Re: [PATCH v2 11/22] coredump: always chunk writes

[R Nageswara Sastry <rnsastry@linux.ibm.com>]

On 20.08.2026 4:39 AM, Christian Brauner wrote:
> Right now dump_emit() is the only coredump helper that writes buffers
> larger than a page in one call. For elf notes that can easily blow past
> PAGE_SIZE. That's annoying because neither pipes nor af_unix sockets
> take such writes in one piece.
>
> If a signal arrives while the writer is waiting they drop a short write.
> With the coredump records work coming up that means header and its data
> are desynchronized. A write that fits in one pipe buffer or one skb
> doesn't suffer from this.
>
> So split all writes up, including elf notes, and cap every write at a
> page. The coredump socket already raises sk_sndbuf far enough for a page
> to fit a single skb and pipes always work that way.
>
> That means dump_interrupted() is now checked once per page. So a large
> coredump stops earlier (good). An empty write no longer issues a
> zero-length write. The rlimit core check stays where it was. It
> continues refusing whole writes.
>
> Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Tested-by: R Nageswara Sastry <rnsastry@linux.ibm.com>

System: ppc64le LPAR (IBM POWER), Linux 7.3-rc2
> ---
>   fs/coredump.c | 37 ++++++++++++++++++++++++++++++-------
>   1 file changed, 30 insertions(+), 7 deletions(-)
>
> diff --git a/fs/coredump.c b/fs/coredump.c
> index d837819031ff..d61f36239f91 100644
> --- a/fs/coredump.c
> +++ b/fs/coredump.c
> @@ -1215,19 +1215,21 @@ void vfs_coredump(const kernel_siginfo_t *siginfo)
>    * do on a core-file: use only these functions to write out all the
>    * necessary info.
>    */
> -static bool __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
> +/* One write, never more than a page. See __dump_emit(). */
> +static bool dump_emit_chunk(struct coredump_params *cprm, const void *addr,
> +			    int nr)
>   {
>   	struct file *file = cprm->file;
>   	loff_t pos = file->f_pos;
>   	ssize_t n;
>   
> -	if (cprm->written + nr > cprm->limit)
> -		return false;
>   	if (dump_interrupted())
>   		return false;
> +
>   	n = __kernel_write(file, addr, nr, &pos);
>   	if (n != nr)
>   		return false;
> +
>   	file->f_pos = pos;
>   	cprm->written += n;
>   	cprm->pos += n;
> @@ -1235,6 +1237,24 @@ static bool __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
>   	return true;
>   }
>   
> +static bool __dump_emit(struct coredump_params *cprm, const void *addr, int nr)
> +{
> +	if (cprm->written + nr > cprm->limit)
> +		return false;
> +
> +	while (nr) {
> +		int chunk = min_t(int, nr, PAGE_SIZE);
> +
> +		if (!dump_emit_chunk(cprm, addr, chunk))
> +			return false;
> +
> +		addr += chunk;
> +		nr -= chunk;
> +	}
> +
> +	return true;
> +}
> +
>   static bool __dump_skip(struct coredump_params *cprm, size_t nr)
>   {
>   	static char zeroes[PAGE_SIZE];
> @@ -1247,13 +1267,16 @@ static bool __dump_skip(struct coredump_params *cprm, size_t nr)
>   		return true;
>   	}
>   
> -	while (nr > PAGE_SIZE) {
> -		if (!__dump_emit(cprm, zeroes, PAGE_SIZE))
> +	while (nr) {
> +		size_t chunk = min_t(size_t, nr, PAGE_SIZE);
> +
> +		if (!__dump_emit(cprm, zeroes, chunk))
>   			return false;
> -		nr -= PAGE_SIZE;
> +
> +		nr -= chunk;
>   	}
>   
> -	return __dump_emit(cprm, zeroes, nr);
> +	return true;
>   }
>   
>   /* Flush the accumulated hole before writing data. */
>
-- 
Thanks and Regards
R.Nageswara Sastry