continue;
 404
 405                 set_phys_to_machine(__pa(vaddr(
 406                         pending_req, i)) >> PAGE_SHIFT,
 407                         FOREIGN_FRAME(map[i].dev_bus_addr >> PAGE_SHIFT));
 408                 seg[i].buf  = map[i].dev_bus_addr |
 409                         (req->seg[i].first_sect << 9);
 410         }
 411
 412         if (ret)
 413                 goto fail_flush;
 414
 415         if (vbd_translate(&preq, blkif, operation) != 0) {
 416                 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
 417                         operation == READ ? "read" : "write",
 418                         preq.sector_number,
 419                         preq.sector_number + preq.nr_sects, preq.dev);
 420                 goto fail_flush;
 421         }
 422
 423         for (i = 0; i < nseg; i++) {
 424                 if (((int)preq.sector_number|(int)seg[i].nsec) &
 425                     ((bdev_hardsect_size(preq.bdev) >> 9) - 1)) {
 426                         DPRINTK("Misaligned I/O request from domain %d",
 427                                 blkif->domid);
 428                         goto fail_put_bio;
 429                 }
 430
 431                 while ((bio == NULL) ||
 432                        (bio_add_page(bio,
 433                                      virt_to_page(vaddr(pending_req, i)),
 434                                      seg[i].nsec << 9,
 435                                      seg[i].buf & ~PAGE_MASK) == 0)) {
 436                         bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, nseg-i);
 437                         if (unlikely(bio == NULL))
 438                                 goto fail_put_bio;
 439
 440                         bio->bi_bdev    = preq.bdev;
 441                         bio->bi_private = pending_req;
 442                         bio->bi_end_io  = end_block_io_op;
 443                         bio->bi_sector  = preq.sector_number;
 444                 }
 445
 446                 preq.sector_number += seg[i].nsec;
 447         }
 448
 449         plug_queue(blkif, bio);
 450         atomic_set(&pending_req->pendcnt, nbio);
 451         blkif_get(blkif);
 452
 453         for (i = 0; i < nbio; i++)
 454                 submit_bio(operation, biolist[i]);
 455
 456         return;
 457
 458  fail_put_bio:
 459         for (i = 0; i < (nbio-1); i++)
 460                 bio_put(biolist[i]);
 461  fail_flush:
 462         fast_flush_area(pending_req);
 463  fail_response:
 464         make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
 465         free_req(pending_req);
 466 }
 467
 468
 469
 470 /******************************************************************
 471  * MISCELLANEOUS SETUP / TEARDOWN / DEBUGGING
 472  */
 473
 474
 475 static void make_response(blkif_t *blkif, unsigned long id,
 476                           unsigned short op, int st)
 477 {
 478         blkif_response_t *resp;
 479         unsigned long     flags;
 480         blkif_back_ring_t *blk_ring = &blkif->blk_ring;
 481         int more_to_do = 0;
 482         int notify;
 483
 484         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
 485
 486         /* Place on the response ring for the relevant domain. */
 487         resp = RING_GET_RESPONSE(blk_ring, blk_ring->rsp_prod_pvt);
 488         resp->id        = id;
 489         resp->operation = op;
 490         resp->status    = st;
 491         blk_ring->rsp_prod_pvt++;
 492         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(blk_ring, notify);
 493
 494         if (blk_ring->rsp_prod_pvt == blk_ring->req_cons) {
 495                 /*
 496                  * Tail check for pending requests. Allows frontend to avoid
 497                  * notifications if requests are already in flight (lower
 498                  * overheads and promotes batching).
 499                  */
 500                 RING_FINAL_CHECK_FOR_REQUESTS(blk_ring, more_to_do);
 501
 502         } else if (RING_HAS_UNCONSUMED_REQUESTS(blk_ring)) {
 503                 more_to_do = 1;
 504
 505         }
 506         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
 507
 508         if (more_to_do)
 509                 blkif_notify_work(blkif);
 510         if (notify)
 511                 notify_remote_via_irq(blkif->irq);
 512 }
 513
 514 static int __init blkif_init(void)
 515 {
 516         int i, mmap_pages;
 517
 518         if (!is_running_on_xen())
 519                 return -ENODEV;
 520
 521         mmap_pages = blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
 522
 523         pending_reqs          = kmalloc(sizeof(pending_reqs[0]) *
 524                                         blkif_reqs, GFP_KERNEL);
 525         pending_grant_handles = kmalloc(sizeof(pending_grant_handles[0]) *
 526                                         mmap_pages, GFP_KERNEL);
 527         pending_pages         = alloc_empty_pages_and_pagevec(mmap_pages);
 528
 529         if (!pending_reqs || !pending_grant_handles || !pending_pages)
 530                 goto out_of_memory;
 531
 532         for (i = 0; i < mmap_pages; i++)
 533                 pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
 534
 535         blkif_interface_init();
 536
 537         memset(pending_reqs, 0, sizeof(pending_reqs));
 538         INIT_LIST_HEAD(&pending_free);
 539
 540         for (i = 0; i < blkif_reqs; i++)
 541                 list_add_tail(&pending_reqs[i].free_list, &pending_free);
 542
 543         blkif_xenbus_init();
 544
 545         return 0;
 546
 547  out_of_memory:
 548         kfree(pending_reqs);
 549         kfree(pending_grant_handles);
 550         free_empty_pages_and_pagevec(pending_pages, mmap_pages);
 551         printk("%s: out of memory\n", __FUNCTION__);
 552         return -ENOMEM;
 553 }
 554
 555 module_init(blkif_init);
 556
 557 MODULE_LICENSE("Dual BSD/GPL");