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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lustre/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #include <linux/kernel.h>
  42 #include <linux/mm.h>
  43 #include <linux/string.h>
  44 #include <linux/stat.h>
  45 #include <linux/errno.h>
  46 #include <linux/unistd.h>
  47 #include <asm/uaccess.h>
  48
  49 #ifdef HAVE_MIGRATE_H
  50 #include <linux/migrate.h>
  51 #elif defined(HAVE_MIGRATE_MODE_H)
  52 #include <linux/migrate_mode.h>
  53 #endif
  54 #include <linux/fs.h>
  55 #include <linux/buffer_head.h>
  56 #include <linux/writeback.h>
  57 #include <linux/stat.h>
  58 #include <asm/uaccess.h>
  59 #include <linux/mm.h>
  60 #include <linux/pagemap.h>
  61
  62 #define DEBUG_SUBSYSTEM S_LLITE
  63
  64 #include <lustre_lite.h>
  65 #include "llite_internal.h"
  66 #include <linux/lustre_compat25.h>
  67
  68 /**
  69  * Implements Linux VM address_space::invalidatepage() method. This method is
  70  * called when the page is truncate from a file, either as a result of
  71  * explicit truncate, or when inode is removed from memory (as a result of
  72  * final iput(), umount, or memory pressure induced icache shrinking).
  73  *
  74  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  75  * aligned truncate). Lustre leaves partially truncated page in the cache,
  76  * relying on struct inode::i_size to limit further accesses.
  77  */
  78 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
  79 {
  80         struct inode     *inode;
  81         struct lu_env    *env;
  82         struct cl_page   *page;
  83         struct cl_object *obj;
  84
  85         int refcheck;
  86
  87         LASSERT(PageLocked(vmpage));
  88         LASSERT(!PageWriteback(vmpage));
  89
  90         /*
  91          * It is safe to not check anything in invalidatepage/releasepage
  92          * below because they are run with page locked and all our io is
  93          * happening with locked page too
  94          */
  95         if (offset == 0) {
  96                 env = cl_env_get(&refcheck);
  97                 if (!IS_ERR(env)) {
  98                         inode = vmpage->mapping->host;
  99                         obj = ll_i2info(inode)->lli_clob;
 100                         if (obj != NULL) {
 101                                 page = cl_vmpage_page(vmpage, obj);
 102                                 if (page != NULL) {
 103                                         lu_ref_add(&page->cp_reference,
 104                                                    "delete", vmpage);
 105                                         cl_page_delete(env, page);
 106                                         lu_ref_del(&page->cp_reference,
 107                                                    "delete", vmpage);
 108                                         cl_page_put(env, page);
 109                                 }
 110                         } else
 111                                 LASSERT(vmpage->private == 0);
 112                         cl_env_put(env, &refcheck);
 113                 }
 114         }
 115 }
 116
 117 #ifdef HAVE_RELEASEPAGE_WITH_INT
 118 #define RELEASEPAGE_ARG_TYPE int
 119 #else
 120 #define RELEASEPAGE_ARG_TYPE gfp_t
 121 #endif
 122 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 123 {
 124         struct cl_env_nest nest;
 125         struct lu_env     *env;
 126         struct cl_object  *obj;
 127         struct cl_page    *page;
 128         struct address_space *mapping;
 129         int result;
 130
 131         LASSERT(PageLocked(vmpage));
 132         if (PageWriteback(vmpage) || PageDirty(vmpage))
 133                 return 0;
 134
 135         mapping = vmpage->mapping;
 136         if (mapping == NULL)
 137                 return 1;
 138
 139         obj = ll_i2info(mapping->host)->lli_clob;
 140         if (obj == NULL)
 141                 return 1;
 142
 143         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 144         if (page_count(vmpage) > 3)
 145                 return 0;
 146
 147         /* TODO: determine what gfp should be used by @gfp_mask. */
 148         env = cl_env_nested_get(&nest);
 149         if (IS_ERR(env))
 150                 /* If we can't allocate an env we won't call cl_page_put()
 151                  * later on which further means it's impossible to drop
 152                  * page refcount by cl_page, so ask kernel to not free
 153                  * this page. */
 154                 return 0;
 155
 156         page = cl_vmpage_page(vmpage, obj);
 157         result = page == NULL;
 158         if (page != NULL) {
 159                 if (cfs_atomic_read(&page->cp_ref) == 1) {
 160                         result = 1;
 161                         cl_page_delete(env, page);
 162                 }
 163                 cl_page_put(env, page);
 164         }
 165         cl_env_nested_put(&nest, env);
 166         return result;
 167 }
 168
 169 static int ll_set_page_dirty(struct page *vmpage)
 170 {
 171 #if 0
 172         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 173         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 174         struct vvp_page   *cpg;
 175
 176         /*
 177          * XXX should page method be called here?
 178          */
 179         LASSERT(&obj->co_cl == page->cp_obj);
 180         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 181         /*
 182          * XXX cannot do much here, because page is possibly not locked:
 183          * sys_munmap()->...
 184          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 185          */
 186         vvp_write_pending(obj, cpg);
 187 #endif
 188         RETURN(__set_page_dirty_nobuffers(vmpage));
 189 }
 190
 191 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 192
 193 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 194                                     size_t size, struct page ***pages,
 195                                     int *max_pages)
 196 {
 197         int result = -ENOMEM;
 198
 199         /* set an arbitrary limit to prevent arithmetic overflow */
 200         if (size > MAX_DIRECTIO_SIZE) {
 201                 *pages = NULL;
 202                 return -EFBIG;
 203         }
 204
 205         *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 206         *max_pages -= user_addr >> CFS_PAGE_SHIFT;
 207
 208         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 209         if (*pages) {
 210                 down_read(&current->mm->mmap_sem);
 211                 result = get_user_pages(current, current->mm, user_addr,
 212                                         *max_pages, (rw == READ), 0, *pages,
 213                                         NULL);
 214                 up_read(&current->mm->mmap_sem);
 215                 if (unlikely(result <= 0))
 216                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 217         }
 218
 219         return result;
 220 }
 221
 222 /*  ll_free_user_pages - tear down page struct array
 223  *  @pages: array of page struct pointers underlying target buffer */
 224 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 225 {
 226         int i;
 227
 228         for (i = 0; i < npages; i++) {
 229                 if (pages[i] == NULL)
 230                         break;
 231                 if (do_dirty)
 232                         set_page_dirty_lock(pages[i]);
 233                 page_cache_release(pages[i]);
 234         }
 235
 236         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 237 }
 238
 239 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 240                            int rw, struct inode *inode,
 241                            struct ll_dio_pages *pv)
 242 {
 243         struct cl_page    *clp;
 244         struct cl_2queue  *queue;
 245         struct cl_object  *obj = io->ci_obj;
 246         int i;
 247         ssize_t rc = 0;
 248         loff_t file_offset  = pv->ldp_start_offset;
 249         long size           = pv->ldp_size;
 250         int page_count      = pv->ldp_nr;
 251         struct page **pages = pv->ldp_pages;
 252         long page_size      = cl_page_size(obj);
 253         bool do_io;
 254         int  io_pages       = 0;
 255         ENTRY;
 256
 257         queue = &io->ci_queue;
 258         cl_2queue_init(queue);
 259         for (i = 0; i < page_count; i++) {
 260                 if (pv->ldp_offsets)
 261                     file_offset = pv->ldp_offsets[i];
 262
 263                 LASSERT(!(file_offset & (page_size - 1)));
 264                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 265                                    pv->ldp_pages[i], CPT_TRANSIENT);
 266                 if (IS_ERR(clp)) {
 267                         rc = PTR_ERR(clp);
 268                         break;
 269                 }
 270
 271                 rc = cl_page_own(env, io, clp);
 272                 if (rc) {
 273                         LASSERT(clp->cp_state == CPS_FREEING);
 274                         cl_page_put(env, clp);
 275                         break;
 276                 }
 277
 278                 do_io = true;
 279
 280                 /* check the page type: if the page is a host page, then do
 281                  * write directly */
 282                 if (clp->cp_type == CPT_CACHEABLE) {
 283                         cfs_page_t *vmpage = cl_page_vmpage(env, clp);
 284                         cfs_page_t *src_page;
 285                         cfs_page_t *dst_page;
 286                         void       *src;
 287                         void       *dst;
 288
 289                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 290                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 291
 292                         src = ll_kmap_atomic(src_page, KM_USER0);
 293                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 294                         memcpy(dst, src, min(page_size, size));
 295                         ll_kunmap_atomic(dst, KM_USER1);
 296                         ll_kunmap_atomic(src, KM_USER0);
 297
 298                         /* make sure page will be added to the transfer by
 299                          * cl_io_submit()->...->vvp_page_prep_write(). */
 300                         if (rw == WRITE)
 301                                 set_page_dirty(vmpage);
 302
 303                         if (rw == READ) {
 304                                 /* do not issue the page for read, since it
 305                                  * may reread a ra page which has NOT uptodate
 306                                  * bit set. */
 307                                 cl_page_disown(env, io, clp);
 308                                 do_io = false;
 309                         }
 310                 }
 311
 312                 if (likely(do_io)) {
 313                         cl_2queue_add(queue, clp);
 314
 315                         /*
 316                          * Set page clip to tell transfer formation engine
 317                          * that page has to be sent even if it is beyond KMS.
 318                          */
 319                         cl_page_clip(env, clp, 0, min(size, page_size));
 320
 321                         ++io_pages;
 322                 }
 323
 324                 /* drop the reference count for cl_page_find */
 325                 cl_page_put(env, clp);
 326                 size -= page_size;
 327                 file_offset += page_size;
 328         }
 329
 330         if (rc == 0 && io_pages) {
 331                 rc = cl_io_submit_sync(env, io,
 332                                        rw == READ ? CRT_READ : CRT_WRITE,
 333                                        queue, 0);
 334         }
 335         if (rc == 0)
 336                 rc = pv->ldp_size;
 337
 338         cl_2queue_discard(env, io, queue);
 339         cl_2queue_disown(env, io, queue);
 340         cl_2queue_fini(env, queue);
 341         RETURN(rc);
 342 }
 343 EXPORT_SYMBOL(ll_direct_rw_pages);
 344
 345 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 346                                    int rw, struct inode *inode,
 347                                    struct address_space *mapping,
 348                                    size_t size, loff_t file_offset,
 349                                    struct page **pages, int page_count)
 350 {
 351     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 352                                  .ldp_nr           = page_count,
 353                                  .ldp_size         = size,
 354                                  .ldp_offsets      = NULL,
 355                                  .ldp_start_offset = file_offset
 356                                };
 357
 358     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 359 }
 360
 361 #ifdef KMALLOC_MAX_SIZE
 362 #define MAX_MALLOC KMALLOC_MAX_SIZE
 363 #else
 364 #define MAX_MALLOC (128 * 1024)
 365 #endif
 366
 367 /* This is the maximum size of a single O_DIRECT request, based on the
 368  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 369  * representing PAGE_SIZE worth of user data, into a single buffer, and
 370  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 371  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 372 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 373                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 374 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 375                                const struct iovec *iov, loff_t file_offset,
 376                                unsigned long nr_segs)
 377 {
 378         struct lu_env *env;
 379         struct cl_io *io;
 380         struct file *file = iocb->ki_filp;
 381         struct inode *inode = file->f_mapping->host;
 382         struct ccc_object *obj = cl_inode2ccc(inode);
 383         long count = iov_length(iov, nr_segs);
 384         long tot_bytes = 0, result = 0;
 385         struct ll_inode_info *lli = ll_i2info(inode);
 386         unsigned long seg = 0;
 387         long size = MAX_DIO_SIZE;
 388         int refcheck;
 389         ENTRY;
 390
 391         if (!lli->lli_has_smd)
 392                 RETURN(-EBADF);
 393
 394         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 395         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 396                 RETURN(-EINVAL);
 397
 398         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
 399                "offset=%lld=%llx, pages %lu (max %lu)\n",
 400                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 401                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 402                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 403
 404         /* Check that all user buffers are aligned as well */
 405         for (seg = 0; seg < nr_segs; seg++) {
 406                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 407                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 408                         RETURN(-EINVAL);
 409         }
 410
 411         env = cl_env_get(&refcheck);
 412         LASSERT(!IS_ERR(env));
 413         io = ccc_env_io(env)->cui_cl.cis_io;
 414         LASSERT(io != NULL);
 415
 416         /* 0. Need locking between buffered and direct access. and race with
 417          *    size changing by concurrent truncates and writes.
 418          * 1. Need inode mutex to operate transient pages.
 419          */
 420         if (rw == READ)
 421                 mutex_lock(&inode->i_mutex);
 422
 423         LASSERT(obj->cob_transient_pages == 0);
 424         for (seg = 0; seg < nr_segs; seg++) {
 425                 long iov_left = iov[seg].iov_len;
 426                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 427
 428                 if (rw == READ) {
 429                         if (file_offset >= i_size_read(inode))
 430                                 break;
 431                         if (file_offset + iov_left > i_size_read(inode))
 432                                 iov_left = i_size_read(inode) - file_offset;
 433                 }
 434
 435                 while (iov_left > 0) {
 436                         struct page **pages;
 437                         int page_count, max_pages = 0;
 438                         long bytes;
 439
 440                         bytes = min(size, iov_left);
 441                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 442                                                        &pages, &max_pages);
 443                         if (likely(page_count > 0)) {
 444                                 if (unlikely(page_count <  max_pages))
 445                                         bytes = page_count << CFS_PAGE_SHIFT;
 446                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 447                                                              file->f_mapping,
 448                                                              bytes, file_offset,
 449                                                              pages, page_count);
 450                                 ll_free_user_pages(pages, max_pages, rw==READ);
 451                         } else if (page_count == 0) {
 452                                 GOTO(out, result = -EFAULT);
 453                         } else {
 454                                 result = page_count;
 455                         }
 456                         if (unlikely(result <= 0)) {
 457                                 /* If we can't allocate a large enough buffer
 458                                  * for the request, shrink it to a smaller
 459                                  * PAGE_SIZE multiple and try again.
 460                                  * We should always be able to kmalloc for a
 461                                  * page worth of page pointers = 4MB on i386. */
 462                                 if (result == -ENOMEM &&
 463                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 464                                            CFS_PAGE_SIZE) {
 465                                         size = ((((size / 2) - 1) |
 466                                                  ~CFS_PAGE_MASK) + 1) &
 467                                                 CFS_PAGE_MASK;
 468                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 469                                                size);
 470                                         continue;
 471                                 }
 472
 473                                 GOTO(out, result);
 474                         }
 475
 476                         tot_bytes += result;
 477                         file_offset += result;
 478                         iov_left -= result;
 479                         user_addr += result;
 480                 }
 481         }
 482 out:
 483         LASSERT(obj->cob_transient_pages == 0);
 484         if (rw == READ)
 485                 mutex_unlock(&inode->i_mutex);
 486
 487         if (tot_bytes > 0) {
 488                 if (rw == WRITE) {
 489                         struct lov_stripe_md *lsm;
 490
 491                         lsm = ccc_inode_lsm_get(inode);
 492                         LASSERT(lsm != NULL);
 493                         lov_stripe_lock(lsm);
 494                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 495                         lov_stripe_unlock(lsm);
 496                         ccc_inode_lsm_put(inode, lsm);
 497                 }
 498         }
 499
 500         cl_env_put(env, &refcheck);
 501         RETURN(tot_bytes ? : result);
 502 }
 503
 504 #if defined(HAVE_KERNEL_WRITE_BEGIN_END) || defined(MS_HAS_NEW_AOPS)
 505 static int ll_write_begin(struct file *file, struct address_space *mapping,
 506                          loff_t pos, unsigned len, unsigned flags,
 507                          struct page **pagep, void **fsdata)
 508 {
 509         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 510         struct page *page;
 511         int rc;
 512         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 513         ENTRY;
 514
 515         page = grab_cache_page_write_begin(mapping, index, flags);
 516         if (!page)
 517                 RETURN(-ENOMEM);
 518
 519         *pagep = page;
 520
 521         rc = ll_prepare_write(file, page, from, from + len);
 522         if (rc) {
 523                 unlock_page(page);
 524                 page_cache_release(page);
 525         }
 526         RETURN(rc);
 527 }
 528
 529 static int ll_write_end(struct file *file, struct address_space *mapping,
 530                         loff_t pos, unsigned len, unsigned copied,
 531                         struct page *page, void *fsdata)
 532 {
 533         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 534         int rc;
 535
 536         rc = ll_commit_write(file, page, from, from + copied);
 537         unlock_page(page);
 538         page_cache_release(page);
 539
 540         return rc ?: copied;
 541 }
 542 #endif
 543
 544 #ifdef CONFIG_MIGRATION
 545 int ll_migratepage(struct address_space *mapping,
 546                 struct page *newpage, struct page *page
 547 #ifdef HAVE_MIGRATEPAGE_4ARGS
 548                 , enum migrate_mode mode
 549 #endif
 550                 )
 551 {
 552         /* Always fail page migration until we have a proper implementation */
 553         return -EIO;
 554 }
 555 #endif
 556
 557 #ifndef MS_HAS_NEW_AOPS
 558 struct address_space_operations ll_aops = {
 559         .readpage       = ll_readpage,
 560 //        .readpages      = ll_readpages,
 561         .direct_IO      = ll_direct_IO_26,
 562         .writepage      = ll_writepage,
 563         .writepages     = ll_writepages,
 564         .set_page_dirty = ll_set_page_dirty,
 565 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
 566         .write_begin    = ll_write_begin,
 567         .write_end      = ll_write_end,
 568 #else
 569         .prepare_write  = ll_prepare_write,
 570         .commit_write   = ll_commit_write,
 571 #endif
 572         .invalidatepage = ll_invalidatepage,
 573         .releasepage    = (void *)ll_releasepage,
 574 #ifdef CONFIG_MIGRATION
 575         .migratepage    = ll_migratepage,
 576 #endif
 577         .bmap           = NULL
 578 };
 579 #else
 580 struct address_space_operations_ext ll_aops = {
 581         .orig_aops.readpage       = ll_readpage,
 582 //        .orig_aops.readpages      = ll_readpages,
 583         .orig_aops.direct_IO      = ll_direct_IO_26,
 584         .orig_aops.writepage      = ll_writepage,
 585         .orig_aops.writepages     = ll_writepages,
 586         .orig_aops.set_page_dirty = ll_set_page_dirty,
 587         .orig_aops.prepare_write  = ll_prepare_write,
 588         .orig_aops.commit_write   = ll_commit_write,
 589         .orig_aops.invalidatepage = ll_invalidatepage,
 590         .orig_aops.releasepage    = ll_releasepage,
 591 #ifdef CONFIG_MIGRATION
 592         .orig_aops.migratepage    = ll_migratepage,
 593 #endif
 594         .orig_aops.bmap           = NULL,
 595         .write_begin    = ll_write_begin,
 596         .write_end      = ll_write_end
 597 };
 598 #endif