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        int                     numblks = 0;
 491        int                     i;
 492
 493        for (i = 0; i < nmaps; i++)
 494                numblks += map[i].bm_len;
 495        numbytes = BBTOB(numblks);
 496
 497        /* Check for IOs smaller than the sector size / not sector aligned */
 498        ASSERT(!(numbytes < (1 << btp->bt_sshift)));
 499        ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask));
 500
 501        /* get tree root */
 502        pag = xfs_perag_get(btp->bt_mount,
 503                                xfs_daddr_to_agno(btp->bt_mount, blkno));
 504
 505        /* walk tree */
 506        spin_lock(&pag->pag_buf_lock);
 507        rbp = &pag->pag_buf_tree.rb_node;
 508        parent = NULL;
 509        bp = NULL;
 510        while (*rbp) {
 511                parent = *rbp;
 512                bp = rb_entry(parent, struct xfs_buf, b_rbnode);
 513
 514                if (blkno < bp->b_bn)
 515                        rbp = &(*rbp)->rb_left;
 516                else if (blkno > bp->b_bn)
 517                        rbp = &(*rbp)->rb_right;
 518                else {
 519                        /*
 520                         * found a block number match. If the range doesn't
 521                         * match, the only way this is allowed is if the buffer
 522                         * in the cache is stale and the transaction that made
 523                         * it stale has not yet committed. i.e. we are
 524                         * reallocating a busy extent. Skip this buffer and
 525                         * continue searching to the right for an exact match.
 526                         */
 527                        if (bp->b_length != numblks) {
 528                                ASSERT(bp->b_flags & XBF_STALE);
 529                                rbp = &(*rbp)->rb_right;
 530                                continue;
 531                        }
 532                        atomic_inc(&bp->b_hold);
 533                        goto found;
 534                }
 535        }
 536
 537        /* No match found */
 538        if (new_bp) {
 539                rb_link_node(&new_bp->b_rbnode, parent, rbp);
 540                rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
 541                /* the buffer keeps the perag reference until it is freed */
 542                new_bp->b_pag = pag;
 543                spin_unlock(&pag->pag_buf_lock);
 544        } else {
 545                XFS_STATS_INC(xb_miss_locked);
 546                spin_unlock(&pag->pag_buf_lock);
 547                xfs_perag_put(pag);
 548        }
 549        return new_bp;
 550
 551found:
 552        spin_unlock(&pag->pag_buf_lock);
 553        xfs_perag_put(pag);
 554
 555        if (!xfs_buf_trylock(bp)) {
 556                if (flags & XBF_TRYLOCK) {
 557                        xfs_buf_rele(bp);
 558                        XFS_STATS_INC(xb_busy_locked);
 559                        return NULL;
 560                }
 561                xfs_buf_lock(bp);
 562                XFS_STATS_INC(xb_get_locked_waited);
 563        }
 564
 565        /*
 566         * if the buffer is stale, clear all the external state associated with
 567         * it. We need to keep flags such as how we allocated the buffer memory
 568         * intact here.
 569         */
 570        if (bp->b_flags & XBF_STALE) {
 571                ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 572                bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
 573        }
 574
 575        trace_xfs_buf_find(bp, flags, _RET_IP_);
 576        XFS_STATS_INC(xb_get_locked);
 577        return bp;
 578}
 579
 580/*
 581 * Assembles a buffer covering the specified range. The code is optimised for
 582 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
 583 * more hits than misses.
 584 */
 585struct xfs_buf *
 586xfs_buf_get_map(
 587        struct xfs_buftarg      *target,
 588        struct xfs_buf_map      *map,
 589        int                     nmaps,
 590        xfs_buf_flags_t         flags)
 591{
 592        struct xfs_buf          *bp;
 593        struct xfs_buf          *new_bp;
 594        int                     error = 0;
 595
 596        bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
 597        if (likely(bp))
 598                goto found;
 599
 600        new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
 601        if (unlikely(!new_bp))
 602                return NULL;
 603
 604        error = xfs_buf_allocate_memory(new_bp, flags);
 605        if (error) {
 606                xfs_buf_free(new_bp);
 607                return NULL;
 608        }
 609
 610        bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
 611        if (!bp) {
 612                xfs_buf_free(new_bp);
 613                return NULL;
 614        }
 615
 616        if (bp != new_bp)
 617                xfs_buf_free(new_bp);
 618
 619found:
 620        if (!bp->b_addr) {
 621                error = _xfs_buf_map_pages(bp, flags);
 622                if (unlikely(error)) {
 623                        xfs_warn(target->bt_mount,
 624                                "%s: failed to map pages\n", __func__);
 625                        xfs_buf_relse(bp);
 626                        return NULL;
 627                }
 628        }
 629
 630        XFS_STATS_INC(xb_get);
 631        trace_xfs_buf_get(bp, flags, _RET_IP_);
 632        return bp;
 633}
 634
 635STATIC int
 636_xfs_buf_read(
 637        xfs_buf_t               *bp,
 638        xfs_buf_flags_t         flags)
 639{
 640        ASSERT(!(flags & XBF_WRITE));
 641        ASSERT(bp->b_map.bm_bn != XFS_BUF_DADDR_NULL);
 642
 643        bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
 644        bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 645
 646        xfs_buf_iorequest(bp);
 647        if (flags & XBF_ASYNC)
 648                return 0;
 649        return xfs_buf_iowait(bp);
 650}
 651
 652xfs_buf_t *
 653xfs_buf_read_map(
 654        struct xfs_buftarg      *target,
 655        struct xfs_buf_map      *map,
 656        int                     nmaps,
 657        xfs_buf_flags_t         flags)
 658{
 659        struct xfs_buf          *bp;
 660
 661        flags |= XBF_READ;
 662
 663        bp = xfs_buf_get_map(target, map, nmaps, flags);
 664        if (bp) {
 665                trace_xfs_buf_read(bp, flags, _RET_IP_);
 666
 667                if (!XFS_BUF_ISDONE(bp)) {
 668                        XFS_STATS_INC(xb_get_read);
 669                        _xfs_buf_read(bp, flags);
 670                } else if (flags & XBF_ASYNC) {
 671                        /*
 672                         * Read ahead call which is already satisfied,
 673                         * drop the buffer
 674                         */
 675                        xfs_buf_relse(bp);
 676                        return NULL;
 677                } else {
 678                        /* We do not want read in the flags */
 679                        bp->b_flags &= ~XBF_READ;
 680                }
 681        }
 682
 683        return bp;
 684}
 685
 686/*
 687 *      If we are not low on memory then do the readahead in a deadlock
 688 *      safe manner.
 689 */
 690void
 691xfs_buf_readahead_map(
 692        struct xfs_buftarg      *target,
 693        struct xfs_buf_map      *map,
 694        int                     nmaps)
 695{
 696        if (bdi_read_congested(target->bt_bdi))
 697                return;
 698
 699        xfs_buf_read_map(target, map, nmaps,
 700                     XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 701}
 702
 703/*
 704 * Read an uncached buffer from disk. Allocates and returns a locked
 705 * buffer containing the disk contents or nothing.
 706 */
 707struct xfs_buf *
 708xfs_buf_read_uncached(
 709        struct xfs_buftarg      *target,
 710        xfs_daddr_t             daddr,
 711        size_t                  numblks,
 712        int                     flags)
 713{
 714        xfs_buf_t               *bp;
 715        int                     error;
 716
 717        bp = xfs_buf_get_uncached(target, numblks, flags);
 718        if (!bp)
 719                return NULL;
 720
 721        /* set up the buffer for a read IO */
 722        ASSERT(bp->b_map_count == 1);
 723        bp->b_bn = daddr;
 724        bp->b_maps[0].bm_bn = daddr;
 725        bp->b_flags |= XBF_READ;
 726
 727        xfsbdstrat(target->bt_mount, bp);
 728        error = xfs_buf_iowait(bp);
 729        if (error) {
 730                xfs_buf_relse(bp);
 731                return NULL;
 732        }
 733        return bp;
 734}
 735
 736/*
 737 * Return a buffer allocated as an empty buffer and associated to external
 738 * memory via xfs_buf_associate_memory() back to it's empty state.
 739 */
 740void
 741xfs_buf_set_empty(
 742        struct xfs_buf          *bp,
 743        size_t                  numblks)
 744{
 745        if (bp->b_pages)
 746                _xfs_buf_free_pages(bp);
 747
 748        bp->b_pages = NULL;
 749        bp->b_page_count = 0;
 750        bp->b_addr = NULL;
 751        bp->b_length = numblks;
 752        bp->b_io_length = numblks;
 753
 754        ASSERT(bp->b_map_count == 1);
 755        bp->b_bn = XFS_BUF_DADDR_NULL;
 756        bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
 757        bp->b_maps[0].bm_len = bp->b_length;
 758}
 759
 760static inline struct page *
 761mem_to_page(
 762        void                    *addr)
 763{
 764        if ((!is_vmalloc_addr(addr))) {
 765                return virt_to_page(addr);
 766        } else {
 767                return vmalloc_to_page(addr);
 768        }
 769}
 770
 771int
 772xfs_buf_associate_memory(
 773        xfs_buf_t               *bp,
 774        void                    *mem,
 775        size_t                  len)
 776{
 777        int                     rval;
 778        int                     i = 0;
 779        unsigned long           pageaddr;
 780        unsigned long           offset;
 781        size_t                  buflen;
 782        int                     page_count;
 783
 784        pageaddr = (unsigned long)mem & PAGE_MASK;
 785        offset = (unsigned long)mem - pageaddr;
 786        buflen = PAGE_ALIGN(len + offset);
 787        page_count = buflen >> PAGE_SHIFT;
 788
 789        /* Free any previous set of page pointers */
 790        if (bp->b_pages)
 791                _xfs_buf_free_pages(bp);
 792
 793        bp->b_pages = NULL;
 794        bp->b_addr = mem;
 795
 796        rval = _xfs_buf_get_pages(bp, page_count, 0);
 797        if (rval)
 798                return rval;
 799
 800        bp->b_offset = offset;
 801
 802        for (i = 0; i < bp->b_page_count; i++) {
 803                bp->b_pages[i] = mem_to_page((void *)pageaddr);
 804                pageaddr += PAGE_SIZE;
 805        }
 806
 807        bp->b_io_length = BTOBB(len);
 808        bp->b_length = BTOBB(buflen);
 809
 810        return 0;
 811}
 812
 813xfs_buf_t *
 814xfs_buf_get_uncached(
 815        struct xfs_buftarg      *target,
 816        size_t                  numblks,
 817        int                     flags)
 818{
 819        unsigned long           page_count;
 820        int                     error, i;
 821        struct xfs_buf          *bp;
 822        DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
 823
 824        bp = _xfs_buf_alloc(target, &map, 1, 0);
 825        if (unlikely(bp == NULL))
 826                goto fail;
 827
 828        page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
 829        error = _xfs_buf_get_pages(bp, page_count, 0);
 830        if (error)
 831                goto fail_free_buf;
 832
 833        for (i = 0; i < page_count; i++) {
 834                bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 835                if (!bp->b_pages[i])
 836                        goto fail_free_mem;
 837        }
 838        bp->b_flags |= _XBF_PAGES;
 839
 840        error = _xfs_buf_map_pages(bp, 0);
 841        if (unlikely(error)) {
 842                xfs_warn(target->bt_mount,
 843                        "%s: failed to map pages\n", __func__);
 844                goto fail_free_mem;
 845        }
 846
 847        trace_xfs_buf_get_uncached(bp, _RET_IP_);
 848        return bp;
 849
 850 fail_free_mem:
 851        while (--i >= 0)
 852                __free_page(bp->b_pages[i]);
 853        _xfs_buf_free_pages(bp);
 854 fail_free_buf:
 855        xfs_buf_free_maps(bp);
 856        kmem_zone_free(xfs_buf_zone, bp);
 857 fail:
 858        return NULL;
 859}
 860
 861/*
 862 *      Increment reference count on buffer, to hold the buffer concurrently
 863 *      with another thread which may release (free) the buffer asynchronously.
 864 *      Must hold the buffer already to call this function.
 865 */
 866void
 867xfs_buf_hold(
 868        xfs_buf_t               *bp)
 869{
 870        trace_xfs_buf_hold(bp, _RET_IP_);
 871        atomic_inc(&bp->b_hold);
 872}
 873
 874/*
 875 *      Releases a hold on the specified buffer.  If the
 876 *      the hold count is 1, calls xfs_buf_free.
 877 */
 878void
 879xfs_buf_rele(
 880        xfs_buf_t               *bp)
 881{
 882        struct xfs_perag        *pag = bp->b_pag;
 883
 884        trace_xfs_buf_rele(bp, _RET_IP_);
 885
 886        if (!pag) {
 887                ASSERT(list_empty(&bp->b_lru));
 888                ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
 889                if (atomic_dec_and_test(&bp->b_hold))
 890                        xfs_buf_free(bp);
 891                return;
 892        }
 893
 894        ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
 895
 896        ASSERT(atomic_read(&bp->b_hold) > 0);
 897        if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
 898                if (!(bp->b_flags & XBF_STALE) &&
 899                           atomic_read(&bp->b_lru_ref)) {
 900                        xfs_buf_lru_add(bp);
 901                        spin_unlock(&pag->pag_buf_lock);
 902                } else {
 903                        xfs_buf_lru_del(bp);
 904                        ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
 905                        rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 906                        spin_unlock(&pag->pag_buf_lock);
 907                        xfs_perag_put(pag);
 908                        xfs_buf_free(bp);
 909                }
 910        }
 911}
 912
 913
 914/*
 915 *      Lock a buffer object, if it is not already locked.
 916 *
 917 *      If we come across a stale, pinned, locked buffer, we know that we are
 918 *      being asked to lock a buffer that has been reallocated. Because it is
 919 *      pinned, we know that the log has not been pushed to disk and hence it
 920 *      will still be locked.  Rather than continuing to have trylock attempts
 921 *      fail until someone else pushes the log, push it ourselves before
 922 *      returning.  This means that the xfsaild will not get stuck trying
 923 *      to push on stale inode buffers.
 924 */
 925int
 926xfs_buf_trylock(
 927        struct xfs_buf          *bp)
 928{
 929        int                     locked;
 930
 931        locked = down_trylock(&bp->b_sema) == 0;
 932        if (locked)
 933                XB_SET_OWNER(bp);
 934        else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 935                xfs_log_force(bp->b_target->bt_mount, 0);
 936
 937        trace_xfs_buf_trylock(bp, _RET_IP_);
 938        return locked;
 939}
 940
 941/*
 942 *      Lock a buffer object.
 943 *
 944 *      If we come across a stale, pinned, locked buffer, we know that we
 945 *      are being asked to lock a buffer that has been reallocated. Because
 946 *      it is pinned, we know that the log has not been pushed to disk and
 947 *      hence it will still be locked. Rather than sleeping until someone
 948 *      else pushes the log, push it ourselves before trying to get the lock.
 949 */
 950void
 951xfs_buf_lock(
 952        struct xfs_buf          *bp)
 953{
 954        trace_xfs_buf_lock(bp, _RET_IP_);
 955
 956        if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 957                xfs_log_force(bp->b_target->bt_mount, 0);
 958        down(&bp->b_sema);
 959        XB_SET_OWNER(bp);
 960
 961        trace_xfs_buf_lock_done(bp, _RET_IP_);
 962}
 963
 964void
 965xfs_buf_unlock(
 966        struct xfs_buf          *bp)
 967{
 968        XB_CLEAR_OWNER(bp);
 969        up(&bp->b_sema);
 970
 971        trace_xfs_buf_unlock(bp, _RET_IP_);
 972}
 973
 974STATIC void
 975xfs_buf_wait_unpin(
 976        xfs_buf_t               *bp)
 977{
 978        DECLARE_WAITQUEUE       (wait, current);
 979
 980        if (atomic_read(&bp->b_pin_count) == 0)
 981                return;
 982
 983        add_wait_queue(&bp->b_waiters, &wait);
 984        for (;;) {
 985                set_current_state(TASK_UNINTERRUPTIBLE);
 986                if (atomic_read(&bp->b_pin_count) == 0)
 987                        break;
 988                io_schedule();
 989        }
 990        remove_wait_queue(&bp->b_waiters, &wait);
 991        set_current_state(TASK_RUNNING);
 992}
 993
 994/*
 995 *      Buffer Utility Routines
 996 */
 997
 998STATIC void
 999xfs_buf_iodone_work(
1000        struct work_struct      *work)
1001{
1002        xfs_buf_t               *bp =
1003                container_of(work, xfs_buf_t, b_iodone_work);
1004
1005        if (bp->b_iodone)
1006                (*(bp->b_iodone))(bp);
1007        else if (bp->b_flags & XBF_ASYNC)
1008                xfs_buf_relse(bp);
1009}
1010
1011void
1012xfs_buf_ioend(
1013        xfs_buf_t               *bp,
1014        int                     schedule)
1015{
1016        trace_xfs_buf_iodone(bp, _RET_IP_);
1017
1018        bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1019        if (bp->b_error == 0)
1020                bp->b_flags |= XBF_DONE;
1021
1022        if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1023                if (schedule) {
1024                        INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1025                        queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1026                } else {
1027                        xfs_buf_iodone_work(&bp->b_iodone_work);
1028                }
1029        } else {
1030                complete(&bp->b_iowait);
1031        }
1032}
1033
1034void
1035xfs_buf_ioerror(
1036        xfs_buf_t               *bp,
1037        int                     error)
1038{
1039        ASSERT(error >= 0 && error <= 0xffff);
1040        bp->b_error = (unsigned short)error;
1041        trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1042}
1043
1044void
1045xfs_buf_ioerror_alert(
1046        struct xfs_buf          *bp,
1047        const char              *func)
1048{
1049        xfs_alert(bp->b_target->bt_mount,
1050"metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1051                (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1052}
1053
1054/*
1055 * Called when we want to stop a buffer from getting written or read.
1056 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1057 * so that the proper iodone callbacks get called.
1058 */
1059STATIC int
1060xfs_bioerror(
1061        xfs_buf_t *bp)
1062{
1063#ifdef XFSERRORDEBUG
1064        ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1065#endif
1066
1067        /*
1068         * No need to wait until the buffer is unpinned, we aren't flushing it.
1069         */
1070        xfs_buf_ioerror(bp, EIO);
1071
1072        /*
1073         * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1074         */
1075        XFS_BUF_UNREAD(bp);
1076        XFS_BUF_UNDONE(bp);
1077        xfs_buf_stale(bp);
1078
1079        xfs_buf_ioend(bp, 0);
1080
1081        return EIO;
1082}
1083
1084/*
1085 * Same as xfs_bioerror, except that we are releasing the buffer
1086 * here ourselves, and avoiding the xfs_buf_ioend call.
1087 * This is meant for userdata errors; metadata bufs come with
1088 * iodone functions attached, so that we can track down errors.
1089 */
1090STATIC int
1091xfs_bioerror_relse(
1092        struct xfs_buf  *bp)
1093{
1094        int64_t         fl = bp->b_flags;
1095        /*
1096         * No need to wait until the buffer is unpinned.
1097         * We aren't flushing it.
1098         *
1099         * chunkhold expects B_DONE to be set, whether
1100         * we actually finish the I/O or not. We don't want to
1101         * change that interface.
1102         */
1103        XFS_BUF_UNREAD(bp);
1104        XFS_BUF_DONE(bp);
1105        xfs_buf_stale(bp);
1106        bp->b_iodone = NULL;
1107        if (!(fl & XBF_ASYNC)) {
1108                /*
1109                 * Mark b_error and B_ERROR _both_.
1110                 * Lot's of chunkcache code assumes that.
1111                 * There's no reason to mark error for
1112                 * ASYNC buffers.
1113                 */
1114                xfs_buf_ioerror(bp, EIO);
1115                complete(&bp->b_iowait);
1116        } else {
1117                xfs_buf_relse(bp);
1118        }
1119
1120        return EIO;
1121}
1122
1123STATIC int
1124xfs_bdstrat_cb(
1125        struct xfs_buf  *bp)
1126{
1127        if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1128                trace_xfs_bdstrat_shut(bp, _RET_IP_);
1129                /*
1130                 * Metadata write that didn't get logged but
1131                 * written delayed anyway. These aren't associated
1132                 * with a transaction, and can be ignored.
1133                 */
1134                if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1135                        return xfs_bioerror_relse(bp);
1136                else
1137                        return xfs_bioerror(bp);
1138        }
1139
1140        xfs_buf_iorequest(bp);
1141        return 0;
1142}
1143
1144int
1145xfs_bwrite(
1146        struct xfs_buf          *bp)
1147{
1148        int                     error;
1149
1150        ASSERT(xfs_buf_islocked(bp));
1151
1152        bp->b_flags |= XBF_WRITE;
1153        bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
1154
1155        xfs_bdstrat_cb(bp);
1156
1157        error = xfs_buf_iowait(bp);
1158        if (error) {
1159                xfs_force_shutdown(bp->b_target->bt_mount,
1160                                   SHUTDOWN_META_IO_ERROR);
1161        }
1162        return error;
1163}
1164
1165/*
1166 * Wrapper around bdstrat so that we can stop data from going to disk in case
1167 * we are shutting down the filesystem.  Typically user data goes thru this
1168 * path; one of the exceptions is the superblock.
1169 */
1170void
1171xfsbdstrat(
1172        struct xfs_mount        *mp,
1173        struct xfs_buf          *bp)
1174{
1175        if (XFS_FORCED_SHUTDOWN(mp)) {
1176                trace_xfs_bdstrat_shut(bp, _RET_IP_);
1177                xfs_bioerror_relse(bp);
1178                return;
1179        }
1180
1181        xfs_buf_iorequest(bp);
1182}
1183
1184STATIC void
1185_xfs_buf_ioend(
1186        xfs_buf_t               *bp,
1187        int                     schedule)
1188{
1189        if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1190                xfs_buf_ioend(bp, schedule);
1191}
1192
1193STATIC void
1194xfs_buf_bio_end_io(
1195        struct bio              *bio,
1196        int                     error)
1197{
1198        xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1199
1200        xfs_buf_ioerror(bp, -error);
1201
1202        if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1203                invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1204
1205        _xfs_buf_ioend(bp, 1);
1206        bio_put(bio);
1207}
1208
1209static void
1210xfs_buf_ioapply_map(
1211        struct xfs_buf  *bp,
1212        int             map,
1213        int             *buf_offset,
1214        int             *count,
1215        int             rw)
1216{
1217        int             page_index;
1218        int             total_nr_pages = bp->b_page_count;
1219        int             nr_pages;
1220        struct bio      *bio;
1221        sector_t        sector =  bp->b_maps[map].bm_bn;
1222        int             size;
1223        int             offset;
1224
1225        total_nr_pages = bp->b_page_count;
1226
1227        /* skip the pages in the buffer before the start offset */
1228        page_index = 0;
1229        offset = *buf_offset;
1230        while (offset >= PAGE_SIZE) {
1231                page_index++;
1232                offset -= PAGE_SIZE;
1233        }
1234
1235        /*
1236         * Limit the IO size to the length of the current vector, and update the
1237         * remaining IO count for the next time around.
1238         */
1239        size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1240        *count -= size;
1241        *buf_offset += size;
1242
1243next_chunk:
1244        atomic_inc(&bp->b_io_remaining);
1245        nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1246        if (nr_pages > total_nr_pages)
1247                nr_pages = total_nr_pages;
1248
1249        bio = bio_alloc(GFP_NOIO, nr_pages);
1250        bio->bi_bdev = bp->b_target->bt_bdev;
1251        bio->bi_sector = sector;
1252        bio->bi_end_io = xfs_buf_bio_end_io;
1253        bio->bi_private = bp;
1254
1255
1256        for (; size && nr_pages; nr_pages--, page_index++) {
1257                int     rbytes, nbytes = PAGE_SIZE - offset;
1258
1259                if (nbytes > size)
1260                        nbytes = size;
1261
1262                rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1263                                      offset);
1264                if (rbytes < nbytes)
1265                        break;
1266
1267                offset = 0;
1268                sector += BTOBB(nbytes);
1269                size -= nbytes;
1270                total_nr_pages--;
1271        }
1272
1273        if (likely(bio->bi_size)) {
1274                if (xfs_buf_is_vmapped(bp)) {
1275                        flush_kernel_vmap_range(bp->b_addr,
1276                                                xfs_buf_vmap_len(bp));
1277                }
1278                submit_bio(rw, bio);
1279                if (size)
1280                        goto next_chunk;
1281        } else {
1282                xfs_buf_ioerror(bp, EIO);
1283                bio_put(bio);
1284        }
1285
1286}
1287
1288STATIC void
1289_xfs_buf_ioapply(
1290        struct xfs_buf  *bp)
1291{
1292        struct blk_plug plug;
1293        int             rw;
1294        int             offset;
1295        int             size;
1296        int             i;
1297
1298        if (bp->b_flags & XBF_WRITE) {
1299                if (bp->b_flags & XBF_SYNCIO)
1300                        rw = WRITE_SYNC;
1301                else
1302                        rw = WRITE;
1303                if (bp->b_flags & XBF_FUA)
1304                        rw |= REQ_FUA;
1305                if (bp->b_flags & XBF_FLUSH)
1306                        rw |= REQ_FLUSH;
1307        } else if (bp->b_flags & XBF_READ_AHEAD) {
1308                rw = READA;
1309        } else {
1310                rw = READ;
1311        }
1312
1313        /* we only use the buffer cache for meta-data */
1314        rw |= REQ_META;
1315
1316        /*
1317         * Walk all the vectors issuing IO on them. Set up the initial offset
1318         * into the buffer and the desired IO size before we start -
1319         * _xfs_buf_ioapply_vec() will modify them appropriately for each
1320         * subsequent call.
1321         */
1322        offset = bp->b_offset;
1323        size = BBTOB(bp->b_io_length);
1324        blk_start_plug(&plug);
1325        for (i = 0; i < bp->b_map_count; i++) {
1326                xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1327                if (bp->b_error)
1328                        break;
1329                if (size <= 0)
1330                        break;  /* all done */
1331        }
1332        blk_finish_plug(&plug);
1333}
1334
1335void
1336xfs_buf_iorequest(
1337        xfs_buf_t               *bp)
1338{
1339        trace_xfs_buf_iorequest(bp, _RET_IP_);
1340
1341        ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1342
1343        if (bp->b_flags & XBF_WRITE)
1344                xfs_buf_wait_unpin(bp);
1345        xfs_buf_hold(bp);
1346
1347        /* Set the count to 1 initially, this will stop an I/O
1348         * completion callout which happens before we have started
1349         * all the I/O from calling xfs_buf_ioend too early.
1350         */
1351        atomic_set(&bp->b_io_remaining, 1);
1352        _xfs_buf_ioapply(bp);
1353        _xfs_buf_ioend(bp, 1);
1354
1355        xfs_buf_rele(bp);
1356}
1357
1358/*
1359 * Waits for I/O to complete on the buffer supplied.  It returns immediately if
1360 * no I/O is pending or there is already a pending error on the buffer.  It
1361 * returns the I/O error code, if any, or 0 if there was no error.
1362 */
1363int
1364xfs_buf_iowait(
1365        xfs_buf_t               *bp)
1366{
1367        trace_xfs_buf_iowait(bp, _RET_IP_);
1368
1369        if (!bp->b_error)
1370                wait_for_completion(&bp->b_iowait);
1371
1372        trace_xfs_buf_iowait_done(bp, _RET_IP_);
1373        return bp->b_error;
1374}
1375
1376xfs_caddr_t
1377xfs_buf_offset(
1378        xfs_buf_t               *bp,
1379        size_t                  offset)
1380{
1381        struct page             *page;
1382
1383        if (bp->b_addr)
1384                return bp->b_addr + offset;
1385
1386        offset += bp->b_offset;
1387        page = bp->b_pages[offset >> PAGE_SHIFT];
1388        return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1389}
1390
1391/*
1392 *      Move data into or out of a buffer.
1393 */
1394void
1395xfs_buf_iomove(
1396        xfs_buf_t               *bp,    /* buffer to process            */
1397        size_t                  boff,   /* starting buffer offset       */
1398        size_t                  bsize,  /* length to copy               */
1399        void                    *data,  /* data address                 */
1400        xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1401{
1402        size_t                  bend;
1403
1404        bend = boff + bsize;
1405        while (boff < bend) {
1406                struct page     *page;
1407                int             page_index, page_offset, csize;
1408
1409                page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1410                page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1411                page = bp->b_pages[page_index];
1412                csize = min_t(size_t, PAGE_SIZE - page_offset,
1413                                      BBTOB(bp->b_io_length) - boff);
1414
1415                ASSERT((csize + page_offset) <= PAGE_SIZE);
1416
1417                switch (mode) {
1418                case XBRW_ZERO:
1419                        memset(page_address(page) + page_offset, 0, csize);
1420                        break;
1421                case XBRW_READ:
1422                        memcpy(data, page_address(page) + page_offset, csize);
1423                        break;
1424                case XBRW_WRITE:
1425                        memcpy(page_address(page) + page_offset, data, csize);
1426                }
1427
1428                boff += csize;
1429                data += csize;
1430        }
1431}
1432
1433/*
1434 *      Handling of buffer targets (buftargs).
1435 */
1436
1437/*
1438 * Wait for any bufs with callbacks that have been submitted but have not yet
1439 * returned. These buffers will have an elevated hold count, so wait on those
1440 * while freeing all the buffers only held by the LRU.
1441 */
1442void
1443xfs_wait_buftarg(
1444        struct xfs_buftarg      *btp)
1445{
1446        struct xfs_buf          *bp;
1447
1448restart:
1449        spin_lock(&btp->bt_lru_lock);
1450        while (!list_empty(&btp->bt_lru)) {
1451                bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1452                if (atomic_read(&bp->b_hold) > 1) {
1453                        spin_unlock(&btp->bt_lru_lock);
1454                        delay(100);
1455                        goto restart;
1456                }
1457                /*
1458                 * clear the LRU reference count so the buffer doesn't get
1459                 * ignored in xfs_buf_rele().
1460                 */
1461                atomic_set(&bp->b_lru_ref, 0);
1462                spin_unlock(&btp->bt_lru_lock);
1463                xfs_buf_rele(bp);
1464                spin_lock(&btp->bt_lru_lock);
1465        }
1466        spin_unlock(&btp->bt_lru_lock);
1467}
1468
1469int
1470xfs_buftarg_shrink(
1471        struct shrinker         *shrink,
1472        struct shrink_control   *sc)
1473{
1474        struct xfs_buftarg      *btp = container_of(shrink,
1475                                        struct xfs_buftarg, bt_shrinker);
1476        struct xfs_buf          *bp;
1477        int nr_to_scan = sc->nr_to_scan;
1478        LIST_HEAD(dispose);
1479
1480        if (!nr_to_scan)
1481                return btp->bt_lru_nr;
1482
1483        spin_lock(&btp->bt_lru_lock);
1484        while (!list_empty(&btp->bt_lru)) {
1485                if (nr_to_scan-- <= 0)
1486                        break;
1487
1488                bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1489
1490                /*
1491                 * Decrement the b_lru_ref count unless the value is already
1492                 * zero. If the value is already zero, we need to reclaim the
1493                 * buffer, otherwise it gets another trip through the LRU.
1494                 */
1495                if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1496                        list_move_tail(&bp->b_lru, &btp->bt_lru);
1497                        continue;
1498                }
1499
1500                /*
1501                 * remove the buffer from the LRU now to avoid needing another
1502                 * lock round trip inside xfs_buf_rele().
1503                 */
1504                list_move(&bp->b_lru, &dispose);
1505                btp->bt_lru_nr--;
1506                bp->b_lru_flags |= _XBF_LRU_DISPOSE;
1507        }
1508        spin_unlock(&btp->bt_lru_lock);
1509
1510        while (!list_empty(&dispose)) {
1511                bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1512                list_del_init(&bp->b_lru);
1513                xfs_buf_rele(bp);
1514        }
1515
1516        return btp->bt_lru_nr;
1517}
1518
1519void
1520xfs_free_buftarg(
1521        struct xfs_mount        *mp,
1522        struct xfs_buftarg      *btp)
1523{
1524        unregister_shrinker(&btp->bt_shrinker);
1525
1526        if (mp->m_flags & XFS_MOUNT_BARRIER)
1527                xfs_blkdev_issue_flush(btp);
1528
1529        kmem_free(btp);
1530}
1531
1532STATIC int
1533xfs_setsize_buftarg_flags(
1534        xfs_buftarg_t           *btp,
1535        unsigned int            blocksize,
1536        unsigned int            sectorsize,
1537        int                     verbose)
1538{
1539        btp->bt_bsize = blocksize;
1540        btp->bt_sshift = ffs(sectorsize) - 1;
1541        btp->bt_smask = sectorsize - 1;
1542
1543        if (set_blocksize(btp->bt_bdev, sectorsize)) {
1544                char name[BDEVNAME_SIZE];
1545
1546                bdevname(btp->bt_bdev, name);
1547
1548                xfs_warn(btp->bt_mount,
1549                        "Cannot set_blocksize to %u on device %s\n",
1550                        sectorsize, name);
1551                return EINVAL;
1552        }
1553
1554        return 0;
1555}
1556
1557/*
1558 *      When allocating the initial buffer target we have not yet
1559 *      read in the superblock, so don't know what sized sectors
1560 *      are being used is at this early stage.  Play safe.
1561 */
1562STATIC int
1563xfs_setsize_buftarg_early(
1564        xfs_buftarg_t           *btp,
1565        struct block_device     *bdev)
1566{
1567        return xfs_setsize_buftarg_flags(btp,
1568                        PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1569}
1570
1571int
1572xfs_setsize_buftarg(
1573        xfs_buftarg_t           *btp,
1574        unsigned int            blocksize,
1575        unsigned int            sectorsize)
1576{
1577        return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1578}
1579
1580xfs_buftarg_t *
1581xfs_alloc_buftarg(
1582        struct xfs_mount        *mp,
1583        struct block_device     *bdev,
1584        int                     external,
1585        const char              *fsname)
1586{
1587        xfs_buftarg_t           *btp;
1588
1589        btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1590
1591        btp->bt_mount = mp;
1592        btp->bt_dev =  bdev->bd_dev;
1593        btp->bt_bdev = bdev;
1594        btp->bt_bdi = blk_get_backing_dev_info(bdev);
1595        if (!btp->bt_bdi)
1596                goto error;
1597
1598        INIT_LIST_HEAD(&btp->bt_lru);
1599        spin_lock_init(&btp->bt_lru_lock);
1600        if (xfs_setsize_buftarg_early(btp, bdev))
1601                goto error;
1602        btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1603        btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1604        register_shrinker(&btp->bt_shrinker);
1605        return btp;
1606
1607error:
1608        kmem_free(btp);
1609        return NULL;
1610}
1611
1612/*
1613 * Add a buffer to the delayed write list.
1614 *
1615 * This queues a buffer for writeout if it hasn't already been.  Note that
1616 * neither this routine nor the buffer list submission functions perform
1617 * any internal synchronization.  It is expected that the lists are thread-local
1618 * to the callers.
1619 *
1620 * Returns true if we queued up the buffer, or false if it already had
1621 * been on the buffer list.
1622 */
1623bool
1624xfs_buf_delwri_queue(
1625        struct xfs_buf          *bp,
1626        struct list_head        *list)
1627{
1628        ASSERT(xfs_buf_islocked(bp));
1629        ASSERT(!(bp->b_flags & XBF_READ));
1630
1631        /*
1632         * If the buffer is already marked delwri it already is queued up
1633         * by someone else for imediate writeout.  Just ignore it in that
1634         * case.
1635         */
1636        if (bp->b_flags & _XBF_DELWRI_Q) {
1637                trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1638                return false;
1639        }
1640
1641        trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1642
1643        /*
1644         * If a buffer gets written out synchronously or marked stale while it
1645         * is on a delwri list we lazily remove it. To do this, the other party
1646         * clears the  _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1647         * It remains referenced and on the list.  In a rare corner case it
1648         * might get readded to a delwri list after the synchronous writeout, in
1649         * which case we need just need to re-add the flag here.
1650         */
1651        bp->b_flags |= _XBF_DELWRI_Q;
1652        if (list_empty(&bp->b_list)) {
1653                atomic_inc(&bp->b_hold);
1654                list_add_tail(&bp->b_list, list);
1655        }
1656
1657        return true;
1658}
1659
1660/*
1661 * Compare function is more complex than it needs to be because
1662 * the return value is only 32 bits and we are doing comparisons
1663 * on 64 bit values
1664 */
1665static int
1666xfs_buf_cmp(
1667        void            *priv,
1668        struct list_head *a,
1669        struct list_head *b)
1670{
1671        struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1672        struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1673        xfs_daddr_t             diff;
1674
1675        diff = ap->b_map.bm_bn - bp->b_map.bm_bn;
1676        if (diff < 0)
1677                return -1;
1678        if (diff > 0)
1679                return 1;
1680        return 0;
1681}
1682
1683static int
1684__xfs_buf_delwri_submit(
1685        struct list_head        *buffer_list,
1686        struct list_head        *io_list,
1687        bool                    wait)
1688{
1689        struct blk_plug         plug;
1690        struct xfs_buf          *bp, *n;
1691        int                     pinned = 0;
1692
1693        list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1694                if (!wait) {
1695                        if (xfs_buf_ispinned(bp)) {
1696                                pinned++;
1697                                continue;
1698                        }
1699                        if (!xfs_buf_trylock(bp))
1700                                continue;
1701                } else {
1702                        xfs_buf_lock(bp);
1703                }
1704
1705                /*
1706                 * Someone else might have written the buffer synchronously or
1707                 * marked it stale in the meantime.  In that case only the
1708                 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1709                 * reference and remove it from the list here.
1710                 */
1711                if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1712                        list_del_init(&bp->b_list);
1713                        xfs_buf_relse(bp);
1714                        continue;
1715                }
1716
1717                list_move_tail(&bp->b_list, io_list);
1718                trace_xfs_buf_delwri_split(bp, _RET_IP_);
1719        }
1720
1721        list_sort(NULL, io_list, xfs_buf_cmp);
1722
1723        blk_start_plug(&plug);
1724        list_for_each_entry_safe(bp, n, io_list, b_list) {
1725                bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
1726                bp->b_flags |= XBF_WRITE;
1727
1728                if (!wait) {
1729                        bp->b_flags |= XBF_ASYNC;
1730                        list_del_init(&bp->b_list);
1731                }
1732                xfs_bdstrat_cb(bp);
1733        }
1734        blk_finish_plug(&plug);
1735
1736        return pinned;
1737}
1738
1739/*
1740 * Write out a buffer list asynchronously.
1741 *
1742 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1743 * out and not wait for I/O completion on any of the buffers.  This interface
1744 * is only safely useable for callers that can track I/O completion by higher
1745 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1746 * function.
1747 */
1748int
1749xfs_buf_delwri_submit_nowait(
1750        struct list_head        *buffer_list)
1751{
1752        LIST_HEAD               (io_list);
1753        return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1754}
1755
1756/*
1757 * Write out a buffer list synchronously.
1758 *
1759 * This will take the @buffer_list, write all buffers out and wait for I/O
1760 * completion on all of the buffers. @buffer_list is consumed by the function,
1761 * so callers must have some other way of tracking buffers if they require such
1762 * functionality.
1763 */
1764int
1765xfs_buf_delwri_submit(
1766        struct list_head        *buffer_list)
1767{
1768        LIST_HEAD               (io_list);
1769        int                     error = 0, error2;
1770        struct xfs_buf          *bp;
1771
1772        __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1773
1774        /* Wait for IO to complete. */
1775        while (!list_empty(&io_list)) {
1776                bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1777
1778                list_del_init(&bp->b_list);
1779                error2 = xfs_buf_iowait(bp);
1780                xfs_buf_relse(bp);
1781                if (!error)
1782                        error = error2;
1783        }
1784
1785        return error;
1786}
1787
1788int __init
1789xfs_buf_init(void)
1790{
1791        xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1792                                                KM_ZONE_HWALIGN, NULL);
1793        if (!xfs_buf_zone)
1794                goto out;
1795
1796        xfslogd_workqueue = alloc_workqueue("xfslogd",
1797                                        WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1798        if (!xfslogd_workqueue)
1799                goto out_free_buf_zone;
1800
1801        return 0;
1802
1803 out_free_buf_zone:
1804        kmem_zone_destroy(xfs_buf_zone);
1805 out:
1806        return -ENOMEM;
1807}
1808
1809void
1810xfs_buf_terminate(void)
1811{
1812        destroy_workqueue(xfslogd_workqueue);
1813        kmem_zone_destroy(xfs_buf_zone);
1814}
1815
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.