On Thu 13-02-20 12:23:36, Mike Christie wrote:
> This patch documents the PR_SET_IO_FLUSHER and PR_GET_IO_FLUSHER
> prctl commands added to the linux kernel for 5.6 in commit:
>
> commit 8d19f1c8e1937baf74e1962aae9f90fa3aeab463
> Author: Mike Christie <mchristi@xxxxxxxxxx>
> Date: Mon Nov 11 18:19:00 2019 -0600
>
> prctl: PR_{G,S}ET_IO_FLUSHER to support controlling memory reclaim
>
> Signed-off-by: Mike Christie <mchristi@xxxxxxxxxx>
> Reviewed-by: Damien Le Moal <damien.lemoal@xxxxxxx>
> ---
>
> V3:
> - Replace emulation device example.
>
> V2:
> - My initial patch for this was very bad. This version is almost 100%
> taken word for word from Dave Chinner's review comments.
>
> Signed-off-by: Mike Christie <mchristi@xxxxxxxxxx>
> ---
> man2/prctl.2 | 25 +++++++++++++++++++++++++
> 1 file changed, 25 insertions(+)
>
> diff --git a/man2/prctl.2 b/man2/prctl.2
> index 720ec04e4..58d77bf2e 100644
> --- a/man2/prctl.2
> +++ b/man2/prctl.2
> @@ -1381,6 +1381,30 @@ system call on Tru64).
> for information on versions and architectures.)
> Return unaligned access control bits, in the location pointed to by
> .IR "(unsigned int\ *) arg2" .
> +.TP
> +.B PR_SET_IO_FLUSHER (Since Linux 5.6)
> +An IO_FLUSHER is a user process that the kernel uses to issue IO
> +that cleans dirty page cache data and/or filesystem metadata. The
> +kernel may need to clean this memory when under memory pressure in
> +order to free it. This means there is potential for a memory reclaim
> +recursion deadlock if the user process attempts to allocate memory
> +and the kernel then blocks waiting for it to clean memory before it
> +can make reclaim progress.
> +
> +The kernel avoids these recursion problems internally via a special
> +process state that prevents recursive reclaim from issuing new IO.
I would refrain from describing the internal implementation. The
important part is that this flag tells the kernel that IO_FLUSHER
process gets a special treatment to workaround the deadlock.
So anytime a process is involved in the IO path and the kernel cannot
make a forward progress without it then the flag should be set.
--
Michal Hocko
SUSE Labs
