From: Jiucheng Xu <jiucheng.xu@amlogic.com>

The bio completion path in the process context (e.g. dm-verity)
will directly call into decompression rather than trigger another
workqueue context for minimal scheduling latencies, which can
then call vm_map_ram() with GFP_KERNEL.

Due to insufficient memory, vm_map_ram() may generate memory
swapping I/O, which can cause submit_bio_wait to deadlock
in some scenarios.

Trimmed down the call stack, as follows:

f2fs_submit_read_io
  submit_bio                      //bio_list is initialized.
    mmc_blk_mq_recovery
      z_erofs_endio
        vm_map_ram
          __pte_alloc_kernel
            __alloc_pages_direct_reclaim
              shrink_folio_list
                __swap_writepage
                  submit_bio_wait  //bio_list is non-NULL, hang!!!

Use memalloc_noio_{save,restore}() to wrap up this path.

Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Signed-off-by: Jiucheng Xu <jiucheng.xu@amlogic.com>
---
Changes in v2:
- Reword the commit message.
- Link to v1: https://lore.kernel.org/r/20260311-origin-dev-v1-1-40524ef07ff0@amlogic.com
---
 fs/erofs/zdata.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
index 3977e42b9516861bf3d59c072b6b8aaa6898dd8a..fe8121df9ef2f2404fc6e3f0fbbd6367f9ec2c67 100644
--- a/fs/erofs/zdata.c
+++ b/fs/erofs/zdata.c
@@ -1445,6 +1445,7 @@ static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 				       int bios)
 {
 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
+	int gfp_flag;
 
 	/* wake up the caller thread for sync decompression */
 	if (io->sync) {
@@ -1477,7 +1478,9 @@ static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 			sbi->sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
 		return;
 	}
+	gfp_flag = memalloc_noio_save();
 	z_erofs_decompressqueue_work(&io->u.work);
+	memalloc_noio_restore(gfp_flag);
 }
 
 static void z_erofs_fill_bio_vec(struct bio_vec *bvec,

---
base-commit: 11439c4635edd669ae435eec308f4ab8a0804808
change-id: 20260311-origin-dev-1c9665798204

Best regards,
-- 
Jiucheng Xu <jiucheng.xu@amlogic.com>

On 2026/3/11 17:11, Jiucheng Xu via B4 Relay wrote:
> From: Jiucheng Xu <jiucheng.xu@amlogic.com>
> 
> The bio completion path in the process context (e.g. dm-verity)
> will directly call into decompression rather than trigger another
> workqueue context for minimal scheduling latencies, which can
> then call vm_map_ram() with GFP_KERNEL.
> 
> Due to insufficient memory, vm_map_ram() may generate memory
> swapping I/O, which can cause submit_bio_wait to deadlock
> in some scenarios.
> 
> Trimmed down the call stack, as follows:
> 
> f2fs_submit_read_io
>    submit_bio                      //bio_list is initialized.
>      mmc_blk_mq_recovery
>        z_erofs_endio
>          vm_map_ram
>            __pte_alloc_kernel
>              __alloc_pages_direct_reclaim
>                shrink_folio_list
>                  __swap_writepage
>                    submit_bio_wait  //bio_list is non-NULL, hang!!!
> 
> Use memalloc_noio_{save,restore}() to wrap up this path.
> 
> Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
> Signed-off-by: Jiucheng Xu <jiucheng.xu@amlogic.com>

Reviewed-by: Chao Yu <chao@kernel.org>

Thanks,