From 35d4c683a3dad51450db3281c9331b876f8313c5 Mon Sep 17 00:00:00 2001 From: Mr NeilBrown Date: Fri, 15 May 2020 14:06:29 +1000 Subject: [PATCH] LU-11310 ldiskfs: Repair support for SUSE 15 again A recent patch split ext4-htree-lock.patch out from the various ext4-pdirop.patch, but this happened before various files were copies and modified for extra sles15 support. The patch makes the necessary checks to bring the htree-lock split to sle15. Fixes: 46ed28c0d10a ("LU-11310 ldiskfs: Repair support for SUSE 15 GA and SP1") Fixes: 42880f9502ba ("LU-13054 ldiskfs: split htree_lock as separate patch") Signed-off-by: Mr NeilBrown Change-Id: I464aaebc52d5b2b73fdd748c7d4dbbaf43f1ac49 Reviewed-on: https://review.whamcloud.com/38611 Reviewed-by: Yang Sheng Reviewed-by: Shaun Tancheff Reviewed-by: James Simmons Tested-by: jenkins Reviewed-by: Oleg Drokin --- .../patches/sles15sp1/ext4-pdirop.patch | 1086 -------------------- .../series/ldiskfs-4.12-sles15-22.series | 1 + .../series/ldiskfs-4.12-sles15sp1-7.series | 1 + .../series/ldiskfs-4.12-sles15sp1.series | 1 + 4 files changed, 3 insertions(+), 1086 deletions(-) diff --git a/ldiskfs/kernel_patches/patches/sles15sp1/ext4-pdirop.patch b/ldiskfs/kernel_patches/patches/sles15sp1/ext4-pdirop.patch index 537545d..b65c947 100644 --- a/ldiskfs/kernel_patches/patches/sles15sp1/ext4-pdirop.patch +++ b/ldiskfs/kernel_patches/patches/sles15sp1/ext4-pdirop.patch @@ -17,10 +17,8 @@ This patch contains: --- fs/ext4/Makefile | 1 fs/ext4/ext4.h | 78 +++ - fs/ext4/htree_lock.c | 891 +++++++++++++++++++++++++++++++++++++++++++++ fs/ext4/namei.c | 446 ++++++++++++++++++++-- fs/ext4/super.c | 1 - include/linux/htree_lock.h | 187 +++++++++ 6 files changed, 1567 insertions(+), 37 deletions(-) --- a/fs/ext4/Makefile @@ -146,900 +144,6 @@ This patch contains: } static const unsigned char ext4_filetype_table[] = { DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK ---- /dev/null -+++ b/fs/ext4/htree_lock.c -@@ -0,0 +1,891 @@ -+/* -+ * fs/ext4/htree_lock.c -+ * -+ * Copyright (c) 2011, 2012, Intel Corporation. -+ * -+ * Author: Liang Zhen -+ */ -+#include -+#include -+#include -+#include -+ -+enum { -+ HTREE_LOCK_BIT_EX = (1 << HTREE_LOCK_EX), -+ HTREE_LOCK_BIT_PW = (1 << HTREE_LOCK_PW), -+ HTREE_LOCK_BIT_PR = (1 << HTREE_LOCK_PR), -+ HTREE_LOCK_BIT_CW = (1 << HTREE_LOCK_CW), -+ HTREE_LOCK_BIT_CR = (1 << HTREE_LOCK_CR), -+}; -+ -+enum { -+ HTREE_LOCK_COMPAT_EX = 0, -+ HTREE_LOCK_COMPAT_PW = HTREE_LOCK_COMPAT_EX | HTREE_LOCK_BIT_CR, -+ HTREE_LOCK_COMPAT_PR = HTREE_LOCK_COMPAT_PW | HTREE_LOCK_BIT_PR, -+ HTREE_LOCK_COMPAT_CW = HTREE_LOCK_COMPAT_PW | HTREE_LOCK_BIT_CW, -+ HTREE_LOCK_COMPAT_CR = HTREE_LOCK_COMPAT_CW | HTREE_LOCK_BIT_PR | -+ HTREE_LOCK_BIT_PW, -+}; -+ -+static int htree_lock_compat[] = { -+ [HTREE_LOCK_EX] HTREE_LOCK_COMPAT_EX, -+ [HTREE_LOCK_PW] HTREE_LOCK_COMPAT_PW, -+ [HTREE_LOCK_PR] HTREE_LOCK_COMPAT_PR, -+ [HTREE_LOCK_CW] HTREE_LOCK_COMPAT_CW, -+ [HTREE_LOCK_CR] HTREE_LOCK_COMPAT_CR, -+}; -+ -+/* max allowed htree-lock depth. -+ * We only need depth=3 for ext4 although user can have higher value. */ -+#define HTREE_LOCK_DEP_MAX 16 -+ -+#ifdef HTREE_LOCK_DEBUG -+ -+static char *hl_name[] = { -+ [HTREE_LOCK_EX] "EX", -+ [HTREE_LOCK_PW] "PW", -+ [HTREE_LOCK_PR] "PR", -+ [HTREE_LOCK_CW] "CW", -+ [HTREE_LOCK_CR] "CR", -+}; -+ -+/* lock stats */ -+struct htree_lock_node_stats { -+ unsigned long long blocked[HTREE_LOCK_MAX]; -+ unsigned long long granted[HTREE_LOCK_MAX]; -+ unsigned long long retried[HTREE_LOCK_MAX]; -+ unsigned long long events; -+}; -+ -+struct htree_lock_stats { -+ struct htree_lock_node_stats nodes[HTREE_LOCK_DEP_MAX]; -+ unsigned long long granted[HTREE_LOCK_MAX]; -+ unsigned long long blocked[HTREE_LOCK_MAX]; -+}; -+ -+static struct htree_lock_stats hl_stats; -+ -+void htree_lock_stat_reset(void) -+{ -+ memset(&hl_stats, 0, sizeof(hl_stats)); -+} -+ -+void htree_lock_stat_print(int depth) -+{ -+ int i; -+ int j; -+ -+ printk(KERN_DEBUG "HTREE LOCK STATS:\n"); -+ for (i = 0; i < HTREE_LOCK_MAX; i++) { -+ printk(KERN_DEBUG "[%s]: G [%10llu], B [%10llu]\n", -+ hl_name[i], hl_stats.granted[i], hl_stats.blocked[i]); -+ } -+ for (i = 0; i < depth; i++) { -+ printk(KERN_DEBUG "HTREE CHILD [%d] STATS:\n", i); -+ for (j = 0; j < HTREE_LOCK_MAX; j++) { -+ printk(KERN_DEBUG -+ "[%s]: G [%10llu], B [%10llu], R [%10llu]\n", -+ hl_name[j], hl_stats.nodes[i].granted[j], -+ hl_stats.nodes[i].blocked[j], -+ hl_stats.nodes[i].retried[j]); -+ } -+ } -+} -+ -+#define lk_grant_inc(m) do { hl_stats.granted[m]++; } while (0) -+#define lk_block_inc(m) do { hl_stats.blocked[m]++; } while (0) -+#define ln_grant_inc(d, m) do { hl_stats.nodes[d].granted[m]++; } while (0) -+#define ln_block_inc(d, m) do { hl_stats.nodes[d].blocked[m]++; } while (0) -+#define ln_retry_inc(d, m) do { hl_stats.nodes[d].retried[m]++; } while (0) -+#define ln_event_inc(d) do { hl_stats.nodes[d].events++; } while (0) -+ -+#else /* !DEBUG */ -+ -+void htree_lock_stat_reset(void) {} -+void htree_lock_stat_print(int depth) {} -+ -+#define lk_grant_inc(m) do {} while (0) -+#define lk_block_inc(m) do {} while (0) -+#define ln_grant_inc(d, m) do {} while (0) -+#define ln_block_inc(d, m) do {} while (0) -+#define ln_retry_inc(d, m) do {} while (0) -+#define ln_event_inc(d) do {} while (0) -+ -+#endif /* DEBUG */ -+ -+EXPORT_SYMBOL(htree_lock_stat_reset); -+EXPORT_SYMBOL(htree_lock_stat_print); -+ -+#define HTREE_DEP_ROOT (-1) -+ -+#define htree_spin_lock(lhead, dep) \ -+ bit_spin_lock((dep) + 1, &(lhead)->lh_lock) -+#define htree_spin_unlock(lhead, dep) \ -+ bit_spin_unlock((dep) + 1, &(lhead)->lh_lock) -+ -+#define htree_key_event_ignore(child, ln) \ -+ (!((child)->lc_events & (1 << (ln)->ln_mode))) -+ -+static int -+htree_key_list_empty(struct htree_lock_node *ln) -+{ -+ return list_empty(&ln->ln_major_list) && list_empty(&ln->ln_minor_list); -+} -+ -+static void -+htree_key_list_del_init(struct htree_lock_node *ln) -+{ -+ struct htree_lock_node *tmp = NULL; -+ -+ if (!list_empty(&ln->ln_minor_list)) { -+ tmp = list_entry(ln->ln_minor_list.next, -+ struct htree_lock_node, ln_minor_list); -+ list_del_init(&ln->ln_minor_list); -+ } -+ -+ if (list_empty(&ln->ln_major_list)) -+ return; -+ -+ if (tmp == NULL) { /* not on minor key list */ -+ list_del_init(&ln->ln_major_list); -+ } else { -+ BUG_ON(!list_empty(&tmp->ln_major_list)); -+ list_replace_init(&ln->ln_major_list, &tmp->ln_major_list); -+ } -+} -+ -+static void -+htree_key_list_replace_init(struct htree_lock_node *old, -+ struct htree_lock_node *new) -+{ -+ if (!list_empty(&old->ln_major_list)) -+ list_replace_init(&old->ln_major_list, &new->ln_major_list); -+ -+ if (!list_empty(&old->ln_minor_list)) -+ list_replace_init(&old->ln_minor_list, &new->ln_minor_list); -+} -+ -+static void -+htree_key_event_enqueue(struct htree_lock_child *child, -+ struct htree_lock_node *ln, int dep, void *event) -+{ -+ struct htree_lock_node *tmp; -+ -+ /* NB: ALWAYS called holding lhead::lh_lock(dep) */ -+ BUG_ON(ln->ln_mode == HTREE_LOCK_NL); -+ if (event == NULL || htree_key_event_ignore(child, ln)) -+ return; -+ -+ /* shouldn't be a very long list */ -+ list_for_each_entry(tmp, &ln->ln_alive_list, ln_alive_list) { -+ if (tmp->ln_mode == HTREE_LOCK_NL) { -+ ln_event_inc(dep); -+ if (child->lc_callback != NULL) -+ child->lc_callback(tmp->ln_ev_target, event); -+ } -+ } -+} -+ -+static int -+htree_node_lock_enqueue(struct htree_lock *newlk, struct htree_lock *curlk, -+ unsigned dep, int wait, void *event) -+{ -+ struct htree_lock_child *child = &newlk->lk_head->lh_children[dep]; -+ struct htree_lock_node *newln = &newlk->lk_nodes[dep]; -+ struct htree_lock_node *curln = &curlk->lk_nodes[dep]; -+ -+ /* NB: ALWAYS called holding lhead::lh_lock(dep) */ -+ /* NB: we only expect PR/PW lock mode at here, only these two modes are -+ * allowed for htree_node_lock(asserted in htree_node_lock_internal), -+ * NL is only used for listener, user can't directly require NL mode */ -+ if ((curln->ln_mode == HTREE_LOCK_NL) || -+ (curln->ln_mode != HTREE_LOCK_PW && -+ newln->ln_mode != HTREE_LOCK_PW)) { -+ /* no conflict, attach it on granted list of @curlk */ -+ if (curln->ln_mode != HTREE_LOCK_NL) { -+ list_add(&newln->ln_granted_list, -+ &curln->ln_granted_list); -+ } else { -+ /* replace key owner */ -+ htree_key_list_replace_init(curln, newln); -+ } -+ -+ list_add(&newln->ln_alive_list, &curln->ln_alive_list); -+ htree_key_event_enqueue(child, newln, dep, event); -+ ln_grant_inc(dep, newln->ln_mode); -+ return 1; /* still hold lh_lock */ -+ } -+ -+ if (!wait) { /* can't grant and don't want to wait */ -+ ln_retry_inc(dep, newln->ln_mode); -+ newln->ln_mode = HTREE_LOCK_INVAL; -+ return -1; /* don't wait and just return -1 */ -+ } -+ -+ newlk->lk_task = current; -+ set_current_state(TASK_UNINTERRUPTIBLE); -+ /* conflict, attach it on blocked list of curlk */ -+ list_add_tail(&newln->ln_blocked_list, &curln->ln_blocked_list); -+ list_add(&newln->ln_alive_list, &curln->ln_alive_list); -+ ln_block_inc(dep, newln->ln_mode); -+ -+ htree_spin_unlock(newlk->lk_head, dep); -+ /* wait to be given the lock */ -+ if (newlk->lk_task != NULL) -+ schedule(); -+ /* granted, no doubt, wake up will set me RUNNING */ -+ if (event == NULL || htree_key_event_ignore(child, newln)) -+ return 0; /* granted without lh_lock */ -+ -+ htree_spin_lock(newlk->lk_head, dep); -+ htree_key_event_enqueue(child, newln, dep, event); -+ return 1; /* still hold lh_lock */ -+} -+ -+/* -+ * get PR/PW access to particular tree-node according to @dep and @key, -+ * it will return -1 if @wait is false and can't immediately grant this lock. -+ * All listeners(HTREE_LOCK_NL) on @dep and with the same @key will get -+ * @event if it's not NULL. -+ * NB: ALWAYS called holding lhead::lh_lock -+ */ -+static int -+htree_node_lock_internal(struct htree_lock_head *lhead, struct htree_lock *lck, -+ htree_lock_mode_t mode, u32 key, unsigned dep, -+ int wait, void *event) -+{ -+ LIST_HEAD(list); -+ struct htree_lock *tmp; -+ struct htree_lock *tmp2; -+ u16 major; -+ u16 minor; -+ u8 reverse; -+ u8 ma_bits; -+ u8 mi_bits; -+ -+ BUG_ON(mode != HTREE_LOCK_PW && mode != HTREE_LOCK_PR); -+ BUG_ON(htree_node_is_granted(lck, dep)); -+ -+ key = hash_long(key, lhead->lh_hbits); -+ -+ mi_bits = lhead->lh_hbits >> 1; -+ ma_bits = lhead->lh_hbits - mi_bits; -+ -+ lck->lk_nodes[dep].ln_major_key = major = key & ((1U << ma_bits) - 1); -+ lck->lk_nodes[dep].ln_minor_key = minor = key >> ma_bits; -+ lck->lk_nodes[dep].ln_mode = mode; -+ -+ /* -+ * The major key list is an ordered list, so searches are started -+ * at the end of the list that is numerically closer to major_key, -+ * so at most half of the list will be walked (for well-distributed -+ * keys). The list traversal aborts early if the expected key -+ * location is passed. -+ */ -+ reverse = (major >= (1 << (ma_bits - 1))); -+ -+ if (reverse) { -+ list_for_each_entry_reverse(tmp, -+ &lhead->lh_children[dep].lc_list, -+ lk_nodes[dep].ln_major_list) { -+ if (tmp->lk_nodes[dep].ln_major_key == major) { -+ goto search_minor; -+ -+ } else if (tmp->lk_nodes[dep].ln_major_key < major) { -+ /* attach _after_ @tmp */ -+ list_add(&lck->lk_nodes[dep].ln_major_list, -+ &tmp->lk_nodes[dep].ln_major_list); -+ goto out_grant_major; -+ } -+ } -+ -+ list_add(&lck->lk_nodes[dep].ln_major_list, -+ &lhead->lh_children[dep].lc_list); -+ goto out_grant_major; -+ -+ } else { -+ list_for_each_entry(tmp, &lhead->lh_children[dep].lc_list, -+ lk_nodes[dep].ln_major_list) { -+ if (tmp->lk_nodes[dep].ln_major_key == major) { -+ goto search_minor; -+ -+ } else if (tmp->lk_nodes[dep].ln_major_key > major) { -+ /* insert _before_ @tmp */ -+ list_add_tail(&lck->lk_nodes[dep].ln_major_list, -+ &tmp->lk_nodes[dep].ln_major_list); -+ goto out_grant_major; -+ } -+ } -+ -+ list_add_tail(&lck->lk_nodes[dep].ln_major_list, -+ &lhead->lh_children[dep].lc_list); -+ goto out_grant_major; -+ } -+ -+ search_minor: -+ /* -+ * NB: minor_key list doesn't have a "head", @list is just a -+ * temporary stub for helping list searching, make sure it's removed -+ * after searching. -+ * minor_key list is an ordered list too. -+ */ -+ list_add_tail(&list, &tmp->lk_nodes[dep].ln_minor_list); -+ -+ reverse = (minor >= (1 << (mi_bits - 1))); -+ -+ if (reverse) { -+ list_for_each_entry_reverse(tmp2, &list, -+ lk_nodes[dep].ln_minor_list) { -+ if (tmp2->lk_nodes[dep].ln_minor_key == minor) { -+ goto out_enqueue; -+ -+ } else if (tmp2->lk_nodes[dep].ln_minor_key < minor) { -+ /* attach _after_ @tmp2 */ -+ list_add(&lck->lk_nodes[dep].ln_minor_list, -+ &tmp2->lk_nodes[dep].ln_minor_list); -+ goto out_grant_minor; -+ } -+ } -+ -+ list_add(&lck->lk_nodes[dep].ln_minor_list, &list); -+ -+ } else { -+ list_for_each_entry(tmp2, &list, -+ lk_nodes[dep].ln_minor_list) { -+ if (tmp2->lk_nodes[dep].ln_minor_key == minor) { -+ goto out_enqueue; -+ -+ } else if (tmp2->lk_nodes[dep].ln_minor_key > minor) { -+ /* insert _before_ @tmp2 */ -+ list_add_tail(&lck->lk_nodes[dep].ln_minor_list, -+ &tmp2->lk_nodes[dep].ln_minor_list); -+ goto out_grant_minor; -+ } -+ } -+ -+ list_add_tail(&lck->lk_nodes[dep].ln_minor_list, &list); -+ } -+ -+ out_grant_minor: -+ if (list.next == &lck->lk_nodes[dep].ln_minor_list) { -+ /* new lock @lck is the first one on minor_key list, which -+ * means it has the smallest minor_key and it should -+ * replace @tmp as minor_key owner */ -+ list_replace_init(&tmp->lk_nodes[dep].ln_major_list, -+ &lck->lk_nodes[dep].ln_major_list); -+ } -+ /* remove the temporary head */ -+ list_del(&list); -+ -+ out_grant_major: -+ ln_grant_inc(dep, lck->lk_nodes[dep].ln_mode); -+ return 1; /* granted with holding lh_lock */ -+ -+ out_enqueue: -+ list_del(&list); /* remove temprary head */ -+ return htree_node_lock_enqueue(lck, tmp2, dep, wait, event); -+} -+ -+/* -+ * release the key of @lck at level @dep, and grant any blocked locks. -+ * caller will still listen on @key if @event is not NULL, which means -+ * caller can see a event (by event_cb) while granting any lock with -+ * the same key at level @dep. -+ * NB: ALWAYS called holding lhead::lh_lock -+ * NB: listener will not block anyone because listening mode is HTREE_LOCK_NL -+ */ -+static void -+htree_node_unlock_internal(struct htree_lock_head *lhead, -+ struct htree_lock *curlk, unsigned dep, void *event) -+{ -+ struct htree_lock_node *curln = &curlk->lk_nodes[dep]; -+ struct htree_lock *grtlk = NULL; -+ struct htree_lock_node *grtln; -+ struct htree_lock *poslk; -+ struct htree_lock *tmplk; -+ -+ if (!htree_node_is_granted(curlk, dep)) -+ return; -+ -+ if (!list_empty(&curln->ln_granted_list)) { -+ /* there is another granted lock */ -+ grtlk = list_entry(curln->ln_granted_list.next, -+ struct htree_lock, -+ lk_nodes[dep].ln_granted_list); -+ list_del_init(&curln->ln_granted_list); -+ } -+ -+ if (grtlk == NULL && !list_empty(&curln->ln_blocked_list)) { -+ /* -+ * @curlk is the only granted lock, so we confirmed: -+ * a) curln is key owner (attached on major/minor_list), -+ * so if there is any blocked lock, it should be attached -+ * on curln->ln_blocked_list -+ * b) we always can grant the first blocked lock -+ */ -+ grtlk = list_entry(curln->ln_blocked_list.next, -+ struct htree_lock, -+ lk_nodes[dep].ln_blocked_list); -+ BUG_ON(grtlk->lk_task == NULL); -+ wake_up_process(grtlk->lk_task); -+ } -+ -+ if (event != NULL && -+ lhead->lh_children[dep].lc_events != HTREE_EVENT_DISABLE) { -+ curln->ln_ev_target = event; -+ curln->ln_mode = HTREE_LOCK_NL; /* listen! */ -+ } else { -+ curln->ln_mode = HTREE_LOCK_INVAL; -+ } -+ -+ if (grtlk == NULL) { /* I must be the only one locking this key */ -+ struct htree_lock_node *tmpln; -+ -+ BUG_ON(htree_key_list_empty(curln)); -+ -+ if (curln->ln_mode == HTREE_LOCK_NL) /* listening */ -+ return; -+ -+ /* not listening */ -+ if (list_empty(&curln->ln_alive_list)) { /* no more listener */ -+ htree_key_list_del_init(curln); -+ return; -+ } -+ -+ tmpln = list_entry(curln->ln_alive_list.next, -+ struct htree_lock_node, ln_alive_list); -+ -+ BUG_ON(tmpln->ln_mode != HTREE_LOCK_NL); -+ -+ htree_key_list_replace_init(curln, tmpln); -+ list_del_init(&curln->ln_alive_list); -+ -+ return; -+ } -+ -+ /* have a granted lock */ -+ grtln = &grtlk->lk_nodes[dep]; -+ if (!list_empty(&curln->ln_blocked_list)) { -+ /* only key owner can be on both lists */ -+ BUG_ON(htree_key_list_empty(curln)); -+ -+ if (list_empty(&grtln->ln_blocked_list)) { -+ list_add(&grtln->ln_blocked_list, -+ &curln->ln_blocked_list); -+ } -+ list_del_init(&curln->ln_blocked_list); -+ } -+ /* -+ * NB: this is the tricky part: -+ * We have only two modes for child-lock (PR and PW), also, -+ * only owner of the key (attached on major/minor_list) can be on -+ * both blocked_list and granted_list, so @grtlk must be one -+ * of these two cases: -+ * -+ * a) @grtlk is taken from granted_list, which means we've granted -+ * more than one lock so @grtlk has to be PR, the first blocked -+ * lock must be PW and we can't grant it at all. -+ * So even @grtlk is not owner of the key (empty blocked_list), -+ * we don't care because we can't grant any lock. -+ * b) we just grant a new lock which is taken from head of blocked -+ * list, and it should be the first granted lock, and it should -+ * be the first one linked on blocked_list. -+ * -+ * Either way, we can get correct result by iterating blocked_list -+ * of @grtlk, and don't have to bother on how to find out -+ * owner of current key. -+ */ -+ list_for_each_entry_safe(poslk, tmplk, &grtln->ln_blocked_list, -+ lk_nodes[dep].ln_blocked_list) { -+ if (grtlk->lk_nodes[dep].ln_mode == HTREE_LOCK_PW || -+ poslk->lk_nodes[dep].ln_mode == HTREE_LOCK_PW) -+ break; -+ /* grant all readers */ -+ list_del_init(&poslk->lk_nodes[dep].ln_blocked_list); -+ list_add(&poslk->lk_nodes[dep].ln_granted_list, -+ &grtln->ln_granted_list); -+ -+ BUG_ON(poslk->lk_task == NULL); -+ wake_up_process(poslk->lk_task); -+ } -+ -+ /* if @curln is the owner of this key, replace it with @grtln */ -+ if (!htree_key_list_empty(curln)) -+ htree_key_list_replace_init(curln, grtln); -+ -+ if (curln->ln_mode == HTREE_LOCK_INVAL) -+ list_del_init(&curln->ln_alive_list); -+} -+ -+/* -+ * it's just wrapper of htree_node_lock_internal, it returns 1 on granted -+ * and 0 only if @wait is false and can't grant it immediately -+ */ -+int -+htree_node_lock_try(struct htree_lock *lck, htree_lock_mode_t mode, -+ u32 key, unsigned dep, int wait, void *event) -+{ -+ struct htree_lock_head *lhead = lck->lk_head; -+ int rc; -+ -+ BUG_ON(dep >= lck->lk_depth); -+ BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL); -+ -+ htree_spin_lock(lhead, dep); -+ rc = htree_node_lock_internal(lhead, lck, mode, key, dep, wait, event); -+ if (rc != 0) -+ htree_spin_unlock(lhead, dep); -+ return rc >= 0; -+} -+EXPORT_SYMBOL(htree_node_lock_try); -+ -+/* it's wrapper of htree_node_unlock_internal */ -+void -+htree_node_unlock(struct htree_lock *lck, unsigned dep, void *event) -+{ -+ struct htree_lock_head *lhead = lck->lk_head; -+ -+ BUG_ON(dep >= lck->lk_depth); -+ BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL); -+ -+ htree_spin_lock(lhead, dep); -+ htree_node_unlock_internal(lhead, lck, dep, event); -+ htree_spin_unlock(lhead, dep); -+} -+EXPORT_SYMBOL(htree_node_unlock); -+ -+/* stop listening on child-lock level @dep */ -+void -+htree_node_stop_listen(struct htree_lock *lck, unsigned dep) -+{ -+ struct htree_lock_node *ln = &lck->lk_nodes[dep]; -+ struct htree_lock_node *tmp; -+ -+ BUG_ON(htree_node_is_granted(lck, dep)); -+ BUG_ON(!list_empty(&ln->ln_blocked_list)); -+ BUG_ON(!list_empty(&ln->ln_granted_list)); -+ -+ if (!htree_node_is_listening(lck, dep)) -+ return; -+ -+ htree_spin_lock(lck->lk_head, dep); -+ ln->ln_mode = HTREE_LOCK_INVAL; -+ ln->ln_ev_target = NULL; -+ -+ if (htree_key_list_empty(ln)) { /* not owner */ -+ list_del_init(&ln->ln_alive_list); -+ goto out; -+ } -+ -+ /* I'm the owner... */ -+ if (list_empty(&ln->ln_alive_list)) { /* no more listener */ -+ htree_key_list_del_init(ln); -+ goto out; -+ } -+ -+ tmp = list_entry(ln->ln_alive_list.next, -+ struct htree_lock_node, ln_alive_list); -+ -+ BUG_ON(tmp->ln_mode != HTREE_LOCK_NL); -+ htree_key_list_replace_init(ln, tmp); -+ list_del_init(&ln->ln_alive_list); -+ out: -+ htree_spin_unlock(lck->lk_head, dep); -+} -+EXPORT_SYMBOL(htree_node_stop_listen); -+ -+/* release all child-locks if we have any */ -+static void -+htree_node_release_all(struct htree_lock *lck) -+{ -+ int i; -+ -+ for (i = 0; i < lck->lk_depth; i++) { -+ if (htree_node_is_granted(lck, i)) -+ htree_node_unlock(lck, i, NULL); -+ else if (htree_node_is_listening(lck, i)) -+ htree_node_stop_listen(lck, i); -+ } -+} -+ -+/* -+ * obtain htree lock, it could be blocked inside if there's conflict -+ * with any granted or blocked lock and @wait is true. -+ * NB: ALWAYS called holding lhead::lh_lock -+ */ -+static int -+htree_lock_internal(struct htree_lock *lck, int wait) -+{ -+ struct htree_lock_head *lhead = lck->lk_head; -+ int granted = 0; -+ int blocked = 0; -+ int i; -+ -+ for (i = 0; i < HTREE_LOCK_MAX; i++) { -+ if (lhead->lh_ngranted[i] != 0) -+ granted |= 1 << i; -+ if (lhead->lh_nblocked[i] != 0) -+ blocked |= 1 << i; -+ } -+ if ((htree_lock_compat[lck->lk_mode] & granted) != granted || -+ (htree_lock_compat[lck->lk_mode] & blocked) != blocked) { -+ /* will block current lock even it just conflicts with any -+ * other blocked lock, so lock like EX wouldn't starve */ -+ if (!wait) -+ return -1; -+ lhead->lh_nblocked[lck->lk_mode]++; -+ lk_block_inc(lck->lk_mode); -+ -+ lck->lk_task = current; -+ list_add_tail(&lck->lk_blocked_list, &lhead->lh_blocked_list); -+ -+retry: -+ set_current_state(TASK_UNINTERRUPTIBLE); -+ htree_spin_unlock(lhead, HTREE_DEP_ROOT); -+ /* wait to be given the lock */ -+ if (lck->lk_task != NULL) -+ schedule(); -+ /* granted, no doubt. wake up will set me RUNNING. -+ * Since thread would be waken up accidentally, -+ * so we need check lock whether granted or not again. */ -+ if (!list_empty(&lck->lk_blocked_list)) { -+ htree_spin_lock(lhead, HTREE_DEP_ROOT); -+ if (list_empty(&lck->lk_blocked_list)) { -+ htree_spin_unlock(lhead, HTREE_DEP_ROOT); -+ return 0; -+ } -+ goto retry; -+ } -+ return 0; /* without lh_lock */ -+ } -+ lhead->lh_ngranted[lck->lk_mode]++; -+ lk_grant_inc(lck->lk_mode); -+ return 1; -+} -+ -+/* release htree lock. NB: ALWAYS called holding lhead::lh_lock */ -+static void -+htree_unlock_internal(struct htree_lock *lck) -+{ -+ struct htree_lock_head *lhead = lck->lk_head; -+ struct htree_lock *tmp; -+ struct htree_lock *tmp2; -+ int granted = 0; -+ int i; -+ -+ BUG_ON(lhead->lh_ngranted[lck->lk_mode] == 0); -+ -+ lhead->lh_ngranted[lck->lk_mode]--; -+ lck->lk_mode = HTREE_LOCK_INVAL; -+ -+ for (i = 0; i < HTREE_LOCK_MAX; i++) { -+ if (lhead->lh_ngranted[i] != 0) -+ granted |= 1 << i; -+ } -+ list_for_each_entry_safe(tmp, tmp2, -+ &lhead->lh_blocked_list, lk_blocked_list) { -+ /* conflict with any granted lock? */ -+ if ((htree_lock_compat[tmp->lk_mode] & granted) != granted) -+ break; -+ -+ list_del_init(&tmp->lk_blocked_list); -+ -+ BUG_ON(lhead->lh_nblocked[tmp->lk_mode] == 0); -+ -+ lhead->lh_nblocked[tmp->lk_mode]--; -+ lhead->lh_ngranted[tmp->lk_mode]++; -+ granted |= 1 << tmp->lk_mode; -+ -+ BUG_ON(tmp->lk_task == NULL); -+ wake_up_process(tmp->lk_task); -+ } -+} -+ -+/* it's wrapper of htree_lock_internal and exported interface. -+ * It always return 1 with granted lock if @wait is true, it can return 0 -+ * if @wait is false and locking request can't be granted immediately */ -+int -+htree_lock_try(struct htree_lock *lck, struct htree_lock_head *lhead, -+ htree_lock_mode_t mode, int wait) -+{ -+ int rc; -+ -+ BUG_ON(lck->lk_depth > lhead->lh_depth); -+ BUG_ON(lck->lk_head != NULL); -+ BUG_ON(lck->lk_task != NULL); -+ -+ lck->lk_head = lhead; -+ lck->lk_mode = mode; -+ -+ htree_spin_lock(lhead, HTREE_DEP_ROOT); -+ rc = htree_lock_internal(lck, wait); -+ if (rc != 0) -+ htree_spin_unlock(lhead, HTREE_DEP_ROOT); -+ return rc >= 0; -+} -+EXPORT_SYMBOL(htree_lock_try); -+ -+/* it's wrapper of htree_unlock_internal and exported interface. -+ * It will release all htree_node_locks and htree_lock */ -+void -+htree_unlock(struct htree_lock *lck) -+{ -+ BUG_ON(lck->lk_head == NULL); -+ BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL); -+ -+ htree_node_release_all(lck); -+ -+ htree_spin_lock(lck->lk_head, HTREE_DEP_ROOT); -+ htree_unlock_internal(lck); -+ htree_spin_unlock(lck->lk_head, HTREE_DEP_ROOT); -+ lck->lk_head = NULL; -+ lck->lk_task = NULL; -+} -+EXPORT_SYMBOL(htree_unlock); -+ -+/* change lock mode */ -+void -+htree_change_mode(struct htree_lock *lck, htree_lock_mode_t mode) -+{ -+ BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL); -+ lck->lk_mode = mode; -+} -+EXPORT_SYMBOL(htree_change_mode); -+ -+/* release htree lock, and lock it again with new mode. -+ * This function will first release all htree_node_locks and htree_lock, -+ * then try to gain htree_lock with new @mode. -+ * It always return 1 with granted lock if @wait is true, it can return 0 -+ * if @wait is false and locking request can't be granted immediately */ -+int -+htree_change_lock_try(struct htree_lock *lck, htree_lock_mode_t mode, int wait) -+{ -+ struct htree_lock_head *lhead = lck->lk_head; -+ int rc; -+ -+ BUG_ON(lhead == NULL); -+ BUG_ON(lck->lk_mode == mode); -+ BUG_ON(lck->lk_mode == HTREE_LOCK_INVAL || mode == HTREE_LOCK_INVAL); -+ -+ htree_node_release_all(lck); -+ -+ htree_spin_lock(lhead, HTREE_DEP_ROOT); -+ htree_unlock_internal(lck); -+ lck->lk_mode = mode; -+ rc = htree_lock_internal(lck, wait); -+ if (rc != 0) -+ htree_spin_unlock(lhead, HTREE_DEP_ROOT); -+ return rc >= 0; -+} -+EXPORT_SYMBOL(htree_change_lock_try); -+ -+/* create a htree_lock head with @depth levels (number of child-locks), -+ * it is a per resoruce structure */ -+struct htree_lock_head * -+htree_lock_head_alloc(unsigned depth, unsigned hbits, unsigned priv) -+{ -+ struct htree_lock_head *lhead; -+ int i; -+ -+ if (depth > HTREE_LOCK_DEP_MAX) { -+ printk(KERN_ERR "%d is larger than max htree_lock depth %d\n", -+ depth, HTREE_LOCK_DEP_MAX); -+ return NULL; -+ } -+ -+ lhead = kzalloc(offsetof(struct htree_lock_head, -+ lh_children[depth]) + priv, GFP_NOFS); -+ if (lhead == NULL) -+ return NULL; -+ -+ if (hbits < HTREE_HBITS_MIN) -+ lhead->lh_hbits = HTREE_HBITS_MIN; -+ else if (hbits > HTREE_HBITS_MAX) -+ lhead->lh_hbits = HTREE_HBITS_MAX; -+ -+ lhead->lh_lock = 0; -+ lhead->lh_depth = depth; -+ INIT_LIST_HEAD(&lhead->lh_blocked_list); -+ if (priv > 0) { -+ lhead->lh_private = (void *)lhead + -+ offsetof(struct htree_lock_head, lh_children[depth]); -+ } -+ -+ for (i = 0; i < depth; i++) { -+ INIT_LIST_HEAD(&lhead->lh_children[i].lc_list); -+ lhead->lh_children[i].lc_events = HTREE_EVENT_DISABLE; -+ } -+ return lhead; -+} -+EXPORT_SYMBOL(htree_lock_head_alloc); -+ -+/* free the htree_lock head */ -+void -+htree_lock_head_free(struct htree_lock_head *lhead) -+{ -+ int i; -+ -+ BUG_ON(!list_empty(&lhead->lh_blocked_list)); -+ for (i = 0; i < lhead->lh_depth; i++) -+ BUG_ON(!list_empty(&lhead->lh_children[i].lc_list)); -+ kfree(lhead); -+} -+EXPORT_SYMBOL(htree_lock_head_free); -+ -+/* register event callback for @events of child-lock at level @dep */ -+void -+htree_lock_event_attach(struct htree_lock_head *lhead, unsigned dep, -+ unsigned events, htree_event_cb_t callback) -+{ -+ BUG_ON(lhead->lh_depth <= dep); -+ lhead->lh_children[dep].lc_events = events; -+ lhead->lh_children[dep].lc_callback = callback; -+} -+EXPORT_SYMBOL(htree_lock_event_attach); -+ -+/* allocate a htree_lock, which is per-thread structure, @pbytes is some -+ * extra-bytes as private data for caller */ -+struct htree_lock * -+htree_lock_alloc(unsigned depth, unsigned pbytes) -+{ -+ struct htree_lock *lck; -+ int i = offsetof(struct htree_lock, lk_nodes[depth]); -+ -+ if (depth > HTREE_LOCK_DEP_MAX) { -+ printk(KERN_ERR "%d is larger than max htree_lock depth %d\n", -+ depth, HTREE_LOCK_DEP_MAX); -+ return NULL; -+ } -+ lck = kzalloc(i + pbytes, GFP_NOFS); -+ if (lck == NULL) -+ return NULL; -+ -+ if (pbytes != 0) -+ lck->lk_private = (void *)lck + i; -+ lck->lk_mode = HTREE_LOCK_INVAL; -+ lck->lk_depth = depth; -+ INIT_LIST_HEAD(&lck->lk_blocked_list); -+ -+ for (i = 0; i < depth; i++) { -+ struct htree_lock_node *node = &lck->lk_nodes[i]; -+ -+ node->ln_mode = HTREE_LOCK_INVAL; -+ INIT_LIST_HEAD(&node->ln_major_list); -+ INIT_LIST_HEAD(&node->ln_minor_list); -+ INIT_LIST_HEAD(&node->ln_alive_list); -+ INIT_LIST_HEAD(&node->ln_blocked_list); -+ INIT_LIST_HEAD(&node->ln_granted_list); -+ } -+ -+ return lck; -+} -+EXPORT_SYMBOL(htree_lock_alloc); -+ -+/* free htree_lock node */ -+void -+htree_lock_free(struct htree_lock *lck) -+{ -+ BUG_ON(lck->lk_mode != HTREE_LOCK_INVAL); -+ kfree(lck); -+} -+EXPORT_SYMBOL(htree_lock_free); --- a/fs/ext4/namei.c +++ b/fs/ext4/namei.c @@ -53,6 +53,7 @@ struct buffer_head *ext4_append(handle_t @@ -1755,193 +859,3 @@ This patch contains: INIT_LIST_HEAD(&ei->i_prealloc_list); spin_lock_init(&ei->i_prealloc_lock); ext4_es_init_tree(&ei->i_es_tree); ---- /dev/null -+++ b/include/linux/htree_lock.h -@@ -0,0 +1,187 @@ -+/* -+ * include/linux/htree_lock.h -+ * -+ * Copyright (c) 2011, 2012, Intel Corporation. -+ * -+ * Author: Liang Zhen -+ */ -+ -+/* -+ * htree lock -+ * -+ * htree_lock is an advanced lock, it can support five lock modes (concept is -+ * taken from DLM) and it's a sleeping lock. -+ * -+ * most common use case is: -+ * - create a htree_lock_head for data -+ * - each thread (contender) creates it's own htree_lock -+ * - contender needs to call htree_lock(lock_node, mode) to protect data and -+ * call htree_unlock to release lock -+ * -+ * Also, there is advanced use-case which is more complex, user can have -+ * PW/PR lock on particular key, it's mostly used while user holding shared -+ * lock on the htree (CW, CR) -+ * -+ * htree_lock(lock_node, HTREE_LOCK_CR); lock the htree with CR -+ * htree_node_lock(lock_node, HTREE_LOCK_PR, key...); lock @key with PR -+ * ... -+ * htree_node_unlock(lock_node);; unlock the key -+ * -+ * Another tip is, we can have N-levels of this kind of keys, all we need to -+ * do is specifying N-levels while creating htree_lock_head, then we can -+ * lock/unlock a specific level by: -+ * htree_node_lock(lock_node, mode1, key1, level1...); -+ * do something; -+ * htree_node_lock(lock_node, mode1, key2, level2...); -+ * do something; -+ * htree_node_unlock(lock_node, level2); -+ * htree_node_unlock(lock_node, level1); -+ * -+ * NB: for multi-level, should be careful about locking order to avoid deadlock -+ */ -+ -+#ifndef _LINUX_HTREE_LOCK_H -+#define _LINUX_HTREE_LOCK_H -+ -+#include -+#include -+#include -+ -+/* -+ * Lock Modes -+ * more details can be found here: -+ * http://en.wikipedia.org/wiki/Distributed_lock_manager -+ */ -+typedef enum { -+ HTREE_LOCK_EX = 0, /* exclusive lock: incompatible with all others */ -+ HTREE_LOCK_PW, /* protected write: allows only CR users */ -+ HTREE_LOCK_PR, /* protected read: allow PR, CR users */ -+ HTREE_LOCK_CW, /* concurrent write: allow CR, CW users */ -+ HTREE_LOCK_CR, /* concurrent read: allow all but EX users */ -+ HTREE_LOCK_MAX, /* number of lock modes */ -+} htree_lock_mode_t; -+ -+#define HTREE_LOCK_NL HTREE_LOCK_MAX -+#define HTREE_LOCK_INVAL 0xdead10c -+ -+enum { -+ HTREE_HBITS_MIN = 2, -+ HTREE_HBITS_DEF = 14, -+ HTREE_HBITS_MAX = 32, -+}; -+ -+enum { -+ HTREE_EVENT_DISABLE = (0), -+ HTREE_EVENT_RD = (1 << HTREE_LOCK_PR), -+ HTREE_EVENT_WR = (1 << HTREE_LOCK_PW), -+ HTREE_EVENT_RDWR = (HTREE_EVENT_RD | HTREE_EVENT_WR), -+}; -+ -+struct htree_lock; -+ -+typedef void (*htree_event_cb_t)(void *target, void *event); -+ -+struct htree_lock_child { -+ struct list_head lc_list; /* granted list */ -+ htree_event_cb_t lc_callback; /* event callback */ -+ unsigned lc_events; /* event types */ -+}; -+ -+struct htree_lock_head { -+ unsigned long lh_lock; /* bits lock */ -+ /* blocked lock list (htree_lock) */ -+ struct list_head lh_blocked_list; -+ /* # key levels */ -+ u16 lh_depth; -+ /* hash bits for key and limit number of locks */ -+ u16 lh_hbits; -+ /* counters for blocked locks */ -+ u16 lh_nblocked[HTREE_LOCK_MAX]; -+ /* counters for granted locks */ -+ u16 lh_ngranted[HTREE_LOCK_MAX]; -+ /* private data */ -+ void *lh_private; -+ /* array of children locks */ -+ struct htree_lock_child lh_children[0]; -+}; -+ -+/* htree_lock_node_t is child-lock for a specific key (ln_value) */ -+struct htree_lock_node { -+ htree_lock_mode_t ln_mode; -+ /* major hash key */ -+ u16 ln_major_key; -+ /* minor hash key */ -+ u16 ln_minor_key; -+ struct list_head ln_major_list; -+ struct list_head ln_minor_list; -+ /* alive list, all locks (granted, blocked, listening) are on it */ -+ struct list_head ln_alive_list; -+ /* blocked list */ -+ struct list_head ln_blocked_list; -+ /* granted list */ -+ struct list_head ln_granted_list; -+ void *ln_ev_target; -+}; -+ -+struct htree_lock { -+ struct task_struct *lk_task; -+ struct htree_lock_head *lk_head; -+ void *lk_private; -+ unsigned lk_depth; -+ htree_lock_mode_t lk_mode; -+ struct list_head lk_blocked_list; -+ struct htree_lock_node lk_nodes[0]; -+}; -+ -+/* create a lock head, which stands for a resource */ -+struct htree_lock_head *htree_lock_head_alloc(unsigned depth, -+ unsigned hbits, unsigned priv); -+/* free a lock head */ -+void htree_lock_head_free(struct htree_lock_head *lhead); -+/* register event callback for child lock at level @depth */ -+void htree_lock_event_attach(struct htree_lock_head *lhead, unsigned depth, -+ unsigned events, htree_event_cb_t callback); -+/* create a lock handle, which stands for a thread */ -+struct htree_lock *htree_lock_alloc(unsigned depth, unsigned pbytes); -+/* free a lock handle */ -+void htree_lock_free(struct htree_lock *lck); -+/* lock htree, when @wait is true, 0 is returned if the lock can't -+ * be granted immediately */ -+int htree_lock_try(struct htree_lock *lck, struct htree_lock_head *lhead, -+ htree_lock_mode_t mode, int wait); -+/* unlock htree */ -+void htree_unlock(struct htree_lock *lck); -+/* unlock and relock htree with @new_mode */ -+int htree_change_lock_try(struct htree_lock *lck, -+ htree_lock_mode_t new_mode, int wait); -+void htree_change_mode(struct htree_lock *lck, htree_lock_mode_t mode); -+/* require child lock (key) of htree at level @dep, @event will be sent to all -+ * listeners on this @key while lock being granted */ -+int htree_node_lock_try(struct htree_lock *lck, htree_lock_mode_t mode, -+ u32 key, unsigned dep, int wait, void *event); -+/* release child lock at level @dep, this lock will listen on it's key -+ * if @event isn't NULL, event_cb will be called against @lck while granting -+ * any other lock at level @dep with the same key */ -+void htree_node_unlock(struct htree_lock *lck, unsigned dep, void *event); -+/* stop listening on child lock at level @dep */ -+void htree_node_stop_listen(struct htree_lock *lck, unsigned dep); -+/* for debug */ -+void htree_lock_stat_print(int depth); -+void htree_lock_stat_reset(void); -+ -+#define htree_lock(lck, lh, mode) htree_lock_try(lck, lh, mode, 1) -+#define htree_change_lock(lck, mode) htree_change_lock_try(lck, mode, 1) -+ -+#define htree_lock_mode(lck) ((lck)->lk_mode) -+ -+#define htree_node_lock(lck, mode, key, dep) \ -+ htree_node_lock_try(lck, mode, key, dep, 1, NULL) -+/* this is only safe in thread context of lock owner */ -+#define htree_node_is_granted(lck, dep) \ -+ ((lck)->lk_nodes[dep].ln_mode != HTREE_LOCK_INVAL && \ -+ (lck)->lk_nodes[dep].ln_mode != HTREE_LOCK_NL) -+/* this is only safe in thread context of lock owner */ -+#define htree_node_is_listening(lck, dep) \ -+ ((lck)->lk_nodes[dep].ln_mode == HTREE_LOCK_NL) -+ -+#endif diff --git a/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15-22.series b/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15-22.series index 9a4f001..b6ecf97 100644 --- a/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15-22.series +++ b/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15-22.series @@ -13,6 +13,7 @@ suse15/ext4-large-dir.patch suse15/ext4-large-eas.patch suse15/ext4-disable-mb-cache.patch ubuntu18/ext4-nocmtime.patch +base/ext4-htree-lock.patch rhel8/ext4-pdirop.patch suse15/ext4-max-dir-size.patch suse15/ext4-corrupted-inode-block-bitmaps-handling-patches.patch diff --git a/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1-7.series b/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1-7.series index 9a4f001..b6ecf97 100644 --- a/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1-7.series +++ b/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1-7.series @@ -13,6 +13,7 @@ suse15/ext4-large-dir.patch suse15/ext4-large-eas.patch suse15/ext4-disable-mb-cache.patch ubuntu18/ext4-nocmtime.patch +base/ext4-htree-lock.patch rhel8/ext4-pdirop.patch suse15/ext4-max-dir-size.patch suse15/ext4-corrupted-inode-block-bitmaps-handling-patches.patch diff --git a/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1.series b/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1.series index 1d3ad84..a4c2371 100644 --- a/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1.series +++ b/ldiskfs/kernel_patches/series/ldiskfs-4.12-sles15sp1.series @@ -13,6 +13,7 @@ suse15/ext4-large-dir.patch sles15sp1/ext4-large-eas.patch suse15/ext4-disable-mb-cache.patch ubuntu18/ext4-nocmtime.patch +base/ext4-htree-lock.patch sles15sp1/ext4-pdirop.patch suse15/ext4-max-dir-size.patch suse15/ext4-corrupted-inode-block-bitmaps-handling-patches.patch -- 1.8.3.1