linux/fs/ceph/addr.c
<<
>>
Prefs
   1#include <linux/ceph/ceph_debug.h>
   2
   3#include <linux/backing-dev.h>
   4#include <linux/fs.h>
   5#include <linux/mm.h>
   6#include <linux/pagemap.h>
   7#include <linux/writeback.h>    /* generic_writepages */
   8#include <linux/slab.h>
   9#include <linux/pagevec.h>
  10#include <linux/task_io_accounting_ops.h>
  11
  12#include "super.h"
  13#include "mds_client.h"
  14#include <linux/ceph/osd_client.h>
  15
  16/*
  17 * Ceph address space ops.
  18 *
  19 * There are a few funny things going on here.
  20 *
  21 * The page->private field is used to reference a struct
  22 * ceph_snap_context for _every_ dirty page.  This indicates which
  23 * snapshot the page was logically dirtied in, and thus which snap
  24 * context needs to be associated with the osd write during writeback.
  25 *
  26 * Similarly, struct ceph_inode_info maintains a set of counters to
  27 * count dirty pages on the inode.  In the absence of snapshots,
  28 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  29 *
  30 * When a snapshot is taken (that is, when the client receives
  31 * notification that a snapshot was taken), each inode with caps and
  32 * with dirty pages (dirty pages implies there is a cap) gets a new
  33 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  34 * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  35 * moved to capsnap->dirty. (Unless a sync write is currently in
  36 * progress.  In that case, the capsnap is said to be "pending", new
  37 * writes cannot start, and the capsnap isn't "finalized" until the
  38 * write completes (or fails) and a final size/mtime for the inode for
  39 * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  40 *
  41 * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  42 * we look for the first capsnap in i_cap_snaps and write out pages in
  43 * that snap context _only_.  Then we move on to the next capsnap,
  44 * eventually reaching the "live" or "head" context (i.e., pages that
  45 * are not yet snapped) and are writing the most recently dirtied
  46 * pages.
  47 *
  48 * Invalidate and so forth must take care to ensure the dirty page
  49 * accounting is preserved.
  50 */
  51
  52#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  53#define CONGESTION_OFF_THRESH(congestion_kb)                            \
  54        (CONGESTION_ON_THRESH(congestion_kb) -                          \
  55         (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  56
  57static inline struct ceph_snap_context *page_snap_context(struct page *page)
  58{
  59        if (PagePrivate(page))
  60                return (void *)page->private;
  61        return NULL;
  62}
  63
  64/*
  65 * Dirty a page.  Optimistically adjust accounting, on the assumption
  66 * that we won't race with invalidate.  If we do, readjust.
  67 */
  68static int ceph_set_page_dirty(struct page *page)
  69{
  70        struct address_space *mapping = page->mapping;
  71        struct inode *inode;
  72        struct ceph_inode_info *ci;
  73        int undo = 0;
  74        struct ceph_snap_context *snapc;
  75
  76        if (unlikely(!mapping))
  77                return !TestSetPageDirty(page);
  78
  79        if (TestSetPageDirty(page)) {
  80                dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  81                     mapping->host, page, page->index);
  82                return 0;
  83        }
  84
  85        inode = mapping->host;
  86        ci = ceph_inode(inode);
  87
  88        /*
  89         * Note that we're grabbing a snapc ref here without holding
  90         * any locks!
  91         */
  92        snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  93
  94        /* dirty the head */
  95        spin_lock(&ci->i_ceph_lock);
  96        if (ci->i_head_snapc == NULL)
  97                ci->i_head_snapc = ceph_get_snap_context(snapc);
  98        ++ci->i_wrbuffer_ref_head;
  99        if (ci->i_wrbuffer_ref == 0)
 100                ihold(inode);
 101        ++ci->i_wrbuffer_ref;
 102        dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 103             "snapc %p seq %lld (%d snaps)\n",
 104             mapping->host, page, page->index,
 105             ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 106             ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 107             snapc, snapc->seq, snapc->num_snaps);
 108        spin_unlock(&ci->i_ceph_lock);
 109
 110        /* now adjust page */
 111        spin_lock_irq(&mapping->tree_lock);
 112        if (page->mapping) {    /* Race with truncate? */
 113                WARN_ON_ONCE(!PageUptodate(page));
 114                account_page_dirtied(page, page->mapping);
 115                radix_tree_tag_set(&mapping->page_tree,
 116                                page_index(page), PAGECACHE_TAG_DIRTY);
 117
 118                /*
 119                 * Reference snap context in page->private.  Also set
 120                 * PagePrivate so that we get invalidatepage callback.
 121                 */
 122                page->private = (unsigned long)snapc;
 123                SetPagePrivate(page);
 124        } else {
 125                dout("ANON set_page_dirty %p (raced truncate?)\n", page);
 126                undo = 1;
 127        }
 128
 129        spin_unlock_irq(&mapping->tree_lock);
 130
 131        if (undo)
 132                /* whoops, we failed to dirty the page */
 133                ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 134
 135        __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 136
 137        BUG_ON(!PageDirty(page));
 138        return 1;
 139}
 140
 141/*
 142 * If we are truncating the full page (i.e. offset == 0), adjust the
 143 * dirty page counters appropriately.  Only called if there is private
 144 * data on the page.
 145 */
 146static void ceph_invalidatepage(struct page *page, unsigned long offset)
 147{
 148        struct inode *inode;
 149        struct ceph_inode_info *ci;
 150        struct ceph_snap_context *snapc = page_snap_context(page);
 151
 152        BUG_ON(!PageLocked(page));
 153        BUG_ON(!PagePrivate(page));
 154        BUG_ON(!page->mapping);
 155
 156        inode = page->mapping->host;
 157
 158        /*
 159         * We can get non-dirty pages here due to races between
 160         * set_page_dirty and truncate_complete_page; just spit out a
 161         * warning, in case we end up with accounting problems later.
 162         */
 163        if (!PageDirty(page))
 164                pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 165
 166        if (offset == 0)
 167                ClearPageChecked(page);
 168
 169        ci = ceph_inode(inode);
 170        if (offset == 0) {
 171                dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
 172                     inode, page, page->index, offset);
 173                ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 174                ceph_put_snap_context(snapc);
 175                page->private = 0;
 176                ClearPagePrivate(page);
 177        } else {
 178                dout("%p invalidatepage %p idx %lu partial dirty page\n",
 179                     inode, page, page->index);
 180        }
 181}
 182
 183/* just a sanity check */
 184static int ceph_releasepage(struct page *page, gfp_t g)
 185{
 186        struct inode *inode = page->mapping ? page->mapping->host : NULL;
 187        dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 188        WARN_ON(PageDirty(page));
 189        WARN_ON(PagePrivate(page));
 190        return 0;
 191}
 192
 193/*
 194 * read a single page, without unlocking it.
 195 */
 196static int readpage_nounlock(struct file *filp, struct page *page)
 197{
 198        struct inode *inode = filp->f_dentry->d_inode;
 199        struct ceph_inode_info *ci = ceph_inode(inode);
 200        struct ceph_osd_client *osdc = 
 201                &ceph_inode_to_client(inode)->client->osdc;
 202        int err = 0;
 203        u64 len = PAGE_CACHE_SIZE;
 204
 205        dout("readpage inode %p file %p page %p index %lu\n",
 206             inode, filp, page, page->index);
 207        err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 208                                  (u64) page_offset(page), &len,
 209                                  ci->i_truncate_seq, ci->i_truncate_size,
 210                                  &page, 1, 0);
 211        if (err == -ENOENT)
 212                err = 0;
 213        if (err < 0) {
 214                SetPageError(page);
 215                goto out;
 216        } else if (err < PAGE_CACHE_SIZE) {
 217                /* zero fill remainder of page */
 218                zero_user_segment(page, err, PAGE_CACHE_SIZE);
 219        }
 220        SetPageUptodate(page);
 221
 222out:
 223        return err < 0 ? err : 0;
 224}
 225
 226static int ceph_readpage(struct file *filp, struct page *page)
 227{
 228        int r = readpage_nounlock(filp, page);
 229        unlock_page(page);
 230        return r;
 231}
 232
 233/*
 234 * Finish an async read(ahead) op.
 235 */
 236static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 237{
 238        struct inode *inode = req->r_inode;
 239        struct ceph_osd_reply_head *replyhead;
 240        int rc, bytes;
 241        int i;
 242
 243        /* parse reply */
 244        replyhead = msg->front.iov_base;
 245        WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 246        rc = le32_to_cpu(replyhead->result);
 247        bytes = le32_to_cpu(msg->hdr.data_len);
 248
 249        dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 250
 251        /* unlock all pages, zeroing any data we didn't read */
 252        for (i = 0; i < req->r_num_pages; i++, bytes -= PAGE_CACHE_SIZE) {
 253                struct page *page = req->r_pages[i];
 254
 255                if (bytes < (int)PAGE_CACHE_SIZE) {
 256                        /* zero (remainder of) page */
 257                        int s = bytes < 0 ? 0 : bytes;
 258                        zero_user_segment(page, s, PAGE_CACHE_SIZE);
 259                }
 260                dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 261                     page->index);
 262                flush_dcache_page(page);
 263                SetPageUptodate(page);
 264                unlock_page(page);
 265                page_cache_release(page);
 266        }
 267        kfree(req->r_pages);
 268}
 269
 270static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 271{
 272        int i;
 273
 274        for (i = 0; i < num_pages; i++)
 275                unlock_page(pages[i]);
 276}
 277
 278/*
 279 * start an async read(ahead) operation.  return nr_pages we submitted
 280 * a read for on success, or negative error code.
 281 */
 282static int start_read(struct inode *inode, struct list_head *page_list, int max)
 283{
 284        struct ceph_osd_client *osdc =
 285                &ceph_inode_to_client(inode)->client->osdc;
 286        struct ceph_inode_info *ci = ceph_inode(inode);
 287        struct page *page = list_entry(page_list->prev, struct page, lru);
 288        struct ceph_osd_request *req;
 289        u64 off;
 290        u64 len;
 291        int i;
 292        struct page **pages;
 293        pgoff_t next_index;
 294        int nr_pages = 0;
 295        int ret;
 296
 297        off = (u64) page_offset(page);
 298
 299        /* count pages */
 300        next_index = page->index;
 301        list_for_each_entry_reverse(page, page_list, lru) {
 302                if (page->index != next_index)
 303                        break;
 304                nr_pages++;
 305                next_index++;
 306                if (max && nr_pages == max)
 307                        break;
 308        }
 309        len = nr_pages << PAGE_CACHE_SHIFT;
 310        dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 311             off, len);
 312
 313        req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
 314                                    off, &len,
 315                                    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 316                                    NULL, 0,
 317                                    ci->i_truncate_seq, ci->i_truncate_size,
 318                                    NULL, false, 1, 0);
 319        if (IS_ERR(req))
 320                return PTR_ERR(req);
 321
 322        /* build page vector */
 323        nr_pages = len >> PAGE_CACHE_SHIFT;
 324        pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 325        ret = -ENOMEM;
 326        if (!pages)
 327                goto out;
 328        for (i = 0; i < nr_pages; ++i) {
 329                page = list_entry(page_list->prev, struct page, lru);
 330                BUG_ON(PageLocked(page));
 331                list_del(&page->lru);
 332                
 333                dout("start_read %p adding %p idx %lu\n", inode, page,
 334                     page->index);
 335                if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 336                                          GFP_NOFS)) {
 337                        page_cache_release(page);
 338                        dout("start_read %p add_to_page_cache failed %p\n",
 339                             inode, page);
 340                        nr_pages = i;
 341                        goto out_pages;
 342                }
 343                pages[i] = page;
 344        }
 345        req->r_pages = pages;
 346        req->r_num_pages = nr_pages;
 347        req->r_callback = finish_read;
 348        req->r_inode = inode;
 349
 350        dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 351        ret = ceph_osdc_start_request(osdc, req, false);
 352        if (ret < 0)
 353                goto out_pages;
 354        ceph_osdc_put_request(req);
 355        return nr_pages;
 356
 357out_pages:
 358        ceph_unlock_page_vector(pages, nr_pages);
 359        ceph_release_page_vector(pages, nr_pages);
 360out:
 361        ceph_osdc_put_request(req);
 362        return ret;
 363}
 364
 365
 366/*
 367 * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 368 * the caller (VM) cleans them up.
 369 */
 370static int ceph_readpages(struct file *file, struct address_space *mapping,
 371                          struct list_head *page_list, unsigned nr_pages)
 372{
 373        struct inode *inode = file->f_dentry->d_inode;
 374        struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 375        int rc = 0;
 376        int max = 0;
 377
 378        if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 379                max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 380                        >> PAGE_SHIFT;
 381
 382        dout("readpages %p file %p nr_pages %d max %d\n", inode, file, nr_pages,
 383             max);
 384        while (!list_empty(page_list)) {
 385                rc = start_read(inode, page_list, max);
 386                if (rc < 0)
 387                        goto out;
 388                BUG_ON(rc == 0);
 389        }
 390out:
 391        dout("readpages %p file %p ret %d\n", inode, file, rc);
 392        return rc;
 393}
 394
 395/*
 396 * Get ref for the oldest snapc for an inode with dirty data... that is, the
 397 * only snap context we are allowed to write back.
 398 */
 399static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 400                                                    u64 *snap_size)
 401{
 402        struct ceph_inode_info *ci = ceph_inode(inode);
 403        struct ceph_snap_context *snapc = NULL;
 404        struct ceph_cap_snap *capsnap = NULL;
 405
 406        spin_lock(&ci->i_ceph_lock);
 407        list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 408                dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 409                     capsnap->context, capsnap->dirty_pages);
 410                if (capsnap->dirty_pages) {
 411                        snapc = ceph_get_snap_context(capsnap->context);
 412                        if (snap_size)
 413                                *snap_size = capsnap->size;
 414                        break;
 415                }
 416        }
 417        if (!snapc && ci->i_wrbuffer_ref_head) {
 418                snapc = ceph_get_snap_context(ci->i_head_snapc);
 419                dout(" head snapc %p has %d dirty pages\n",
 420                     snapc, ci->i_wrbuffer_ref_head);
 421        }
 422        spin_unlock(&ci->i_ceph_lock);
 423        return snapc;
 424}
 425
 426/*
 427 * Write a single page, but leave the page locked.
 428 *
 429 * If we get a write error, set the page error bit, but still adjust the
 430 * dirty page accounting (i.e., page is no longer dirty).
 431 */
 432static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 433{
 434        struct inode *inode;
 435        struct ceph_inode_info *ci;
 436        struct ceph_fs_client *fsc;
 437        struct ceph_osd_client *osdc;
 438        loff_t page_off = page_offset(page);
 439        int len = PAGE_CACHE_SIZE;
 440        loff_t i_size;
 441        int err = 0;
 442        struct ceph_snap_context *snapc, *oldest;
 443        u64 snap_size = 0;
 444        long writeback_stat;
 445
 446        dout("writepage %p idx %lu\n", page, page->index);
 447
 448        if (!page->mapping || !page->mapping->host) {
 449                dout("writepage %p - no mapping\n", page);
 450                return -EFAULT;
 451        }
 452        inode = page->mapping->host;
 453        ci = ceph_inode(inode);
 454        fsc = ceph_inode_to_client(inode);
 455        osdc = &fsc->client->osdc;
 456
 457        /* verify this is a writeable snap context */
 458        snapc = page_snap_context(page);
 459        if (snapc == NULL) {
 460                dout("writepage %p page %p not dirty?\n", inode, page);
 461                goto out;
 462        }
 463        oldest = get_oldest_context(inode, &snap_size);
 464        if (snapc->seq > oldest->seq) {
 465                dout("writepage %p page %p snapc %p not writeable - noop\n",
 466                     inode, page, snapc);
 467                /* we should only noop if called by kswapd */
 468                WARN_ON((current->flags & PF_MEMALLOC) == 0);
 469                ceph_put_snap_context(oldest);
 470                goto out;
 471        }
 472        ceph_put_snap_context(oldest);
 473
 474        /* is this a partial page at end of file? */
 475        if (snap_size)
 476                i_size = snap_size;
 477        else
 478                i_size = i_size_read(inode);
 479        if (i_size < page_off + len)
 480                len = i_size - page_off;
 481
 482        dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 483             inode, page, page->index, page_off, len, snapc);
 484
 485        writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 486        if (writeback_stat >
 487            CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 488                set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 489
 490        set_page_writeback(page);
 491        err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 492                                   &ci->i_layout, snapc,
 493                                   page_off, len,
 494                                   ci->i_truncate_seq, ci->i_truncate_size,
 495                                   &inode->i_mtime,
 496                                   &page, 1, 0, 0, true);
 497        if (err < 0) {
 498                dout("writepage setting page/mapping error %d %p\n", err, page);
 499                SetPageError(page);
 500                mapping_set_error(&inode->i_data, err);
 501                if (wbc)
 502                        wbc->pages_skipped++;
 503        } else {
 504                dout("writepage cleaned page %p\n", page);
 505                err = 0;  /* vfs expects us to return 0 */
 506        }
 507        page->private = 0;
 508        ClearPagePrivate(page);
 509        end_page_writeback(page);
 510        ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 511        ceph_put_snap_context(snapc);  /* page's reference */
 512out:
 513        return err;
 514}
 515
 516static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 517{
 518        int err;
 519        struct inode *inode = page->mapping->host;
 520        BUG_ON(!inode);
 521        ihold(inode);
 522        err = writepage_nounlock(page, wbc);
 523        unlock_page(page);
 524        iput(inode);
 525        return err;
 526}
 527
 528
 529/*
 530 * lame release_pages helper.  release_pages() isn't exported to
 531 * modules.
 532 */
 533static void ceph_release_pages(struct page **pages, int num)
 534{
 535        struct pagevec pvec;
 536        int i;
 537
 538        pagevec_init(&pvec, 0);
 539        for (i = 0; i < num; i++) {
 540                if (pagevec_add(&pvec, pages[i]) == 0)
 541                        pagevec_release(&pvec);
 542        }
 543        pagevec_release(&pvec);
 544}
 545
 546
 547/*
 548 * async writeback completion handler.
 549 *
 550 * If we get an error, set the mapping error bit, but not the individual
 551 * page error bits.
 552 */
 553static void writepages_finish(struct ceph_osd_request *req,
 554                              struct ceph_msg *msg)
 555{
 556        struct inode *inode = req->r_inode;
 557        struct ceph_osd_reply_head *replyhead;
 558        struct ceph_osd_op *op;
 559        struct ceph_inode_info *ci = ceph_inode(inode);
 560        unsigned wrote;
 561        struct page *page;
 562        int i;
 563        struct ceph_snap_context *snapc = req->r_snapc;
 564        struct address_space *mapping = inode->i_mapping;
 565        __s32 rc = -EIO;
 566        u64 bytes = 0;
 567        struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 568        long writeback_stat;
 569        unsigned issued = ceph_caps_issued(ci);
 570
 571        /* parse reply */
 572        replyhead = msg->front.iov_base;
 573        WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
 574        op = (void *)(replyhead + 1);
 575        rc = le32_to_cpu(replyhead->result);
 576        bytes = le64_to_cpu(op->extent.length);
 577
 578        if (rc >= 0) {
 579                /*
 580                 * Assume we wrote the pages we originally sent.  The
 581                 * osd might reply with fewer pages if our writeback
 582                 * raced with a truncation and was adjusted at the osd,
 583                 * so don't believe the reply.
 584                 */
 585                wrote = req->r_num_pages;
 586        } else {
 587                wrote = 0;
 588                mapping_set_error(mapping, rc);
 589        }
 590        dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 591             inode, rc, bytes, wrote);
 592
 593        /* clean all pages */
 594        for (i = 0; i < req->r_num_pages; i++) {
 595                page = req->r_pages[i];
 596                BUG_ON(!page);
 597                WARN_ON(!PageUptodate(page));
 598
 599                writeback_stat =
 600                        atomic_long_dec_return(&fsc->writeback_count);
 601                if (writeback_stat <
 602                    CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 603                        clear_bdi_congested(&fsc->backing_dev_info,
 604                                            BLK_RW_ASYNC);
 605
 606                ceph_put_snap_context(page_snap_context(page));
 607                page->private = 0;
 608                ClearPagePrivate(page);
 609                dout("unlocking %d %p\n", i, page);
 610                end_page_writeback(page);
 611
 612                /*
 613                 * We lost the cache cap, need to truncate the page before
 614                 * it is unlocked, otherwise we'd truncate it later in the
 615                 * page truncation thread, possibly losing some data that
 616                 * raced its way in
 617                 */
 618                if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 619                        generic_error_remove_page(inode->i_mapping, page);
 620
 621                unlock_page(page);
 622        }
 623        dout("%p wrote+cleaned %d pages\n", inode, wrote);
 624        ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
 625
 626        ceph_release_pages(req->r_pages, req->r_num_pages);
 627        if (req->r_pages_from_pool)
 628                mempool_free(req->r_pages,
 629                             ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 630        else
 631                kfree(req->r_pages);
 632        ceph_osdc_put_request(req);
 633}
 634
 635/*
 636 * allocate a page vec, either directly, or if necessary, via a the
 637 * mempool.  we avoid the mempool if we can because req->r_num_pages
 638 * may be less than the maximum write size.
 639 */
 640static void alloc_page_vec(struct ceph_fs_client *fsc,
 641                           struct ceph_osd_request *req)
 642{
 643        req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
 644                               GFP_NOFS);
 645        if (!req->r_pages) {
 646                req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
 647                req->r_pages_from_pool = 1;
 648                WARN_ON(!req->r_pages);
 649        }
 650}
 651
 652/*
 653 * initiate async writeback
 654 */
 655static int ceph_writepages_start(struct address_space *mapping,
 656                                 struct writeback_control *wbc)
 657{
 658        struct inode *inode = mapping->host;
 659        struct ceph_inode_info *ci = ceph_inode(inode);
 660        struct ceph_fs_client *fsc;
 661        pgoff_t index, start, end;
 662        int range_whole = 0;
 663        int should_loop = 1;
 664        pgoff_t max_pages = 0, max_pages_ever = 0;
 665        struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 666        struct pagevec pvec;
 667        int done = 0;
 668        int rc = 0;
 669        unsigned wsize = 1 << inode->i_blkbits;
 670        struct ceph_osd_request *req = NULL;
 671        int do_sync;
 672        u64 snap_size = 0;
 673
 674        /*
 675         * Include a 'sync' in the OSD request if this is a data
 676         * integrity write (e.g., O_SYNC write or fsync()), or if our
 677         * cap is being revoked.
 678         */
 679        do_sync = wbc->sync_mode == WB_SYNC_ALL;
 680        if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 681                do_sync = 1;
 682        dout("writepages_start %p dosync=%d (mode=%s)\n",
 683             inode, do_sync,
 684             wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 685             (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 686
 687        fsc = ceph_inode_to_client(inode);
 688        if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 689                pr_warning("writepage_start %p on forced umount\n", inode);
 690                return -EIO; /* we're in a forced umount, don't write! */
 691        }
 692        if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 693                wsize = fsc->mount_options->wsize;
 694        if (wsize < PAGE_CACHE_SIZE)
 695                wsize = PAGE_CACHE_SIZE;
 696        max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 697
 698        pagevec_init(&pvec, 0);
 699
 700        /* where to start/end? */
 701        if (wbc->range_cyclic) {
 702                start = mapping->writeback_index; /* Start from prev offset */
 703                end = -1;
 704                dout(" cyclic, start at %lu\n", start);
 705        } else {
 706                start = wbc->range_start >> PAGE_CACHE_SHIFT;
 707                end = wbc->range_end >> PAGE_CACHE_SHIFT;
 708                if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 709                        range_whole = 1;
 710                should_loop = 0;
 711                dout(" not cyclic, %lu to %lu\n", start, end);
 712        }
 713        index = start;
 714
 715retry:
 716        /* find oldest snap context with dirty data */
 717        ceph_put_snap_context(snapc);
 718        snapc = get_oldest_context(inode, &snap_size);
 719        if (!snapc) {
 720                /* hmm, why does writepages get called when there
 721                   is no dirty data? */
 722                dout(" no snap context with dirty data?\n");
 723                goto out;
 724        }
 725        dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 726             snapc, snapc->seq, snapc->num_snaps);
 727        if (last_snapc && snapc != last_snapc) {
 728                /* if we switched to a newer snapc, restart our scan at the
 729                 * start of the original file range. */
 730                dout("  snapc differs from last pass, restarting at %lu\n",
 731                     index);
 732                index = start;
 733        }
 734        last_snapc = snapc;
 735
 736        while (!done && index <= end) {
 737                unsigned i;
 738                int first;
 739                pgoff_t next;
 740                int pvec_pages, locked_pages;
 741                struct page *page;
 742                int want;
 743                u64 offset, len;
 744                struct ceph_osd_request_head *reqhead;
 745                struct ceph_osd_op *op;
 746                long writeback_stat;
 747
 748                next = 0;
 749                locked_pages = 0;
 750                max_pages = max_pages_ever;
 751
 752get_more_pages:
 753                first = -1;
 754                want = min(end - index,
 755                           min((pgoff_t)PAGEVEC_SIZE,
 756                               max_pages - (pgoff_t)locked_pages) - 1)
 757                        + 1;
 758                pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 759                                                PAGECACHE_TAG_DIRTY,
 760                                                want);
 761                dout("pagevec_lookup_tag got %d\n", pvec_pages);
 762                if (!pvec_pages && !locked_pages)
 763                        break;
 764                for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 765                        page = pvec.pages[i];
 766                        dout("? %p idx %lu\n", page, page->index);
 767                        if (locked_pages == 0)
 768                                lock_page(page);  /* first page */
 769                        else if (!trylock_page(page))
 770                                break;
 771
 772                        /* only dirty pages, or our accounting breaks */
 773                        if (unlikely(!PageDirty(page)) ||
 774                            unlikely(page->mapping != mapping)) {
 775                                dout("!dirty or !mapping %p\n", page);
 776                                unlock_page(page);
 777                                break;
 778                        }
 779                        if (!wbc->range_cyclic && page->index > end) {
 780                                dout("end of range %p\n", page);
 781                                done = 1;
 782                                unlock_page(page);
 783                                break;
 784                        }
 785                        if (next && (page->index != next)) {
 786                                dout("not consecutive %p\n", page);
 787                                unlock_page(page);
 788                                break;
 789                        }
 790                        if (wbc->sync_mode != WB_SYNC_NONE) {
 791                                dout("waiting on writeback %p\n", page);
 792                                wait_on_page_writeback(page);
 793                        }
 794                        if ((snap_size && page_offset(page) > snap_size) ||
 795                            (!snap_size &&
 796                             page_offset(page) > i_size_read(inode))) {
 797                                dout("%p page eof %llu\n", page, snap_size ?
 798                                     snap_size : i_size_read(inode));
 799                                done = 1;
 800                                unlock_page(page);
 801                                break;
 802                        }
 803                        if (PageWriteback(page)) {
 804                                dout("%p under writeback\n", page);
 805                                unlock_page(page);
 806                                break;
 807                        }
 808
 809                        /* only if matching snap context */
 810                        pgsnapc = page_snap_context(page);
 811                        if (pgsnapc->seq > snapc->seq) {
 812                                dout("page snapc %p %lld > oldest %p %lld\n",
 813                                     pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 814                                unlock_page(page);
 815                                if (!locked_pages)
 816                                        continue; /* keep looking for snap */
 817                                break;
 818                        }
 819
 820                        if (!clear_page_dirty_for_io(page)) {
 821                                dout("%p !clear_page_dirty_for_io\n", page);
 822                                unlock_page(page);
 823                                break;
 824                        }
 825
 826                        /* ok */
 827                        if (locked_pages == 0) {
 828                                /* prepare async write request */
 829                                offset = (u64) page_offset(page);
 830                                len = wsize;
 831                                req = ceph_osdc_new_request(&fsc->client->osdc,
 832                                            &ci->i_layout,
 833                                            ceph_vino(inode),
 834                                            offset, &len,
 835                                            CEPH_OSD_OP_WRITE,
 836                                            CEPH_OSD_FLAG_WRITE |
 837                                                    CEPH_OSD_FLAG_ONDISK,
 838                                            snapc, do_sync,
 839                                            ci->i_truncate_seq,
 840                                            ci->i_truncate_size,
 841                                            &inode->i_mtime, true, 1, 0);
 842
 843                                if (IS_ERR(req)) {
 844                                        rc = PTR_ERR(req);
 845                                        unlock_page(page);
 846                                        break;
 847                                }
 848
 849                                max_pages = req->r_num_pages;
 850
 851                                alloc_page_vec(fsc, req);
 852                                req->r_callback = writepages_finish;
 853                                req->r_inode = inode;
 854                        }
 855
 856                        /* note position of first page in pvec */
 857                        if (first < 0)
 858                                first = i;
 859                        dout("%p will write page %p idx %lu\n",
 860                             inode, page, page->index);
 861
 862                        writeback_stat =
 863                               atomic_long_inc_return(&fsc->writeback_count);
 864                        if (writeback_stat > CONGESTION_ON_THRESH(
 865                                    fsc->mount_options->congestion_kb)) {
 866                                set_bdi_congested(&fsc->backing_dev_info,
 867                                                  BLK_RW_ASYNC);
 868                        }
 869
 870                        set_page_writeback(page);
 871                        req->r_pages[locked_pages] = page;
 872                        locked_pages++;
 873                        next = page->index + 1;
 874                }
 875
 876                /* did we get anything? */
 877                if (!locked_pages)
 878                        goto release_pvec_pages;
 879                if (i) {
 880                        int j;
 881                        BUG_ON(!locked_pages || first < 0);
 882
 883                        if (pvec_pages && i == pvec_pages &&
 884                            locked_pages < max_pages) {
 885                                dout("reached end pvec, trying for more\n");
 886                                pagevec_reinit(&pvec);
 887                                goto get_more_pages;
 888                        }
 889
 890                        /* shift unused pages over in the pvec...  we
 891                         * will need to release them below. */
 892                        for (j = i; j < pvec_pages; j++) {
 893                                dout(" pvec leftover page %p\n",
 894                                     pvec.pages[j]);
 895                                pvec.pages[j-i+first] = pvec.pages[j];
 896                        }
 897                        pvec.nr -= i-first;
 898                }
 899
 900                /* submit the write */
 901                offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
 902                len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
 903                          (u64)locked_pages << PAGE_CACHE_SHIFT);
 904                dout("writepages got %d pages at %llu~%llu\n",
 905                     locked_pages, offset, len);
 906
 907                /* revise final length, page count */
 908                req->r_num_pages = locked_pages;
 909                reqhead = req->r_request->front.iov_base;
 910                op = (void *)(reqhead + 1);
 911                op->extent.length = cpu_to_le64(len);
 912                op->payload_len = cpu_to_le32(len);
 913                req->r_request->hdr.data_len = cpu_to_le32(len);
 914
 915                rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 916                BUG_ON(rc);
 917                req = NULL;
 918
 919                /* continue? */
 920                index = next;
 921                wbc->nr_to_write -= locked_pages;
 922                if (wbc->nr_to_write <= 0)
 923                        done = 1;
 924
 925release_pvec_pages:
 926                dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 927                     pvec.nr ? pvec.pages[0] : NULL);
 928                pagevec_release(&pvec);
 929
 930                if (locked_pages && !done)
 931                        goto retry;
 932        }
 933
 934        if (should_loop && !done) {
 935                /* more to do; loop back to beginning of file */
 936                dout("writepages looping back to beginning of file\n");
 937                should_loop = 0;
 938                index = 0;
 939                goto retry;
 940        }
 941
 942        if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 943                mapping->writeback_index = index;
 944
 945out:
 946        if (req)
 947                ceph_osdc_put_request(req);
 948        ceph_put_snap_context(snapc);
 949        dout("writepages done, rc = %d\n", rc);
 950        return rc;
 951}
 952
 953
 954
 955/*
 956 * See if a given @snapc is either writeable, or already written.
 957 */
 958static int context_is_writeable_or_written(struct inode *inode,
 959                                           struct ceph_snap_context *snapc)
 960{
 961        struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
 962        int ret = !oldest || snapc->seq <= oldest->seq;
 963
 964        ceph_put_snap_context(oldest);
 965        return ret;
 966}
 967
 968/*
 969 * We are only allowed to write into/dirty the page if the page is
 970 * clean, or already dirty within the same snap context.
 971 *
 972 * called with page locked.
 973 * return success with page locked,
 974 * or any failure (incl -EAGAIN) with page unlocked.
 975 */
 976static int ceph_update_writeable_page(struct file *file,
 977                            loff_t pos, unsigned len,
 978                            struct page *page)
 979{
 980        struct inode *inode = file->f_dentry->d_inode;
 981        struct ceph_inode_info *ci = ceph_inode(inode);
 982        struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 983        loff_t page_off = pos & PAGE_CACHE_MASK;
 984        int pos_in_page = pos & ~PAGE_CACHE_MASK;
 985        int end_in_page = pos_in_page + len;
 986        loff_t i_size;
 987        int r;
 988        struct ceph_snap_context *snapc, *oldest;
 989
 990retry_locked:
 991        /* writepages currently holds page lock, but if we change that later, */
 992        wait_on_page_writeback(page);
 993
 994        /* check snap context */
 995        BUG_ON(!ci->i_snap_realm);
 996        down_read(&mdsc->snap_rwsem);
 997        BUG_ON(!ci->i_snap_realm->cached_context);
 998        snapc = page_snap_context(page);
 999        if (snapc && snapc != ci->i_head_snapc) {
1000                /*
1001                 * this page is already dirty in another (older) snap
1002                 * context!  is it writeable now?
1003                 */
1004                oldest = get_oldest_context(inode, NULL);
1005                up_read(&mdsc->snap_rwsem);
1006
1007                if (snapc->seq > oldest->seq) {
1008                        ceph_put_snap_context(oldest);
1009                        dout(" page %p snapc %p not current or oldest\n",
1010                             page, snapc);
1011                        /*
1012                         * queue for writeback, and wait for snapc to
1013                         * be writeable or written
1014                         */
1015                        snapc = ceph_get_snap_context(snapc);
1016                        unlock_page(page);
1017                        ceph_queue_writeback(inode);
1018                        r = wait_event_interruptible(ci->i_cap_wq,
1019                               context_is_writeable_or_written(inode, snapc));
1020                        ceph_put_snap_context(snapc);
1021                        if (r == -ERESTARTSYS)
1022                                return r;
1023                        return -EAGAIN;
1024                }
1025                ceph_put_snap_context(oldest);
1026
1027                /* yay, writeable, do it now (without dropping page lock) */
1028                dout(" page %p snapc %p not current, but oldest\n",
1029                     page, snapc);
1030                if (!clear_page_dirty_for_io(page))
1031                        goto retry_locked;
1032                r = writepage_nounlock(page, NULL);
1033                if (r < 0)
1034                        goto fail_nosnap;
1035                goto retry_locked;
1036        }
1037
1038        if (PageUptodate(page)) {
1039                dout(" page %p already uptodate\n", page);
1040                return 0;
1041        }
1042
1043        /* full page? */
1044        if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1045                return 0;
1046
1047        /* past end of file? */
1048        i_size = inode->i_size;   /* caller holds i_mutex */
1049
1050        if (i_size + len > inode->i_sb->s_maxbytes) {
1051                /* file is too big */
1052                r = -EINVAL;
1053                goto fail;
1054        }
1055
1056        if (page_off >= i_size ||
1057            (pos_in_page == 0 && (pos+len) >= i_size &&
1058             end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1059                dout(" zeroing %p 0 - %d and %d - %d\n",
1060                     page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1061                zero_user_segments(page,
1062                                   0, pos_in_page,
1063                                   end_in_page, PAGE_CACHE_SIZE);
1064                return 0;
1065        }
1066
1067        /* we need to read it. */
1068        up_read(&mdsc->snap_rwsem);
1069        r = readpage_nounlock(file, page);
1070        if (r < 0)
1071                goto fail_nosnap;
1072        goto retry_locked;
1073
1074fail:
1075        up_read(&mdsc->snap_rwsem);
1076fail_nosnap:
1077        unlock_page(page);
1078        return r;
1079}
1080
1081/*
1082 * We are only allowed to write into/dirty the page if the page is
1083 * clean, or already dirty within the same snap context.
1084 */
1085static int ceph_write_begin(struct file *file, struct address_space *mapping,
1086                            loff_t pos, unsigned len, unsigned flags,
1087                            struct page **pagep, void **fsdata)
1088{
1089        struct inode *inode = file->f_dentry->d_inode;
1090        struct page *page;
1091        pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1092        int r;
1093
1094        do {
1095                /* get a page */
1096                page = grab_cache_page_write_begin(mapping, index, 0);
1097                if (!page)
1098                        return -ENOMEM;
1099                *pagep = page;
1100
1101                dout("write_begin file %p inode %p page %p %d~%d\n", file,
1102                     inode, page, (int)pos, (int)len);
1103
1104                r = ceph_update_writeable_page(file, pos, len, page);
1105        } while (r == -EAGAIN);
1106
1107        return r;
1108}
1109
1110/*
1111 * we don't do anything in here that simple_write_end doesn't do
1112 * except adjust dirty page accounting and drop read lock on
1113 * mdsc->snap_rwsem.
1114 */
1115static int ceph_write_end(struct file *file, struct address_space *mapping,
1116                          loff_t pos, unsigned len, unsigned copied,
1117                          struct page *page, void *fsdata)
1118{
1119        struct inode *inode = file->f_dentry->d_inode;
1120        struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1121        struct ceph_mds_client *mdsc = fsc->mdsc;
1122        unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1123        int check_cap = 0;
1124
1125        dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1126             inode, page, (int)pos, (int)copied, (int)len);
1127
1128        /* zero the stale part of the page if we did a short copy */
1129        if (copied < len)
1130                zero_user_segment(page, from+copied, len);
1131
1132        /* did file size increase? */
1133        /* (no need for i_size_read(); we caller holds i_mutex */
1134        if (pos+copied > inode->i_size)
1135                check_cap = ceph_inode_set_size(inode, pos+copied);
1136
1137        if (!PageUptodate(page))
1138                SetPageUptodate(page);
1139
1140        set_page_dirty(page);
1141
1142        unlock_page(page);
1143        up_read(&mdsc->snap_rwsem);
1144        page_cache_release(page);
1145
1146        if (check_cap)
1147                ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1148
1149        return copied;
1150}
1151
1152/*
1153 * we set .direct_IO to indicate direct io is supported, but since we
1154 * intercept O_DIRECT reads and writes early, this function should
1155 * never get called.
1156 */
1157static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1158                              const struct iovec *iov,
1159                              loff_t pos, unsigned long nr_segs)
1160{
1161        WARN_ON(1);
1162        return -EINVAL;
1163}
1164
1165const struct address_space_operations ceph_aops = {
1166        .readpage = ceph_readpage,
1167        .readpages = ceph_readpages,
1168        .writepage = ceph_writepage,
1169        .writepages = ceph_writepages_start,
1170        .write_begin = ceph_write_begin,
1171        .write_end = ceph_write_end,
1172        .set_page_dirty = ceph_set_page_dirty,
1173        .invalidatepage = ceph_invalidatepage,
1174        .releasepage = ceph_releasepage,
1175        .direct_IO = ceph_direct_io,
1176};
1177
1178
1179/*
1180 * vm ops
1181 */
1182
1183/*
1184 * Reuse write_begin here for simplicity.
1185 */
1186static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1187{
1188        struct inode *inode = vma->vm_file->f_dentry->d_inode;
1189        struct page *page = vmf->page;
1190        struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1191        loff_t off = page_offset(page);
1192        loff_t size, len;
1193        int ret;
1194
1195        /* Update time before taking page lock */
1196        file_update_time(vma->vm_file);
1197
1198        size = i_size_read(inode);
1199        if (off + PAGE_CACHE_SIZE <= size)
1200                len = PAGE_CACHE_SIZE;
1201        else
1202                len = size & ~PAGE_CACHE_MASK;
1203
1204        dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1205             off, len, page, page->index);
1206
1207        lock_page(page);
1208
1209        ret = VM_FAULT_NOPAGE;
1210        if ((off > size) ||
1211            (page->mapping != inode->i_mapping))
1212                goto out;
1213
1214        ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1215        if (ret == 0) {
1216                /* success.  we'll keep the page locked. */
1217                set_page_dirty(page);
1218                up_read(&mdsc->snap_rwsem);
1219                ret = VM_FAULT_LOCKED;
1220        } else {
1221                if (ret == -ENOMEM)
1222                        ret = VM_FAULT_OOM;
1223                else
1224                        ret = VM_FAULT_SIGBUS;
1225        }
1226out:
1227        dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1228        if (ret != VM_FAULT_LOCKED)
1229                unlock_page(page);
1230        return ret;
1231}
1232
1233static struct vm_operations_struct ceph_vmops = {
1234        .fault          = filemap_fault,
1235        .page_mkwrite   = ceph_page_mkwrite,
1236        .remap_pages    = generic_file_remap_pages,
1237};
1238
1239int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1240{
1241        struct address_space *mapping = file->f_mapping;
1242
1243        if (!mapping->a_ops->readpage)
1244                return -ENOEXEC;
1245        file_accessed(file);
1246        vma->vm_ops = &ceph_vmops;
1247        return 0;
1248}
1249
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.