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