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);
151 CERROR("lock on inode without such object\n");
154 #warning "fix l_lock() using here!"
155 // l_lock(&lock->l_resource->lr_namespace->ns_lock);
156 kms = ldlm_extent_shift_kms(lock,
157 lsm->lsm_oinfo[stripe].loi_kms);
158 // l_unlock(&lock->l_resource->lr_namespace->ns_lock);
159 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
160 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
161 lsm->lsm_oinfo[stripe].loi_kms, kms);
162 lsm->lsm_oinfo[stripe].loi_kms = kms;
174 static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp)
176 struct ptlrpc_request *req = reqp;
177 struct inode *inode = llu_inode_from_lock(lock);
178 struct llu_inode_info *lli;
180 int rc, size = sizeof(*lvb), stripe = 0;
184 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
185 lli = llu_i2info(inode);
187 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
188 if (lli->lli_smd == NULL)
189 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
191 /* First, find out which stripe index this lock corresponds to. */
192 if (lli->lli_smd->lsm_stripe_count > 1) {
193 stripe = llu_lock_to_stripe_offset(inode, lock);
195 CWARN("lock on inode without such object\n");
196 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
200 rc = lustre_pack_reply(req, 1, &size, NULL);
202 CERROR("lustre_pack_reply: %d\n", rc);
206 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
207 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
209 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
210 lli->lli_st_size, stripe, lvb->lvb_size);
214 /* These errors are normal races, so we don't want to fill the console
215 * with messages by calling ptlrpc_error() */
216 if (rc == -ELDLM_NO_LOCK_DATA)
217 lustre_pack_reply(req, 0, NULL, NULL);
223 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms_only);
224 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
225 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
227 /* NB: lov_merge_size will prefer locally cached writes if they extend the
228 * file (because it prefers KMS over RSS when larger) */
229 int llu_glimpse_size(struct inode *inode)
231 struct llu_inode_info *lli = llu_i2info(inode);
232 struct llu_sb_info *sbi = llu_i2sbi(inode);
233 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
234 struct lustre_handle lockh = { 0 };
235 int rc, flags = LDLM_FL_HAS_INTENT;
238 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino);
240 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
241 LCK_PR, &flags, llu_extent_lock_callback,
242 ldlm_completion_ast, llu_glimpse_callback, inode,
243 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
245 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
246 RETURN(rc > 0 ? -EIO : rc);
249 lli->lli_st_size = lov_merge_size(lli->lli_smd, 0);
250 lli->lli_st_blocks = lov_merge_blocks(lli->lli_smd);
251 //lli->lli_st_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
253 CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %lu\n",
254 lli->lli_st_size, lli->lli_st_blocks);
256 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
261 int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
262 struct lov_stripe_md *lsm, int mode,
263 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
266 struct llu_sb_info *sbi = llu_i2sbi(inode);
267 struct llu_inode_info *lli = llu_i2info(inode);
271 LASSERT(lockh->cookie == 0);
273 /* XXX phil: can we do this? won't it screw the file size up? */
274 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
275 (sbi->ll_flags & LL_SBI_NOLCK))
278 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
279 lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end);
281 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
282 &ast_flags, llu_extent_lock_callback,
283 ldlm_completion_ast, llu_glimpse_callback, inode,
284 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
288 if (policy->l_extent.start == 0 &&
289 policy->l_extent.end == OBD_OBJECT_EOF)
290 lli->lli_st_size = lov_merge_size(lsm, 1);
292 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
297 int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode,
298 struct lov_stripe_md *lsm, int mode,
299 struct lustre_handle *lockh)
301 struct llu_sb_info *sbi = llu_i2sbi(inode);
305 /* XXX phil: can we do this? won't it screw the file size up? */
306 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
307 (sbi->ll_flags & LL_SBI_NOLCK))
310 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
315 #define LLAP_MAGIC 12346789
317 struct ll_async_page {
321 struct page *llap_page;
322 struct inode *llap_inode;
325 static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
327 struct ll_async_page *llap;
329 struct lov_stripe_md *lsm;
330 obd_valid valid_flags;
333 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);
354 llap->llap_queued = 0;
355 page = llap->llap_page;
358 if (cmd == OBD_BRW_WRITE)
359 CERROR("writeback error on page %p index %ld: %d\n",
360 page, page->index, rc);
365 static struct obd_async_page_ops llu_async_page_ops = {
366 .ap_make_ready = NULL,
367 .ap_refresh_count = NULL,
368 .ap_fill_obdo = llu_ap_fill_obdo,
369 .ap_completion = llu_ap_completion,
372 static int llu_queue_pio(int cmd, struct llu_io_group *group,
373 char *buf, size_t count, loff_t pos)
375 struct llu_inode_info *lli = llu_i2info(group->lig_inode);
376 struct lov_stripe_md *lsm = lli->lli_smd;
377 struct obd_export *exp = llu_i2dtexp(group->lig_inode);
378 struct page *pages = &group->lig_pages[group->lig_npages];
379 struct ll_async_page *llap = &group->lig_llap[group->lig_npages];
380 int i, rc, npages = 0, ret_bytes = 0;
386 /* prepare the pages array */
388 unsigned long index, offset, bytes;
390 offset = (pos & ~PAGE_CACHE_MASK);
391 index = pos >> PAGE_CACHE_SHIFT;
392 bytes = PAGE_CACHE_SIZE - offset;
396 /* prevent read beyond file range */
397 if ((cmd == OBD_BRW_READ) &&
398 (pos + bytes) >= lli->lli_st_size) {
399 if (pos >= lli->lli_st_size)
401 bytes = lli->lli_st_size - pos;
404 /* prepare page for this index */
405 pages[npages].index = index;
406 pages[npages].addr = buf - offset;
408 pages[npages]._offset = offset;
409 pages[npages]._count = bytes;
416 group->lig_rwcount += bytes;
420 group->lig_npages += npages;
422 for (i = 0; i < npages; i++) {
423 llap[i].llap_magic = LLAP_MAGIC;
424 rc = obd_prep_async_page(exp, lsm, NULL, &pages[i],
425 (obd_off)pages[i].index << PAGE_SHIFT,
427 &llap[i], &llap[i].llap_cookie);
430 llap[i].llap_cookie = NULL;
433 CDEBUG(D_CACHE, "llap %p page %p group %p obj off "LPU64"\n",
434 &llap[i], &pages[i], llap[i].llap_cookie,
435 (obd_off)pages[i].index << PAGE_SHIFT);
436 pages[i].private = (unsigned long)&llap[i];
437 llap[i].llap_page = &pages[i];
438 llap[i].llap_inode = group->lig_inode;
440 rc = obd_queue_group_io(exp, lsm, NULL, group->lig_oig,
441 llap[i].llap_cookie, cmd,
442 pages[i]._offset, pages[i]._count, 0,
443 ASYNC_READY | ASYNC_URGENT |
444 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
450 llap[i].llap_queued = 1;
457 struct llu_io_group * get_io_group(struct inode *inode, int maxpages)
459 struct llu_io_group *group;
462 OBD_ALLOC(group, LLU_IO_GROUP_SIZE(maxpages));
464 return ERR_PTR(-ENOMEM);
467 group->lig_inode = inode;
468 group->lig_maxpages = maxpages;
469 group->lig_llap = (struct ll_async_page *)(group + 1);
470 group->lig_pages = (struct page *) (group->lig_llap + maxpages);
472 rc = oig_init(&group->lig_oig);
474 OBD_FREE(group, LLU_IO_GROUP_SIZE(maxpages));
481 static int max_io_pages(ssize_t len, int iovlen)
483 return (((len + PAGE_SIZE -1) / PAGE_SIZE) + 2 + iovlen - 1);
487 void put_io_group(struct llu_io_group *group)
489 struct lov_stripe_md *lsm = llu_i2info(group->lig_inode)->lli_smd;
490 struct obd_export *exp = llu_i2dtexp(group->lig_inode);
491 struct ll_async_page *llap = group->lig_llap;
494 for (i = 0; i< group->lig_npages; i++) {
495 if (llap[i].llap_cookie)
496 obd_teardown_async_page(exp, lsm, NULL,
497 llap[i].llap_cookie);
500 I_RELE(group->lig_inode);
502 oig_release(group->lig_oig);
503 OBD_FREE(group, LLU_IO_GROUP_SIZE(group->lig_maxpages));
507 ssize_t llu_file_prwv(const struct iovec *iovec, int iovlen,
508 _SYSIO_OFF_T pos, ssize_t len,
511 struct llu_io_session *session = (struct llu_io_session *) private;
512 struct inode *inode = session->lis_inode;
513 struct llu_inode_info *lli = llu_i2info(inode);
514 struct ll_file_data *fd = lli->lli_file_data;
515 struct lustre_handle lockh = {0};
516 struct lov_stripe_md *lsm = lli->lli_smd;
517 struct obd_export *exp = NULL;
518 ldlm_policy_data_t policy;
519 struct llu_io_group *iogroup;
520 int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
521 LDLM_FL_BLOCK_NOWAIT : 0;
523 int err, is_read, lock_mode, iovidx, ret;
526 /* in a large iov read/write we'll be repeatedly called.
527 * so give a chance to answer cancel ast here
529 liblustre_wait_event(0);
531 exp = llu_i2dtexp(inode);
535 if (len == 0 || iovlen == 0)
538 if (pos + len > lli->lli_maxbytes)
541 iogroup = get_io_group(inode, max_io_pages(len, iovlen));
543 RETURN(PTR_ERR(iogroup));
545 is_read = session->lis_cmd == OBD_BRW_READ;
546 lock_mode = is_read ? LCK_PR : LCK_PW;
548 if (!is_read && (lli->lli_open_flags & O_APPEND)) {
549 policy.l_extent.start = 0;
550 policy.l_extent.end = OBD_OBJECT_EOF;
552 policy.l_extent.start = pos;
553 policy.l_extent.end = pos + len - 1;
556 err = llu_extent_lock(fd, inode, lsm, lock_mode, &policy,
562 kms = lov_merge_size(lsm, 1);
563 if (policy.l_extent.end > kms) {
564 /* A glimpse is necessary to determine whether we
565 * return a short read or some zeroes at the end of
567 if ((err = llu_glimpse_size(inode))) {
568 llu_extent_unlock(fd, inode, lsm,
573 lli->lli_st_size = kms;
576 if (lli->lli_open_flags & O_APPEND)
577 pos = lli->lli_st_size;
580 for (iovidx = 0; iovidx < iovlen; iovidx++) {
581 char *buf = (char *) iovec[iovidx].iov_base;
582 size_t count = iovec[iovidx].iov_len;
588 if (IS_BAD_PTR(buf) || IS_BAD_PTR(buf + count)) {
589 llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
590 GOTO(err_put, err = -EFAULT);
594 if (pos >= lli->lli_st_size)
597 if (pos >= lli->lli_maxbytes) {
598 llu_extent_unlock(fd, inode, lsm, lock_mode,
600 GOTO(err_put, err = -EFBIG);
602 if (pos + count >= lli->lli_maxbytes)
603 count = lli->lli_maxbytes - pos;
606 ret = llu_queue_pio(session->lis_cmd, iogroup, buf, count, pos);
608 llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
609 GOTO(err_put, err = ret);
613 LASSERT(ret == count);
614 obd_adjust_kms(exp, lsm, pos, 0);
615 /* file size grow immediately */
616 if (pos > lli->lli_st_size)
617 lli->lli_st_size = pos;
624 LASSERT(len == 0 || is_read); /* libsysio should guarantee this */
626 err = llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
628 CERROR("extent unlock error %d\n", err);
630 err = obd_trigger_group_io(exp, lsm, NULL, iogroup->lig_oig);
634 session->lis_groups[session->lis_ngroups++] = iogroup;
637 put_io_group(iogroup);
638 RETURN((ssize_t)err);
642 struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd)
644 struct llu_io_session *session;
646 OBD_ALLOC(session, LLU_IO_SESSION_SIZE(ngroups));
651 session->lis_inode = ino;
652 session->lis_max_groups = ngroups;
653 session->lis_cmd = cmd;
657 static void put_io_session(struct llu_io_session *session)
661 for (i = 0; i < session->lis_ngroups; i++) {
662 if (session->lis_groups[i]) {
663 put_io_group(session->lis_groups[i]);
664 session->lis_groups[i] = NULL;
668 I_RELE(session->lis_inode);
669 OBD_FREE(session, LLU_IO_SESSION_SIZE(session->lis_max_groups));
672 static int llu_file_rwx(struct inode *ino,
676 struct llu_io_session *session;
678 int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE;
681 LASSERT(ioctx->ioctx_xtvlen >= 0);
682 LASSERT(ioctx->ioctx_iovlen >= 0);
684 liblustre_wait_event(0);
686 if (!ioctx->ioctx_xtvlen)
689 /* XXX consider other types later */
690 LASSERT(S_ISREG(llu_i2info(ino)->lli_st_mode));
692 session = get_io_session(ino, ioctx->ioctx_xtvlen * 2, cmd);
696 cc = _sysio_enumerate_extents(ioctx->ioctx_xtv, ioctx->ioctx_xtvlen,
697 ioctx->ioctx_iov, ioctx->ioctx_iovlen,
698 llu_file_prwv, session);
701 LASSERT(!ioctx->ioctx_cc);
702 ioctx->ioctx_private = session;
705 put_io_session(session);
710 int llu_iop_read(struct inode *ino,
713 return llu_file_rwx(ino, ioctx, 1);
716 int llu_iop_write(struct inode *ino,
722 memset(&iattr, 0, sizeof(iattr));
723 iattr.ia_mtime = iattr.ia_atime = CURRENT_TIME;
724 iattr.ia_valid = ATTR_MTIME | ATTR_ATIME | ATTR_RAW;
726 liblustre_wait_event(0);
727 rc = llu_setattr_raw(ino, &iattr);
729 CERROR("failed to set mtime/atime during write: %d", rc);
730 /* XXX should continue or return error? */
733 return llu_file_rwx(ino, ioctx, 0);
736 int llu_iop_iodone(struct ioctx *ioctx)
738 struct llu_io_session *session;
739 struct llu_io_group *group;
740 int i, err = 0, rc = 0;
743 liblustre_wait_event(0);
745 session = (struct llu_io_session *) ioctx->ioctx_private;
747 LASSERT(!IS_ERR(session));
749 for (i = 0; i < session->lis_ngroups; i++) {
750 group = session->lis_groups[i];
753 err = oig_wait(group->lig_oig);
758 ioctx->ioctx_cc += group->lig_rwcount;
760 session->lis_groups[i] = NULL;
766 ioctx->ioctx_cc = -1;
767 ioctx->ioctx_errno = -rc;
770 put_io_session(session);
771 ioctx->ioctx_private = NULL;