lustre/liblustre/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Light block IO
   5  *
   6  *  Copyright (c) 2002, 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 #define DEBUG_SUBSYSTEM S_LLITE
  25
  26 #include <stdlib.h>
  27 #include <string.h>
  28 #include <assert.h>
  29 #include <time.h>
  30 #include <sys/types.h>
  31 #include <sys/queue.h>
  32 #include <fcntl.h>
  33
  34 #include <sysio.h>
  35 #include <fs.h>
  36 #include <mount.h>
  37 #include <inode.h>
  38 #include <file.h>
  39
  40 #undef LIST_HEAD
  41
  42 #include "llite_lib.h"
  43
  44 static int llu_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
  45 {
  46         struct llu_inode_info *lli = llu_i2info(inode);
  47         struct lov_stripe_md *lsm = lli->lli_smd;
  48         struct obd_export *exp = llu_i2obdexp(inode);
  49         struct {
  50                 char name[16];
  51                 struct ldlm_lock *lock;
  52                 struct lov_stripe_md *lsm;
  53         } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
  54         __u32 stripe, vallen = sizeof(stripe);
  55         int rc;
  56         ENTRY;
  57
  58         if (lsm->lsm_stripe_count == 1)
  59                 RETURN(0);
  60
  61         /* get our offset in the lov */
  62         rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
  63         if (rc != 0) {
  64                 CERROR("obd_get_info: rc = %d\n", rc);
  65                 LBUG();
  66         }
  67         LASSERT(stripe < lsm->lsm_stripe_count);
  68         RETURN(stripe);
  69 }
  70
  71 static int llu_extent_lock_callback(struct ldlm_lock *lock,
  72                                     struct ldlm_lock_desc *new, void *data,
  73                                     int flag)
  74 {
  75         struct lustre_handle lockh = { 0 };
  76         int rc;
  77         ENTRY;
  78
  79         if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
  80                 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
  81                 LBUG();
  82         }
  83
  84         switch (flag) {
  85         case LDLM_CB_BLOCKING:
  86                 ldlm_lock2handle(lock, &lockh);
  87                 rc = ldlm_cli_cancel(&lockh);
  88                 if (rc != ELDLM_OK)
  89                         CERROR("ldlm_cli_cancel failed: %d\n", rc);
  90                 break;
  91         case LDLM_CB_CANCELING: {
  92                 struct inode *inode;
  93                 struct llu_inode_info *lli;
  94                 struct lov_stripe_md *lsm;
  95                 __u32 stripe;
  96                 __u64 kms;
  97
  98                 /* This lock wasn't granted, don't try to evict pages */
  99                 if (lock->l_req_mode != lock->l_granted_mode)
 100                         RETURN(0);
 101
 102                 inode = llu_inode_from_lock(lock);
 103                 if (!inode)
 104                         RETURN(0);
 105                 lli= llu_i2info(inode);
 106                 if (!lli)
 107                         goto iput;
 108                 if (!lli->lli_smd)
 109                         goto iput;
 110                 lsm = lli->lli_smd;
 111
 112                 stripe = llu_lock_to_stripe_offset(inode, lock);
 113                 kms = ldlm_extent_shift_kms(lock,
 114                                             lsm->lsm_oinfo[stripe].loi_kms);
 115                 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
 116                         LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
 117                                    lsm->lsm_oinfo[stripe].loi_kms, kms);
 118                 lsm->lsm_oinfo[stripe].loi_kms = kms;
 119 iput:
 120                 I_RELE(inode);
 121                 break;
 122         }
 123         default:
 124                 LBUG();
 125         }
 126
 127         RETURN(0);
 128 }
 129
 130 static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp)
 131 {
 132         struct ptlrpc_request *req = reqp;
 133         struct inode *inode = llu_inode_from_lock(lock);
 134         struct llu_inode_info *lli;
 135         struct ost_lvb *lvb;
 136         int rc, size = sizeof(*lvb), stripe = 0;
 137         ENTRY;
 138
 139         if (inode == NULL)
 140                 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
 141         lli = llu_i2info(inode);
 142         if (lli == NULL)
 143                 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
 144         if (lli->lli_smd == NULL)
 145                 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
 146
 147         /* First, find out which stripe index this lock corresponds to. */
 148         if (lli->lli_smd->lsm_stripe_count > 1)
 149                 stripe = llu_lock_to_stripe_offset(inode, lock);
 150
 151         rc = lustre_pack_reply(req, 1, &size, NULL);
 152         if (rc) {
 153                 CERROR("lustre_pack_reply: %d\n", rc);
 154                 GOTO(iput, rc);
 155         }
 156
 157         lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
 158         lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
 159
 160         LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
 161                    lli->lli_st_size, stripe, lvb->lvb_size);
 162  iput:
 163         I_RELE(inode);
 164  out:
 165         /* These errors are normal races, so we don't want to fill the console
 166          * with messages by calling ptlrpc_error() */
 167         if (rc == -ELDLM_NO_LOCK_DATA)
 168                 lustre_pack_reply(req, 0, NULL, NULL);
 169
 170         req->rq_status = rc;
 171         return rc;
 172 }
 173
 174 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 175 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
 176
 177 /* NB: lov_merge_size will prefer locally cached writes if they extend the
 178  * file (because it prefers KMS over RSS when larger) */
 179 int llu_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
 180 {
 181         struct llu_inode_info *lli = llu_i2info(inode);
 182         struct llu_sb_info *sbi = llu_i2sbi(inode);
 183         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
 184         struct lustre_handle lockh = { 0 };
 185         int rc, flags = LDLM_FL_HAS_INTENT;
 186         ENTRY;
 187
 188         CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino);
 189
 190         rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
 191                          LCK_PR, &flags, llu_extent_lock_callback,
 192                          ldlm_completion_ast, llu_glimpse_callback, inode,
 193                          sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
 194         if (rc > 0)
 195                 RETURN(-EIO);
 196
 197         lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
 198         //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
 199
 200         CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size);
 201
 202         obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
 203
 204         RETURN(rc);
 205 }
 206
 207 int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
 208                     struct lov_stripe_md *lsm, int mode,
 209                     ldlm_policy_data_t *policy, struct lustre_handle *lockh,
 210                     int ast_flags)
 211 {
 212         struct llu_sb_info *sbi = llu_i2sbi(inode);
 213         struct llu_inode_info *lli = llu_i2info(inode);
 214         int rc;
 215         ENTRY;
 216
 217         LASSERT(lockh->cookie == 0);
 218
 219         /* XXX phil: can we do this?  won't it screw the file size up? */
 220         if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
 221             (sbi->ll_flags & LL_SBI_NOLCK))
 222                 RETURN(0);
 223
 224         CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
 225                lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end);
 226
 227         rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
 228                          &ast_flags, llu_extent_lock_callback,
 229                          ldlm_completion_ast, llu_glimpse_callback, inode,
 230                          sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
 231         if (rc > 0)
 232                 rc = -EIO;
 233
 234         if (policy->l_extent.start == 0 &&
 235             policy->l_extent.end == OBD_OBJECT_EOF)
 236                 lli->lli_st_size = lov_merge_size(lsm, 1);
 237
 238         //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
 239
 240         RETURN(rc);
 241 }
 242
 243 #if 0
 244 int llu_extent_lock_no_validate(struct ll_file_data *fd,
 245                                 struct inode *inode,
 246                                 struct lov_stripe_md *lsm,
 247                                 int mode,
 248                                 struct ldlm_extent *extent,
 249                                 struct lustre_handle *lockh,
 250                                 int ast_flags)
 251 {
 252         struct llu_sb_info *sbi = llu_i2sbi(inode);
 253         struct llu_inode_info *lli = llu_i2info(inode);
 254         int rc;
 255         ENTRY;
 256
 257         LASSERT(lockh->cookie == 0);
 258
 259         /* XXX phil: can we do this?  won't it screw the file size up? */
 260         if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
 261             (sbi->ll_flags & LL_SBI_NOLCK))
 262                 RETURN(0);
 263
 264         CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
 265                lli->lli_st_ino, extent->start, extent->end);
 266
 267         rc = obd_enqueue(sbi->ll_osc_exp, lsm, NULL, LDLM_EXTENT, extent,
 268                          sizeof(extent), mode, &ast_flags,
 269                          llu_extent_lock_callback, inode, lockh);
 270
 271         RETURN(rc);
 272 }
 273
 274 /*
 275  * this grabs a lock and manually implements behaviour that makes it look like
 276  * the OST is returning the file size with each lock acquisition.
 277  */
 278 int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
 279                     struct lov_stripe_md *lsm, int mode,
 280                     struct ldlm_extent *extent, struct lustre_handle *lockh,
 281                     int nonblock)
 282 {
 283         struct llu_inode_info *lli = llu_i2info(inode);
 284         struct obd_export *exp = llu_i2obdexp(inode);
 285         struct ldlm_extent size_lock;
 286         struct lustre_handle match_lockh = {0};
 287         int flags, rc, matched;
 288         int astflags = nonblock ? LDLM_FL_BLOCK_NOWAIT : 0;
 289         ENTRY;
 290
 291         rc = llu_extent_lock_no_validate(fd, inode, lsm, mode, extent,
 292                                          lockh, astflags);
 293         if (rc != ELDLM_OK)
 294                 RETURN(rc);
 295
 296         if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags))
 297                 RETURN(0);
 298
 299         rc = llu_inode_getattr(inode, lsm);
 300         if (rc) {
 301                 llu_extent_unlock(fd, inode, lsm, mode, lockh);
 302                 RETURN(rc);
 303         }
 304
 305         size_lock.start = lli->lli_st_size;
 306         size_lock.end = OBD_OBJECT_EOF;
 307
 308         /* XXX I bet we should be checking the lock ignore flags.. */
 309         /* FIXME use LDLM_FL_TEST_LOCK instead */
 310         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
 311         matched = obd_match(exp, lsm, LDLM_EXTENT, &size_lock,
 312                             sizeof(size_lock), LCK_PR, &flags, inode,
 313                             &match_lockh);
 314
 315         /* hey, alright, we hold a size lock that covers the size we
 316          * just found, its not going to change for a while.. */
 317         if (matched == 1) {
 318                 set_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags);
 319                 obd_cancel(exp, lsm, LCK_PR, &match_lockh);
 320         }
 321
 322         RETURN(0);
 323 }
 324 #endif
 325
 326 int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode,
 327                 struct lov_stripe_md *lsm, int mode,
 328                 struct lustre_handle *lockh)
 329 {
 330         struct llu_sb_info *sbi = llu_i2sbi(inode);
 331         int rc;
 332         ENTRY;
 333
 334         /* XXX phil: can we do this?  won't it screw the file size up? */
 335         if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
 336             (sbi->ll_flags & LL_SBI_NOLCK))
 337                 RETURN(0);
 338
 339         rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
 340
 341         RETURN(rc);
 342 }
 343
 344 #define LLAP_MAGIC 12346789
 345
 346 struct ll_async_page {
 347         int             llap_magic;
 348         void           *llap_cookie;
 349         int             llap_queued;
 350         struct page    *llap_page;
 351         struct inode   *llap_inode;
 352 };
 353
 354 static struct ll_async_page *llap_from_cookie(void *cookie)
 355 {
 356         struct ll_async_page *llap = cookie;
 357         if (llap->llap_magic != LLAP_MAGIC)
 358                 return ERR_PTR(-EINVAL);
 359         return llap;
 360 };
 361
 362 static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 363 {
 364         struct ll_async_page *llap;
 365         struct inode *inode;
 366         struct lov_stripe_md *lsm;
 367         obd_flag valid_flags;
 368         ENTRY;
 369
 370         llap = llap_from_cookie(data);
 371         if (IS_ERR(llap)) {
 372                 EXIT;
 373                 return;
 374         }
 375
 376         inode = llap->llap_inode;
 377         lsm = llu_i2info(inode)->lli_smd;
 378
 379         oa->o_id = lsm->lsm_object_id;
 380         oa->o_valid = OBD_MD_FLID;
 381         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 382         if (cmd == OBD_BRW_WRITE)
 383                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 384
 385         obdo_from_inode(oa, inode, valid_flags);
 386         EXIT;
 387 }
 388
 389 /* called for each page in a completed rpc.*/
 390 static void llu_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 391 {
 392         struct ll_async_page *llap;
 393         struct page *page;
 394
 395         llap = llap_from_cookie(data);
 396         if (IS_ERR(llap)) {
 397                 EXIT;
 398                 return;
 399         }
 400
 401         llap->llap_queued = 0;
 402         page = llap->llap_page;
 403
 404         if (rc != 0) {
 405                 if (cmd == OBD_BRW_WRITE)
 406                         CERROR("writeback error on page %p index %ld: %d\n",
 407                                page, page->index, rc);
 408         }
 409         EXIT;
 410 }
 411
 412 static struct obd_async_page_ops llu_async_page_ops = {
 413         .ap_make_ready =        NULL,
 414         .ap_refresh_count =     NULL,
 415         .ap_fill_obdo =         llu_ap_fill_obdo,
 416         .ap_completion =        llu_ap_completion,
 417 };
 418
 419 static
 420 struct llu_sysio_cookie* get_sysio_cookie(struct inode *inode, int maxpages)
 421 {
 422         struct llu_sysio_cookie *cookie;
 423         int rc;
 424
 425         OBD_ALLOC(cookie, LLU_SYSIO_COOKIE_SIZE(maxpages));
 426         if (cookie == NULL)
 427                 goto out;
 428
 429         I_REF(inode);
 430         cookie->lsc_inode = inode;
 431         cookie->lsc_maxpages = maxpages;
 432         cookie->lsc_llap = (struct ll_async_page *)(cookie + 1);
 433         cookie->lsc_pages = (struct page *) (cookie->lsc_llap + maxpages);
 434
 435         rc = oig_init(&cookie->lsc_oig);
 436         if (rc) {
 437                 OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(maxpages));
 438                 cookie = NULL;
 439         }
 440
 441 out:
 442         return cookie;
 443 }
 444
 445 static
 446 void put_sysio_cookie(struct llu_sysio_cookie *cookie)
 447 {
 448         struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd;
 449         struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
 450         struct ll_async_page *llap = cookie->lsc_llap;
 451 #ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE
 452         struct page *pages = cookie->lsc_pages;
 453 #endif
 454         int i;
 455
 456         for (i = 0; i< cookie->lsc_maxpages; i++) {
 457                 if (llap[i].llap_cookie)
 458                         obd_teardown_async_page(exp, lsm, NULL,
 459                                                 llap[i].llap_cookie);
 460 #ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE
 461                 if (pages[i]._managed) {
 462                         free(pages[i].addr);
 463                         pages[i]._managed = 0;
 464                 }
 465 #endif
 466         }
 467
 468         I_RELE(cookie->lsc_inode);
 469
 470         oig_release(cookie->lsc_oig);
 471         OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(cookie->lsc_maxpages));
 472 }
 473
 474 #ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE
 475 /* Note: these code should be removed finally, don't need
 476  * more cleanup
 477  */
 478 static
 479 int prepare_unaligned_write(struct llu_sysio_cookie *cookie)
 480 {
 481         struct inode *inode = cookie->lsc_inode;
 482         struct llu_inode_info *lli = llu_i2info(inode);
 483         struct lov_stripe_md *lsm = lli->lli_smd;
 484         struct obdo oa;
 485         struct page *pages = cookie->lsc_pages;
 486         int i, pgidx[2] = {0, cookie->lsc_npages-1};
 487         int rc;
 488         ENTRY;
 489
 490         for (i = 0; i < 2; i++) {
 491                 struct page *oldpage = &pages[pgidx[i]];
 492                 struct page newpage;
 493                 struct brw_page pg;
 494                 char *newbuf;
 495
 496                 if (i == 0 && pgidx[0] == pgidx[1])
 497                         continue;
 498
 499                 LASSERT(oldpage->_offset + oldpage->_count <= PAGE_CACHE_SIZE);
 500
 501                 if (oldpage->_count == PAGE_CACHE_SIZE)
 502                         continue;
 503
 504                 if (oldpage->index << PAGE_CACHE_SHIFT >=
 505                     lli->lli_st_size)
 506                         continue;
 507
 508                 newbuf = malloc(PAGE_CACHE_SIZE);
 509                 if (!newbuf)
 510                         return -ENOMEM;
 511
 512                 newpage.index = oldpage->index;
 513                 newpage.addr = newbuf;
 514
 515                 pg.pg = &newpage;
 516                 pg.off = ((obd_off)newpage.index << PAGE_CACHE_SHIFT);
 517                 if (pg.off + PAGE_CACHE_SIZE > lli->lli_st_size)
 518                         pg.count = lli->lli_st_size % PAGE_CACHE_SIZE;
 519                 else
 520                         pg.count = PAGE_CACHE_SIZE;
 521                 pg.flag = 0;
 522
 523                 oa.o_id = lsm->lsm_object_id;
 524                 oa.o_mode = lli->lli_st_mode;
 525                 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 526
 527                 /* issue read */
 528                 rc = obd_brw(OBD_BRW_READ, llu_i2obdexp(inode), &oa, lsm, 1, &pg, NULL);
 529                 if (rc) {
 530                         free(newbuf);
 531                         RETURN(rc);
 532                 }
 533
 534                 /* copy page content, and reset page params */
 535                 memcpy(newbuf + oldpage->_offset,
 536                        (char*)oldpage->addr + oldpage->_offset,
 537                        oldpage->_count);
 538
 539                 oldpage->addr = newbuf;
 540                 if ((((obd_off)oldpage->index << PAGE_CACHE_SHIFT) +
 541                     oldpage->_offset + oldpage->_count) > lli->lli_st_size)
 542                         oldpage->_count += oldpage->_offset;
 543                 else
 544                         oldpage->_count = PAGE_CACHE_SIZE;
 545                 oldpage->_offset = 0;
 546                 oldpage->_managed = 1;
 547         }
 548
 549         RETURN(0);
 550 }
 551 #endif
 552
 553 static
 554 int llu_prep_async_io(struct llu_sysio_cookie *cookie, int cmd,
 555                       char *buf, loff_t pos, size_t count)
 556 {
 557         struct llu_inode_info *lli = llu_i2info(cookie->lsc_inode);
 558         struct lov_stripe_md *lsm = lli->lli_smd;
 559         struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
 560         struct page *pages = cookie->lsc_pages;
 561         struct ll_async_page *llap = cookie->lsc_llap;
 562         int i, rc, npages = 0;
 563         ENTRY;
 564
 565         if (!exp)
 566                 RETURN(-EINVAL);
 567
 568         /* prepare the pages array */
 569         do {
 570                 unsigned long index, offset, bytes;
 571
 572                 offset = (pos & ~PAGE_CACHE_MASK);
 573                 index = pos >> PAGE_CACHE_SHIFT;
 574                 bytes = PAGE_CACHE_SIZE - offset;
 575                 if (bytes > count)
 576                         bytes = count;
 577
 578                 /* prevent read beyond file range */
 579                 if ((cmd == OBD_BRW_READ) &&
 580                     (pos + bytes) >= lli->lli_st_size) {
 581                         if (pos >= lli->lli_st_size)
 582                                 break;
 583                         bytes = lli->lli_st_size - pos;
 584                 }
 585
 586                 /* prepare page for this index */
 587                 pages[npages].index = index;
 588                 pages[npages].addr = buf - offset;
 589
 590                 pages[npages]._offset = offset;
 591                 pages[npages]._count = bytes;
 592
 593                 npages++;
 594                 count -= bytes;
 595                 pos += bytes;
 596                 buf += bytes;
 597
 598                 cookie->lsc_rwcount += bytes;
 599         } while (count);
 600
 601         cookie->lsc_npages = npages;
 602
 603 #ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE
 604         if (cmd == OBD_BRW_WRITE) {
 605                 rc = prepare_unaligned_write(cookie);
 606                 if (rc)
 607                         RETURN(rc);
 608         }
 609 #endif
 610
 611         for (i = 0; i < npages; i++) {
 612                 llap[i].llap_magic = LLAP_MAGIC;
 613                 rc = obd_prep_async_page(exp, lsm, NULL, &pages[i],
 614                                          (obd_off)pages[i].index << PAGE_SHIFT,
 615                                          &llu_async_page_ops,
 616                                          &llap[i], &llap[i].llap_cookie);
 617                 if (rc) {
 618                         llap[i].llap_cookie = NULL;
 619                         RETURN(rc);
 620                 }
 621                 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n",
 622                        &llap[i], &pages[i], llap[i].llap_cookie,
 623                        (obd_off)pages[i].index << PAGE_SHIFT);
 624                 pages[i].private = (unsigned long)&llap[i];
 625                 llap[i].llap_page = &pages[i];
 626                 llap[i].llap_inode = cookie->lsc_inode;
 627
 628                 rc = obd_queue_group_io(exp, lsm, NULL, cookie->lsc_oig,
 629                                         llap[i].llap_cookie, cmd,
 630                                         pages[i]._offset, pages[i]._count, 0,
 631                                         ASYNC_READY | ASYNC_URGENT |
 632                                         ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 633                 if (rc)
 634                         RETURN(rc);
 635
 636                 llap[i].llap_queued = 1;
 637         }
 638
 639         RETURN(0);
 640 }
 641
 642 static
 643 int llu_start_async_io(struct llu_sysio_cookie *cookie)
 644 {
 645         struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd;
 646         struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
 647
 648         return obd_trigger_group_io(exp, lsm, NULL, cookie->lsc_oig);
 649 }
 650
 651 /*
 652  * read/write a continuous buffer for an inode (zero-copy)
 653  */
 654 struct llu_sysio_cookie*
 655 llu_rw(int cmd, struct inode *inode, char *buf, size_t count, loff_t pos)
 656 {
 657         struct llu_sysio_cookie *cookie;
 658         int max_pages, rc;
 659         ENTRY;
 660
 661         max_pages = (count >> PAGE_SHIFT) + 2;
 662
 663         cookie = get_sysio_cookie(inode, max_pages);
 664         if (!cookie)
 665                 RETURN(ERR_PTR(-ENOMEM));
 666
 667         rc = llu_prep_async_io(cookie, cmd, buf, pos, count);
 668         if (rc)
 669                 GOTO(out_cleanup, rc);
 670
 671         rc = llu_start_async_io(cookie);
 672         if (rc)
 673                 GOTO(out_cleanup, rc);
 674
 675 /*
 676         rc = oig_wait(&oig);
 677         if (rc) {
 678                 CERROR("file i/o error!\n");
 679                 rw_count = rc;
 680         }
 681 */
 682         RETURN(cookie);
 683
 684 out_cleanup:
 685         put_sysio_cookie(cookie);
 686         RETURN(ERR_PTR(rc));
 687 }
 688
 689 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 690                       obd_off size);
 691
 692 struct llu_sysio_callback_args*
 693 llu_file_write(struct inode *inode, const struct iovec *iovec,
 694                size_t iovlen, loff_t pos)
 695 {
 696         struct llu_inode_info *lli = llu_i2info(inode);
 697         struct ll_file_data *fd = lli->lli_file_data;
 698         struct lustre_handle lockh = {0};
 699         struct lov_stripe_md *lsm = lli->lli_smd;
 700         struct obd_export *exp = NULL;
 701         ldlm_policy_data_t policy;
 702         struct llu_sysio_callback_args *lsca;
 703         struct llu_sysio_cookie *cookie;
 704         int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
 705                        LDLM_FL_BLOCK_NOWAIT : 0;
 706         ldlm_error_t err;
 707         int iovidx;
 708         ENTRY;
 709
 710         /* XXX consider other types later */
 711         if (!S_ISREG(lli->lli_st_mode))
 712                 LBUG();
 713
 714         LASSERT(iovlen <= MAX_IOVEC);
 715
 716         exp = llu_i2obdexp(inode);
 717         if (exp == NULL)
 718                 RETURN(ERR_PTR(-EINVAL));
 719
 720         OBD_ALLOC(lsca, sizeof(*lsca));
 721         if (!lsca)
 722                 RETURN(ERR_PTR(-ENOMEM));
 723
 724         /* FIXME optimize the following extent locking */
 725         for (iovidx = 0; iovidx < iovlen; iovidx++) {
 726                 char *buf = (char*)iovec[iovidx].iov_base;
 727                 size_t count = iovec[iovidx].iov_len;
 728
 729                 if (count == 0)
 730                         continue;
 731
 732                 if (pos + count > lli->lli_maxbytes)
 733                         GOTO(err_out, err = -ERANGE);
 734
 735                 /* FIXME libsysio haven't handle O_APPEND?? */
 736                 policy.l_extent.start = pos;
 737                 policy.l_extent.end = pos + count - 1;
 738
 739                 err = llu_extent_lock(fd, inode, lsm, LCK_PW, &policy,
 740                                       &lockh, astflag);
 741                 if (err != ELDLM_OK)
 742                         GOTO(err_out, err = -ENOLCK);
 743
 744                 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
 745                        lli->lli_st_ino, count, pos);
 746
 747                 cookie = llu_rw(OBD_BRW_WRITE, inode, buf, count, pos);
 748                 if (!IS_ERR(cookie)) {
 749                         /* save cookie */
 750                         lsca->cookies[lsca->ncookies++] = cookie;
 751                         pos += count;
 752                         lov_increase_kms(exp, lsm, pos);
 753                         /* file size grow */
 754                         if (pos > lli->lli_st_size)
 755                                 lli->lli_st_size = pos;
 756                 } else {
 757                         llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
 758                         GOTO(err_out, err = PTR_ERR(cookie));
 759                 }
 760
 761                 /* XXX errors? */
 762                 err = llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
 763                 if (err)
 764                         CERROR("extent unlock error %d\n", err);
 765         }
 766
 767         RETURN(lsca);
 768
 769 err_out:
 770         /* teardown all async stuff */
 771         while (lsca->ncookies--) {
 772                 put_sysio_cookie(lsca->cookies[lsca->ncookies]);
 773         }
 774         OBD_FREE(lsca, sizeof(*lsca));
 775
 776         RETURN(ERR_PTR(err));
 777 }
 778
 779 #if 0
 780 static void llu_update_atime(struct inode *inode)
 781 {
 782         struct llu_inode_info *lli = llu_i2info(inode);
 783
 784 #ifdef USE_ATIME
 785         struct iattr attr;
 786
 787         attr.ia_atime = LTIME_S(CURRENT_TIME);
 788         attr.ia_valid = ATTR_ATIME;
 789
 790         if (lli->lli_st_atime == attr.ia_atime) return;
 791         if (IS_RDONLY(inode)) return;
 792         if (IS_NOATIME(inode)) return;
 793
 794         /* ll_inode_setattr() sets inode->i_atime from attr.ia_atime */
 795         llu_inode_setattr(inode, &attr, 0);
 796 #else
 797         /* update atime, but don't explicitly write it out just this change */
 798         inode->i_atime = CURRENT_TIME;
 799 #endif
 800 }
 801 #endif
 802
 803 struct llu_sysio_callback_args*
 804 llu_file_read(struct inode *inode, const struct iovec *iovec,
 805               size_t iovlen, loff_t pos)
 806 {
 807         struct llu_inode_info *lli = llu_i2info(inode);
 808         struct ll_file_data *fd = lli->lli_file_data;
 809         struct lov_stripe_md *lsm = lli->lli_smd;
 810         struct lustre_handle lockh = { 0 };
 811         ldlm_policy_data_t policy;
 812         struct llu_sysio_callback_args *lsca;
 813         struct llu_sysio_cookie *cookie;
 814         int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
 815                        LDLM_FL_BLOCK_NOWAIT : 0;
 816         __u64 kms;
 817         int iovidx;
 818
 819         ldlm_error_t err;
 820         ENTRY;
 821
 822         OBD_ALLOC(lsca, sizeof(*lsca));
 823         if (!lsca)
 824                 RETURN(ERR_PTR(-ENOMEM));
 825
 826         for (iovidx = 0; iovidx < iovlen; iovidx++) {
 827                 char *buf = iovec[iovidx].iov_base;
 828                 size_t count = iovec[iovidx].iov_len;
 829
 830                 /* "If nbyte is 0, read() will return 0 and have no other results."
 831                  *                      -- Single Unix Spec */
 832                 if (count == 0)
 833                         continue;
 834
 835                 policy.l_extent.start = pos;
 836                 policy.l_extent.end = pos + count - 1;
 837
 838                 err = llu_extent_lock(fd, inode, lsm, LCK_PR, &policy,
 839                                       &lockh, astflag);
 840                 if (err != ELDLM_OK)
 841                         GOTO(err_out, err = -ENOLCK);
 842
 843                 kms = lov_merge_size(lsm, 1);
 844                 if (policy.l_extent.end > kms) {
 845                         /* A glimpse is necessary to determine whether we
 846                          * return a short read or some zeroes at the end of
 847                          * the buffer */
 848                         struct ost_lvb lvb;
 849                         if (llu_glimpse_size(inode, &lvb)) {
 850                                 llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
 851                                 GOTO(err_out, err = -ENOLCK);
 852                         }
 853                         lli->lli_st_size = lvb.lvb_size;
 854                 } else {
 855                         lli->lli_st_size = kms;
 856                 }
 857
 858                 CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld, "
 859                        "i_size "LPU64"\n", lli->lli_st_ino, count, pos,
 860                        lli->lli_st_size);
 861
 862                 if (pos >= lli->lli_st_size) {
 863                         llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
 864                         break;
 865                 }
 866
 867                 cookie = llu_rw(OBD_BRW_READ, inode, buf, count, pos);
 868                 if (!IS_ERR(cookie)) {
 869                         /* save cookie */
 870                         lsca->cookies[lsca->ncookies++] = cookie;
 871                         pos += count;
 872                 } else {
 873                         llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
 874                         GOTO(err_out, err = PTR_ERR(cookie));
 875                 }
 876
 877                 /* XXX errors? */
 878                 err = llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
 879                 if (err)
 880                         CERROR("extent_unlock fail: %d\n", err);
 881         }
 882 #if 0
 883         if (readed > 0)
 884                 llu_update_atime(inode);
 885 #endif
 886         RETURN(lsca);
 887
 888 err_out:
 889         /* teardown all async stuff */
 890         while (lsca->ncookies--) {
 891                 put_sysio_cookie(lsca->cookies[lsca->ncookies]);
 892         }
 893         OBD_FREE(lsca, sizeof(*lsca));
 894
 895         RETURN(ERR_PTR(err));
 896 }
 897
 898 int llu_iop_iodone(struct ioctx *ioctxp)
 899 {
 900         struct llu_sysio_callback_args *lsca = ioctxp->ioctx_private;
 901         struct llu_sysio_cookie *cookie;
 902         int i, err = 0, rc = 0;
 903         ENTRY;
 904
 905         /* write/read(fd, buf, 0) */
 906         if (!lsca) {
 907                 ioctxp->ioctx_cc = 0;
 908                 RETURN(1);
 909         }
 910
 911         LASSERT(!IS_ERR(lsca));
 912
 913         for (i = 0; i < lsca->ncookies; i++) {
 914                 cookie = lsca->cookies[i];
 915                 if (cookie) {
 916                         err = oig_wait(cookie->lsc_oig);
 917                         if (err && !rc)
 918                                 rc = err;
 919                         if (!rc)
 920                                 ioctxp->ioctx_cc += cookie->lsc_rwcount;
 921                         put_sysio_cookie(cookie);
 922                 }
 923         }
 924
 925         if (rc) {
 926                 LASSERT(rc < 0);
 927                 ioctxp->ioctx_cc = -1;
 928                 ioctxp->ioctx_errno = -rc;
 929         }
 930
 931         OBD_FREE(lsca, sizeof(*lsca));
 932         ioctxp->ioctx_private = NULL;
 933
 934         RETURN(1);
 935 }