Re: [PATCH 1/4] mm/gup: Add FOLL_INTERRUPTIBLE

[John Hubbard <jhubbard@nvidia.com>]

On 6/22/22 14:36, Peter Xu wrote:
> We have had FAULT_FLAG_INTERRUPTIBLE but it was never applied to GUPs.  One
> issue with it is that not all GUP paths are able to handle signal delivers
> besides SIGKILL.
> 
> That's not ideal for the GUP users who are actually able to handle these
> cases, like KVM.
> 
> KVM uses GUP extensively on faulting guest pages, during which we've got
> existing infrastructures to retry a page fault at a later time.  Allowing
> the GUP to be interrupted by generic signals can make KVM related threads
> to be more responsive.  For examples:
> 
>    (1) SIGUSR1: which QEMU/KVM uses to deliver an inter-process IPI,
>        e.g. when the admin issues a vm_stop QMP command, SIGUSR1 can be
>        generated to kick the vcpus out of kernel context immediately,
> 
>    (2) SIGINT: which can be used with interactive hypervisor users to stop a
>        virtual machine with Ctrl-C without any delays/hangs,
> 
>    (3) SIGTRAP: which grants GDB capability even during page faults that are
>        stuck for a long time.
> 
> Normally hypervisor will be able to receive these signals properly, but not
> if we're stuck in a GUP for a long time for whatever reason.  It happens
> easily with a stucked postcopy migration when e.g. a network temp failure
> happens, then some vcpu threads can hang death waiting for the pages.  With
> the new FOLL_INTERRUPTIBLE, we can allow GUP users like KVM to selectively
> enable the ability to trap these signals.
> 
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
>   include/linux/mm.h |  1 +
>   mm/gup.c           | 33 +++++++++++++++++++++++++++++----
>   2 files changed, 30 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index bc8f326be0ce..ebdf8a6b86c1 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -2941,6 +2941,7 @@ struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
>   #define FOLL_SPLIT_PMD	0x20000	/* split huge pmd before returning */
>   #define FOLL_PIN	0x40000	/* pages must be released via unpin_user_page */
>   #define FOLL_FAST_ONLY	0x80000	/* gup_fast: prevent fall-back to slow gup */
> +#define FOLL_INTERRUPTIBLE  0x100000 /* allow interrupts from generic signals */

Perhaps, s/generic/non-fatal/ ?
>   
>   /*
>    * FOLL_PIN and FOLL_LONGTERM may be used in various combinations with each
> diff --git a/mm/gup.c b/mm/gup.c
> index 551264407624..ad74b137d363 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -933,8 +933,17 @@ static int faultin_page(struct vm_area_struct *vma,
>   		fault_flags |= FAULT_FLAG_WRITE;
>   	if (*flags & FOLL_REMOTE)
>   		fault_flags |= FAULT_FLAG_REMOTE;
> -	if (locked)
> +	if (locked) {
>   		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
> +		/*
> +		 * We should only grant FAULT_FLAG_INTERRUPTIBLE when we're
> +		 * (at least) killable.  It also mostly means we're not
> +		 * with NOWAIT.  Otherwise ignore FOLL_INTERRUPTIBLE since
> +		 * it won't make a lot of sense to be used alone.
> +		 */

This comment seems a little confusing due to its location. We've just
checked "locked", but the comment is talking about other constraints.

Not sure what to suggest. Maybe move it somewhere else?

> +		if (*flags & FOLL_INTERRUPTIBLE)
> +			fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
> +	}
>   	if (*flags & FOLL_NOWAIT)
>   		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
>   	if (*flags & FOLL_TRIED) {
> @@ -1322,6 +1331,22 @@ int fixup_user_fault(struct mm_struct *mm,
>   }
>   EXPORT_SYMBOL_GPL(fixup_user_fault);
>   
> +/*
> + * GUP always responds to fatal signals.  When FOLL_INTERRUPTIBLE is
> + * specified, it'll also respond to generic signals.  The caller of GUP
> + * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
> + */
> +static bool gup_signal_pending(unsigned int flags)
> +{
> +	if (fatal_signal_pending(current))
> +		return true;
> +
> +	if (!(flags & FOLL_INTERRUPTIBLE))
> +		return false;
> +
> +	return signal_pending(current);
> +}
> +

OK.

>   /*
>    * Please note that this function, unlike __get_user_pages will not
>    * return 0 for nr_pages > 0 without FOLL_NOWAIT
> @@ -1403,11 +1428,11 @@ static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
>   		 * Repeat on the address that fired VM_FAULT_RETRY
>   		 * with both FAULT_FLAG_ALLOW_RETRY and
>   		 * FAULT_FLAG_TRIED.  Note that GUP can be interrupted
> -		 * by fatal signals, so we need to check it before we
> +		 * by fatal signals of even common signals, depending on
> +		 * the caller's request. So we need to check it before we
>   		 * start trying again otherwise it can loop forever.
>   		 */
> -
> -		if (fatal_signal_pending(current)) {
> +		if (gup_signal_pending(flags)) {

This is new and bold. :) Signals that an application was prepared to
handle can now cause gup to quit early. I wonder if that will break any
use cases out there (SIGPIPE...) ?

Generally, gup callers handle failures pretty well, so it's probably
not too bad. But I wanted to mention the idea that handled interrupts
might be a little surprising here.

thanks,
-- 
John Hubbard
NVIDIA

>   			if (!pages_done)
>   				pages_done = -EINTR;
>   			break;