summaryrefslogtreecommitdiff
path: root/fs/erofs/utils.c
diff options
context:
space:
mode:
authorGao Xiang <gaoxiang25@huawei.com>2020-02-20 05:46:42 +0300
committerGao Xiang <gaoxiang25@huawei.com>2020-03-03 18:27:25 +0300
commit64094a04414f0b2fb7e31e1c57a67e91e390d66c (patch)
treee39ebae96a3290a69216c187d4298c7da8bdc3a8 /fs/erofs/utils.c
parent98d54f81e36ba3bf92172791eba5ca5bd813989b (diff)
downloadlinux-64094a04414f0b2fb7e31e1c57a67e91e390d66c.tar.xz
erofs: convert workstn to XArray
XArray has friendly APIs and it will replace the old radix tree in the near future. This convert makes use of __xa_cmpxchg when inserting on a just inserted item by other thread. In detail, instead of totally looking up again as what we did for the old radix tree, it will try to legitimize the current in-tree item in the XArray therefore more effective. In addition, naming is rather a challenge for non-English speaker like me. The basic idea of workstn is to provide a runtime sparse array with items arranged in the physical block number order. Such items (was called workgroup) can be used to record compress clusters or for later new features. However, both workgroup and workstn seem not good names from whatever point of view, so I'd like to rename them as pslot and managed_pslots to stand for physical slots. This patch handles the second as a part of the radix tree convert. Cc: Matthew Wilcox <willy@infradead.org> Link: https://lore.kernel.org/r/20200220024642.91529-1-gaoxiang25@huawei.com Reviewed-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Gao Xiang <gaoxiang25@huawei.com>
Diffstat (limited to 'fs/erofs/utils.c')
-rw-r--r--fs/erofs/utils.c88
1 files changed, 32 insertions, 56 deletions
diff --git a/fs/erofs/utils.c b/fs/erofs/utils.c
index fddc5059c930..4396b424373f 100644
--- a/fs/erofs/utils.c
+++ b/fs/erofs/utils.c
@@ -37,9 +37,6 @@ void *erofs_get_pcpubuf(unsigned int pagenr)
/* global shrink count (for all mounted EROFS instances) */
static atomic_long_t erofs_global_shrink_cnt;
-#define __erofs_workgroup_get(grp) atomic_inc(&(grp)->refcount)
-#define __erofs_workgroup_put(grp) atomic_dec(&(grp)->refcount)
-
static int erofs_workgroup_get(struct erofs_workgroup *grp)
{
int o;
@@ -66,7 +63,7 @@ struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
repeat:
rcu_read_lock();
- grp = radix_tree_lookup(&sbi->workstn_tree, index);
+ grp = xa_load(&sbi->managed_pslots, index);
if (grp) {
if (erofs_workgroup_get(grp)) {
/* prefer to relax rcu read side */
@@ -80,43 +77,37 @@ repeat:
return grp;
}
-int erofs_register_workgroup(struct super_block *sb,
- struct erofs_workgroup *grp)
+struct erofs_workgroup *erofs_insert_workgroup(struct super_block *sb,
+ struct erofs_workgroup *grp)
{
- struct erofs_sb_info *sbi;
- int err;
-
- /* grp shouldn't be broken or used before */
- if (atomic_read(&grp->refcount) != 1) {
- DBG_BUGON(1);
- return -EINVAL;
- }
-
- err = radix_tree_preload(GFP_NOFS);
- if (err)
- return err;
-
- sbi = EROFS_SB(sb);
- xa_lock(&sbi->workstn_tree);
+ struct erofs_sb_info *const sbi = EROFS_SB(sb);
+ struct erofs_workgroup *pre;
/*
- * Bump up reference count before making this workgroup
- * visible to other users in order to avoid potential UAF
- * without serialized by workstn_lock.
+ * Bump up a reference count before making this visible
+ * to others for the XArray in order to avoid potential
+ * UAF without serialized by xa_lock.
*/
- __erofs_workgroup_get(grp);
-
- err = radix_tree_insert(&sbi->workstn_tree, grp->index, grp);
- if (err)
- /*
- * it's safe to decrease since the workgroup isn't visible
- * and refcount >= 2 (cannot be freezed).
- */
- __erofs_workgroup_put(grp);
+ atomic_inc(&grp->refcount);
- xa_unlock(&sbi->workstn_tree);
- radix_tree_preload_end();
- return err;
+repeat:
+ xa_lock(&sbi->managed_pslots);
+ pre = __xa_cmpxchg(&sbi->managed_pslots, grp->index,
+ NULL, grp, GFP_NOFS);
+ if (pre) {
+ if (xa_is_err(pre)) {
+ pre = ERR_PTR(xa_err(pre));
+ } else if (erofs_workgroup_get(pre)) {
+ /* try to legitimize the current in-tree one */
+ xa_unlock(&sbi->managed_pslots);
+ cond_resched();
+ goto repeat;
+ }
+ atomic_dec(&grp->refcount);
+ grp = pre;
+ }
+ xa_unlock(&sbi->managed_pslots);
+ return grp;
}
static void __erofs_workgroup_free(struct erofs_workgroup *grp)
@@ -155,7 +146,7 @@ static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
/*
* Note that all cached pages should be unattached
- * before deleted from the radix tree. Otherwise some
+ * before deleted from the XArray. Otherwise some
* cached pages could be still attached to the orphan
* old workgroup when the new one is available in the tree.
*/
@@ -169,7 +160,7 @@ static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
* however in order to avoid some race conditions, add a
* DBG_BUGON to observe this in advance.
*/
- DBG_BUGON(radix_tree_delete(&sbi->workstn_tree, grp->index) != grp);
+ DBG_BUGON(xa_erase(&sbi->managed_pslots, grp->index) != grp);
/*
* If managed cache is on, last refcount should indicate
@@ -182,22 +173,11 @@ static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
unsigned long nr_shrink)
{
- pgoff_t first_index = 0;
- void *batch[PAGEVEC_SIZE];
+ struct erofs_workgroup *grp;
unsigned int freed = 0;
+ unsigned long index;
- int i, found;
-repeat:
- xa_lock(&sbi->workstn_tree);
-
- found = radix_tree_gang_lookup(&sbi->workstn_tree,
- batch, first_index, PAGEVEC_SIZE);
-
- for (i = 0; i < found; ++i) {
- struct erofs_workgroup *grp = batch[i];
-
- first_index = grp->index + 1;
-
+ xa_for_each(&sbi->managed_pslots, index, grp) {
/* try to shrink each valid workgroup */
if (!erofs_try_to_release_workgroup(sbi, grp))
continue;
@@ -206,10 +186,6 @@ repeat:
if (!--nr_shrink)
break;
}
- xa_unlock(&sbi->workstn_tree);
-
- if (i && nr_shrink)
- goto repeat;
return freed;
}