Hi Lucas,

On 2026/3/2 23:22, Lucas Karpinski wrote:
On 2026-03-02 8:03 a.m., lishixian wrote:
When rebuilding from source EROFS images, erofs_read_xattrs_from_disk()
is called for inodes that have xattr. At that point inode->sbi points to
the source image's sbi, which is opened read-only and never gets
erofs_xattr_init(), so sbi->xamgr is NULL. get_xattritem(sbi) then
dereferences xamgr and crashes with SIGSEGV.

Fix by using the build target's xamgr when initializing src's sbi.

Reported-by: Yixiao Chen <[email protected]>
Fixes: https://github.com/erofs/erofs-utils/issues/42
Signed-off-by: lishixian <[email protected]>
Reviewed-by: Yifan Zhao <[email protected]>
---
  lib/rebuild.c | 1 +
  mkfs/main.c   | 1 +
  2 files changed, 2 insertions(+)

diff --git a/lib/rebuild.c b/lib/rebuild.c
index f89a17c..f1e79c1 100644
--- a/lib/rebuild.c
+++ b/lib/rebuild.c
@@ -437,6 +437,7 @@ int erofs_rebuild_load_tree(struct erofs_inode *root, 
struct erofs_sb_info *sbi,
                erofs_err("failed to read superblock of %s", fsid);
                return ret;
        }
+       sbi->xamgr = g_sbi.xamgr;
inode.nid = sbi->root_nid;
        inode.sbi = sbi;
diff --git a/mkfs/main.c b/mkfs/main.c
index b84d1b4..cb0f0cc 100644
--- a/mkfs/main.c
+++ b/mkfs/main.c
@@ -1011,6 +1011,7 @@ static void erofs_rebuild_cleanup(void)
list_for_each_entry_safe(src, n, &rebuild_src_list, list) {
                list_del(&src->list);
+               src->xamgr = NULL; /* borrowed from g_sbi, do not free */
                erofs_put_super(src);
                erofs_dev_close(src);
                free(src);

I was similarly looking at this issue in my patchset so I can confirm it
fixes the seg fault.

Tested-by: Lucas Karpinski <[email protected]>

Thanks for this, but as I said to lishixian we shouldn't use
global g_sbi in the liberofs anymore.

Could we try to assign sbi->xamgr in the caller instead?

And

in my patchset

Do you have more urgent fixes? I'm about to release
erofs-utils 1.9.1 since there are some urgent fixes
so fixes would be better to be sent out now.

Also I think we should have a basic testcase to cover
this, I will try to add one this week.

Thanks,
Gao Xiang


Reply via email to