4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
33 #include <linux/sched.h>
34 #include <linux/kthread.h>
36 #include <linux/highmem.h>
37 #include <linux/pagemap.h>
38 #include <linux/delay.h>
40 #define DEBUG_SUBSYSTEM S_LLITE
42 #include <obd_support.h>
43 #include <lustre_dlm.h>
44 #include "llite_internal.h"
46 #define SA_OMITTED_ENTRY_MAX 8ULL
49 /** negative values are for error cases */
50 SA_ENTRY_INIT = 0, /** init entry */
51 SA_ENTRY_SUCC = 1, /** stat succeed */
52 SA_ENTRY_INVA = 2, /** invalid entry */
56 * sa_entry is not refcounted: statahead thread allocates it and do async stat,
57 * and in async stat callback ll_statahead_interpret() will prepare the inode
58 * and set lock data in the ptlrpcd context. Then the scanner process will be
59 * woken up if this entry is the waiting one, can access and free it.
62 /* link into sai_entries */
63 struct list_head se_list;
64 /* link into sai hash table locally */
65 struct list_head se_hash;
66 /* entry index in the sai */
68 /* low layer ldlm lock handle */
72 /* entry size, contains name */
74 /* pointer to the target inode */
75 struct inode *se_inode;
76 /* pointer to @sai per process struct */
77 struct ll_statahead_info *se_sai;
84 static unsigned int sai_generation;
85 static DEFINE_SPINLOCK(sai_generation_lock);
87 static inline int sa_unhashed(struct sa_entry *entry)
89 return list_empty(&entry->se_hash);
92 /* sa_entry is ready to use */
93 static inline int sa_ready(struct sa_entry *entry)
95 /* Make sure sa_entry is updated and ready to use */
97 return (entry->se_state != SA_ENTRY_INIT);
100 /* hash value to put in sai_cache */
101 static inline int sa_hash(int val)
103 return val & LL_SA_CACHE_MASK;
106 /* hash entry into sax_cache */
108 sa_rehash(struct ll_statahead_context *ctx, struct sa_entry *entry)
110 int i = sa_hash(entry->se_qstr.hash);
112 spin_lock(&ctx->sax_cache_lock[i]);
113 list_add_tail(&entry->se_hash, &ctx->sax_cache[i]);
114 spin_unlock(&ctx->sax_cache_lock[i]);
117 /* unhash entry from sai_cache */
119 sa_unhash(struct ll_statahead_context *ctx, struct sa_entry *entry)
121 int i = sa_hash(entry->se_qstr.hash);
123 spin_lock(&ctx->sax_cache_lock[i]);
124 list_del_init(&entry->se_hash);
125 spin_unlock(&ctx->sax_cache_lock[i]);
128 static inline int agl_should_run(struct ll_statahead_info *sai,
131 return inode && S_ISREG(inode->i_mode) && sai->sai_agl_task;
134 static inline struct ll_inode_info *
135 agl_first_entry(struct ll_statahead_info *sai)
137 return list_first_entry(&sai->sai_agls, struct ll_inode_info,
141 /* statahead window is full */
142 static inline int sa_sent_full(struct ll_statahead_info *sai)
144 return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
147 /* Batch metadata handle */
148 static inline bool sa_has_batch_handle(struct ll_statahead_info *sai)
150 return sai->sai_bh != NULL;
153 static inline void ll_statahead_flush_nowait(struct ll_statahead_info *sai)
155 if (sa_has_batch_handle(sai)) {
156 sai->sai_index_end = sai->sai_index - 1;
157 (void) md_batch_flush(ll_i2mdexp(sai->sai_dentry->d_inode),
162 static inline int agl_list_empty(struct ll_statahead_info *sai)
164 return list_empty(&sai->sai_agls);
168 * (1) hit ratio less than 80%
170 * (2) consecutive miss more than 8
171 * then means low hit.
173 static inline int sa_low_hit(struct ll_statahead_info *sai)
175 return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
176 (sai->sai_consecutive_miss > 8));
180 * if the given index is behind of statahead window more than
181 * SA_OMITTED_ENTRY_MAX, then it is old.
183 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
185 return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
189 /* allocate sa_entry and hash it to allow scanner process to find it */
190 static struct sa_entry *
191 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
192 const char *name, int len, const struct lu_fid *fid)
194 struct ll_inode_info *lli;
195 struct sa_entry *entry;
201 entry_size = sizeof(struct sa_entry) +
202 round_up(len + 1 /* for trailing NUL */, 4);
203 OBD_ALLOC(entry, entry_size);
204 if (unlikely(!entry))
205 RETURN(ERR_PTR(-ENOMEM));
207 CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
208 len, name, entry, index);
210 entry->se_index = index;
213 entry->se_state = SA_ENTRY_INIT;
214 entry->se_size = entry_size;
215 dname = (char *)entry + sizeof(struct sa_entry);
216 memcpy(dname, name, len);
218 entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
219 entry->se_qstr.len = len;
220 entry->se_qstr.name = dname;
223 entry->se_fid = *fid;
225 lli = ll_i2info(sai->sai_dentry->d_inode);
226 spin_lock(&lli->lli_sa_lock);
227 INIT_LIST_HEAD(&entry->se_list);
228 sa_rehash(lli->lli_sax, entry);
229 spin_unlock(&lli->lli_sa_lock);
231 atomic_inc(&sai->sai_cache_count);
236 /* free sa_entry, which should have been unhashed and not in any list */
237 static void sa_free(struct ll_statahead_context *ctx, struct sa_entry *entry)
239 CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
240 entry->se_qstr.len, entry->se_qstr.name, entry,
243 LASSERT(list_empty(&entry->se_list));
244 LASSERT(sa_unhashed(entry));
246 OBD_FREE(entry, entry->se_size);
250 * find sa_entry by name, used by directory scanner, lock is not needed because
251 * only scanner can remove the entry from cache.
253 static struct sa_entry *
254 sa_get(struct ll_statahead_context *ctx, const struct qstr *qstr)
256 struct sa_entry *entry;
257 int i = sa_hash(qstr->hash);
259 spin_lock(&ctx->sax_cache_lock[i]);
260 list_for_each_entry(entry, &ctx->sax_cache[i], se_hash) {
261 if (entry->se_qstr.hash == qstr->hash &&
262 entry->se_qstr.len == qstr->len &&
263 memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0) {
264 spin_unlock(&ctx->sax_cache_lock[i]);
268 spin_unlock(&ctx->sax_cache_lock[i]);
272 /* unhash and unlink sa_entry, and then free it */
274 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry, bool locked)
276 struct inode *dir = sai->sai_dentry->d_inode;
277 struct ll_inode_info *lli = ll_i2info(dir);
278 struct ll_statahead_context *ctx = lli->lli_sax;
280 LASSERT(!sa_unhashed(entry));
281 LASSERT(!list_empty(&entry->se_list));
282 LASSERT(sa_ready(entry));
284 sa_unhash(ctx, entry);
287 spin_lock(&lli->lli_sa_lock);
288 list_del_init(&entry->se_list);
289 spin_unlock(&lli->lli_sa_lock);
291 iput(entry->se_inode);
292 atomic_dec(&sai->sai_cache_count);
295 spin_lock(&lli->lli_sa_lock);
298 /* called by scanner after use, sa_entry will be killed */
300 sa_put(struct inode *dir, struct ll_statahead_info *sai, struct sa_entry *entry)
302 struct ll_inode_info *lli = ll_i2info(dir);
303 struct sa_entry *tmp;
306 if (entry && entry->se_state == SA_ENTRY_SUCC) {
307 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
310 sai->sai_consecutive_miss = 0;
311 if (sai->sai_max < sbi->ll_sa_max) {
312 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
314 } else if (sai->sai_max_batch_count > 0) {
315 if (sai->sai_max >= sai->sai_max_batch_count &&
316 (sai->sai_index_end - entry->se_index) %
317 sai->sai_max_batch_count == 0) {
319 } else if (entry->se_index == sai->sai_index_end) {
327 sai->sai_consecutive_miss++;
332 sa_kill(sai, entry, false);
336 * kill old completed entries. Maybe kicking old entries can
339 spin_lock(&lli->lli_sa_lock);
340 while ((tmp = list_first_entry_or_null(&sai->sai_entries,
341 struct sa_entry, se_list))) {
342 if (!is_omitted_entry(sai, tmp->se_index))
345 /* ll_sa_lock is dropped by sa_kill(), restart list */
346 sa_kill(sai, tmp, true);
348 spin_unlock(&lli->lli_sa_lock);
351 spin_lock(&lli->lli_sa_lock);
352 if (wakeup && sai->sai_task)
353 wake_up_process(sai->sai_task);
354 spin_unlock(&lli->lli_sa_lock);
358 * update state and sort add entry to sai_entries by index, return true if
359 * scanner is waiting on this entry.
362 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
365 struct list_head *pos = &sai->sai_entries;
366 __u64 index = entry->se_index;
368 LASSERT(!sa_ready(entry));
369 LASSERT(list_empty(&entry->se_list));
371 list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
372 if (se->se_index < entry->se_index) {
377 list_add(&entry->se_list, pos);
379 * LU-9210: ll_statahead_interpet must be able to see this before
382 smp_store_release(&entry->se_state,
383 ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
385 return (index == sai->sai_index_wait);
388 /* finish async stat RPC arguments */
389 static void sa_fini_data(struct md_op_item *item)
391 struct md_op_data *op_data = &item->mop_data;
393 if (op_data->op_flags & MF_OPNAME_KMALLOCED)
394 /* allocated via ll_setup_filename called from sa_prep_data */
395 kfree(op_data->op_name);
396 ll_unlock_md_op_lsm(&item->mop_data);
398 if (item->mop_subpill_allocated)
399 OBD_FREE_PTR(item->mop_pill);
403 static int ll_statahead_interpret(struct md_op_item *item, int rc);
406 * prepare arguments for async stat RPC.
408 static struct md_op_item *
409 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
411 struct md_op_item *item;
412 struct ldlm_enqueue_info *einfo;
413 struct md_op_data *op_data;
417 return ERR_PTR(-ENOMEM);
419 op_data = ll_prep_md_op_data(&item->mop_data, dir, child,
420 entry->se_qstr.name, entry->se_qstr.len, 0,
421 LUSTRE_OPC_ANY, NULL);
422 if (IS_ERR(op_data)) {
424 return (struct md_op_item *)op_data;
428 op_data->op_fid2 = entry->se_fid;
430 item->mop_opc = MD_OP_GETATTR;
431 item->mop_it.it_op = IT_GETATTR;
432 item->mop_dir = igrab(dir);
433 item->mop_cb = ll_statahead_interpret;
434 item->mop_cbdata = entry;
436 einfo = &item->mop_einfo;
437 einfo->ei_type = LDLM_IBITS;
438 einfo->ei_mode = it_to_lock_mode(&item->mop_it);
439 einfo->ei_cb_bl = ll_md_blocking_ast;
440 einfo->ei_cb_cp = ldlm_completion_ast;
441 einfo->ei_cb_gl = NULL;
442 einfo->ei_cbdata = NULL;
443 einfo->ei_req_slot = 1;
449 * release resources used in async stat RPC, update entry state and wakeup if
450 * scanner process it waiting on this entry.
453 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
455 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
458 spin_lock(&lli->lli_sa_lock);
459 wakeup = __sa_make_ready(sai, entry, ret);
460 spin_unlock(&lli->lli_sa_lock);
463 wake_up(&sai->sai_waitq);
466 /* insert inode into the list of sai_agls */
467 static void ll_agl_add(struct ll_statahead_info *sai,
468 struct inode *inode, int index)
470 struct ll_inode_info *child = ll_i2info(inode);
471 struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
473 spin_lock(&child->lli_agl_lock);
474 if (child->lli_agl_index == 0) {
475 child->lli_agl_index = index;
476 spin_unlock(&child->lli_agl_lock);
478 LASSERT(list_empty(&child->lli_agl_list));
480 spin_lock(&parent->lli_agl_lock);
481 /* Re-check under the lock */
482 if (agl_should_run(sai, inode)) {
483 if (agl_list_empty(sai))
484 wake_up_process(sai->sai_agl_task);
486 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
488 child->lli_agl_index = 0;
489 spin_unlock(&parent->lli_agl_lock);
491 spin_unlock(&child->lli_agl_lock);
496 static struct ll_statahead_context *ll_sax_alloc(struct inode *dir)
498 struct ll_statahead_context *ctx;
507 ctx->sax_inode = igrab(dir);
508 atomic_set(&ctx->sax_refcount, 1);
509 for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
510 INIT_LIST_HEAD(&ctx->sax_cache[i]);
511 spin_lock_init(&ctx->sax_cache_lock[i]);
517 static inline void ll_sax_free(struct ll_statahead_context *ctx)
519 LASSERT(ctx->sax_inode != NULL);
520 iput(ctx->sax_inode);
524 static inline void __ll_sax_get(struct ll_statahead_context *ctx)
526 atomic_inc(&ctx->sax_refcount);
529 static inline struct ll_statahead_context *ll_sax_get(struct inode *dir)
531 struct ll_inode_info *lli = ll_i2info(dir);
532 struct ll_statahead_context *ctx = NULL;
534 spin_lock(&lli->lli_sa_lock);
538 spin_unlock(&lli->lli_sa_lock);
543 static inline void ll_sax_put(struct inode *dir,
544 struct ll_statahead_context *ctx)
546 struct ll_inode_info *lli = ll_i2info(dir);
548 if (atomic_dec_and_lock(&ctx->sax_refcount, &lli->lli_sa_lock)) {
551 if (lli->lli_sa_pattern & (LSA_PATTERN_ADVISE |
552 LSA_PATTERN_FNAME)) {
553 lli->lli_opendir_key = NULL;
554 lli->lli_opendir_pid = 0;
555 lli->lli_sa_enabled = 0;
557 lli->lli_sa_pattern = LSA_PATTERN_NONE;
558 spin_unlock(&lli->lli_sa_lock);
565 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
567 struct ll_statahead_info *sai;
568 struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
576 sai->sai_dentry = dget(dentry);
577 atomic_set(&sai->sai_refcount, 1);
578 sai->sai_max = ll_i2sbi(dentry->d_inode)->ll_sa_min;
580 init_waitqueue_head(&sai->sai_waitq);
582 INIT_LIST_HEAD(&sai->sai_entries);
583 INIT_LIST_HEAD(&sai->sai_agls);
585 atomic_set(&sai->sai_cache_count, 0);
587 spin_lock(&sai_generation_lock);
588 lli->lli_sa_generation = ++sai_generation;
589 if (unlikely(sai_generation == 0))
590 lli->lli_sa_generation = ++sai_generation;
591 spin_unlock(&sai_generation_lock);
597 static inline void ll_sai_free(struct ll_statahead_info *sai)
599 LASSERT(sai->sai_dentry != NULL);
600 dput(sai->sai_dentry);
604 static inline struct ll_statahead_info *
605 __ll_sai_get(struct ll_statahead_info *sai)
607 atomic_inc(&sai->sai_refcount);
612 * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
615 static void ll_sai_put(struct ll_statahead_info *sai)
617 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
619 if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
620 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
621 struct sa_entry *entry, *next;
624 spin_unlock(&lli->lli_sa_lock);
626 LASSERT(!sai->sai_task);
627 LASSERT(!sai->sai_agl_task);
628 LASSERT(sai->sai_sent == sai->sai_replied);
630 list_for_each_entry_safe(entry, next, &sai->sai_entries,
632 sa_kill(sai, entry, false);
634 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
635 LASSERT(agl_list_empty(sai));
638 atomic_dec(&sbi->ll_sa_running);
642 /* Do NOT forget to drop inode refcount when into sai_agls. */
643 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
645 struct ll_inode_info *lli = ll_i2info(inode);
646 u64 index = lli->lli_agl_index;
652 LASSERT(list_empty(&lli->lli_agl_list));
654 /* AGL maybe fall behind statahead with one entry */
655 if (is_omitted_entry(sai, index + 1)) {
656 lli->lli_agl_index = 0;
662 * In case of restore, the MDT has the right size and has already
663 * sent it back without granting the layout lock, inode is up-to-date.
664 * Then AGL (async glimpse lock) is useless.
665 * Also to glimpse we need the layout, in case of a runninh restore
666 * the MDT holds the layout lock so the glimpse will block up to the
667 * end of restore (statahead/agl will block)
669 if (test_bit(LLIF_FILE_RESTORING, &lli->lli_flags)) {
670 lli->lli_agl_index = 0;
675 /* Someone is in glimpse (sync or async), do nothing. */
676 rc = down_write_trylock(&lli->lli_glimpse_sem);
678 lli->lli_agl_index = 0;
684 * Someone triggered glimpse within 1 sec before.
685 * 1) The former glimpse succeeded with glimpse lock granted by OST, and
686 * if the lock is still cached on client, AGL needs to do nothing. If
687 * it is cancelled by other client, AGL maybe cannot obtaion new lock
688 * for no glimpse callback triggered by AGL.
689 * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
690 * Under such case, it is quite possible that the OST will not grant
691 * glimpse lock for AGL also.
692 * 3) The former glimpse failed, compared with other two cases, it is
693 * relative rare. AGL can ignore such case, and it will not muchly
694 * affect the performance.
696 expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
697 if (ktime_to_ns(lli->lli_glimpse_time) &&
698 ktime_before(expire, lli->lli_glimpse_time)) {
699 up_write(&lli->lli_glimpse_sem);
700 lli->lli_agl_index = 0;
706 "Handling (init) async glimpse: inode = " DFID", idx = %llu\n",
707 PFID(&lli->lli_fid), index);
710 lli->lli_agl_index = 0;
711 lli->lli_glimpse_time = ktime_get();
712 up_write(&lli->lli_glimpse_sem);
715 "Handled (init) async glimpse: inode= " DFID", idx = %llu, rc = %d\n",
716 PFID(&lli->lli_fid), index, rc);
723 static void ll_statahead_interpret_fini(struct ll_inode_info *lli,
724 struct ll_statahead_info *sai,
725 struct md_op_item *item,
726 struct sa_entry *entry,
727 struct ptlrpc_request *req,
731 * First it will drop ldlm ibits lock refcount by calling
732 * ll_intent_drop_lock() in spite of failures. Do not worry about
733 * calling ll_intent_drop_lock() more than once.
735 ll_intent_release(&item->mop_it);
738 ptlrpc_req_finished(req);
739 sa_make_ready(sai, entry, rc);
741 spin_lock(&lli->lli_sa_lock);
743 spin_unlock(&lli->lli_sa_lock);
746 static void ll_statahead_interpret_work(struct work_struct *work)
748 struct md_op_item *item = container_of(work, struct md_op_item,
750 struct req_capsule *pill = item->mop_pill;
751 struct inode *dir = item->mop_dir;
752 struct ll_inode_info *lli = ll_i2info(dir);
753 struct ll_statahead_info *sai;
754 struct lookup_intent *it;
755 struct sa_entry *entry;
756 struct mdt_body *body;
762 entry = (struct sa_entry *)item->mop_cbdata;
763 LASSERT(entry->se_handle != 0);
767 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
769 GOTO(out, rc = -EFAULT);
771 child = entry->se_inode;
772 /* revalidate; unlinked and re-created with the same name */
773 if (unlikely(!fid_is_zero(&item->mop_data.op_fid2) &&
774 !lu_fid_eq(&item->mop_data.op_fid2, &body->mbo_fid1))) {
776 entry->se_inode = NULL;
779 /* The mdt_body is invalid. Skip this entry */
780 GOTO(out, rc = -EAGAIN);
783 it->it_lock_handle = entry->se_handle;
784 rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
786 GOTO(out, rc = -EAGAIN);
788 rc = ll_prep_inode(&child, pill, dir->i_sb, it);
790 CERROR("%s: getattr callback for %.*s "DFID": rc = %d\n",
791 ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
792 entry->se_qstr.name, PFID(&entry->se_fid), rc);
796 /* If encryption context was returned by MDT, put it in
797 * inode now to save an extra getxattr.
799 if (body->mbo_valid & OBD_MD_ENCCTX) {
800 void *encctx = req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
801 __u32 encctxlen = req_capsule_get_size(pill, &RMF_FILE_ENCCTX,
806 "server returned encryption ctx for "DFID"\n",
807 PFID(ll_inode2fid(child)));
808 rc = ll_xattr_cache_insert(child,
809 xattr_for_enc(child),
812 CWARN("%s: cannot set enc ctx for "DFID": rc = %d\n",
813 ll_i2sbi(child)->ll_fsname,
814 PFID(ll_inode2fid(child)), rc);
818 CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
819 ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
820 entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
821 ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
823 entry->se_inode = child;
825 if (agl_should_run(sai, child))
826 ll_agl_add(sai, child, entry->se_index);
828 ll_statahead_interpret_fini(lli, sai, item, entry, pill->rc_req, rc);
832 * Callback for async stat RPC, this is called in ptlrpcd context. It prepares
833 * the inode and set lock data directly in the ptlrpcd context. It will wake up
834 * the directory listing process if the dentry is the waiting one.
836 static int ll_statahead_interpret(struct md_op_item *item, int rc)
838 struct req_capsule *pill = item->mop_pill;
839 struct lookup_intent *it = &item->mop_it;
840 struct inode *dir = item->mop_dir;
841 struct ll_inode_info *lli = ll_i2info(dir);
842 struct sa_entry *entry = (struct sa_entry *)item->mop_cbdata;
843 struct work_struct *work = &item->mop_work;
844 struct ll_statahead_info *sai;
845 struct mdt_body *body;
851 if (it_disposition(it, DISP_LOOKUP_NEG))
855 * because statahead thread will wait for all inflight RPC to finish,
856 * sai should be always valid, no need to refcount
858 LASSERT(entry != NULL);
860 LASSERT(sai != NULL);
862 CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
863 entry->se_qstr.len, entry->se_qstr.name, rc);
868 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
870 GOTO(out, rc = -EFAULT);
872 child = entry->se_inode;
874 * revalidate; unlinked and re-created with the same name.
875 * exclude the case where FID is zero as it was from statahead with
876 * regularized file name pattern and had no idea for the FID of the
879 if (unlikely(!fid_is_zero(&item->mop_data.op_fid2) &&
880 !lu_fid_eq(&item->mop_data.op_fid2, &body->mbo_fid1))) {
882 entry->se_inode = NULL;
885 /* The mdt_body is invalid. Skip this entry */
886 GOTO(out, rc = -EAGAIN);
889 entry->se_handle = it->it_lock_handle;
891 * In ptlrpcd context, it is not allowed to generate new RPCs
892 * especially for striped directories or regular files with layout
896 * release ibits lock ASAP to avoid deadlock when statahead
897 * thread enqueues lock on parent in readdir and another
898 * process enqueues lock on child with parent lock held, eg.
901 handle = it->it_lock_handle;
902 ll_intent_drop_lock(it);
903 ll_unlock_md_op_lsm(&item->mop_data);
906 * If the statahead entry is a striped directory or regular file with
907 * layout change, it will generate a new RPC and long wait in the
909 * However, it is dangerous of blocking in ptlrpcd thread.
910 * Here we use work queue or the separate statahead thread to handle
911 * the extra RPC and long wait:
912 * (@ll_prep_inode->@lmv_revalidate_slaves);
913 * (@ll_prep_inode->@lov_layout_change->osc_cache_wait_range);
915 INIT_WORK(work, ll_statahead_interpret_work);
916 ptlrpc_request_addref(pill->rc_req);
920 ll_statahead_interpret_fini(lli, sai, item, entry, NULL, rc);
924 static inline int sa_getattr(struct ll_statahead_info *sai, struct inode *dir,
925 struct md_op_item *item)
929 if (sa_has_batch_handle(sai))
930 rc = md_batch_add(ll_i2mdexp(dir), sai->sai_bh, item);
932 rc = md_intent_getattr_async(ll_i2mdexp(dir), item);
937 /* async stat for file not found in dcache */
938 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
940 struct md_op_item *item;
945 item = sa_prep_data(dir, NULL, entry);
947 RETURN(PTR_ERR(item));
949 rc = sa_getattr(entry->se_sai, dir, item);
957 * async stat for file found in dcache, similar to .revalidate
959 * \retval 1 dentry valid, no RPC sent
960 * \retval 0 dentry invalid, will send async stat RPC
961 * \retval negative number upon error
963 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
964 struct dentry *dentry)
966 struct inode *inode = dentry->d_inode;
967 struct lookup_intent it = { .it_op = IT_GETATTR,
968 .it_lock_handle = 0 };
969 struct md_op_item *item;
974 if (unlikely(!inode))
977 if (d_mountpoint(dentry))
980 item = sa_prep_data(dir, inode, entry);
982 RETURN(PTR_ERR(item));
984 entry->se_inode = igrab(inode);
985 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
988 entry->se_handle = it.it_lock_handle;
989 ll_intent_release(&it);
994 rc = sa_getattr(entry->se_sai, dir, item);
996 entry->se_inode = NULL;
1004 /* async stat for file with @name */
1005 static void sa_statahead(struct ll_statahead_info *sai, struct dentry *parent,
1006 const char *name, int len, const struct lu_fid *fid)
1008 struct inode *dir = parent->d_inode;
1009 struct dentry *dentry = NULL;
1010 struct sa_entry *entry;
1015 entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
1019 dentry = d_lookup(parent, &entry->se_qstr);
1021 rc = sa_lookup(dir, entry);
1023 rc = sa_revalidate(dir, entry, dentry);
1024 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
1025 ll_agl_add(sai, dentry->d_inode, entry->se_index);
1032 sa_make_ready(sai, entry, rc);
1038 if (sa_sent_full(sai))
1039 ll_statahead_flush_nowait(sai);
1044 /* async glimpse (agl) thread main function */
1045 static int ll_agl_thread(void *arg)
1048 * We already own this reference, so it is safe to take it
1051 struct ll_statahead_info *sai = (struct ll_statahead_info *)arg;
1052 struct dentry *parent = sai->sai_dentry;
1053 struct inode *dir = parent->d_inode;
1054 struct ll_inode_info *plli = ll_i2info(dir);
1055 struct ll_inode_info *clli;
1059 CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
1062 while (({set_current_state(TASK_IDLE);
1063 !kthread_should_stop(); })) {
1064 spin_lock(&plli->lli_agl_lock);
1065 clli = list_first_entry_or_null(&sai->sai_agls,
1066 struct ll_inode_info,
1069 __set_current_state(TASK_RUNNING);
1070 list_del_init(&clli->lli_agl_list);
1071 spin_unlock(&plli->lli_agl_lock);
1072 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1075 spin_unlock(&plli->lli_agl_lock);
1079 __set_current_state(TASK_RUNNING);
1083 static void ll_stop_agl(struct ll_statahead_info *sai)
1085 struct dentry *parent = sai->sai_dentry;
1086 struct ll_inode_info *plli = ll_i2info(parent->d_inode);
1087 struct ll_inode_info *clli;
1088 struct task_struct *agl_task;
1090 spin_lock(&plli->lli_agl_lock);
1091 agl_task = sai->sai_agl_task;
1092 sai->sai_agl_task = NULL;
1093 spin_unlock(&plli->lli_agl_lock);
1097 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1098 sai, (unsigned int)agl_task->pid);
1099 kthread_stop(agl_task);
1101 spin_lock(&plli->lli_agl_lock);
1102 while ((clli = list_first_entry_or_null(&sai->sai_agls,
1103 struct ll_inode_info,
1104 lli_agl_list)) != NULL) {
1105 list_del_init(&clli->lli_agl_list);
1106 spin_unlock(&plli->lli_agl_lock);
1107 clli->lli_agl_index = 0;
1108 iput(&clli->lli_vfs_inode);
1109 spin_lock(&plli->lli_agl_lock);
1111 spin_unlock(&plli->lli_agl_lock);
1112 CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
1117 /* start agl thread */
1118 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1120 int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
1121 struct ll_inode_info *plli;
1122 struct task_struct *task;
1126 CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
1129 plli = ll_i2info(parent->d_inode);
1130 task = kthread_create_on_node(ll_agl_thread, sai, node, "ll_agl_%d",
1131 plli->lli_opendir_pid);
1133 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1136 sai->sai_agl_task = task;
1137 atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
1138 /* Get an extra reference that the thread holds */
1141 wake_up_process(task);
1146 static int ll_statahead_by_list(struct dentry *parent)
1148 struct inode *dir = parent->d_inode;
1149 struct ll_inode_info *lli = ll_i2info(dir);
1150 struct ll_statahead_info *sai = lli->lli_sai;
1151 struct ll_sb_info *sbi = ll_i2sbi(dir);
1152 struct md_op_data *op_data;
1153 struct page *page = NULL;
1160 CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1163 OBD_ALLOC_PTR(op_data);
1167 while (pos != MDS_DIR_END_OFF &&
1168 /* matches smp_store_release() in ll_deauthorize_statahead() */
1169 smp_load_acquire(&sai->sai_task) &&
1170 lli->lli_sa_enabled) {
1171 struct lu_dirpage *dp;
1172 struct lu_dirent *ent;
1174 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1175 LUSTRE_OPC_ANY, dir);
1176 if (IS_ERR(op_data)) {
1177 rc = PTR_ERR(op_data);
1181 page = ll_get_dir_page(dir, op_data, pos, NULL);
1182 ll_unlock_md_op_lsm(op_data);
1186 "error reading dir "DFID" at %llu /%llu opendir_pid = %u: rc = %d\n",
1187 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1188 lli->lli_opendir_pid, rc);
1192 dp = page_address(page);
1193 for (ent = lu_dirent_start(dp);
1194 /* matches smp_store_release() in ll_deauthorize_statahead() */
1195 ent != NULL && smp_load_acquire(&sai->sai_task) &&
1196 !sa_low_hit(sai) && lli->lli_sa_enabled;
1197 ent = lu_dirent_next(ent)) {
1202 struct llcrypt_str lltr = LLTR_INIT(NULL, 0);
1204 hash = le64_to_cpu(ent->lde_hash);
1205 if (unlikely(hash < pos))
1207 * Skip until we find target hash value.
1211 namelen = le16_to_cpu(ent->lde_namelen);
1212 if (unlikely(namelen == 0))
1214 * Skip dummy record.
1218 name = ent->lde_name;
1219 if (name[0] == '.') {
1225 } else if (name[1] == '.' && namelen == 2) {
1230 } else if (!sai->sai_ls_all) {
1232 * skip hidden files.
1234 sai->sai_skip_hidden++;
1240 * don't stat-ahead first entry.
1242 if (unlikely(++first == 1))
1245 fid_le_to_cpu(&fid, &ent->lde_fid);
1247 while (({set_current_state(TASK_IDLE);
1248 /* matches smp_store_release() in
1249 * ll_deauthorize_statahead()
1251 smp_load_acquire(&sai->sai_task); })) {
1254 spin_lock(&lli->lli_agl_lock);
1255 while (sa_sent_full(sai) &&
1256 !agl_list_empty(sai)) {
1257 struct ll_inode_info *clli;
1259 __set_current_state(TASK_RUNNING);
1260 clli = agl_first_entry(sai);
1261 list_del_init(&clli->lli_agl_list);
1262 spin_unlock(&lli->lli_agl_lock);
1264 ll_agl_trigger(&clli->lli_vfs_inode,
1267 spin_lock(&lli->lli_agl_lock);
1269 spin_unlock(&lli->lli_agl_lock);
1271 if (!sa_sent_full(sai))
1275 * If the thread is not doing stat in
1276 * @sbi->ll_sa_timeout (30s) then it probably
1277 * does not care too much about performance,
1278 * or is no longer using this directory.
1279 * Stop the statahead thread in this case.
1281 timeout = schedule_timeout(
1282 cfs_time_seconds(sbi->ll_sa_timeout));
1284 lli->lli_sa_enabled = 0;
1288 __set_current_state(TASK_RUNNING);
1290 if (IS_ENCRYPTED(dir)) {
1291 struct llcrypt_str de_name =
1292 LLTR_INIT(ent->lde_name, namelen);
1295 rc = llcrypt_fname_alloc_buffer(dir, NAME_MAX,
1300 fid_le_to_cpu(&fid, &ent->lde_fid);
1301 if (ll_fname_disk_to_usr(dir, 0, 0, &de_name,
1303 llcrypt_fname_free_buffer(&lltr);
1311 sa_statahead(sai, parent, name, namelen, &fid);
1312 llcrypt_fname_free_buffer(&lltr);
1315 pos = le64_to_cpu(dp->ldp_hash_end);
1316 ll_release_page(dir, page,
1317 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1319 if (sa_low_hit(sai)) {
1321 atomic_inc(&sbi->ll_sa_wrong);
1323 "Statahead for dir "DFID" hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stoppingstatahead thread: pid %d\n",
1324 PFID(&lli->lli_fid), sai->sai_hit,
1325 sai->sai_miss, sai->sai_sent,
1326 sai->sai_replied, current->pid);
1330 ll_finish_md_op_data(op_data);
1335 static void ll_statahead_handle(struct ll_statahead_info *sai,
1336 struct dentry *parent, const char *name,
1337 int len, const struct lu_fid *fid)
1339 struct inode *dir = parent->d_inode;
1340 struct ll_inode_info *lli = ll_i2info(dir);
1341 struct ll_sb_info *sbi = ll_i2sbi(dir);
1344 while (({set_current_state(TASK_IDLE);
1345 /* matches smp_store_release() in ll_deauthorize_statahead() */
1346 smp_load_acquire(&sai->sai_task); })) {
1347 spin_lock(&lli->lli_agl_lock);
1348 while (sa_sent_full(sai) && !agl_list_empty(sai)) {
1349 struct ll_inode_info *clli;
1351 __set_current_state(TASK_RUNNING);
1352 clli = agl_first_entry(sai);
1353 list_del_init(&clli->lli_agl_list);
1354 spin_unlock(&lli->lli_agl_lock);
1356 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1358 spin_lock(&lli->lli_agl_lock);
1360 spin_unlock(&lli->lli_agl_lock);
1362 if (!sa_sent_full(sai))
1366 * If the thread is not doing a stat in 30s then it probably
1367 * does not care too much about performance, or is no longer
1368 * using this directory. Stop the statahead thread in this case.
1370 timeout = schedule_timeout(
1371 cfs_time_seconds(sbi->ll_sa_timeout));
1373 lli->lli_sa_enabled = 0;
1377 __set_current_state(TASK_RUNNING);
1379 sa_statahead(sai, parent, name, len, fid);
1382 static int ll_statahead_by_advise(struct ll_statahead_info *sai,
1383 struct dentry *parent)
1385 struct inode *dir = parent->d_inode;
1386 struct ll_inode_info *lli = ll_i2info(dir);
1387 struct ll_sb_info *sbi = ll_i2sbi(dir);
1397 CDEBUG(D_READA, "%s: ADVISE statahead: parent %pd fname prefix %s\n",
1398 sbi->ll_fsname, parent, sai->sai_fname);
1400 OBD_ALLOC(fname, NAME_MAX);
1404 len = strlen(sai->sai_fname);
1405 memcpy(fname, sai->sai_fname, len);
1406 max_len = sizeof(sai->sai_fname) - len;
1409 /* matches smp_store_release() in ll_deauthorize_statahead() */
1410 while (smp_load_acquire(&sai->sai_task) && lli->lli_sa_enabled) {
1413 numlen = snprintf(ptr, max_len, "%llu",
1414 sai->sai_fstart + i);
1416 ll_statahead_handle(sai, parent, fname, len + numlen, NULL);
1417 if (++i >= sai->sai_fend)
1421 OBD_FREE(fname, NAME_MAX);
1425 static int ll_statahead_by_fname(struct ll_statahead_info *sai,
1426 struct dentry *parent)
1428 struct inode *dir = parent->d_inode;
1429 struct ll_inode_info *lli = ll_i2info(dir);
1430 struct ll_sb_info *sbi = ll_i2sbi(dir);
1439 CDEBUG(D_READA, "%s: FNAME statahead: parent %pd fname prefix %s\n",
1440 sbi->ll_fsname, parent, sai->sai_fname);
1442 OBD_ALLOC(fname, NAME_MAX);
1446 len = strlen(sai->sai_fname);
1447 memcpy(fname, sai->sai_fname, len);
1448 max_len = sizeof(sai->sai_fname) - len;
1451 /* matches smp_store_release() in ll_deauthorize_statahead() */
1452 while (smp_load_acquire(&sai->sai_task) && lli->lli_sa_enabled) {
1455 if (sai->sai_fname_zeroed_len)
1456 numlen = snprintf(ptr, max_len, "%0*llu",
1457 sai->sai_fname_zeroed_len,
1458 ++sai->sai_fname_index);
1460 numlen = snprintf(ptr, max_len, "%llu",
1461 ++sai->sai_fname_index);
1463 ll_statahead_handle(sai, parent, fname, len + numlen, NULL);
1465 if (sa_low_hit(sai)) {
1467 atomic_inc(&sbi->ll_sa_wrong);
1468 CDEBUG(D_CACHE, "%s: low hit ratio for %pd "DFID": hit=%llu miss=%llu sent=%llu replied=%llu, stopping PID %d\n",
1469 sbi->ll_fsname, parent, PFID(ll_inode2fid(dir)),
1470 sai->sai_hit, sai->sai_miss, sai->sai_sent,
1471 sai->sai_replied, current->pid);
1476 OBD_FREE(fname, NAME_MAX);
1480 /* statahead thread main function */
1481 static int ll_statahead_thread(void *arg)
1483 struct ll_statahead_info *sai = (struct ll_statahead_info *)arg;
1484 struct dentry *parent = sai->sai_dentry;
1485 struct inode *dir = parent->d_inode;
1486 struct ll_inode_info *lli = ll_i2info(dir);
1487 struct ll_sb_info *sbi = ll_i2sbi(dir);
1488 struct lu_batch *bh = NULL;
1493 CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1496 sai->sai_max_batch_count = sbi->ll_sa_batch_max;
1497 if (sai->sai_max_batch_count) {
1498 bh = md_batch_create(ll_i2mdexp(dir), BATCH_FL_RDONLY,
1499 sai->sai_max_batch_count);
1501 GOTO(out_stop_agl, rc = PTR_ERR(bh));
1506 switch (lli->lli_sa_pattern & LSA_PATTERN_MASK) {
1507 case LSA_PATTERN_LIST:
1508 rc = ll_statahead_by_list(parent);
1510 case LSA_PATTERN_ADVISE:
1511 rc = ll_statahead_by_advise(sai, parent);
1513 case LSA_PATTERN_FNAME:
1514 rc = ll_statahead_by_fname(sai, parent);
1522 spin_lock(&lli->lli_sa_lock);
1523 sai->sai_task = NULL;
1524 lli->lli_sa_enabled = 0;
1525 spin_unlock(&lli->lli_sa_lock);
1528 ll_statahead_flush_nowait(sai);
1531 * statahead is finished, but statahead entries need to be cached, wait
1532 * for file release closedir() call to stop me.
1534 while (({set_current_state(TASK_IDLE);
1535 /* matches smp_store_release() in ll_deauthorize_statahead() */
1536 smp_load_acquire(&sai->sai_task) && lli->lli_sa_enabled; })) {
1539 __set_current_state(TASK_RUNNING);
1544 rc = md_batch_stop(ll_i2mdexp(dir), sai->sai_bh);
1552 * wait for inflight statahead RPCs to finish, and then we can free sai
1553 * safely because statahead RPC will access sai data
1555 while (sai->sai_sent != sai->sai_replied)
1556 /* in case we're not woken up, timeout wait */
1559 CDEBUG(D_READA, "%s: statahead thread stopped: sai %p, parent %pd hit %llu miss %llu\n",
1560 sbi->ll_fsname, sai, parent, sai->sai_hit, sai->sai_miss);
1562 spin_lock(&lli->lli_sa_lock);
1563 sai->sai_task = NULL;
1564 spin_unlock(&lli->lli_sa_lock);
1565 wake_up(&sai->sai_waitq);
1567 atomic_add(sai->sai_hit, &sbi->ll_sa_hit_total);
1568 atomic_add(sai->sai_miss, &sbi->ll_sa_miss_total);
1571 ll_sax_put(dir, lli->lli_sax);
1576 /* authorize opened dir handle @key to statahead */
1577 void ll_authorize_statahead(struct inode *dir, void *key)
1579 struct ll_inode_info *lli = ll_i2info(dir);
1581 spin_lock(&lli->lli_sa_lock);
1582 if (!lli->lli_opendir_key && !lli->lli_sai) {
1584 * if lli_sai is not NULL, it means previous statahead is not
1585 * finished yet, we'd better not start a new statahead for now.
1587 lli->lli_opendir_key = key;
1588 lli->lli_opendir_pid = current->pid;
1589 lli->lli_sa_enabled = 1;
1591 spin_unlock(&lli->lli_sa_lock);
1594 static void ll_deauthorize_statahead_advise(struct inode *dir, void *key)
1596 struct ll_inode_info *lli = ll_i2info(dir);
1597 struct ll_file_data *fd = (struct ll_file_data *)key;
1598 struct ll_statahead_info *sai = fd->fd_sai;
1603 spin_lock(&lli->lli_sa_lock);
1604 if (sai->sai_task) {
1605 struct task_struct *task = sai->sai_task;
1607 /* matches smp_load_acquire() in ll_statahead_thread() */
1608 smp_store_release(&sai->sai_task, NULL);
1609 wake_up_process(task);
1612 spin_unlock(&lli->lli_sa_lock);
1614 LASSERT(lli->lli_sax != NULL);
1615 ll_sax_put(dir, lli->lli_sax);
1619 * deauthorize opened dir handle @key to statahead, and notify statahead thread
1620 * to quit if it's running.
1622 void ll_deauthorize_statahead(struct inode *dir, void *key)
1624 struct ll_inode_info *lli = ll_i2info(dir);
1625 struct ll_statahead_info *sai;
1627 CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1628 PFID(&lli->lli_fid));
1630 if (lli->lli_sa_pattern & LSA_PATTERN_ADVISE) {
1631 ll_deauthorize_statahead_advise(dir, key);
1635 LASSERT(lli->lli_opendir_pid != 0);
1636 LASSERT(lli->lli_opendir_key == key);
1637 spin_lock(&lli->lli_sa_lock);
1638 lli->lli_opendir_key = NULL;
1639 lli->lli_opendir_pid = 0;
1640 lli->lli_sa_enabled = 0;
1641 lli->lli_sa_pattern = LSA_PATTERN_NONE;
1642 lli->lli_sa_fname_index = 0;
1643 lli->lli_sa_match_count = 0;
1645 if (sai && sai->sai_task) {
1647 * statahead thread may not have quit yet because it needs to
1648 * cache entries, now it's time to tell it to quit.
1650 * wake_up_process() provides the necessary barriers
1651 * to pair with set_current_state().
1653 struct task_struct *task = sai->sai_task;
1655 /* matches smp_load_acquire() in ll_statahead_thread() */
1656 smp_store_release(&sai->sai_task, NULL);
1657 wake_up_process(task);
1659 spin_unlock(&lli->lli_sa_lock);
1664 * not first dirent, or is "."
1666 LS_NOT_FIRST_DE = 0,
1668 * the first non-hidden dirent
1672 * the first hidden dirent, that is "."
1677 /* file is first dirent under @dir */
1678 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1680 struct qstr *target = &dentry->d_name;
1681 struct md_op_data *op_data;
1683 struct page *page = NULL;
1684 int rc = LS_NOT_FIRST_DE;
1686 struct llcrypt_str lltr = LLTR_INIT(NULL, 0);
1690 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1691 LUSTRE_OPC_ANY, dir);
1692 if (IS_ERR(op_data))
1693 RETURN(PTR_ERR(op_data));
1695 if (IS_ENCRYPTED(dir)) {
1696 int rc2 = llcrypt_fname_alloc_buffer(dir, NAME_MAX, &lltr);
1703 *FIXME choose the start offset of the readdir
1706 page = ll_get_dir_page(dir, op_data, 0, NULL);
1709 struct lu_dirpage *dp;
1710 struct lu_dirent *ent;
1713 struct ll_inode_info *lli = ll_i2info(dir);
1716 CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
1717 ll_i2sbi(dir)->ll_fsname,
1718 PFID(ll_inode2fid(dir)), pos,
1719 lli->lli_opendir_pid, rc);
1723 dp = page_address(page);
1724 for (ent = lu_dirent_start(dp); ent != NULL;
1725 ent = lu_dirent_next(ent)) {
1730 hash = le64_to_cpu(ent->lde_hash);
1732 * The ll_get_dir_page() can return any page containing
1733 * the given hash which may be not the start hash.
1735 if (unlikely(hash < pos))
1738 namelen = le16_to_cpu(ent->lde_namelen);
1739 if (unlikely(namelen == 0))
1741 * skip dummy record.
1745 name = ent->lde_name;
1746 if (name[0] == '.') {
1752 else if (name[1] == '.' && namelen == 2)
1763 if (dot_de && target->name[0] != '.') {
1764 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1765 target->len, target->name,
1770 if (IS_ENCRYPTED(dir)) {
1771 struct llcrypt_str de_name =
1772 LLTR_INIT(ent->lde_name, namelen);
1775 fid_le_to_cpu(&fid, &ent->lde_fid);
1776 if (ll_fname_disk_to_usr(dir, 0, 0, &de_name,
1783 if (target->len != namelen ||
1784 memcmp(target->name, name, namelen) != 0)
1785 rc = LS_NOT_FIRST_DE;
1789 rc = LS_FIRST_DOT_DE;
1791 ll_release_page(dir, page, false);
1794 pos = le64_to_cpu(dp->ldp_hash_end);
1795 if (pos == MDS_DIR_END_OFF) {
1797 * End of directory reached.
1799 ll_release_page(dir, page, false);
1803 * chain is exhausted
1804 * Normal case: continue to the next page.
1806 ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1808 page = ll_get_dir_page(dir, op_data, pos, NULL);
1813 llcrypt_fname_free_buffer(&lltr);
1814 ll_finish_md_op_data(op_data);
1820 * revalidate @dentryp from statahead cache
1822 * \param[in] dir parent directory
1823 * \param[in] sai sai structure
1824 * \param[out] dentryp pointer to dentry which will be revalidated
1825 * \param[in] unplug unplug statahead window only (normally for negative
1827 * \retval 1 on success, dentry is saved in @dentryp
1828 * \retval 0 if revalidation failed (no proper lock on client)
1829 * \retval negative number upon error
1831 static int revalidate_statahead_dentry(struct inode *dir,
1832 struct ll_statahead_context *ctx,
1833 struct dentry **dentryp,
1836 struct sa_entry *entry = NULL;
1837 struct ll_inode_info *lli = ll_i2info(dir);
1838 struct ll_statahead_info *sai = lli->lli_sai;
1843 if (sai && (*dentryp)->d_name.name[0] == '.') {
1844 if (sai->sai_ls_all ||
1845 sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1847 * Hidden dentry is the first one, or statahead
1848 * thread does not skip so many hidden dentries
1849 * before "sai_ls_all" enabled as below.
1852 if (!sai->sai_ls_all)
1854 * It maybe because hidden dentry is not
1855 * the first one, "sai_ls_all" was not
1856 * set, then "ls -al" missed. Enable
1857 * "sai_ls_all" for such case.
1859 sai->sai_ls_all = 1;
1862 * Such "getattr" has been skipped before
1863 * "sai_ls_all" enabled as above.
1865 sai->sai_miss_hidden++;
1873 entry = sa_get(ctx, &(*dentryp)->d_name);
1875 GOTO(out, rc = -EAGAIN);
1877 if (lli->lli_sa_pattern & LSA_PATTERN_LIST ||
1878 lli->lli_sa_pattern & LSA_PATTERN_FNAME)
1879 LASSERT(sai == entry->se_sai);
1880 else if (lli->lli_sa_pattern == LSA_PATTERN_ADVISE)
1881 sai = entry->se_sai;
1883 LASSERT(sai != NULL);
1884 if (!sa_ready(entry)) {
1885 spin_lock(&lli->lli_sa_lock);
1886 sai->sai_index_wait = entry->se_index;
1887 spin_unlock(&lli->lli_sa_lock);
1888 rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
1889 cfs_time_seconds(30));
1892 * entry may not be ready, so it may be used by inflight
1893 * statahead RPC, don't free it.
1896 GOTO(out, rc = -EAGAIN);
1901 * We need to see the value that was set immediately before we
1904 if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1906 struct inode *inode = entry->se_inode;
1907 struct lookup_intent it = { .it_op = IT_GETATTR,
1912 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1913 ll_inode2fid(inode), &bits);
1915 if (!(*dentryp)->d_inode) {
1916 struct dentry *alias;
1918 alias = ll_splice_alias(inode, *dentryp);
1919 if (IS_ERR(alias)) {
1920 ll_intent_release(&it);
1921 GOTO(out, rc = PTR_ERR(alias));
1925 * statahead prepared this inode, transfer inode
1926 * refcount from sa_entry to dentry
1928 entry->se_inode = NULL;
1929 } else if ((*dentryp)->d_inode != inode) {
1930 /* revalidate, but inode is recreated */
1932 "%s: stale dentry %pd inode " DFID", statahead inode "DFID "\n",
1933 ll_i2sbi(inode)->ll_fsname, *dentryp,
1934 PFID(ll_inode2fid((*dentryp)->d_inode)),
1935 PFID(ll_inode2fid(inode)));
1936 ll_intent_release(&it);
1937 GOTO(out, rc = -ESTALE);
1940 if (bits & MDS_INODELOCK_LOOKUP) {
1941 d_lustre_revalidate(*dentryp);
1942 if (S_ISDIR(inode->i_mode))
1943 ll_update_dir_depth_dmv(dir, *dentryp);
1946 ll_intent_release(&it);
1951 * statahead cached sa_entry can be used only once, and will be killed
1952 * right after use, so if lookup/revalidate accessed statahead cache,
1953 * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1954 * stat this file again, we know we've done statahead before, see
1955 * dentry_may_statahead().
1957 if (lld_is_init(*dentryp))
1958 ll_d2d(*dentryp)->lld_sa_generation = lli->lli_sa_generation;
1959 sa_put(dir, sai, entry);
1965 sa_pattern_list_detect(struct inode *dir, struct dentry *dchild, int *first)
1967 struct ll_inode_info *lli = ll_i2info(dir);
1969 if (lli->lli_opendir_pid == 0)
1972 if (lli->lli_sa_enabled == 0)
1975 if (lli->lli_sa_pattern & LSA_PATTERN_LS_NOT_FIRST_DE)
1978 *first = is_first_dirent(dir, dchild);
1979 if (*first == LS_NOT_FIRST_DE) {
1981 * It is not "ls -{a}l" operation, no need statahead for it.
1982 * Disable statahead so that subsequent stat() won't waste
1985 spin_lock(&lli->lli_sa_lock);
1986 if (lli->lli_opendir_pid == current->pid) {
1987 lli->lli_sa_enabled = 0;
1988 lli->lli_sa_pattern |= LSA_PATTERN_LS_NOT_FIRST_DE;
1990 spin_unlock(&lli->lli_sa_lock);
1994 spin_lock(&lli->lli_sa_lock);
1995 lli->lli_sa_pattern |= LSA_PATTERN_LIST;
1996 spin_unlock(&lli->lli_sa_lock);
2001 sa_pattern_fname_detect(struct inode *dir, struct dentry *dchild)
2003 struct ll_inode_info *lli = ll_i2info(dir);
2004 struct qstr *dname = &dchild->d_name;
2005 const unsigned char *name = dname->name;
2009 if (ll_i2sbi(dir)->ll_enable_statahead_fname == 0)
2013 * Parse the format of the file name to determine whether it matches
2014 * the supported file name pattern for statahead (i.e. mdtest.$i).
2017 if (isdigit(name[i])) {
2021 while (--i >= 0 && isdigit(name[i]))
2024 ret = kstrtol(&name[i], 0, &num);
2029 * The traversing program do multiple stat() calls on the same
2030 * children entry. i.e. ls $dir*.
2032 if (lli->lli_sa_fname_index == num)
2035 if (lli->lli_sa_match_count == 0 ||
2036 num == lli->lli_sa_fname_index + 1) {
2037 lli->lli_sa_match_count++;
2038 lli->lli_sa_fname_index = num;
2040 if (lli->lli_sa_match_count > LSA_FN_MATCH_HIT) {
2041 spin_lock(&lli->lli_sa_lock);
2042 lli->lli_sa_pattern |= LSA_PATTERN_FN_UNIQUE;
2043 spin_unlock(&lli->lli_sa_lock);
2044 GOTO(out, rc = true);
2051 spin_lock(&lli->lli_sa_lock);
2053 lli->lli_sa_pattern |= LSA_PATTERN_FNAME;
2055 lli->lli_sa_pattern = LSA_PATTERN_NONE;
2056 lli->lli_sa_match_count = 0;
2057 lli->lli_sa_fname_index = 0;
2058 lli->lli_sa_enabled = 0;
2060 spin_unlock(&lli->lli_sa_lock);
2065 /* detect the statahead pattern. */
2067 sa_pattern_detect(struct inode *dir, struct dentry *dchild, int *first)
2069 return sa_pattern_list_detect(dir, dchild, first) ||
2070 sa_pattern_fname_detect(dir, dchild);
2074 * start statahead thread
2076 * \param[in] dir parent directory
2077 * \param[in] dentry dentry that triggers statahead, normally the first
2079 * \param[in] agl indicate whether AGL is needed
2080 * \retval -EAGAIN on success, because when this function is
2081 * called, it's already in lookup call, so client should
2082 * do it itself instead of waiting for statahead thread
2083 * to do it asynchronously.
2084 * \retval negative number upon error
2086 static int start_statahead_thread(struct inode *dir, struct dentry *dentry,
2089 int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
2090 struct ll_inode_info *lli = ll_i2info(dir);
2091 struct ll_statahead_info *sai = NULL;
2092 struct ll_statahead_context *ctx = NULL;
2093 struct dentry *parent;
2094 struct task_struct *task;
2095 struct ll_sb_info *sbi;
2096 int first = LS_FIRST_DE;
2101 if (sa_pattern_detect(dir, dentry, &first) == false)
2104 parent = dget_parent(dentry);
2105 sbi = ll_i2sbi(d_inode(parent));
2106 if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
2107 sbi->ll_sa_running_max)) {
2109 "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
2111 GOTO(out, rc = -EMFILE);
2114 /* on success ll_sai_alloc holds a ref on parent */
2115 sai = ll_sai_alloc(parent);
2118 GOTO(out, rc = -ENOMEM);
2120 ctx = ll_sax_alloc(dir);
2122 GOTO(out, rc = -ENOMEM);
2124 sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
2126 if (lli->lli_sa_pattern & LSA_PATTERN_FNAME) {
2127 struct qstr *dname = &dentry->d_name;
2128 const unsigned char *name = dname->name;
2132 if (dname->len >= sizeof(sai->sai_fname))
2133 GOTO(out, rc = -ERANGE);
2136 while (--i >= 0 && isdigit(name[i]))
2140 memcpy(sai->sai_fname, dname->name, i);
2141 sai->sai_fname[i] = '\0';
2142 sai->sai_fname_index = lli->lli_sa_fname_index;
2143 /* The front part of the file name is zeroed padding. */
2145 sai->sai_fname_zeroed_len = dname->len - i;
2149 * if current lli_opendir_key was deauthorized, or dir re-opened by
2150 * another process, don't start statahead, otherwise the newly spawned
2151 * statahead thread won't be notified to quit.
2153 spin_lock(&lli->lli_sa_lock);
2154 if (unlikely(lli->lli_sai ||
2155 ((lli->lli_sa_pattern & LSA_PATTERN_LIST) &&
2156 !lli->lli_opendir_key &&
2157 lli->lli_opendir_pid != current->pid))) {
2158 spin_unlock(&lli->lli_sa_lock);
2159 GOTO(out, rc = -EPERM);
2163 spin_unlock(&lli->lli_sa_lock);
2165 CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
2166 current->pid, parent);
2168 task = kthread_create_on_node(ll_statahead_thread, sai, node,
2169 "ll_sa_%u", lli->lli_opendir_pid);
2171 spin_lock(&lli->lli_sa_lock);
2172 lli->lli_sai = NULL;
2173 spin_unlock(&lli->lli_sa_lock);
2175 CERROR("can't start ll_sa thread, rc: %d\n", rc);
2179 if (test_bit(LL_SBI_AGL_ENABLED, sbi->ll_flags) && agl)
2180 ll_start_agl(parent, sai);
2182 atomic_inc(&sbi->ll_sa_total);
2183 if (lli->lli_sa_pattern & LSA_PATTERN_LIST)
2184 atomic_inc(&sbi->ll_sa_list_total);
2185 else if (lli->lli_sa_pattern & LSA_PATTERN_FNAME)
2186 atomic_inc(&sbi->ll_sa_fname_total);
2188 sai->sai_task = task;
2189 wake_up_process(task);
2191 * We don't stat-ahead for the first dirent since we are already in
2198 * once we start statahead thread failed, disable statahead so that
2199 * subsequent stat won't waste time to try it.
2201 spin_lock(&lli->lli_sa_lock);
2202 if (lli->lli_opendir_pid == current->pid)
2203 lli->lli_sa_enabled = 0;
2204 spin_unlock(&lli->lli_sa_lock);
2212 if (first != LS_NOT_FIRST_DE)
2213 atomic_dec(&sbi->ll_sa_running);
2219 * Check whether statahead for @dir was started.
2221 static inline bool ll_statahead_started(struct inode *dir, bool agl)
2223 struct ll_inode_info *lli = ll_i2info(dir);
2224 struct ll_statahead_context *ctx;
2225 struct ll_statahead_info *sai;
2227 spin_lock(&lli->lli_sa_lock);
2230 if (sai && (sai->sai_agl_task != NULL) != agl)
2232 "%s: Statahead AGL hint changed from %d to %d\n",
2233 ll_i2sbi(dir)->ll_fsname,
2234 sai->sai_agl_task != NULL, agl);
2235 spin_unlock(&lli->lli_sa_lock);
2241 * statahead entry function, this is called when client getattr on a file, it
2242 * will start statahead thread if this is the first dir entry, else revalidate
2243 * dentry from statahead cache.
2245 * \param[in] dir parent directory
2246 * \param[out] dentryp dentry to getattr
2247 * \param[in] agl whether start the agl thread
2249 * \retval 1 on success
2250 * \retval 0 revalidation from statahead cache failed, caller needs
2251 * to getattr from server directly
2252 * \retval negative number on error, caller often ignores this and
2253 * then getattr from server
2255 int ll_start_statahead(struct inode *dir, struct dentry *dentry, bool agl)
2257 if (!ll_statahead_started(dir, agl))
2258 return start_statahead_thread(dir, dentry, agl);
2263 * revalidate dentry from statahead cache.
2265 * \param[in] dir parent directory
2266 * \param[out] dentryp dentry to getattr
2267 * \param[in] unplug unplug statahead window only (normally for negative
2269 * \retval 1 on success
2270 * \retval 0 revalidation from statahead cache failed, caller needs
2271 * to getattr from server directly
2272 * \retval negative number on error, caller often ignores this and
2273 * then getattr from server
2275 int ll_revalidate_statahead(struct inode *dir, struct dentry **dentryp,
2278 struct ll_inode_info *lli = ll_i2info(dir);
2279 struct ll_statahead_context *ctx;
2280 struct ll_statahead_info *sai = NULL;
2283 spin_lock(&lli->lli_sa_lock);
2288 atomic_inc(&sai->sai_refcount);
2289 } else if (lli->lli_sa_pattern & LSA_PATTERN_LIST) {
2290 spin_unlock(&lli->lli_sa_lock);
2295 spin_unlock(&lli->lli_sa_lock);
2297 rc = revalidate_statahead_dentry(dir, ctx, dentryp, unplug);
2298 CDEBUG(D_READA, "revalidate statahead %pd: rc = %d.\n",
2302 ll_sax_put(dir, ctx);
2307 int ll_ioctl_ahead(struct file *file, struct llapi_lu_ladvise2 *ladvise)
2309 int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
2310 struct ll_file_data *fd = file->private_data;
2311 struct dentry *dentry = file_dentry(file);
2312 struct inode *dir = dentry->d_inode;
2313 struct ll_inode_info *lli = ll_i2info(dir);
2314 struct ll_sb_info *sbi = ll_i2sbi(dir);
2315 struct ll_statahead_info *sai = NULL;
2316 struct ll_statahead_context *ctx = NULL;
2317 struct task_struct *task;
2323 if (sbi->ll_sa_max == 0)
2326 if (!S_ISDIR(dir->i_mode))
2331 CWARN("%s: already set statahead hint for dir %pd: rc = %d\n",
2332 sbi->ll_fsname, dentry, rc);
2336 if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
2337 sbi->ll_sa_running_max)) {
2339 "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
2340 GOTO(out, rc = -EMFILE);
2343 sai = ll_sai_alloc(dentry);
2345 GOTO(out, rc = -ENOMEM);
2347 sai->sai_fstart = ladvise->lla_start;
2348 sai->sai_fend = ladvise->lla_end;
2349 sai->sai_ls_all = 0;
2350 sai->sai_max = sbi->ll_sa_max;
2351 strncpy(sai->sai_fname, ladvise->lla_fname, sizeof(sai->sai_fname));
2353 ctx = ll_sax_get(dir);
2355 ctx = ll_sax_alloc(dir);
2357 GOTO(out, rc = -ENOMEM);
2359 spin_lock(&lli->lli_sa_lock);
2360 if (unlikely(lli->lli_sax)) {
2361 struct ll_statahead_context *tmp = ctx;
2363 if (lli->lli_sa_pattern == LSA_PATTERN_NONE ||
2364 lli->lli_sa_pattern == LSA_PATTERN_ADVISE) {
2365 lli->lli_sa_pattern = LSA_PATTERN_ADVISE;
2368 fd->fd_sai = __ll_sai_get(sai);
2372 CWARN("%s: pattern %X is not ADVISE: rc = %d\n",
2373 sbi->ll_fsname, lli->lli_sa_pattern, rc);
2376 spin_unlock(&lli->lli_sa_lock);
2381 lli->lli_sa_pattern = LSA_PATTERN_ADVISE;
2383 fd->fd_sai = __ll_sai_get(sai);
2384 spin_unlock(&lli->lli_sa_lock);
2387 spin_lock(&lli->lli_sa_lock);
2388 if (!(lli->lli_sa_pattern == LSA_PATTERN_ADVISE ||
2389 lli->lli_sa_pattern == LSA_PATTERN_NONE)) {
2390 spin_unlock(&lli->lli_sa_lock);
2391 GOTO(out, rc = -EINVAL);
2394 lli->lli_sa_pattern = LSA_PATTERN_ADVISE;
2395 fd->fd_sai = __ll_sai_get(sai);
2396 spin_unlock(&lli->lli_sa_lock);
2401 "start statahead thread: [pid %d] [parent %pd] sai %p ctx %p\n",
2402 current->pid, dentry, sai, ctx);
2404 task = kthread_create_on_node(ll_statahead_thread, sai, node,
2405 "ll_sa_%u", current->pid);
2408 CERROR("%s: cannot start ll_sa thread: rc = %d\n",
2409 sbi->ll_fsname, rc);
2413 if (test_bit(LL_SBI_AGL_ENABLED, sbi->ll_flags) && agl)
2414 ll_start_agl(dentry, sai);
2416 atomic_inc(&sbi->ll_sa_total);
2417 sai->sai_task = task;
2418 wake_up_process(task);
2424 ll_sax_put(dir, ctx);
2434 atomic_dec(&sbi->ll_sa_running);
2439 * This function is called in each stat() system call to do statahead check.
2440 * When the files' naming of stat() call sequence under a directory follows
2441 * a certain name rule roughly, this directory is considered as an condicant
2443 * For an example, the file naming rule is mdtest.$rank.$i, the suffix of
2444 * the stat() dentry name is number and do stat() for dentries with name
2445 * ending with number more than @LSA_FN_PREDICT_HIT, then the corresponding
2446 * directory is met the requrirement for statahead.
2448 void ll_statahead_enter(struct inode *dir, struct dentry *dchild)
2450 struct ll_inode_info *lli;
2451 struct qstr *dname = &dchild->d_name;
2453 if (ll_i2sbi(dir)->ll_sa_max == 0)
2456 if (ll_i2sbi(dir)->ll_enable_statahead_fname == 0)
2459 lli = ll_i2info(dir);
2460 if (lli->lli_sa_enabled)
2463 if (lli->lli_sa_pattern & (LSA_PATTERN_FN_PREDICT | LSA_PATTERN_LIST))
2467 * Now support number indexing regularized statahead pattern only.
2468 * Quick check whether the last character is digit.
2470 if (!isdigit(dname->name[dname->len - 1])) {
2471 lli->lli_sa_match_count = 0;
2475 lli->lli_sa_match_count++;
2476 if (lli->lli_sa_match_count > LSA_FN_PREDICT_HIT) {
2477 spin_lock(&lli->lli_sa_lock);
2478 lli->lli_sa_pattern |= LSA_PATTERN_FN_PREDICT;
2479 spin_unlock(&lli->lli_sa_lock);
2480 lli->lli_sa_enabled = 1;
2481 lli->lli_sa_match_count = 0;