X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fstatahead.c;h=2874589dc50f1a7307083c878a94641125a959d4;hp=2e7d0997ebf56bd97f0c88fdc4d1c83b4012da41;hb=30165c4024f8a1ee316fb192005d9b2a8b89d0b8;hpb=c4b74bf9e14a113e531d2d38a294a870ef3e0593 diff --git a/lustre/llite/statahead.c b/lustre/llite/statahead.c index 2e7d099..2874589 100644 --- a/lustre/llite/statahead.c +++ b/lustre/llite/statahead.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -37,485 +37,684 @@ #include #include #include -#include #include #include #define DEBUG_SUBSYSTEM S_LLITE #include -#include #include -#include #include "llite_internal.h" -struct ll_sai_entry { - struct list_head se_list; - unsigned int se_index; - int se_stat; - struct ptlrpc_request *se_req; - struct md_enqueue_info *se_minfo; +#define SA_OMITTED_ENTRY_MAX 8ULL + +typedef enum { + /** negative values are for error cases */ + SA_ENTRY_INIT = 0, /** init entry */ + SA_ENTRY_SUCC = 1, /** stat succeed */ + SA_ENTRY_INVA = 2, /** invalid entry */ +} se_state_t; + +/* sa_entry is not refcounted: statahead thread allocates it and do async stat, + * and in async stat callback ll_statahead_interpret() will add it into + * sai_cb_entries, later statahead thread will call sa_handle_callback() to + * instantiate entry and move it into sai_entries, and then only scanner process + * can access and free it. */ +struct sa_entry { + /* link into sai_cb_entries or sai_entries */ + struct list_head se_list; + /* link into sai hash table locally */ + struct list_head se_hash; + /* entry index in the sai */ + __u64 se_index; + /* low layer ldlm lock handle */ + __u64 se_handle; + /* entry status */ + se_state_t se_state; + /* entry size, contains name */ + int se_size; + /* pointer to async getattr enqueue info */ + struct md_enqueue_info *se_minfo; + /* pointer to the async getattr request */ + struct ptlrpc_request *se_req; + /* pointer to the target inode */ + struct inode *se_inode; + /* entry name */ + struct qstr se_qstr; }; -enum { - SA_ENTRY_UNSTATED = 0, - SA_ENTRY_STATED -}; +static unsigned int sai_generation = 0; +static DEFINE_SPINLOCK(sai_generation_lock); -struct dentry_operations ll_sai_d_ops = { - .d_release = ll_release, -}; +static inline int sa_unhashed(struct sa_entry *entry) +{ + return list_empty(&entry->se_hash); +} -static unsigned int sai_generation = 0; -static spinlock_t sai_generation_lock = SPIN_LOCK_UNLOCKED; +/* + * The entry only can be released by the caller, it is necessary to hold lock. + */ +static inline int sa_ready(struct sa_entry *entry) +{ + smp_rmb(); + return (entry->se_state != SA_ENTRY_INIT); +} -/** - * Check whether first entry was stated already or not. - * No need to hold lli_lock, for: - * (1) it is me that remove entry from the list - * (2) the statahead thread only add new entry to the list +static inline int sa_hash(int val) +{ + return val & LL_SA_CACHE_MASK; +} + +/* + * Insert entry to hash SA table. */ -static int ll_sai_entry_stated(struct ll_statahead_info *sai) +static inline void +sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry) { - struct ll_sai_entry *entry; - int rc = 0; - - if (!list_empty(&sai->sai_entries_stated)) { - entry = list_entry(sai->sai_entries_stated.next, - struct ll_sai_entry, se_list); - if (entry->se_index == sai->sai_index_next) - rc = 1; - } - return rc; + int i = sa_hash(entry->se_qstr.hash); + + spin_lock(&sai->sai_cache_lock[i]); + list_add_tail(&entry->se_hash, &sai->sai_cache[i]); + spin_unlock(&sai->sai_cache_lock[i]); +} + +/* + * Remove entry from SA table. + */ +static inline void +sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry) +{ + int i = sa_hash(entry->se_qstr.hash); + + spin_lock(&sai->sai_cache_lock[i]); + list_del_init(&entry->se_hash); + spin_unlock(&sai->sai_cache_lock[i]); } -static inline int sa_received_empty(struct ll_statahead_info *sai) +static inline int agl_should_run(struct ll_statahead_info *sai, + struct inode *inode) { - return list_empty(&sai->sai_entries_received); + return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid); } -static inline int sa_not_full(struct ll_statahead_info *sai) +static inline struct ll_inode_info * +agl_first_entry(struct ll_statahead_info *sai) { - return sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max; + return list_entry(sai->sai_agls.next, struct ll_inode_info, + lli_agl_list); } -static inline int sa_is_running(struct ll_statahead_info *sai) +static inline int sa_sent_full(struct ll_statahead_info *sai) { - return !!(sai->sai_thread.t_flags & SVC_RUNNING); + return atomic_read(&sai->sai_cache_count) >= sai->sai_max; } -static inline int sa_is_stopping(struct ll_statahead_info *sai) +static inline int sa_has_callback(struct ll_statahead_info *sai) { - return !!(sai->sai_thread.t_flags & SVC_STOPPING); + return !list_empty(&sai->sai_interim_entries); } -static inline int sa_is_stopped(struct ll_statahead_info *sai) +static inline int agl_list_empty(struct ll_statahead_info *sai) { - return !!(sai->sai_thread.t_flags & SVC_STOPPED); + return list_empty(&sai->sai_agls); } /** * (1) hit ratio less than 80% * or * (2) consecutive miss more than 8 + * then means low hit. */ static inline int sa_low_hit(struct ll_statahead_info *sai) { - return ((sai->sai_hit < 4 * sai->sai_miss && sai->sai_hit > 7) || + return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) || (sai->sai_consecutive_miss > 8)); } -/** - * process the deleted entry's member and free the entry. - * (1) release intent - * (2) free md_enqueue_info - * (3) drop dentry's ref count - * (4) release request's ref count +/* + * If the given index is behind of statahead window more than + * SA_OMITTED_ENTRY_MAX, then it is old. */ -static void ll_sai_entry_cleanup(struct ll_sai_entry *entry) +static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index) { - struct ptlrpc_request *req = entry->se_req; - struct md_enqueue_info *minfo = entry->se_minfo; - ENTRY; - - if (minfo) { - struct dentry *dentry = minfo->mi_dentry; - struct lookup_intent *it = &minfo->mi_it; - - entry->se_minfo = NULL; - ll_intent_release(it); - OBD_FREE_PTR(minfo); - dput(dentry); - } - if (req) { - entry->se_req = NULL; - ptlrpc_req_finished(req); - } - OBD_FREE_PTR(entry); + return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX < + sai->sai_index); +} - EXIT; +/* allocate sa_entry and add it into hash to let scanner process to find it */ +static struct sa_entry * +sa_alloc(struct ll_statahead_info *sai, __u64 index, const char *name, int len) +{ + struct ll_inode_info *lli; + struct sa_entry *entry; + int entry_size; + char *dname; + ENTRY; + + entry_size = sizeof(struct sa_entry) + (len & ~3) + 4; + OBD_ALLOC(entry, entry_size); + if (unlikely(entry == NULL)) + RETURN(ERR_PTR(-ENOMEM)); + + CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n", + len, name, entry, index); + + entry->se_index = index; + + entry->se_state = SA_ENTRY_INIT; + entry->se_size = entry_size; + dname = (char *)entry + sizeof(struct sa_entry); + memcpy(dname, name, len); + dname[len] = 0; + entry->se_qstr.hash = full_name_hash(name, len); + entry->se_qstr.len = len; + entry->se_qstr.name = dname; + + lli = ll_i2info(sai->sai_inode); + spin_lock(&lli->lli_sa_lock); + INIT_LIST_HEAD(&entry->se_list); + sa_rehash(sai, entry); + spin_unlock(&lli->lli_sa_lock); + + atomic_inc(&sai->sai_cache_count); + + RETURN(entry); } -static struct ll_statahead_info *ll_sai_alloc(void) +/* free sa_entry which should have been unhashed and not in any list */ +static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry) { - struct ll_statahead_info *sai; - - OBD_ALLOC_PTR(sai); - if (!sai) - return NULL; - - spin_lock(&sai_generation_lock); - sai->sai_generation = ++sai_generation; - if (unlikely(sai_generation == 0)) - sai->sai_generation = ++sai_generation; - spin_unlock(&sai_generation_lock); - atomic_set(&sai->sai_refcount, 1); - sai->sai_max = LL_SA_RPC_MIN; - cfs_waitq_init(&sai->sai_waitq); - cfs_waitq_init(&sai->sai_thread.t_ctl_waitq); - CFS_INIT_LIST_HEAD(&sai->sai_entries_sent); - CFS_INIT_LIST_HEAD(&sai->sai_entries_received); - CFS_INIT_LIST_HEAD(&sai->sai_entries_stated); - return sai; + CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n", + entry->se_qstr.len, entry->se_qstr.name, entry, + entry->se_index); + + LASSERT(list_empty(&entry->se_list)); + LASSERT(sa_unhashed(entry)); + + OBD_FREE(entry, entry->se_size); + atomic_dec(&sai->sai_cache_count); } -static inline -struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai) +/* find sa_entry by name, used by directory scanner, lock is not needed because + * only scanner can remove the entry from hash. + */ +static struct sa_entry * +sa_get(struct ll_statahead_info *sai, const struct qstr *qstr) { - LASSERT(sai); - atomic_inc(&sai->sai_refcount); - return sai; + struct sa_entry *entry; + int i = sa_hash(qstr->hash); + + list_for_each_entry(entry, &sai->sai_cache[i], se_hash) { + if (entry->se_qstr.hash == qstr->hash && + entry->se_qstr.len == qstr->len && + memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0) + return entry; + } + return NULL; } -static void ll_sai_put(struct ll_statahead_info *sai) +static inline void +sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry) { - struct inode *inode = sai->sai_inode; - struct ll_inode_info *lli = ll_i2info(inode); - ENTRY; + struct ll_inode_info *lli = ll_i2info(sai->sai_inode); - if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_lock)) { - struct ll_sai_entry *entry, *next; + LASSERT(!sa_unhashed(entry)); + LASSERT(!list_empty(&entry->se_list)); + LASSERT(sa_ready(entry)); - lli->lli_sai = NULL; - spin_unlock(&lli->lli_lock); + sa_unhash(sai, entry); - LASSERT(sa_is_stopped(sai)); + spin_lock(&lli->lli_sa_lock); + list_del_init(&entry->se_list); + spin_unlock(&lli->lli_sa_lock); - if (sai->sai_sent > sai->sai_replied) - CDEBUG(D_READA,"statahead for dir "DFID" does not " - "finish: [sent:%u] [replied:%u]\n", - PFID(&lli->lli_fid), - sai->sai_sent, sai->sai_replied); + if (entry->se_inode != NULL) + iput(entry->se_inode); - list_for_each_entry_safe(entry, next, &sai->sai_entries_sent, - se_list) { - list_del(&entry->se_list); - ll_sai_entry_cleanup(entry); - } - list_for_each_entry_safe(entry, next, &sai->sai_entries_received, - se_list) { - list_del(&entry->se_list); - ll_sai_entry_cleanup(entry); - } - list_for_each_entry_safe(entry, next, &sai->sai_entries_stated, - se_list) { - list_del(&entry->se_list); - ll_sai_entry_cleanup(entry); - } - dput(sai->sai_first); - OBD_FREE_PTR(sai); - iput(inode); - } - EXIT; + sa_free(sai, entry); } -/** - * insert it into sai_entries_sent tail when init. - */ -static struct ll_sai_entry * -ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index) +/* called by scanner after use, sa_entry will be killed */ +static void +sa_put(struct ll_statahead_info *sai, struct sa_entry *entry) +{ + struct sa_entry *tmp, *next; + + if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) { + struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode); + + sai->sai_hit++; + sai->sai_consecutive_miss = 0; + sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max); + } else { + sai->sai_miss++; + sai->sai_consecutive_miss++; + } + + if (entry != NULL) + sa_kill(sai, entry); + + /* kill old completed entries, only scanner process does this, no need + * to lock */ + list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) { + if (!is_omitted_entry(sai, tmp->se_index)) + break; + sa_kill(sai, tmp); + } + + wake_up(&sai->sai_thread.t_ctl_waitq); +} + +/* update state and sort add entry to sai_entries by index, return true if + * scanner is waiting on this entry. */ +static bool +__sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret) { - struct ll_inode_info *lli = ll_i2info(sai->sai_inode); - struct ll_sai_entry *entry; - ENTRY; + struct sa_entry *se; + struct list_head *pos = &sai->sai_entries; - OBD_ALLOC_PTR(entry); - if (entry == NULL) - RETURN(ERR_PTR(-ENOMEM)); + LASSERT(!sa_ready(entry)); + LASSERT(list_empty(&entry->se_list)); - CDEBUG(D_READA, "alloc sai entry %p index %u\n", - entry, index); - entry->se_index = index; - entry->se_stat = SA_ENTRY_UNSTATED; + entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC; - spin_lock(&lli->lli_lock); - list_add_tail(&entry->se_list, &sai->sai_entries_sent); - spin_unlock(&lli->lli_lock); + list_for_each_entry_reverse(se, &sai->sai_entries, se_list) { + if (se->se_index < entry->se_index) { + pos = &se->se_list; + break; + } + } + list_add(&entry->se_list, pos); - RETURN(entry); + return (entry->se_index == sai->sai_index_wait); } -/** - * delete it from sai_entries_stated head when fini, it need not - * to process entry's member. +/* release resources used in async stat RPC, complete entry information and + * wakeup if necessary */ +static void +sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret) +{ + struct ll_inode_info *lli = ll_i2info(sai->sai_inode); + struct md_enqueue_info *minfo = entry->se_minfo; + struct ptlrpc_request *req = entry->se_req; + bool wakeup; + + /* release resources used in RPC */ + if (minfo) { + entry->se_minfo = NULL; + ll_intent_release(&minfo->mi_it); + iput(minfo->mi_dir); + OBD_FREE_PTR(minfo); + } + + if (req) { + entry->se_req = NULL; + ptlrpc_req_finished(req); + } + + spin_lock(&lli->lli_sa_lock); + wakeup = __sa_make_ready(sai, entry, ret); + spin_unlock(&lli->lli_sa_lock); + + if (wakeup) + wake_up(&sai->sai_waitq); +} + +/* + * Insert inode into the list of sai_agls. */ -static void ll_sai_entry_fini(struct ll_statahead_info *sai) +static void ll_agl_add(struct ll_statahead_info *sai, + struct inode *inode, int index) { - struct ll_inode_info *lli = ll_i2info(sai->sai_inode); - struct ll_sai_entry *entry; - ENTRY; - - spin_lock(&lli->lli_lock); - sai->sai_index_next++; - if (likely(!list_empty(&sai->sai_entries_stated))) { - entry = list_entry(sai->sai_entries_stated.next, - struct ll_sai_entry, se_list); - if (entry->se_index < sai->sai_index_next) { - list_del(&entry->se_list); - OBD_FREE_PTR(entry); - } - } else - LASSERT(sa_is_stopped(sai)); - spin_unlock(&lli->lli_lock); + struct ll_inode_info *child = ll_i2info(inode); + struct ll_inode_info *parent = ll_i2info(sai->sai_inode); + int added = 0; + + spin_lock(&child->lli_agl_lock); + if (child->lli_agl_index == 0) { + child->lli_agl_index = index; + spin_unlock(&child->lli_agl_lock); + + LASSERT(list_empty(&child->lli_agl_list)); + + igrab(inode); + spin_lock(&parent->lli_agl_lock); + if (agl_list_empty(sai)) + added = 1; + list_add_tail(&child->lli_agl_list, &sai->sai_agls); + spin_unlock(&parent->lli_agl_lock); + } else { + spin_unlock(&child->lli_agl_lock); + } + + if (added > 0) + wake_up(&sai->sai_agl_thread.t_ctl_waitq); +} - EXIT; +static struct ll_statahead_info *ll_sai_alloc(void) +{ + struct ll_statahead_info *sai; + int i; + ENTRY; + + OBD_ALLOC_PTR(sai); + if (!sai) + RETURN(NULL); + + atomic_set(&sai->sai_refcount, 1); + + spin_lock(&sai_generation_lock); + sai->sai_generation = ++sai_generation; + if (unlikely(sai_generation == 0)) + sai->sai_generation = ++sai_generation; + spin_unlock(&sai_generation_lock); + + sai->sai_max = LL_SA_RPC_MIN; + sai->sai_index = 1; + init_waitqueue_head(&sai->sai_waitq); + init_waitqueue_head(&sai->sai_thread.t_ctl_waitq); + init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq); + + INIT_LIST_HEAD(&sai->sai_interim_entries); + INIT_LIST_HEAD(&sai->sai_entries); + INIT_LIST_HEAD(&sai->sai_agls); + + for (i = 0; i < LL_SA_CACHE_SIZE; i++) { + INIT_LIST_HEAD(&sai->sai_cache[i]); + spin_lock_init(&sai->sai_cache_lock[i]); + } + atomic_set(&sai->sai_cache_count, 0); + + RETURN(sai); } -/** - * inside lli_lock. - * \retval NULL : can not find the entry in sai_entries_sent with the index - * \retval entry: find the entry in sai_entries_sent with the index - */ -static struct ll_sai_entry * -ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat, - struct ptlrpc_request *req, struct md_enqueue_info *minfo) +static inline struct ll_statahead_info *ll_sai_get(struct inode *dir) { - struct ll_sai_entry *entry; - ENTRY; - - if (!list_empty(&sai->sai_entries_sent)) { - list_for_each_entry(entry, &sai->sai_entries_sent, - se_list) { - if (entry->se_index == index) { - entry->se_stat = stat; - entry->se_req = ptlrpc_request_addref(req); - entry->se_minfo = minfo; - RETURN(entry); - } else if (entry->se_index > index) - RETURN(NULL); - } - } - RETURN(NULL); + struct ll_inode_info *lli = ll_i2info(dir); + struct ll_statahead_info *sai = NULL; + + spin_lock(&lli->lli_sa_lock); + sai = lli->lli_sai; + if (sai != NULL) + atomic_inc(&sai->sai_refcount); + spin_unlock(&lli->lli_sa_lock); + + return sai; } -/** - * inside lli_lock. - * Move entry to sai_entries_received and - * insert it into sai_entries_received tail. - */ -static inline void -ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry) +static void ll_sai_put(struct ll_statahead_info *sai) { - if (!list_empty(&entry->se_list)) - list_del_init(&entry->se_list); - list_add_tail(&entry->se_list, &sai->sai_entries_received); + struct ll_inode_info *lli = ll_i2info(sai->sai_inode); + + if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) { + struct sa_entry *entry, *next; + struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode); + + lli->lli_sai = NULL; + spin_unlock(&lli->lli_sa_lock); + + LASSERT(thread_is_stopped(&sai->sai_thread)); + LASSERT(thread_is_stopped(&sai->sai_agl_thread)); + LASSERT(sai->sai_sent == sai->sai_replied); + LASSERT(!sa_has_callback(sai)); + + list_for_each_entry_safe(entry, next, &sai->sai_entries, + se_list) + sa_kill(sai, entry); + + LASSERT(atomic_read(&sai->sai_cache_count) == 0); + LASSERT(agl_list_empty(sai)); + + iput(sai->sai_inode); + OBD_FREE_PTR(sai); + atomic_dec(&sbi->ll_sa_running); + } } -/** - * Move entry to sai_entries_stated and - * sort with the index. - */ -static int -ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry) +/* Do NOT forget to drop inode refcount when into sai_agls. */ +static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai) { - struct ll_inode_info *lli = ll_i2info(sai->sai_inode); - struct ll_sai_entry *se; - ENTRY; - - spin_lock(&lli->lli_lock); - if (!list_empty(&entry->se_list)) - list_del_init(&entry->se_list); - - if (unlikely(entry->se_index < sai->sai_index_next)) { - spin_unlock(&lli->lli_lock); - ll_sai_entry_cleanup(entry); - RETURN(0); + struct ll_inode_info *lli = ll_i2info(inode); + __u64 index = lli->lli_agl_index; + int rc; + ENTRY; + + LASSERT(list_empty(&lli->lli_agl_list)); + + /* AGL maybe fall behind statahead with one entry */ + if (is_omitted_entry(sai, index + 1)) { + lli->lli_agl_index = 0; + iput(inode); + RETURN_EXIT; } - list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) { - if (se->se_index < entry->se_index) { - list_add(&entry->se_list, &se->se_list); - spin_unlock(&lli->lli_lock); - RETURN(1); - } + /* Someone is in glimpse (sync or async), do nothing. */ + rc = down_write_trylock(&lli->lli_glimpse_sem); + if (rc == 0) { + lli->lli_agl_index = 0; + iput(inode); + RETURN_EXIT; } /* - * I am the first entry. + * Someone triggered glimpse within 1 sec before. + * 1) The former glimpse succeeded with glimpse lock granted by OST, and + * if the lock is still cached on client, AGL needs to do nothing. If + * it is cancelled by other client, AGL maybe cannot obtaion new lock + * for no glimpse callback triggered by AGL. + * 2) The former glimpse succeeded, but OST did not grant glimpse lock. + * Under such case, it is quite possible that the OST will not grant + * glimpse lock for AGL also. + * 3) The former glimpse failed, compared with other two cases, it is + * relative rare. AGL can ignore such case, and it will not muchly + * affect the performance. */ - list_add(&entry->se_list, &sai->sai_entries_stated); - spin_unlock(&lli->lli_lock); - RETURN(1); + if (lli->lli_glimpse_time != 0 && + cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) { + up_write(&lli->lli_glimpse_sem); + lli->lli_agl_index = 0; + iput(inode); + RETURN_EXIT; + } + + CDEBUG(D_READA, "Handling (init) async glimpse: inode = " + DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index); + + cl_agl(inode); + lli->lli_agl_index = 0; + lli->lli_glimpse_time = cfs_time_current(); + up_write(&lli->lli_glimpse_sem); + + CDEBUG(D_READA, "Handled (init) async glimpse: inode= " + DFID", idx = "LPU64", rc = %d\n", + PFID(&lli->lli_fid), index, rc); + + iput(inode); + + EXIT; } -/** - * finish lookup/revalidate. - */ -static int do_statahead_interpret(struct ll_statahead_info *sai) +/* prepare inode for sa entry, add it into agl list, now sa_entry is ready + * to be used by scanner process. */ +static void sa_instantiate(struct ll_statahead_info *sai, + struct sa_entry *entry) { - struct ll_inode_info *lli = ll_i2info(sai->sai_inode); - struct ll_sai_entry *entry; - struct ptlrpc_request *req; - struct md_enqueue_info *minfo; - struct dentry *dentry; - struct lookup_intent *it; - int rc = 0; - ENTRY; - - spin_lock(&lli->lli_lock); - LASSERT(!sa_received_empty(sai)); - entry = list_entry(sai->sai_entries_received.next, struct ll_sai_entry, - se_list); - list_del_init(&entry->se_list); - spin_unlock(&lli->lli_lock); - - if (unlikely(entry->se_index < sai->sai_index_next)) { - ll_sai_entry_cleanup(entry); - RETURN(0); - } + struct inode *dir = sai->sai_inode; + struct inode *child; + struct md_enqueue_info *minfo; + struct lookup_intent *it; + struct ptlrpc_request *req; + struct mdt_body *body; + int rc = 0; + ENTRY; + + LASSERT(entry->se_handle != 0); - req = entry->se_req; minfo = entry->se_minfo; - dentry = minfo->mi_dentry; it = &minfo->mi_it; + req = entry->se_req; + body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY); + if (body == NULL) + GOTO(out, rc = -EFAULT); - if (entry->se_stat != SA_ENTRY_STATED) - GOTO(out, rc = entry->se_stat); - - if (dentry->d_inode == NULL) { + child = entry->se_inode; + if (child == NULL) { /* * lookup. */ - struct dentry *save = dentry; - struct it_cb_data icbd = { - .icbd_parent = dentry->d_parent->d_inode, - .icbd_childp = &dentry - }; - LASSERT(fid_is_zero(&minfo->mi_data.op_fid2)); - rc = ll_lookup_it_finish(req, it, &icbd); - if (!rc) - /* - * Here dentry->d_inode might be NULL, - * because the entry may have been removed before - * we start doing stat ahead. - */ - ll_lookup_finish_locks(it, dentry); - - if (dentry != save) { - minfo->mi_dentry = dentry; - dput(save); + /* XXX: No fid in reply, this is probaly cross-ref case. + * SA can't handle it yet. */ + if (body->mbo_valid & OBD_MD_MDS) + GOTO(out, rc = -EAGAIN); + } else { + /* + * revalidate. + */ + /* unlinked and re-created with the same name */ + if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, + &body->mbo_fid1))) { + entry->se_inode = NULL; + iput(child); + child = NULL; } - } else { - /* - * revalidate. - */ - struct mdt_body *body; + } - body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, - sizeof(*body)); - if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) { - ll_unhash_aliases(dentry->d_inode); - GOTO(out, rc = -EAGAIN); - } + it->d.lustre.it_lock_handle = entry->se_handle; + rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL); + if (rc != 1) + GOTO(out, rc = -EAGAIN); - rc = ll_revalidate_it_finish(req, it, dentry); - if (rc) { - ll_unhash_aliases(dentry->d_inode); - GOTO(out, rc); - } + rc = ll_prep_inode(&child, req, dir->i_sb, it); + if (rc) + GOTO(out, rc); + + CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n", + ll_get_fsname(child->i_sb, NULL, 0), + entry->se_qstr.len, entry->se_qstr.name, + PFID(ll_inode2fid(child)), child); + ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL); + + entry->se_inode = child; + + if (agl_should_run(sai, child)) + ll_agl_add(sai, child, entry->se_index); - spin_lock(&dcache_lock); - lock_dentry(dentry); - __d_drop(dentry); -#ifdef DCACHE_LUSTRE_INVALID - dentry->d_flags &= ~DCACHE_LUSTRE_INVALID; -#endif - unlock_dentry(dentry); - d_rehash_cond(dentry, 0); - spin_unlock(&dcache_lock); - - ll_lookup_finish_locks(it, dentry); - } EXIT; out: - if (likely(ll_sai_entry_to_stated(sai, entry))) { - entry->se_minfo = NULL; - entry->se_req = NULL; - cfs_waitq_signal(&sai->sai_waitq); - ll_intent_release(it); - OBD_FREE_PTR(minfo); - dput(dentry); - ptlrpc_req_finished(req); - } - return rc; + /* sa_make_ready() will drop ldlm ibits lock refcount by calling + * ll_intent_drop_lock() in spite of failures. Do not worry about + * calling ll_intent_drop_lock() more than once. */ + sa_make_ready(sai, entry, rc); } +/* once there are async stat replies, instantiate sa_entry */ +static void sa_handle_callback(struct ll_statahead_info *sai) +{ + struct ll_inode_info *lli; + + lli = ll_i2info(sai->sai_inode); + + while (sa_has_callback(sai)) { + struct sa_entry *entry; + + spin_lock(&lli->lli_sa_lock); + if (unlikely(!sa_has_callback(sai))) { + spin_unlock(&lli->lli_sa_lock); + break; + } + entry = list_entry(sai->sai_interim_entries.next, + struct sa_entry, se_list); + list_del_init(&entry->se_list); + spin_unlock(&lli->lli_sa_lock); + + sa_instantiate(sai, entry); + } + + spin_lock(&lli->lli_agl_lock); + while (!agl_list_empty(sai)) { + struct ll_inode_info *clli; + + clli = agl_first_entry(sai); + list_del_init(&clli->lli_agl_list); + spin_unlock(&lli->lli_agl_lock); + + ll_agl_trigger(&clli->lli_vfs_inode, sai); + + spin_lock(&lli->lli_agl_lock); + } + spin_unlock(&lli->lli_agl_lock); +} + +/* callback for async stat, because this is called in ptlrpcd context, we only + * put sa_entry in sai_cb_entries list, and let sa_handle_callback() to really + * prepare inode and instantiate sa_entry later. */ static int ll_statahead_interpret(struct ptlrpc_request *req, - struct md_enqueue_info *minfo, - int rc) + struct md_enqueue_info *minfo, int rc) { - struct dentry *dentry = minfo->mi_dentry; - struct lookup_intent *it = &minfo->mi_it; - struct inode *dir = dentry->d_parent->d_inode; - struct ll_inode_info *lli = ll_i2info(dir); - struct ll_statahead_info *sai; - struct ll_sai_entry *entry; - ENTRY; - - CDEBUG(D_READA, "interpret statahead %.*s rc %d\n", - dentry->d_name.len, dentry->d_name.name, rc); - - spin_lock(&lli->lli_lock); - if (unlikely(lli->lli_sai == NULL || - lli->lli_sai->sai_generation != minfo->mi_generation)) { - spin_unlock(&lli->lli_lock); - ll_intent_release(it); - dput(dentry); - OBD_FREE_PTR(minfo); - RETURN(-ESTALE); - } else { - sai = lli->lli_sai; - if (rc || dir == NULL) - rc = -ESTALE; - - entry = ll_sai_entry_set(sai, - (unsigned int)(long)minfo->mi_cbdata, - rc ? SA_ENTRY_UNSTATED : - SA_ENTRY_STATED, req, minfo); - LASSERT(entry != NULL); - if (likely(sa_is_running(sai))) { - ll_sai_entry_to_received(sai, entry); - sai->sai_replied++; - spin_unlock(&lli->lli_lock); - cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq); - } else { - if (!list_empty(&entry->se_list)) - list_del_init(&entry->se_list); - sai->sai_replied++; - spin_unlock(&lli->lli_lock); - ll_sai_entry_cleanup(entry); - } - RETURN(rc); - } + struct lookup_intent *it = &minfo->mi_it; + struct inode *dir = minfo->mi_dir; + struct ll_inode_info *lli = ll_i2info(dir); + struct ll_statahead_info *sai = lli->lli_sai; + struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata; + __u64 handle = 0; + bool wakeup; + ENTRY; + + if (it_disposition(it, DISP_LOOKUP_NEG)) + rc = -ENOENT; + + /* because statahead thread will wait for all inflight RPC to finish, + * sai should be always valid, no need to refcount */ + LASSERT(sai != NULL); + LASSERT(!thread_is_stopped(&sai->sai_thread)); + LASSERT(entry != NULL); + + CDEBUG(D_READA, "sa_entry %.*s rc %d\n", + entry->se_qstr.len, entry->se_qstr.name, rc); + + if (rc != 0) { + ll_intent_release(it); + iput(dir); + OBD_FREE_PTR(minfo); + } else { + /* release ibits lock ASAP to avoid deadlock when statahead + * thread enqueues lock on parent in readdir and another + * process enqueues lock on child with parent lock held, eg. + * unlink. */ + handle = it->d.lustre.it_lock_handle; + ll_intent_drop_lock(it); + } + + spin_lock(&lli->lli_sa_lock); + if (rc != 0) { + wakeup = __sa_make_ready(sai, entry, rc); + } else { + entry->se_minfo = minfo; + entry->se_req = ptlrpc_request_addref(req); + /* Release the async ibits lock ASAP to avoid deadlock + * when statahead thread tries to enqueue lock on parent + * for readpage and other tries to enqueue lock on child + * with parent's lock held, for example: unlink. */ + entry->se_handle = handle; + wakeup = !sa_has_callback(sai); + list_add_tail(&entry->se_list, &sai->sai_interim_entries); + } + sai->sai_replied++; + if (wakeup) + wake_up(&sai->sai_thread.t_ctl_waitq); + spin_unlock(&lli->lli_sa_lock); + + RETURN(rc); } -static void sa_args_fini(struct md_enqueue_info *minfo, +static void sa_fini_data(struct md_enqueue_info *minfo, struct ldlm_enqueue_info *einfo) { LASSERT(minfo && einfo); + iput(minfo->mi_dir); capa_put(minfo->mi_data.op_capa1); capa_put(minfo->mi_data.op_capa2); OBD_FREE_PTR(minfo); @@ -531,12 +730,12 @@ static void sa_args_fini(struct md_enqueue_info *minfo, * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling * "md_intent_getattr_async". */ -static int sa_args_init(struct inode *dir, struct dentry *dentry, - struct md_enqueue_info **pmi, - struct ldlm_enqueue_info **pei, - struct obd_capa **pcapa) +static int sa_prep_data(struct inode *dir, struct inode *child, + struct sa_entry *entry, struct md_enqueue_info **pmi, + struct ldlm_enqueue_info **pei, + struct obd_capa **pcapa) { - struct ll_inode_info *lli = ll_i2info(dir); + struct qstr *qstr = &entry->se_qstr; struct md_enqueue_info *minfo; struct ldlm_enqueue_info *einfo; struct md_op_data *op_data; @@ -551,20 +750,18 @@ static int sa_args_init(struct inode *dir, struct dentry *dentry, return -ENOMEM; } - op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode, - dentry->d_name.name, dentry->d_name.len, - 0, LUSTRE_OPC_ANY, NULL); + op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name, + qstr->len, 0, LUSTRE_OPC_ANY, NULL); if (IS_ERR(op_data)) { OBD_FREE_PTR(einfo); OBD_FREE_PTR(minfo); return PTR_ERR(op_data); } - minfo->mi_it.it_op = IT_GETATTR; - minfo->mi_dentry = dentry; - minfo->mi_cb = ll_statahead_interpret; - minfo->mi_generation = lli->lli_sai->sai_generation; - minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index; + minfo->mi_it.it_op = IT_GETATTR; + minfo->mi_dir = igrab(dir); + minfo->mi_cb = ll_statahead_interpret; + minfo->mi_cbdata = entry; einfo->ei_type = LDLM_IBITS; einfo->ei_mode = it_to_lock_mode(&minfo->mi_it); @@ -581,30 +778,27 @@ static int sa_args_init(struct inode *dir, struct dentry *dentry, return 0; } -/** - * similar to ll_lookup_it(). - */ -static int do_sa_lookup(struct inode *dir, struct dentry *dentry) +static int sa_lookup(struct inode *dir, struct sa_entry *entry) { - struct md_enqueue_info *minfo; - struct ldlm_enqueue_info *einfo; - struct obd_capa *capas[2]; - int rc; - ENTRY; - - rc = sa_args_init(dir, dentry, &minfo, &einfo, capas); - if (rc) - RETURN(rc); - - rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo); - if (!rc) { - capa_put(capas[0]); - capa_put(capas[1]); - } else { - sa_args_fini(minfo, einfo); - } - - RETURN(rc); + struct md_enqueue_info *minfo; + struct ldlm_enqueue_info *einfo; + struct obd_capa *capas[2]; + int rc; + ENTRY; + + rc = sa_prep_data(dir, NULL, entry, &minfo, &einfo, capas); + if (rc) + RETURN(rc); + + rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo); + if (!rc) { + capa_put(capas[0]); + capa_put(capas[1]); + } else { + sa_fini_data(minfo, einfo); + } + + RETURN(rc); } /** @@ -613,329 +807,442 @@ static int do_sa_lookup(struct inode *dir, struct dentry *dentry) * \retval 0 -- will send stat-ahead request * \retval others -- prepare stat-ahead request failed */ -static int do_sa_revalidate(struct dentry *dentry) +static int sa_revalidate(struct inode *dir, struct sa_entry *entry, + struct dentry *dentry) { - struct inode *inode = dentry->d_inode; - struct inode *dir = dentry->d_parent->d_inode; - struct lookup_intent it = { .it_op = IT_GETATTR }; - struct md_enqueue_info *minfo; - struct ldlm_enqueue_info *einfo; - struct obd_capa *capas[2]; - int rc; - ENTRY; - - if (inode == NULL) - RETURN(1); - - if (d_mountpoint(dentry)) - RETURN(1); - - if (dentry == dentry->d_sb->s_root) - RETURN(1); - - rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode)); - if (rc == 1) { - ll_intent_release(&it); - RETURN(1); - } - - rc = sa_args_init(dir, dentry, &minfo, &einfo, capas); - if (rc) - RETURN(rc); - - rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo); - if (!rc) { - capa_put(capas[0]); - capa_put(capas[1]); - } else { - sa_args_fini(minfo, einfo); - } - - RETURN(rc); + struct inode *inode = dentry->d_inode; + struct lookup_intent it = { .it_op = IT_GETATTR, + .d.lustre.it_lock_handle = 0 }; + struct md_enqueue_info *minfo; + struct ldlm_enqueue_info *einfo; + struct obd_capa *capas[2]; + int rc; + ENTRY; + + if (unlikely(inode == NULL)) + RETURN(1); + + if (d_mountpoint(dentry)) + RETURN(1); + + entry->se_inode = igrab(inode); + rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode), + NULL); + if (rc == 1) { + entry->se_handle = it.d.lustre.it_lock_handle; + ll_intent_release(&it); + RETURN(1); + } + + rc = sa_prep_data(dir, inode, entry, &minfo, &einfo, capas); + if (rc) { + entry->se_inode = NULL; + iput(inode); + RETURN(rc); + } + + rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo); + if (!rc) { + capa_put(capas[0]); + capa_put(capas[1]); + } else { + entry->se_inode = NULL; + iput(inode); + sa_fini_data(minfo, einfo); + } + + RETURN(rc); } -static inline void ll_name2qstr(struct qstr *this, const char *name, int namelen) +static void sa_statahead(struct dentry *parent, const char *name, int len) { - unsigned long hash = init_name_hash(); - unsigned int c; - - this->name = name; - this->len = namelen; - for (; namelen > 0; namelen--, name++) { - c = *(const unsigned char *)name; - hash = partial_name_hash(c, hash); - } - this->hash = end_name_hash(hash); + struct inode *dir = parent->d_inode; + struct ll_inode_info *lli = ll_i2info(dir); + struct ll_statahead_info *sai = lli->lli_sai; + struct dentry *dentry = NULL; + struct sa_entry *entry; + int rc; + ENTRY; + + entry = sa_alloc(sai, sai->sai_index, name, len); + if (IS_ERR(entry)) + RETURN_EXIT; + + dentry = d_lookup(parent, &entry->se_qstr); + if (!dentry) { + rc = sa_lookup(dir, entry); + } else { + rc = sa_revalidate(dir, entry, dentry); + if (rc == 1 && agl_should_run(sai, dentry->d_inode)) + ll_agl_add(sai, dentry->d_inode, entry->se_index); + } + + if (dentry != NULL) + dput(dentry); + + if (rc != 0) + sa_make_ready(sai, entry, rc); + else + sai->sai_sent++; + + sai->sai_index++; + + EXIT; } -static int ll_statahead_one(struct dentry *parent, const char* entry_name, - int entry_name_len) +static int ll_agl_thread(void *arg) { - struct inode *dir = parent->d_inode; - struct ll_inode_info *lli = ll_i2info(dir); - struct ll_statahead_info *sai = lli->lli_sai; - struct qstr name; - struct dentry *dentry; - struct ll_sai_entry *se; - int rc; - ENTRY; - -#ifdef DCACHE_LUSTRE_INVALID - if (parent->d_flags & DCACHE_LUSTRE_INVALID) { -#else - if (d_unhashed(parent)) { -#endif - CDEBUG(D_READA, "parent dentry@%p %.*s is " - "invalid, skip statahead\n", - parent, parent->d_name.len, parent->d_name.name); - RETURN(-EINVAL); - } + struct dentry *parent = (struct dentry *)arg; + struct inode *dir = parent->d_inode; + struct ll_inode_info *plli = ll_i2info(dir); + struct ll_inode_info *clli; + struct ll_sb_info *sbi = ll_i2sbi(dir); + struct ll_statahead_info *sai; + struct ptlrpc_thread *thread; + struct l_wait_info lwi = { 0 }; + ENTRY; + + + sai = ll_sai_get(dir); + thread = &sai->sai_agl_thread; + thread->t_pid = current_pid(); + CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n", + sai, parent->d_name.len, parent->d_name.name); + + atomic_inc(&sbi->ll_agl_total); + spin_lock(&plli->lli_agl_lock); + sai->sai_agl_valid = 1; + if (thread_is_init(thread)) + /* If someone else has changed the thread state + * (e.g. already changed to SVC_STOPPING), we can't just + * blindly overwrite that setting. */ + thread_set_flags(thread, SVC_RUNNING); + spin_unlock(&plli->lli_agl_lock); + wake_up(&thread->t_ctl_waitq); - se = ll_sai_entry_init(sai, sai->sai_index); - if (IS_ERR(se)) - RETURN(PTR_ERR(se)); - - ll_name2qstr(&name, entry_name, entry_name_len); - dentry = d_lookup(parent, &name); - if (!dentry) { - dentry = d_alloc(parent, &name); - if (dentry) { - rc = do_sa_lookup(dir, dentry); - if (rc) - dput(dentry); - } else { - GOTO(out, rc = -ENOMEM); - } - } else { - rc = do_sa_revalidate(dentry); - if (rc) - dput(dentry); - } - - EXIT; + while (1) { + l_wait_event(thread->t_ctl_waitq, + !agl_list_empty(sai) || + !thread_is_running(thread), + &lwi); -out: - if (rc) { - CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n", - se, se->se_index, se->se_stat, rc); - se->se_stat = rc; - if (ll_sai_entry_to_stated(sai, se)) - cfs_waitq_signal(&sai->sai_waitq); - } else { - sai->sai_sent++; - } + if (!thread_is_running(thread)) + break; - sai->sai_index++; - return rc; + spin_lock(&plli->lli_agl_lock); + /* The statahead thread maybe help to process AGL entries, + * so check whether list empty again. */ + if (!agl_list_empty(sai)) { + clli = agl_first_entry(sai); + list_del_init(&clli->lli_agl_list); + spin_unlock(&plli->lli_agl_lock); + ll_agl_trigger(&clli->lli_vfs_inode, sai); + } else { + spin_unlock(&plli->lli_agl_lock); + } + } + + spin_lock(&plli->lli_agl_lock); + sai->sai_agl_valid = 0; + while (!agl_list_empty(sai)) { + clli = agl_first_entry(sai); + list_del_init(&clli->lli_agl_list); + spin_unlock(&plli->lli_agl_lock); + clli->lli_agl_index = 0; + iput(&clli->lli_vfs_inode); + spin_lock(&plli->lli_agl_lock); + } + thread_set_flags(thread, SVC_STOPPED); + spin_unlock(&plli->lli_agl_lock); + wake_up(&thread->t_ctl_waitq); + ll_sai_put(sai); + CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n", + sai, parent->d_name.len, parent->d_name.name); + RETURN(0); } -struct ll_sa_thread_args { - struct dentry *sta_parent; - pid_t sta_pid; -}; +static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai) +{ + struct ptlrpc_thread *thread = &sai->sai_agl_thread; + struct l_wait_info lwi = { 0 }; + struct ll_inode_info *plli; + struct task_struct *task; + ENTRY; + + CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n", + sai, parent->d_name.len, parent->d_name.name); + + plli = ll_i2info(parent->d_inode); + task = kthread_run(ll_agl_thread, parent, + "ll_agl_%u", plli->lli_opendir_pid); + if (IS_ERR(task)) { + CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task)); + thread_set_flags(thread, SVC_STOPPED); + RETURN_EXIT; + } + + l_wait_event(thread->t_ctl_waitq, + thread_is_running(thread) || thread_is_stopped(thread), + &lwi); + EXIT; +} static int ll_statahead_thread(void *arg) { - struct ll_sa_thread_args *sta = arg; - struct dentry *parent = dget(sta->sta_parent); - struct inode *dir = parent->d_inode; - struct ll_inode_info *lli = ll_i2info(dir); - struct ll_sb_info *sbi = ll_i2sbi(dir); - struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai); - struct ptlrpc_thread *thread = &sai->sai_thread; - struct page *page; - __u64 pos = 0; - int first = 0; - int rc = 0; - struct ll_dir_chain chain; - ENTRY; - - { - char pname[16]; - snprintf(pname, 15, "ll_sa_%u", sta->sta_pid); - cfs_daemonize(pname); - } - - sbi->ll_sa_total++; - spin_lock(&lli->lli_lock); - thread->t_flags = SVC_RUNNING; - spin_unlock(&lli->lli_lock); - cfs_waitq_signal(&thread->t_ctl_waitq); - CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name); - - ll_dir_chain_init(&chain); - page = ll_get_dir_page(dir, pos, 0, &chain); - - while (1) { - struct l_wait_info lwi = { 0 }; - struct lu_dirpage *dp; - struct lu_dirent *ent; - - if (IS_ERR(page)) { - rc = PTR_ERR(page); - CERROR("error reading dir "DFID" at "LPU64"/%u: rc %d\n", - PFID(ll_inode2fid(dir)), pos, - sai->sai_index, rc); - break; - } - - dp = page_address(page); - for (ent = lu_dirent_start(dp); ent != NULL; - ent = lu_dirent_next(ent)) { - char *name = ent->lde_name; - int namelen = le16_to_cpu(ent->lde_namelen); - - if (namelen == 0) - /* - * Skip dummy record. - */ - continue; - - if (name[0] == '.') { - if (namelen == 1) { - /* - * skip "." - */ - continue; - } else if (name[1] == '.' && namelen == 2) { - /* - * skip ".." - */ - continue; - } else if (!sai->sai_ls_all) { - /* - * skip hidden files. - */ - sai->sai_skip_hidden++; - continue; - } - } - - /* - * don't stat-ahead first entry. - */ - if (unlikely(!first)) { - first++; - continue; - } - -keep_de: - l_wait_event(thread->t_ctl_waitq, - !sa_is_running(sai) || sa_not_full(sai) || - !sa_received_empty(sai), - &lwi); - - while (!sa_received_empty(sai) && sa_is_running(sai)) - do_statahead_interpret(sai); - - if (unlikely(!sa_is_running(sai))) { - ll_put_page(page); - GOTO(out, rc); - } - - if (!sa_not_full(sai)) - /* - * do not skip the current de. - */ - goto keep_de; - - rc = ll_statahead_one(parent, name, namelen); - if (rc < 0) { - ll_put_page(page); - GOTO(out, rc); - } - } - pos = le64_to_cpu(dp->ldp_hash_end); - ll_put_page(page); - if (pos == DIR_END_OFF) { - /* - * End of directory reached. - */ - while (1) { - l_wait_event(thread->t_ctl_waitq, - !sa_is_running(sai) || - !sa_received_empty(sai) || - sai->sai_sent == sai->sai_replied, - &lwi); - if (!sa_received_empty(sai) && - sa_is_running(sai)) - do_statahead_interpret(sai); - else - GOTO(out, rc); - } - } else if (1) { - /* - * chain is exhausted. - * Normal case: continue to the next page. - */ - page = ll_get_dir_page(dir, pos, 1, &chain); - } else { - /* - * go into overflow page. - */ - } - } - EXIT; - + struct dentry *parent = (struct dentry *)arg; + struct inode *dir = parent->d_inode; + struct ll_inode_info *lli = ll_i2info(dir); + struct ll_sb_info *sbi = ll_i2sbi(dir); + struct ll_statahead_info *sai; + struct ptlrpc_thread *thread; + struct ptlrpc_thread *agl_thread; + int first = 0; + struct md_op_data *op_data; + struct ll_dir_chain chain; + struct l_wait_info lwi = { 0 }; + struct page *page = NULL; + __u64 pos = 0; + int rc = 0; + ENTRY; + + sai = ll_sai_get(dir); + thread = &sai->sai_thread; + agl_thread = &sai->sai_agl_thread; + thread->t_pid = current_pid(); + CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n", + sai, parent->d_name.len, parent->d_name.name); + + op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0, + LUSTRE_OPC_ANY, dir); + if (IS_ERR(op_data)) + GOTO(out, rc = PTR_ERR(op_data)); + + op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages; + + if (sbi->ll_flags & LL_SBI_AGL_ENABLED) + ll_start_agl(parent, sai); + + atomic_inc(&sbi->ll_sa_total); + spin_lock(&lli->lli_sa_lock); + if (thread_is_init(thread)) + /* If someone else has changed the thread state + * (e.g. already changed to SVC_STOPPING), we can't just + * blindly overwrite that setting. */ + thread_set_flags(thread, SVC_RUNNING); + spin_unlock(&lli->lli_sa_lock); + wake_up(&thread->t_ctl_waitq); + + ll_dir_chain_init(&chain); + while (pos != MDS_DIR_END_OFF && thread_is_running(thread)) { + struct lu_dirpage *dp; + struct lu_dirent *ent; + + sai->sai_in_readpage = 1; + page = ll_get_dir_page(dir, op_data, pos, &chain); + sai->sai_in_readpage = 0; + if (IS_ERR(page)) { + rc = PTR_ERR(page); + CDEBUG(D_READA, "error reading dir "DFID" at "LPU64 + "/"LPU64" opendir_pid = %u: rc = %d\n", + PFID(ll_inode2fid(dir)), pos, sai->sai_index, + lli->lli_opendir_pid, rc); + break; + } + + dp = page_address(page); + for (ent = lu_dirent_start(dp); + ent != NULL && thread_is_running(thread) && + !sa_low_hit(sai); + ent = lu_dirent_next(ent)) { + __u64 hash; + int namelen; + char *name; + + hash = le64_to_cpu(ent->lde_hash); + if (unlikely(hash < pos)) + /* + * Skip until we find target hash value. + */ + continue; + + namelen = le16_to_cpu(ent->lde_namelen); + if (unlikely(namelen == 0)) + /* + * Skip dummy record. + */ + continue; + + name = ent->lde_name; + if (name[0] == '.') { + if (namelen == 1) { + /* + * skip "." + */ + continue; + } else if (name[1] == '.' && namelen == 2) { + /* + * skip ".." + */ + continue; + } else if (!sai->sai_ls_all) { + /* + * skip hidden files. + */ + sai->sai_skip_hidden++; + continue; + } + } + + /* + * don't stat-ahead first entry. + */ + if (unlikely(++first == 1)) + continue; + + /* wait for spare statahead window */ + do { + l_wait_event(thread->t_ctl_waitq, + !sa_sent_full(sai) || + sa_has_callback(sai) || + !agl_list_empty(sai) || + !thread_is_running(thread), + &lwi); + + sa_handle_callback(sai); + } while (sa_sent_full(sai) && + thread_is_running(thread)); + + sa_statahead(parent, name, namelen); + } + + pos = le64_to_cpu(dp->ldp_hash_end); + ll_release_page(dir, page, + le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE); + + if (sa_low_hit(sai)) { + rc = -EFAULT; + atomic_inc(&sbi->ll_sa_wrong); + CDEBUG(D_READA, "Statahead for dir "DFID" hit " + "ratio too low: hit/miss "LPU64"/"LPU64 + ", sent/replied "LPU64"/"LPU64", stopping " + "statahead thread: pid %d\n", + PFID(&lli->lli_fid), sai->sai_hit, + sai->sai_miss, sai->sai_sent, + sai->sai_replied, current_pid()); + break; + } + } + ll_dir_chain_fini(&chain); + ll_finish_md_op_data(op_data); + + if (rc < 0) { + spin_lock(&lli->lli_sa_lock); + thread_set_flags(thread, SVC_STOPPING); + lli->lli_sa_enabled = 0; + spin_unlock(&lli->lli_sa_lock); + } + + /* statahead is finished, but statahead entries need to be cached, wait + * for file release to stop me. */ + while (thread_is_running(thread)) { + l_wait_event(thread->t_ctl_waitq, + sa_has_callback(sai) || + !agl_list_empty(sai) || + !thread_is_running(thread), + &lwi); + + sa_handle_callback(sai); + } + + EXIT; out: - ll_dir_chain_fini(&chain); - spin_lock(&lli->lli_lock); - thread->t_flags = SVC_STOPPED; - spin_unlock(&lli->lli_lock); - cfs_waitq_signal(&sai->sai_waitq); - cfs_waitq_signal(&thread->t_ctl_waitq); + if (sai->sai_agl_valid) { + spin_lock(&lli->lli_agl_lock); + thread_set_flags(agl_thread, SVC_STOPPING); + spin_unlock(&lli->lli_agl_lock); + wake_up(&agl_thread->t_ctl_waitq); + + CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n", + sai, (unsigned int)agl_thread->t_pid); + l_wait_event(agl_thread->t_ctl_waitq, + thread_is_stopped(agl_thread), + &lwi); + } else { + /* Set agl_thread flags anyway. */ + thread_set_flags(agl_thread, SVC_STOPPED); + } + + /* wait for inflight statahead RPCs to finish, and then we can free sai + * safely because statahead RPC will access sai data */ + while (sai->sai_sent != sai->sai_replied) { + /* in case we're not woken up, timeout wait */ + lwi = LWI_TIMEOUT(HZ >> 3, NULL, NULL); + l_wait_event(thread->t_ctl_waitq, + sai->sai_sent == sai->sai_replied, &lwi); + } + + /* release resources held by statahead RPCs */ + sa_handle_callback(sai); + + spin_lock(&lli->lli_sa_lock); + thread_set_flags(thread, SVC_STOPPED); + spin_unlock(&lli->lli_sa_lock); + + wake_up(&sai->sai_waitq); + wake_up(&thread->t_ctl_waitq); ll_sai_put(sai); - dput(parent); - CDEBUG(D_READA, "statahead thread stopped, pid %d\n", - cfs_curproc_pid()); - return rc; + CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n", + sai, parent->d_name.len, parent->d_name.name); + dput(parent); + return rc; } -/** - * called in ll_file_release(). - */ -void ll_stop_statahead(struct inode *inode, void *key) +/* authorize opened dir handle @key to statahead later */ +void ll_authorize_statahead(struct inode *dir, void *key) { - struct ll_inode_info *lli = ll_i2info(inode); - struct ptlrpc_thread *thread; - - spin_lock(&lli->lli_lock); - if (lli->lli_opendir_pid == 0 || - unlikely(lli->lli_opendir_key != key)) { - spin_unlock(&lli->lli_lock); - return; - } - - lli->lli_opendir_key = NULL; - lli->lli_opendir_pid = 0; - - if (lli->lli_sai) { - struct l_wait_info lwi = { 0 }; - - thread = &lli->lli_sai->sai_thread; - if (!sa_is_stopped(lli->lli_sai)) { - thread->t_flags = SVC_STOPPING; - spin_unlock(&lli->lli_lock); - cfs_waitq_signal(&thread->t_ctl_waitq); - - CDEBUG(D_READA, "stopping statahead thread, pid %d\n", - cfs_curproc_pid()); - l_wait_event(thread->t_ctl_waitq, - sa_is_stopped(lli->lli_sai), - &lwi); - } else { - spin_unlock(&lli->lli_lock); - } + struct ll_inode_info *lli = ll_i2info(dir); + + spin_lock(&lli->lli_sa_lock); + if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) { + /* + * if lli_sai is not NULL, it means previous statahead is not + * finished yet, we'd better not start a new statahead for now. + */ + LASSERT(lli->lli_opendir_pid == 0); + lli->lli_opendir_key = key; + lli->lli_opendir_pid = current_pid(); + lli->lli_sa_enabled = 1; + } + spin_unlock(&lli->lli_sa_lock); +} - /* - * Put the ref which was held when first statahead_enter. - * It maybe not the last ref for some statahead requests - * maybe inflight. - */ - ll_sai_put(lli->lli_sai); - return; - } - spin_unlock(&lli->lli_lock); +/* deauthorize opened dir handle @key to statahead, but statahead thread may + * still be running, notify it to quit. */ +void ll_deauthorize_statahead(struct inode *dir, void *key) +{ + struct ll_inode_info *lli = ll_i2info(dir); + struct ll_statahead_info *sai; + + LASSERT(lli->lli_opendir_key == key); + LASSERT(lli->lli_opendir_pid != 0); + + CDEBUG(D_READA, "deauthorize statahead for "DFID"\n", + PFID(&lli->lli_fid)); + + spin_lock(&lli->lli_sa_lock); + lli->lli_opendir_key = NULL; + lli->lli_opendir_pid = 0; + lli->lli_sa_enabled = 0; + sai = lli->lli_sai; + if (sai != NULL && thread_is_running(&sai->sai_thread)) { + /* + * statahead thread may not quit yet because it needs to cache + * entries, now it's time to tell it to quit. + */ + thread_set_flags(&sai->sai_thread, SVC_STOPPING); + wake_up(&sai->sai_thread.t_ctl_waitq); + } + spin_unlock(&lli->lli_sa_lock); } enum { @@ -948,308 +1255,353 @@ enum { */ LS_FIRST_DE, /** - * the first hidden dirent, that is ".xxx + * the first hidden dirent, that is "." */ LS_FIRST_DOT_DE }; static int is_first_dirent(struct inode *dir, struct dentry *dentry) { - struct ll_dir_chain chain; - struct qstr *target = &dentry->d_name; - struct page *page; - __u64 pos = 0; - int dot_de; - int rc = LS_NONE_FIRST_DE; - ENTRY; - - ll_dir_chain_init(&chain); - page = ll_get_dir_page(dir, pos, 0, &chain); - - while (1) { - struct lu_dirpage *dp; - struct lu_dirent *ent; - - if (IS_ERR(page)) { - rc = PTR_ERR(page); - CERROR("error reading dir "DFID" at "LPU64": rc %d\n", - PFID(ll_inode2fid(dir)), pos, rc); - break; - } - - dp = page_address(page); - for (ent = lu_dirent_start(dp); ent != NULL; - ent = lu_dirent_next(ent)) { - char *name = ent->lde_name; - int namelen = le16_to_cpu(ent->lde_namelen); - - if (namelen == 0) - /* - * skip dummy record. - */ - continue; - - if (name[0] == '.') { - if (namelen == 1) - /* - * skip "." - */ - continue; - else if (name[1] == '.' && namelen == 2) - /* - * skip ".." - */ - continue; - else - dot_de = 1; - } else { - dot_de = 0; - } - - if (dot_de && target->name[0] != '.') { - CDEBUG(D_READA, "%.*s skip hidden file %.*s\n", - target->len, target->name, - namelen, name); - continue; - } - - if (target->len == namelen && - !strncmp(target->name, name, target->len)) - rc = LS_FIRST_DE + dot_de; - else - rc = LS_NONE_FIRST_DE; - ll_put_page(page); - GOTO(out, rc); - } - pos = le64_to_cpu(dp->ldp_hash_end); - ll_put_page(page); - if (pos == DIR_END_OFF) { - /* - * End of directory reached. - */ - break; - } else if (1) { - /* - * chain is exhausted - * Normal case: continue to the next page. - */ - page = ll_get_dir_page(dir, pos, 1, &chain); - } else { - /* - * go into overflow page. - */ - } - } - EXIT; - + struct ll_dir_chain chain; + struct qstr *target = &dentry->d_name; + struct md_op_data *op_data; + int dot_de; + struct page *page = NULL; + int rc = LS_NONE_FIRST_DE; + __u64 pos = 0; + ENTRY; + + op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0, + LUSTRE_OPC_ANY, dir); + if (IS_ERR(op_data)) + RETURN(PTR_ERR(op_data)); + /** + *FIXME choose the start offset of the readdir + */ + op_data->op_stripe_offset = 0; + op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages; + + ll_dir_chain_init(&chain); + page = ll_get_dir_page(dir, op_data, 0, &chain); + + while (1) { + struct lu_dirpage *dp; + struct lu_dirent *ent; + + if (IS_ERR(page)) { + struct ll_inode_info *lli = ll_i2info(dir); + + rc = PTR_ERR(page); + CERROR("%s: reading dir "DFID" at "LPU64 + "opendir_pid = %u : rc = %d\n", + ll_get_fsname(dir->i_sb, NULL, 0), + PFID(ll_inode2fid(dir)), pos, + lli->lli_opendir_pid, rc); + break; + } + + dp = page_address(page); + for (ent = lu_dirent_start(dp); ent != NULL; + ent = lu_dirent_next(ent)) { + __u64 hash; + int namelen; + char *name; + + hash = le64_to_cpu(ent->lde_hash); + /* The ll_get_dir_page() can return any page containing + * the given hash which may be not the start hash. */ + if (unlikely(hash < pos)) + continue; + + namelen = le16_to_cpu(ent->lde_namelen); + if (unlikely(namelen == 0)) + /* + * skip dummy record. + */ + continue; + + name = ent->lde_name; + if (name[0] == '.') { + if (namelen == 1) + /* + * skip "." + */ + continue; + else if (name[1] == '.' && namelen == 2) + /* + * skip ".." + */ + continue; + else + dot_de = 1; + } else { + dot_de = 0; + } + + if (dot_de && target->name[0] != '.') { + CDEBUG(D_READA, "%.*s skip hidden file %.*s\n", + target->len, target->name, + namelen, name); + continue; + } + + if (target->len != namelen || + memcmp(target->name, name, namelen) != 0) + rc = LS_NONE_FIRST_DE; + else if (!dot_de) + rc = LS_FIRST_DE; + else + rc = LS_FIRST_DOT_DE; + + ll_release_page(dir, page, false); + GOTO(out, rc); + } + pos = le64_to_cpu(dp->ldp_hash_end); + if (pos == MDS_DIR_END_OFF) { + /* + * End of directory reached. + */ + ll_release_page(dir, page, false); + GOTO(out, rc); + } else { + /* + * chain is exhausted + * Normal case: continue to the next page. + */ + ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) & + LDF_COLLIDE); + page = ll_get_dir_page(dir, op_data, pos, &chain); + } + } + EXIT; out: - ll_dir_chain_fini(&chain); + ll_dir_chain_fini(&chain); + ll_finish_md_op_data(op_data); return rc; } -/** - * Start statahead thread if this is the first dir entry. - * Otherwise if a thread is started already, wait it until it is ahead of me. - * \retval 0 -- stat ahead thread process such dentry, for lookup, it miss - * \retval 1 -- stat ahead thread process such dentry, for lookup, it hit - * \retval -EEXIST -- stat ahead thread started, and this is the first dentry - * \retval -EBADFD -- statahead thread exit and not dentry available - * \retval others -- error - */ -int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup) +static int revalidate_statahead_dentry(struct inode *dir, + struct ll_statahead_info *sai, + struct dentry **dentryp, + int only_unplug) { - struct ll_sb_info *sbi = ll_i2sbi(dir); - struct ll_inode_info *lli = ll_i2info(dir); - struct ll_statahead_info *sai = lli->lli_sai; - struct ll_sa_thread_args sta; - struct l_wait_info lwi = { 0 }; - int rc; - ENTRY; - - LASSERT(lli->lli_opendir_pid == cfs_curproc_pid()); - - if (sai) { - if (unlikely(sa_is_stopped(sai) && - list_empty(&sai->sai_entries_stated))) - RETURN(-EBADFD); - - /* - * skip the first dentry. - */ - if (unlikely((*dentryp)->d_name.len == - sai->sai_first->d_name.len && - !strncmp((*dentryp)->d_name.name, - sai->sai_first->d_name.name, - sai->sai_first->d_name.len))) - RETURN(-EEXIST); - - if ((*dentryp)->d_name.name[0] == '.') { - if (likely(sai->sai_ls_all || - sai->sai_miss_hidden >= sai->sai_skip_hidden)) { - /* - * Hidden dentry is the first one, or statahead - * thread does not skip so many hidden dentries - * before "sai_ls_all" enabled as below. - */ - } else { - if (!sai->sai_ls_all) - /* - * It maybe because hidden dentry is not - * the first one, "sai_ls_all" was not - * set, then "ls -al" missed. Enable - * "sai_ls_all" for such case. - */ - sai->sai_ls_all = 1; - - /* - * Such "getattr" has been skipped before - * "sai_ls_all" enabled as above. - */ - sai->sai_miss_hidden++; - RETURN(-ENOENT); - } - } - - if (ll_sai_entry_stated(sai)) { - sbi->ll_sa_cached++; - } else { - sbi->ll_sa_blocked++; - /* - * thread started already, avoid double-stat. - */ - l_wait_event(sai->sai_waitq, - ll_sai_entry_stated(sai) || sa_is_stopped(sai), - &lwi); - } - - if (lookup) { - struct dentry *result; - - result = d_lookup((*dentryp)->d_parent, - &(*dentryp)->d_name); - if (result) { - LASSERT(result != *dentryp); - dput(*dentryp); - *dentryp = result; - RETURN(1); - } - } - /* - * do nothing for revalidate. - */ - RETURN(0); - } - - /* - * I am the "lli_opendir_pid" owner, only me can set "lli_sai". - */ - LASSERT(lli->lli_sai == NULL); - - rc = is_first_dirent(dir, *dentryp); - if (rc == LS_NONE_FIRST_DE) { - /* - * It is not "ls -{a}l" operation, no need statahead for it. - */ - spin_lock(&lli->lli_lock); - lli->lli_opendir_key = NULL; - lli->lli_opendir_pid = 0; - spin_unlock(&lli->lli_lock); - RETURN(-EBADF); - } - - sai = ll_sai_alloc(); - if (sai == NULL) - RETURN(-ENOMEM); - - sai->sai_inode = igrab(dir); - sai->sai_first = dget(*dentryp); - sai->sai_ls_all = (rc == LS_FIRST_DOT_DE); + struct sa_entry *entry = NULL; + struct l_wait_info lwi = { 0 }; + int rc = 0; + ENTRY; + + if ((*dentryp)->d_name.name[0] == '.') { + if (sai->sai_ls_all || + sai->sai_miss_hidden >= sai->sai_skip_hidden) { + /* + * Hidden dentry is the first one, or statahead + * thread does not skip so many hidden dentries + * before "sai_ls_all" enabled as below. + */ + } else { + if (!sai->sai_ls_all) + /* + * It maybe because hidden dentry is not + * the first one, "sai_ls_all" was not + * set, then "ls -al" missed. Enable + * "sai_ls_all" for such case. + */ + sai->sai_ls_all = 1; + + /* + * Such "getattr" has been skipped before + * "sai_ls_all" enabled as above. + */ + sai->sai_miss_hidden++; + RETURN(-EAGAIN); + } + } + + entry = sa_get(sai, &(*dentryp)->d_name); + if (entry == NULL || only_unplug) { + sa_put(sai, entry); + RETURN(entry ? 1 : -EAGAIN); + } + + /* if statahead is busy in readdir, help it do post-work */ + if (!sa_ready(entry) && sai->sai_in_readpage) + sa_handle_callback(sai); + + if (!sa_ready(entry)) { + sai->sai_index_wait = entry->se_index; + lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL, + LWI_ON_SIGNAL_NOOP, NULL); + rc = l_wait_event(sai->sai_waitq, + sa_ready(entry) || + thread_is_stopped(&sai->sai_thread), + &lwi); + if (rc < 0) { + sa_put(sai, entry); + RETURN(-EAGAIN); + } + } + + if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) { + struct inode *inode = entry->se_inode; + struct lookup_intent it = { .it_op = IT_GETATTR, + .d.lustre.it_lock_handle = + entry->se_handle }; + __u64 bits; + + rc = md_revalidate_lock(ll_i2mdexp(dir), &it, + ll_inode2fid(inode), &bits); + if (rc == 1) { + if ((*dentryp)->d_inode == NULL) { + struct dentry *alias; + + alias = ll_splice_alias(inode, *dentryp); + if (IS_ERR(alias)) { + sa_put(sai, entry); + RETURN(PTR_ERR(alias)); + } + *dentryp = alias; + } else if ((*dentryp)->d_inode != inode) { + /* revalidate, but inode is recreated */ + CDEBUG(D_READA, + "%s: stale dentry %.*s inode " + DFID", statahead inode "DFID + "\n", + ll_get_fsname((*dentryp)->d_inode->i_sb, + NULL, 0), + (*dentryp)->d_name.len, + (*dentryp)->d_name.name, + PFID(ll_inode2fid((*dentryp)->d_inode)), + PFID(ll_inode2fid(inode))); + sa_put(sai, entry); + RETURN(-ESTALE); + } else { + iput(inode); + } + entry->se_inode = NULL; + + if ((bits & MDS_INODELOCK_LOOKUP) && + d_lustre_invalid(*dentryp)) + d_lustre_revalidate(*dentryp); + ll_intent_release(&it); + } + } + + sa_put(sai, entry); + RETURN(rc); +} - sta.sta_parent = (*dentryp)->d_parent; - sta.sta_pid = cfs_curproc_pid(); +static int start_statahead_thread(struct inode *dir, struct dentry *dentry) +{ + struct ll_inode_info *lli = ll_i2info(dir); + struct ll_statahead_info *sai = NULL; + struct dentry *parent; + struct ptlrpc_thread *thread; + struct l_wait_info lwi = { 0 }; + struct task_struct *task; + int rc; + ENTRY; + + /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */ + rc = is_first_dirent(dir, dentry); + if (rc == LS_NONE_FIRST_DE) + /* It is not "ls -{a}l" operation, no need statahead for it. */ + GOTO(out, rc = -EAGAIN); + + sai = ll_sai_alloc(); + if (sai == NULL) + GOTO(out, rc = -ENOMEM); + + sai->sai_ls_all = (rc == LS_FIRST_DOT_DE); + sai->sai_inode = igrab(dir); + if (unlikely(sai->sai_inode == NULL)) { + CWARN("Do not start stat ahead on dying inode "DFID"\n", + PFID(&lli->lli_fid)); + GOTO(out, rc = -ESTALE); + } + + /* get parent reference count here, and put it in ll_statahead_thread */ + parent = dget(dentry->d_parent); + if (unlikely(sai->sai_inode != parent->d_inode)) { + struct ll_inode_info *nlli = ll_i2info(parent->d_inode); + + CWARN("Race condition, someone changed %.*s just now: " + "old parent "DFID", new parent "DFID"\n", + dentry->d_name.len, dentry->d_name.name, + PFID(&lli->lli_fid), PFID(&nlli->lli_fid)); + dput(parent); + iput(sai->sai_inode); + GOTO(out, rc = -EAGAIN); + } + + CDEBUG(D_READA, "start statahead thread: sai %p, parent %.*s\n", + sai, parent->d_name.len, parent->d_name.name); lli->lli_sai = sai; - rc = cfs_kernel_thread(ll_statahead_thread, &sta, 0); - if (rc < 0) { - CERROR("can't start ll_sa thread, rc: %d\n", rc); - sai->sai_thread.t_flags = SVC_STOPPED; - ll_sai_put(sai); - LASSERT(lli->lli_sai == NULL); - RETURN(rc); - } - l_wait_event(sai->sai_thread.t_ctl_waitq, - sa_is_running(sai) || sa_is_stopped(sai), - &lwi); + task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u", + lli->lli_opendir_pid); + thread = &sai->sai_thread; + if (IS_ERR(task)) { + rc = PTR_ERR(task); + CERROR("cannot start ll_sa thread: rc = %d\n", rc); + dput(parent); + lli->lli_opendir_key = NULL; + thread_set_flags(thread, SVC_STOPPED); + thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED); + ll_sai_put(sai); + LASSERT(lli->lli_sai == NULL); + RETURN(-EAGAIN); + } + + l_wait_event(thread->t_ctl_waitq, + thread_is_running(thread) || thread_is_stopped(thread), + &lwi); + atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running); + ll_sai_put(sai); + + /* + * We don't stat-ahead for the first dirent since we are already in + * lookup. + */ + RETURN(-EAGAIN); - /* - * We don't stat-ahead for the first dirent since we are already in - * lookup, and -EEXIST also indicates that this is the first dirent. - */ - RETURN(-EEXIST); +out: + if (sai != NULL) + OBD_FREE_PTR(sai); + spin_lock(&lli->lli_sa_lock); + lli->lli_opendir_key = NULL; + lli->lli_opendir_pid = 0; + lli->lli_sa_enabled = 0; + spin_unlock(&lli->lli_sa_lock); + + RETURN(rc); } /** - * update hit/miss count. + * Start statahead thread if this is the first dir entry. + * Otherwise if a thread is started already, wait it until it is ahead of me. + * \retval 1 -- find entry with lock in cache, the caller needs to do + * nothing. + * \retval 0 -- find entry in cache, but without lock, the caller needs + * refresh from MDS. + * \retval others -- the caller need to process as non-statahead. */ -int ll_statahead_exit(struct dentry *dentry, int result) +int do_statahead_enter(struct inode *dir, struct dentry **dentryp, + int only_unplug) { - struct dentry *parent = dentry->d_parent; - struct ll_inode_info *lli = ll_i2info(parent->d_inode); - struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode); - struct ll_dentry_data *ldd = ll_d2d(dentry); - ENTRY; - - if (lli->lli_opendir_pid != cfs_curproc_pid()) - RETURN(-EBADFD); - - if (lli->lli_sai) { - struct ll_statahead_info *sai = lli->lli_sai; - - if (result >= 1) { - sbi->ll_sa_hit++; - sai->sai_hit++; - sai->sai_consecutive_miss = 0; - sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max); - } else { - sbi->ll_sa_miss++; - sai->sai_miss++; - sai->sai_consecutive_miss++; - if (sa_low_hit(sai) && sa_is_running(sai)) { - sbi->ll_sa_wrong++; - CDEBUG(D_READA, "statahead for dir %.*s hit " - "ratio too low: hit/miss %u/%u, " - "sent/replied %u/%u. stopping statahead " - "thread: pid %d\n", - parent->d_name.len, parent->d_name.name, - sai->sai_hit, sai->sai_miss, - sai->sai_sent, sai->sai_replied, - cfs_curproc_pid()); - spin_lock(&lli->lli_lock); - if (!sa_is_stopped(sai)) - sai->sai_thread.t_flags = SVC_STOPPING; - spin_unlock(&lli->lli_lock); - } - } + struct ll_statahead_info *sai; - if (!sa_is_stopped(sai)) - cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq); - ll_sai_entry_fini(sai); - - if (unlikely(ldd == NULL)) { - ll_set_dd(dentry); - ldd = ll_d2d(dentry); - if (ldd != NULL && dentry->d_op == NULL) { - lock_dentry(dentry); - dentry->d_op = dentry->d_op ? : &ll_sai_d_ops; - unlock_dentry(dentry); - } - } + sai = ll_sai_get(dir); + if (sai != NULL) { + int rc; - if (likely(ldd != NULL)) - ldd->lld_sa_generation = sai->sai_generation; - else - RETURN(-ENOMEM); - } - RETURN(0); + rc = revalidate_statahead_dentry(dir, sai, dentryp, + only_unplug); + CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n", + (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc); + ll_sai_put(sai); + return rc; + } + + return start_statahead_thread(dir, *dentryp); }