]> git.itanic.dy.fi Git - linux-stable/commitdiff
jffs2: GC deadlock reading a page that is used in jffs2_write_begin()
authorKyeong Yoo <kyeong.yoo@alliedtelesis.co.nz>
Tue, 4 Jul 2017 04:22:38 +0000 (16:22 +1200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 27 Jan 2022 07:47:38 +0000 (08:47 +0100)
[ Upstream commit aa39cc675799bc92da153af9a13d6f969c348e82 ]

GC task can deadlock in read_cache_page() because it may attempt
to release a page that is actually allocated by another task in
jffs2_write_begin().
The reason is that in jffs2_write_begin() there is a small window
a cache page is allocated for use but not set Uptodate yet.

This ends up with a deadlock between two tasks:
1) A task (e.g. file copy)
   - jffs2_write_begin() locks a cache page
   - jffs2_write_end() tries to lock "alloc_sem" from
 jffs2_reserve_space() <-- STUCK
2) GC task (jffs2_gcd_mtd3)
   - jffs2_garbage_collect_pass() locks "alloc_sem"
   - try to lock the same cache page in read_cache_page() <-- STUCK

So to avoid this deadlock, hold "alloc_sem" in jffs2_write_begin()
while reading data in a cache page.

Signed-off-by: Kyeong Yoo <kyeong.yoo@alliedtelesis.co.nz>
Signed-off-by: Richard Weinberger <richard@nod.at>
Signed-off-by: Sasha Levin <sashal@kernel.org>
fs/jffs2/file.c

index c12476e309c67c26883d5d2ae95166595359d5ab..eb4e4d784d26e6ef4c20d7407079e83d1c26c430 100644 (file)
@@ -135,20 +135,15 @@ static int jffs2_write_begin(struct file *filp, struct address_space *mapping,
        struct page *pg;
        struct inode *inode = mapping->host;
        struct jffs2_inode_info *f = JFFS2_INODE_INFO(inode);
+       struct jffs2_sb_info *c = JFFS2_SB_INFO(inode->i_sb);
        pgoff_t index = pos >> PAGE_SHIFT;
        uint32_t pageofs = index << PAGE_SHIFT;
        int ret = 0;
 
-       pg = grab_cache_page_write_begin(mapping, index, flags);
-       if (!pg)
-               return -ENOMEM;
-       *pagep = pg;
-
        jffs2_dbg(1, "%s()\n", __func__);
 
        if (pageofs > inode->i_size) {
                /* Make new hole frag from old EOF to new page */
-               struct jffs2_sb_info *c = JFFS2_SB_INFO(inode->i_sb);
                struct jffs2_raw_inode ri;
                struct jffs2_full_dnode *fn;
                uint32_t alloc_len;
@@ -159,7 +154,7 @@ static int jffs2_write_begin(struct file *filp, struct address_space *mapping,
                ret = jffs2_reserve_space(c, sizeof(ri), &alloc_len,
                                          ALLOC_NORMAL, JFFS2_SUMMARY_INODE_SIZE);
                if (ret)
-                       goto out_page;
+                       goto out_err;
 
                mutex_lock(&f->sem);
                memset(&ri, 0, sizeof(ri));
@@ -189,7 +184,7 @@ static int jffs2_write_begin(struct file *filp, struct address_space *mapping,
                        ret = PTR_ERR(fn);
                        jffs2_complete_reservation(c);
                        mutex_unlock(&f->sem);
-                       goto out_page;
+                       goto out_err;
                }
                ret = jffs2_add_full_dnode_to_inode(c, f, fn);
                if (f->metadata) {
@@ -204,13 +199,26 @@ static int jffs2_write_begin(struct file *filp, struct address_space *mapping,
                        jffs2_free_full_dnode(fn);
                        jffs2_complete_reservation(c);
                        mutex_unlock(&f->sem);
-                       goto out_page;
+                       goto out_err;
                }
                jffs2_complete_reservation(c);
                inode->i_size = pageofs;
                mutex_unlock(&f->sem);
        }
 
+       /*
+        * While getting a page and reading data in, lock c->alloc_sem until
+        * the page is Uptodate. Otherwise GC task may attempt to read the same
+        * page in read_cache_page(), which causes a deadlock.
+        */
+       mutex_lock(&c->alloc_sem);
+       pg = grab_cache_page_write_begin(mapping, index, flags);
+       if (!pg) {
+               ret = -ENOMEM;
+               goto release_sem;
+       }
+       *pagep = pg;
+
        /*
         * Read in the page if it wasn't already present. Cannot optimize away
         * the whole page write case until jffs2_write_end can handle the
@@ -220,15 +228,17 @@ static int jffs2_write_begin(struct file *filp, struct address_space *mapping,
                mutex_lock(&f->sem);
                ret = jffs2_do_readpage_nolock(inode, pg);
                mutex_unlock(&f->sem);
-               if (ret)
-                       goto out_page;
+               if (ret) {
+                       unlock_page(pg);
+                       put_page(pg);
+                       goto release_sem;
+               }
        }
        jffs2_dbg(1, "end write_begin(). pg->flags %lx\n", pg->flags);
-       return ret;
 
-out_page:
-       unlock_page(pg);
-       put_page(pg);
+release_sem:
+       mutex_unlock(&c->alloc_sem);
+out_err:
        return ret;
 }