Date:   Thu, 26 Jan 2023 15:30:47 -0500
From:   Peter Xu <peterx@redhat.com>
To:     James Houghton <jthoughton@google.com>
Cc:     Mike Kravetz <mike.kravetz@oracle.com>,
        David Hildenbrand <david@redhat.com>,
        Muchun Song <songmuchun@bytedance.com>,
        David Rientjes <rientjes@google.com>,
        Axel Rasmussen <axelrasmussen@google.com>,
        Mina Almasry <almasrymina@google.com>,
        Zach O'Keefe <zokeefe@google.com>,
        Manish Mishra <manish.mishra@nutanix.com>,
        Naoya Horiguchi <naoya.horiguchi@nec.com>,
        "Dr . David Alan Gilbert" <dgilbert@redhat.com>,
        "Matthew Wilcox (Oracle)" <willy@infradead.org>,
        Vlastimil Babka <vbabka@suse.cz>,
        Baolin Wang <baolin.wang@linux.alibaba.com>,
        Miaohe Lin <linmiaohe@huawei.com>,
        Yang Shi <shy828301@gmail.com>,
        Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH 21/46] hugetlb: use struct hugetlb_pte for
 walk_hugetlb_range
Message-ID: <Y9Li93O6Ffwcr+vn@x1n>
References: <06423461-c543-56fe-cc63-cabda6871104@redhat.com>
 <CADrL8HUdg1Fr=tLEQRkDjeTzNzzSM6EPhvDgzURxSZSBMLgjoQ@mail.gmail.com>
 <6548b3b3-30c9-8f64-7d28-8a434e0a0b80@redhat.com>
 <Y8gRpEonhXgqfb41@x1n>
 <CADrL8HVGMTowH4trJhS+eM_EwZKoUgu7LmfwyTGyGRnNnwL3Zg@mail.gmail.com>
 <Y8hITxr/BBMuO6WX@monkey>
 <CADrL8HUggALQET-09Zw3BhFjZdw_G9+v6CU=qtGtK=KZ_DeAsw@mail.gmail.com>
 <Y8l+f2wNp2gAjvYg@monkey>
 <CADrL8HVdL_NMdNq2mEemNCfwkYBAWnbqwyjsAYdQ2fF0iz34Hw@mail.gmail.com>
 <CADrL8HV92DaNm5bUwcOxsG8Lg4yLT6F13KWSbjkySPNAsgCfpA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <CADrL8HV92DaNm5bUwcOxsG8Lg4yLT6F13KWSbjkySPNAsgCfpA@mail.gmail.com>
Precedence: bulk

James,

On Thu, Jan 26, 2023 at 08:58:51AM -0800, James Houghton wrote:
> It turns out that the THP-like scheme significantly slows down
> MADV_COLLAPSE: decrementing the mapcounts for the 4K subpages becomes
> the vast majority of the time spent in MADV_COLLAPSE when collapsing
> 1G mappings. It is doing 262k atomic decrements, so this makes sense.
> 
> This is only really a problem because this is done between
> mmu_notifier_invalidate_range_start() and
> mmu_notifier_invalidate_range_end(), so KVM won't allow vCPUs to
> access any of the 1G page while we're doing this (and it can take like
> ~1 second for each 1G, at least on the x86 server I was testing on).

Did you try to measure the time, or it's a quick observation from perf?

IIRC I used to measure some atomic ops, it is not as drastic as I thought.
But maybe it depends on many things.

I'm curious how the 1sec is provisioned between the procedures.  E.g., I
would expect mmu_notifier_invalidate_range_start() to also take some time
too as it should walk the smally mapped EPT pgtables.

Since we'll still keep the intermediate levels around - from application
POV, one other thing to remedy this is further shrink the size of COLLAPSE
so potentially for a very large page we can start with building 2M layers.
But then collapse will need to be run at least two rounds.

-- 
Peter Xu