Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 17 Dec 2020 16:50:48 -0400
From:   Jason Gunthorpe <jgg@ziepe.ca>
To:     Pavel Tatashin <pasha.tatashin@soleen.com>
Cc:     linux-kernel@vger.kernel.org, linux-mm@kvack.org,
        akpm@linux-foundation.org, vbabka@suse.cz, mhocko@suse.com,
        david@redhat.com, osalvador@suse.de, dan.j.williams@intel.com,
        sashal@kernel.org, tyhicks@linux.microsoft.com,
        iamjoonsoo.kim@lge.com, mike.kravetz@oracle.com,
        rostedt@goodmis.org, mingo@redhat.com, peterz@infradead.org,
        mgorman@suse.de, willy@infradead.org, rientjes@google.com,
        jhubbard@nvidia.com, linux-doc@vger.kernel.org,
        ira.weiny@intel.com, linux-kselftest@vger.kernel.org
Subject: Re: [PATCH v4 08/10] mm/gup: limit number of gup migration failures,
 honor failures
Message-ID: <20201217205048.GL5487@ziepe.ca>
References: <20201217185243.3288048-1-pasha.tatashin@soleen.com>
 <20201217185243.3288048-9-pasha.tatashin@soleen.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20201217185243.3288048-9-pasha.tatashin@soleen.com>
Precedence: bulk

On Thu, Dec 17, 2020 at 01:52:41PM -0500, Pavel Tatashin wrote:
> +/*
> + * Verify that there are no unpinnable (movable) pages, if so return true.
> + * Otherwise an unpinnable pages is found return false, and unpin all pages.
> + */
> +static bool check_and_unpin_pages(unsigned long nr_pages, struct page **pages,
> +				  unsigned int gup_flags)
> +{
> +	unsigned long i, step;
> +
> +	for (i = 0; i < nr_pages; i += step) {
> +		struct page *head = compound_head(pages[i]);
> +
> +		step = compound_nr(head) - (pages[i] - head);

You can't assume that all of a compound head is in the pages array,
this assumption would only work inside the page walkers if the page
was found in a PMD or something.

> +	if (gup_flags & FOLL_PIN) {
> +		unpin_user_pages(pages, nr_pages);

So we throw everything away? Why? That isn't how the old algorithm worked

> @@ -1654,22 +1664,55 @@ static long __gup_longterm_locked(struct mm_struct *mm,
>  				  struct vm_area_struct **vmas,
>  				  unsigned int gup_flags)
>  {
> -	unsigned long flags = 0;
> +	int migrate_retry = 0;
> +	int isolate_retry = 0;
> +	unsigned int flags;
>  	long rc;
>  
> -	if (gup_flags & FOLL_LONGTERM)
> -		flags = memalloc_pin_save();
> +	if (!(gup_flags & FOLL_LONGTERM))
> +		return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
> +					       NULL, gup_flags);
>  
> -	rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas, NULL,
> -				     gup_flags);
> +	/*
> +	 * Without FOLL_WRITE fault handler may return zero page, which can
> +	 * be in a movable zone, and also will fail to isolate during migration,
> +	 * thus the longterm pin will fail.
> +	 */
> +	gup_flags &= FOLL_WRITE;

Is &= what you mean here? |= right?

Seems like we've ended up in a weird place if FOLL_LONGTERM always
includes FOLL_WRITE. Putting the zero page in ZONE_MOVABLE seems like
a bad idea, no?

> +	/*
> +	 * Migration may fail, we retry before giving up. Also, because after
> +	 * migration pages[] becomes outdated, we unpin and repin all pages
> +	 * in the range, so pages array is repopulated with new values.
> +	 * Also, because of this we cannot retry migration failures in a loop
> +	 * without pinning/unpinnig pages.
> +	 */

The old algorithm made continuous forward progress and only went back
to the first migration point.

> +	for (; ; ) {

while (true)?

> +		rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
> +					     NULL, gup_flags);

> +		/* Return if error or if all pages are pinnable */
> +		if (rc <= 0 || check_and_unpin_pages(rc, pages, gup_flags))
> +			break;

So we sweep the pages list twice now?

> +		/* Some pages are not pinnable, migrate them */
> +		rc = migrate_movable_pages(rc, pages);
> +
> +		/*
> +		 * If there is an error, and we tried maximum number of times
> +		 * bail out. Notice: we return an error code, and all pages are
> +		 * unpinned
> +		 */
> +		if (rc < 0 && migrate_retry++ >= PINNABLE_MIGRATE_MAX) {
> +			break;
> +		} else if (rc > 0 && isolate_retry++ >= PINNABLE_ISOLATE_MAX) {
> +			rc = -EBUSY;

I don't like this at all. It shouldn't be so flakey

Can you do migration without the LRU?

Jason