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/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/mpage.h>
  58 #include <linux/writeback.h>
  59 #include <linux/stat.h>
  60 #include <asm/uaccess.h>
  61 #include <linux/mm.h>
  62 #include <linux/pagemap.h>
  63 #include <linux/smp_lock.h>
  64
  65 #define DEBUG_SUBSYSTEM S_LLITE
  66
  67 //#include <lustre_mdc.h>
  68 #include <lustre_lite.h>
  69 #include "llite_internal.h"
  70 #include <linux/lustre_compat25.h>
  71
  72 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  73 {
  74         return ll_writepage(page);
  75 }
  76
  77 /* It is safe to not check anything in invalidatepage/releasepage below
  78    because they are run with page locked and all our io is happening with
  79    locked page too */
  80 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  81 static int ll_invalidatepage(struct page *page, unsigned long offset)
  82 {
  83         if (offset)
  84                 return 0;
  85         if (PagePrivate(page))
  86                 ll_removepage(page);
  87         return 1;
  88 }
  89 #else
  90 static void ll_invalidatepage(struct page *page, unsigned long offset)
  91 {
  92         if (offset == 0 && PagePrivate(page))
  93                 ll_removepage(page);
  94 }
  95 #endif
  96
  97 #ifdef HAVE_RELEASEPAGE_WITH_INT
  98 #define RELEASEPAGE_ARG_TYPE int
  99 #else
 100 #define RELEASEPAGE_ARG_TYPE gfp_t
 101 #endif
 102 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
 103 {
 104         if (PagePrivate(page))
 105                 ll_removepage(page);
 106         return 1;
 107 }
 108
 109 static int ll_set_page_dirty(struct page *page)
 110 {
 111         struct ll_async_page *llap;
 112         ENTRY;
 113
 114         llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
 115         if (IS_ERR(llap))
 116                 RETURN(PTR_ERR(llap));
 117
 118         llap_write_pending(page->mapping->host, llap);
 119         RETURN(__set_page_dirty_nobuffers(page));
 120 }
 121
 122 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 123
 124 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 125                                     size_t size, struct page ***pages)
 126 {
 127         int result = -ENOMEM;
 128         int page_count;
 129
 130         /* set an arbitrary limit to prevent arithmetic overflow */
 131         if (size > MAX_DIRECTIO_SIZE) {
 132                 *pages = NULL;
 133                 return -EFBIG;
 134         }
 135
 136         page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 137         page_count -= user_addr >> CFS_PAGE_SHIFT;
 138
 139         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 140         if (*pages) {
 141                 down_read(&current->mm->mmap_sem);
 142                 result = get_user_pages(current, current->mm, user_addr,
 143                                         page_count, (rw == READ), 0, *pages,
 144                                         NULL);
 145                 up_read(&current->mm->mmap_sem);
 146                 if (result < 0)
 147                         OBD_FREE(*pages, page_count * sizeof(**pages));
 148         }
 149
 150         return result;
 151 }
 152
 153 /*  ll_free_user_pages - tear down page struct array
 154  *  @pages: array of page struct pointers underlying target buffer */
 155 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 156 {
 157         int i;
 158
 159         for (i = 0; i < npages; i++) {
 160                 if (do_dirty)
 161                         set_page_dirty_lock(pages[i]);
 162                 page_cache_release(pages[i]);
 163         }
 164
 165         OBD_FREE(pages, npages * sizeof(*pages));
 166 }
 167
 168 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 169                                    struct address_space *mapping,
 170                                    struct obd_info *oinfo,
 171                                    struct ptlrpc_request_set *set,
 172                                    size_t size, loff_t file_offset,
 173                                    struct page **pages, int page_count)
 174 {
 175         struct brw_page *pga;
 176         int i, rc = 0;
 177         size_t length;
 178         ENTRY;
 179
 180         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 181         if (!pga) {
 182                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 183                        (int)sizeof(*pga), page_count);
 184                 RETURN(-ENOMEM);
 185         }
 186
 187         for (i = 0, length = size; length > 0;
 188              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 189                 pga[i].pg = pages[i];
 190                 pga[i].off = file_offset;
 191                 /* To the end of the page, or the length, whatever is less */
 192                 pga[i].count = min_t(int, CFS_PAGE_SIZE -
 193                                           (file_offset & ~CFS_PAGE_MASK),
 194                                      length);
 195                 pga[i].flag = 0;
 196                 if (rw == READ)
 197                         POISON_PAGE(pages[i], 0x0d);
 198         }
 199
 200         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 201                                 ll_i2dtexp(inode), oinfo, page_count,
 202                                 pga, NULL, set);
 203         if (rc == 0)
 204                 rc = size;
 205
 206         OBD_FREE(pga, sizeof(*pga) * page_count);
 207         RETURN(rc);
 208 }
 209
 210 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 211  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 212  * representing PAGE_SIZE worth of user data, into a single buffer, and
 213  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 214 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 215                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 216 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 217                                const struct iovec *iov, loff_t file_offset,
 218                                unsigned long nr_segs)
 219 {
 220         struct file *file = iocb->ki_filp;
 221         struct inode *inode = file->f_mapping->host;
 222         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 223         struct ll_inode_info *lli = ll_i2info(inode);
 224         struct lov_stripe_md *lsm = lli->lli_smd;
 225         struct ptlrpc_request_set *set;
 226         struct obd_info oinfo;
 227         struct obdo oa;
 228         unsigned long seg = 0;
 229         size_t size = MAX_DIO_SIZE;
 230         int opc;
 231         ENTRY;
 232
 233         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 234                 RETURN(-EBADF);
 235
 236         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 237         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 238                 RETURN(-EINVAL);
 239
 240         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 241                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 242                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 243                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 244                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 245
 246         if (rw == WRITE) {
 247                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 248                 opc = CAPA_OPC_OSS_WRITE;
 249                 llap_write_pending(inode, NULL);
 250         } else {
 251                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 252                 opc = CAPA_OPC_OSS_RW;
 253         }
 254
 255         /* Check that all user buffers are aligned as well */
 256         for (seg = 0; seg < nr_segs; seg++) {
 257                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 258                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 259                         RETURN(-EINVAL);
 260         }
 261
 262         set = ptlrpc_prep_set();
 263         if (set == NULL)
 264                 RETURN(-ENOMEM);
 265
 266         ll_inode_fill_obdo(inode, rw, &oa);
 267         oinfo.oi_oa = &oa;
 268         oinfo.oi_md = lsm;
 269         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 270
 271         /* need locking between buffered and direct access. and race with
 272          *size changing by concurrent truncates and writes. */
 273         if (rw == READ)
 274                 LOCK_INODE_MUTEX(inode);
 275
 276         for (seg = 0; seg < nr_segs; seg++) {
 277                 size_t iov_left = iov[seg].iov_len;
 278                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 279
 280                 if (rw == READ) {
 281                         if (file_offset >= inode->i_size)
 282                                 break;
 283                         if (file_offset + iov_left > inode->i_size)
 284                                 iov_left = inode->i_size - file_offset;
 285                 }
 286
 287                 while (iov_left > 0) {
 288                         struct page **pages;
 289                         int page_count;
 290                         ssize_t result;
 291
 292                         page_count = ll_get_user_pages(rw, user_addr,
 293                                                        min(size, iov_left),
 294                                                        &pages);
 295                         LASSERT(page_count != 0);
 296                         if (page_count > 0) {
 297                                 result = ll_direct_IO_26_seg(rw, inode,
 298                                                              file->f_mapping,
 299                                                              &oinfo, set,
 300                                                              min(size,iov_left),
 301                                                              file_offset, pages,
 302                                                              page_count);
 303                                 ll_free_user_pages(pages, page_count, rw==READ);
 304                         } else {
 305                                 result = 0;
 306                         }
 307                         if (page_count < 0 || result <= 0) {
 308                                 /* If we can't allocate a large enough buffer
 309                                  * for the request, shrink it to a smaller
 310                                  * PAGE_SIZE multiple and try again.
 311                                  * We should always be able to kmalloc for a
 312                                  * page worth of page pointers = 4MB on i386. */
 313                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 314                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 315                                            CFS_PAGE_SIZE) {
 316                                         size = ((((size / 2) - 1) |
 317                                                  ~CFS_PAGE_MASK) + 1) &
 318                                                 CFS_PAGE_MASK;
 319                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 320                                                (int)size);
 321                                         continue;
 322                                 }
 323
 324                                 if (tot_bytes <= 0)
 325                                         tot_bytes = page_count < 0 ? page_count : result;
 326                                 GOTO(out, tot_bytes);
 327                         }
 328
 329                         tot_bytes += result;
 330                         file_offset += result;
 331                         iov_left -= result;
 332                         user_addr += result;
 333                 }
 334         }
 335 out:
 336         if (rw == READ)
 337                 UNLOCK_INODE_MUTEX(inode);
 338
 339         if (tot_bytes > 0) {
 340                 int rc;
 341
 342                 rc = ptlrpc_set_wait(set);
 343                 if (rc) {
 344                         tot_bytes = rc;
 345                 } else if (rw == WRITE) {
 346                         lov_stripe_lock(lsm);
 347                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 348                         lov_stripe_unlock(lsm);
 349                 }
 350         }
 351
 352         capa_put(oinfo.oi_capa);
 353         ptlrpc_set_destroy(set);
 354         RETURN(tot_bytes);
 355 }
 356
 357 struct address_space_operations ll_aops = {
 358         .readpage       = ll_readpage,
 359 //        .readpages      = ll_readpages,
 360         .direct_IO      = ll_direct_IO_26,
 361         .writepage      = ll_writepage_26,
 362         .writepages     = generic_writepages,
 363         .set_page_dirty = ll_set_page_dirty,
 364         .sync_page      = NULL,
 365         .prepare_write  = ll_prepare_write,
 366         .commit_write   = ll_commit_write,
 367         .invalidatepage = ll_invalidatepage,
 368         .releasepage    = ll_releasepage,
 369         .bmap           = NULL
 370 };