Re: [PATCH RFC 2/4] md/raid5-ppl: replace async_tx with synchronous xor

[Alex Naylor <alex.r.naylor@gmail.com>]

Hi Frank,

I encountered a PPL recovery failure that appears relevant to this 
proposal, on a QNAP TS-410 (Marvell Kirkwood, ARMv5).

The system runs a custom Linux 6.1-based kernel with CONFIG_MV_XOR=y and 
CONFIG_ASYNC_TX_DMA=y, plus a local PCIe startup-reset patch. Storage is 
a four-member RAID5 with metadata 1.2, 512 KiB chunks and PPL enabled.

After a power outage, array startup hit:

kernel BUG at crypto/async_tx/async_tx.c:269!
PC: async_tx_quiesce
LR: ppl_xor [raid456]

The call trace included:

async_tx_quiesce
ppl_xor
ppl_recover_entry
ppl_init_log
raid5_run
md_run
do_md_run

All four members had matching event counts and valid superblock 
checksums, with no members marked failed.

The apparent issue is that ppl_xor() submits with ASYNC_TX_ACK, then 
passes the returned descriptor to async_tx_quiesce(), which asserts that 
it is not already acknowledged.

Disabling both Marvell XOR engines through their device-tree status 
properties, while retaining the same kernel binary and modules, allowed 
the existing array to assemble and the filesystem to recover. No forced 
assembly, array recreation or PPL-policy change was used. File transfers 
are working again.

I have not tested this RFC or reproduced the crash on a current kernel. 
Source inspection found the same ACK/quiesce sequence in v6.1.190 and 
current mainline.

I’m sharing this because the ppl_xor() change below removes the exact 
ACK/quiesce sequence implicated in my crash. I haven’t tested your 
patch, but disabling hardware XOR allowed recovery to complete on the 
same kernel.

Thanks,

Alex

On 2026-10-15 12:22, Frank.Li@oss.nxp.com wrote:
> From: Frank Li <Frank.Li@nxp.com>
>
> Remove the async_tx dependency from ops_run_partial_parity() and
> ppl_xor().  The function now returns void and operates synchronously:
>
> - count==1: copy_page() replaces async_memcpy
> - count>1:  memcpy + xor_blocks() replaces async_xor
> - ppl_xor(): xor_blocks() replaces async_xor + async_tx_quiesce
>
> Drop #include <linux/async_tx.h> and add #include <linux/xor.h>.
>
> Signed-off-by: Frank Li <Frank.Li@nxp.com>
> ---
>   drivers/md/raid5-ppl.c | 40 ++++++++++++++++------------------------
>   1 file changed, 16 insertions(+), 24 deletions(-)
>
> diff --git a/drivers/md/raid5-ppl.c b/drivers/md/raid5-ppl.c
> index 7f8a9d3fd578d..bd844b4e8b372 100644
> --- a/drivers/md/raid5-ppl.c
> +++ b/drivers/md/raid5-ppl.c
> @@ -8,7 +8,7 @@
>   #include <linux/blkdev.h>
>   #include <linux/slab.h>
>   #include <linux/crc32c.h>
> -#include <linux/async_tx.h>
> +#include <linux/raid/xor.h>
>   #include <linux/raid/md_p.h>
>   #include "md.h"
>   #include "raid5.h"
> @@ -154,14 +154,11 @@ struct ppl_io_unit {
>   	struct bio_vec biovec[PPL_IO_INLINE_BVECS];
>   };
>   
> -struct dma_async_tx_descriptor *
> -ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu,
> -		       struct dma_async_tx_descriptor *tx)
> +void ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu)
>   {
>   	int disks = sh->disks;
>   	struct page **srcs = percpu->scribble;
>   	int count = 0, pd_idx = sh->pd_idx, i;
> -	struct async_submit_ctl submit;
>   
>   	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
>   
> @@ -186,20 +183,20 @@ ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu,
>   				srcs[count++] = dev->page;
>   		}
>   	} else {
> -		return tx;
> +		return;
>   	}
>   
> -	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, tx,
> -			  NULL, sh, (void *) (srcs + sh->disks + 2));
> -
> -	if (count == 1)
> -		tx = async_memcpy(sh->ppl_page, srcs[0], 0, 0, PAGE_SIZE,
> -				  &submit);
> -	else
> -		tx = async_xor(sh->ppl_page, srcs, 0, count, PAGE_SIZE,
> -			       &submit);
> +	if (count == 1) {
> +		copy_page(page_address(sh->ppl_page), page_address(srcs[0]));
> +	} else {
> +		void *dest = page_address(sh->ppl_page);
> +		void **ptrs = (void **)srcs;
> +		int j;
>   
> -	return tx;
> +		for (j = 0; j < count; j++)
> +			ptrs[j] = page_address(srcs[j]);
> +		xor_gen(dest, ptrs, count, PAGE_SIZE);
> +	}
>   }
>   
>   static void *ppl_io_pool_alloc(gfp_t gfp_mask, void *pool_data)
> @@ -706,15 +703,10 @@ void ppl_stripe_write_finished(struct stripe_head *sh)
>   
>   static void ppl_xor(int size, struct page *page1, struct page *page2)
>   {
> -	struct async_submit_ctl submit;
> -	struct dma_async_tx_descriptor *tx;
> -	struct page *xor_srcs[] = { page1, page2 };
> -
> -	init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
> -			  NULL, NULL, NULL, NULL);
> -	tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
> +	void *dest = page_address(page1);
> +	void *ptrs[2] = { dest, page_address(page2) };
>   
> -	async_tx_quiesce(&tx);
> +	xor_gen(dest, ptrs, 2, size);
>   }
>   
>   /*
>