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