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>
44 #include "llite_lib.h"
48 struct obd_io_group *lig_oig;
49 struct inode *lig_inode;
53 struct ll_async_page *lig_llap;
54 struct page *lig_pages;
57 #define LLU_IO_GROUP_SIZE(x) \
58 (sizeof(struct llu_io_group) + \
59 sizeof(struct ll_async_page) * (x) + \
60 sizeof(struct page) * (x))
64 struct inode *lis_inode;
68 struct llu_io_group *lis_groups[0];
70 #define LLU_IO_SESSION_SIZE(x) \
71 (sizeof(struct llu_io_session) + (x) * 2 * sizeof(void *))
74 typedef ssize_t llu_file_piov_t(const struct iovec *iovec, int iovlen,
75 _SYSIO_OFF_T pos, ssize_t len,
78 static int llu_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
80 struct llu_inode_info *lli = llu_i2info(inode);
81 struct lov_stripe_md *lsm = lli->lli_smd;
82 struct obd_export *exp = llu_i2obdexp(inode);
85 struct ldlm_lock *lock;
86 struct lov_stripe_md *lsm;
87 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
88 __u32 stripe, vallen = sizeof(stripe);
92 if (lsm->lsm_stripe_count == 1)
95 /* get our offset in the lov */
96 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
98 CERROR("obd_get_info: rc = %d\n", rc);
101 LASSERT(stripe < lsm->lsm_stripe_count);
105 static int llu_extent_lock_callback(struct ldlm_lock *lock,
106 struct ldlm_lock_desc *new, void *data,
109 struct lustre_handle lockh = { 0 };
113 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
114 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
119 case LDLM_CB_BLOCKING:
120 ldlm_lock2handle(lock, &lockh);
121 rc = ldlm_cli_cancel(&lockh);
123 CERROR("ldlm_cli_cancel failed: %d\n", rc);
125 case LDLM_CB_CANCELING: {
127 struct llu_inode_info *lli;
128 struct lov_stripe_md *lsm;
132 /* This lock wasn't granted, don't try to evict pages */
133 if (lock->l_req_mode != lock->l_granted_mode)
136 inode = llu_inode_from_lock(lock);
139 lli= llu_i2info(inode);
146 stripe = llu_lock_to_stripe_offset(inode, lock);
147 l_lock(&lock->l_resource->lr_namespace->ns_lock);
148 kms = ldlm_extent_shift_kms(lock,
149 lsm->lsm_oinfo[stripe].loi_kms);
150 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
151 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
152 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
153 lsm->lsm_oinfo[stripe].loi_kms, kms);
154 lsm->lsm_oinfo[stripe].loi_kms = kms;
166 static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp)
168 struct ptlrpc_request *req = reqp;
169 struct inode *inode = llu_inode_from_lock(lock);
170 struct llu_inode_info *lli;
172 int rc, size = sizeof(*lvb), stripe = 0;
176 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
177 lli = llu_i2info(inode);
179 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
180 if (lli->lli_smd == NULL)
181 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
183 /* First, find out which stripe index this lock corresponds to. */
184 if (lli->lli_smd->lsm_stripe_count > 1)
185 stripe = llu_lock_to_stripe_offset(inode, lock);
187 rc = lustre_pack_reply(req, 1, &size, NULL);
189 CERROR("lustre_pack_reply: %d\n", rc);
193 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
194 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
196 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
197 lli->lli_st_size, stripe, lvb->lvb_size);
201 /* These errors are normal races, so we don't want to fill the console
202 * with messages by calling ptlrpc_error() */
203 if (rc == -ELDLM_NO_LOCK_DATA)
204 lustre_pack_reply(req, 0, NULL, NULL);
210 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms_only);
211 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
212 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
214 /* NB: lov_merge_size will prefer locally cached writes if they extend the
215 * file (because it prefers KMS over RSS when larger) */
216 int llu_glimpse_size(struct inode *inode)
218 struct llu_inode_info *lli = llu_i2info(inode);
219 struct llu_sb_info *sbi = llu_i2sbi(inode);
220 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
221 struct lustre_handle lockh = { 0 };
222 int rc, flags = LDLM_FL_HAS_INTENT;
225 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino);
227 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
228 LCK_PR, &flags, llu_extent_lock_callback,
229 ldlm_completion_ast, llu_glimpse_callback, inode,
230 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
232 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
233 RETURN(rc > 0 ? -EIO : rc);
236 lli->lli_st_size = lov_merge_size(lli->lli_smd, 0);
237 lli->lli_st_blocks = lov_merge_blocks(lli->lli_smd);
238 //lli->lli_st_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
240 CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %lu\n",
241 lli->lli_st_size, lli->lli_st_blocks);
243 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
248 int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
249 struct lov_stripe_md *lsm, int mode,
250 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
253 struct llu_sb_info *sbi = llu_i2sbi(inode);
254 struct llu_inode_info *lli = llu_i2info(inode);
258 LASSERT(lockh->cookie == 0);
260 /* XXX phil: can we do this? won't it screw the file size up? */
261 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
262 (sbi->ll_flags & LL_SBI_NOLCK))
265 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
266 lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end);
268 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
269 &ast_flags, llu_extent_lock_callback,
270 ldlm_completion_ast, llu_glimpse_callback, inode,
271 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
275 if (policy->l_extent.start == 0 &&
276 policy->l_extent.end == OBD_OBJECT_EOF)
277 lli->lli_st_size = lov_merge_size(lsm, 1);
279 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
284 int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode,
285 struct lov_stripe_md *lsm, int mode,
286 struct lustre_handle *lockh)
288 struct llu_sb_info *sbi = llu_i2sbi(inode);
292 /* XXX phil: can we do this? won't it screw the file size up? */
293 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
294 (sbi->ll_flags & LL_SBI_NOLCK))
297 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
302 #define LLAP_MAGIC 12346789
304 struct ll_async_page {
308 struct page *llap_page;
309 struct inode *llap_inode;
312 static struct ll_async_page *llap_from_cookie(void *cookie)
314 struct ll_async_page *llap = cookie;
315 if (llap->llap_magic != LLAP_MAGIC)
316 return ERR_PTR(-EINVAL);
320 static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
322 struct ll_async_page *llap;
324 struct lov_stripe_md *lsm;
325 obd_flag valid_flags;
328 llap = llap_from_cookie(data);
334 inode = llap->llap_inode;
335 lsm = llu_i2info(inode)->lli_smd;
337 oa->o_id = lsm->lsm_object_id;
338 oa->o_valid = OBD_MD_FLID;
339 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
340 if (cmd == OBD_BRW_WRITE)
341 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
343 obdo_from_inode(oa, inode, valid_flags);
347 /* called for each page in a completed rpc.*/
348 static void llu_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
350 struct ll_async_page *llap;
353 llap = llap_from_cookie(data);
359 llap->llap_queued = 0;
360 page = llap->llap_page;
363 if (cmd == OBD_BRW_WRITE)
364 CERROR("writeback error on page %p index %ld: %d\n",
365 page, page->index, rc);
370 static struct obd_async_page_ops llu_async_page_ops = {
371 .ap_make_ready = NULL,
372 .ap_refresh_count = NULL,
373 .ap_fill_obdo = llu_ap_fill_obdo,
374 .ap_completion = llu_ap_completion,
377 static int llu_queue_pio(int cmd, struct llu_io_group *group,
378 char *buf, size_t count, loff_t pos)
380 struct llu_inode_info *lli = llu_i2info(group->lig_inode);
381 struct lov_stripe_md *lsm = lli->lli_smd;
382 struct obd_export *exp = llu_i2obdexp(group->lig_inode);
383 struct page *pages = &group->lig_pages[group->lig_npages];
384 struct ll_async_page *llap = &group->lig_llap[group->lig_npages];
385 int i, rc, npages = 0, ret_bytes = 0;
391 /* prepare the pages array */
393 unsigned long index, offset, bytes;
395 offset = (pos & ~PAGE_CACHE_MASK);
396 index = pos >> PAGE_CACHE_SHIFT;
397 bytes = PAGE_CACHE_SIZE - offset;
401 /* prevent read beyond file range */
402 if ((cmd == OBD_BRW_READ) &&
403 (pos + bytes) >= lli->lli_st_size) {
404 if (pos >= lli->lli_st_size)
406 bytes = lli->lli_st_size - pos;
409 /* prepare page for this index */
410 pages[npages].index = index;
411 pages[npages].addr = buf - offset;
413 pages[npages]._offset = offset;
414 pages[npages]._count = bytes;
421 group->lig_rwcount += bytes;
425 group->lig_npages += npages;
427 for (i = 0; i < npages; i++) {
428 llap[i].llap_magic = LLAP_MAGIC;
429 rc = obd_prep_async_page(exp, lsm, NULL, &pages[i],
430 (obd_off)pages[i].index << PAGE_SHIFT,
432 &llap[i], &llap[i].llap_cookie);
435 llap[i].llap_cookie = NULL;
438 CDEBUG(D_CACHE, "llap %p page %p group %p obj off "LPU64"\n",
439 &llap[i], &pages[i], llap[i].llap_cookie,
440 (obd_off)pages[i].index << PAGE_SHIFT);
441 pages[i].private = (unsigned long)&llap[i];
442 llap[i].llap_page = &pages[i];
443 llap[i].llap_inode = group->lig_inode;
445 rc = obd_queue_group_io(exp, lsm, NULL, group->lig_oig,
446 llap[i].llap_cookie, cmd,
447 pages[i]._offset, pages[i]._count, 0,
448 ASYNC_READY | ASYNC_URGENT |
449 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
455 llap[i].llap_queued = 1;
462 struct llu_io_group * get_io_group(struct inode *inode, int maxpages)
464 struct llu_io_group *group;
467 OBD_ALLOC(group, LLU_IO_GROUP_SIZE(maxpages));
469 return ERR_PTR(-ENOMEM);
472 group->lig_inode = inode;
473 group->lig_maxpages = maxpages;
474 group->lig_llap = (struct ll_async_page *)(group + 1);
475 group->lig_pages = (struct page *) (group->lig_llap + maxpages);
477 rc = oig_init(&group->lig_oig);
479 OBD_FREE(group, LLU_IO_GROUP_SIZE(maxpages));
486 static int max_io_pages(ssize_t len, int iovlen)
488 return (((len + PAGE_SIZE -1) / PAGE_SIZE) + 2 + iovlen - 1);
492 void put_io_group(struct llu_io_group *group)
494 struct lov_stripe_md *lsm = llu_i2info(group->lig_inode)->lli_smd;
495 struct obd_export *exp = llu_i2obdexp(group->lig_inode);
496 struct ll_async_page *llap = group->lig_llap;
499 for (i = 0; i< group->lig_npages; i++) {
500 if (llap[i].llap_cookie)
501 obd_teardown_async_page(exp, lsm, NULL,
502 llap[i].llap_cookie);
505 I_RELE(group->lig_inode);
507 oig_release(group->lig_oig);
508 OBD_FREE(group, LLU_IO_GROUP_SIZE(group->lig_maxpages));
511 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
515 ssize_t llu_file_prwv(const struct iovec *iovec, int iovlen,
516 _SYSIO_OFF_T pos, ssize_t len,
519 struct llu_io_session *session = (struct llu_io_session *) private;
520 struct inode *inode = session->lis_inode;
521 struct llu_inode_info *lli = llu_i2info(inode);
522 struct ll_file_data *fd = lli->lli_file_data;
523 struct lustre_handle lockh = {0};
524 struct lov_stripe_md *lsm = lli->lli_smd;
525 struct obd_export *exp = NULL;
526 ldlm_policy_data_t policy;
527 struct llu_io_group *iogroup;
528 int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
529 LDLM_FL_BLOCK_NOWAIT : 0;
531 int err, is_read, lock_mode, iovidx, ret;
534 /* in a large iov read/write we'll be repeatedly called.
535 * so give a chance to answer cancel ast here
537 liblustre_wait_event(0);
539 exp = llu_i2obdexp(inode);
543 if (len == 0 || iovlen == 0)
546 if (pos + len > lli->lli_maxbytes)
549 iogroup = get_io_group(inode, max_io_pages(len, iovlen));
551 RETURN(PTR_ERR(iogroup));
553 is_read = session->lis_cmd == OBD_BRW_READ;
554 lock_mode = is_read ? LCK_PR : LCK_PW;
556 if (!is_read && (lli->lli_open_flags & O_APPEND)) {
557 policy.l_extent.start = 0;
558 policy.l_extent.end = OBD_OBJECT_EOF;
560 policy.l_extent.start = pos;
561 policy.l_extent.end = pos + len - 1;
564 err = llu_extent_lock(fd, inode, lsm, lock_mode, &policy,
570 kms = lov_merge_size(lsm, 1);
571 if (policy.l_extent.end > kms) {
572 /* A glimpse is necessary to determine whether we
573 * return a short read or some zeroes at the end of
575 if ((err = llu_glimpse_size(inode))) {
576 llu_extent_unlock(fd, inode, lsm,
581 lli->lli_st_size = kms;
584 if (lli->lli_open_flags & O_APPEND)
585 pos = lli->lli_st_size;
588 for (iovidx = 0; iovidx < iovlen; iovidx++) {
589 char *buf = (char *) iovec[iovidx].iov_base;
590 size_t count = iovec[iovidx].iov_len;
596 if (IS_BAD_PTR(buf) || IS_BAD_PTR(buf + count)) {
597 llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
598 GOTO(err_put, err = -EFAULT);
602 if (pos >= lli->lli_st_size)
605 if (pos >= lli->lli_maxbytes) {
606 llu_extent_unlock(fd, inode, lsm, lock_mode,
608 GOTO(err_put, err = -EFBIG);
610 if (pos + count >= lli->lli_maxbytes)
611 count = lli->lli_maxbytes - pos;
614 ret = llu_queue_pio(session->lis_cmd, iogroup, buf, count, pos);
616 llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
617 GOTO(err_put, err = ret);
621 LASSERT(ret == count);
622 lov_increase_kms(exp, lsm, pos);
623 /* file size grow immediately */
624 if (pos > lli->lli_st_size)
625 lli->lli_st_size = pos;
632 LASSERT(len == 0 || is_read); /* libsysio should guarantee this */
634 err = llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
636 CERROR("extent unlock error %d\n", err);
638 err = obd_trigger_group_io(exp, lsm, NULL, iogroup->lig_oig);
642 session->lis_groups[session->lis_ngroups++] = iogroup;
645 put_io_group(iogroup);
646 RETURN((ssize_t)err);
650 struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd)
652 struct llu_io_session *session;
654 OBD_ALLOC(session, LLU_IO_SESSION_SIZE(ngroups));
659 session->lis_inode = ino;
660 session->lis_max_groups = ngroups;
661 session->lis_cmd = cmd;
665 static void put_io_session(struct llu_io_session *session)
669 for (i = 0; i < session->lis_ngroups; i++) {
670 if (session->lis_groups[i]) {
671 put_io_group(session->lis_groups[i]);
672 session->lis_groups[i] = NULL;
676 I_RELE(session->lis_inode);
677 OBD_FREE(session, LLU_IO_SESSION_SIZE(session->lis_max_groups));
680 static int llu_file_rwx(struct inode *ino,
684 struct llu_io_session *session;
686 int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE;
689 LASSERT(ioctx->ioctx_xtvlen >= 0);
690 LASSERT(ioctx->ioctx_iovlen >= 0);
692 liblustre_wait_event(0);
694 if (!ioctx->ioctx_xtvlen)
697 /* XXX consider other types later */
698 LASSERT(S_ISREG(llu_i2info(ino)->lli_st_mode));
700 session = get_io_session(ino, ioctx->ioctx_xtvlen * 2, cmd);
704 cc = _sysio_enumerate_extents(ioctx->ioctx_xtv, ioctx->ioctx_xtvlen,
705 ioctx->ioctx_iov, ioctx->ioctx_iovlen,
706 llu_file_prwv, session);
709 LASSERT(!ioctx->ioctx_cc);
710 ioctx->ioctx_private = session;
713 put_io_session(session);
718 int llu_iop_read(struct inode *ino,
721 return llu_file_rwx(ino, ioctx, 1);
724 int llu_iop_write(struct inode *ino,
730 memset(&iattr, 0, sizeof(iattr));
731 iattr.ia_mtime = iattr.ia_atime = CURRENT_TIME;
732 iattr.ia_valid = ATTR_MTIME | ATTR_ATIME | ATTR_RAW;
734 liblustre_wait_event(0);
735 rc = llu_setattr_raw(ino, &iattr);
737 CERROR("failed to set mtime/atime during write: %d", rc);
738 /* XXX should continue or return error? */
741 return llu_file_rwx(ino, ioctx, 0);
744 int llu_iop_iodone(struct ioctx *ioctx)
746 struct llu_io_session *session;
747 struct llu_io_group *group;
748 int i, err = 0, rc = 0;
751 liblustre_wait_event(0);
753 session = (struct llu_io_session *) ioctx->ioctx_private;
755 LASSERT(!IS_ERR(session));
757 for (i = 0; i < session->lis_ngroups; i++) {
758 group = session->lis_groups[i];
761 err = oig_wait(group->lig_oig);
766 ioctx->ioctx_cc += group->lig_rwcount;
768 session->lis_groups[i] = NULL;
774 ioctx->ioctx_cc = -1;
775 ioctx->ioctx_errno = -rc;
778 put_io_session(session);
779 ioctx->ioctx_private = NULL;