drivers/staging/lustre/lustre/llite/rw26.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2012, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/lustre/llite/rw26.c
  33  *
  34  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  35  */
  36
  37 #include <linux/kernel.h>
  38 #include <linux/mm.h>
  39 #include <linux/string.h>
  40 #include <linux/stat.h>
  41 #include <linux/errno.h>
  42 #include <linux/unistd.h>
  43 #include <linux/uaccess.h>
  44
  45 #include <linux/migrate.h>
  46 #include <linux/fs.h>
  47 #include <linux/buffer_head.h>
  48 #include <linux/mpage.h>
  49 #include <linux/writeback.h>
  50 #include <linux/pagemap.h>
  51
  52 #define DEBUG_SUBSYSTEM S_LLITE
  53
  54 #include "llite_internal.h"
  55
  56 /**
  57  * Implements Linux VM address_space::invalidatepage() method. This method is
  58  * called when the page is truncate from a file, either as a result of
  59  * explicit truncate, or when inode is removed from memory (as a result of
  60  * final iput(), umount, or memory pressure induced icache shrinking).
  61  *
  62  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  63  * aligned truncate). Lustre leaves partially truncated page in the cache,
  64  * relying on struct inode::i_size to limit further accesses.
  65  */
  66 static void ll_invalidatepage(struct page *vmpage, unsigned int offset,
  67                               unsigned int length)
  68 {
  69         struct inode     *inode;
  70         struct lu_env    *env;
  71         struct cl_page   *page;
  72         struct cl_object *obj;
  73
  74         LASSERT(PageLocked(vmpage));
  75         LASSERT(!PageWriteback(vmpage));
  76
  77         /*
  78          * It is safe to not check anything in invalidatepage/releasepage
  79          * below because they are run with page locked and all our io is
  80          * happening with locked page too
  81          */
  82         if (offset == 0 && length == PAGE_SIZE) {
  83                 /* See the comment in ll_releasepage() */
  84                 env = cl_env_percpu_get();
  85                 LASSERT(!IS_ERR(env));
  86                 inode = vmpage->mapping->host;
  87                 obj = ll_i2info(inode)->lli_clob;
  88                 if (obj) {
  89                         page = cl_vmpage_page(vmpage, obj);
  90                         if (page) {
  91                                 cl_page_delete(env, page);
  92                                 cl_page_put(env, page);
  93                         }
  94                 } else {
  95                         LASSERT(vmpage->private == 0);
  96                 }
  97                 cl_env_percpu_put(env);
  98         }
  99 }
 100
 101 static int ll_releasepage(struct page *vmpage, gfp_t gfp_mask)
 102 {
 103         struct lu_env     *env;
 104         struct cl_object  *obj;
 105         struct cl_page    *page;
 106         struct address_space *mapping;
 107         int result = 0;
 108
 109         LASSERT(PageLocked(vmpage));
 110         if (PageWriteback(vmpage) || PageDirty(vmpage))
 111                 return 0;
 112
 113         mapping = vmpage->mapping;
 114         if (!mapping)
 115                 return 1;
 116
 117         obj = ll_i2info(mapping->host)->lli_clob;
 118         if (!obj)
 119                 return 1;
 120
 121         /* 1 for caller, 1 for cl_page and 1 for page cache */
 122         if (page_count(vmpage) > 3)
 123                 return 0;
 124
 125         page = cl_vmpage_page(vmpage, obj);
 126         if (!page)
 127                 return 1;
 128
 129         env = cl_env_percpu_get();
 130         LASSERT(!IS_ERR(env));
 131
 132         if (!cl_page_in_use(page)) {
 133                 result = 1;
 134                 cl_page_delete(env, page);
 135         }
 136
 137         /* To use percpu env array, the call path can not be rescheduled;
 138          * otherwise percpu array will be messed if ll_releaspage() called
 139          * again on the same CPU.
 140          *
 141          * If this page holds the last refc of cl_object, the following
 142          * call path may cause reschedule:
 143          *   cl_page_put -> cl_page_free -> cl_object_put ->
 144          *     lu_object_put -> lu_object_free -> lov_delete_raid0.
 145          *
 146          * However, the kernel can't get rid of this inode until all pages have
 147          * been cleaned up. Now that we hold page lock here, it's pretty safe
 148          * that we won't get into object delete path.
 149          */
 150         LASSERT(cl_object_refc(obj) > 1);
 151         cl_page_put(env, page);
 152
 153         cl_env_percpu_put(env);
 154         return result;
 155 }
 156
 157 #define MAX_DIRECTIO_SIZE (2 * 1024 * 1024 * 1024UL)
 158
 159 /*  ll_free_user_pages - tear down page struct array
 160  *  @pages: array of page struct pointers underlying target buffer
 161  */
 162 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 163 {
 164         int i;
 165
 166         for (i = 0; i < npages; i++) {
 167                 if (do_dirty)
 168                         set_page_dirty_lock(pages[i]);
 169                 put_page(pages[i]);
 170         }
 171         kvfree(pages);
 172 }
 173
 174 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 175                            int rw, struct inode *inode,
 176                            struct ll_dio_pages *pv)
 177 {
 178         struct cl_page    *clp;
 179         struct cl_2queue  *queue;
 180         struct cl_object  *obj = io->ci_obj;
 181         int i;
 182         ssize_t rc = 0;
 183         loff_t file_offset  = pv->ldp_start_offset;
 184         size_t size = pv->ldp_size;
 185         int page_count      = pv->ldp_nr;
 186         struct page **pages = pv->ldp_pages;
 187         size_t page_size = cl_page_size(obj);
 188         bool do_io;
 189         int  io_pages       = 0;
 190
 191         queue = &io->ci_queue;
 192         cl_2queue_init(queue);
 193         for (i = 0; i < page_count; i++) {
 194                 if (pv->ldp_offsets)
 195                         file_offset = pv->ldp_offsets[i];
 196
 197                 LASSERT(!(file_offset & (page_size - 1)));
 198                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 199                                    pv->ldp_pages[i], CPT_TRANSIENT);
 200                 if (IS_ERR(clp)) {
 201                         rc = PTR_ERR(clp);
 202                         break;
 203                 }
 204
 205                 rc = cl_page_own(env, io, clp);
 206                 if (rc) {
 207                         LASSERT(clp->cp_state == CPS_FREEING);
 208                         cl_page_put(env, clp);
 209                         break;
 210                 }
 211
 212                 do_io = true;
 213
 214                 /* check the page type: if the page is a host page, then do
 215                  * write directly
 216                  */
 217                 if (clp->cp_type == CPT_CACHEABLE) {
 218                         struct page *vmpage = cl_page_vmpage(clp);
 219                         struct page *src_page;
 220                         struct page *dst_page;
 221                         void       *src;
 222                         void       *dst;
 223
 224                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 225                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 226
 227                         src = kmap_atomic(src_page);
 228                         dst = kmap_atomic(dst_page);
 229                         memcpy(dst, src, min(page_size, size));
 230                         kunmap_atomic(dst);
 231                         kunmap_atomic(src);
 232
 233                         /* make sure page will be added to the transfer by
 234                          * cl_io_submit()->...->vvp_page_prep_write().
 235                          */
 236                         if (rw == WRITE)
 237                                 set_page_dirty(vmpage);
 238
 239                         if (rw == READ) {
 240                                 /* do not issue the page for read, since it
 241                                  * may reread a ra page which has NOT uptodate
 242                                  * bit set.
 243                                  */
 244                                 cl_page_disown(env, io, clp);
 245                                 do_io = false;
 246                         }
 247                 }
 248
 249                 if (likely(do_io)) {
 250                         /*
 251                          * Add a page to the incoming page list of 2-queue.
 252                          */
 253                         cl_page_list_add(&queue->c2_qin, clp);
 254
 255                         /*
 256                          * Set page clip to tell transfer formation engine
 257                          * that page has to be sent even if it is beyond KMS.
 258                          */
 259                         cl_page_clip(env, clp, 0, min(size, page_size));
 260
 261                         ++io_pages;
 262                 }
 263
 264                 /* drop the reference count for cl_page_find */
 265                 cl_page_put(env, clp);
 266                 size -= page_size;
 267                 file_offset += page_size;
 268         }
 269
 270         if (rc == 0 && io_pages) {
 271                 rc = cl_io_submit_sync(env, io,
 272                                        rw == READ ? CRT_READ : CRT_WRITE,
 273                                        queue, 0);
 274         }
 275         if (rc == 0)
 276                 rc = pv->ldp_size;
 277
 278         cl_2queue_discard(env, io, queue);
 279         cl_2queue_disown(env, io, queue);
 280         cl_2queue_fini(env, queue);
 281         return rc;
 282 }
 283 EXPORT_SYMBOL(ll_direct_rw_pages);
 284
 285 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 286                                    int rw, struct inode *inode,
 287                                    struct address_space *mapping,
 288                                    size_t size, loff_t file_offset,
 289                                    struct page **pages, int page_count)
 290 {
 291         struct ll_dio_pages pvec = {
 292                 .ldp_pages      = pages,
 293                 .ldp_nr         = page_count,
 294                 .ldp_size       = size,
 295                 .ldp_offsets    = NULL,
 296                 .ldp_start_offset = file_offset
 297         };
 298
 299         return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 300 }
 301
 302 /* This is the maximum size of a single O_DIRECT request, based on the
 303  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 304  * representing PAGE_SIZE worth of user data, into a single buffer, and
 305  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 306  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc.
 307  */
 308 #define MAX_DIO_SIZE ((KMALLOC_MAX_SIZE / sizeof(struct brw_page) *       \
 309                        PAGE_SIZE) & ~(DT_MAX_BRW_SIZE - 1))
 310 static ssize_t ll_direct_IO_26(struct kiocb *iocb, struct iov_iter *iter)
 311 {
 312         struct ll_cl_context *lcc;
 313         const struct lu_env *env;
 314         struct cl_io *io;
 315         struct file *file = iocb->ki_filp;
 316         struct inode *inode = file->f_mapping->host;
 317         loff_t file_offset = iocb->ki_pos;
 318         ssize_t count = iov_iter_count(iter);
 319         ssize_t tot_bytes = 0, result = 0;
 320         long size = MAX_DIO_SIZE;
 321
 322         /* Check EOF by ourselves */
 323         if (iov_iter_rw(iter) == READ && file_offset >= i_size_read(inode))
 324                 return 0;
 325
 326         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 327         if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
 328                 return -EINVAL;
 329
 330         CDEBUG(D_VFSTRACE, "VFS Op:inode=" DFID "(%p), size=%zd (max %lu), offset=%lld=%llx, pages %zd (max %lu)\n",
 331                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 332                file_offset, file_offset, count >> PAGE_SHIFT,
 333                MAX_DIO_SIZE >> PAGE_SHIFT);
 334
 335         /* Check that all user buffers are aligned as well */
 336         if (iov_iter_alignment(iter) & ~PAGE_MASK)
 337                 return -EINVAL;
 338
 339         lcc = ll_cl_find(file);
 340         if (!lcc)
 341                 return -EIO;
 342
 343         env = lcc->lcc_env;
 344         LASSERT(!IS_ERR(env));
 345         io = lcc->lcc_io;
 346         LASSERT(io);
 347
 348         while (iov_iter_count(iter)) {
 349                 struct page **pages;
 350                 size_t offs;
 351
 352                 count = min_t(size_t, iov_iter_count(iter), size);
 353                 if (iov_iter_rw(iter) == READ) {
 354                         if (file_offset >= i_size_read(inode))
 355                                 break;
 356                         if (file_offset + count > i_size_read(inode))
 357                                 count = i_size_read(inode) - file_offset;
 358                 }
 359
 360                 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
 361                 if (likely(result > 0)) {
 362                         int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
 363
 364                         result = ll_direct_IO_26_seg(env, io, iov_iter_rw(iter),
 365                                                      inode, file->f_mapping,
 366                                                      result, file_offset, pages,
 367                                                      n);
 368                         ll_free_user_pages(pages, n, iov_iter_rw(iter) == READ);
 369                 }
 370                 if (unlikely(result <= 0)) {
 371                         /* If we can't allocate a large enough buffer
 372                          * for the request, shrink it to a smaller
 373                          * PAGE_SIZE multiple and try again.
 374                          * We should always be able to kmalloc for a
 375                          * page worth of page pointers = 4MB on i386.
 376                          */
 377                         if (result == -ENOMEM &&
 378                             size > (PAGE_SIZE / sizeof(*pages)) *
 379                             PAGE_SIZE) {
 380                                 size = ((((size / 2) - 1) |
 381                                          ~PAGE_MASK) + 1) &
 382                                         PAGE_MASK;
 383                                 CDEBUG(D_VFSTRACE, "DIO size now %lu\n",
 384                                        size);
 385                                 continue;
 386                         }
 387
 388                         goto out;
 389                 }
 390                 iov_iter_advance(iter, result);
 391                 tot_bytes += result;
 392                 file_offset += result;
 393         }
 394 out:
 395         if (tot_bytes > 0) {
 396                 struct vvp_io *vio = vvp_env_io(env);
 397
 398                 /* no commit async for direct IO */
 399                 vio->u.write.vui_written += tot_bytes;
 400         }
 401
 402         return tot_bytes ? tot_bytes : result;
 403 }
 404
 405 /**
 406  * Prepare partially written-to page for a write.
 407  */
 408 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 409                                    struct cl_page *pg)
 410 {
 411         struct cl_attr *attr   = vvp_env_thread_attr(env);
 412         struct cl_object *obj  = io->ci_obj;
 413         struct vvp_page *vpg   = cl_object_page_slice(obj, pg);
 414         loff_t          offset = cl_offset(obj, vvp_index(vpg));
 415         int             result;
 416
 417         cl_object_attr_lock(obj);
 418         result = cl_object_attr_get(env, obj, attr);
 419         cl_object_attr_unlock(obj);
 420         if (result == 0) {
 421                 /*
 422                  * If are writing to a new page, no need to read old data.
 423                  * The extent locking will have updated the KMS, and for our
 424                  * purposes here we can treat it like i_size.
 425                  */
 426                 if (attr->cat_kms <= offset) {
 427                         char *kaddr = kmap_atomic(vpg->vpg_page);
 428
 429                         memset(kaddr, 0, cl_page_size(obj));
 430                         kunmap_atomic(kaddr);
 431                 } else if (vpg->vpg_defer_uptodate) {
 432                         vpg->vpg_ra_used = 1;
 433                 } else {
 434                         result = ll_page_sync_io(env, io, pg, CRT_READ);
 435                 }
 436         }
 437         return result;
 438 }
 439
 440 static int ll_write_begin(struct file *file, struct address_space *mapping,
 441                           loff_t pos, unsigned int len, unsigned int flags,
 442                           struct page **pagep, void **fsdata)
 443 {
 444         struct ll_cl_context *lcc;
 445         const struct lu_env *env = NULL;
 446         struct cl_io   *io;
 447         struct cl_page *page = NULL;
 448         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 449         pgoff_t index = pos >> PAGE_SHIFT;
 450         struct page *vmpage = NULL;
 451         unsigned int from = pos & (PAGE_SIZE - 1);
 452         unsigned int to = from + len;
 453         int result = 0;
 454
 455         CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
 456
 457         lcc = ll_cl_find(file);
 458         if (!lcc) {
 459                 io = NULL;
 460                 result = -EIO;
 461                 goto out;
 462         }
 463
 464         env = lcc->lcc_env;
 465         io  = lcc->lcc_io;
 466
 467         /* To avoid deadlock, try to lock page first. */
 468         vmpage = grab_cache_page_nowait(mapping, index);
 469         if (unlikely(!vmpage || PageDirty(vmpage) || PageWriteback(vmpage))) {
 470                 struct vvp_io *vio = vvp_env_io(env);
 471                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 472
 473                 /* if the page is already in dirty cache, we have to commit
 474                  * the pages right now; otherwise, it may cause deadlock
 475                  * because it holds page lock of a dirty page and request for
 476                  * more grants. It's okay for the dirty page to be the first
 477                  * one in commit page list, though.
 478                  */
 479                 if (vmpage && plist->pl_nr > 0) {
 480                         unlock_page(vmpage);
 481                         put_page(vmpage);
 482                         vmpage = NULL;
 483                 }
 484
 485                 /* commit pages and then wait for page lock */
 486                 result = vvp_io_write_commit(env, io);
 487                 if (result < 0)
 488                         goto out;
 489
 490                 if (!vmpage) {
 491                         vmpage = grab_cache_page_write_begin(mapping, index,
 492                                                              flags);
 493                         if (!vmpage) {
 494                                 result = -ENOMEM;
 495                                 goto out;
 496                         }
 497                 }
 498         }
 499
 500         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 501         if (IS_ERR(page)) {
 502                 result = PTR_ERR(page);
 503                 goto out;
 504         }
 505
 506         lcc->lcc_page = page;
 507         lu_ref_add(&page->cp_reference, "cl_io", io);
 508
 509         cl_page_assume(env, io, page);
 510         if (!PageUptodate(vmpage)) {
 511                 /*
 512                  * We're completely overwriting an existing page,
 513                  * so _don't_ set it up to date until commit_write
 514                  */
 515                 if (from == 0 && to == PAGE_SIZE) {
 516                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 517                         POISON_PAGE(vmpage, 0x11);
 518                 } else {
 519                         /* TODO: can be optimized at OSC layer to check if it
 520                          * is a lockless IO. In that case, it's not necessary
 521                          * to read the data.
 522                          */
 523                         result = ll_prepare_partial_page(env, io, page);
 524                         if (result == 0)
 525                                 SetPageUptodate(vmpage);
 526                 }
 527         }
 528         if (result < 0)
 529                 cl_page_unassume(env, io, page);
 530 out:
 531         if (result < 0) {
 532                 if (vmpage) {
 533                         unlock_page(vmpage);
 534                         put_page(vmpage);
 535                 }
 536                 if (!IS_ERR_OR_NULL(page)) {
 537                         lu_ref_del(&page->cp_reference, "cl_io", io);
 538                         cl_page_put(env, page);
 539                 }
 540                 if (io)
 541                         io->ci_result = result;
 542         } else {
 543                 *pagep = vmpage;
 544                 *fsdata = lcc;
 545         }
 546         return result;
 547 }
 548
 549 static int ll_write_end(struct file *file, struct address_space *mapping,
 550                         loff_t pos, unsigned int len, unsigned int copied,
 551                         struct page *vmpage, void *fsdata)
 552 {
 553         struct ll_cl_context *lcc = fsdata;
 554         const struct lu_env *env;
 555         struct cl_io *io;
 556         struct vvp_io *vio;
 557         struct cl_page *page;
 558         unsigned int from = pos & (PAGE_SIZE - 1);
 559         bool unplug = false;
 560         int result = 0;
 561
 562         put_page(vmpage);
 563
 564         env  = lcc->lcc_env;
 565         page = lcc->lcc_page;
 566         io   = lcc->lcc_io;
 567         vio  = vvp_env_io(env);
 568
 569         LASSERT(cl_page_is_owned(page, io));
 570         if (copied > 0) {
 571                 struct cl_page_list *plist = &vio->u.write.vui_queue;
 572
 573                 lcc->lcc_page = NULL; /* page will be queued */
 574
 575                 /* Add it into write queue */
 576                 cl_page_list_add(plist, page);
 577                 if (plist->pl_nr == 1) /* first page */
 578                         vio->u.write.vui_from = from;
 579                 else
 580                         LASSERT(from == 0);
 581                 vio->u.write.vui_to = from + copied;
 582
 583                 /*
 584                  * To address the deadlock in balance_dirty_pages() where
 585                  * this dirty page may be written back in the same thread.
 586                  */
 587                 if (PageDirty(vmpage))
 588                         unplug = true;
 589
 590                 /* We may have one full RPC, commit it soon */
 591                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
 592                         unplug = true;
 593
 594                 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
 595                               "queued page: %d.\n", plist->pl_nr);
 596         } else {
 597                 cl_page_disown(env, io, page);
 598
 599                 lcc->lcc_page = NULL;
 600                 lu_ref_del(&page->cp_reference, "cl_io", io);
 601                 cl_page_put(env, page);
 602
 603                 /* page list is not contiguous now, commit it now */
 604                 unplug = true;
 605         }
 606
 607         if (unplug ||
 608             file->f_flags & O_SYNC || IS_SYNC(file_inode(file)))
 609                 result = vvp_io_write_commit(env, io);
 610
 611         if (result < 0)
 612                 io->ci_result = result;
 613         return result >= 0 ? copied : result;
 614 }
 615
 616 #ifdef CONFIG_MIGRATION
 617 static int ll_migratepage(struct address_space *mapping,
 618                           struct page *newpage, struct page *page,
 619                           enum migrate_mode mode
 620                 )
 621 {
 622         /* Always fail page migration until we have a proper implementation */
 623         return -EIO;
 624 }
 625 #endif
 626
 627 const struct address_space_operations ll_aops = {
 628         .readpage       = ll_readpage,
 629         .direct_IO      = ll_direct_IO_26,
 630         .writepage      = ll_writepage,
 631         .writepages     = ll_writepages,
 632         .set_page_dirty = __set_page_dirty_nobuffers,
 633         .write_begin    = ll_write_begin,
 634         .write_end      = ll_write_end,
 635         .invalidatepage = ll_invalidatepage,
 636         .releasepage    = (void *)ll_releasepage,
 637 #ifdef CONFIG_MIGRATION
 638         .migratepage    = ll_migratepage,
 639 #endif
 640 };