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