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