erofs: avoid allocating DEFLATE streams before mounting

Currently, each DEFLATE stream takes one 32 KiB permanent internal
window buffer even if there is no running instance which uses DEFLATE
algorithm.

It's unexpected and wasteful on embedded devices with limited resources
and servers with hundreds of CPU cores if DEFLATE is enabled but unused.

Fixes: ffa09b3 ("erofs: DEFLATE compression support")
Cc: <stable@vger.kernel.org> # 6.6+
Reviewed-by: Sandeep Dhavale <dhavale@google.com>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240520090106.2898681-1-hsiangkao@linux.alibaba.com

Author: hsiangkao

fs/erofs/decompressor_deflate.c

@@ -46,39 +46,15 @@ int __init z_erofs_deflate_init(void)
 	/* by default, use # of possible CPUs instead */
 	if (!z_erofs_deflate_nstrms)
 		z_erofs_deflate_nstrms = num_possible_cpus();
-
-	for (; z_erofs_deflate_avail_strms < z_erofs_deflate_nstrms;
-	     ++z_erofs_deflate_avail_strms) {
-		struct z_erofs_deflate *strm;
-
-		strm = kzalloc(sizeof(*strm), GFP_KERNEL);
-		if (!strm)
-			goto out_failed;
-
-		/* XXX: in-kernel zlib cannot shrink windowbits currently */
-		strm->z.workspace = vmalloc(zlib_inflate_workspacesize());
-		if (!strm->z.workspace) {
-			kfree(strm);
-			goto out_failed;
-		}
-
-		spin_lock(&z_erofs_deflate_lock);
-		strm->next = z_erofs_deflate_head;
-		z_erofs_deflate_head = strm;
-		spin_unlock(&z_erofs_deflate_lock);
-	}
 	return 0;
-
-out_failed:
-	erofs_err(NULL, "failed to allocate zlib workspace");
-	z_erofs_deflate_exit();
-	return -ENOMEM;
 }
 
 int z_erofs_load_deflate_config(struct super_block *sb,
 			struct erofs_super_block *dsb, void *data, int size)
 {
 	struct z_erofs_deflate_cfgs *dfl = data;
+	static DEFINE_MUTEX(deflate_resize_mutex);
+	static bool inited;
 
 	if (!dfl || size < sizeof(struct z_erofs_deflate_cfgs)) {
 		erofs_err(sb, "invalid deflate cfgs, size=%u", size);
@@ -89,9 +65,36 @@ int z_erofs_load_deflate_config(struct super_block *sb,
 		erofs_err(sb, "unsupported windowbits %u", dfl->windowbits);
 		return -EOPNOTSUPP;
 	}
+	mutex_lock(&deflate_resize_mutex);
+	if (!inited) {
+		for (; z_erofs_deflate_avail_strms < z_erofs_deflate_nstrms;
+		     ++z_erofs_deflate_avail_strms) {
+			struct z_erofs_deflate *strm;
+
+			strm = kzalloc(sizeof(*strm), GFP_KERNEL);
+			if (!strm)
+				goto failed;
+			/* XXX: in-kernel zlib cannot customize windowbits */
+			strm->z.workspace = vmalloc(zlib_inflate_workspacesize());
+			if (!strm->z.workspace) {
+				kfree(strm);
+				goto failed;
+			}
 
+			spin_lock(&z_erofs_deflate_lock);
+			strm->next = z_erofs_deflate_head;
+			z_erofs_deflate_head = strm;
+			spin_unlock(&z_erofs_deflate_lock);
+		}
+		inited = true;
+	}
+	mutex_unlock(&deflate_resize_mutex);
 	erofs_info(sb, "EXPERIMENTAL DEFLATE feature in use. Use at your own risk!");
 	return 0;
+failed:
+	mutex_unlock(&deflate_resize_mutex);
+	z_erofs_deflate_exit();
+	return -ENOMEM;
 }
 
 int z_erofs_deflate_decompress(struct z_erofs_decompress_req *rq,