lustre/llite/rw26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23 #ifdef HAVE_KERNEL_CONFIG_H
  24 #include <linux/config.h>
  25 #endif
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/string.h>
  29 #include <linux/stat.h>
  30 #include <linux/errno.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/unistd.h>
  33 #include <linux/version.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36
  37 #include <linux/fs.h>
  38 #include <linux/buffer_head.h>
  39 #include <linux/mpage.h>
  40 #include <linux/writeback.h>
  41 #include <linux/stat.h>
  42 #include <asm/uaccess.h>
  43 #include <asm/segment.h>
  44 #include <linux/mm.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/smp_lock.h>
  47
  48 #define DEBUG_SUBSYSTEM S_LLITE
  49
  50 #include <lustre_lite.h>
  51 #include "llite_internal.h"
  52 #include <linux/lustre_compat25.h>
  53
  54 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  55 {
  56         return ll_writepage(page);
  57 }
  58
  59 /* It is safe to not check anything in invalidatepage/releasepage below
  60    because they are run with page locked and all our io is happening with
  61    locked page too */
  62 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  63 static int ll_invalidatepage(struct page *page, unsigned long offset)
  64 {
  65         if (offset)
  66                 return 0;
  67         if (PagePrivate(page))
  68                 ll_removepage(page);
  69         return 1;
  70 }
  71 #else
  72 static void ll_invalidatepage(struct page *page, unsigned long offset)
  73 {
  74         if (offset)
  75                 return;
  76         if (PagePrivate(page))
  77                 ll_removepage(page);
  78 }
  79 #endif
  80
  81 static int ll_releasepage(struct page *page, gfp_t gfp_mask)
  82 {
  83         if (PagePrivate(page))
  84                 ll_removepage(page);
  85         return 1;
  86 }
  87
  88 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
  89
  90 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
  91                                     size_t size, struct page ***pages)
  92 {
  93         int result = -ENOMEM;
  94         unsigned long page_count;
  95
  96         /* set an arbitrary limit to prevent arithmetic overflow */
  97         if (size > MAX_DIRECTIO_SIZE) {
  98                 *pages = NULL;
  99                 return -EFBIG;
 100         }
 101
 102         page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 103         page_count -= user_addr >> CFS_PAGE_SHIFT;
 104
 105         OBD_ALLOC_GFP(*pages, page_count * sizeof(**pages), GFP_KERNEL);
 106         if (*pages) {
 107                 down_read(&current->mm->mmap_sem);
 108                 result = get_user_pages(current, current->mm, user_addr,
 109                                         page_count, (rw == READ), 0, *pages,
 110                                         NULL);
 111                 up_read(&current->mm->mmap_sem);
 112         }
 113
 114         return result;
 115 }
 116
 117 /*  ll_free_user_pages - tear down page struct array
 118  *  @pages: array of page struct pointers underlying target buffer */
 119 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 120 {
 121         int i;
 122
 123         for (i = 0; i < npages; i++) {
 124                 if (do_dirty)
 125                         set_page_dirty_lock(pages[i]);
 126                 page_cache_release(pages[i]);
 127         }
 128
 129         OBD_FREE(pages, npages * sizeof(*pages));
 130 }
 131
 132 static ssize_t ll_direct_IO_26_seg(int rw, struct file *file,
 133                                    struct address_space *mapping,
 134                                    struct inode *inode,
 135                                    struct lov_stripe_md *lsm,
 136                                    unsigned long user_addr, size_t size,
 137                                    loff_t file_offset, struct page **pages,
 138                                    int page_count)
 139 {
 140         struct brw_page *pga;
 141         struct obdo oa;
 142         int i, rc = 0;
 143         size_t length;
 144         ENTRY;
 145
 146         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 147         if (!pga)
 148                 RETURN(-ENOMEM);
 149
 150         for (i = 0, length = size; length > 0;
 151              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 152                 pga[i].pg = pages[i];
 153                 pga[i].off = file_offset;
 154                 /* To the end of the page, or the length, whatever is less */
 155                 pga[i].count = min_t(int, CFS_PAGE_SIZE -(file_offset & ~CFS_PAGE_MASK),
 156                                      length);
 157                 pga[i].flag = 0;
 158                 if (rw == READ)
 159                         POISON_PAGE(pages[i], 0x0d);
 160         }
 161
 162         ll_inode_fill_obdo(inode, rw, &oa);
 163
 164         if (rw == WRITE)
 165                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 166                                     LPROC_LL_DIRECT_WRITE, size);
 167         else
 168                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 169                                     LPROC_LL_DIRECT_READ, size);
 170         rc = obd_brw_rqset(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 171                            ll_i2obdexp(inode), &oa, lsm, page_count, pga, NULL);
 172         if (rc == 0) {
 173                 rc = size;
 174                 if (rw == WRITE) {
 175                         lov_stripe_lock(lsm);
 176                         obd_adjust_kms(ll_i2obdexp(inode), lsm, file_offset, 0);
 177                         lov_stripe_unlock(lsm);
 178                 }
 179         }
 180
 181         OBD_FREE(pga, sizeof(*pga) * page_count);
 182         RETURN(rc);
 183 }
 184
 185 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 186                                const struct iovec *iov, loff_t file_offset,
 187                                unsigned long nr_segs)
 188 {
 189         struct file *file = iocb->ki_filp;
 190         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 191         struct ll_inode_info *lli = ll_i2info(file->f_mapping->host);
 192         unsigned long seg;
 193         ENTRY;
 194
 195         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 196                 RETURN(-EBADF);
 197
 198         /* FIXME: io smaller than CFS_PAGE_SIZE is broken on ia64 ??? */
 199         if ((file_offset & (~CFS_PAGE_MASK)) || (count & ~CFS_PAGE_MASK))
 200                 RETURN(-EINVAL);
 201
 202         /* Check that all user buffers are aligned as well */
 203         for (seg = 0; seg < nr_segs; seg++) {
 204                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 205                 size_t size = iov[seg].iov_len;
 206                 if ((user_addr & ~CFS_PAGE_MASK) || (size & ~CFS_PAGE_MASK))
 207                         RETURN(-EINVAL);
 208         }
 209
 210         seg = 0;
 211         while ((seg < nr_segs) && (tot_bytes >= 0)) {
 212                 const struct iovec *vec = &iov[seg++];
 213                 unsigned long user_addr = (unsigned long)vec->iov_base;
 214                 size_t size = vec->iov_len;
 215                 struct page **pages;
 216                 int page_count;
 217                 ssize_t result;
 218
 219                 page_count = ll_get_user_pages(rw, user_addr, size, &pages);
 220                 if (page_count < 0) {
 221                         ll_free_user_pages(pages, 0, 0);
 222                         if (tot_bytes > 0)
 223                                 break;
 224                         return page_count;
 225                 }
 226
 227                 result = ll_direct_IO_26_seg(rw, file, file->f_mapping,
 228                                              file->f_mapping->host,
 229                                              lli->lli_smd, user_addr, size,
 230                                              file_offset, pages, page_count);
 231                 ll_free_user_pages(pages, page_count, rw == READ);
 232
 233                 if (result <= 0) {
 234                         if (tot_bytes > 0)
 235                                 break;
 236                         return result;
 237                 }
 238
 239                 tot_bytes += result;
 240                 file_offset += result;
 241                 if (result < size)
 242                         break;
 243         }
 244         return tot_bytes;
 245 }
 246
 247 struct address_space_operations ll_aops = {
 248         .readpage       = ll_readpage,
 249 //        .readpages      = ll_readpages,
 250         .direct_IO      = ll_direct_IO_26,
 251         .writepage      = ll_writepage_26,
 252         .writepages     = generic_writepages,
 253         .set_page_dirty = __set_page_dirty_nobuffers,
 254         .sync_page      = NULL,
 255         .prepare_write  = ll_prepare_write,
 256         .commit_write   = ll_commit_write,
 257         .invalidatepage = ll_invalidatepage,
 258         .releasepage    = ll_releasepage,
 259         .bmap           = NULL
 260 };
 261
 262 static int wait_on_page_locked_range(struct address_space *mapping,
 263                                      pgoff_t start, pgoff_t end)
 264 {
 265         pgoff_t index;
 266         struct page *page;
 267         int ret = 0;
 268
 269         if (end < start)
 270                 return 0;
 271
 272         for(index = start; index < end; index++) {
 273                 page = find_get_page(mapping, index);
 274                 if (page == NULL)
 275                         continue;
 276
 277                 wait_on_page_locked(page);
 278                 if (PageError(page))
 279                         ret = -EIO;
 280                 cond_resched();
 281         }
 282
 283         /* Check for outstanding write errors */
 284         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 285                 ret = -ENOSPC;
 286         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 287                 ret = -EIO;
 288
 289         return ret;
 290 }
 291
 292 int ll_sync_page_range(struct inode *inode, struct address_space *mapping,
 293                        loff_t pos, size_t count)
 294 {
 295         pgoff_t start = pos >> CFS_PAGE_SHIFT;
 296         pgoff_t end = (pos + count - 1) >> CFS_PAGE_SHIFT;
 297         struct writeback_control wbc;
 298         int ret;
 299
 300         wbc.sync_mode = WB_SYNC_ALL;
 301         wbc.nr_to_write = mapping->nrpages * 2;
 302         wbc.start = start;
 303         wbc.end = end;
 304         ret = generic_writepages(mapping, &wbc);
 305
 306         if (ret == 0)
 307                 ret = wait_on_page_locked_range(mapping, start, end);
 308         return ret;
 309 }
 310