linux/fs/xfs/xfs_buf.c
<<
>>
Prefs
   1/*
   2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include <linux/stddef.h>
  20#include <linux/errno.h>
  21#include <linux/gfp.h>
  22#include <linux/pagemap.h>
  23#include <linux/init.h>
  24#include <linux/vmalloc.h>
  25#include <linux/bio.h>
  26#include <linux/sysctl.h>
  27#include <linux/proc_fs.h>
  28#include <linux/workqueue.h>
  29#include <linux/percpu.h>
  30#include <linux/blkdev.h>
  31#include <linux/hash.h>
  32#include <linux/kthread.h>
  33#include <linux/migrate.h>
  34#include <linux/backing-dev.h>
  35#include <linux/freezer.h>
  36
  37#include "xfs_sb.h"
  38#include "xfs_log.h"
  39#include "xfs_ag.h"
  40#include "xfs_mount.h"
  41#include "xfs_trace.h"
  42
  43static kmem_zone_t *xfs_buf_zone;
  44
  45static struct workqueue_struct *xfslogd_workqueue;
  46
  47#ifdef XFS_BUF_LOCK_TRACKING
  48# define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  49# define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  50# define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  51#else
  52# define XB_SET_OWNER(bp)       do { } while (0)
  53# define XB_CLEAR_OWNER(bp)     do { } while (0)
  54# define XB_GET_OWNER(bp)       do { } while (0)
  55#endif
  56
  57#define xb_to_gfp(flags) \
  58        ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
  59
  60
  61static inline int
  62xfs_buf_is_vmapped(
  63        struct xfs_buf  *bp)
  64{
  65        /*
  66         * Return true if the buffer is vmapped.
  67         *
  68         * b_addr is null if the buffer is not mapped, but the code is clever
  69         * enough to know it doesn't have to map a single page, so the check has
  70         * to be both for b_addr and bp->b_page_count > 1.
  71         */
  72        return bp->b_addr && bp->b_page_count > 1;
  73}
  74
  75static inline int
  76xfs_buf_vmap_len(
  77        struct xfs_buf  *bp)
  78{
  79        return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
  80}
  81
  82/*
  83 * xfs_buf_lru_add - add a buffer to the LRU.
  84 *
  85 * The LRU takes a new reference to the buffer so that it will only be freed
  86 * once the shrinker takes the buffer off the LRU.
  87 */
  88STATIC void
  89xfs_buf_lru_add(
  90        struct xfs_buf  *bp)
  91{
  92        struct xfs_buftarg *btp = bp->b_target;
  93
  94        spin_lock(&btp->bt_lru_lock);
  95        if (list_empty(&bp->b_lru)) {
  96                atomic_inc(&bp->b_hold);
  97                list_add_tail(&bp->b_lru, &btp->bt_lru);
  98                btp->bt_lru_nr++;
  99                bp->b_lru_flags &= ~_XBF_LRU_DISPOSE;
 100        }
 101        spin_unlock(&btp->bt_lru_lock);
 102}
 103
 104/*
 105 * xfs_buf_lru_del - remove a buffer from the LRU
 106 *
 107 * The unlocked check is safe here because it only occurs when there are not
 108 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
 109 * to optimise the shrinker removing the buffer from the LRU and calling
 110 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
 111 * bt_lru_lock.
 112 */
 113STATIC void
 114xfs_buf_lru_del(
 115        struct xfs_buf  *bp)
 116{
 117        struct xfs_buftarg *btp = bp->b_target;
 118
 119        if (list_empty(&bp->b_lru))
 120                return;
 121
 122        spin_lock(&btp->bt_lru_lock);
 123        if (!list_empty(&bp->b_lru)) {
 124                list_del_init(&bp->b_lru);
 125                btp->bt_lru_nr--;
 126        }
 127        spin_unlock(&btp->bt_lru_lock);
 128}
 129
 130/*
 131 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
 132 * b_lru_ref count so that the buffer is freed immediately when the buffer
 133 * reference count falls to zero. If the buffer is already on the LRU, we need
 134 * to remove the reference that LRU holds on the buffer.
 135 *
 136 * This prevents build-up of stale buffers on the LRU.
 137 */
 138void
 139xfs_buf_stale(
 140        struct xfs_buf  *bp)
 141{
 142        ASSERT(xfs_buf_islocked(bp));
 143
 144        bp->b_flags |= XBF_STALE;
 145
 146        /*
 147         * Clear the delwri status so that a delwri queue walker will not
 148         * flush this buffer to disk now that it is stale. The delwri queue has
 149         * a reference to the buffer, so this is safe to do.
 150         */
 151        bp->b_flags &= ~_XBF_DELWRI_Q;
 152
 153        atomic_set(&(bp)->b_lru_ref, 0);
 154        if (!list_empty(&bp->b_lru)) {
 155                struct xfs_buftarg *btp = bp->b_target;
 156
 157                spin_lock(&btp->bt_lru_lock);
 158                if (!list_empty(&bp->b_lru) &&
 159                    !(bp->b_lru_flags & _XBF_LRU_DISPOSE)) {
 160                        list_del_init(&bp->b_lru);
 161                        btp->bt_lru_nr--;
 162                        atomic_dec(&bp->b_hold);
 163                }
 164                spin_unlock(&btp->bt_lru_lock);
 165        }
 166        ASSERT(atomic_read(&bp->b_hold) >= 1);
 167}
 168
 169static int
 170xfs_buf_get_maps(
 171        struct xfs_buf          *bp,
 172        int                     map_count)
 173{
 174        ASSERT(bp->b_maps == NULL);
 175        bp->b_map_count = map_count;
 176
 177        if (map_count == 1) {
 178                bp->b_maps = &bp->b_map;
 179                return 0;
 180        }
 181
 182        bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
 183                                KM_NOFS);
 184        if (!bp->b_maps)
 185                return ENOMEM;
 186        return 0;
 187}
 188
 189/*
 190 *      Frees b_pages if it was allocated.
 191 */
 192static void
 193xfs_buf_free_maps(
 194        struct xfs_buf  *bp)
 195{
 196        if (bp->b_maps != &bp->b_map) {
 197                kmem_free(bp->b_maps);
 198                bp->b_maps = NULL;
 199        }
 200}
 201
 202struct xfs_buf *
 203_xfs_buf_alloc(
 204        struct xfs_buftarg      *target,
 205        struct xfs_buf_map      *map,
 206        int                     nmaps,
 207        xfs_buf_flags_t         flags)
 208{
 209        struct xfs_buf          *bp;
 210        int                     error;
 211        int                     i;
 212
 213        bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
 214        if (unlikely(!bp))
 215                return NULL;
 216
 217        /*
 218         * We don't want certain flags to appear in b_flags unless they are
 219         * specifically set by later operations on the buffer.
 220         */
 221        flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
 222
 223        atomic_set(&bp->b_hold, 1);
 224        atomic_set(&bp->b_lru_ref, 1);
 225        init_completion(&bp->b_iowait);
 226        INIT_LIST_HEAD(&bp->b_lru);
 227        INIT_LIST_HEAD(&bp->b_list);
 228        RB_CLEAR_NODE(&bp->b_rbnode);
 229        sema_init(&bp->b_sema, 0); /* held, no waiters */
 230        XB_SET_OWNER(bp);
 231        bp->b_target = target;
 232        bp->b_flags = flags;
 233
 234        /*
 235         * Set length and io_length to the same value initially.
 236         * I/O routines should use io_length, which will be the same in
 237         * most cases but may be reset (e.g. XFS recovery).
 238         */
 239        error = xfs_buf_get_maps(bp, nmaps);
 240        if (error)  {
 241                kmem_zone_free(xfs_buf_zone, bp);
 242                return NULL;
 243        }
 244
 245        bp->b_bn = map[0].bm_bn;
 246        bp->b_length = 0;
 247        for (i = 0; i < nmaps; i++) {
 248                bp->b_maps[i].bm_bn = map[i].bm_bn;
 249                bp->b_maps[i].bm_len = map[i].bm_len;
 250                bp->b_length += map[i].bm_len;
 251        }
 252        bp->b_io_length = bp->b_length;
 253
 254        atomic_set(&bp->b_pin_count, 0);
 255        init_waitqueue_head(&bp->b_waiters);
 256
 257        XFS_STATS_INC(xb_create);
 258        trace_xfs_buf_init(bp, _RET_IP_);
 259
 260        return bp;
 261}
 262
 263/*
 264 *      Allocate a page array capable of holding a specified number
 265 *      of pages, and point the page buf at it.
 266 */
 267STATIC int
 268_xfs_buf_get_pages(
 269        xfs_buf_t               *bp,
 270        int                     page_count,
 271        xfs_buf_flags_t         flags)
 272{
 273        /* Make sure that we have a page list */
 274        if (bp->b_pages == NULL) {
 275                bp->b_page_count = page_count;
 276                if (page_count <= XB_PAGES) {
 277                        bp->b_pages = bp->b_page_array;
 278                } else {
 279                        bp->b_pages = kmem_alloc(sizeof(struct page *) *
 280                                                 page_count, KM_NOFS);
 281                        if (bp->b_pages == NULL)
 282                                return -ENOMEM;
 283                }
 284                memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 285        }
 286        return 0;
 287}
 288
 289/*
 290 *      Frees b_pages if it was allocated.
 291 */
 292STATIC void
 293_xfs_buf_free_pages(
 294        xfs_buf_t       *bp)
 295{
 296        if (bp->b_pages != bp->b_page_array) {
 297                kmem_free(bp->b_pages);
 298                bp->b_pages = NULL;
 299        }
 300}
 301
 302/*
 303 *      Releases the specified buffer.
 304 *
 305 *      The modification state of any associated pages is left unchanged.
 306 *      The buffer most not be on any hash - use xfs_buf_rele instead for
 307 *      hashed and refcounted buffers
 308 */
 309void
 310xfs_buf_free(
 311        xfs_buf_t               *bp)
 312{
 313        trace_xfs_buf_free(bp, _RET_IP_);
 314
 315        ASSERT(list_empty(&bp->b_lru));
 316
 317        if (bp->b_flags & _XBF_PAGES) {
 318                uint            i;
 319
 320                if (xfs_buf_is_vmapped(bp))
 321                        vm_unmap_ram(bp->b_addr - bp->b_offset,
 322                                        bp->b_page_count);
 323
 324                for (i = 0; i < bp->b_page_count; i++) {
 325                        struct page     *page = bp->b_pages[i];
 326
 327                        __free_page(page);
 328                }
 329        } else if (bp->b_flags & _XBF_KMEM)
 330                kmem_free(bp->b_addr);
 331        _xfs_buf_free_pages(bp);
 332        xfs_buf_free_maps(bp);
 333        kmem_zone_free(xfs_buf_zone, bp);
 334}
 335
 336/*
 337 * Allocates all the pages for buffer in question and builds it's page list.
 338 */
 339STATIC int
 340xfs_buf_allocate_memory(
 341        xfs_buf_t               *bp,
 342        uint                    flags)
 343{
 344        size_t                  size;
 345        size_t                  nbytes, offset;
 346        gfp_t                   gfp_mask = xb_to_gfp(flags);
 347        unsigned short          page_count, i;
 348        xfs_off_t               start, end;
 349        int                     error;
 350
 351        /*
 352         * for buffers that are contained within a single page, just allocate
 353         * the memory from the heap - there's no need for the complexity of
 354         * page arrays to keep allocation down to order 0.
 355         */
 356        size = BBTOB(bp->b_length);
 357        if (size < PAGE_SIZE) {
 358                bp->b_addr = kmem_alloc(size, KM_NOFS);
 359                if (!bp->b_addr) {
 360                        /* low memory - use alloc_page loop instead */
 361                        goto use_alloc_page;
 362                }
 363
 364                if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
 365                    ((unsigned long)bp->b_addr & PAGE_MASK)) {
 366                        /* b_addr spans two pages - use alloc_page instead */
 367                        kmem_free(bp->b_addr);
 368                        bp->b_addr = NULL;
 369                        goto use_alloc_page;
 370                }
 371                bp->b_offset = offset_in_page(bp->b_addr);
 372                bp->b_pages = bp->b_page_array;
 373                bp->b_pages[0] = virt_to_page(bp->b_addr);
 374                bp->b_page_count = 1;
 375                bp->b_flags |= _XBF_KMEM;
 376                return 0;
 377        }
 378
 379use_alloc_page:
 380        start = BBTOB(bp->b_map.bm_bn) >> PAGE_SHIFT;
 381        end = (BBTOB(bp->b_map.bm_bn + bp->b_length) + PAGE_SIZE - 1)
 382                                                                >> PAGE_SHIFT;
 383        page_count = end - start;
 384        error = _xfs_buf_get_pages(bp, page_count, flags);
 385        if (unlikely(error))
 386                return error;
 387
 388        offset = bp->b_offset;
 389        bp->b_flags |= _XBF_PAGES;
 390
 391        for (i = 0; i < bp->b_page_count; i++) {
 392                struct page     *page;
 393                uint            retries = 0;
 394retry:
 395                page = alloc_page(gfp_mask);
 396                if (unlikely(page == NULL)) {
 397                        if (flags & XBF_READ_AHEAD) {
 398                                bp->b_page_count = i;
 399                                error = ENOMEM;
 400                                goto out_free_pages;
 401                        }
 402
 403                        /*
 404                         * This could deadlock.
 405                         *
 406                         * But until all the XFS lowlevel code is revamped to
 407                         * handle buffer allocation failures we can't do much.
 408                         */
 409                        if (!(++retries % 100))
 410                                xfs_err(NULL,
 411                "possible memory allocation deadlock in %s (mode:0x%x)",
 412                                        __func__, gfp_mask);
 413
 414                        XFS_STATS_INC(xb_page_retries);
 415                        congestion_wait(BLK_RW_ASYNC, HZ/50);
 416                        goto retry;
 417                }
 418
 419                XFS_STATS_INC(xb_page_found);
 420
 421                nbytes = min_t(size_t, size, PAGE_SIZE - offset);
 422                size -= nbytes;
 423                bp->b_pages[i] = page;
 424                offset = 0;
 425        }
 426        return 0;
 427
 428out_free_pages:
 429        for (i = 0; i < bp->b_page_count; i++)
 430                __free_page(bp->b_pages[i]);
 431        return error;
 432}
 433
 434/*
 435 *      Map buffer into kernel address-space if necessary.
 436 */
 437STATIC int
 438_xfs_buf_map_pages(
 439        xfs_buf_t               *bp,
 440        uint                    flags)
 441{
 442        ASSERT(bp->b_flags & _XBF_PAGES);
 443        if (bp->b_page_count == 1) {
 444                /* A single page buffer is always mappable */
 445                bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 446        } else if (flags & XBF_UNMAPPED) {
 447                bp->b_addr = NULL;
 448        } else {
 449                int retried = 0;
 450
 451                do {
 452                        bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 453                                                -1, PAGE_KERNEL);
 454                        if (bp->b_addr)
 455                                break;
 456                        vm_unmap_aliases();
 457                } while (retried++ <= 1);
 458
 459                if (!bp->b_addr)
 460                        return -ENOMEM;
 461                bp->b_addr += bp->b_offset;
 462        }
 463
 464        return 0;
 465}
 466
 467/*
 468 *      Finding and Reading Buffers
 469 */
 470
 471/*
 472 *      Look up, and creates if absent, a lockable buffer for
 473 *      a given range of an inode.  The buffer is returned
 474 *      locked. No I/O is implied by this call.
 475 */
 476xfs_buf_t *
 477_xfs_buf_find(
 478        struct xfs_buftarg      *btp,
 479        struct xfs_buf_map      *map,
 480        int                     nmaps,
 481        xfs_buf_flags_t         flags,
 482        xfs_buf_t               *new_bp)
 483{
 484        size_t                  numbytes;
 485        struct xfs_perag        *pag;
 486        struct rb_node          **rbp;
 487        struct rb_node          *parent;
 488        xfs_buf_t               *bp;
 489        xfs_daddr_t             blkno = map[0].bm_bn;
 490        xfs_daddr_t             eofs;
 491        int                     numblks = 0;
 492        int                     i;
 493
 494        for (i = 0; i < nmaps; i++)
 495                numblks += map[i].bm_len;
 496        numbytes = BBTOB(numblks);
 497
 498        /* Check for IOs smaller than the sector size / not sector aligned */
 499        ASSERT(!(numbytes < (1 << btp->bt_sshift)));
 500        ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask));
 501
 502        /*
 503         * Corrupted block numbers can get through to here, unfortunately, so we
 504         * have to check that the buffer falls within the filesystem bounds.
 505         */
 506        eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
 507        if (blkno >= eofs) {
 508                /*
 509                 * XXX (dgc): we should really be returning EFSCORRUPTED here,
 510                 * but none of the higher level infrastructure supports
 511                 * returning a specific error on buffer lookup failures.
 512                 */
 513                xfs_alert(btp->bt_mount,
 514                          "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
 515                          __func__, blkno, eofs);
 516                return NULL;
 517        }
 518
 519        /* get tree root */
 520        pag = xfs_perag_get(btp->bt_mount,
 521                                xfs_daddr_to_agno(btp->bt_mount, blkno));
 522
 523        /* walk tree */
 524        spin_lock(&pag->pag_buf_lock);
 525        rbp = &pag->pag_buf_tree.rb_node;
 526        parent = NULL;
 527        bp = NULL;
 528        while (*rbp) {
 529                parent = *rbp;
 530                bp = rb_entry(parent, struct xfs_buf, b_rbnode);
 531
 532                if (blkno < bp->b_bn)
 533                        rbp = &(*rbp)->rb_left;
 534                else if (blkno > bp->b_bn)
 535                        rbp = &(*rbp)->rb_right;
 536                else {
 537                        /*
 538                         * found a block number match. If the range doesn't
 539                         * match, the only way this is allowed is if the buffer
 540                         * in the cache is stale and the transaction that made
 541                         * it stale has not yet committed. i.e. we are
 542                         * reallocating a busy extent. Skip this buffer and
 543                         * continue searching to the right for an exact match.
 544                         */
 545                        if (bp->b_length != numblks) {
 546                                ASSERT(bp->b_flags & XBF_STALE);
 547                                rbp = &(*rbp)->rb_right;
 548                                continue;
 549                        }
 550                        atomic_inc(&bp->b_hold);
 551                        goto found;
 552                }
 553        }
 554
 555        /* No match found */
 556        if (new_bp) {
 557                rb_link_node(&new_bp->b_rbnode, parent, rbp);
 558                rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
 559                /* the buffer keeps the perag reference until it is freed */
 560                new_bp->b_pag = pag;
 561                spin_unlock(&pag->pag_buf_lock);
 562        } else {
 563                XFS_STATS_INC(xb_miss_locked);
 564                spin_unlock(&pag->pag_buf_lock);
 565                xfs_perag_put(pag);
 566        }
 567        return new_bp;
 568
 569found:
 570        spin_unlock(&pag->pag_buf_lock);
 571        xfs_perag_put(pag);
 572
 573        if (!xfs_buf_trylock(bp)) {
 574                if (flags & XBF_TRYLOCK) {
 575                        xfs_buf_rele(bp);
 576                        XFS_STATS_INC(xb_busy_locked);
 577                        return NULL;
 578                }
 579                xfs_buf_lock(bp);
 580                XFS_STATS_INC(xb_get_locked_waited);
 581        }
 582
 583        /*
 584         * if the buffer is stale, clear all the external state associated with
 585         * it. We need to keep flags such as how we allocated the buffer memory
 586         * intact here.
 587         */
 588        if (bp->b_flags & XBF_STALE) {
 589                ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 590                bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
 591        }
 592
 593        trace_xfs_buf_find(bp, flags, _RET_IP_);
 594        XFS_STATS_INC(xb_get_locked);
 595        return bp;
 596}
 597
 598/*
 599 * Assembles a buffer covering the specified range. The code is optimised for
 600 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
 601 * more hits than misses.
 602 */
 603struct xfs_buf *
 604xfs_buf_get_map(
 605        struct xfs_buftarg      *target,
 606        struct xfs_buf_map      *map,
 607        int                     nmaps,
 608        xfs_buf_flags_t         flags)
 609{
 610        struct xfs_buf          *bp;
 611        struct xfs_buf          *new_bp;
 612        int                     error = 0;
 613
 614        bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
 615        if (likely(bp))
 616                goto found;
 617
 618        new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
 619        if (unlikely(!new_bp))
 620                return NULL;
 621
 622        error = xfs_buf_allocate_memory(new_bp, flags);
 623        if (error) {
 624                xfs_buf_free(new_bp);
 625                return NULL;
 626        }
 627
 628        bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
 629        if (!bp) {
 630                xfs_buf_free(new_bp);
 631                return NULL;
 632        }
 633
 634        if (bp != new_bp)
 635                xfs_buf_free(new_bp);
 636
 637found:
 638        if (!bp->b_addr) {
 639                error = _xfs_buf_map_pages(bp, flags);
 640                if (unlikely(error)) {
 641                        xfs_warn(target->bt_mount,
 642                                "%s: failed to map pages\n", __func__);
 643                        xfs_buf_relse(bp);
 644                        return NULL;
 645                }
 646        }
 647
 648        XFS_STATS_INC(xb_get);
 649        trace_xfs_buf_get(bp, flags, _RET_IP_);
 650        return bp;
 651}
 652
 653STATIC int
 654_xfs_buf_read(
 655        xfs_buf_t               *bp,
 656        xfs_buf_flags_t         flags)
 657{
 658        ASSERT(!(flags & XBF_WRITE));
 659        ASSERT(bp->b_map.bm_bn != XFS_BUF_DADDR_NULL);
 660
 661        bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
 662        bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 663
 664        xfs_buf_iorequest(bp);
 665        if (flags & XBF_ASYNC)
 666                return 0;
 667        return xfs_buf_iowait(bp);
 668}
 669
 670xfs_buf_t *
 671xfs_buf_read_map(
 672        struct xfs_buftarg      *target,
 673        struct xfs_buf_map      *map,
 674        int                     nmaps,
 675        xfs_buf_flags_t         flags)
 676{
 677        struct xfs_buf          *bp;
 678
 679        flags |= XBF_READ;
 680
 681        bp = xfs_buf_get_map(target, map, nmaps, flags);
 682        if (bp) {
 683                trace_xfs_buf_read(bp, flags, _RET_IP_);
 684
 685                if (!XFS_BUF_ISDONE(bp)) {
 686                        XFS_STATS_INC(xb_get_read);
 687                        _xfs_buf_read(bp, flags);
 688                } else if (flags & XBF_ASYNC) {
 689                        /*
 690                         * Read ahead call which is already satisfied,
 691                         * drop the buffer
 692                         */
 693                        xfs_buf_relse(bp);
 694                        return NULL;
 695                } else {
 696                        /* We do not want read in the flags */
 697                        bp->b_flags &= ~XBF_READ;
 698                }
 699        }
 700
 701        return bp;
 702}
 703
 704/*
 705 *      If we are not low on memory then do the readahead in a deadlock
 706 *      safe manner.
 707 */
 708void
 709xfs_buf_readahead_map(
 710        struct xfs_buftarg      *target,
 711        struct xfs_buf_map      *map,
 712        int                     nmaps)
 713{
 714        if (bdi_read_congested(target->bt_bdi))
 715                return;
 716
 717        xfs_buf_read_map(target, map, nmaps,
 718                     XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 719}
 720
 721/*
 722 * Read an uncached buffer from disk. Allocates and returns a locked
 723 * buffer containing the disk contents or nothing.
 724 */
 725struct xfs_buf *
 726xfs_buf_read_uncached(
 727        struct xfs_buftarg      *target,
 728        xfs_daddr_t             daddr,
 729        size_t                  numblks,
 730        int                     flags)
 731{
 732        xfs_buf_t               *bp;
 733        int                     error;
 734
 735        bp = xfs_buf_get_uncached(target, numblks, flags);
 736        if (!bp)
 737                return NULL;
 738
 739        /* set up the buffer for a read IO */
 740        ASSERT(bp->b_map_count == 1);
 741        bp->b_bn = daddr;
 742        bp->b_maps[0].bm_bn = daddr;
 743        bp->b_flags |= XBF_READ;
 744
 745        xfsbdstrat(target->bt_mount, bp);
 746        error = xfs_buf_iowait(bp);
 747        if (error) {
 748                xfs_buf_relse(bp);
 749                return NULL;
 750        }
 751        return bp;
 752}
 753
 754/*
 755 * Return a buffer allocated as an empty buffer and associated to external
 756 * memory via xfs_buf_associate_memory() back to it's empty state.
 757 */
 758void
 759xfs_buf_set_empty(
 760        struct xfs_buf          *bp,
 761        size_t                  numblks)
 762{
 763        if (bp->b_pages)
 764                _xfs_buf_free_pages(bp);
 765
 766        bp->b_pages = NULL;
 767        bp->b_page_count = 0;
 768        bp->b_addr = NULL;
 769        bp->b_length = numblks;
 770        bp->b_io_length = numblks;
 771
 772        ASSERT(bp->b_map_count == 1);
 773        bp->b_bn = XFS_BUF_DADDR_NULL;
 774        bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
 775        bp->b_maps[0].bm_len = bp->b_length;
 776}
 777
 778static inline struct page *
 779mem_to_page(
 780        void                    *addr)
 781{
 782        if ((!is_vmalloc_addr(addr))) {
 783                return virt_to_page(addr);
 784        } else {
 785                return vmalloc_to_page(addr);
 786        }
 787}
 788
 789int
 790xfs_buf_associate_memory(
 791        xfs_buf_t               *bp,
 792        void                    *mem,
 793        size_t                  len)
 794{
 795        int                     rval;
 796        int                     i = 0;
 797        unsigned long           pageaddr;
 798        unsigned long           offset;
 799        size_t                  buflen;
 800        int                     page_count;
 801
 802        pageaddr = (unsigned long)mem & PAGE_MASK;
 803        offset = (unsigned long)mem - pageaddr;
 804        buflen = PAGE_ALIGN(len + offset);
 805        page_count = buflen >> PAGE_SHIFT;
 806
 807        /* Free any previous set of page pointers */
 808        if (bp->b_pages)
 809                _xfs_buf_free_pages(bp);
 810
 811        bp->b_pages = NULL;
 812        bp->b_addr = mem;
 813
 814        rval = _xfs_buf_get_pages(bp, page_count, 0);
 815        if (rval)
 816                return rval;
 817
 818        bp->b_offset = offset;
 819
 820        for (i = 0; i < bp->b_page_count; i++) {
 821                bp->b_pages[i] = mem_to_page((void *)pageaddr);
 822                pageaddr += PAGE_SIZE;
 823        }
 824
 825        bp->b_io_length = BTOBB(len);
 826        bp->b_length = BTOBB(buflen);
 827
 828        return 0;
 829}
 830
 831xfs_buf_t *
 832xfs_buf_get_uncached(
 833        struct xfs_buftarg      *target,
 834        size_t                  numblks,
 835        int                     flags)
 836{
 837        unsigned long           page_count;
 838        int                     error, i;
 839        struct xfs_buf          *bp;
 840        DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
 841
 842        bp = _xfs_buf_alloc(target, &map, 1, 0);
 843        if (unlikely(bp == NULL))
 844                goto fail;
 845
 846        page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
 847        error = _xfs_buf_get_pages(bp, page_count, 0);
 848        if (error)
 849                goto fail_free_buf;
 850
 851        for (i = 0; i < page_count; i++) {
 852                bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 853                if (!bp->b_pages[i])
 854                        goto fail_free_mem;
 855        }
 856        bp->b_flags |= _XBF_PAGES;
 857
 858        error = _xfs_buf_map_pages(bp, 0);
 859        if (unlikely(error)) {
 860                xfs_warn(target->bt_mount,
 861                        "%s: failed to map pages\n", __func__);
 862                goto fail_free_mem;
 863        }
 864
 865        trace_xfs_buf_get_uncached(bp, _RET_IP_);
 866        return bp;
 867
 868 fail_free_mem:
 869        while (--i >= 0)
 870                __free_page(bp->b_pages[i]);
 871        _xfs_buf_free_pages(bp);
 872 fail_free_buf:
 873        xfs_buf_free_maps(bp);
 874        kmem_zone_free(xfs_buf_zone, bp);
 875 fail:
 876        return NULL;
 877}
 878
 879/*
 880 *      Increment reference count on buffer, to hold the buffer concurrently
 881 *      with another thread which may release (free) the buffer asynchronously.
 882 *      Must hold the buffer already to call this function.
 883 */
 884void
 885xfs_buf_hold(
 886        xfs_buf_t               *bp)
 887{
 888        trace_xfs_buf_hold(bp, _RET_IP_);
 889        atomic_inc(&bp->b_hold);
 890}
 891
 892/*
 893 *      Releases a hold on the specified buffer.  If the
 894 *      the hold count is 1, calls xfs_buf_free.
 895 */
 896void
 897xfs_buf_rele(
 898        xfs_buf_t               *bp)
 899{
 900        struct xfs_perag        *pag = bp->b_pag;
 901
 902        trace_xfs_buf_rele(bp, _RET_IP_);
 903
 904        if (!pag) {
 905                ASSERT(list_empty(&bp->b_lru));
 906                ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
 907                if (atomic_dec_and_test(&bp->b_hold))
 908                        xfs_buf_free(bp);
 909                return;
 910        }
 911
 912        ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
 913
 914        ASSERT(atomic_read(&bp->b_hold) > 0);
 915        if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
 916                if (!(bp->b_flags & XBF_STALE) &&
 917                           atomic_read(&bp->b_lru_ref)) {
 918                        xfs_buf_lru_add(bp);
 919                        spin_unlock(&pag->pag_buf_lock);
 920                } else {
 921                        xfs_buf_lru_del(bp);
 922                        ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
 923                        rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 924                        spin_unlock(&pag->pag_buf_lock);
 925                        xfs_perag_put(pag);
 926                        xfs_buf_free(bp);
 927                }
 928        }
 929}
 930
 931
 932/*
 933 *      Lock a buffer object, if it is not already locked.
 934 *
 935 *      If we come across a stale, pinned, locked buffer, we know that we are
 936 *      being asked to lock a buffer that has been reallocated. Because it is
 937 *      pinned, we know that the log has not been pushed to disk and hence it
 938 *      will still be locked.  Rather than continuing to have trylock attempts
 939 *      fail until someone else pushes the log, push it ourselves before
 940 *      returning.  This means that the xfsaild will not get stuck trying
 941 *      to push on stale inode buffers.
 942 */
 943int
 944xfs_buf_trylock(
 945        struct xfs_buf          *bp)
 946{
 947        int                     locked;
 948
 949        locked = down_trylock(&bp->b_sema) == 0;
 950        if (locked)
 951                XB_SET_OWNER(bp);
 952        else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 953                xfs_log_force(bp->b_target->bt_mount, 0);
 954
 955        trace_xfs_buf_trylock(bp, _RET_IP_);
 956        return locked;
 957}
 958
 959/*
 960 *      Lock a buffer object.
 961 *
 962 *      If we come across a stale, pinned, locked buffer, we know that we
 963 *      are being asked to lock a buffer that has been reallocated. Because
 964 *      it is pinned, we know that the log has not been pushed to disk and
 965 *      hence it will still be locked. Rather than sleeping until someone
 966 *      else pushes the log, push it ourselves before trying to get the lock.
 967 */
 968void
 969xfs_buf_lock(
 970        struct xfs_buf          *bp)
 971{
 972        trace_xfs_buf_lock(bp, _RET_IP_);
 973
 974        if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 975                xfs_log_force(bp->b_target->bt_mount, 0);
 976        down(&bp->b_sema);
 977        XB_SET_OWNER(bp);
 978
 979        trace_xfs_buf_lock_done(bp, _RET_IP_);
 980}
 981
 982void
 983xfs_buf_unlock(
 984        struct xfs_buf          *bp)
 985{
 986        XB_CLEAR_OWNER(bp);
 987        up(&bp->b_sema);
 988
 989        trace_xfs_buf_unlock(bp, _RET_IP_);
 990}
 991
 992STATIC void
 993xfs_buf_wait_unpin(
 994        xfs_buf_t               *bp)
 995{
 996        DECLARE_WAITQUEUE       (wait, current);
 997
 998        if (atomic_read(&bp->b_pin_count) == 0)
 999                return;
1000
1001        add_wait_queue(&bp->b_waiters, &wait);
1002        for (;;) {
1003                set_current_state(TASK_UNINTERRUPTIBLE);
1004                if (atomic_read(&bp->b_pin_count) == 0)
1005                        break;
1006                io_schedule();
1007        }
1008        remove_wait_queue(&bp->b_waiters, &wait);
1009        set_current_state(TASK_RUNNING);
1010}
1011
1012/*
1013 *      Buffer Utility Routines
1014 */
1015
1016STATIC void
1017xfs_buf_iodone_work(
1018        struct work_struct      *work)
1019{
1020        xfs_buf_t               *bp =
1021                container_of(work, xfs_buf_t, b_iodone_work);
1022
1023        if (bp->b_iodone)
1024                (*(bp->b_iodone))(bp);
1025        else if (bp->b_flags & XBF_ASYNC)
1026                xfs_buf_relse(bp);
1027}
1028
1029void
1030xfs_buf_ioend(
1031        xfs_buf_t               *bp,
1032        int                     schedule)
1033{
1034        trace_xfs_buf_iodone(bp, _RET_IP_);
1035
1036        bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1037        if (bp->b_error == 0)
1038                bp->b_flags |= XBF_DONE;
1039
1040        if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1041                if (schedule) {
1042                        INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1043                        queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1044                } else {
1045                        xfs_buf_iodone_work(&bp->b_iodone_work);
1046                }
1047        } else {
1048                complete(&bp->b_iowait);
1049        }
1050}
1051
1052void
1053xfs_buf_ioerror(
1054        xfs_buf_t               *bp,
1055        int                     error)
1056{
1057        ASSERT(error >= 0 && error <= 0xffff);
1058        bp->b_error = (unsigned short)error;
1059        trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1060}
1061
1062void
1063xfs_buf_ioerror_alert(
1064        struct xfs_buf          *bp,
1065        const char              *func)
1066{
1067        xfs_alert(bp->b_target->bt_mount,
1068"metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1069                (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1070}
1071
1072/*
1073 * Called when we want to stop a buffer from getting written or read.
1074 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1075 * so that the proper iodone callbacks get called.
1076 */
1077STATIC int
1078xfs_bioerror(
1079        xfs_buf_t *bp)
1080{
1081#ifdef XFSERRORDEBUG
1082        ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1083#endif
1084
1085        /*
1086         * No need to wait until the buffer is unpinned, we aren't flushing it.
1087         */
1088        xfs_buf_ioerror(bp, EIO);
1089
1090        /*
1091         * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1092         */
1093        XFS_BUF_UNREAD(bp);
1094        XFS_BUF_UNDONE(bp);
1095        xfs_buf_stale(bp);
1096
1097        xfs_buf_ioend(bp, 0);
1098
1099        return EIO;
1100}
1101
1102/*
1103 * Same as xfs_bioerror, except that we are releasing the buffer
1104 * here ourselves, and avoiding the xfs_buf_ioend call.
1105 * This is meant for userdata errors; metadata bufs come with
1106 * iodone functions attached, so that we can track down errors.
1107 */
1108STATIC int
1109xfs_bioerror_relse(
1110        struct xfs_buf  *bp)
1111{
1112        int64_t         fl = bp->b_flags;
1113        /*
1114         * No need to wait until the buffer is unpinned.
1115         * We aren't flushing it.
1116         *
1117         * chunkhold expects B_DONE to be set, whether
1118         * we actually finish the I/O or not. We don't want to
1119         * change that interface.
1120         */
1121        XFS_BUF_UNREAD(bp);
1122        XFS_BUF_DONE(bp);
1123        xfs_buf_stale(bp);
1124        bp->b_iodone = NULL;
1125        if (!(fl & XBF_ASYNC)) {
1126                /*
1127                 * Mark b_error and B_ERROR _both_.
1128                 * Lot's of chunkcache code assumes that.
1129                 * There's no reason to mark error for
1130                 * ASYNC buffers.
1131                 */
1132                xfs_buf_ioerror(bp, EIO);
1133                complete(&bp->b_iowait);
1134        } else {
1135                xfs_buf_relse(bp);
1136        }
1137
1138        return EIO;
1139}
1140
1141STATIC int
1142xfs_bdstrat_cb(
1143        struct xfs_buf  *bp)
1144{
1145        if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1146                trace_xfs_bdstrat_shut(bp, _RET_IP_);
1147                /*
1148                 * Metadata write that didn't get logged but
1149                 * written delayed anyway. These aren't associated
1150                 * with a transaction, and can be ignored.
1151                 */
1152                if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1153                        return xfs_bioerror_relse(bp);
1154                else
1155                        return xfs_bioerror(bp);
1156        }
1157
1158        xfs_buf_iorequest(bp);
1159        return 0;
1160}
1161
1162int
1163xfs_bwrite(
1164        struct xfs_buf          *bp)
1165{
1166        int                     error;
1167
1168        ASSERT(xfs_buf_islocked(bp));
1169
1170        bp->b_flags |= XBF_WRITE;
1171        bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1172
1173        xfs_bdstrat_cb(bp);
1174
1175        error = xfs_buf_iowait(bp);
1176        if (error) {
1177                xfs_force_shutdown(bp->b_target->bt_mount,
1178                                   SHUTDOWN_META_IO_ERROR);
1179        }
1180        return error;
1181}
1182
1183/*
1184 * Wrapper around bdstrat so that we can stop data from going to disk in case
1185 * we are shutting down the filesystem.  Typically user data goes thru this
1186 * path; one of the exceptions is the superblock.
1187 */
1188void
1189xfsbdstrat(
1190        struct xfs_mount        *mp,
1191        struct xfs_buf          *bp)
1192{
1193        if (XFS_FORCED_SHUTDOWN(mp)) {
1194                trace_xfs_bdstrat_shut(bp, _RET_IP_);
1195                xfs_bioerror_relse(bp);
1196                return;
1197        }
1198
1199        xfs_buf_iorequest(bp);
1200}
1201
1202STATIC void
1203_xfs_buf_ioend(
1204        xfs_buf_t               *bp,
1205        int                     schedule)
1206{
1207        if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1208                xfs_buf_ioend(bp, schedule);
1209}
1210
1211STATIC void
1212xfs_buf_bio_end_io(
1213        struct bio              *bio,
1214        int                     error)
1215{
1216        xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1217
1218        /*
1219         * don't overwrite existing errors - otherwise we can lose errors on
1220         * buffers that require multiple bios to complete.
1221         */
1222        if (!bp->b_error)
1223                xfs_buf_ioerror(bp, -error);
1224
1225        if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1226                invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1227
1228        _xfs_buf_ioend(bp, 1);
1229        bio_put(bio);
1230}
1231
1232static void
1233xfs_buf_ioapply_map(
1234        struct xfs_buf  *bp,
1235        int             map,
1236        int             *buf_offset,
1237        int             *count,
1238        int             rw)
1239{
1240        int             page_index;
1241        int             total_nr_pages = bp->b_page_count;
1242        int             nr_pages;
1243        struct bio      *bio;
1244        sector_t        sector =  bp->b_maps[map].bm_bn;
1245        int             size;
1246        int             offset;
1247
1248        total_nr_pages = bp->b_page_count;
1249
1250        /* skip the pages in the buffer before the start offset */
1251        page_index = 0;
1252        offset = *buf_offset;
1253        while (offset >= PAGE_SIZE) {
1254                page_index++;
1255                offset -= PAGE_SIZE;
1256        }
1257
1258        /*
1259         * Limit the IO size to the length of the current vector, and update the
1260         * remaining IO count for the next time around.
1261         */
1262        size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1263        *count -= size;
1264        *buf_offset += size;
1265
1266next_chunk:
1267        atomic_inc(&bp->b_io_remaining);
1268        nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1269        if (nr_pages > total_nr_pages)
1270                nr_pages = total_nr_pages;
1271
1272        bio = bio_alloc(GFP_NOIO, nr_pages);
1273        bio->bi_bdev = bp->b_target->bt_bdev;
1274        bio->bi_sector = sector;
1275        bio->bi_end_io = xfs_buf_bio_end_io;
1276        bio->bi_private = bp;
1277
1278
1279        for (; size && nr_pages; nr_pages--, page_index++) {
1280                int     rbytes, nbytes = PAGE_SIZE - offset;
1281
1282                if (nbytes > size)
1283                        nbytes = size;
1284
1285                rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1286                                      offset);
1287                if (rbytes < nbytes)
1288                        break;
1289
1290                offset = 0;
1291                sector += BTOBB(nbytes);
1292                size -= nbytes;
1293                total_nr_pages--;
1294        }
1295
1296        if (likely(bio->bi_size)) {
1297                if (xfs_buf_is_vmapped(bp)) {
1298                        flush_kernel_vmap_range(bp->b_addr,
1299                                                xfs_buf_vmap_len(bp));
1300                }
1301                submit_bio(rw, bio);
1302                if (size)
1303                        goto next_chunk;
1304        } else {
1305                /*
1306                 * This is guaranteed not to be the last io reference count
1307                 * because the caller (xfs_buf_iorequest) holds a count itself.
1308                 */
1309                atomic_dec(&bp->b_io_remaining);
1310                xfs_buf_ioerror(bp, EIO);
1311                bio_put(bio);
1312        }
1313
1314}
1315
1316STATIC void
1317_xfs_buf_ioapply(
1318        struct xfs_buf  *bp)
1319{
1320        struct blk_plug plug;
1321        int             rw;
1322        int             offset;
1323        int             size;
1324        int             i;
1325
1326        if (bp->b_flags & XBF_WRITE) {
1327                if (bp->b_flags & XBF_SYNCIO)
1328                        rw = WRITE_SYNC;
1329                else
1330                        rw = WRITE;
1331                if (bp->b_flags & XBF_FUA)
1332                        rw |= REQ_FUA;
1333                if (bp->b_flags & XBF_FLUSH)
1334                        rw |= REQ_FLUSH;
1335        } else if (bp->b_flags & XBF_READ_AHEAD) {
1336                rw = READA;
1337        } else {
1338                rw = READ;
1339        }
1340
1341        /* we only use the buffer cache for meta-data */
1342        rw |= REQ_META;
1343
1344        /*
1345         * Walk all the vectors issuing IO on them. Set up the initial offset
1346         * into the buffer and the desired IO size before we start -
1347         * _xfs_buf_ioapply_vec() will modify them appropriately for each
1348         * subsequent call.
1349         */
1350        offset = bp->b_offset;
1351        size = BBTOB(bp->b_io_length);
1352        blk_start_plug(&plug);
1353        for (i = 0; i < bp->b_map_count; i++) {
1354                xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1355                if (bp->b_error)
1356                        break;
1357                if (size <= 0)
1358                        break;  /* all done */
1359        }
1360        blk_finish_plug(&plug);
1361}
1362
1363void
1364xfs_buf_iorequest(
1365        xfs_buf_t               *bp)
1366{
1367        trace_xfs_buf_iorequest(bp, _RET_IP_);
1368
1369        ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1370
1371        if (bp->b_flags & XBF_WRITE)
1372                xfs_buf_wait_unpin(bp);
1373        xfs_buf_hold(bp);
1374
1375        /* Set the count to 1 initially, this will stop an I/O
1376         * completion callout which happens before we have started
1377         * all the I/O from calling xfs_buf_ioend too early.
1378         */
1379        atomic_set(&bp->b_io_remaining, 1);
1380        _xfs_buf_ioapply(bp);
1381        _xfs_buf_ioend(bp, 1);
1382
1383        xfs_buf_rele(bp);
1384}
1385
1386/*
1387 * Waits for I/O to complete on the buffer supplied.  It returns immediately if
1388 * no I/O is pending or there is already a pending error on the buffer.  It
1389 * returns the I/O error code, if any, or 0 if there was no error.
1390 */
1391int
1392xfs_buf_iowait(
1393        xfs_buf_t               *bp)
1394{
1395        trace_xfs_buf_iowait(bp, _RET_IP_);
1396
1397        if (!bp->b_error)
1398                wait_for_completion(&bp->b_iowait);
1399
1400        trace_xfs_buf_iowait_done(bp, _RET_IP_);
1401        return bp->b_error;
1402}
1403
1404xfs_caddr_t
1405xfs_buf_offset(
1406        xfs_buf_t               *bp,
1407        size_t                  offset)
1408{
1409        struct page             *page;
1410
1411        if (bp->b_addr)
1412                return bp->b_addr + offset;
1413
1414        offset += bp->b_offset;
1415        page = bp->b_pages[offset >> PAGE_SHIFT];
1416        return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1417}
1418
1419/*
1420 *      Move data into or out of a buffer.
1421 */
1422void
1423xfs_buf_iomove(
1424        xfs_buf_t               *bp,    /* buffer to process            */
1425        size_t                  boff,   /* starting buffer offset       */
1426        size_t                  bsize,  /* length to copy               */
1427        void                    *data,  /* data address                 */
1428        xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1429{
1430        size_t                  bend;
1431
1432        bend = boff + bsize;
1433        while (boff < bend) {
1434                struct page     *page;
1435                int             page_index, page_offset, csize;
1436
1437                page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1438                page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1439                page = bp->b_pages[page_index];
1440                csize = min_t(size_t, PAGE_SIZE - page_offset,
1441                                      BBTOB(bp->b_io_length) - boff);
1442
1443                ASSERT((csize + page_offset) <= PAGE_SIZE);
1444
1445                switch (mode) {
1446                case XBRW_ZERO:
1447                        memset(page_address(page) + page_offset, 0, csize);
1448                        break;
1449                case XBRW_READ:
1450                        memcpy(data, page_address(page) + page_offset, csize);
1451                        break;
1452                case XBRW_WRITE:
1453                        memcpy(page_address(page) + page_offset, data, csize);
1454                }
1455
1456                boff += csize;
1457                data += csize;
1458        }
1459}
1460
1461/*
1462 *      Handling of buffer targets (buftargs).
1463 */
1464
1465/*
1466 * Wait for any bufs with callbacks that have been submitted but have not yet
1467 * returned. These buffers will have an elevated hold count, so wait on those
1468 * while freeing all the buffers only held by the LRU.
1469 */
1470void
1471xfs_wait_buftarg(
1472        struct xfs_buftarg      *btp)
1473{
1474        struct xfs_buf          *bp;
1475
1476restart:
1477        spin_lock(&btp->bt_lru_lock);
1478        while (!list_empty(&btp->bt_lru)) {
1479                bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1480                if (atomic_read(&bp->b_hold) > 1) {
1481                        spin_unlock(&btp->bt_lru_lock);
1482                        delay(100);
1483                        goto restart;
1484                }
1485                /*
1486                 * clear the LRU reference count so the buffer doesn't get
1487                 * ignored in xfs_buf_rele().
1488                 */
1489                atomic_set(&bp->b_lru_ref, 0);
1490                spin_unlock(&btp->bt_lru_lock);
1491                xfs_buf_rele(bp);
1492                spin_lock(&btp->bt_lru_lock);
1493        }
1494        spin_unlock(&btp->bt_lru_lock);
1495}
1496
1497int
1498xfs_buftarg_shrink(
1499        struct shrinker         *shrink,
1500        struct shrink_control   *sc)
1501{
1502        struct xfs_buftarg      *btp = container_of(shrink,
1503                                        struct xfs_buftarg, bt_shrinker);
1504        struct xfs_buf          *bp;
1505        int nr_to_scan = sc->nr_to_scan;
1506        LIST_HEAD(dispose);
1507
1508        if (!nr_to_scan)
1509                return btp->bt_lru_nr;
1510
1511        spin_lock(&btp->bt_lru_lock);
1512        while (!list_empty(&btp->bt_lru)) {
1513                if (nr_to_scan-- <= 0)
1514                        break;
1515
1516                bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1517
1518                /*
1519                 * Decrement the b_lru_ref count unless the value is already
1520                 * zero. If the value is already zero, we need to reclaim the
1521                 * buffer, otherwise it gets another trip through the LRU.
1522                 */
1523                if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1524                        list_move_tail(&bp->b_lru, &btp->bt_lru);
1525                        continue;
1526                }
1527
1528                /*
1529                 * remove the buffer from the LRU now to avoid needing another
1530                 * lock round trip inside xfs_buf_rele().
1531                 */
1532                list_move(&bp->b_lru, &dispose);
1533                btp->bt_lru_nr--;
1534                bp->b_lru_flags |= _XBF_LRU_DISPOSE;
1535        }
1536        spin_unlock(&btp->bt_lru_lock);
1537
1538        while (!list_empty(&dispose)) {
1539                bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1540                list_del_init(&bp->b_lru);
1541                xfs_buf_rele(bp);
1542        }
1543
1544        return btp->bt_lru_nr;
1545}
1546
1547void
1548xfs_free_buftarg(
1549        struct xfs_mount        *mp,
1550        struct xfs_buftarg      *btp)
1551{
1552        unregister_shrinker(&btp->bt_shrinker);
1553
1554        if (mp->m_flags & XFS_MOUNT_BARRIER)
1555                xfs_blkdev_issue_flush(btp);
1556
1557        kmem_free(btp);
1558}
1559
1560STATIC int
1561xfs_setsize_buftarg_flags(
1562        xfs_buftarg_t           *btp,
1563        unsigned int            blocksize,
1564        unsigned int            sectorsize,
1565        int                     verbose)
1566{
1567        btp->bt_bsize = blocksize;
1568        btp->bt_sshift = ffs(sectorsize) - 1;
1569        btp->bt_smask = sectorsize - 1;
1570
1571        if (set_blocksize(btp->bt_bdev, sectorsize)) {
1572                char name[BDEVNAME_SIZE];
1573
1574                bdevname(btp->bt_bdev, name);
1575
1576                xfs_warn(btp->bt_mount,
1577                        "Cannot set_blocksize to %u on device %s\n",
1578                        sectorsize, name);
1579                return EINVAL;
1580        }
1581
1582        return 0;
1583}
1584
1585/*
1586 *      When allocating the initial buffer target we have not yet
1587 *      read in the superblock, so don't know what sized sectors
1588 *      are being used is at this early stage.  Play safe.
1589 */
1590STATIC int
1591xfs_setsize_buftarg_early(
1592        xfs_buftarg_t           *btp,
1593        struct block_device     *bdev)
1594{
1595        return xfs_setsize_buftarg_flags(btp,
1596                        PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1597}
1598
1599int
1600xfs_setsize_buftarg(
1601        xfs_buftarg_t           *btp,
1602        unsigned int            blocksize,
1603        unsigned int            sectorsize)
1604{
1605        return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1606}
1607
1608xfs_buftarg_t *
1609xfs_alloc_buftarg(
1610        struct xfs_mount        *mp,
1611        struct block_device     *bdev,
1612        int                     external,
1613        const char              *fsname)
1614{
1615        xfs_buftarg_t           *btp;
1616
1617        btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1618
1619        btp->bt_mount = mp;
1620        btp->bt_dev =  bdev->bd_dev;
1621        btp->bt_bdev = bdev;
1622        btp->bt_bdi = blk_get_backing_dev_info(bdev);
1623        if (!btp->bt_bdi)
1624                goto error;
1625
1626        INIT_LIST_HEAD(&btp->bt_lru);
1627        spin_lock_init(&btp->bt_lru_lock);
1628        if (xfs_setsize_buftarg_early(btp, bdev))
1629                goto error;
1630        btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1631        btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1632        register_shrinker(&btp->bt_shrinker);
1633        return btp;
1634
1635error:
1636        kmem_free(btp);
1637        return NULL;
1638}
1639
1640/*
1641 * Add a buffer to the delayed write list.
1642 *
1643 * This queues a buffer for writeout if it hasn't already been.  Note that
1644 * neither this routine nor the buffer list submission functions perform
1645 * any internal synchronization.  It is expected that the lists are thread-local
1646 * to the callers.
1647 *
1648 * Returns true if we queued up the buffer, or false if it already had
1649 * been on the buffer list.
1650 */
1651bool
1652xfs_buf_delwri_queue(
1653        struct xfs_buf          *bp,
1654        struct list_head        *list)
1655{
1656        ASSERT(xfs_buf_islocked(bp));
1657        ASSERT(!(bp->b_flags & XBF_READ));
1658
1659        /*
1660         * If the buffer is already marked delwri it already is queued up
1661         * by someone else for imediate writeout.  Just ignore it in that
1662         * case.
1663         */
1664        if (bp->b_flags & _XBF_DELWRI_Q) {
1665                trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1666                return false;
1667        }
1668
1669        trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1670
1671        /*
1672         * If a buffer gets written out synchronously or marked stale while it
1673         * is on a delwri list we lazily remove it. To do this, the other party
1674         * clears the  _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1675         * It remains referenced and on the list.  In a rare corner case it
1676         * might get readded to a delwri list after the synchronous writeout, in
1677         * which case we need just need to re-add the flag here.
1678         */
1679        bp->b_flags |= _XBF_DELWRI_Q;
1680        if (list_empty(&bp->b_list)) {
1681                atomic_inc(&bp->b_hold);
1682                list_add_tail(&bp->b_list, list);
1683        }
1684
1685        return true;
1686}
1687
1688/*
1689 * Compare function is more complex than it needs to be because
1690 * the return value is only 32 bits and we are doing comparisons
1691 * on 64 bit values
1692 */
1693static int
1694xfs_buf_cmp(
1695        void            *priv,
1696        struct list_head *a,
1697        struct list_head *b)
1698{
1699        struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1700        struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1701        xfs_daddr_t             diff;
1702
1703        diff = ap->b_map.bm_bn - bp->b_map.bm_bn;
1704        if (diff < 0)
1705                return -1;
1706        if (diff > 0)
1707                return 1;
1708        return 0;
1709}
1710
1711static int
1712__xfs_buf_delwri_submit(
1713        struct list_head        *buffer_list,
1714        struct list_head        *io_list,
1715        bool                    wait)
1716{
1717        struct blk_plug         plug;
1718        struct xfs_buf          *bp, *n;
1719        int                     pinned = 0;
1720
1721        list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1722                if (!wait) {
1723                        if (xfs_buf_ispinned(bp)) {
1724                                pinned++;
1725                                continue;
1726                        }
1727                        if (!xfs_buf_trylock(bp))
1728                                continue;
1729                } else {
1730                        xfs_buf_lock(bp);
1731                }
1732
1733                /*
1734                 * Someone else might have written the buffer synchronously or
1735                 * marked it stale in the meantime.  In that case only the
1736                 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1737                 * reference and remove it from the list here.
1738                 */
1739                if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1740                        list_del_init(&bp->b_list);
1741                        xfs_buf_relse(bp);
1742                        continue;
1743                }
1744
1745                list_move_tail(&bp->b_list, io_list);
1746                trace_xfs_buf_delwri_split(bp, _RET_IP_);
1747        }
1748
1749        list_sort(NULL, io_list, xfs_buf_cmp);
1750
1751        blk_start_plug(&plug);
1752        list_for_each_entry_safe(bp, n, io_list, b_list) {
1753                bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1754                bp->b_flags |= XBF_WRITE;
1755
1756                if (!wait) {
1757                        bp->b_flags |= XBF_ASYNC;
1758                        list_del_init(&bp->b_list);
1759                }
1760                xfs_bdstrat_cb(bp);
1761        }
1762        blk_finish_plug(&plug);
1763
1764        return pinned;
1765}
1766
1767/*
1768 * Write out a buffer list asynchronously.
1769 *
1770 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1771 * out and not wait for I/O completion on any of the buffers.  This interface
1772 * is only safely useable for callers that can track I/O completion by higher
1773 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1774 * function.
1775 */
1776int
1777xfs_buf_delwri_submit_nowait(
1778        struct list_head        *buffer_list)
1779{
1780        LIST_HEAD               (io_list);
1781        return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1782}
1783
1784/*
1785 * Write out a buffer list synchronously.
1786 *
1787 * This will take the @buffer_list, write all buffers out and wait for I/O
1788 * completion on all of the buffers. @buffer_list is consumed by the function,
1789 * so callers must have some other way of tracking buffers if they require such
1790 * functionality.
1791 */
1792int
1793xfs_buf_delwri_submit(
1794        struct list_head        *buffer_list)
1795{
1796        LIST_HEAD               (io_list);
1797        int                     error = 0, error2;
1798        struct xfs_buf          *bp;
1799
1800        __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1801
1802        /* Wait for IO to complete. */
1803        while (!list_empty(&io_list)) {
1804                bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1805
1806                list_del_init(&bp->b_list);
1807                error2 = xfs_buf_iowait(bp);
1808                xfs_buf_relse(bp);
1809                if (!error)
1810                        error = error2;
1811        }
1812
1813        return error;
1814}
1815
1816int __init
1817xfs_buf_init(void)
1818{
1819        xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1820                                                KM_ZONE_HWALIGN, NULL);
1821        if (!xfs_buf_zone)
1822                goto out;
1823
1824        xfslogd_workqueue = alloc_workqueue("xfslogd",
1825                                        WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1826        if (!xfslogd_workqueue)
1827                goto out_free_buf_zone;
1828
1829        return 0;
1830
1831 out_free_buf_zone:
1832        kmem_zone_destroy(xfs_buf_zone);
1833 out:
1834        return -ENOMEM;
1835}
1836
1837void
1838xfs_buf_terminate(void)
1839{
1840        destroy_workqueue(xfslogd_workqueue);
1841        kmem_zone_destroy(xfs_buf_zone);
1842}
1843
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.