1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Light block IO
6 * Copyright (c) 2002-2004 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
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.
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.
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.
24 #define DEBUG_SUBSYSTEM S_LLITE
30 #include <sys/types.h>
31 #include <sys/queue.h>
47 #include "llite_lib.h"
51 struct obd_io_group *lig_oig;
52 struct inode *lig_inode;
56 struct ll_async_page *lig_llap;
57 struct page *lig_pages;
60 #define LLU_IO_GROUP_SIZE(x) \
61 (sizeof(struct llu_io_group) + \
62 sizeof(struct ll_async_page) * (x) + \
63 sizeof(struct page) * (x))
67 struct inode *lis_inode;
71 struct llu_io_group *lis_groups[0];
73 #define LLU_IO_SESSION_SIZE(x) \
74 (sizeof(struct llu_io_session) + (x) * 2 * sizeof(void *))
77 typedef ssize_t llu_file_piov_t(const struct iovec *iovec, int iovlen,
78 _SYSIO_OFF_T pos, ssize_t len,
81 static int llu_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
83 struct llu_inode_info *lli = llu_i2info(inode);
84 struct lov_stripe_md *lsm = lli->lli_smd;
85 struct obd_export *exp = llu_i2dtexp(inode);
88 struct ldlm_lock *lock;
89 struct lov_stripe_md *lsm;
90 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
91 __u32 stripe, vallen = sizeof(stripe);
95 if (lsm->lsm_stripe_count == 1)
98 /* get our offset in the lov */
99 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
101 CERROR("obd_get_info: rc = %d\n", rc);
104 LASSERT(stripe < lsm->lsm_stripe_count);
108 static int llu_extent_lock_callback(struct ldlm_lock *lock,
109 struct ldlm_lock_desc *new, void *data,
112 struct lustre_handle lockh = { 0 };
116 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
117 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
122 case LDLM_CB_BLOCKING:
123 ldlm_lock2handle(lock, &lockh);
124 rc = ldlm_cli_cancel(&lockh);
126 CERROR("ldlm_cli_cancel failed: %d\n", rc);
128 case LDLM_CB_CANCELING: {
130 struct llu_inode_info *lli;
131 struct lov_stripe_md *lsm;
135 /* This lock wasn't granted, don't try to evict pages */
136 if (lock->l_req_mode != lock->l_granted_mode)
139 inode = llu_inode_from_lock(lock);
142 lli= llu_i2info(inode);
149 stripe = llu_lock_to_stripe_offset(inode, lock);
150 l_lock(&lock->l_resource->lr_namespace->ns_lock);
151 kms = ldlm_extent_shift_kms(lock,
152 lsm->lsm_oinfo[stripe].loi_kms);
153 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
154 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
155 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
156 lsm->lsm_oinfo[stripe].loi_kms, kms);
157 lsm->lsm_oinfo[stripe].loi_kms = kms;
169 static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp)
171 struct ptlrpc_request *req = reqp;
172 struct inode *inode = llu_inode_from_lock(lock);
173 struct llu_inode_info *lli;
175 int rc, size = sizeof(*lvb), stripe = 0;
179 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
180 lli = llu_i2info(inode);
182 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
183 if (lli->lli_smd == NULL)
184 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
186 /* First, find out which stripe index this lock corresponds to. */
187 if (lli->lli_smd->lsm_stripe_count > 1)
188 stripe = llu_lock_to_stripe_offset(inode, lock);
190 rc = lustre_pack_reply(req, 1, &size, NULL);
192 CERROR("lustre_pack_reply: %d\n", rc);
196 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
197 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
199 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
200 lli->lli_st_size, stripe, lvb->lvb_size);
204 /* These errors are normal races, so we don't want to fill the console
205 * with messages by calling ptlrpc_error() */
206 if (rc == -ELDLM_NO_LOCK_DATA)
207 lustre_pack_reply(req, 0, NULL, NULL);
213 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms_only);
214 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
215 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
217 /* NB: lov_merge_size will prefer locally cached writes if they extend the
218 * file (because it prefers KMS over RSS when larger) */
219 int llu_glimpse_size(struct inode *inode)
221 struct llu_inode_info *lli = llu_i2info(inode);
222 struct llu_sb_info *sbi = llu_i2sbi(inode);
223 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
224 struct lustre_handle lockh = { 0 };
225 int rc, flags = LDLM_FL_HAS_INTENT;
228 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino);
230 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
231 LCK_PR, &flags, llu_extent_lock_callback,
232 ldlm_completion_ast, llu_glimpse_callback, inode,
233 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
235 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
236 RETURN(rc > 0 ? -EIO : rc);
239 lli->lli_st_size = lov_merge_size(lli->lli_smd, 0);
240 lli->lli_st_blocks = lov_merge_blocks(lli->lli_smd);
241 //lli->lli_st_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
243 CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %lu\n",
244 lli->lli_st_size, lli->lli_st_blocks);
246 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
251 int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
252 struct lov_stripe_md *lsm, int mode,
253 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
256 struct llu_sb_info *sbi = llu_i2sbi(inode);
257 struct llu_inode_info *lli = llu_i2info(inode);
261 LASSERT(lockh->cookie == 0);
263 /* XXX phil: can we do this? won't it screw the file size up? */
264 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
265 (sbi->ll_flags & LL_SBI_NOLCK))
268 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
269 lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end);
271 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
272 &ast_flags, llu_extent_lock_callback,
273 ldlm_completion_ast, llu_glimpse_callback, inode,
274 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
278 if (policy->l_extent.start == 0 &&
279 policy->l_extent.end == OBD_OBJECT_EOF)
280 lli->lli_st_size = lov_merge_size(lsm, 1);
282 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
287 int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode,
288 struct lov_stripe_md *lsm, int mode,
289 struct lustre_handle *lockh)
291 struct llu_sb_info *sbi = llu_i2sbi(inode);
295 /* XXX phil: can we do this? won't it screw the file size up? */
296 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
297 (sbi->ll_flags & LL_SBI_NOLCK))
300 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
305 #define LLAP_MAGIC 12346789
307 struct ll_async_page {
311 struct page *llap_page;
312 struct inode *llap_inode;
315 static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
317 struct ll_async_page *llap;
319 struct lov_stripe_md *lsm;
320 obd_valid valid_flags;
323 llap = LLAP_FROM_COOKIE(data);
324 inode = llap->llap_inode;
325 lsm = llu_i2info(inode)->lli_smd;
327 oa->o_id = lsm->lsm_object_id;
328 oa->o_valid = OBD_MD_FLID;
329 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
330 if (cmd == OBD_BRW_WRITE)
331 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
333 obdo_from_inode(oa, inode, valid_flags);
337 /* called for each page in a completed rpc.*/
338 static void llu_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
340 struct ll_async_page *llap;
343 llap = LLAP_FROM_COOKIE(data);
344 llap->llap_queued = 0;
345 page = llap->llap_page;
348 if (cmd == OBD_BRW_WRITE)
349 CERROR("writeback error on page %p index %ld: %d\n",
350 page, page->index, rc);
355 static struct obd_async_page_ops llu_async_page_ops = {
356 .ap_make_ready = NULL,
357 .ap_refresh_count = NULL,
358 .ap_fill_obdo = llu_ap_fill_obdo,
359 .ap_completion = llu_ap_completion,
362 static int llu_queue_pio(int cmd, struct llu_io_group *group,
363 char *buf, size_t count, loff_t pos)
365 struct llu_inode_info *lli = llu_i2info(group->lig_inode);
366 struct lov_stripe_md *lsm = lli->lli_smd;
367 struct obd_export *exp = llu_i2dtexp(group->lig_inode);
368 struct page *pages = &group->lig_pages[group->lig_npages];
369 struct ll_async_page *llap = &group->lig_llap[group->lig_npages];
370 int i, rc, npages = 0, ret_bytes = 0;
376 /* prepare the pages array */
378 unsigned long index, offset, bytes;
380 offset = (pos & ~PAGE_CACHE_MASK);
381 index = pos >> PAGE_CACHE_SHIFT;
382 bytes = PAGE_CACHE_SIZE - offset;
386 /* prevent read beyond file range */
387 if ((cmd == OBD_BRW_READ) &&
388 (pos + bytes) >= lli->lli_st_size) {
389 if (pos >= lli->lli_st_size)
391 bytes = lli->lli_st_size - pos;
394 /* prepare page for this index */
395 pages[npages].index = index;
396 pages[npages].addr = buf - offset;
398 pages[npages]._offset = offset;
399 pages[npages]._count = bytes;
406 group->lig_rwcount += bytes;
410 group->lig_npages += npages;
412 for (i = 0; i < npages; i++) {
413 llap[i].llap_magic = LLAP_MAGIC;
414 rc = obd_prep_async_page(exp, lsm, NULL, &pages[i],
415 (obd_off)pages[i].index << PAGE_SHIFT,
417 &llap[i], &llap[i].llap_cookie);
420 llap[i].llap_cookie = NULL;
423 CDEBUG(D_CACHE, "llap %p page %p group %p obj off "LPU64"\n",
424 &llap[i], &pages[i], llap[i].llap_cookie,
425 (obd_off)pages[i].index << PAGE_SHIFT);
426 pages[i].private = (unsigned long)&llap[i];
427 llap[i].llap_page = &pages[i];
428 llap[i].llap_inode = group->lig_inode;
430 rc = obd_queue_group_io(exp, lsm, NULL, group->lig_oig,
431 llap[i].llap_cookie, cmd,
432 pages[i]._offset, pages[i]._count, 0,
433 ASYNC_READY | ASYNC_URGENT |
434 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
440 llap[i].llap_queued = 1;
447 struct llu_io_group * get_io_group(struct inode *inode, int maxpages)
449 struct llu_io_group *group;
452 OBD_ALLOC(group, LLU_IO_GROUP_SIZE(maxpages));
454 return ERR_PTR(-ENOMEM);
457 group->lig_inode = inode;
458 group->lig_maxpages = maxpages;
459 group->lig_llap = (struct ll_async_page *)(group + 1);
460 group->lig_pages = (struct page *) (group->lig_llap + maxpages);
462 rc = oig_init(&group->lig_oig);
464 OBD_FREE(group, LLU_IO_GROUP_SIZE(maxpages));
471 static int max_io_pages(ssize_t len, int iovlen)
473 return (((len + PAGE_SIZE -1) / PAGE_SIZE) + 2 + iovlen - 1);
477 void put_io_group(struct llu_io_group *group)
479 struct lov_stripe_md *lsm = llu_i2info(group->lig_inode)->lli_smd;
480 struct obd_export *exp = llu_i2dtexp(group->lig_inode);
481 struct ll_async_page *llap = group->lig_llap;
484 for (i = 0; i< group->lig_npages; i++) {
485 if (llap[i].llap_cookie)
486 obd_teardown_async_page(exp, lsm, NULL,
487 llap[i].llap_cookie);
490 I_RELE(group->lig_inode);
492 oig_release(group->lig_oig);
493 OBD_FREE(group, LLU_IO_GROUP_SIZE(group->lig_maxpages));
497 ssize_t llu_file_prwv(const struct iovec *iovec, int iovlen,
498 _SYSIO_OFF_T pos, ssize_t len,
501 struct llu_io_session *session = (struct llu_io_session *) private;
502 struct inode *inode = session->lis_inode;
503 struct llu_inode_info *lli = llu_i2info(inode);
504 struct ll_file_data *fd = lli->lli_file_data;
505 struct lustre_handle lockh = {0};
506 struct lov_stripe_md *lsm = lli->lli_smd;
507 struct obd_export *exp = NULL;
508 ldlm_policy_data_t policy;
509 struct llu_io_group *iogroup;
510 int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
511 LDLM_FL_BLOCK_NOWAIT : 0;
513 int err, is_read, lock_mode, iovidx, ret;
516 /* in a large iov read/write we'll be repeatedly called.
517 * so give a chance to answer cancel ast here
519 liblustre_wait_event(0);
521 exp = llu_i2dtexp(inode);
525 if (len == 0 || iovlen == 0)
528 if (pos + len > lli->lli_maxbytes)
531 iogroup = get_io_group(inode, max_io_pages(len, iovlen));
533 RETURN(PTR_ERR(iogroup));
535 is_read = session->lis_cmd == OBD_BRW_READ;
536 lock_mode = is_read ? LCK_PR : LCK_PW;
538 if (!is_read && (lli->lli_open_flags & O_APPEND)) {
539 policy.l_extent.start = 0;
540 policy.l_extent.end = OBD_OBJECT_EOF;
542 policy.l_extent.start = pos;
543 policy.l_extent.end = pos + len - 1;
546 err = llu_extent_lock(fd, inode, lsm, lock_mode, &policy,
552 kms = lov_merge_size(lsm, 1);
553 if (policy.l_extent.end > kms) {
554 /* A glimpse is necessary to determine whether we
555 * return a short read or some zeroes at the end of
557 if ((err = llu_glimpse_size(inode))) {
558 llu_extent_unlock(fd, inode, lsm,
563 lli->lli_st_size = kms;
566 if (lli->lli_open_flags & O_APPEND)
567 pos = lli->lli_st_size;
570 for (iovidx = 0; iovidx < iovlen; iovidx++) {
571 char *buf = (char *) iovec[iovidx].iov_base;
572 size_t count = iovec[iovidx].iov_len;
578 if (IS_BAD_PTR(buf) || IS_BAD_PTR(buf + count)) {
579 llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
580 GOTO(err_put, err = -EFAULT);
584 if (pos >= lli->lli_st_size)
587 if (pos >= lli->lli_maxbytes) {
588 llu_extent_unlock(fd, inode, lsm, lock_mode,
590 GOTO(err_put, err = -EFBIG);
592 if (pos + count >= lli->lli_maxbytes)
593 count = lli->lli_maxbytes - pos;
596 ret = llu_queue_pio(session->lis_cmd, iogroup, buf, count, pos);
598 llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
599 GOTO(err_put, err = ret);
603 LASSERT(ret == count);
604 obd_adjust_kms(exp, lsm, pos, 0);
605 /* file size grow immediately */
606 if (pos > lli->lli_st_size)
607 lli->lli_st_size = pos;
614 LASSERT(len == 0 || is_read); /* libsysio should guarantee this */
616 err = llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
618 CERROR("extent unlock error %d\n", err);
620 err = obd_trigger_group_io(exp, lsm, NULL, iogroup->lig_oig);
624 session->lis_groups[session->lis_ngroups++] = iogroup;
627 put_io_group(iogroup);
628 RETURN((ssize_t)err);
632 struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd)
634 struct llu_io_session *session;
636 OBD_ALLOC(session, LLU_IO_SESSION_SIZE(ngroups));
641 session->lis_inode = ino;
642 session->lis_max_groups = ngroups;
643 session->lis_cmd = cmd;
647 static void put_io_session(struct llu_io_session *session)
651 for (i = 0; i < session->lis_ngroups; i++) {
652 if (session->lis_groups[i]) {
653 put_io_group(session->lis_groups[i]);
654 session->lis_groups[i] = NULL;
658 I_RELE(session->lis_inode);
659 OBD_FREE(session, LLU_IO_SESSION_SIZE(session->lis_max_groups));
662 static int llu_file_rwx(struct inode *ino,
666 struct llu_io_session *session;
668 int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE;
671 LASSERT(ioctx->ioctx_xtvlen >= 0);
672 LASSERT(ioctx->ioctx_iovlen >= 0);
674 liblustre_wait_event(0);
676 if (!ioctx->ioctx_xtvlen)
679 /* XXX consider other types later */
680 LASSERT(S_ISREG(llu_i2info(ino)->lli_st_mode));
682 session = get_io_session(ino, ioctx->ioctx_xtvlen * 2, cmd);
686 cc = _sysio_enumerate_extents(ioctx->ioctx_xtv, ioctx->ioctx_xtvlen,
687 ioctx->ioctx_iov, ioctx->ioctx_iovlen,
688 llu_file_prwv, session);
691 LASSERT(!ioctx->ioctx_cc);
692 ioctx->ioctx_private = session;
695 put_io_session(session);
700 int llu_iop_read(struct inode *ino,
703 return llu_file_rwx(ino, ioctx, 1);
706 int llu_iop_write(struct inode *ino,
712 memset(&iattr, 0, sizeof(iattr));
713 iattr.ia_mtime = iattr.ia_atime = CURRENT_TIME;
714 iattr.ia_valid = ATTR_MTIME | ATTR_ATIME | ATTR_RAW;
716 liblustre_wait_event(0);
717 rc = llu_setattr_raw(ino, &iattr);
719 CERROR("failed to set mtime/atime during write: %d", rc);
720 /* XXX should continue or return error? */
723 return llu_file_rwx(ino, ioctx, 0);
726 int llu_iop_iodone(struct ioctx *ioctx)
728 struct llu_io_session *session;
729 struct llu_io_group *group;
730 int i, err = 0, rc = 0;
733 liblustre_wait_event(0);
735 session = (struct llu_io_session *) ioctx->ioctx_private;
737 LASSERT(!IS_ERR(session));
739 for (i = 0; i < session->lis_ngroups; i++) {
740 group = session->lis_groups[i];
743 err = oig_wait(group->lig_oig);
748 ioctx->ioctx_cc += group->lig_rwcount;
750 session->lis_groups[i] = NULL;
756 ioctx->ioctx_cc = -1;
757 ioctx->ioctx_errno = -rc;
760 put_io_session(session);
761 ioctx->ioctx_private = NULL;