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, Intel Corporation.
  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 <asm/uaccess.h>
  48
  49 #ifdef HAVE_MIGRATE_H
  50 #include <linux/migrate.h>
  51 #elif defined(HAVE_MIGRATE_MODE_H)
  52 #include <linux/migrate_mode.h>
  53 #endif
  54 #include <linux/fs.h>
  55 #include <linux/buffer_head.h>
  56 #include <linux/mpage.h>
  57 #include <linux/writeback.h>
  58 #include <linux/stat.h>
  59 #include <asm/uaccess.h>
  60 #include <linux/mm.h>
  61 #include <linux/pagemap.h>
  62
  63 #define DEBUG_SUBSYSTEM S_LLITE
  64
  65 #include <lustre_lite.h>
  66 #include "llite_internal.h"
  67 #include <linux/lustre_compat25.h>
  68
  69 /**
  70  * Implements Linux VM address_space::invalidatepage() method. This method is
  71  * called when the page is truncate from a file, either as a result of
  72  * explicit truncate, or when inode is removed from memory (as a result of
  73  * final iput(), umount, or memory pressure induced icache shrinking).
  74  *
  75  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  76  * aligned truncate). Lustre leaves partially truncated page in the cache,
  77  * relying on struct inode::i_size to limit further accesses.
  78  */
  79 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
  80 {
  81         struct inode     *inode;
  82         struct lu_env    *env;
  83         struct cl_page   *page;
  84         struct cl_object *obj;
  85
  86         int refcheck;
  87
  88         LASSERT(PageLocked(vmpage));
  89         LASSERT(!PageWriteback(vmpage));
  90
  91         /*
  92          * It is safe to not check anything in invalidatepage/releasepage
  93          * below because they are run with page locked and all our io is
  94          * happening with locked page too
  95          */
  96         if (offset == 0) {
  97                 env = cl_env_get(&refcheck);
  98                 if (!IS_ERR(env)) {
  99                         inode = vmpage->mapping->host;
 100                         obj = ll_i2info(inode)->lli_clob;
 101                         if (obj != NULL) {
 102                                 page = cl_vmpage_page(vmpage, obj);
 103                                 if (page != NULL) {
 104                                         cl_page_delete(env, page);
 105                                         cl_page_put(env, page);
 106                                 }
 107                         } else
 108                                 LASSERT(vmpage->private == 0);
 109                         cl_env_put(env, &refcheck);
 110                 }
 111         }
 112 }
 113
 114 #ifdef HAVE_RELEASEPAGE_WITH_INT
 115 #define RELEASEPAGE_ARG_TYPE int
 116 #else
 117 #define RELEASEPAGE_ARG_TYPE gfp_t
 118 #endif
 119 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 120 {
 121         struct lu_env           *env;
 122         void                    *cookie;
 123         struct cl_object        *obj;
 124         struct cl_page          *page;
 125         struct address_space    *mapping;
 126         int result = 0;
 127
 128         LASSERT(PageLocked(vmpage));
 129         if (PageWriteback(vmpage) || PageDirty(vmpage))
 130                 return 0;
 131
 132         mapping = vmpage->mapping;
 133         if (mapping == NULL)
 134                 return 1;
 135
 136         obj = ll_i2info(mapping->host)->lli_clob;
 137         if (obj == NULL)
 138                 return 1;
 139
 140         /* 1 for caller, 1 for cl_page and 1 for page cache */
 141         if (page_count(vmpage) > 3)
 142                 return 0;
 143
 144         page = cl_vmpage_page(vmpage, obj);
 145         if (page == NULL)
 146                 return 1;
 147
 148         cookie = cl_env_reenter();
 149         env = cl_env_percpu_get();
 150         LASSERT(!IS_ERR(env));
 151
 152         if (!cl_page_in_use(page)) {
 153                 result = 1;
 154                 cl_page_delete(env, page);
 155         }
 156
 157         /* To use percpu env array, the call path can not be rescheduled;
 158          * otherwise percpu array will be messed if ll_releaspage() called
 159          * again on the same CPU.
 160          *
 161          * If this page holds the last refc of cl_object, the following
 162          * call path may cause reschedule:
 163          *   cl_page_put -> cl_page_free -> cl_object_put ->
 164          *     lu_object_put -> lu_object_free -> lov_delete_raid0 ->
 165          *     cl_locks_prune.
 166          *
 167          * However, the kernel can't get rid of this inode until all pages have
 168          * been cleaned up. Now that we hold page lock here, it's pretty safe
 169          * that we won't get into object delete path.
 170          */
 171         LASSERT(cl_object_refc(obj) > 1);
 172         cl_page_put(env, page);
 173
 174         cl_env_percpu_put(env);
 175         cl_env_reexit(cookie);
 176         return result;
 177 }
 178
 179 static int ll_set_page_dirty(struct page *vmpage)
 180 {
 181 #if 0
 182         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 183         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 184         struct vvp_page   *cpg;
 185
 186         /*
 187          * XXX should page method be called here?
 188          */
 189         LASSERT(&obj->co_cl == page->cp_obj);
 190         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 191         /*
 192          * XXX cannot do much here, because page is possibly not locked:
 193          * sys_munmap()->...
 194          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 195          */
 196         vvp_write_pending(obj, cpg);
 197 #endif
 198         RETURN(__set_page_dirty_nobuffers(vmpage));
 199 }
 200
 201 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 202
 203 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 204                                     size_t size, struct page ***pages,
 205                                     int *max_pages)
 206 {
 207         int result = -ENOMEM;
 208
 209         /* set an arbitrary limit to prevent arithmetic overflow */
 210         if (size > MAX_DIRECTIO_SIZE) {
 211                 *pages = NULL;
 212                 return -EFBIG;
 213         }
 214
 215         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
 216                      PAGE_CACHE_SHIFT;
 217         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 218
 219         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 220         if (*pages) {
 221                 down_read(&current->mm->mmap_sem);
 222                 result = get_user_pages(current, current->mm, user_addr,
 223                                         *max_pages, (rw == READ), 0, *pages,
 224                                         NULL);
 225                 up_read(&current->mm->mmap_sem);
 226                 if (unlikely(result <= 0))
 227                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 228         }
 229
 230         return result;
 231 }
 232
 233 /*  ll_free_user_pages - tear down page struct array
 234  *  @pages: array of page struct pointers underlying target buffer */
 235 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 236 {
 237         int i;
 238
 239         for (i = 0; i < npages; i++) {
 240                 if (pages[i] == NULL)
 241                         break;
 242                 if (do_dirty)
 243                         set_page_dirty_lock(pages[i]);
 244                 page_cache_release(pages[i]);
 245         }
 246
 247         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 248 }
 249
 250 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 251                            int rw, struct inode *inode,
 252                            struct ll_dio_pages *pv)
 253 {
 254         struct cl_page    *clp;
 255         struct cl_2queue  *queue;
 256         struct cl_object  *obj = io->ci_obj;
 257         int i;
 258         ssize_t rc = 0;
 259         loff_t file_offset  = pv->ldp_start_offset;
 260         long size           = pv->ldp_size;
 261         int page_count      = pv->ldp_nr;
 262         struct page **pages = pv->ldp_pages;
 263         long page_size      = cl_page_size(obj);
 264         bool do_io;
 265         int  io_pages       = 0;
 266         ENTRY;
 267
 268         queue = &io->ci_queue;
 269         cl_2queue_init(queue);
 270         for (i = 0; i < page_count; i++) {
 271                 if (pv->ldp_offsets)
 272                     file_offset = pv->ldp_offsets[i];
 273
 274                 LASSERT(!(file_offset & (page_size - 1)));
 275                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 276                                    pv->ldp_pages[i], CPT_TRANSIENT);
 277                 if (IS_ERR(clp)) {
 278                         rc = PTR_ERR(clp);
 279                         break;
 280                 }
 281
 282                 rc = cl_page_own(env, io, clp);
 283                 if (rc) {
 284                         LASSERT(clp->cp_state == CPS_FREEING);
 285                         cl_page_put(env, clp);
 286                         break;
 287                 }
 288
 289                 do_io = true;
 290
 291                 /* check the page type: if the page is a host page, then do
 292                  * write directly */
 293                 if (clp->cp_type == CPT_CACHEABLE) {
 294                         struct page *vmpage = cl_page_vmpage(env, clp);
 295                         struct page *src_page;
 296                         struct page *dst_page;
 297                         void       *src;
 298                         void       *dst;
 299
 300                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 301                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 302
 303                         src = ll_kmap_atomic(src_page, KM_USER0);
 304                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 305                         memcpy(dst, src, min(page_size, size));
 306                         ll_kunmap_atomic(dst, KM_USER1);
 307                         ll_kunmap_atomic(src, KM_USER0);
 308
 309                         /* make sure page will be added to the transfer by
 310                          * cl_io_submit()->...->vvp_page_prep_write(). */
 311                         if (rw == WRITE)
 312                                 set_page_dirty(vmpage);
 313
 314                         if (rw == READ) {
 315                                 /* do not issue the page for read, since it
 316                                  * may reread a ra page which has NOT uptodate
 317                                  * bit set. */
 318                                 cl_page_disown(env, io, clp);
 319                                 do_io = false;
 320                         }
 321                 }
 322
 323                 if (likely(do_io)) {
 324                         cl_2queue_add(queue, clp);
 325
 326                         /*
 327                          * Set page clip to tell transfer formation engine
 328                          * that page has to be sent even if it is beyond KMS.
 329                          */
 330                         cl_page_clip(env, clp, 0, min(size, page_size));
 331
 332                         ++io_pages;
 333                 }
 334
 335                 /* drop the reference count for cl_page_find */
 336                 cl_page_put(env, clp);
 337                 size -= page_size;
 338                 file_offset += page_size;
 339         }
 340
 341         if (rc == 0 && io_pages) {
 342                 rc = cl_io_submit_sync(env, io,
 343                                        rw == READ ? CRT_READ : CRT_WRITE,
 344                                        queue, 0);
 345         }
 346         if (rc == 0)
 347                 rc = pv->ldp_size;
 348
 349         cl_2queue_discard(env, io, queue);
 350         cl_2queue_disown(env, io, queue);
 351         cl_2queue_fini(env, queue);
 352         RETURN(rc);
 353 }
 354 EXPORT_SYMBOL(ll_direct_rw_pages);
 355
 356 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 357                                    int rw, struct inode *inode,
 358                                    struct address_space *mapping,
 359                                    size_t size, loff_t file_offset,
 360                                    struct page **pages, int page_count)
 361 {
 362     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 363                                  .ldp_nr           = page_count,
 364                                  .ldp_size         = size,
 365                                  .ldp_offsets      = NULL,
 366                                  .ldp_start_offset = file_offset
 367                                };
 368
 369     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 370 }
 371
 372 #ifdef KMALLOC_MAX_SIZE
 373 #define MAX_MALLOC KMALLOC_MAX_SIZE
 374 #else
 375 #define MAX_MALLOC (128 * 1024)
 376 #endif
 377
 378 /* This is the maximum size of a single O_DIRECT request, based on the
 379  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 380  * representing PAGE_SIZE worth of user data, into a single buffer, and
 381  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 382  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 383 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 384                       ~(DT_MAX_BRW_SIZE - 1))
 385 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 386                                const struct iovec *iov, loff_t file_offset,
 387                                unsigned long nr_segs)
 388 {
 389         struct lu_env *env;
 390         struct cl_io *io;
 391         struct file *file = iocb->ki_filp;
 392         struct inode *inode = file->f_mapping->host;
 393         struct ccc_object *obj = cl_inode2ccc(inode);
 394         long count = iov_length(iov, nr_segs);
 395         long tot_bytes = 0, result = 0;
 396         struct ll_inode_info *lli = ll_i2info(inode);
 397         unsigned long seg = 0;
 398         long size = MAX_DIO_SIZE;
 399         int refcheck;
 400         ENTRY;
 401
 402         if (!lli->lli_has_smd)
 403                 RETURN(-EBADF);
 404
 405         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 406         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 407                 RETURN(-EINVAL);
 408
 409         CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%lu (max %lu), "
 410                "offset=%lld=%llx, pages %lu (max %lu)\n",
 411                PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
 412                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 413                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 414
 415         /* Check that all user buffers are aligned as well */
 416         for (seg = 0; seg < nr_segs; seg++) {
 417                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 418                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 419                         RETURN(-EINVAL);
 420         }
 421
 422         env = cl_env_get(&refcheck);
 423         LASSERT(!IS_ERR(env));
 424         io = ccc_env_io(env)->cui_cl.cis_io;
 425         LASSERT(io != NULL);
 426
 427         /* 0. Need locking between buffered and direct access. and race with
 428          *    size changing by concurrent truncates and writes.
 429          * 1. Need inode mutex to operate transient pages.
 430          */
 431         if (rw == READ)
 432                 mutex_lock(&inode->i_mutex);
 433
 434         LASSERT(obj->cob_transient_pages == 0);
 435         for (seg = 0; seg < nr_segs; seg++) {
 436                 long iov_left = iov[seg].iov_len;
 437                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 438
 439                 if (rw == READ) {
 440                         if (file_offset >= i_size_read(inode))
 441                                 break;
 442                         if (file_offset + iov_left > i_size_read(inode))
 443                                 iov_left = i_size_read(inode) - file_offset;
 444                 }
 445
 446                 while (iov_left > 0) {
 447                         struct page **pages;
 448                         int page_count, max_pages = 0;
 449                         long bytes;
 450
 451                         bytes = min(size, iov_left);
 452                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 453                                                        &pages, &max_pages);
 454                         if (likely(page_count > 0)) {
 455                                 if (unlikely(page_count <  max_pages))
 456                                         bytes = page_count << PAGE_CACHE_SHIFT;
 457                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 458                                                              file->f_mapping,
 459                                                              bytes, file_offset,
 460                                                              pages, page_count);
 461                                 ll_free_user_pages(pages, max_pages, rw==READ);
 462                         } else if (page_count == 0) {
 463                                 GOTO(out, result = -EFAULT);
 464                         } else {
 465                                 result = page_count;
 466                         }
 467                         if (unlikely(result <= 0)) {
 468                                 /* If we can't allocate a large enough buffer
 469                                  * for the request, shrink it to a smaller
 470                                  * PAGE_SIZE multiple and try again.
 471                                  * We should always be able to kmalloc for a
 472                                  * page worth of page pointers = 4MB on i386. */
 473                                 if (result == -ENOMEM &&
 474                                     size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
 475                                            PAGE_CACHE_SIZE) {
 476                                         size = ((((size / 2) - 1) |
 477                                                  ~CFS_PAGE_MASK) + 1) &
 478                                                 CFS_PAGE_MASK;
 479                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 480                                                size);
 481                                         continue;
 482                                 }
 483
 484                                 GOTO(out, result);
 485                         }
 486
 487                         tot_bytes += result;
 488                         file_offset += result;
 489                         iov_left -= result;
 490                         user_addr += result;
 491                 }
 492         }
 493 out:
 494         LASSERT(obj->cob_transient_pages == 0);
 495         if (rw == READ)
 496                 mutex_unlock(&inode->i_mutex);
 497
 498         if (tot_bytes > 0) {
 499                 if (rw == WRITE) {
 500                         struct lov_stripe_md *lsm;
 501
 502                         lsm = ccc_inode_lsm_get(inode);
 503                         LASSERT(lsm != NULL);
 504                         lov_stripe_lock(lsm);
 505                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 506                         lov_stripe_unlock(lsm);
 507                         ccc_inode_lsm_put(inode, lsm);
 508                 }
 509         }
 510
 511         cl_env_put(env, &refcheck);
 512         RETURN(tot_bytes ? tot_bytes : result);
 513 }
 514
 515 static int ll_write_begin(struct file *file, struct address_space *mapping,
 516                          loff_t pos, unsigned len, unsigned flags,
 517                          struct page **pagep, void **fsdata)
 518 {
 519         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 520         struct page *page;
 521         int rc;
 522         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 523         ENTRY;
 524
 525         page = grab_cache_page_write_begin(mapping, index, flags);
 526         if (!page)
 527                 RETURN(-ENOMEM);
 528
 529         *pagep = page;
 530
 531         rc = ll_prepare_write(file, page, from, from + len);
 532         if (rc) {
 533                 unlock_page(page);
 534                 page_cache_release(page);
 535         }
 536         RETURN(rc);
 537 }
 538
 539 static int ll_write_end(struct file *file, struct address_space *mapping,
 540                         loff_t pos, unsigned len, unsigned copied,
 541                         struct page *page, void *fsdata)
 542 {
 543         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 544         int rc;
 545
 546         rc = ll_commit_write(file, page, from, from + copied);
 547         unlock_page(page);
 548         page_cache_release(page);
 549
 550         return rc ?: copied;
 551 }
 552
 553 #ifdef CONFIG_MIGRATION
 554 int ll_migratepage(struct address_space *mapping,
 555                 struct page *newpage, struct page *page
 556 #ifdef HAVE_MIGRATEPAGE_4ARGS
 557                 , enum migrate_mode mode
 558 #endif
 559                 )
 560 {
 561         /* Always fail page migration until we have a proper implementation */
 562         return -EIO;
 563 }
 564 #endif
 565
 566 #ifndef MS_HAS_NEW_AOPS
 567 struct address_space_operations ll_aops = {
 568         .readpage       = ll_readpage,
 569 //        .readpages      = ll_readpages,
 570         .direct_IO      = ll_direct_IO_26,
 571         .writepage      = ll_writepage,
 572         .writepages     = ll_writepages,
 573         .set_page_dirty = ll_set_page_dirty,
 574         .write_begin    = ll_write_begin,
 575         .write_end      = ll_write_end,
 576         .invalidatepage = ll_invalidatepage,
 577         .releasepage    = (void *)ll_releasepage,
 578 #ifdef CONFIG_MIGRATION
 579         .migratepage    = ll_migratepage,
 580 #endif
 581         .bmap           = NULL
 582 };
 583 #else
 584 struct address_space_operations_ext ll_aops = {
 585         .orig_aops.readpage       = ll_readpage,
 586 //        .orig_aops.readpages      = ll_readpages,
 587         .orig_aops.direct_IO      = ll_direct_IO_26,
 588         .orig_aops.writepage      = ll_writepage,
 589         .orig_aops.writepages     = ll_writepages,
 590         .orig_aops.set_page_dirty = ll_set_page_dirty,
 591         .orig_aops.prepare_write  = ll_prepare_write,
 592         .orig_aops.commit_write   = ll_commit_write,
 593         .orig_aops.invalidatepage = ll_invalidatepage,
 594         .orig_aops.releasepage    = ll_releasepage,
 595 #ifdef CONFIG_MIGRATION
 596         .orig_aops.migratepage    = ll_migratepage,
 597 #endif
 598         .orig_aops.bmap           = NULL,
 599         .write_begin    = ll_write_begin,
 600         .write_end      = ll_write_end
 601 };
 602 #endif