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;
To:     Qi Zheng <zhengqi.arch@bytedance.com>, akpm@linux-foundation.org,
        tglx@linutronix.de, hannes@cmpxchg.org, mhocko@kernel.org,
        vdavydov.dev@gmail.com, kirill.shutemov@linux.intel.com,
        mika.penttila@nextfour.com
Cc:     linux-doc@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org, songmuchun@bytedance.com
References: <20210819031858.98043-1-zhengqi.arch@bytedance.com>
 <20210819031858.98043-7-zhengqi.arch@bytedance.com>
From:   David Hildenbrand <david@redhat.com>
Organization: Red Hat
Subject: Re: [PATCH v2 6/9] mm: free user PTE page table pages
Message-ID: <5aa3020c-fcf2-87bd-31fe-e2b5c2aafcf2@redhat.com>
Date:   Thu, 19 Aug 2021 09:01:40 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101
 Thunderbird/78.11.0
MIME-Version: 1.0
In-Reply-To: <20210819031858.98043-7-zhengqi.arch@bytedance.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 19.08.21 05:18, Qi Zheng wrote:
> Some malloc libraries(e.g. jemalloc or tcmalloc) usually
> allocate the amount of VAs by mmap() and do not unmap
> those VAs. They will use madvise(MADV_DONTNEED) to free
> physical memory if they want. But the page tables do not
> be freed by madvise(), so it can produce many page tables
> when the process touches an enormous virtual address space.
> 
> The following figures are a memory usage snapshot of one
> process which actually happened on our server:
> 
>          VIRT:  55t
>          RES:   590g
>          VmPTE: 110g
> 
> As we can see, the PTE page tables size is 110g, while the
> RES is 590g. In theory, the process only need 1.2g PTE page
> tables to map those physical memory. The reason why PTE page
> tables occupy a lot of memory is that madvise(MADV_DONTNEED)
> only empty the PTE and free physical memory but doesn't free
> the PTE page table pages. So we can free those empty PTE page
> tables to save memory. In the above cases, we can save memory
> about 108g(best case). And the larger the difference between
> the size of VIRT and RES, the more memory we save.
> 
> In this patch series, we add a pte_refcount field to the
> struct page of page table to track how many users of PTE page
> table. Similar to the mechanism of page refcount, the user of
> PTE page table should hold a refcount to it before accessing.
> The PTE page table page will be freed when the last refcount
> is dropped.
> 
> While we access ->pte_refcount of a PTE page table, any of the
> following ensures the pmd entry corresponding to the PTE page
> table stability:
> 
> 	- mmap_lock
> 	- anon_lock
> 	- i_mmap_lock
> 	- parallel threads are excluded by other means which
> 	  can make ->pmd stable(e.g. gup case)
> 
> This patch does not support THP temporarily, it will be
> supported in the next patch.

Can you clarify (and document here) who exactly takes a reference on the 
page table? Do I understand correctly that

a) each !pte_none() entry inside a page table take a reference to the 
page it's containted in.
b) each page table walker temporarily grabs a page table reference
c) The PMD tables the PTE is referenced in (->currently only ever a 
single one) does *not* take a reference.

So if there are no PTE entries left and nobody walks the page tables, 
you can remove it? You should really extend the 
description/documentation to make it clearer how exactly it's supposed 
to work.


It feels kind of strange to not introduce the CONFIG_FREE_USER_PTE 
Kconfig option in this patch. At least it took me a while to identify it 
in the previous patch.

Maybe you should introduce the empty stubs and use them in a separate 
patch, and then have this patch just introduce CONFIG_FREE_USER_PTE 
along with the actual refcounting magic inside the !stub implementation.

-- 
Thanks,

David / dhildenb