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  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #ifndef AUTOCONF_INCLUDED
  42 #include <linux/config.h>
  43 #endif
  44 #include <linux/kernel.h>
  45 #include <linux/mm.h>
  46 #include <linux/string.h>
  47 #include <linux/stat.h>
  48 #include <linux/errno.h>
  49 #include <linux/smp_lock.h>
  50 #include <linux/unistd.h>
  51 #include <linux/version.h>
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54
  55 #include <linux/fs.h>
  56 #include <linux/buffer_head.h>
  57 #include <linux/writeback.h>
  58 #include <linux/stat.h>
  59 #include <asm/uaccess.h>
  60 #ifdef HAVE_SEGMENT_H
  61 # include <asm/segment.h>
  62 #endif
  63 #include <linux/mm.h>
  64 #include <linux/pagemap.h>
  65 #include <linux/smp_lock.h>
  66
  67 #define DEBUG_SUBSYSTEM S_LLITE
  68
  69 #include <lustre_lite.h>
  70 #include "llite_internal.h"
  71 #include <linux/lustre_compat25.h>
  72
  73 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  74 {
  75         return ll_writepage(page);
  76 }
  77
  78 /* It is safe to not check anything in invalidatepage/releasepage below
  79    because they are run with page locked and all our io is happening with
  80    locked page too */
  81 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  82 static int ll_invalidatepage(struct page *page, unsigned long offset)
  83 {
  84         if (offset)
  85                 return 0;
  86         if (PagePrivate(page))
  87                 ll_removepage(page);
  88         return 1;
  89 }
  90 #else
  91 static void ll_invalidatepage(struct page *page, unsigned long offset)
  92 {
  93         if (offset)
  94                 return;
  95         if (PagePrivate(page))
  96                 ll_removepage(page);
  97 }
  98 #endif
  99
 100 #ifdef HAVE_RELEASEPAGE_WITH_GFP
 101 #define RELEASEPAGE_ARG_TYPE gfp_t
 102 #else
 103 #define RELEASEPAGE_ARG_TYPE int
 104 #endif
 105 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
 106 {
 107         if (PagePrivate(page))
 108                 ll_removepage(page);
 109         return 1;
 110 }
 111
 112 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 113
 114 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 115                                     size_t size, struct page ***pages,
 116                                     int *max_pages)
 117 {
 118         int result = -ENOMEM;
 119
 120         /* set an arbitrary limit to prevent arithmetic overflow */
 121         if (size > MAX_DIRECTIO_SIZE) {
 122                 *pages = NULL;
 123                 return -EFBIG;
 124         }
 125
 126         *max_pages = ((user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT)-
 127                       (user_addr >> CFS_PAGE_SHIFT);
 128
 129         OBD_ALLOC_WAIT(*pages, *max_pages * sizeof(**pages));
 130         if (*pages) {
 131                 down_read(&current->mm->mmap_sem);
 132                 result = get_user_pages(current, current->mm, user_addr,
 133                                         *max_pages, (rw == READ), 0, *pages,
 134                                         NULL);
 135                 up_read(&current->mm->mmap_sem);
 136                 if (unlikely(result <= 0))
 137                         OBD_FREE(*pages, *max_pages * sizeof(**pages));
 138         }
 139
 140         return result;
 141 }
 142
 143 /*  ll_free_user_pages - tear down page struct array
 144  *  @pages: array of page struct pointers underlying target buffer */
 145 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 146 {
 147         int i;
 148
 149         for (i = 0; i < npages; i++) {
 150                 if (pages[i] == NULL)
 151                         break;
 152                 if (do_dirty)
 153                         set_page_dirty_lock(pages[i]);
 154                 page_cache_release(pages[i]);
 155         }
 156         OBD_FREE(pages, npages * sizeof(*pages));
 157 }
 158
 159 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 160                                    struct address_space *mapping,
 161                                    struct obd_info *oinfo,
 162                                    struct ptlrpc_request_set *set,
 163                                    size_t size, loff_t file_offset,
 164                                    struct page **pages, int page_count,
 165                                    unsigned long user_addr, int locked)
 166 {
 167         struct brw_page *pga;
 168         int i, rc = 0, pshift;
 169         size_t length;
 170         ENTRY;
 171
 172         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 173         if (!pga) {
 174                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 175                       (int)sizeof(*pga), page_count);
 176                 RETURN(-ENOMEM);
 177         }
 178
 179         /*
 180          * pshift is something we'll add to ->off to get the in-memory offset,
 181          * also see the OSC_FILE2MEM_OFF macro
 182          */
 183         pshift = (user_addr & ~CFS_PAGE_MASK) - (file_offset & ~CFS_PAGE_MASK);
 184
 185         for (i = 0, length = size; length > 0; i++) {/*i last!*/
 186                 LASSERT(i < page_count);
 187
 188                 pga[i].pg = pages[i];
 189                 pga[i].off = file_offset;
 190                 /* To the end of the page, or the length, whatever is less */
 191                 pga[i].count = min_t(int, CFS_PAGE_SIZE -(user_addr & ~CFS_PAGE_MASK),
 192                                      length);
 193
 194                 pga[i].flag = OBD_BRW_SYNC;
 195                 if (!locked)
 196                         pga[i].flag |= OBD_BRW_SRVLOCK;
 197
 198                 if (rw == READ)
 199                         POISON_PAGE(pages[i], 0x0d);
 200
 201                 length -= pga[i].count;
 202                 file_offset += pga[i].count;
 203                 user_addr += pga[i].count;
 204         }
 205
 206         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 207                            ll_i2obdexp(inode), oinfo, page_count,
 208                            pga, NULL, set, pshift);
 209         if (rc == 0)
 210                 rc = size;
 211
 212         OBD_FREE(pga, sizeof(*pga) * page_count);
 213         RETURN(rc);
 214 }
 215
 216 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 217  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 218  * representing PAGE_SIZE worth of user data, into a single buffer, and
 219  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 220 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 221                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 222
 223 ssize_t ll_direct_IO(int rw, struct file *file,
 224                      const struct iovec *iov, loff_t file_offset,
 225                      unsigned long nr_segs, int locked)
 226 {
 227         struct inode *inode = file->f_mapping->host;
 228         ssize_t count = iov_length(iov, nr_segs);
 229         ssize_t tot_bytes = 0, result = 0;
 230         struct ll_inode_info *lli = ll_i2info(inode);
 231         struct lov_stripe_md *lsm = lli->lli_smd;
 232         struct ptlrpc_request_set *set;
 233         struct obd_info oinfo;
 234         struct obdo oa = { 0 };
 235         unsigned long seg;
 236         size_t size = MAX_DIO_SIZE;
 237         ENTRY;
 238
 239         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 240                 RETURN(-EBADF);
 241
 242         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 243                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 244                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 245                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 246                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 247
 248         if (rw == WRITE)
 249                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 250         else
 251                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 252
 253         set = ptlrpc_prep_set();
 254         if (set == NULL)
 255                 RETURN(-ENOMEM);
 256
 257         ll_inode_fill_obdo(inode, rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ, &oa);
 258         oinfo.oi_oa = &oa;
 259         oinfo.oi_md = lsm;
 260
 261         /* need locking between buffered and direct access. and race with
 262          *size changing by concurrent truncates and writes. */
 263         if (rw == READ)
 264                 LOCK_INODE_MUTEX(inode);
 265         for (seg = 0; seg < nr_segs; seg++) {
 266                 size_t iov_left = iov[seg].iov_len;
 267                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 268
 269                 if (rw == READ) {
 270                         if (file_offset >= inode->i_size)
 271                                 break;
 272                         if (file_offset + iov_left > inode->i_size)
 273                                 iov_left = inode->i_size - file_offset;
 274                 }
 275
 276                 while (iov_left > 0) {
 277                         struct page **pages;
 278                         int page_count, max_pages = 0;
 279                         size_t bytes;
 280
 281                         bytes = min(size,iov_left);
 282                         page_count = ll_get_user_pages(rw, user_addr,
 283                                                        bytes,
 284                                                        &pages, &max_pages);
 285                         if (likely(page_count > 0)) {
 286                                 if (unlikely(page_count <  max_pages))
 287                                         bytes = page_count << CFS_PAGE_SHIFT;
 288                                 result = ll_direct_IO_26_seg(rw, inode,
 289                                                              file->f_mapping,
 290                                                              &oinfo, set,
 291                                                              bytes,
 292                                                              file_offset, pages,
 293                                                              page_count,
 294                                                              user_addr, locked);
 295                                 ll_free_user_pages(pages, max_pages, rw==READ);
 296                         } else if (page_count == 0) {
 297                                 GOTO(out, result = -EFAULT);
 298                         } else {
 299                                 result = page_count;
 300                         }
 301                         if (unlikely(result <= 0)) {
 302                                 /* If we can't allocate a large enough buffer
 303                                  * for the request, shrink it to a smaller
 304                                  * PAGE_SIZE multiple and try again.
 305                                  * We should always be able to kmalloc for a
 306                                  * page worth of page pointers = 4MB on i386. */
 307                                 if (result == -ENOMEM &&
 308                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 309                                            CFS_PAGE_SIZE) {
 310                                         size = ((((size / 2) - 1) |
 311                                                  ~CFS_PAGE_MASK) + 1) &
 312                                                 CFS_PAGE_MASK;
 313                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 314                                                (int)size);
 315                                         continue;
 316                                 }
 317                                 GOTO(out, result);
 318                         }
 319                         tot_bytes += result;
 320                         file_offset += result;
 321                         iov_left -= result;
 322                         user_addr += result;
 323                 }
 324         }
 325 out:
 326         if (likely(tot_bytes > 0)) {
 327                 int rc;
 328
 329                 rc = ptlrpc_set_wait(set);
 330                 if (unlikely(rc != 0))
 331                         GOTO(unlock_mutex, tot_bytes = rc);
 332                 if (rw == WRITE && locked) {
 333                         lov_stripe_lock(lsm);
 334                         obd_adjust_kms(ll_i2obdexp(inode),
 335                                        lsm, file_offset, 0);
 336                         lov_stripe_unlock(lsm);
 337                 }
 338         } else {
 339                 tot_bytes = result;
 340         }
 341 unlock_mutex:
 342         if (rw == READ)
 343                 UNLOCK_INODE_MUTEX(inode);
 344
 345         ptlrpc_set_destroy(set);
 346         RETURN(tot_bytes);
 347 }
 348
 349 static ssize_t ll_direct_IO_26(int rw, struct kiocb *kiocb,
 350                                const struct iovec *iov, loff_t file_offset,
 351                                unsigned long nr_segs)
 352 {
 353         return ll_direct_IO(rw, kiocb->ki_filp, iov, file_offset, nr_segs, 1);
 354 }
 355
 356 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
 357 static int ll_write_begin(struct file *file, struct address_space *mapping,
 358                          loff_t pos, unsigned len, unsigned flags,
 359                          struct page **pagep, void **fsdata)
 360 {
 361         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 362         struct page *page;
 363         int rc;
 364         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 365         ENTRY;
 366
 367         page = grab_cache_page_write_begin(mapping, index, flags);
 368         if (!page)
 369                 RETURN(-ENOMEM);
 370
 371         *pagep = page;
 372
 373         rc = ll_prepare_write(file, page, from, from + len);
 374         if (rc) {
 375                 unlock_page(page);
 376                 page_cache_release(page);
 377         }
 378         RETURN(rc);
 379 }
 380
 381 static int ll_write_end(struct file *file, struct address_space *mapping,
 382                         loff_t pos, unsigned len, unsigned copied,
 383                         struct page *page, void *fsdata)
 384 {
 385         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 386         int rc;
 387         rc = ll_commit_write(file, page, from, from + copied);
 388
 389         unlock_page(page);
 390         page_cache_release(page);
 391         return rc?rc:copied;
 392 }
 393 #endif
 394
 395 struct address_space_operations ll_aops = {
 396         .readpage       = ll_readpage,
 397 //        .readpages      = ll_readpages,
 398         .direct_IO      = ll_direct_IO_26,
 399         .writepage      = ll_writepage_26,
 400         .writepages     = generic_writepages,
 401         .set_page_dirty = __set_page_dirty_nobuffers,
 402         .sync_page      = NULL,
 403 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
 404         .write_begin    = ll_write_begin,
 405         .write_end      = ll_write_end,
 406 #else
 407         .prepare_write  = ll_prepare_write,
 408         .commit_write   = ll_commit_write,
 409 #endif
 410         .invalidatepage = ll_invalidatepage,
 411         .releasepage    = ll_releasepage,
 412         .bmap           = NULL
 413 };