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 int cl_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 result;
  90         int refcheck;
  91
  92         LASSERT(PageLocked(vmpage));
  93         LASSERT(!PageWriteback(vmpage));
  94
  95         /*
  96          * It is safe to not check anything in invalidatepage/releasepage
  97          * below because they are run with page locked and all our io is
  98          * happening with locked page too
  99          */
 100         result = 0;
 101         if (offset == 0) {
 102                 env = cl_env_get(&refcheck);
 103                 if (!IS_ERR(env)) {
 104                         inode = vmpage->mapping->host;
 105                         obj = ll_i2info(inode)->lli_clob;
 106                         if (obj != NULL) {
 107                                 page = cl_vmpage_page(vmpage, obj);
 108                                 if (page != NULL) {
 109                                         lu_ref_add(&page->cp_reference,
 110                                                    "delete", vmpage);
 111                                         cl_page_delete(env, page);
 112                                         result = 1;
 113                                         lu_ref_del(&page->cp_reference,
 114                                                    "delete", vmpage);
 115                                         cl_page_put(env, page);
 116                                 }
 117                         } else
 118                                 LASSERT(vmpage->private == 0);
 119                         cl_env_put(env, &refcheck);
 120                 }
 121         }
 122         return result;
 123 }
 124
 125 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
 126 static int ll_invalidatepage(struct page *page, unsigned long offset)
 127 {
 128         return cl_invalidatepage(page, offset);
 129 }
 130 #else /* !HAVE_INVALIDATEPAGE_RETURN_INT */
 131 static void ll_invalidatepage(struct page *page, unsigned long offset)
 132 {
 133         cl_invalidatepage(page, offset);
 134 }
 135 #endif
 136
 137 #ifdef HAVE_RELEASEPAGE_WITH_INT
 138 #define RELEASEPAGE_ARG_TYPE int
 139 #else
 140 #define RELEASEPAGE_ARG_TYPE gfp_t
 141 #endif
 142 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 143 {
 144         struct cl_env_nest nest;
 145         struct lu_env     *env;
 146         struct cl_object  *obj;
 147         struct cl_page    *page;
 148         struct address_space *mapping;
 149         int result;
 150
 151         LASSERT(PageLocked(vmpage));
 152         if (PageWriteback(vmpage) || PageDirty(vmpage))
 153                 return 0;
 154
 155         mapping = vmpage->mapping;
 156         if (mapping == NULL)
 157                 return 1;
 158
 159         obj = ll_i2info(mapping->host)->lli_clob;
 160         if (obj == NULL)
 161                 return 1;
 162
 163         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 164         if (page_count(vmpage) > 3)
 165                 return 0;
 166
 167         /* TODO: determine what gfp should be used by @gfp_mask. */
 168         env = cl_env_nested_get(&nest);
 169         if (IS_ERR(env))
 170                 /* If we can't allocate an env we won't call cl_page_put()
 171                  * later on which further means it's impossible to drop
 172                  * page refcount by cl_page, so ask kernel to not free
 173                  * this page. */
 174                 return 0;
 175
 176         page = cl_vmpage_page(vmpage, obj);
 177         result = page == NULL;
 178         if (page != NULL) {
 179                 if (cfs_atomic_read(&page->cp_ref) == 1) {
 180                         result = 1;
 181                         cl_page_delete(env, page);
 182                 }
 183                 cl_page_put(env, page);
 184         }
 185         cl_env_nested_put(&nest, env);
 186         return result;
 187 }
 188
 189 static int ll_set_page_dirty(struct page *vmpage)
 190 {
 191 #if 0
 192         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 193         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 194         struct vvp_page   *cpg;
 195
 196         /*
 197          * XXX should page method be called here?
 198          */
 199         LASSERT(&obj->co_cl == page->cp_obj);
 200         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 201         /*
 202          * XXX cannot do much here, because page is possibly not locked:
 203          * sys_munmap()->...
 204          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 205          */
 206         vvp_write_pending(obj, cpg);
 207 #endif
 208         RETURN(__set_page_dirty_nobuffers(vmpage));
 209 }
 210
 211 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 212
 213 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 214                                     size_t size, struct page ***pages,
 215                                     int *max_pages)
 216 {
 217         int result = -ENOMEM;
 218
 219         /* set an arbitrary limit to prevent arithmetic overflow */
 220         if (size > MAX_DIRECTIO_SIZE) {
 221                 *pages = NULL;
 222                 return -EFBIG;
 223         }
 224
 225         *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 226         *max_pages -= user_addr >> CFS_PAGE_SHIFT;
 227
 228         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 229         if (*pages) {
 230                 down_read(&current->mm->mmap_sem);
 231                 result = get_user_pages(current, current->mm, user_addr,
 232                                         *max_pages, (rw == READ), 0, *pages,
 233                                         NULL);
 234                 up_read(&current->mm->mmap_sem);
 235                 if (unlikely(result <= 0))
 236                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 237         }
 238
 239         return result;
 240 }
 241
 242 /*  ll_free_user_pages - tear down page struct array
 243  *  @pages: array of page struct pointers underlying target buffer */
 244 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 245 {
 246         int i;
 247
 248         for (i = 0; i < npages; i++) {
 249                 if (pages[i] == NULL)
 250                         break;
 251                 if (do_dirty)
 252                         set_page_dirty_lock(pages[i]);
 253                 page_cache_release(pages[i]);
 254         }
 255
 256         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 257 }
 258
 259 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 260                            int rw, struct inode *inode,
 261                            struct ll_dio_pages *pv)
 262 {
 263         struct cl_page    *clp;
 264         struct cl_2queue  *queue;
 265         struct cl_object  *obj = io->ci_obj;
 266         int i;
 267         ssize_t rc = 0;
 268         loff_t file_offset  = pv->ldp_start_offset;
 269         long size           = pv->ldp_size;
 270         int page_count      = pv->ldp_nr;
 271         struct page **pages = pv->ldp_pages;
 272         long page_size      = cl_page_size(obj);
 273         bool do_io;
 274         int  io_pages       = 0;
 275         ENTRY;
 276
 277         queue = &io->ci_queue;
 278         cl_2queue_init(queue);
 279         for (i = 0; i < page_count; i++) {
 280                 if (pv->ldp_offsets)
 281                     file_offset = pv->ldp_offsets[i];
 282
 283                 LASSERT(!(file_offset & (page_size - 1)));
 284                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 285                                    pv->ldp_pages[i], CPT_TRANSIENT);
 286                 if (IS_ERR(clp)) {
 287                         rc = PTR_ERR(clp);
 288                         break;
 289                 }
 290
 291                 rc = cl_page_own(env, io, clp);
 292                 if (rc) {
 293                         LASSERT(clp->cp_state == CPS_FREEING);
 294                         cl_page_put(env, clp);
 295                         break;
 296                 }
 297
 298                 do_io = true;
 299
 300                 /* check the page type: if the page is a host page, then do
 301                  * write directly */
 302                 if (clp->cp_type == CPT_CACHEABLE) {
 303                         cfs_page_t *vmpage = cl_page_vmpage(env, clp);
 304                         cfs_page_t *src_page;
 305                         cfs_page_t *dst_page;
 306                         void       *src;
 307                         void       *dst;
 308
 309                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 310                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 311
 312                         src = kmap_atomic(src_page, KM_USER0);
 313                         dst = kmap_atomic(dst_page, KM_USER1);
 314                         memcpy(dst, src, min(page_size, size));
 315                         kunmap_atomic(dst, KM_USER1);
 316                         kunmap_atomic(src, KM_USER0);
 317
 318                         /* make sure page will be added to the transfer by
 319                          * cl_io_submit()->...->vvp_page_prep_write(). */
 320                         if (rw == WRITE)
 321                                 set_page_dirty(vmpage);
 322
 323                         if (rw == READ) {
 324                                 /* do not issue the page for read, since it
 325                                  * may reread a ra page which has NOT uptodate
 326                                  * bit set. */
 327                                 cl_page_disown(env, io, clp);
 328                                 do_io = false;
 329                         }
 330                 }
 331
 332                 if (likely(do_io)) {
 333                         cl_2queue_add(queue, clp);
 334
 335                         /*
 336                          * Set page clip to tell transfer formation engine
 337                          * that page has to be sent even if it is beyond KMS.
 338                          */
 339                         cl_page_clip(env, clp, 0, min(size, page_size));
 340
 341                         ++io_pages;
 342                 }
 343
 344                 /* drop the reference count for cl_page_find */
 345                 cl_page_put(env, clp);
 346                 size -= page_size;
 347                 file_offset += page_size;
 348         }
 349
 350         if (rc == 0 && io_pages) {
 351                 rc = cl_io_submit_sync(env, io,
 352                                        rw == READ ? CRT_READ : CRT_WRITE,
 353                                        queue, CRP_NORMAL, 0);
 354         }
 355         if (rc == 0)
 356                 rc = pv->ldp_size;
 357
 358         cl_2queue_discard(env, io, queue);
 359         cl_2queue_disown(env, io, queue);
 360         cl_2queue_fini(env, queue);
 361         RETURN(rc);
 362 }
 363 EXPORT_SYMBOL(ll_direct_rw_pages);
 364
 365 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 366                                    int rw, struct inode *inode,
 367                                    struct address_space *mapping,
 368                                    size_t size, loff_t file_offset,
 369                                    struct page **pages, int page_count)
 370 {
 371     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 372                                  .ldp_nr           = page_count,
 373                                  .ldp_size         = size,
 374                                  .ldp_offsets      = NULL,
 375                                  .ldp_start_offset = file_offset
 376                                };
 377
 378     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 379 }
 380
 381 #ifdef KMALLOC_MAX_SIZE
 382 #define MAX_MALLOC KMALLOC_MAX_SIZE
 383 #else
 384 #define MAX_MALLOC (128 * 1024)
 385 #endif
 386
 387 /* This is the maximum size of a single O_DIRECT request, based on the
 388  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 389  * representing PAGE_SIZE worth of user data, into a single buffer, and
 390  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 391  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 392 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 393                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 394 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 395                                const struct iovec *iov, loff_t file_offset,
 396                                unsigned long nr_segs)
 397 {
 398         struct lu_env *env;
 399         struct cl_io *io;
 400         struct file *file = iocb->ki_filp;
 401         struct inode *inode = file->f_mapping->host;
 402         struct ccc_object *obj = cl_inode2ccc(inode);
 403         long count = iov_length(iov, nr_segs);
 404         long tot_bytes = 0, result = 0;
 405         struct ll_inode_info *lli = ll_i2info(inode);
 406         struct lov_stripe_md *lsm = lli->lli_smd;
 407         unsigned long seg = 0;
 408         long size = MAX_DIO_SIZE;
 409         int refcheck;
 410         ENTRY;
 411
 412         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 413                 RETURN(-EBADF);
 414
 415         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 416         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 417                 RETURN(-EINVAL);
 418
 419         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
 420                "offset=%lld=%llx, pages %lu (max %lu)\n",
 421                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 422                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 423                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 424
 425         /* Check that all user buffers are aligned as well */
 426         for (seg = 0; seg < nr_segs; seg++) {
 427                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 428                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 429                         RETURN(-EINVAL);
 430         }
 431
 432         env = cl_env_get(&refcheck);
 433         LASSERT(!IS_ERR(env));
 434         io = ccc_env_io(env)->cui_cl.cis_io;
 435         LASSERT(io != NULL);
 436
 437         /* 0. Need locking between buffered and direct access. and race with
 438          *    size changing by concurrent truncates and writes.
 439          * 1. Need inode sem to operate transient pages. */
 440         if (rw == READ)
 441                 LOCK_INODE_MUTEX(inode);
 442
 443         LASSERT(obj->cob_transient_pages == 0);
 444         for (seg = 0; seg < nr_segs; seg++) {
 445                 long iov_left = iov[seg].iov_len;
 446                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 447
 448                 if (rw == READ) {
 449                         if (file_offset >= i_size_read(inode))
 450                                 break;
 451                         if (file_offset + iov_left > i_size_read(inode))
 452                                 iov_left = i_size_read(inode) - file_offset;
 453                 }
 454
 455                 while (iov_left > 0) {
 456                         struct page **pages;
 457                         int page_count, max_pages = 0;
 458                         long bytes;
 459
 460                         bytes = min(size, iov_left);
 461                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 462                                                        &pages, &max_pages);
 463                         if (likely(page_count > 0)) {
 464                                 if (unlikely(page_count <  max_pages))
 465                                         bytes = page_count << CFS_PAGE_SHIFT;
 466                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 467                                                              file->f_mapping,
 468                                                              bytes, file_offset,
 469                                                              pages, page_count);
 470                                 ll_free_user_pages(pages, max_pages, rw==READ);
 471                         } else if (page_count == 0) {
 472                                 GOTO(out, result = -EFAULT);
 473                         } else {
 474                                 result = page_count;
 475                         }
 476                         if (unlikely(result <= 0)) {
 477                                 /* If we can't allocate a large enough buffer
 478                                  * for the request, shrink it to a smaller
 479                                  * PAGE_SIZE multiple and try again.
 480                                  * We should always be able to kmalloc for a
 481                                  * page worth of page pointers = 4MB on i386. */
 482                                 if (result == -ENOMEM &&
 483                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 484                                            CFS_PAGE_SIZE) {
 485                                         size = ((((size / 2) - 1) |
 486                                                  ~CFS_PAGE_MASK) + 1) &
 487                                                 CFS_PAGE_MASK;
 488                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 489                                                size);
 490                                         continue;
 491                                 }
 492
 493                                 GOTO(out, result);
 494                         }
 495
 496                         tot_bytes += result;
 497                         file_offset += result;
 498                         iov_left -= result;
 499                         user_addr += result;
 500                 }
 501         }
 502 out:
 503         LASSERT(obj->cob_transient_pages == 0);
 504         if (rw == READ)
 505                 UNLOCK_INODE_MUTEX(inode);
 506
 507         if (tot_bytes > 0) {
 508                 if (rw == WRITE) {
 509                         lov_stripe_lock(lsm);
 510                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 511                         lov_stripe_unlock(lsm);
 512                 }
 513         }
 514
 515         cl_env_put(env, &refcheck);
 516         RETURN(tot_bytes ? : result);
 517 }
 518
 519 #if defined(HAVE_KERNEL_WRITE_BEGIN_END) || defined(MS_HAS_NEW_AOPS)
 520 static int ll_write_begin(struct file *file, struct address_space *mapping,
 521                          loff_t pos, unsigned len, unsigned flags,
 522                          struct page **pagep, void **fsdata)
 523 {
 524         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 525         struct page *page;
 526         int rc;
 527         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 528         ENTRY;
 529
 530         page = grab_cache_page_write_begin(mapping, index, flags);
 531         if (!page)
 532                 RETURN(-ENOMEM);
 533
 534         *pagep = page;
 535
 536         rc = ll_prepare_write(file, page, from, from + len);
 537         if (rc) {
 538                 unlock_page(page);
 539                 page_cache_release(page);
 540         }
 541         RETURN(rc);
 542 }
 543
 544 static int ll_write_end(struct file *file, struct address_space *mapping,
 545                         loff_t pos, unsigned len, unsigned copied,
 546                         struct page *page, void *fsdata)
 547 {
 548         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 549         int rc;
 550
 551         rc = ll_commit_write(file, page, from, from + copied);
 552         unlock_page(page);
 553         page_cache_release(page);
 554
 555         return rc ?: copied;
 556 }
 557 #endif
 558
 559 #ifdef CONFIG_MIGRATION
 560 int ll_migratepage(struct address_space *mapping,
 561                    struct page *newpage, struct page *page)
 562 {
 563         /* Always fail page migration until we have a proper implementation */
 564         return -EIO;
 565 }
 566 #endif
 567
 568 #ifndef MS_HAS_NEW_AOPS
 569 struct address_space_operations ll_aops = {
 570         .readpage       = ll_readpage,
 571 //        .readpages      = ll_readpages,
 572         .direct_IO      = ll_direct_IO_26,
 573         .writepage      = ll_writepage,
 574         .writepages     = generic_writepages,
 575         .set_page_dirty = ll_set_page_dirty,
 576         .sync_page      = NULL,
 577 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
 578         .write_begin    = ll_write_begin,
 579         .write_end      = ll_write_end,
 580 #else
 581         .prepare_write  = ll_prepare_write,
 582         .commit_write   = ll_commit_write,
 583 #endif
 584         .invalidatepage = ll_invalidatepage,
 585         .releasepage    = (void *)ll_releasepage,
 586 #ifdef CONFIG_MIGRATION
 587         .migratepage    = ll_migratepage,
 588 #endif
 589         .bmap           = NULL
 590 };
 591 #else
 592 struct address_space_operations_ext ll_aops = {
 593         .orig_aops.readpage       = ll_readpage,
 594 //        .orig_aops.readpages      = ll_readpages,
 595         .orig_aops.direct_IO      = ll_direct_IO_26,
 596         .orig_aops.writepage      = ll_writepage,
 597         .orig_aops.writepages     = generic_writepages,
 598         .orig_aops.set_page_dirty = ll_set_page_dirty,
 599         .orig_aops.sync_page      = NULL,
 600         .orig_aops.prepare_write  = ll_prepare_write,
 601         .orig_aops.commit_write   = ll_commit_write,
 602         .orig_aops.invalidatepage = ll_invalidatepage,
 603         .orig_aops.releasepage    = ll_releasepage,
 604 #ifdef CONFIG_MIGRATION
 605         .orig_aops.migratepage    = ll_migratepage,
 606 #endif
 607         .orig_aops.bmap           = NULL,
 608         .write_begin    = ll_write_begin,
 609         .write_end      = ll_write_end
 610 };
 611 #endif