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
  24 #ifndef AUTOCONF_INCLUDED
  25 #include <linux/config.h>
  26 #endif
  27 #include <linux/kernel.h>
  28 #include <linux/mm.h>
  29 #include <linux/string.h>
  30 #include <linux/stat.h>
  31 #include <linux/errno.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/unistd.h>
  34 #include <linux/version.h>
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37
  38 #include <linux/fs.h>
  39 #include <linux/buffer_head.h>
  40 #include <linux/mpage.h>
  41 #include <linux/writeback.h>
  42 #include <linux/stat.h>
  43 #include <asm/uaccess.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_mdc.h>
  51 #include <lustre_lite.h>
  52 #include "llite_internal.h"
  53 #include <linux/lustre_compat25.h>
  54
  55 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  56 {
  57         return ll_writepage(page);
  58 }
  59
  60 /* It is safe to not check anything in invalidatepage/releasepage below
  61    because they are run with page locked and all our io is happening with
  62    locked page too */
  63 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  64 static int ll_invalidatepage(struct page *page, unsigned long offset)
  65 {
  66         if (offset)
  67                 return 0;
  68         if (PagePrivate(page))
  69                 ll_removepage(page);
  70         return 1;
  71 }
  72 #else
  73 static void ll_invalidatepage(struct page *page, unsigned long offset)
  74 {
  75         if (offset == 0 && PagePrivate(page))
  76                 ll_removepage(page);
  77 }
  78 #endif
  79 static int ll_releasepage(struct page *page, gfp_t gfp_mask)
  80 {
  81         if (PagePrivate(page))
  82                 ll_removepage(page);
  83         return 1;
  84 }
  85
  86 static int ll_set_page_dirty(struct page *page)
  87 {
  88         struct ll_async_page *llap;
  89         ENTRY;
  90
  91         llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
  92         if (IS_ERR(llap))
  93                 RETURN(PTR_ERR(llap));
  94
  95         llap_write_pending(page->mapping->host, llap);
  96         RETURN(__set_page_dirty_nobuffers(page));
  97 }
  98
  99 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 100
 101 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 102                                     size_t size, struct page ***pages)
 103 {
 104         int result = -ENOMEM;
 105         int page_count;
 106
 107         /* set an arbitrary limit to prevent arithmetic overflow */
 108         if (size > MAX_DIRECTIO_SIZE) {
 109                 *pages = NULL;
 110                 return -EFBIG;
 111         }
 112
 113         page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 114         page_count -= user_addr >> CFS_PAGE_SHIFT;
 115
 116         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 117         if (*pages) {
 118                 down_read(&current->mm->mmap_sem);
 119                 result = get_user_pages(current, current->mm, user_addr,
 120                                         page_count, (rw == READ), 0, *pages,
 121                                         NULL);
 122                 up_read(&current->mm->mmap_sem);
 123                 if (result < 0)
 124                         OBD_FREE(*pages, page_count * sizeof(**pages));
 125         }
 126
 127         return result;
 128 }
 129
 130 /*  ll_free_user_pages - tear down page struct array
 131  *  @pages: array of page struct pointers underlying target buffer */
 132 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 133 {
 134         int i;
 135
 136         for (i = 0; i < npages; i++) {
 137                 if (do_dirty)
 138                         set_page_dirty_lock(pages[i]);
 139                 page_cache_release(pages[i]);
 140         }
 141
 142         OBD_FREE(pages, npages * sizeof(*pages));
 143 }
 144
 145 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 146                                    struct address_space *mapping,
 147                                    struct obd_info *oinfo,
 148                                    struct ptlrpc_request_set *set,
 149                                    size_t size, loff_t file_offset,
 150                                    struct page **pages, int page_count)
 151 {
 152         struct brw_page *pga;
 153         int i, rc = 0;
 154         size_t length;
 155         ENTRY;
 156
 157         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 158         if (!pga) {
 159                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 160                        (int)sizeof(*pga), page_count);
 161                 RETURN(-ENOMEM);
 162         }
 163
 164         for (i = 0, length = size; length > 0;
 165              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 166                 pga[i].pg = pages[i];
 167                 pga[i].off = file_offset;
 168                 /* To the end of the page, or the length, whatever is less */
 169                 pga[i].count = min_t(int, CFS_PAGE_SIZE -
 170                                           (file_offset & ~CFS_PAGE_MASK),
 171                                      length);
 172                 pga[i].flag = 0;
 173                 if (rw == READ)
 174                         POISON_PAGE(pages[i], 0x0d);
 175         }
 176
 177         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 178                                 ll_i2dtexp(inode), oinfo, page_count,
 179                                 pga, NULL, set);
 180         if (rc == 0)
 181                 rc = size;
 182
 183         OBD_FREE(pga, sizeof(*pga) * page_count);
 184         RETURN(rc);
 185 }
 186
 187 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 188  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 189  * representing PAGE_SIZE worth of user data, into a single buffer, and
 190  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 191 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 192                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 193 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 194                                const struct iovec *iov, loff_t file_offset,
 195                                unsigned long nr_segs)
 196 {
 197         struct file *file = iocb->ki_filp;
 198         struct inode *inode = file->f_mapping->host;
 199         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 200         struct ll_inode_info *lli = ll_i2info(inode);
 201         struct lov_stripe_md *lsm = lli->lli_smd;
 202         struct ptlrpc_request_set *set;
 203         struct obd_info oinfo;
 204         struct obdo oa;
 205         unsigned long seg = 0;
 206         size_t size = MAX_DIO_SIZE;
 207         int opc;
 208         ENTRY;
 209
 210         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 211                 RETURN(-EBADF);
 212
 213         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 214         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 215                 RETURN(-EINVAL);
 216
 217         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 218                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 219                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 220                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 221                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 222
 223         if (rw == WRITE) {
 224                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 225                 opc = CAPA_OPC_OSS_WRITE;
 226                 llap_write_pending(inode, NULL);
 227         } else {
 228                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 229                 opc = CAPA_OPC_OSS_RW;
 230         }
 231
 232         /* Check that all user buffers are aligned as well */
 233         for (seg = 0; seg < nr_segs; seg++) {
 234                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 235                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 236                         RETURN(-EINVAL);
 237         }
 238
 239         set = ptlrpc_prep_set();
 240         if (set == NULL)
 241                 RETURN(-ENOMEM);
 242
 243         ll_inode_fill_obdo(inode, rw, &oa);
 244         oinfo.oi_oa = &oa;
 245         oinfo.oi_md = lsm;
 246         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 247
 248         /* need locking between buffered and direct access. and race with
 249          *size changing by concurrent truncates and writes. */
 250         if (rw == READ)
 251                 LOCK_INODE_MUTEX(inode);
 252
 253         for (seg = 0; seg < nr_segs; seg++) {
 254                 size_t iov_left = iov[seg].iov_len;
 255                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 256
 257                 if (rw == READ) {
 258                         if (file_offset >= inode->i_size)
 259                                 break;
 260                         if (file_offset + iov_left > inode->i_size)
 261                                 iov_left = inode->i_size - file_offset;
 262                 }
 263
 264                 while (iov_left > 0) {
 265                         struct page **pages;
 266                         int page_count;
 267                         ssize_t result;
 268
 269                         page_count = ll_get_user_pages(rw, user_addr,
 270                                                        min(size, iov_left),
 271                                                        &pages);
 272                         LASSERT(page_count != 0);
 273                         if (page_count > 0) {
 274                                 result = ll_direct_IO_26_seg(rw, inode,
 275                                                              file->f_mapping,
 276                                                              &oinfo, set,
 277                                                              min(size,iov_left),
 278                                                              file_offset, pages,
 279                                                              page_count);
 280                                 ll_free_user_pages(pages, page_count, rw==READ);
 281                         } else {
 282                                 result = 0;
 283                         }
 284                         if (page_count < 0 || result <= 0) {
 285                                 /* If we can't allocate a large enough buffer
 286                                  * for the request, shrink it to a smaller
 287                                  * PAGE_SIZE multiple and try again.
 288                                  * We should always be able to kmalloc for a
 289                                  * page worth of page pointers = 4MB on i386. */
 290                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 291                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 292                                            CFS_PAGE_SIZE) {
 293                                         size = ((((size / 2) - 1) |
 294                                                  ~CFS_PAGE_MASK) + 1) &
 295                                                 CFS_PAGE_MASK;
 296                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 297                                                (int)size);
 298                                         continue;
 299                                 }
 300
 301                                 if (tot_bytes <= 0)
 302                                         tot_bytes = page_count < 0 ? page_count : result;
 303                                 GOTO(out, tot_bytes);
 304                         }
 305
 306                         tot_bytes += result;
 307                         file_offset += result;
 308                         iov_left -= result;
 309                         user_addr += result;
 310                 }
 311         }
 312 out:
 313         if (rw == READ)
 314                 UNLOCK_INODE_MUTEX(inode);
 315
 316         if (tot_bytes > 0) {
 317                 int rc;
 318
 319                 rc = ptlrpc_set_wait(set);
 320                 if (rc) {
 321                         tot_bytes = rc;
 322                 } else if (rw == WRITE) {
 323                         lov_stripe_lock(lsm);
 324                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 325                         lov_stripe_unlock(lsm);
 326                 }
 327         }
 328
 329         capa_put(oinfo.oi_capa);
 330         ptlrpc_set_destroy(set);
 331         RETURN(tot_bytes);
 332 }
 333
 334 struct address_space_operations ll_aops = {
 335         .readpage       = ll_readpage,
 336 //        .readpages      = ll_readpages,
 337         .direct_IO      = ll_direct_IO_26,
 338         .writepage      = ll_writepage_26,
 339         .writepages     = generic_writepages,
 340         .set_page_dirty = ll_set_page_dirty,
 341         .sync_page      = NULL,
 342         .prepare_write  = ll_prepare_write,
 343         .commit_write   = ll_commit_write,
 344         .invalidatepage = ll_invalidatepage,
 345         .releasepage    = ll_releasepage,
 346         .bmap           = NULL
 347 };