[PATCH 0/2] fsck.erofs: implement concurrent extraction and decompression

[PATCH 0/2] fsck.erofs: implement concurrent extraction and decompression

[Nithurshen]

fsck.erofs currently traverses the filesystem and verifies inodes
synchronously. While data decompression can be offloaded, large
CPU-intensive workloads block the main thread, bottlenecking the
I/O-heavy directory walk. 

This series resolves thread starvation by decoupling traversal
producers from decompression consumers.

Key Changes:
  1/2 Localizes shared global states (fsckcfg.extract_path,
      fsckcfg.dirstack) into task payloads and dispatches them to
      a dedicated traversal worker pool.

  2/2 Introduces a per-CPU worker pool and asynchronous reading
      context to batch and offload decompression of files >128KB,
      utilizing pwrite() for concurrent out-of-order output.

Benchmark setup:
  - Test hosts: Acer Nitro 5 (WSL), Raspberry Pi 4, Mac Book Pro M4 Pro
  - Formats: lz4hc, zstd, lzma; cluster sizes [4096, 8192, 16384, 32768, 65536]
  - 10 runs per configuration

Results (best of 10 runs; ST Base = current master, MT Trav = 1/2,
MT Comb = 1/2 + 2/2; ImpP1 = improvement from 1/2 alone,
ImpP2 = additional improvement from 2/2, ImpAll = total vs master):

Minimum Execution Times (Best of 10 runs)
|Alg  |Clust|ST Base|MT Trav|MT Comb|ImpP1|ImpP2|ImpAll|
|-----|-----|-------|-------|-------|-----|-----|------|
|lz4hc| 4096|  3.80s|  2.75s|  2.74s|27.6%| 0.4%| 27.9%|
|lz4hc| 8192|  3.55s|  2.68s|  2.60s|24.5%| 3.0%| 26.8%|
|lz4hc|16384|  3.43s|  2.59s|  2.58s|24.5%| 0.4%| 24.8%|
|lz4hc|32768|  3.52s|  2.57s|  2.44s|27.0%| 5.1%| 30.7%|
|lz4hc|65536|  3.37s|  2.33s|  2.33s|30.9%| 0.0%| 30.9%|
|zstd | 4096|  4.33s|  3.00s|  2.86s|30.7%| 4.7%| 33.9%|
|zstd | 8192|  4.21s|  2.88s|  2.84s|31.6%| 1.4%| 32.5%|
|zstd |16384|  4.00s|  2.75s|  2.69s|31.2%| 2.2%| 32.8%|
|zstd |32768|  3.98s|  2.61s|  2.58s|34.4%| 1.1%| 35.2%|
|zstd |65536|  3.89s|  2.59s|  2.56s|33.4%| 1.2%| 34.2%|
|lzma | 4096| 51.44s| 33.39s| 14.23s|35.1%|57.4%| 72.3%|
|lzma | 8192| 61.08s| 42.71s| 15.11s|30.1%|64.6%| 75.3%|
|lzma |16384| 65.37s| 47.38s| 15.77s|27.5%|66.7%| 75.9%|
|lzma |32768| 70.51s| 52.73s| 16.44s|25.2%|68.8%| 76.7%|
|lzma |65536| 76.19s| 58.38s| 17.83s|23.4%|69.5%| 76.6%|

Mean +/- Standard Deviation (10 runs)
|Alg  |Clust| ST Base  | MT Trav  | MT Comb  |ImpP1|ImpP2|ImpAll|
|-----|-----|----------|----------|----------|-----|-----|------|
|lz4hc| 4096|5.37+/-6.98s|4.21+/-0.99s|3.30+/-0.85s|21.6%|21.6%| 38.5%|
|lz4hc| 8192|7.42+/-13.7s|3.98+/-0.51s|3.14+/-0.77s|46.4%|21.1%| 57.7%|
|lz4hc|16384|4.06+/-0.95s|3.27+/-0.86s|3.02+/-0.77s|19.5%| 7.6%| 25.6%|
|lz4hc|32768|5.33+/-3.29s|2.88+/-0.47s|2.76+/-0.28s|46.0%| 4.2%| 48.2%|
|lz4hc|65536|4.40+/-2.11s|3.11+/-1.19s|2.65+/-0.56s|29.4%|14.7%| 39.8%|
|zstd | 4096|5.96+/-3.65s|4.01+/-1.97s|3.08+/-0.36s|32.7%|23.2%| 48.3%|
|zstd | 8192|4.99+/-2.23s|3.08+/-0.33s|3.03+/-0.16s|38.4%| 1.7%| 39.4%|
|zstd |16384|5.42+/-4.06s|3.13+/-0.53s|3.03+/-0.46s|42.2%| 3.1%| 44.0%|
|zstd |32768|5.34+/-3.90s|3.96+/-3.19s|3.05+/-0.52s|25.8%|23.0%| 42.9%|
|zstd |65536|6.08+/-4.36s|3.86+/-3.24s|2.87+/-0.39s|36.5%|25.6%| 52.8%|
|lzma | 4096|52.16+/-0.6s|34.38+/-2.1s|16.06+/-3.1s|34.1%|53.3%| 69.2%|
|lzma | 8192|61.45+/-0.4s|43.34+/-0.9s|15.85+/-0.6s|29.5%|63.4%| 74.2%|
|lzma |16384|65.69+/-0.5s|47.78+/-0.4s|16.30+/-0.4s|27.3%|65.9%| 75.2%|
|lzma |32768|70.79+/-0.3s|53.02+/-0.3s|17.15+/-0.4s|25.1%|67.6%| 75.8%|
|lzma |65536|76.50+/-0.4s|58.71+/-0.3s|18.10+/-0.2s|23.3%|69.2%| 76.3%|

Summary of Improvements:
Patch 1/2 alone provides a reliable 24-35% reduction in minimum
execution times across all tested formats and cluster sizes simply by
parallelizing the traversal walk. 

Patch 2/2 targets CPU-heavy decompression, delivering a massive
secondary impact for LZMA. LZMA extraction drops by a further
57-69% on top of patch 1/2, resulting in a 72-76% total execution
time reduction compared to the current master. Lz4hc and zstd are
fast enough to bypass the background worker pool, so their ImpP2
remains negligible, as intended.

Nithurshen (2):
  fsck.erofs: implement concurrent directory traversal
  fsck.erofs: add multi-threaded decompression

 fsck/main.c              | 580 +++++++++++++++++++++++++++++----------
 include/erofs/cond.h     |  31 +++
 include/erofs/internal.h |  16 +-
 include/erofs/lock.h     |   2 +
 lib/data.c               | 377 +++++++++++++++++++++++--
 lib/global.c             |  12 +-
 6 files changed, 848 insertions(+), 170 deletions(-)
 create mode 100644 include/erofs/cond.h

-- 
2.52.0

[PATCH 1/2] fsck.erofs: implement concurrent directory traversal

[Nithurshen]

Currently, fsck.erofs traverses the filesystem tree and verifies
inodes synchronously on the main thread. While data decompression is
offloaded, the main thread remains a bottleneck during the I/O-heavy
directory walk.

This patch parallelizes the directory traversal and inode extraction
processes. To achieve this safely, globally shared states such as
fsckcfg.extract_path and fsckcfg.dirstack are decoupled and localized
into individual struct erofsfsck_inode_task payloads. These payloads
are dispatched to a dedicated traversal worker pool.

Global statistics and hardlink tables are now secured using native
erofs_mutex_t primitives. By isolating the traversal producers from
the pcluster decompression consumers, the pipeline scales across
directories without thread pool starvation.

Signed-off-by: Nithurshen <nithurshen.dev@gmail.com>
---
 fsck/main.c          | 465 ++++++++++++++++++++++++++++++++-----------
 include/erofs/cond.h |  31 +++
 include/erofs/lock.h |   2 +
 3 files changed, 382 insertions(+), 116 deletions(-)
 create mode 100644 include/erofs/cond.h

diff --git a/fsck/main.c b/fsck/main.c
index b2d8f1a5..a520af91 100644
--- a/fsck/main.c
+++ b/fsck/main.c
@@ -13,25 +13,51 @@
 #include "erofs/decompress.h"
 #include "erofs/dir.h"
 #include "erofs/xattr.h"
+#include "erofs/internal.h"
+#include "erofs/lock.h"
+#include "erofs/cond.h"
 #include "../lib/compressor.h"
 #include "../lib/liberofs_compress.h"
 #include "../lib/liberofs_sha256.h"
 
-static int erofsfsck_check_inode(erofs_nid_t pnid, erofs_nid_t nid);
-
-static char erofsfsck_nullstr[] = "";
-
 struct erofsfsck_dirstack {
 	erofs_nid_t dirs[PATH_MAX];
 	int top;
 };
 
-struct erofsfsck_cfg {
+static int erofsfsck_check_inode(erofs_nid_t pnid, erofs_nid_t nid,
+	const char *path, struct erofsfsck_dirstack *dirstack);
+
+#ifdef EROFS_MT_ENABLED
+struct erofsfsck_inode_task {
+	struct erofsfsck_inode_task *next;
+	erofs_nid_t pnid;
+	erofs_nid_t nid;
+	char *path;
 	struct erofsfsck_dirstack dirstack;
+};
+
+static erofs_mutex_t traverse_mtx;
+static erofs_cond_t traverse_cond;
+static erofs_cond_t traverse_cond_finish;
+static erofs_mutex_t hardlink_mtx;
+static erofs_mutex_t stats_mtx;
+
+static struct erofsfsck_inode_task *traverse_head;
+static struct erofsfsck_inode_task *traverse_tail;
+static int traverse_pending_tasks;
+static bool traverse_shutdown;
+static pthread_t *traverse_workers;
+static int traverse_num_workers;
+static int traverse_final_err;
+#endif
+
+static char erofsfsck_nullstr[] = "";
+
+struct erofsfsck_cfg {
 	u64 physical_blocks;
 	u64 logical_blocks;
 	char *extract_path;
-	size_t extract_pos;
 	char *digest_xattr_name;
 	mode_t umask;
 	bool superuser;
@@ -205,7 +231,6 @@ static int erofsfsck_parse_options_cfg(int argc, char **argv)
 				/* if path is root, start writing from position 0 */
 				if (len == 1 && fsckcfg.extract_path[0] == '/')
 					len = 0;
-				fsckcfg.extract_pos = len;
 			}
 			break;
 		case 3:
@@ -276,7 +301,7 @@ static int erofsfsck_parse_options_cfg(int argc, char **argv)
 	}
 
 	if (fsckcfg.extract_path) {
-		if (!fsckcfg.extract_pos && !fsckcfg.force) {
+		if (!strcmp(fsckcfg.extract_path, "/") && !fsckcfg.force) {
 			erofs_err("--extract=/ must be used together with --force");
 			return -EINVAL;
 		}
@@ -431,7 +456,7 @@ out:
 	return ret;
 }
 
-static int erofsfsck_dump_xattrs(struct erofs_inode *inode)
+static int erofsfsck_dump_xattrs(struct erofs_inode *inode, const char *path)
 {
 	static bool ignore_xattrs = false;
 	char *keylst, *key;
@@ -477,9 +502,8 @@ static int erofsfsck_dump_xattrs(struct erofs_inode *inode)
 			free(value);
 			break;
 		}
-		if (fsckcfg.extract_path)
-			ret = erofs_sys_lsetxattr(fsckcfg.extract_path, key,
-						  value, size);
+		if (path)
+			ret = erofs_sys_lsetxattr(path, key, value, size);
 		else
 			ret = 0;
 		free(value);
@@ -648,9 +672,19 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 	}
 
 	if (fsckcfg.print_comp_ratio) {
+		u64 log_blocks = 0;
 		if (!erofs_is_packed_inode(inode))
-			fsckcfg.logical_blocks += BLK_ROUND_UP(inode->sbi, inode->i_size);
-		fsckcfg.physical_blocks += BLK_ROUND_UP(inode->sbi, pchunk_len);
+			log_blocks = BLK_ROUND_UP(inode->sbi, inode->i_size);
+		u64 phys_blocks = BLK_ROUND_UP(inode->sbi, pchunk_len);
+
+#ifdef EROFS_MT_ENABLED
+		erofs_mutex_lock(&stats_mtx);
+#endif
+		fsckcfg.logical_blocks += log_blocks;
+		fsckcfg.physical_blocks += phys_blocks;
+#ifdef EROFS_MT_ENABLED
+		erofs_mutex_unlock(&stats_mtx);
+#endif
 	}
 out:
 	if (raw)
@@ -666,11 +700,11 @@ fail_eio:
 	goto out;
 }
 
-static inline int erofs_extract_dir(struct erofs_inode *inode)
+static inline int erofs_extract_dir(struct erofs_inode *inode, const char *path)
 {
 	int ret;
 
-	erofs_dbg("create directory %s", fsckcfg.extract_path);
+	erofs_dbg("create directory %s", path);
 
 	/* verify data chunk layout */
 	ret = erofs_verify_inode_data(inode, -1, NULL);
@@ -683,19 +717,18 @@ static inline int erofs_extract_dir(struct erofs_inode *inode)
 	 * write/execute permission.  These are fixed up later in
 	 * erofsfsck_set_attributes().
 	 */
-	if (mkdir(fsckcfg.extract_path, 0700) < 0) {
+	if (mkdir(path, 0700) < 0) {
 		struct stat st;
 
 		if (errno != EEXIST) {
 			erofs_err("failed to create directory: %s (%s)",
-				  fsckcfg.extract_path, strerror(errno));
+				  path, strerror(errno));
 			return -errno;
 		}
 
-		if (lstat(fsckcfg.extract_path, &st) ||
+		if (lstat(path, &st) ||
 		    !S_ISDIR(st.st_mode)) {
-			erofs_err("path is not a directory: %s",
-				  fsckcfg.extract_path);
+			erofs_err("path is not a directory: %s", path);
 			return -ENOTDIR;
 		}
 
@@ -703,25 +736,185 @@ static inline int erofs_extract_dir(struct erofs_inode *inode)
 		 * Try to change permissions of existing directory so
 		 * that we can write to it
 		 */
-		if (chmod(fsckcfg.extract_path, 0700) < 0) {
+		if (chmod(path, 0700) < 0) {
 			erofs_err("failed to set permissions: %s (%s)",
-				  fsckcfg.extract_path, strerror(errno));
+				path, strerror(errno));
 			return -errno;
 		}
 	}
 	return 0;
 }
 
+static int erofsfsck_enqueue_task(erofs_nid_t pnid, erofs_nid_t nid,
+	const char *path, struct erofsfsck_dirstack *dirstack)
+{
+#ifdef EROFS_MT_ENABLED
+	if (traverse_num_workers > 1) {
+		struct erofsfsck_inode_task *task = malloc(sizeof(*task));
+
+		if (!task)
+			return -ENOMEM;
+
+		task->pnid = pnid;
+		task->nid = nid;
+		task->path = path ? strdup(path) : NULL;
+		task->dirstack = *dirstack;
+		task->next = NULL;
+
+		erofs_mutex_lock(&traverse_mtx);
+		if (traverse_final_err) {
+			erofs_mutex_unlock(&traverse_mtx);
+			if (task->path)
+				free(task->path);
+			free(task);
+			return traverse_final_err;
+		}
+
+		if (!traverse_tail)
+			traverse_head = traverse_tail = task;
+		else {
+			traverse_tail->next = task;
+			traverse_tail = task;
+		}
+		traverse_pending_tasks++;
+		erofs_cond_signal(&traverse_cond);
+		erofs_mutex_unlock(&traverse_mtx);
+		return 0;
+	}
+#endif
+	int err = erofsfsck_check_inode(pnid, nid, path, dirstack);
+#ifdef EROFS_MT_ENABLED
+	if (err && !traverse_final_err)
+		traverse_final_err = err;
+#endif
+	return err;
+}
+
+#ifdef EROFS_MT_ENABLED
+static void *erofsfsck_traverse_worker(void *arg)
+{
+	while (1) {
+		struct erofsfsck_inode_task *task;
+
+		erofs_mutex_lock(&traverse_mtx);
+		while (!traverse_head && !traverse_shutdown)
+			erofs_cond_wait(&traverse_cond, &traverse_mtx);
+
+		if (traverse_shutdown && !traverse_head) {
+			erofs_mutex_unlock(&traverse_mtx);
+			break;
+		}
+
+		task = traverse_head;
+		traverse_head = task->next;
+		if (!traverse_head)
+			traverse_tail = NULL;
+		erofs_mutex_unlock(&traverse_mtx);
+
+		int err = erofsfsck_check_inode(task->pnid, task->nid, task->path, &task->dirstack);
+
+		erofs_mutex_lock(&traverse_mtx);
+		if (err && !traverse_final_err)
+			traverse_final_err = err;
+		traverse_pending_tasks--;
+		if (traverse_pending_tasks == 0)
+			erofs_cond_signal(&traverse_cond_finish);
+		erofs_mutex_unlock(&traverse_mtx);
+
+		if (task->path)
+			free(task->path);
+		free(task);
+	}
+	return NULL;
+}
+#endif
+
+static int erofsfsck_traverse_mt_init(void)
+{
+#ifdef EROFS_MT_ENABLED
+	erofs_mutex_init(&traverse_mtx);
+	erofs_cond_init(&traverse_cond);
+	erofs_cond_init(&traverse_cond_finish);
+	erofs_mutex_init(&hardlink_mtx);
+	erofs_mutex_init(&stats_mtx);
+
+	traverse_num_workers = erofs_get_available_processors();
+	if (traverse_num_workers <= 1)
+		return 0;
+
+	traverse_workers = calloc(traverse_num_workers, sizeof(pthread_t));
+	if (!traverse_workers)
+		return -ENOMEM;
+
+	for (int i = 0; i < traverse_num_workers; i++)
+		pthread_create(&traverse_workers[i], NULL, erofsfsck_traverse_worker, NULL);
+#endif
+	return 0;
+}
+
+static void erofsfsck_traverse_mt_exit(void)
+{
+#ifdef EROFS_MT_ENABLED
+	if (traverse_workers) {
+		erofs_mutex_lock(&traverse_mtx);
+		traverse_shutdown = true;
+		erofs_cond_broadcast(&traverse_cond);
+		erofs_mutex_unlock(&traverse_mtx);
+
+		for (int i = 0; i < traverse_num_workers; i++)
+			pthread_join(traverse_workers[i], NULL);
+
+		free(traverse_workers);
+		traverse_workers = NULL;
+	}
+	erofs_mutex_destroy(&traverse_mtx);
+	erofs_cond_destroy(&traverse_cond);
+	erofs_cond_destroy(&traverse_cond_finish);
+	erofs_mutex_destroy(&hardlink_mtx);
+	erofs_mutex_destroy(&stats_mtx);
+#endif
+}
+
+static int erofsfsck_traverse_mt_wait(void)
+{
+#ifdef EROFS_MT_ENABLED
+	if (traverse_num_workers <= 1)
+		return traverse_final_err;
+
+	erofs_mutex_lock(&traverse_mtx);
+	while (traverse_pending_tasks > 0)
+		erofs_cond_wait(&traverse_cond_finish, &traverse_mtx);
+
+	int err = traverse_final_err;
+
+	erofs_mutex_unlock(&traverse_mtx);
+	return err;
+#else
+	return 0;
+#endif
+}
+
 static char *erofsfsck_hardlink_find(erofs_nid_t nid)
 {
 	struct list_head *head =
 			&erofsfsck_link_hashtable[nid % NR_HARDLINK_HASHTABLE];
 	struct erofsfsck_hardlink_entry *entry;
+	char *path = NULL;
+
+#ifdef EROFS_MT_ENABLED
+	erofs_mutex_lock(&hardlink_mtx);
+#endif
+	list_for_each_entry(entry, head, list) {
+		if (entry->nid == nid) {
+			path = entry->path;
+			break;
+		}
+	}
+#ifdef EROFS_MT_ENABLED
+	erofs_mutex_unlock(&hardlink_mtx);
+#endif
+	return path;
 
-	list_for_each_entry(entry, head, list)
-		if (entry->nid == nid)
-			return entry->path;
-	return NULL;
 }
 
 static int erofsfsck_hardlink_insert(erofs_nid_t nid, const char *path)
@@ -739,8 +932,17 @@ static int erofsfsck_hardlink_insert(erofs_nid_t nid, const char *path)
 		return -ENOMEM;
 	}
 
+#ifdef EROFS_MT_ENABLED
+	erofs_mutex_lock(&hardlink_mtx);
+#endif
+
 	list_add_tail(&entry->list,
 		      &erofsfsck_link_hashtable[nid % NR_HARDLINK_HASHTABLE]);
+
+#ifdef EROFS_MT_ENABLED
+	erofs_mutex_unlock(&hardlink_mtx);
+#endif
+
 	return 0;
 }
 
@@ -819,38 +1021,36 @@ static int erofsfsck_calc_inode_data(struct erofs_inode *inode, int outfd)
 	return erofs_verify_inode_data(inode, outfd, NULL);
 }
 
-static inline int erofs_extract_file(struct erofs_inode *inode)
+static inline int erofs_extract_file(struct erofs_inode *inode, const char *path)
 {
 	bool tryagain = true;
 	int ret, fd;
 
-	erofs_dbg("extract file to path: %s", fsckcfg.extract_path);
+	erofs_dbg("extract file to path: %s", path);
 
 again:
-	fd = open(fsckcfg.extract_path,
+	fd = open(path,
 		  O_WRONLY | O_CREAT | O_NOFOLLOW |
 			(fsckcfg.overwrite ? O_TRUNC : O_EXCL), 0700);
 	if (fd < 0) {
 		if (fsckcfg.overwrite && tryagain) {
 			if (errno == EISDIR) {
-				erofs_warn("try to forcely remove directory %s",
-					   fsckcfg.extract_path);
-				if (rmdir(fsckcfg.extract_path) < 0) {
-					erofs_err("failed to remove: %s (%s)",
-						  fsckcfg.extract_path, strerror(errno));
+				erofs_warn("try to forcely remove directory %s", path);
+				if (rmdir(path) < 0) {
+					erofs_err("failed to remove: %s (%s)", path,
+						strerror(errno));
 					return -EISDIR;
 				}
 			} else if (errno == EACCES &&
-				   chmod(fsckcfg.extract_path, 0700) < 0) {
-				erofs_err("failed to set permissions: %s (%s)",
-					  fsckcfg.extract_path, strerror(errno));
+				   chmod(path, 0700) < 0) {
+				erofs_err("failed to set permissions: %s (%s)", path,
+					strerror(errno));
 				return -errno;
 			}
 			tryagain = false;
 			goto again;
 		}
-		erofs_err("failed to open: %s (%s)", fsckcfg.extract_path,
-			  strerror(errno));
+		erofs_err("failed to open: %s (%s)", path, strerror(errno));
 		return -errno;
 	}
 
@@ -859,7 +1059,7 @@ again:
 	return ret;
 }
 
-static inline int erofs_extract_symlink(struct erofs_inode *inode)
+static inline int erofs_extract_symlink(struct erofs_inode *inode, const char *path)
 {
 	struct erofs_vfile vf;
 	bool tryagain = true;
@@ -867,7 +1067,7 @@ static inline int erofs_extract_symlink(struct erofs_inode *inode)
 	int ret;
 	char *buf = NULL;
 
-	erofs_dbg("extract symlink to path: %s", fsckcfg.extract_path);
+	erofs_dbg("extract symlink to path: %s", path);
 
 	/* verify data chunk layout */
 	ret = erofs_verify_inode_data(inode, -1, NULL);
@@ -893,21 +1093,18 @@ static inline int erofs_extract_symlink(struct erofs_inode *inode)
 
 	buf[inode->i_size] = '\0';
 again:
-	if (symlink(buf, fsckcfg.extract_path) < 0) {
+	if (symlink(buf, path) < 0) {
 		if (errno == EEXIST && fsckcfg.overwrite && tryagain) {
-			erofs_warn("try to forcely remove file %s",
-				   fsckcfg.extract_path);
-			if (unlink(fsckcfg.extract_path) < 0) {
-				erofs_err("failed to remove: %s",
-					  fsckcfg.extract_path);
+			erofs_warn("try to forcely remove file %s", path);
+			if (unlink(path) < 0) {
+				erofs_err("failed to remove: %s", path);
 				ret = -errno;
 				goto out;
 			}
 			tryagain = false;
 			goto again;
 		}
-		erofs_err("failed to create symlink: %s",
-			  fsckcfg.extract_path);
+		erofs_err("failed to create symlink: %s", path);
 		ret = -errno;
 	}
 out:
@@ -916,12 +1113,12 @@ out:
 	return ret;
 }
 
-static int erofs_extract_special(struct erofs_inode *inode)
+static int erofs_extract_special(struct erofs_inode *inode, const char *path)
 {
 	bool tryagain = true;
 	int ret;
 
-	erofs_dbg("extract special to path: %s", fsckcfg.extract_path);
+	erofs_dbg("extract special to path: %s", path);
 
 	/* verify data chunk layout */
 	ret = erofs_verify_inode_data(inode, -1, NULL);
@@ -929,25 +1126,21 @@ static int erofs_extract_special(struct erofs_inode *inode)
 		return ret;
 
 again:
-	if (mknod(fsckcfg.extract_path, inode->i_mode, inode->u.i_rdev) < 0) {
+	if (mknod(path, inode->i_mode, inode->u.i_rdev) < 0) {
 		if (errno == EEXIST && fsckcfg.overwrite && tryagain) {
-			erofs_warn("try to forcely remove file %s",
-				   fsckcfg.extract_path);
-			if (unlink(fsckcfg.extract_path) < 0) {
-				erofs_err("failed to remove: %s",
-					  fsckcfg.extract_path);
+			erofs_warn("try to forcely remove file %s", path);
+			if (unlink(path) < 0) {
+				erofs_err("failed to remove: %s", path);
 				return -errno;
 			}
 			tryagain = false;
 			goto again;
 		}
 		if (errno == EEXIST || fsckcfg.superuser) {
-			erofs_err("failed to create special file: %s",
-				  fsckcfg.extract_path);
+			erofs_err("failed to create special file: %s", path);
 			ret = -errno;
 		} else {
-			erofs_warn("failed to create special file: %s, skipped",
-				   fsckcfg.extract_path);
+			erofs_warn("failed to create special file: %s, skipped", path);
 			ret = -ECANCELED;
 		}
 	}
@@ -970,42 +1163,53 @@ static int erofsfsck_get_parent_cb(struct erofs_dir_context *ctx)
 	return 0;
 }
 
+struct erofsfsck_dir_ctx {
+	struct erofs_dir_context ctx;
+	const char *path;
+	struct erofsfsck_dirstack *dirstack;
+};
+
 static int erofsfsck_dirent_iter(struct erofs_dir_context *ctx)
 {
+	struct erofsfsck_dir_ctx *fctx = (void *)ctx;
 	int ret;
-	size_t prev_pos, curr_pos;
+	char *newpath = NULL;
 
 	if (ctx->dot_dotdot)
 		return 0;
 
-	prev_pos = fsckcfg.extract_pos;
-	curr_pos = prev_pos;
+	if (fctx->path) {
+		size_t prev_len = strlen(fctx->path);
+		size_t curr_len = prev_len + ctx->de_namelen + 1;
 
-	if (prev_pos + ctx->de_namelen + 1 >= PATH_MAX) {
-		erofs_err("unable to fsck since the path is too long (%llu)",
-			  (curr_pos + ctx->de_namelen + 1) | 0ULL);
-		return -EOPNOTSUPP;
-	}
+		if (curr_len >= PATH_MAX) {
+			erofs_err("unable to fsck since the path is too long (%llu)",
+				(unsigned long long)curr_len);
+			return -EOPNOTSUPP;
+		}
+		newpath = malloc(curr_len + 1);
+		if (!newpath)
+			return -ENOMEM;
 
-	if (fsckcfg.extract_path) {
-		fsckcfg.extract_path[curr_pos++] = '/';
-		strncpy(fsckcfg.extract_path + curr_pos, ctx->dname,
-			ctx->de_namelen);
-		curr_pos += ctx->de_namelen;
-		fsckcfg.extract_path[curr_pos] = '\0';
-	} else {
-		curr_pos += ctx->de_namelen + 1;
+		bool is_root = (fctx->path[0] == '/' && fctx->path[1] == '\0');
+		size_t offset = prev_len;
+
+		memcpy(newpath, fctx->path, prev_len);
+		if (!is_root)
+			newpath[offset++] = '/';
+
+		memcpy(newpath + offset, ctx->dname, ctx->de_namelen);
+		offset += ctx->de_namelen;
+		newpath[offset] = '\0';
 	}
-	fsckcfg.extract_pos = curr_pos;
-	ret = erofsfsck_check_inode(ctx->dir->nid, ctx->de_nid);
 
-	if (fsckcfg.extract_path)
-		fsckcfg.extract_path[prev_pos] = '\0';
-	fsckcfg.extract_pos = prev_pos;
+	ret = erofsfsck_enqueue_task(ctx->dir->nid, ctx->de_nid, newpath, fctx->dirstack);
+	if (newpath)
+		free(newpath);
 	return ret;
 }
 
-static int erofsfsck_extract_inode(struct erofs_inode *inode)
+static int erofsfsck_extract_inode(struct erofs_inode *inode, const char *path)
 {
 	char *oldpath;
 	int ret;
@@ -1018,9 +1222,8 @@ verify:
 
 	oldpath = erofsfsck_hardlink_find(inode->nid);
 	if (oldpath) {
-		if (link(oldpath, fsckcfg.extract_path) == -1) {
-			erofs_err("failed to extract hard link: %s (%s)",
-				  fsckcfg.extract_path, strerror(errno));
+		if (link(oldpath, path) == -1) {
+			erofs_err("failed to extract hard link: %s (%s)", path, strerror(errno));
 			return -errno;
 		}
 		return 0;
@@ -1028,19 +1231,19 @@ verify:
 
 	switch (inode->i_mode & S_IFMT) {
 	case S_IFDIR:
-		ret = erofs_extract_dir(inode);
+		ret = erofs_extract_dir(inode, path);
 		break;
 	case S_IFREG:
-		ret = erofs_extract_file(inode);
+		ret = erofs_extract_file(inode, path);
 		break;
 	case S_IFLNK:
-		ret = erofs_extract_symlink(inode);
+		ret = erofs_extract_symlink(inode, path);
 		break;
 	case S_IFCHR:
 	case S_IFBLK:
 	case S_IFIFO:
 	case S_IFSOCK:
-		ret = erofs_extract_special(inode);
+		ret = erofs_extract_special(inode, path);
 		break;
 	default:
 		erofs_warn("unsupported file type %o @ nid %llu, skipped extraction",
@@ -1053,12 +1256,12 @@ verify:
 
 	/* record nid and old path for hardlink */
 	if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
-		ret = erofsfsck_hardlink_insert(inode->nid,
-						fsckcfg.extract_path);
+		ret = erofsfsck_hardlink_insert(inode->nid, path);
 	return ret;
 }
 
-static int erofsfsck_check_inode(erofs_nid_t pnid, erofs_nid_t nid)
+static int erofsfsck_check_inode(erofs_nid_t pnid, erofs_nid_t nid,
+	const char *path, struct erofsfsck_dirstack *dirstack)
 {
 	int ret, i;
 	struct erofs_inode inode = {.sbi = &g_sbi, .nid = nid};
@@ -1079,43 +1282,57 @@ static int erofsfsck_check_inode(erofs_nid_t pnid, erofs_nid_t nid)
 			goto out;
 	}
 
-	ret = erofsfsck_extract_inode(&inode);
+	ret = erofsfsck_extract_inode(&inode, path);
 	if (ret && ret != -ECANCELED)
 		goto out;
 
 	if (fsckcfg.check_decomp && fsckcfg.dump_xattrs) {
-		ret = erofsfsck_dump_xattrs(&inode);
+		ret = erofsfsck_dump_xattrs(&inode, path);
 		if (ret)
 			return ret;
 	}
 
 	if (S_ISDIR(inode.i_mode)) {
-		struct erofs_dir_context ctx = {
-			.flags = EROFS_READDIR_VALID_PNID,
-			.pnid = pnid,
-			.dir = &inode,
-			.cb = erofsfsck_dirent_iter,
+		struct erofsfsck_dir_ctx ctx = {
+			.ctx.flags = EROFS_READDIR_VALID_PNID,
+			.ctx.pnid = pnid,
+			.ctx.dir = &inode,
+			.ctx.cb = erofsfsck_dirent_iter,
+			.path = path,
+			.dirstack = dirstack,
 		};
+		struct erofsfsck_dirstack next_dirstack;
 
-		/* XXX: support the deeper cases later */
-		if (fsckcfg.dirstack.top >= ARRAY_SIZE(fsckcfg.dirstack.dirs))
+		if (dirstack->top >= ARRAY_SIZE(dirstack->dirs))
 			return -ENAMETOOLONG;
-		for (i = 0; i < fsckcfg.dirstack.top; ++i)
-			if (inode.nid == fsckcfg.dirstack.dirs[i])
+
+		for (i = 0; i < dirstack->top; ++i) {
+			if (inode.nid == dirstack->dirs[i])
 				return -ELOOP;
-		fsckcfg.dirstack.dirs[fsckcfg.dirstack.top++] = pnid;
-		ret = erofs_iterate_dir(&ctx, true);
-		--fsckcfg.dirstack.top;
+		}
+
+		next_dirstack = *dirstack;
+		next_dirstack.dirs[next_dirstack.top++] = pnid;
+		ctx.dirstack = &next_dirstack;
+
+		ret = erofs_iterate_dir(&ctx.ctx, true);
+		if (ret == -ECANCELED)
+			ret = 0;
 	}
 
 	if (!ret && !erofs_is_packed_inode(&inode))
-		erofsfsck_set_attributes(&inode, fsckcfg.extract_path);
+		erofsfsck_set_attributes(&inode, (char *)path);
 
-	if (ret == -ECANCELED)
-		ret = 0;
 out:
-	if (ret && ret != -EIO)
+	if (ret && ret != -EIO) {
+#ifdef EROFS_MT_ENABLED
+		erofs_mutex_lock(&traverse_mtx);
+#endif
 		fsckcfg.corrupted = true;
+#ifdef EROFS_MT_ENABLED
+		erofs_mutex_unlock(&traverse_mtx);
+#endif
+	}
 	return ret;
 }
 
@@ -1132,7 +1349,6 @@ int main(int argc, char *argv[])
 	fsckcfg.physical_blocks = 0;
 	fsckcfg.logical_blocks = 0;
 	fsckcfg.extract_path = NULL;
-	fsckcfg.extract_pos = 0;
 	fsckcfg.umask = umask(0);
 	fsckcfg.superuser = geteuid() == 0;
 	fsckcfg.corrupted = false;
@@ -1199,6 +1415,10 @@ int main(int argc, char *argv[])
 	if (fsckcfg.extract_path)
 		erofsfsck_hardlink_init();
 
+	err = erofsfsck_traverse_mt_init();
+	if (err)
+		goto exit_hardlink;
+
 	if (fsckcfg.inode_path) {
 		struct erofs_inode inode = { .sbi = &g_sbi };
 
@@ -1218,10 +1438,12 @@ int main(int argc, char *argv[])
 			if (err) {
 				erofs_err("failed to initialize packedfile: %s",
 					  erofs_strerror(err));
-				goto exit_hardlink;
+				goto exit_mt_exit;
 			}
+			struct erofsfsck_dirstack empty_dirstack = {0};
 
-			err = erofsfsck_check_inode(g_sbi.packed_nid, g_sbi.packed_nid);
+			err = erofsfsck_enqueue_task(g_sbi.packed_nid, g_sbi.packed_nid,
+				NULL, &empty_dirstack);
 			if (err) {
 				erofs_err("failed to verify packed file");
 				goto exit_packedinode;
@@ -1232,6 +1454,7 @@ int main(int argc, char *argv[])
 	{
 		erofs_nid_t pnid = fsckcfg.nid;
 
+
 		if (fsckcfg.nid != g_sbi.root_nid) {
 			struct erofs_inode inode = { .sbi = &g_sbi, .nid = fsckcfg.nid };
 
@@ -1246,7 +1469,15 @@ int main(int argc, char *argv[])
 					pnid = ctx.pnid;
 			}
 		}
-		err = erofsfsck_check_inode(pnid, fsckcfg.nid);
+		struct erofsfsck_dirstack empty_dirstack = {0};
+
+		err = erofsfsck_enqueue_task(pnid, fsckcfg.nid, fsckcfg.extract_path,
+			&empty_dirstack);
+
+		int wait_err = erofsfsck_traverse_mt_wait();
+
+		if (wait_err && !err)
+			err = wait_err;
 	}
 
 	if (fsckcfg.corrupted) {
@@ -1272,6 +1503,8 @@ int main(int argc, char *argv[])
 
 exit_packedinode:
 	erofs_packedfile_exit(&g_sbi);
+exit_mt_exit:
+	erofsfsck_traverse_mt_exit();
 exit_hardlink:
 	if (fsckcfg.extract_path)
 		erofsfsck_hardlink_exit();
diff --git a/include/erofs/cond.h b/include/erofs/cond.h
new file mode 100644
index 00000000..c5e20610
--- /dev/null
+++ b/include/erofs/cond.h
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: GPL-2.0+ OR Apache-2.0 */
+#ifndef __EROFS_COND_H
+#define __EROFS_COND_H
+
+#include "lock.h"
+
+#if defined(HAVE_PTHREAD_H) && defined(EROFS_MT_ENABLED)
+#include <pthread.h>
+
+typedef pthread_cond_t erofs_cond_t;
+
+static inline void erofs_cond_init(erofs_cond_t *cond)
+{
+	pthread_cond_init(cond, NULL);
+}
+#define erofs_cond_wait		pthread_cond_wait
+#define erofs_cond_signal	pthread_cond_signal
+#define erofs_cond_broadcast	pthread_cond_broadcast
+#define erofs_cond_destroy	pthread_cond_destroy
+
+#else
+typedef struct {} erofs_cond_t;
+
+static inline void erofs_cond_init(erofs_cond_t *cond) {}
+static inline int erofs_cond_wait(erofs_cond_t *cond, erofs_mutex_t *mutex) { return 0; }
+static inline int erofs_cond_signal(erofs_cond_t *cond) { return 0; }
+static inline int erofs_cond_broadcast(erofs_cond_t *cond) { return 0; }
+static inline int erofs_cond_destroy(erofs_cond_t *cond) { return 0; }
+#endif
+
+#endif
diff --git a/include/erofs/lock.h b/include/erofs/lock.h
index 884f23ea..673465fb 100644
--- a/include/erofs/lock.h
+++ b/include/erofs/lock.h
@@ -15,6 +15,7 @@ static inline void erofs_mutex_init(erofs_mutex_t *lock)
 }
 #define erofs_mutex_lock	pthread_mutex_lock
 #define erofs_mutex_unlock	pthread_mutex_unlock
+#define erofs_mutex_destroy pthread_mutex_destroy
 
 #define EROFS_DEFINE_MUTEX(lock)	\
 	erofs_mutex_t lock = PTHREAD_MUTEX_INITIALIZER
@@ -35,6 +36,7 @@ typedef struct {} erofs_mutex_t;
 static inline void erofs_mutex_init(erofs_mutex_t *lock) {}
 static inline void erofs_mutex_lock(erofs_mutex_t *lock) {}
 static inline void erofs_mutex_unlock(erofs_mutex_t *lock) {}
+static inline void erofs_mutex_destroy(erofs_mutex_t *lock) {}
 
 #define EROFS_DEFINE_MUTEX(lock)	\
 	erofs_mutex_t lock = {}
-- 
2.52.0

[PATCH 2/2] fsck.erofs: add multi-threaded decompression

[Nithurshen]

Currently, fsck.erofs verifies and extracts inode data synchronously
during directory traversal. While directory walking can now run
concurrently across worker threads, CPU-intensive decompression (such
as LZMA or DEFLATE) for large files still blocks traversal threads,
creating a processing bottleneck during extraction and verification.

This patch introduces a multi-threaded decompression pipeline using a
per-CPU worker pool and an asynchronous reading context
(struct z_erofs_mt_read_ctx). For large compressed files exceeding
EROFS_MT_THRESHOLD (128KB), decompression workloads are batched and
offloaded to dedicated background workers.

Decompressed chunks are directly written out at calculated offsets via
pwrite(), allowing concurrent out-of-order decompression while
maintaining output correctness. Global initialization and teardown of
the per-CPU decompression workers are wired into liberofs_global_init()
and liberofs_global_exit().

Signed-off-by: Nithurshen <nithurshen.dev@gmail.com>
---
 fsck/main.c              | 117 +++++++++---
 include/erofs/internal.h |  16 +-
 lib/data.c               | 377 ++++++++++++++++++++++++++++++++++++---
 lib/global.c             |  12 +-
 4 files changed, 467 insertions(+), 55 deletions(-)

diff --git a/fsck/main.c b/fsck/main.c
index a520af91..0336335b 100644
--- a/fsck/main.c
+++ b/fsck/main.c
@@ -20,6 +20,8 @@
 #include "../lib/liberofs_compress.h"
 #include "../lib/liberofs_sha256.h"
 
+#define EROFS_MT_THRESHOLD (128 * 1024)
+
 struct erofsfsck_dirstack {
 	erofs_nid_t dirs[PATH_MAX];
 	int top;
@@ -550,6 +552,7 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 	u64 pchunk_len = 0;
 	u64 raw_size = 0, buffer_size = 0;
 	char *raw = NULL, *buffer = NULL;
+	struct z_erofs_mt_read_ctx *ctx = NULL;
 
 	erofs_dbg("verify data chunk of nid(%llu): type(%d)",
 		  inode->nid | 0ULL, inode->datalayout);
@@ -563,6 +566,16 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 		if (ret)
 			goto out;
 
+		if (!ctx && !digest && inode->i_size > EROFS_MT_THRESHOLD && compressed &&
+		    (map.m_algorithmformat == Z_EROFS_COMPRESSION_LZMA ||
+		     map.m_algorithmformat == Z_EROFS_COMPRESSION_DEFLATE)) {
+			ctx = z_erofs_mt_read_ctx_alloc(outfd, true);
+			if (!ctx) {
+				ret = -ENOMEM;
+				goto out;
+			}
+		}
+
 		if (!compressed && map.m_llen != map.m_plen) {
 			erofs_err("broken chunk length m_la %" PRIu64 " m_llen %" PRIu64 " m_plen %" PRIu64,
 				  map.m_la, map.m_llen, map.m_plen);
@@ -593,7 +606,7 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 						(const u8 *)zeros, chunk);
 					remain -= chunk;
 				}
-			} else if (outfd >= 0) {
+			} else if (!ctx && outfd >= 0) {
 				ret = lseek(outfd, map.m_llen, SEEK_CUR);
 				if (ret < 0) {
 					ret = -errno;
@@ -616,19 +629,33 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 			alloc_rawsize = map.m_plen;
 		}
 
-		if (alloc_rawsize > raw_size) {
-			char *newraw = realloc(raw, alloc_rawsize);
+		if (!ctx) {
+			if (alloc_rawsize > raw_size) {
+				char *newraw = realloc(raw, alloc_rawsize);
 
-			if (!newraw) {
-				ret = -ENOMEM;
-				goto out;
+				if (!newraw) {
+					ret = -ENOMEM;
+					goto out;
+				}
+				raw = newraw;
+				raw_size = alloc_rawsize;
 			}
-			raw = newraw;
-			raw_size = alloc_rawsize;
 		}
 
 		if (compressed) {
-			if (map.m_llen > buffer_size) {
+			char *c_raw = raw;
+			char *c_buf = buffer;
+
+			if (ctx) {
+				c_raw = malloc(alloc_rawsize);
+				c_buf = malloc(map.m_llen);
+				if (!c_raw || !c_buf) {
+					free(c_raw);
+					free(c_buf);
+					ret = -ENOMEM;
+					goto out;
+				}
+			} else if (map.m_llen > buffer_size) {
 				char *newbuffer;
 
 				buffer_size = map.m_llen;
@@ -638,39 +665,58 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 					goto out;
 				}
 				buffer = newbuffer;
+				c_buf = buffer;
 			}
-			ret = z_erofs_read_one_data(inode, &map, raw, buffer,
-						    0, map.m_llen, false);
-			if (ret)
+
+			ret = z_erofs_read_one_data(inode, &map, c_raw, c_buf,
+						    0, map.m_llen, false,
+						    map.m_la, ctx);
+			if (ret < 0)
 				goto out;
 
-			if (digest)
-				erofs_sha256_process(digest,
-					(const u8 *)buffer, map.m_llen);
-			if (outfd >= 0 && write(outfd, buffer, map.m_llen) < 0)
-				goto fail_eio;
+			if (!ctx) {
+				if (digest)
+					erofs_sha256_process(digest,
+						(const u8 *)c_buf, map.m_llen);
+				if (outfd >= 0 && write(outfd, c_buf, map.m_llen) < 0)
+					goto fail_eio;
+			}
 		} else {
 			u64 p = 0;
+			erofs_off_t m_llen = map.m_llen;
 
 			do {
-				u64 count = min_t(u64, alloc_rawsize,
-						  map.m_llen);
+				u64 count = min_t(u64, alloc_rawsize, m_llen);
 
 				ret = erofs_read_one_data(inode, &map, raw, p, count);
-				if (ret)
+				if (ret < 0)
 					goto out;
 
-				if (digest)
-					erofs_sha256_process(digest,
-						(const u8 *)raw, count);
-				if (outfd >= 0 && write(outfd, raw, count) < 0)
-					goto fail_eio;
-				map.m_llen -= count;
+				if (!ctx) {
+					if (digest)
+						erofs_sha256_process(digest,
+							(const u8 *)raw, count);
+					if (outfd >= 0 && write(outfd, raw, count) < 0)
+						goto fail_eio;
+				} else if (outfd >= 0) {
+					if (pwrite(outfd, raw, count, map.m_la + p) < 0)
+						goto fail_eio;
+				}
+				m_llen -= count;
 				p += count;
-			} while (map.m_llen);
+			} while (m_llen);
 		}
 	}
 
+	if (ctx) {
+		int wait_err;
+
+		z_erofs_mt_read_enqueue(ctx);
+		wait_err = z_erofs_mt_read_ctx_wait(ctx);
+		if (wait_err < 0 && ret == 0)
+			ret = wait_err;
+	}
+
 	if (fsckcfg.print_comp_ratio) {
 		u64 log_blocks = 0;
 		if (!erofs_is_packed_inode(inode))
@@ -687,6 +733,17 @@ static int erofs_verify_inode_data(struct erofs_inode *inode, int outfd,
 #endif
 	}
 out:
+	if (ctx && outfd >= 0 && ret == 0) {
+		if (ftruncate(outfd, inode->i_size) < 0) {
+			erofs_err("failed to truncate file to %llu: %d",
+				  (unsigned long long)inode->i_size, errno);
+			if (ret == 0)
+				ret = -errno;
+		}
+	}
+
+	if (ctx)
+		z_erofs_mt_read_ctx_free(ctx);
 	if (raw)
 		free(raw);
 	if (buffer)
@@ -1344,7 +1401,9 @@ int main(int argc, char *argv[])
 {
 	int err;
 
-	erofs_init_configure();
+	err = liberofs_global_init();
+	if (err)
+		return 1;
 
 	fsckcfg.physical_blocks = 0;
 	fsckcfg.logical_blocks = 0;
@@ -1516,7 +1575,7 @@ exit_dev_close:
 	erofs_dev_close(&g_sbi);
 exit:
 	erofs_blob_closeall(&g_sbi);
-	erofs_exit_configure();
+	liberofs_global_exit();
 	return err ? 1 : 0;
 }
 
diff --git a/include/erofs/internal.h b/include/erofs/internal.h
index 2cc9cc80..d663f57e 100644
--- a/include/erofs/internal.h
+++ b/include/erofs/internal.h
@@ -63,6 +63,8 @@ struct erofs_buf {
 #define BLK_ROUND_UP(sbi, addr)	\
 	(roundup(addr, erofs_blksiz(sbi)) >> (sbi)->blkszbits)
 
+#define Z_EROFS_PCLUSTER_MAX_BATCH_SIZE 32
+
 struct erofs_buffer_head;
 struct erofs_bufmgr;
 
@@ -451,6 +453,17 @@ struct z_erofs_paramset {
 	char *extraopts;
 };
 
+struct z_erofs_decompress_task;
+struct z_erofs_mt_read_ctx;
+
+struct z_erofs_mt_read_ctx *z_erofs_mt_read_ctx_alloc(int outfd, bool free_out);
+int z_erofs_mt_read_ctx_wait(struct z_erofs_mt_read_ctx *ctx);
+void z_erofs_mt_read_ctx_free(struct z_erofs_mt_read_ctx *ctx);
+void z_erofs_mt_read_enqueue(struct z_erofs_mt_read_ctx *ctx);
+
+int z_erofs_mt_workers_init(void);
+void z_erofs_mt_workers_exit(void);
+
 int liberofs_global_init(void);
 void liberofs_global_exit(void);
 
@@ -487,7 +500,8 @@ int erofs_read_one_data(struct erofs_inode *inode, struct erofs_map_blocks *map,
 			char *buffer, u64 offset, size_t len);
 int z_erofs_read_one_data(struct erofs_inode *inode,
 			struct erofs_map_blocks *map, char *raw, char *buffer,
-			erofs_off_t skip, erofs_off_t length, bool trimmed);
+			erofs_off_t skip, erofs_off_t length, bool trimmed,
+			erofs_off_t out_offset, struct z_erofs_mt_read_ctx *ctx);
 void *erofs_read_metadata(struct erofs_sb_info *sbi, erofs_nid_t nid,
 			  erofs_off_t *offset, int *lengthp);
 int z_erofs_parse_cfgs(struct erofs_sb_info *sbi, struct erofs_super_block *dsb);
diff --git a/lib/data.c b/lib/data.c
index 1bb9269c..3deaa137 100644
--- a/lib/data.c
+++ b/lib/data.c
@@ -9,6 +9,262 @@
 #include "erofs/trace.h"
 #include "erofs/decompress.h"
 #include "liberofs_fragments.h"
+#include "erofs/lock.h"
+#include "erofs/cond.h"
+
+#ifdef EROFS_MT_ENABLED
+#include "erofs/config.h"
+
+struct z_erofs_percpu_queue {
+	erofs_mutex_t lock;
+	erofs_cond_t cond;
+	struct z_erofs_decompress_task *head;
+	struct z_erofs_decompress_task *tail;
+	int pending_tasks;
+	bool shutdown;
+	pthread_t thread;
+} __attribute__((aligned(64)));
+
+static struct z_erofs_percpu_queue *worker_queues;
+static int num_workers;
+static pthread_once_t worker_pool_once = PTHREAD_ONCE_INIT;
+static int worker_pool_init_err;
+static int z_erofs_mt_workers_ensure_init(void);
+#endif
+
+struct z_erofs_mt_read_ctx {
+	erofs_mutex_t lock;
+	erofs_cond_t cond;
+	int pending_tasks;
+	int final_err;
+	int outfd;
+	bool free_out;
+	struct z_erofs_decompress_task *current_task;
+};
+
+struct z_erofs_mt_read_ctx *z_erofs_mt_read_ctx_alloc(int outfd, bool free_out)
+{
+	struct z_erofs_mt_read_ctx *ctx;
+
+#ifdef EROFS_MT_ENABLED
+	z_erofs_mt_workers_ensure_init();
+#endif
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx)
+		return NULL;
+
+	erofs_mutex_init(&ctx->lock);
+	erofs_cond_init(&ctx->cond);
+	ctx->outfd = outfd;
+	ctx->free_out = free_out;
+	return ctx;
+}
+
+int z_erofs_mt_read_ctx_wait(struct z_erofs_mt_read_ctx *ctx)
+{
+	int err;
+
+	if (!ctx)
+		return 0;
+
+	erofs_mutex_lock(&ctx->lock);
+	while (ctx->pending_tasks > 0)
+		erofs_cond_wait(&ctx->cond, &ctx->lock);
+	err = ctx->final_err;
+	erofs_mutex_unlock(&ctx->lock);
+
+	return err;
+}
+
+void z_erofs_mt_read_ctx_free(struct z_erofs_mt_read_ctx *ctx)
+{
+	if (!ctx)
+		return;
+
+	erofs_mutex_destroy(&ctx->lock);
+	erofs_cond_destroy(&ctx->cond);
+	free(ctx);
+}
+
+struct z_erofs_decompress_item {
+	struct z_erofs_decompress_req req;
+	char *raw_buf;
+	char *out_buf;
+	erofs_off_t out_offset;
+	unsigned int out_length;
+};
+
+struct z_erofs_decompress_task {
+	struct z_erofs_decompress_task *next;
+	struct z_erofs_mt_read_ctx *ctx;
+	struct z_erofs_decompress_item items[Z_EROFS_PCLUSTER_MAX_BATCH_SIZE];
+	unsigned int nr_reqs;
+};
+
+static void z_erofs_process_task(struct z_erofs_decompress_task *task)
+{
+	struct z_erofs_mt_read_ctx *ctx = task->ctx;
+	int i, ret = 0, first_err = 0;
+
+	for (i = 0; i < task->nr_reqs; ++i) {
+		struct z_erofs_decompress_item *item = &task->items[i];
+
+		ret = z_erofs_decompress(&item->req);
+
+		if (ret >= 0 && ctx && ctx->outfd >= 0) {
+			if (pwrite(ctx->outfd, item->out_buf,
+				   item->out_length, item->out_offset) < 0)
+				ret = -errno;
+		}
+
+		if (ret < 0 && !first_err)
+			first_err = ret;
+
+		free(item->raw_buf);
+		if (ctx && ctx->free_out)
+			free(item->out_buf);
+	}
+
+	if (ctx) {
+		erofs_mutex_lock(&ctx->lock);
+		if (first_err < 0 && !ctx->final_err)
+			ctx->final_err = first_err;
+		ctx->pending_tasks--;
+		if (!ctx->pending_tasks)
+			erofs_cond_signal(&ctx->cond);
+		erofs_mutex_unlock(&ctx->lock);
+	}
+	free(task);
+}
+
+#ifdef EROFS_MT_ENABLED
+static void *z_erofs_worker_thread(void *arg)
+{
+	struct z_erofs_percpu_queue *q = arg;
+
+	while (1) {
+		erofs_mutex_lock(&q->lock);
+		while (!q->head && !q->shutdown)
+			erofs_cond_wait(&q->cond, &q->lock);
+
+		if (q->shutdown && !q->head) {
+			erofs_mutex_unlock(&q->lock);
+			break;
+		}
+
+		struct z_erofs_decompress_task *tasks = q->head;
+
+		q->head = q->tail = NULL;
+		q->pending_tasks = 0;
+		erofs_mutex_unlock(&q->lock);
+
+		while (tasks) {
+			struct z_erofs_decompress_task *task = tasks;
+
+			tasks = tasks->next;
+			z_erofs_process_task(task);
+		}
+	}
+	return NULL;
+}
+
+static void z_erofs_mt_workers_create_once(void)
+{
+	int i;
+
+	num_workers = erofs_get_available_processors();
+
+	if (num_workers < 1)
+		num_workers = 1;
+
+	worker_queues = calloc(num_workers, sizeof(*worker_queues));
+	if (!worker_queues) {
+		worker_pool_init_err = -ENOMEM;
+		num_workers = 0;
+		return;
+	}
+
+	for (i = 0; i < num_workers; ++i) {
+		erofs_mutex_init(&worker_queues[i].lock);
+		erofs_cond_init(&worker_queues[i].cond);
+		worker_queues[i].head = worker_queues[i].tail = NULL;
+		worker_queues[i].pending_tasks = 0;
+		worker_queues[i].shutdown = false;
+		pthread_create(&worker_queues[i].thread, NULL, z_erofs_worker_thread,
+			&worker_queues[i]);
+	}
+}
+
+static int z_erofs_mt_workers_ensure_init(void)
+{
+	pthread_once(&worker_pool_once, z_erofs_mt_workers_create_once);
+	return worker_pool_init_err;
+}
+
+int z_erofs_mt_workers_init(void)
+{
+	return 0;
+}
+
+void z_erofs_mt_workers_exit(void)
+{
+	int i;
+
+	if (!worker_queues)
+		return;
+	for (i = 0; i < num_workers; ++i) {
+		erofs_mutex_lock(&worker_queues[i].lock);
+		worker_queues[i].shutdown = true;
+		erofs_cond_signal(&worker_queues[i].cond);
+		erofs_mutex_unlock(&worker_queues[i].lock);
+
+		pthread_join(worker_queues[i].thread, NULL);
+		erofs_cond_destroy(&worker_queues[i].cond);
+		erofs_mutex_destroy(&worker_queues[i].lock);
+	}
+	free(worker_queues);
+	worker_queues = NULL;
+}
+#endif
+
+void z_erofs_mt_read_enqueue(struct z_erofs_mt_read_ctx *ctx)
+{
+#ifdef EROFS_MT_ENABLED
+	static int next_worker;
+#endif
+
+	if (!ctx || !ctx->current_task)
+		return;
+
+	struct z_erofs_decompress_task *task = ctx->current_task;
+
+	ctx->current_task = NULL;
+
+#ifdef EROFS_MT_ENABLED
+	if (num_workers > 0) {
+		int target = next_worker;
+
+		next_worker = (next_worker + 1) % num_workers;
+		struct z_erofs_percpu_queue *q = &worker_queues[target];
+
+		erofs_mutex_lock(&q->lock);
+		task->next = NULL;
+		if (!q->tail) {
+			q->head = q->tail = task;
+		} else {
+			q->tail->next = task;
+			q->tail = task;
+		}
+		q->pending_tasks++;
+		erofs_cond_signal(&q->cond);
+		erofs_mutex_unlock(&q->lock);
+		return;
+	}
+#endif
+	task->next = NULL;
+	z_erofs_process_task(task);
+}
 
 void *erofs_bread(struct erofs_buf *buf, erofs_off_t offset, bool need_kmap)
 {
@@ -277,20 +533,29 @@ static int erofs_read_raw_data(struct erofs_inode *inode, char *buffer,
 
 int z_erofs_read_one_data(struct erofs_inode *inode,
 			struct erofs_map_blocks *map, char *raw, char *buffer,
-			erofs_off_t skip, erofs_off_t length, bool trimmed)
+			erofs_off_t skip, erofs_off_t length, bool trimmed,
+			erofs_off_t out_offset, struct z_erofs_mt_read_ctx *ctx)
 {
 	struct erofs_sb_info *sbi = inode->sbi;
 	struct erofs_map_dev mdev;
-	int ret = 0;
+	struct z_erofs_decompress_task *task;
+	struct z_erofs_decompress_item *item;
+	int ret = 0, idx, batch_limit;
 
 	if (map->m_flags & __EROFS_MAP_FRAGMENT) {
 		if (__erofs_unlikely(inode->nid == sbi->packed_nid)) {
 			erofs_err("fragment should not exist in the packed inode %llu",
 				  sbi->packed_nid | 0ULL);
-			return -EFSCORRUPTED;
+			ret = -EFSCORRUPTED;
+			goto err_out;
 		}
-		return erofs_packedfile_read(sbi, buffer, length - skip,
-				   inode->fragmentoff + skip);
+		ret = erofs_packedfile_read(sbi, buffer, length - skip,
+					inode->fragmentoff + skip);
+		if (ret >= 0 && ctx && ctx->outfd >= 0) {
+			if (pwrite(ctx->outfd, buffer, length - skip, out_offset) < 0)
+				ret = -errno;
+		}
+		goto err_out;
 	}
 
 	/* no device id here, thus it will always succeed */
@@ -300,31 +565,95 @@ int z_erofs_read_one_data(struct erofs_inode *inode,
 	ret = erofs_map_dev(sbi, &mdev);
 	if (ret) {
 		DBG_BUGON(1);
-		return ret;
+		goto err_out;
 	}
 
 	ret = erofs_dev_read(sbi, mdev.m_deviceid, raw, mdev.m_pa, map->m_plen);
 	if (ret < 0)
+		goto err_out;
+
+	if (!ctx) {
+		ret = z_erofs_decompress(&(struct z_erofs_decompress_req) {
+				.sbi = sbi,
+				.in = raw,
+				.out = buffer,
+				.decodedskip = skip,
+				.interlaced_offset =
+					map->m_algorithmformat == Z_EROFS_COMPRESSION_INTERLACED ?
+						erofs_blkoff(sbi, map->m_la) : 0,
+				.inputsize = map->m_plen,
+				.decodedlength = length,
+				.alg = map->m_algorithmformat,
+				.partial_decoding = trimmed ? true :
+					!(map->m_flags & EROFS_MAP_FULL_MAPPED) ||
+						(map->m_flags & EROFS_MAP_PARTIAL_REF),
+				});
 		return ret;
+	}
 
-	ret = z_erofs_decompress(&(struct z_erofs_decompress_req) {
-			.sbi = sbi,
-			.in = raw,
-			.out = buffer,
-			.decodedskip = skip,
-			.interlaced_offset =
-				map->m_algorithmformat == Z_EROFS_COMPRESSION_INTERLACED ?
-					erofs_blkoff(sbi, map->m_la) : 0,
-			.inputsize = map->m_plen,
-			.decodedlength = length,
-			.alg = map->m_algorithmformat,
-			.partial_decoding = trimmed ? true :
-				!(map->m_flags & EROFS_MAP_FULL_MAPPED) ||
-					(map->m_flags & EROFS_MAP_PARTIAL_REF),
-			 });
-	if (ret < 0)
-		return ret;
+	task = ctx->current_task;
+	if (!task) {
+		task = malloc(sizeof(*task));
+		if (!task) {
+			ret = -ENOMEM;
+			goto err_out;
+		}
+		task->nr_reqs = 0;
+		task->ctx = ctx;
+		ctx->current_task = task;
+
+		erofs_mutex_lock(&ctx->lock);
+		ctx->pending_tasks++;
+		erofs_mutex_unlock(&ctx->lock);
+	}
+
+	idx = task->nr_reqs++;
+	item = &task->items[idx];
+
+	item->req = (struct z_erofs_decompress_req) {
+		.sbi = sbi,
+		.in = raw,
+		.out = buffer,
+		.decodedskip = skip,
+		.interlaced_offset =
+			map->m_algorithmformat == Z_EROFS_COMPRESSION_INTERLACED ?
+				erofs_blkoff(sbi, map->m_la) : 0,
+		.inputsize = map->m_plen,
+		.decodedlength = length,
+		.alg = map->m_algorithmformat,
+		.partial_decoding = trimmed ? true :
+			!(map->m_flags & EROFS_MAP_FULL_MAPPED) ||
+				(map->m_flags & EROFS_MAP_PARTIAL_REF),
+	};
+	item->raw_buf = raw;
+	item->out_buf = buffer;
+	item->out_offset = out_offset;
+	item->out_length = length;
+
+	batch_limit = (map->m_algorithmformat == Z_EROFS_COMPRESSION_LZ4 ||
+						map->m_algorithmformat == Z_EROFS_COMPRESSION_ZSTD)
+					? Z_EROFS_PCLUSTER_MAX_BATCH_SIZE : 8;
+
+	if (task->nr_reqs >= batch_limit) {
+		/* Execute fast algorithms synchronously on the traversal thread */
+		if (map->m_algorithmformat == Z_EROFS_COMPRESSION_LZ4 ||
+			map->m_algorithmformat == Z_EROFS_COMPRESSION_ZSTD) {
+			ctx->current_task = NULL;
+			task->next = NULL;
+			z_erofs_process_task(task);
+		} else {
+			z_erofs_mt_read_enqueue(ctx);
+		}
+	}
 	return 0;
+
+err_out:
+	if (ctx) {
+		if (ctx->free_out)
+			free(buffer);
+		free(raw);
+	}
+	return ret;
 }
 
 static int z_erofs_read_data(struct erofs_inode *inode, char *buffer,
@@ -387,7 +716,7 @@ static int z_erofs_read_data(struct erofs_inode *inode, char *buffer,
 		}
 
 		ret = z_erofs_read_one_data(inode, &map, raw,
-				buffer + end - offset, skip, length, trimmed);
+				buffer + end - offset, skip, length, trimmed, 0, NULL);
 		if (ret < 0)
 			break;
 	}
diff --git a/lib/global.c b/lib/global.c
index 938aa0a7..39451a35 100644
--- a/lib/global.c
+++ b/lib/global.c
@@ -27,6 +27,13 @@ int liberofs_global_init(void)
 #ifdef S3EROFS_ENABLED
 	xmlInitParser();
 #endif
+
+#ifdef EROFS_MT_ENABLED
+	err = z_erofs_mt_workers_init();
+	if (err)
+		goto out_unlock;
+#endif
+
 #ifdef HAVE_LIBCURL
 	if (!erofs_global_curl_initialized) {
 		if (curl_global_init(CURL_GLOBAL_DEFAULT) != CURLE_OK) {
@@ -35,14 +42,17 @@ int liberofs_global_init(void)
 		}
 		erofs_global_curl_initialized = true;
 	}
-out_unlock:
 #endif
+out_unlock:
 	erofs_mutex_unlock(&erofs_global_mutex);
 	return err;
 }
 
 void liberofs_global_exit(void)
 {
+#ifdef EROFS_MT_ENABLED
+	z_erofs_mt_workers_exit();
+#endif
 	erofs_mutex_lock(&erofs_global_mutex);
 	z_erofs_mt_global_exit();
 #ifdef HAVE_LIBCURL
-- 
2.52.0