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 <linux/version.h>
  48 #include <asm/system.h>
  49 #include <asm/uaccess.h>
  50
  51 #ifdef HAVE_MIGRATE_H
  52 #include <linux/migrate.h>
  53 #elif defined(HAVE_MIGRATE_MODE_H)
  54 #include <linux/migrate_mode.h>
  55 #endif
  56 #include <linux/fs.h>
  57 #include <linux/buffer_head.h>
  58 #include <linux/writeback.h>
  59 #include <linux/stat.h>
  60 #include <asm/uaccess.h>
  61 #include <linux/mm.h>
  62 #include <linux/pagemap.h>
  63
  64 #define DEBUG_SUBSYSTEM S_LLITE
  65
  66 #include <lustre_lite.h>
  67 #include "llite_internal.h"
  68 #include <linux/lustre_compat25.h>
  69
  70 /**
  71  * Implements Linux VM address_space::invalidatepage() method. This method is
  72  * called when the page is truncate from a file, either as a result of
  73  * explicit truncate, or when inode is removed from memory (as a result of
  74  * final iput(), umount, or memory pressure induced icache shrinking).
  75  *
  76  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  77  * aligned truncate). Lustre leaves partially truncated page in the cache,
  78  * relying on struct inode::i_size to limit further accesses.
  79  */
  80 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
  81 {
  82         struct inode     *inode;
  83         struct lu_env    *env;
  84         struct cl_page   *page;
  85         struct cl_object *obj;
  86
  87         int refcheck;
  88
  89         LASSERT(PageLocked(vmpage));
  90         LASSERT(!PageWriteback(vmpage));
  91
  92         /*
  93          * It is safe to not check anything in invalidatepage/releasepage
  94          * below because they are run with page locked and all our io is
  95          * happening with locked page too
  96          */
  97         if (offset == 0) {
  98                 env = cl_env_get(&refcheck);
  99                 if (!IS_ERR(env)) {
 100                         inode = vmpage->mapping->host;
 101                         obj = ll_i2info(inode)->lli_clob;
 102                         if (obj != NULL) {
 103                                 page = cl_vmpage_page(vmpage, obj);
 104                                 if (page != NULL) {
 105                                         lu_ref_add(&page->cp_reference,
 106                                                    "delete", vmpage);
 107                                         cl_page_delete(env, page);
 108                                         lu_ref_del(&page->cp_reference,
 109                                                    "delete", vmpage);
 110                                         cl_page_put(env, page);
 111                                 }
 112                         } else
 113                                 LASSERT(vmpage->private == 0);
 114                         cl_env_put(env, &refcheck);
 115                 }
 116         }
 117 }
 118
 119 #ifdef HAVE_RELEASEPAGE_WITH_INT
 120 #define RELEASEPAGE_ARG_TYPE int
 121 #else
 122 #define RELEASEPAGE_ARG_TYPE gfp_t
 123 #endif
 124 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 125 {
 126         struct cl_env_nest nest;
 127         struct lu_env     *env;
 128         struct cl_object  *obj;
 129         struct cl_page    *page;
 130         struct address_space *mapping;
 131         int result;
 132
 133         LASSERT(PageLocked(vmpage));
 134         if (PageWriteback(vmpage) || PageDirty(vmpage))
 135                 return 0;
 136
 137         mapping = vmpage->mapping;
 138         if (mapping == NULL)
 139                 return 1;
 140
 141         obj = ll_i2info(mapping->host)->lli_clob;
 142         if (obj == NULL)
 143                 return 1;
 144
 145         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 146         if (page_count(vmpage) > 3)
 147                 return 0;
 148
 149         /* TODO: determine what gfp should be used by @gfp_mask. */
 150         env = cl_env_nested_get(&nest);
 151         if (IS_ERR(env))
 152                 /* If we can't allocate an env we won't call cl_page_put()
 153                  * later on which further means it's impossible to drop
 154                  * page refcount by cl_page, so ask kernel to not free
 155                  * this page. */
 156                 return 0;
 157
 158         page = cl_vmpage_page(vmpage, obj);
 159         result = page == NULL;
 160         if (page != NULL) {
 161                 if (cfs_atomic_read(&page->cp_ref) == 1) {
 162                         result = 1;
 163                         cl_page_delete(env, page);
 164                 }
 165                 cl_page_put(env, page);
 166         }
 167         cl_env_nested_put(&nest, env);
 168         return result;
 169 }
 170
 171 static int ll_set_page_dirty(struct page *vmpage)
 172 {
 173 #if 0
 174         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 175         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 176         struct vvp_page   *cpg;
 177
 178         /*
 179          * XXX should page method be called here?
 180          */
 181         LASSERT(&obj->co_cl == page->cp_obj);
 182         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 183         /*
 184          * XXX cannot do much here, because page is possibly not locked:
 185          * sys_munmap()->...
 186          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 187          */
 188         vvp_write_pending(obj, cpg);
 189 #endif
 190         RETURN(__set_page_dirty_nobuffers(vmpage));
 191 }
 192
 193 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 194
 195 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 196                                     size_t size, struct page ***pages,
 197                                     int *max_pages)
 198 {
 199         int result = -ENOMEM;
 200
 201         /* set an arbitrary limit to prevent arithmetic overflow */
 202         if (size > MAX_DIRECTIO_SIZE) {
 203                 *pages = NULL;
 204                 return -EFBIG;
 205         }
 206
 207         *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 208         *max_pages -= user_addr >> CFS_PAGE_SHIFT;
 209
 210         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 211         if (*pages) {
 212                 down_read(&current->mm->mmap_sem);
 213                 result = get_user_pages(current, current->mm, user_addr,
 214                                         *max_pages, (rw == READ), 0, *pages,
 215                                         NULL);
 216                 up_read(&current->mm->mmap_sem);
 217                 if (unlikely(result <= 0))
 218                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 219         }
 220
 221         return result;
 222 }
 223
 224 /*  ll_free_user_pages - tear down page struct array
 225  *  @pages: array of page struct pointers underlying target buffer */
 226 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 227 {
 228         int i;
 229
 230         for (i = 0; i < npages; i++) {
 231                 if (pages[i] == NULL)
 232                         break;
 233                 if (do_dirty)
 234                         set_page_dirty_lock(pages[i]);
 235                 page_cache_release(pages[i]);
 236         }
 237
 238         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 239 }
 240
 241 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 242                            int rw, struct inode *inode,
 243                            struct ll_dio_pages *pv)
 244 {
 245         struct cl_page    *clp;
 246         struct cl_2queue  *queue;
 247         struct cl_object  *obj = io->ci_obj;
 248         int i;
 249         ssize_t rc = 0;
 250         loff_t file_offset  = pv->ldp_start_offset;
 251         long size           = pv->ldp_size;
 252         int page_count      = pv->ldp_nr;
 253         struct page **pages = pv->ldp_pages;
 254         long page_size      = cl_page_size(obj);
 255         bool do_io;
 256         int  io_pages       = 0;
 257         ENTRY;
 258
 259         queue = &io->ci_queue;
 260         cl_2queue_init(queue);
 261         for (i = 0; i < page_count; i++) {
 262                 if (pv->ldp_offsets)
 263                     file_offset = pv->ldp_offsets[i];
 264
 265                 LASSERT(!(file_offset & (page_size - 1)));
 266                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 267                                    pv->ldp_pages[i], CPT_TRANSIENT);
 268                 if (IS_ERR(clp)) {
 269                         rc = PTR_ERR(clp);
 270                         break;
 271                 }
 272
 273                 rc = cl_page_own(env, io, clp);
 274                 if (rc) {
 275                         LASSERT(clp->cp_state == CPS_FREEING);
 276                         cl_page_put(env, clp);
 277                         break;
 278                 }
 279
 280                 do_io = true;
 281
 282                 /* check the page type: if the page is a host page, then do
 283                  * write directly */
 284                 if (clp->cp_type == CPT_CACHEABLE) {
 285                         cfs_page_t *vmpage = cl_page_vmpage(env, clp);
 286                         cfs_page_t *src_page;
 287                         cfs_page_t *dst_page;
 288                         void       *src;
 289                         void       *dst;
 290
 291                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 292                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 293
 294                         src = kmap_atomic(src_page, KM_USER0);
 295                         dst = kmap_atomic(dst_page, KM_USER1);
 296                         memcpy(dst, src, min(page_size, size));
 297                         kunmap_atomic(dst, KM_USER1);
 298                         kunmap_atomic(src, KM_USER0);
 299
 300                         /* make sure page will be added to the transfer by
 301                          * cl_io_submit()->...->vvp_page_prep_write(). */
 302                         if (rw == WRITE)
 303                                 set_page_dirty(vmpage);
 304
 305                         if (rw == READ) {
 306                                 /* do not issue the page for read, since it
 307                                  * may reread a ra page which has NOT uptodate
 308                                  * bit set. */
 309                                 cl_page_disown(env, io, clp);
 310                                 do_io = false;
 311                         }
 312                 }
 313
 314                 if (likely(do_io)) {
 315                         cl_2queue_add(queue, clp);
 316
 317                         /*
 318                          * Set page clip to tell transfer formation engine
 319                          * that page has to be sent even if it is beyond KMS.
 320                          */
 321                         cl_page_clip(env, clp, 0, min(size, page_size));
 322
 323                         ++io_pages;
 324                 }
 325
 326                 /* drop the reference count for cl_page_find */
 327                 cl_page_put(env, clp);
 328                 size -= page_size;
 329                 file_offset += page_size;
 330         }
 331
 332         if (rc == 0 && io_pages) {
 333                 rc = cl_io_submit_sync(env, io,
 334                                        rw == READ ? CRT_READ : CRT_WRITE,
 335                                        queue, 0);
 336         }
 337         if (rc == 0)
 338                 rc = pv->ldp_size;
 339
 340         cl_2queue_discard(env, io, queue);
 341         cl_2queue_disown(env, io, queue);
 342         cl_2queue_fini(env, queue);
 343         RETURN(rc);
 344 }
 345 EXPORT_SYMBOL(ll_direct_rw_pages);
 346
 347 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 348                                    int rw, struct inode *inode,
 349                                    struct address_space *mapping,
 350                                    size_t size, loff_t file_offset,
 351                                    struct page **pages, int page_count)
 352 {
 353     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 354                                  .ldp_nr           = page_count,
 355                                  .ldp_size         = size,
 356                                  .ldp_offsets      = NULL,
 357                                  .ldp_start_offset = file_offset
 358                                };
 359
 360     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 361 }
 362
 363 #ifdef KMALLOC_MAX_SIZE
 364 #define MAX_MALLOC KMALLOC_MAX_SIZE
 365 #else
 366 #define MAX_MALLOC (128 * 1024)
 367 #endif
 368
 369 /* This is the maximum size of a single O_DIRECT request, based on the
 370  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 371  * representing PAGE_SIZE worth of user data, into a single buffer, and
 372  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 373  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 374 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 375                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 376 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 377                                const struct iovec *iov, loff_t file_offset,
 378                                unsigned long nr_segs)
 379 {
 380         struct lu_env *env;
 381         struct cl_io *io;
 382         struct file *file = iocb->ki_filp;
 383         struct inode *inode = file->f_mapping->host;
 384         struct ccc_object *obj = cl_inode2ccc(inode);
 385         long count = iov_length(iov, nr_segs);
 386         long tot_bytes = 0, result = 0;
 387         struct ll_inode_info *lli = ll_i2info(inode);
 388         unsigned long seg = 0;
 389         long size = MAX_DIO_SIZE;
 390         int refcheck;
 391         ENTRY;
 392
 393         if (!lli->lli_has_smd)
 394                 RETURN(-EBADF);
 395
 396         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 397         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 398                 RETURN(-EINVAL);
 399
 400         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
 401                "offset=%lld=%llx, pages %lu (max %lu)\n",
 402                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 403                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 404                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 405
 406         /* Check that all user buffers are aligned as well */
 407         for (seg = 0; seg < nr_segs; seg++) {
 408                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 409                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 410                         RETURN(-EINVAL);
 411         }
 412
 413         env = cl_env_get(&refcheck);
 414         LASSERT(!IS_ERR(env));
 415         io = ccc_env_io(env)->cui_cl.cis_io;
 416         LASSERT(io != NULL);
 417
 418         /* 0. Need locking between buffered and direct access. and race with
 419          *    size changing by concurrent truncates and writes.
 420          * 1. Need inode mutex to operate transient pages.
 421          */
 422         if (rw == READ)
 423                 mutex_lock(&inode->i_mutex);
 424
 425         LASSERT(obj->cob_transient_pages == 0);
 426         for (seg = 0; seg < nr_segs; seg++) {
 427                 long iov_left = iov[seg].iov_len;
 428                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 429
 430                 if (rw == READ) {
 431                         if (file_offset >= i_size_read(inode))
 432                                 break;
 433                         if (file_offset + iov_left > i_size_read(inode))
 434                                 iov_left = i_size_read(inode) - file_offset;
 435                 }
 436
 437                 while (iov_left > 0) {
 438                         struct page **pages;
 439                         int page_count, max_pages = 0;
 440                         long bytes;
 441
 442                         bytes = min(size, iov_left);
 443                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 444                                                        &pages, &max_pages);
 445                         if (likely(page_count > 0)) {
 446                                 if (unlikely(page_count <  max_pages))
 447                                         bytes = page_count << CFS_PAGE_SHIFT;
 448                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 449                                                              file->f_mapping,
 450                                                              bytes, file_offset,
 451                                                              pages, page_count);
 452                                 ll_free_user_pages(pages, max_pages, rw==READ);
 453                         } else if (page_count == 0) {
 454                                 GOTO(out, result = -EFAULT);
 455                         } else {
 456                                 result = page_count;
 457                         }
 458                         if (unlikely(result <= 0)) {
 459                                 /* If we can't allocate a large enough buffer
 460                                  * for the request, shrink it to a smaller
 461                                  * PAGE_SIZE multiple and try again.
 462                                  * We should always be able to kmalloc for a
 463                                  * page worth of page pointers = 4MB on i386. */
 464                                 if (result == -ENOMEM &&
 465                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 466                                            CFS_PAGE_SIZE) {
 467                                         size = ((((size / 2) - 1) |
 468                                                  ~CFS_PAGE_MASK) + 1) &
 469                                                 CFS_PAGE_MASK;
 470                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 471                                                size);
 472                                         continue;
 473                                 }
 474
 475                                 GOTO(out, result);
 476                         }
 477
 478                         tot_bytes += result;
 479                         file_offset += result;
 480                         iov_left -= result;
 481                         user_addr += result;
 482                 }
 483         }
 484 out:
 485         LASSERT(obj->cob_transient_pages == 0);
 486         if (rw == READ)
 487                 mutex_unlock(&inode->i_mutex);
 488
 489         if (tot_bytes > 0) {
 490                 if (rw == WRITE) {
 491                         struct lov_stripe_md *lsm;
 492
 493                         lsm = ccc_inode_lsm_get(inode);
 494                         LASSERT(lsm != NULL);
 495                         lov_stripe_lock(lsm);
 496                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 497                         lov_stripe_unlock(lsm);
 498                         ccc_inode_lsm_put(inode, lsm);
 499                 }
 500         }
 501
 502         cl_env_put(env, &refcheck);
 503         RETURN(tot_bytes ? : result);
 504 }
 505
 506 #if defined(HAVE_KERNEL_WRITE_BEGIN_END) || defined(MS_HAS_NEW_AOPS)
 507 static int ll_write_begin(struct file *file, struct address_space *mapping,
 508                          loff_t pos, unsigned len, unsigned flags,
 509                          struct page **pagep, void **fsdata)
 510 {
 511         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 512         struct page *page;
 513         int rc;
 514         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 515         ENTRY;
 516
 517         page = grab_cache_page_write_begin(mapping, index, flags);
 518         if (!page)
 519                 RETURN(-ENOMEM);
 520
 521         *pagep = page;
 522
 523         rc = ll_prepare_write(file, page, from, from + len);
 524         if (rc) {
 525                 unlock_page(page);
 526                 page_cache_release(page);
 527         }
 528         RETURN(rc);
 529 }
 530
 531 static int ll_write_end(struct file *file, struct address_space *mapping,
 532                         loff_t pos, unsigned len, unsigned copied,
 533                         struct page *page, void *fsdata)
 534 {
 535         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 536         int rc;
 537
 538         rc = ll_commit_write(file, page, from, from + copied);
 539         unlock_page(page);
 540         page_cache_release(page);
 541
 542         return rc ?: copied;
 543 }
 544 #endif
 545
 546 #ifdef CONFIG_MIGRATION
 547 int ll_migratepage(struct address_space *mapping,
 548                 struct page *newpage, struct page *page
 549 #ifdef HAVE_MIGRATEPAGE_4ARGS
 550                 , enum migrate_mode mode
 551 #endif
 552                 )
 553 {
 554         /* Always fail page migration until we have a proper implementation */
 555         return -EIO;
 556 }
 557 #endif
 558
 559 #ifndef MS_HAS_NEW_AOPS
 560 struct address_space_operations ll_aops = {
 561         .readpage       = ll_readpage,
 562 //        .readpages      = ll_readpages,
 563         .direct_IO      = ll_direct_IO_26,
 564         .writepage      = ll_writepage,
 565         .writepages     = ll_writepages,
 566         .set_page_dirty = ll_set_page_dirty,
 567 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
 568         .write_begin    = ll_write_begin,
 569         .write_end      = ll_write_end,
 570 #else
 571         .prepare_write  = ll_prepare_write,
 572         .commit_write   = ll_commit_write,
 573 #endif
 574         .invalidatepage = ll_invalidatepage,
 575         .releasepage    = (void *)ll_releasepage,
 576 #ifdef CONFIG_MIGRATION
 577         .migratepage    = ll_migratepage,
 578 #endif
 579         .bmap           = NULL
 580 };
 581 #else
 582 struct address_space_operations_ext ll_aops = {
 583         .orig_aops.readpage       = ll_readpage,
 584 //        .orig_aops.readpages      = ll_readpages,
 585         .orig_aops.direct_IO      = ll_direct_IO_26,
 586         .orig_aops.writepage      = ll_writepage,
 587         .orig_aops.writepages     = ll_writepages,
 588         .orig_aops.set_page_dirty = ll_set_page_dirty,
 589         .orig_aops.prepare_write  = ll_prepare_write,
 590         .orig_aops.commit_write   = ll_commit_write,
 591         .orig_aops.invalidatepage = ll_invalidatepage,
 592         .orig_aops.releasepage    = ll_releasepage,
 593 #ifdef CONFIG_MIGRATION
 594         .orig_aops.migratepage    = ll_migratepage,
 595 #endif
 596         .orig_aops.bmap           = NULL,
 597         .write_begin    = ll_write_begin,
 598         .write_end      = ll_write_end
 599 };
 600 #endif