lustre/llite/rw26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/lustre/llite/rw26.c
  39  *
  40  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  41  */
  42
  43 #ifndef AUTOCONF_INCLUDED
  44 #include <linux/config.h>
  45 #endif
  46 #include <linux/kernel.h>
  47 #include <linux/mm.h>
  48 #include <linux/string.h>
  49 #include <linux/stat.h>
  50 #include <linux/errno.h>
  51 #include <linux/smp_lock.h>
  52 #include <linux/unistd.h>
  53 #include <linux/version.h>
  54 #include <asm/system.h>
  55 #include <asm/uaccess.h>
  56
  57 #include <linux/fs.h>
  58 #include <linux/buffer_head.h>
  59 #include <linux/writeback.h>
  60 #include <linux/stat.h>
  61 #include <asm/uaccess.h>
  62 #include <linux/mm.h>
  63 #include <linux/pagemap.h>
  64 #include <linux/smp_lock.h>
  65
  66 #define DEBUG_SUBSYSTEM S_LLITE
  67
  68 #include <lustre_lite.h>
  69 #include "llite_internal.h"
  70 #include <linux/lustre_compat25.h>
  71
  72 /**
  73  * Implements Linux VM address_space::invalidatepage() method. This method is
  74  * called when the page is truncate from a file, either as a result of
  75  * explicit truncate, or when inode is removed from memory (as a result of
  76  * final iput(), umount, or memory pressure induced icache shrinking).
  77  *
  78  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  79  * aligned truncate). Lustre leaves partially truncated page in the cache,
  80  * relying on struct inode::i_size to limit further accesses.
  81  */
  82 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
  83 {
  84         struct inode     *inode;
  85         struct lu_env    *env;
  86         struct cl_page   *page;
  87         struct cl_object *obj;
  88
  89         int refcheck;
  90
  91         LASSERT(PageLocked(vmpage));
  92         LASSERT(!PageWriteback(vmpage));
  93
  94         /*
  95          * It is safe to not check anything in invalidatepage/releasepage
  96          * below because they are run with page locked and all our io is
  97          * happening with locked page too
  98          */
  99         if (offset == 0) {
 100                 env = cl_env_get(&refcheck);
 101                 if (!IS_ERR(env)) {
 102                         inode = vmpage->mapping->host;
 103                         obj = ll_i2info(inode)->lli_clob;
 104                         if (obj != NULL) {
 105                                 page = cl_vmpage_page(vmpage, obj);
 106                                 if (page != NULL) {
 107                                         lu_ref_add(&page->cp_reference,
 108                                                    "delete", vmpage);
 109                                         cl_page_delete(env, page);
 110                                         lu_ref_del(&page->cp_reference,
 111                                                    "delete", vmpage);
 112                                         cl_page_put(env, page);
 113                                 }
 114                         } else
 115                                 LASSERT(vmpage->private == 0);
 116                         cl_env_put(env, &refcheck);
 117                 }
 118         }
 119 }
 120
 121 #ifdef HAVE_RELEASEPAGE_WITH_INT
 122 #define RELEASEPAGE_ARG_TYPE int
 123 #else
 124 #define RELEASEPAGE_ARG_TYPE gfp_t
 125 #endif
 126 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 127 {
 128         struct cl_env_nest nest;
 129         struct lu_env     *env;
 130         struct cl_object  *obj;
 131         struct cl_page    *page;
 132         struct address_space *mapping;
 133         int result;
 134
 135         LASSERT(PageLocked(vmpage));
 136         if (PageWriteback(vmpage) || PageDirty(vmpage))
 137                 return 0;
 138
 139         mapping = vmpage->mapping;
 140         if (mapping == NULL)
 141                 return 1;
 142
 143         obj = ll_i2info(mapping->host)->lli_clob;
 144         if (obj == NULL)
 145                 return 1;
 146
 147         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 148         if (page_count(vmpage) > 3)
 149                 return 0;
 150
 151         /* TODO: determine what gfp should be used by @gfp_mask. */
 152         env = cl_env_nested_get(&nest);
 153         if (IS_ERR(env))
 154                 /* If we can't allocate an env we won't call cl_page_put()
 155                  * later on which further means it's impossible to drop
 156                  * page refcount by cl_page, so ask kernel to not free
 157                  * this page. */
 158                 return 0;
 159
 160         page = cl_vmpage_page(vmpage, obj);
 161         result = page == NULL;
 162         if (page != NULL) {
 163                 if (cfs_atomic_read(&page->cp_ref) == 1) {
 164                         result = 1;
 165                         cl_page_delete(env, page);
 166                 }
 167                 cl_page_put(env, page);
 168         }
 169         cl_env_nested_put(&nest, env);
 170         return result;
 171 }
 172
 173 static int ll_set_page_dirty(struct page *vmpage)
 174 {
 175 #if 0
 176         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 177         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 178         struct vvp_page   *cpg;
 179
 180         /*
 181          * XXX should page method be called here?
 182          */
 183         LASSERT(&obj->co_cl == page->cp_obj);
 184         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 185         /*
 186          * XXX cannot do much here, because page is possibly not locked:
 187          * sys_munmap()->...
 188          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 189          */
 190         vvp_write_pending(obj, cpg);
 191 #endif
 192         RETURN(__set_page_dirty_nobuffers(vmpage));
 193 }
 194
 195 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 196
 197 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 198                                     size_t size, struct page ***pages,
 199                                     int *max_pages)
 200 {
 201         int result = -ENOMEM;
 202
 203         /* set an arbitrary limit to prevent arithmetic overflow */
 204         if (size > MAX_DIRECTIO_SIZE) {
 205                 *pages = NULL;
 206                 return -EFBIG;
 207         }
 208
 209         *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 210         *max_pages -= user_addr >> CFS_PAGE_SHIFT;
 211
 212         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 213         if (*pages) {
 214                 down_read(&current->mm->mmap_sem);
 215                 result = get_user_pages(current, current->mm, user_addr,
 216                                         *max_pages, (rw == READ), 0, *pages,
 217                                         NULL);
 218                 up_read(&current->mm->mmap_sem);
 219                 if (unlikely(result <= 0))
 220                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 221         }
 222
 223         return result;
 224 }
 225
 226 /*  ll_free_user_pages - tear down page struct array
 227  *  @pages: array of page struct pointers underlying target buffer */
 228 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 229 {
 230         int i;
 231
 232         for (i = 0; i < npages; i++) {
 233                 if (pages[i] == NULL)
 234                         break;
 235                 if (do_dirty)
 236                         set_page_dirty_lock(pages[i]);
 237                 page_cache_release(pages[i]);
 238         }
 239
 240         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 241 }
 242
 243 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 244                            int rw, struct inode *inode,
 245                            struct ll_dio_pages *pv)
 246 {
 247         struct cl_page    *clp;
 248         struct cl_2queue  *queue;
 249         struct cl_object  *obj = io->ci_obj;
 250         int i;
 251         ssize_t rc = 0;
 252         loff_t file_offset  = pv->ldp_start_offset;
 253         long size           = pv->ldp_size;
 254         int page_count      = pv->ldp_nr;
 255         struct page **pages = pv->ldp_pages;
 256         long page_size      = cl_page_size(obj);
 257         bool do_io;
 258         int  io_pages       = 0;
 259         ENTRY;
 260
 261         queue = &io->ci_queue;
 262         cl_2queue_init(queue);
 263         for (i = 0; i < page_count; i++) {
 264                 if (pv->ldp_offsets)
 265                     file_offset = pv->ldp_offsets[i];
 266
 267                 LASSERT(!(file_offset & (page_size - 1)));
 268                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 269                                    pv->ldp_pages[i], CPT_TRANSIENT);
 270                 if (IS_ERR(clp)) {
 271                         rc = PTR_ERR(clp);
 272                         break;
 273                 }
 274
 275                 rc = cl_page_own(env, io, clp);
 276                 if (rc) {
 277                         LASSERT(clp->cp_state == CPS_FREEING);
 278                         cl_page_put(env, clp);
 279                         break;
 280                 }
 281
 282                 do_io = true;
 283
 284                 /* check the page type: if the page is a host page, then do
 285                  * write directly */
 286                 if (clp->cp_type == CPT_CACHEABLE) {
 287                         cfs_page_t *vmpage = cl_page_vmpage(env, clp);
 288                         cfs_page_t *src_page;
 289                         cfs_page_t *dst_page;
 290                         void       *src;
 291                         void       *dst;
 292
 293                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 294                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 295
 296                         src = kmap_atomic(src_page, KM_USER0);
 297                         dst = kmap_atomic(dst_page, KM_USER1);
 298                         memcpy(dst, src, min(page_size, size));
 299                         kunmap_atomic(dst, KM_USER1);
 300                         kunmap_atomic(src, KM_USER0);
 301
 302                         /* make sure page will be added to the transfer by
 303                          * cl_io_submit()->...->vvp_page_prep_write(). */
 304                         if (rw == WRITE)
 305                                 set_page_dirty(vmpage);
 306
 307                         if (rw == READ) {
 308                                 /* do not issue the page for read, since it
 309                                  * may reread a ra page which has NOT uptodate
 310                                  * bit set. */
 311                                 cl_page_disown(env, io, clp);
 312                                 do_io = false;
 313                         }
 314                 }
 315
 316                 if (likely(do_io)) {
 317                         cl_2queue_add(queue, clp);
 318
 319                         /*
 320                          * Set page clip to tell transfer formation engine
 321                          * that page has to be sent even if it is beyond KMS.
 322                          */
 323                         cl_page_clip(env, clp, 0, min(size, page_size));
 324
 325                         ++io_pages;
 326                 }
 327
 328                 /* drop the reference count for cl_page_find */
 329                 cl_page_put(env, clp);
 330                 size -= page_size;
 331                 file_offset += page_size;
 332         }
 333
 334         if (rc == 0 && io_pages) {
 335                 rc = cl_io_submit_sync(env, io,
 336                                        rw == READ ? CRT_READ : CRT_WRITE,
 337                                        queue, CRP_NORMAL, 0);
 338         }
 339         if (rc == 0)
 340                 rc = pv->ldp_size;
 341
 342         cl_2queue_discard(env, io, queue);
 343         cl_2queue_disown(env, io, queue);
 344         cl_2queue_fini(env, queue);
 345         RETURN(rc);
 346 }
 347 EXPORT_SYMBOL(ll_direct_rw_pages);
 348
 349 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 350                                    int rw, struct inode *inode,
 351                                    struct address_space *mapping,
 352                                    size_t size, loff_t file_offset,
 353                                    struct page **pages, int page_count)
 354 {
 355     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 356                                  .ldp_nr           = page_count,
 357                                  .ldp_size         = size,
 358                                  .ldp_offsets      = NULL,
 359                                  .ldp_start_offset = file_offset
 360                                };
 361
 362     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 363 }
 364
 365 #ifdef KMALLOC_MAX_SIZE
 366 #define MAX_MALLOC KMALLOC_MAX_SIZE
 367 #else
 368 #define MAX_MALLOC (128 * 1024)
 369 #endif
 370
 371 /* This is the maximum size of a single O_DIRECT request, based on the
 372  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 373  * representing PAGE_SIZE worth of user data, into a single buffer, and
 374  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 375  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 376 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 377                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 378 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 379                                const struct iovec *iov, loff_t file_offset,
 380                                unsigned long nr_segs)
 381 {
 382         struct lu_env *env;
 383         struct cl_io *io;
 384         struct file *file = iocb->ki_filp;
 385         struct inode *inode = file->f_mapping->host;
 386         struct ccc_object *obj = cl_inode2ccc(inode);
 387         long count = iov_length(iov, nr_segs);
 388         long tot_bytes = 0, result = 0;
 389         struct ll_inode_info *lli = ll_i2info(inode);
 390         struct lov_stripe_md *lsm = lli->lli_smd;
 391         unsigned long seg = 0;
 392         long size = MAX_DIO_SIZE;
 393         int refcheck;
 394         ENTRY;
 395
 396         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 397                 RETURN(-EBADF);
 398
 399         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 400         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 401                 RETURN(-EINVAL);
 402
 403         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
 404                "offset=%lld=%llx, pages %lu (max %lu)\n",
 405                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 406                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 407                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 408
 409         /* Check that all user buffers are aligned as well */
 410         for (seg = 0; seg < nr_segs; seg++) {
 411                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 412                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 413                         RETURN(-EINVAL);
 414         }
 415
 416         env = cl_env_get(&refcheck);
 417         LASSERT(!IS_ERR(env));
 418         io = ccc_env_io(env)->cui_cl.cis_io;
 419         LASSERT(io != NULL);
 420
 421         /* 0. Need locking between buffered and direct access. and race with
 422          *    size changing by concurrent truncates and writes.
 423          * 1. Need inode sem to operate transient pages. */
 424         if (rw == READ)
 425                 LOCK_INODE_MUTEX(inode);
 426
 427         LASSERT(obj->cob_transient_pages == 0);
 428         for (seg = 0; seg < nr_segs; seg++) {
 429                 long iov_left = iov[seg].iov_len;
 430                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 431
 432                 if (rw == READ) {
 433                         if (file_offset >= i_size_read(inode))
 434                                 break;
 435                         if (file_offset + iov_left > i_size_read(inode))
 436                                 iov_left = i_size_read(inode) - file_offset;
 437                 }
 438
 439                 while (iov_left > 0) {
 440                         struct page **pages;
 441                         int page_count, max_pages = 0;
 442                         long bytes;
 443
 444                         bytes = min(size, iov_left);
 445                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 446                                                        &pages, &max_pages);
 447                         if (likely(page_count > 0)) {
 448                                 if (unlikely(page_count <  max_pages))
 449                                         bytes = page_count << CFS_PAGE_SHIFT;
 450                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 451                                                              file->f_mapping,
 452                                                              bytes, file_offset,
 453                                                              pages, page_count);
 454                                 ll_free_user_pages(pages, max_pages, rw==READ);
 455                         } else if (page_count == 0) {
 456                                 GOTO(out, result = -EFAULT);
 457                         } else {
 458                                 result = page_count;
 459                         }
 460                         if (unlikely(result <= 0)) {
 461                                 /* If we can't allocate a large enough buffer
 462                                  * for the request, shrink it to a smaller
 463                                  * PAGE_SIZE multiple and try again.
 464                                  * We should always be able to kmalloc for a
 465                                  * page worth of page pointers = 4MB on i386. */
 466                                 if (result == -ENOMEM &&
 467                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 468                                            CFS_PAGE_SIZE) {
 469                                         size = ((((size / 2) - 1) |
 470                                                  ~CFS_PAGE_MASK) + 1) &
 471                                                 CFS_PAGE_MASK;
 472                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 473                                                size);
 474                                         continue;
 475                                 }
 476
 477                                 GOTO(out, result);
 478                         }
 479
 480                         tot_bytes += result;
 481                         file_offset += result;
 482                         iov_left -= result;
 483                         user_addr += result;
 484                 }
 485         }
 486 out:
 487         LASSERT(obj->cob_transient_pages == 0);
 488         if (rw == READ)
 489                 UNLOCK_INODE_MUTEX(inode);
 490
 491         if (tot_bytes > 0) {
 492                 if (rw == WRITE) {
 493                         lov_stripe_lock(lsm);
 494                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 495                         lov_stripe_unlock(lsm);
 496                 }
 497         }
 498
 499         cl_env_put(env, &refcheck);
 500         RETURN(tot_bytes ? : result);
 501 }
 502
 503 #if defined(HAVE_KERNEL_WRITE_BEGIN_END) || defined(MS_HAS_NEW_AOPS)
 504 static int ll_write_begin(struct file *file, struct address_space *mapping,
 505                          loff_t pos, unsigned len, unsigned flags,
 506                          struct page **pagep, void **fsdata)
 507 {
 508         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 509         struct page *page;
 510         int rc;
 511         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 512         ENTRY;
 513
 514         page = grab_cache_page_write_begin(mapping, index, flags);
 515         if (!page)
 516                 RETURN(-ENOMEM);
 517
 518         *pagep = page;
 519
 520         rc = ll_prepare_write(file, page, from, from + len);
 521         if (rc) {
 522                 unlock_page(page);
 523                 page_cache_release(page);
 524         }
 525         RETURN(rc);
 526 }
 527
 528 static int ll_write_end(struct file *file, struct address_space *mapping,
 529                         loff_t pos, unsigned len, unsigned copied,
 530                         struct page *page, void *fsdata)
 531 {
 532         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 533         int rc;
 534
 535         rc = ll_commit_write(file, page, from, from + copied);
 536         unlock_page(page);
 537         page_cache_release(page);
 538
 539         return rc ?: copied;
 540 }
 541 #endif
 542
 543 #ifdef CONFIG_MIGRATION
 544 int ll_migratepage(struct address_space *mapping,
 545                    struct page *newpage, struct page *page)
 546 {
 547         /* Always fail page migration until we have a proper implementation */
 548         return -EIO;
 549 }
 550 #endif
 551
 552 #ifndef MS_HAS_NEW_AOPS
 553 struct address_space_operations ll_aops = {
 554         .readpage       = ll_readpage,
 555 //        .readpages      = ll_readpages,
 556         .direct_IO      = ll_direct_IO_26,
 557         .writepage      = ll_writepage,
 558         .writepages     = generic_writepages,
 559         .set_page_dirty = ll_set_page_dirty,
 560         .sync_page      = NULL,
 561 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
 562         .write_begin    = ll_write_begin,
 563         .write_end      = ll_write_end,
 564 #else
 565         .prepare_write  = ll_prepare_write,
 566         .commit_write   = ll_commit_write,
 567 #endif
 568         .invalidatepage = ll_invalidatepage,
 569         .releasepage    = (void *)ll_releasepage,
 570 #ifdef CONFIG_MIGRATION
 571         .migratepage    = ll_migratepage,
 572 #endif
 573         .bmap           = NULL
 574 };
 575 #else
 576 struct address_space_operations_ext ll_aops = {
 577         .orig_aops.readpage       = ll_readpage,
 578 //        .orig_aops.readpages      = ll_readpages,
 579         .orig_aops.direct_IO      = ll_direct_IO_26,
 580         .orig_aops.writepage      = ll_writepage,
 581         .orig_aops.writepages     = generic_writepages,
 582         .orig_aops.set_page_dirty = ll_set_page_dirty,
 583         .orig_aops.sync_page      = NULL,
 584         .orig_aops.prepare_write  = ll_prepare_write,
 585         .orig_aops.commit_write   = ll_commit_write,
 586         .orig_aops.invalidatepage = ll_invalidatepage,
 587         .orig_aops.releasepage    = ll_releasepage,
 588 #ifdef CONFIG_MIGRATION
 589         .orig_aops.migratepage    = ll_migratepage,
 590 #endif
 591         .orig_aops.bmap           = NULL,
 592         .write_begin    = ll_write_begin,
 593         .write_end      = ll_write_end
 594 };
 595 #endif