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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/lustre/llite/rw26.c
  32  *
  33  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  34  */
  35
  36 #include <linux/buffer_head.h>
  37 #include <linux/errno.h>
  38 #include <linux/fs.h>
  39 #include <linux/kernel.h>
  40 #include <linux/mm.h>
  41 #include <linux/mpage.h>
  42 #include <linux/pagemap.h>
  43 #include <linux/string.h>
  44 #include <linux/unistd.h>
  45 #include <linux/writeback.h>
  46 #include <linux/migrate.h>
  47
  48 #define DEBUG_SUBSYSTEM S_LLITE
  49
  50 #include "llite_internal.h"
  51 #include <lustre_compat.h>
  52
  53 #ifdef HAVE_INVALIDATE_FOLIO
  54 /**
  55  * Implements Linux VM address_space::invalidate_folio() method. This method is
  56  * called when the folio is truncated from a file, either as a result of
  57  * explicit truncate, or when inode is removed from memory (as a result of
  58  * final iput(), umount, or memory pressure induced icache shrinking).
  59  *
  60  * [0, off] bytes of the folio remain valid (this is for a case of non-page
  61  * aligned truncate). Lustre leaves partially truncated folios in the cache,
  62  * relying on struct inode::i_size to limit further accesses.
  63  */
  64 static void ll_invalidate_folio(struct folio *folio, size_t offset, size_t len)
  65 {
  66         struct inode *inode;
  67         struct lu_env *env;
  68         struct cl_page *page;
  69         struct cl_object *obj;
  70
  71         LASSERT(!folio_test_writeback(folio));
  72         LASSERT(folio_test_locked(folio));
  73
  74         if (!(offset == 0 && len == folio_size(folio)) &&
  75             !folio_test_large(folio))
  76                 return;
  77
  78         /* Drop the pages from the folio */
  79         env = cl_env_percpu_get();
  80         LASSERT(!IS_ERR(env));
  81
  82         inode = folio_inode(folio);
  83         obj = ll_i2info(inode)->lli_clob;
  84         if (obj != NULL) {
  85                 int n, npgs = folio_nr_pages(folio);
  86
  87                 for (n = 0; n < npgs; n++) {
  88                         struct page *vmpage = folio_page(folio, n);
  89
  90                         LASSERT(PageLocked(vmpage));
  91                         LASSERT(!PageWriteback(vmpage));
  92
  93                         page = cl_vmpage_page(vmpage, obj);
  94                         if (page != NULL) {
  95                                 cl_page_delete(env, page);
  96                                 cl_page_put(env, page);
  97                         }
  98                 }
  99         } else {
 100                 LASSERT(!folio_get_private(folio));
 101         }
 102         cl_env_percpu_put(env);
 103 }
 104 #else
 105
 106 /**
 107  * Implements Linux VM address_space::invalidatepage() method. This method is
 108  * called when the page is truncate from a file, either as a result of
 109  * explicit truncate, or when inode is removed from memory (as a result of
 110  * final iput(), umount, or memory pressure induced icache shrinking).
 111  *
 112  * [0, offset] bytes of the page remain valid (this is for a case of not-page
 113  * aligned truncate). Lustre leaves partially truncated page in the cache,
 114  * relying on struct inode::i_size to limit further accesses.
 115  */
 116 static void ll_invalidatepage(struct page *vmpage,
 117 #ifdef HAVE_INVALIDATE_RANGE
 118                                 unsigned int offset, unsigned int length
 119 #else
 120                                 unsigned long offset
 121 #endif
 122                              )
 123 {
 124         struct inode     *inode;
 125         struct lu_env    *env;
 126         struct cl_page   *page;
 127         struct cl_object *obj;
 128
 129         LASSERT(PageLocked(vmpage));
 130         LASSERT(!PageWriteback(vmpage));
 131
 132         /*
 133          * It is safe to not check anything in invalidatepage/releasepage
 134          * below because they are run with page locked and all our io is
 135          * happening with locked page too
 136          */
 137 #ifdef HAVE_INVALIDATE_RANGE
 138         if (offset == 0 && length == PAGE_SIZE) {
 139 #else
 140         if (offset == 0) {
 141 #endif
 142                 /* See the comment in ll_releasepage() */
 143                 env = cl_env_percpu_get();
 144                 LASSERT(!IS_ERR(env));
 145
 146                 inode = vmpage->mapping->host;
 147                 obj = ll_i2info(inode)->lli_clob;
 148                 if (obj != NULL) {
 149                         page = cl_vmpage_page(vmpage, obj);
 150                         if (page != NULL) {
 151                                 cl_page_delete(env, page);
 152                                 cl_page_put(env, page);
 153                         }
 154                 } else
 155                         LASSERT(vmpage->private == 0);
 156
 157                 cl_env_percpu_put(env);
 158         }
 159
 160         if (CFS_FAIL_PRECHECK(OBD_FAIL_LLITE_PAGE_INVALIDATE_PAUSE)) {
 161                 unlock_page(vmpage);
 162                 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_PAGE_INVALIDATE_PAUSE,
 163                                  cfs_fail_val);
 164                 lock_page(vmpage);
 165         }
 166 }
 167 #endif
 168
 169 static bool do_release_page(struct page *vmpage, gfp_t wait)
 170 {
 171         struct address_space *mapping;
 172         struct cl_object *obj;
 173         struct cl_page *page;
 174         struct lu_env *env;
 175         int result = 0;
 176
 177         ENTRY;
 178
 179         LASSERT(PageLocked(vmpage));
 180         if (PageWriteback(vmpage) || PageDirty(vmpage))
 181                 RETURN(0);
 182
 183         mapping = vmpage->mapping;
 184         if (mapping == NULL)
 185                 RETURN(1);
 186
 187         obj = ll_i2info(mapping->host)->lli_clob;
 188         if (obj == NULL)
 189                 RETURN(1);
 190
 191         page = cl_vmpage_page(vmpage, obj);
 192         if (page == NULL)
 193                 RETURN(1);
 194
 195         env = cl_env_percpu_get();
 196         LASSERT(!IS_ERR(env));
 197
 198         if (!cl_page_in_use(page)) {
 199                 result = 1;
 200                 cl_page_delete(env, page);
 201         }
 202
 203         /* To use percpu env array, the call path can not be rescheduled;
 204          * otherwise percpu array will be messed if ll_releaspage() called
 205          * again on the same CPU.
 206          *
 207          * If this page holds the last refc of cl_object, the following
 208          * call path may cause reschedule:
 209          *   cl_page_put -> cl_page_free -> cl_object_put ->
 210          *     lu_object_put -> lu_object_free -> lov_delete_raid0.
 211          *
 212          * However, the kernel can't get rid of this inode until all pages have
 213          * been cleaned up. Now that we hold page lock here, it's pretty safe
 214          * that we won't get into object delete path.
 215          */
 216         LASSERT(cl_object_refc(obj) > 1);
 217         cl_page_put(env, page);
 218
 219         cl_env_percpu_put(env);
 220         RETURN(result);
 221 }
 222
 223 #ifdef HAVE_AOPS_RELEASE_FOLIO
 224 static bool ll_release_folio(struct folio *folio, gfp_t wait)
 225 {
 226         struct page *vmpage = folio_page(folio, 0);
 227
 228         /* folio_nr_pages(folio) == 1 is fixed with grab_cache_page* */
 229         BUG_ON(folio_nr_pages(folio) != 1);
 230
 231         return do_release_page(vmpage, wait);
 232 }
 233 #else /* !HAVE_AOPS_RELEASE_FOLIO */
 234 #ifdef HAVE_RELEASEPAGE_WITH_INT
 235 #define RELEASEPAGE_ARG_TYPE int
 236 #else
 237 #define RELEASEPAGE_ARG_TYPE gfp_t
 238 #endif
 239 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 240 {
 241         return do_release_page(vmpage, gfp_mask);
 242 }
 243 #endif /* HAVE_AOPS_RELEASE_FOLIO */
 244
 245 static ssize_t ll_get_user_pages(int rw, struct iov_iter *iter,
 246                                 struct ll_dio_pages *pvec,
 247                                 size_t maxsize)
 248 {
 249 #if defined(HAVE_DIO_ITER)
 250         size_t start;
 251         size_t result;
 252
 253         result = iov_iter_get_pages_alloc2(iter, &pvec->ldp_pages, maxsize,
 254                                           &start);
 255         if (result > 0)
 256                 pvec->ldp_count = DIV_ROUND_UP(result + start, PAGE_SIZE);
 257
 258         return result;
 259 #else
 260         unsigned long addr;
 261         size_t page_count;
 262         size_t size;
 263         long result;
 264
 265         if (!maxsize)
 266                 return 0;
 267
 268         if (!iter->nr_segs)
 269                 return 0;
 270
 271         addr = (unsigned long)iter->iov->iov_base + iter->iov_offset;
 272         if (addr & ~PAGE_MASK)
 273                 return -EINVAL;
 274
 275         size = min_t(size_t, maxsize, iter->iov->iov_len);
 276         page_count = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 277         OBD_ALLOC_PTR_ARRAY_LARGE(pvec->ldp_pages, page_count);
 278         if (pvec->ldp_pages == NULL)
 279                 return -ENOMEM;
 280
 281         mmap_read_lock(current->mm);
 282         result = get_user_pages(current, current->mm, addr, page_count,
 283                                 rw == READ, 0, pvec->ldp_pages, NULL);
 284         mmap_read_unlock(current->mm);
 285
 286         if (unlikely(result != page_count)) {
 287                 ll_release_user_pages(pvec->ldp_pages, page_count);
 288                 pvec->ldp_pages = NULL;
 289
 290                 if (result >= 0)
 291                         return -EFAULT;
 292
 293                 return result;
 294         }
 295         pvec->ldp_count = page_count;
 296
 297         return size;
 298 #endif
 299 }
 300
 301 /* iov_iter_alignment() is introduced in 3.16 similar to HAVE_DIO_ITER */
 302 #if defined(HAVE_DIO_ITER)
 303 static unsigned long iov_iter_alignment_vfs(const struct iov_iter *i)
 304 {
 305         return iov_iter_alignment(i);
 306 }
 307 #else /* copied from alignment_iovec() */
 308 static unsigned long iov_iter_alignment_vfs(const struct iov_iter *i)
 309 {
 310         const struct iovec *iov = i->iov;
 311         unsigned long res;
 312         size_t size = i->count;
 313         size_t n;
 314
 315         if (!size)
 316                 return 0;
 317
 318         res = (unsigned long)iov->iov_base + i->iov_offset;
 319         n = iov->iov_len - i->iov_offset;
 320         if (n >= size)
 321                 return res | size;
 322
 323         size -= n;
 324         res |= n;
 325         while (size > (++iov)->iov_len) {
 326                 res |= (unsigned long)iov->iov_base | iov->iov_len;
 327                 size -= iov->iov_len;
 328         }
 329         res |= (unsigned long)iov->iov_base | size;
 330
 331         return res;
 332 }
 333 #endif
 334
 335 /*
 336  * Lustre could relax a bit for alignment, io count is not
 337  * necessary page alignment.
 338  */
 339 static unsigned long ll_iov_iter_alignment(struct iov_iter *i)
 340 {
 341         size_t orig_size = i->count;
 342         size_t count = orig_size & ~PAGE_MASK;
 343         unsigned long res;
 344
 345         if (!count)
 346                 return iov_iter_alignment_vfs(i);
 347
 348         if (orig_size > PAGE_SIZE) {
 349                 iov_iter_truncate(i, orig_size - count);
 350                 res = iov_iter_alignment_vfs(i);
 351                 iov_iter_reexpand(i, orig_size);
 352
 353                 return res;
 354         }
 355
 356         res = iov_iter_alignment_vfs(i);
 357         /* start address is page aligned */
 358         if ((res & ~PAGE_MASK) == orig_size)
 359                 return PAGE_SIZE;
 360
 361         return res;
 362 }
 363
 364 static int
 365 ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io, size_t size,
 366                    int rw, struct inode *inode, struct cl_sub_dio *sdio)
 367 {
 368         struct ll_dio_pages *pv = &sdio->csd_dio_pages;
 369         struct cl_page    *page;
 370         struct cl_2queue  *queue = &io->ci_queue;
 371         struct cl_object  *obj = io->ci_obj;
 372         struct cl_sync_io *anchor = &sdio->csd_sync;
 373         loff_t offset   = pv->ldp_file_offset;
 374         int io_pages    = 0;
 375         ssize_t rc = 0;
 376         int i = 0;
 377
 378         ENTRY;
 379
 380         cl_2queue_init(queue);
 381         while (size > 0) {
 382                 size_t from = offset & ~PAGE_MASK;
 383                 size_t to = min(from + size, PAGE_SIZE);
 384
 385                 page = cl_page_find(env, obj, offset >> PAGE_SHIFT,
 386                                     pv->ldp_pages[i], CPT_TRANSIENT);
 387                 if (IS_ERR(page)) {
 388                         rc = PTR_ERR(page);
 389                         break;
 390                 }
 391
 392                 LASSERT(page->cp_type == CPT_TRANSIENT);
 393                 rc = cl_page_own(env, io, page);
 394                 if (rc) {
 395                         cl_page_put(env, page);
 396                         break;
 397                 }
 398
 399                 page->cp_sync_io = anchor;
 400                 if (inode && IS_ENCRYPTED(inode)) {
 401                         /* In case of Direct IO on encrypted file, we need to
 402                          * add a reference to the inode on the cl_page.
 403                          * This info is required by llcrypt to proceed
 404                          * to encryption/decryption.
 405                          * This is safe because we know these pages are private
 406                          * to the thread doing the Direct IO.
 407                          */
 408                         page->cp_inode = inode;
 409                 }
 410                 /* We keep the refcount from cl_page_find, so we don't need
 411                  * another one here
 412                  */
 413                 cl_page_list_add(&queue->c2_qin, page, false);
 414                 /*
 415                  * Call page clip for incomplete pages, to set range of bytes
 416                  * in the page and to tell transfer formation engine to send
 417                  * the page even if it is beyond KMS (ie, don't trim IO to KMS)
 418                  */
 419                 if (from != 0 || to != PAGE_SIZE)
 420                         cl_page_clip(env, page, from, to);
 421                 ++io_pages;
 422                 i++;
 423
 424                 offset += to - from;
 425                 size -= to - from;
 426         }
 427         /* on success, we should hit every page in the pvec and have no bytes
 428          * left in 'size'
 429          */
 430         LASSERT(ergo(rc == 0, i == pv->ldp_count));
 431         LASSERT(ergo(rc == 0, size == 0));
 432
 433         if (rc == 0 && io_pages > 0) {
 434                 int iot = rw == READ ? CRT_READ : CRT_WRITE;
 435
 436                 atomic_add(io_pages, &anchor->csi_sync_nr);
 437                 /*
 438                  * Avoid out-of-order execution of adding inflight
 439                  * modifications count and io submit.
 440                  */
 441                 smp_mb();
 442                 rc = cl_io_submit_rw(env, io, iot, queue);
 443                 if (rc == 0) {
 444                         cl_page_list_splice(&queue->c2_qout, &sdio->csd_pages);
 445                 } else {
 446                         atomic_add(-queue->c2_qin.pl_nr,
 447                                    &anchor->csi_sync_nr);
 448                         cl_page_list_for_each(page, &queue->c2_qin)
 449                                 page->cp_sync_io = NULL;
 450                 }
 451                 /* handle partially submitted reqs */
 452                 if (queue->c2_qin.pl_nr > 0) {
 453                         CERROR(DFID " failed to submit %d dio pages: %zd\n",
 454                                PFID(lu_object_fid(&obj->co_lu)),
 455                                queue->c2_qin.pl_nr, rc);
 456                         if (rc == 0)
 457                                 rc = -EIO;
 458                 }
 459         }
 460
 461         cl_2queue_discard(env, io, queue);
 462         cl_2queue_disown(env, queue);
 463         cl_2queue_fini(env, queue);
 464         RETURN(rc);
 465 }
 466
 467 #ifdef KMALLOC_MAX_SIZE
 468 #define MAX_MALLOC KMALLOC_MAX_SIZE
 469 #else
 470 #define MAX_MALLOC (128 * 1024)
 471 #endif
 472
 473 /* This is the maximum size of a single O_DIRECT request, based on the
 474  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 475  * representing PAGE_SIZE worth of user data, into a single buffer, and
 476  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 477  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 478 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
 479                       ~((size_t)DT_MAX_BRW_SIZE - 1))
 480
 481 static ssize_t
 482 ll_direct_IO_impl(struct kiocb *iocb, struct iov_iter *iter, int rw)
 483 {
 484         struct ll_cl_context *lcc;
 485         const struct lu_env *env;
 486         struct cl_io *io;
 487         struct file *file = iocb->ki_filp;
 488         struct inode *inode = file->f_mapping->host;
 489         struct cl_dio_aio *ll_dio_aio;
 490         struct cl_sub_dio *sdio;
 491         size_t count = iov_iter_count(iter);
 492         ssize_t tot_bytes = 0, result = 0;
 493         loff_t file_offset = iocb->ki_pos;
 494         bool sync_submit = false;
 495         bool unaligned = false;
 496         struct vvp_io *vio;
 497         ssize_t rc2;
 498
 499         ENTRY;
 500
 501         if (file_offset & ~PAGE_MASK)
 502                 unaligned = true;
 503
 504         if (count & ~PAGE_MASK)
 505                 unaligned = true;
 506
 507         /* Check that all user buffers are aligned as well */
 508         if (ll_iov_iter_alignment(iter) & ~PAGE_MASK)
 509                 unaligned = true;
 510
 511         CDEBUG(D_VFSTRACE,
 512                "VFS Op:inode="DFID"(%p), size=%zd (max %lu), offset=%lld=%llx, pages %zd (max %lu)%s\n",
 513                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 514                file_offset, file_offset,
 515                (count >> PAGE_SHIFT) + !!(count & ~PAGE_MASK),
 516                MAX_DIO_SIZE >> PAGE_SHIFT, unaligned ? ", unaligned" : "");
 517
 518         /* Check EOF by ourselves */
 519         if (rw == READ && file_offset >= i_size_read(inode))
 520                 RETURN(0);
 521
 522         /* unaligned DIO support can be turned off, so is it on? */
 523         if (unaligned && !ll_sbi_has_unaligned_dio(ll_i2sbi(inode)))
 524                 RETURN(-EINVAL);
 525
 526         /* the requirement to not return EIOCBQUEUED for pipes (see bottom of
 527          * this function) plays havoc with the unaligned I/O lifecycle, so
 528          * don't allow unaligned I/O on pipes
 529          */
 530         if (unaligned && iov_iter_is_pipe(iter))
 531                 RETURN(0);
 532
 533         lcc = ll_cl_find(inode);
 534         if (lcc == NULL)
 535                 RETURN(-EIO);
 536
 537         env = lcc->lcc_env;
 538         LASSERT(!IS_ERR(env));
 539         vio = vvp_env_io(env);
 540         io = lcc->lcc_io;
 541         LASSERT(io != NULL);
 542
 543         /* this means we encountered an old server which can't safely support
 544          * unaligned DIO, so we have to disable it
 545          */
 546         if (unaligned && !cl_io_top(io)->ci_allow_unaligned_dio)
 547                 RETURN(-EINVAL);
 548
 549         /* if one part of an I/O is unaligned, just handle all of it that way -
 550          * otherwise we create significant complexities with managing the iovec
 551          * in different ways, etc, all for very marginal benefits
 552          */
 553         if (unaligned)
 554                 io->ci_unaligned_dio = true;
 555         if (io->ci_unaligned_dio)
 556                 unaligned = true;
 557
 558         ll_dio_aio = io->ci_dio_aio;
 559         LASSERT(ll_dio_aio);
 560         LASSERT(ll_dio_aio->cda_iocb == iocb);
 561
 562         /* We cannot do parallel submission of sub-I/Os - for AIO or regular
 563          * DIO - unless lockless because it causes us to release the lock
 564          * early.
 565          *
 566          * There are also several circumstances in which we must disable
 567          * parallel DIO, so we check if it is enabled.
 568          *
 569          * The check for "is_sync_kiocb" excludes AIO, which does not need to
 570          * be disabled in these situations.
 571          */
 572         if (io->ci_dio_lock || (is_sync_kiocb(iocb) && !io->ci_parallel_dio))
 573                 sync_submit = true;
 574
 575         while (iov_iter_count(iter)) {
 576                 struct ll_dio_pages *pvec;
 577
 578                 count = min_t(size_t, iov_iter_count(iter), MAX_DIO_SIZE);
 579                 if (rw == READ) {
 580                         if (file_offset >= i_size_read(inode))
 581                                 break;
 582
 583                         if (file_offset + count > i_size_read(inode))
 584                                 count = i_size_read(inode) - file_offset;
 585                 }
 586
 587                 /* if we are doing sync_submit, then we free this below,
 588                  * otherwise it is freed on the final call to cl_sync_io_note
 589                  * (either in this function or from a ptlrpcd daemon)
 590                  */
 591                 sdio = cl_sub_dio_alloc(ll_dio_aio, iter, rw == WRITE,
 592                                         unaligned, sync_submit);
 593                 if (!sdio)
 594                         GOTO(out, result = -ENOMEM);
 595
 596                 pvec = &sdio->csd_dio_pages;
 597                 pvec->ldp_file_offset = file_offset;
 598
 599                 if (!unaligned) {
 600                         result = ll_get_user_pages(rw, iter, pvec, count);
 601                         /* ll_get_user_pages returns bytes in the IO or error*/
 602                         count = result;
 603                 } else {
 604                         /* same calculation used in ll_get_user_pages */
 605                         count = min_t(size_t, count, iter_iov(iter)->iov_len);
 606                         result = ll_allocate_dio_buffer(pvec, count);
 607                         /* allocate_dio_buffer returns number of pages or
 608                          * error, so do not set count = result
 609                          */
 610                 }
 611
 612                 /* now we have the actual count, so store it in the sdio */
 613                 sdio->csd_bytes = count;
 614
 615                 if (unlikely(result <= 0)) {
 616                         cl_sync_io_note(env, &sdio->csd_sync, result);
 617                         if (sync_submit) {
 618                                 LASSERT(sdio->csd_creator_free);
 619                                 cl_sub_dio_free(sdio);
 620                         }
 621                         GOTO(out, result);
 622                 }
 623
 624                 if (unaligned && rw == WRITE) {
 625                         result = ll_dio_user_copy(sdio, iter);
 626                         if (unlikely(result <= 0)) {
 627                                 cl_sync_io_note(env, &sdio->csd_sync, result);
 628                                 if (sync_submit) {
 629                                         LASSERT(sdio->csd_creator_free);
 630                                         cl_sub_dio_free(sdio);
 631                                 }
 632                                 GOTO(out, result);
 633                         }
 634                 }
 635
 636                 result = ll_direct_rw_pages(env, io, count, rw, inode, sdio);
 637                 /* if the i/o was unsuccessful, we zero the number of bytes to
 638                  * copy back.  Note that partial I/O completion isn't possible
 639                  * here - I/O either completes or fails.  So there's no need to
 640                  * handle short I/O here by changing 'count' with the result
 641                  * from ll_direct_rw_pages.
 642                  *
 643                  * This must be done before we release the reference
 644                  * immediately below, because releasing the reference allows
 645                  * i/o completion (and copyback to userspace, if unaligned) to
 646                  * start.
 647                  */
 648                 if (result != 0)
 649                         sdio->csd_bytes = 0;
 650                 /* We've submitted pages and can now remove the extra
 651                  * reference for that
 652                  */
 653                 cl_sync_io_note(env, &sdio->csd_sync, result);
 654
 655                 if (sync_submit) {
 656                         rc2 = cl_sync_io_wait(env, &sdio->csd_sync,
 657                                              0);
 658                         if (result == 0 && rc2)
 659                                 result = rc2;
 660                         LASSERT(sdio->csd_creator_free);
 661                         cl_sub_dio_free(sdio);
 662                 }
 663                 if (unlikely(result < 0))
 664                         GOTO(out, result);
 665
 666                 iov_iter_advance(iter, count);
 667
 668                 tot_bytes += count;
 669                 file_offset += count;
 670                 CDEBUG(D_VFSTRACE,
 671                        "result %zd tot_bytes %zd count %zd file_offset %lld\n",
 672                        result, tot_bytes, count, file_offset);
 673         }
 674
 675 out:
 676         if (rw == WRITE)
 677                 vio->u.readwrite.vui_written += tot_bytes;
 678         else
 679                 vio->u.readwrite.vui_read += tot_bytes;
 680
 681         /* AIO is not supported on pipes, so we cannot return EIOCBQEUED like
 682          * we normally would for both DIO and AIO here
 683          */
 684         if (result == 0 && !iov_iter_is_pipe(iter))
 685                 result = -EIOCBQUEUED;
 686
 687         RETURN(result);
 688 }
 689
 690 #ifdef HAVE_DIO_ITER
 691 static ssize_t ll_direct_IO(
 692 #ifndef HAVE_IOV_ITER_RW
 693              int rw,
 694 #endif
 695              struct kiocb *iocb, struct iov_iter *iter
 696 #ifndef HAVE_DIRECTIO_2ARGS
 697              , loff_t file_offset
 698 #endif
 699              )
 700 {
 701         int nrw;
 702
 703 #ifndef HAVE_IOV_ITER_RW
 704         nrw = rw;
 705 #else
 706         nrw = iov_iter_rw(iter);
 707 #endif
 708
 709         return ll_direct_IO_impl(iocb, iter, nrw);
 710 }
 711
 712 #else /* !defined(HAVE_DIO_ITER) */
 713
 714 static ssize_t
 715 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
 716              loff_t file_offset, unsigned long nr_segs)
 717 {
 718         struct iov_iter iter;
 719
 720         iov_iter_init(&iter, iov, nr_segs, iov_length(iov, nr_segs), 0);
 721         return ll_direct_IO_impl(iocb, &iter, rw);
 722 }
 723
 724 #endif /* !defined(HAVE_DIO_ITER) */
 725
 726 /**
 727  * Prepare partially written-to page for a write.
 728  * @pg is owned when passed in and disowned when it returns non-zero result to
 729  * the caller.
 730  */
 731 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
 732                                    struct cl_page *pg, struct file *file)
 733 {
 734         struct cl_attr *attr   = vvp_env_thread_attr(env);
 735         struct cl_object *obj  = io->ci_obj;
 736         loff_t offset = cl_page_index(pg) << PAGE_SHIFT;
 737         int result;
 738         ENTRY;
 739
 740         cl_object_attr_lock(obj);
 741         result = cl_object_attr_get(env, obj, attr);
 742         cl_object_attr_unlock(obj);
 743         if (result) {
 744                 cl_page_disown(env, io, pg);
 745                 GOTO(out, result);
 746         }
 747
 748         /*
 749          * If are writing to a new page, no need to read old data.
 750          * The extent locking will have updated the KMS, and for our
 751          * purposes here we can treat it like i_size.
 752          */
 753         if (attr->cat_kms <= offset) {
 754                 char *kaddr = kmap_atomic(pg->cp_vmpage);
 755
 756                 memset(kaddr, 0, PAGE_SIZE);
 757                 kunmap_atomic(kaddr);
 758                 GOTO(out, result = 0);
 759         }
 760
 761         if (pg->cp_defer_uptodate) {
 762                 pg->cp_ra_used = 1;
 763                 GOTO(out, result = 0);
 764         }
 765
 766         result = ll_io_read_page(env, io, pg, file);
 767         if (result)
 768                 GOTO(out, result);
 769
 770         /* ll_io_read_page() disowns the page */
 771         result = cl_page_own(env, io, pg);
 772         if (!result) {
 773                 if (!PageUptodate(cl_page_vmpage(pg))) {
 774                         cl_page_disown(env, io, pg);
 775                         result = -EIO;
 776                 }
 777         } else if (result == -ENOENT) {
 778                 /* page was truncated */
 779                 result = -EAGAIN;
 780         }
 781         EXIT;
 782
 783 out:
 784         return result;
 785 }
 786
 787 static int ll_tiny_write_begin(struct page *vmpage, struct address_space *mapping)
 788 {
 789         /* Page must be present, up to date, dirty, and not in writeback. */
 790         if (!vmpage || !PageUptodate(vmpage) || !PageDirty(vmpage) ||
 791             PageWriteback(vmpage) || vmpage->mapping != mapping)
 792                 return -ENODATA;
 793
 794         return 0;
 795 }
 796
 797 static int ll_write_begin(struct file *file, struct address_space *mapping,
 798                           loff_t pos, unsigned int len,
 799 #ifdef HAVE_GRAB_CACHE_PAGE_WRITE_BEGIN_WITH_FLAGS
 800                           unsigned int flags,
 801 #endif
 802                           struct page **pagep, void **fsdata)
 803 {
 804         struct ll_cl_context *lcc = NULL;
 805         const struct lu_env  *env = NULL;
 806         struct vvp_io *vio;
 807         struct cl_io   *io = NULL;
 808         struct cl_page *page = NULL;
 809         struct inode *inode = file_inode(file);
 810         struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
 811         pgoff_t index = pos >> PAGE_SHIFT;
 812         struct page *vmpage = NULL;
 813         unsigned from = pos & (PAGE_SIZE - 1);
 814         unsigned to = from + len;
 815         int result = 0;
 816         int iocb_flags;
 817         ENTRY;
 818
 819         CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
 820
 821         lcc = ll_cl_find(inode);
 822         if (lcc == NULL) {
 823                 vmpage = grab_cache_page_nowait(mapping, index);
 824                 result = ll_tiny_write_begin(vmpage, mapping);
 825                 GOTO(out, result);
 826         }
 827
 828         env = lcc->lcc_env;
 829         io  = lcc->lcc_io;
 830         vio = vvp_env_io(env);
 831
 832         iocb_flags = iocb_ki_flags_get(file, vio->vui_iocb);
 833         if (iocb_ki_flags_check(iocb_flags, DIRECT)) {
 834                 /* direct IO failed because it couldn't clean up cached pages,
 835                  * this causes a problem for mirror write because the cached
 836                  * page may belong to another mirror, which will result in
 837                  * problem submitting the I/O. */
 838                 if (io->ci_designated_mirror > 0)
 839                         GOTO(out, result = -EBUSY);
 840
 841                 /**
 842                  * Direct write can fall back to buffered read, but DIO is done
 843                  * with lockless i/o, and buffered requires LDLM locking, so
 844                  * in this case we must restart without lockless.
 845                  */
 846                 if (!io->ci_dio_lock) {
 847                         io->ci_dio_lock = 1;
 848                         io->ci_need_restart = 1;
 849                         GOTO(out, result = -ENOLCK);
 850                 }
 851         }
 852 again:
 853         /* To avoid deadlock, try to lock page first. */
 854         vmpage = grab_cache_page_nowait(mapping, index);
 855
 856         if (unlikely(vmpage == NULL ||
 857                      PageDirty(vmpage) || PageWriteback(vmpage))) {
 858                 struct vvp_io *vio = vvp_env_io(env);
 859                 struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
 860
 861                 /* if the page is already in dirty cache, we have to commit
 862                  * the pages right now; otherwise, it may cause deadlock
 863                  * because it holds page lock of a dirty page and request for
 864                  * more grants. It's okay for the dirty page to be the first
 865                  * one in commit page list, though. */
 866                 if (vmpage != NULL && plist->pl_nr > 0) {
 867                         unlock_page(vmpage);
 868                         put_page(vmpage);
 869                         vmpage = NULL;
 870                 }
 871
 872                 /* commit pages and then wait for page lock */
 873                 result = vvp_io_write_commit(env, io);
 874                 if (result < 0)
 875                         GOTO(out, result);
 876
 877                 if (vmpage == NULL) {
 878                         vmpage = grab_cache_page_write_begin(mapping, index
 879 #ifdef HAVE_GRAB_CACHE_PAGE_WRITE_BEGIN_WITH_FLAGS
 880                                                              , flags
 881 #endif
 882                                                              );
 883                         if (vmpage == NULL)
 884                                 GOTO(out, result = -ENOMEM);
 885                 }
 886         }
 887
 888         /* page was truncated */
 889         if (mapping != vmpage->mapping) {
 890                 CDEBUG(D_VFSTRACE, "page: %lu was truncated\n", index);
 891                 unlock_page(vmpage);
 892                 put_page(vmpage);
 893                 vmpage = NULL;
 894                 goto again;
 895         }
 896
 897         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 898         if (IS_ERR(page))
 899                 GOTO(out, result = PTR_ERR(page));
 900
 901         lcc->lcc_page = page;
 902         lu_ref_add(&page->cp_reference, "cl_io", io);
 903
 904         cl_page_assume(env, io, page);
 905         if (!PageUptodate(vmpage)) {
 906                 /*
 907                  * We're completely overwriting an existing page,
 908                  * so _don't_ set it up to date until commit_write
 909                  */
 910                 if (from == 0 && to == PAGE_SIZE) {
 911                         CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
 912                         POISON_PAGE(vmpage, 0x11);
 913                 } else {
 914                         /* TODO: can be optimized at OSC layer to check if it
 915                          * is a lockless IO. In that case, it's not necessary
 916                          * to read the data. */
 917                         result = ll_prepare_partial_page(env, io, page, file);
 918                         if (result) {
 919                                 /* vmpage should have been unlocked */
 920                                 put_page(vmpage);
 921                                 vmpage = NULL;
 922
 923                                 if (result == -EAGAIN)
 924                                         goto again;
 925                                 GOTO(out, result);
 926                         }
 927                 }
 928         }
 929         EXIT;
 930 out:
 931         if (result < 0) {
 932                 if (vmpage != NULL) {
 933                         unlock_page(vmpage);
 934                         put_page(vmpage);
 935                 }
 936                 /* On tiny_write failure, page and io are always null. */
 937                 if (!IS_ERR_OR_NULL(page)) {
 938                         lu_ref_del(&page->cp_reference, "cl_io", io);
 939                         cl_page_put(env, page);
 940                 }
 941                 if (io)
 942                         io->ci_result = result;
 943         } else {
 944                 *pagep = vmpage;
 945                 *fsdata = lcc;
 946         }
 947         RETURN(result);
 948 }
 949
 950 static int ll_tiny_write_end(struct file *file, struct address_space *mapping,
 951                              loff_t pos, unsigned int len, unsigned int copied,
 952                              struct page *vmpage)
 953 {
 954         struct cl_page *clpage = (struct cl_page *) vmpage->private;
 955         loff_t kms = pos+copied;
 956         loff_t to = kms & (PAGE_SIZE-1) ? kms & (PAGE_SIZE-1) : PAGE_SIZE;
 957         __u16 refcheck;
 958         struct lu_env *env = cl_env_get(&refcheck);
 959         int rc = 0;
 960
 961         ENTRY;
 962
 963         if (IS_ERR(env)) {
 964                 rc = PTR_ERR(env);
 965                 goto out;
 966         }
 967
 968         /* This page is dirty in cache, so it should have a cl_page pointer
 969          * set in vmpage->private.
 970          */
 971         LASSERT(clpage != NULL);
 972
 973         if (copied == 0)
 974                 goto out_env;
 975
 976         /* Update the underlying size information in the OSC/LOV objects this
 977          * page is part of.
 978          */
 979         cl_page_touch(env, clpage, to);
 980
 981 out_env:
 982         cl_env_put(env, &refcheck);
 983
 984 out:
 985         /* Must return page unlocked. */
 986         unlock_page(vmpage);
 987
 988         RETURN(rc);
 989 }
 990
 991 static int ll_write_end(struct file *file, struct address_space *mapping,
 992                         loff_t pos, unsigned len, unsigned copied,
 993                         struct page *vmpage, void *fsdata)
 994 {
 995         struct ll_cl_context *lcc = fsdata;
 996         const struct lu_env *env;
 997         struct cl_io *io;
 998         struct vvp_io *vio;
 999         struct cl_page *page;
1000         unsigned from = pos & (PAGE_SIZE - 1);
1001         bool unplug = false;
1002         int result = 0;
1003         ENTRY;
1004
1005         put_page(vmpage);
1006
1007         CDEBUG(D_VFSTRACE, "pos %llu, len %u, copied %u\n", pos, len, copied);
1008
1009         if (lcc == NULL) {
1010                 result = ll_tiny_write_end(file, mapping, pos, len, copied,
1011                                            vmpage);
1012                 GOTO(out, result);
1013         }
1014
1015         LASSERT(lcc != NULL);
1016         env  = lcc->lcc_env;
1017         page = lcc->lcc_page;
1018         io   = lcc->lcc_io;
1019         vio  = vvp_env_io(env);
1020
1021         LASSERT(cl_page_is_owned(page, io));
1022         if (copied > 0) {
1023                 struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
1024
1025                 lcc->lcc_page = NULL; /* page will be queued */
1026
1027                 /* Add it into write queue */
1028                 cl_page_list_add(plist, page, true);
1029                 if (plist->pl_nr == 1) /* first page */
1030                         vio->u.readwrite.vui_from = from;
1031                 else
1032                         LASSERT(from == 0);
1033                 vio->u.readwrite.vui_to = from + copied;
1034
1035                 /* To address the deadlock in balance_dirty_pages() where
1036                  * this dirty page may be written back in the same thread. */
1037                 if (PageDirty(vmpage))
1038                         unplug = true;
1039
1040                 /* We may have one full RPC, commit it soon */
1041                 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
1042                         unplug = true;
1043
1044                 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
1045                               "queued page: %d.\n", plist->pl_nr);
1046         } else {
1047                 cl_page_disown(env, io, page);
1048
1049                 lcc->lcc_page = NULL;
1050                 lu_ref_del(&page->cp_reference, "cl_io", io);
1051                 cl_page_put(env, page);
1052
1053                 /* page list is not contiguous now, commit it now */
1054                 unplug = true;
1055         }
1056         if (unplug || io->u.ci_wr.wr_sync)
1057                 result = vvp_io_write_commit(env, io);
1058
1059         if (result < 0)
1060                 io->ci_result = result;
1061
1062
1063 out:
1064         RETURN(result >= 0 ? copied : result);
1065 }
1066
1067 #ifdef CONFIG_MIGRATION
1068 static int ll_migrate_folio(struct address_space *mapping,
1069                             struct folio_migr *newpage, struct folio_migr *page,
1070                             enum migrate_mode mode)
1071 {
1072         /* Always fail page migration until we have a proper implementation */
1073         return -EIO;
1074 }
1075 #endif
1076
1077 const struct address_space_operations ll_aops = {
1078 #ifdef HAVE_DIRTY_FOLIO
1079         .dirty_folio            = filemap_dirty_folio,
1080 #else
1081         .set_page_dirty         = __set_page_dirty_nobuffers,
1082 #endif
1083 #ifdef HAVE_INVALIDATE_FOLIO
1084         .invalidate_folio       = ll_invalidate_folio,
1085 #else
1086         .invalidatepage         = ll_invalidatepage,
1087 #endif
1088 #ifdef HAVE_AOPS_READ_FOLIO
1089         .read_folio             = ll_read_folio,
1090 #else
1091         .readpage               = ll_readpage,
1092 #endif
1093 #ifdef HAVE_AOPS_RELEASE_FOLIO
1094         .release_folio          = ll_release_folio,
1095 #else
1096         .releasepage            = (void *)ll_releasepage,
1097 #endif
1098         .direct_IO              = ll_direct_IO,
1099         .writepage              = ll_writepage,
1100         .writepages             = ll_writepages,
1101         .write_begin            = ll_write_begin,
1102         .write_end              = ll_write_end,
1103 #ifdef CONFIG_MIGRATION
1104         .migrate_folio          = ll_migrate_folio,
1105 #endif
1106 };