Subject: Re: [PATCH 2/2] mremap: use mmu gather logic for tlb flush in mremap
To: Aaron Lu <aaron.lu@intel.com>,
        Linus Torvalds <torvalds@linux-foundation.org>
References: <026b73f6-ca1d-e7bb-766c-4aaeb7071ce6@intel.com>
 <CA+55aFzHfpZckv8ck19fZSFK+3TmR5eF=BsDzhwVGKrbyEBjEw@mail.gmail.com>
 <c160bc18-7c1b-2d54-8af1-7c5bfcbcefe8@intel.com>
 <20161128083715.GA21738@aaronlu.sh.intel.com>
 <20161128084012.GC21738@aaronlu.sh.intel.com>
Cc: Linux Memory Management List <linux-mm@kvack.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>,
        Huang Ying <ying.huang@intel.com>, linux-kernel@vger.kernel.org
From: Dave Hansen <dave.hansen@intel.com>
Message-ID: <31b1393c-22d2-8a49-842c-9678a1921441@intel.com>
Date: Mon, 28 Nov 2016 09:32:35 -0800
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101
 Thunderbird/45.4.0
MIME-Version: 1.0
In-Reply-To: <20161128084012.GC21738@aaronlu.sh.intel.com>
Content-Type: text/plain; charset=windows-1252
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1548
Lines: 33

On 11/28/2016 12:40 AM, Aaron Lu wrote:
> As suggested by Linus, the same mmu gather logic could be used for tlb
> flush in mremap and this patch just did that.
> 
> Note that since there is no page added to "struct mmu_gather" for free
> during mremap, when tlb needs to be flushed, we can use tlb_flush_mmu or
> tlb_flush_mmu_tlbonly. Using tlb_flush_mmu could also avoid exporting
> tlb_flush_mmu_tlbonly. But tlb_flush_mmu_tlbonly *looks* more clear and
> straightforward so I end up using it.

OK, so the code before this patch was passing around a pointer to
'need_flush', and we basically just pass around an mmu_gather instead.
It doesn't really simplify the code _too_ much, although it does make it
less confusing than when we saw 'need_flush' mixed with 'force_flush' in
the code.

tlb_flush_mmu_tlbonly() has exactly one other use: zap_pte_range() for
flushing the TLB before pte_unmap_unlock():

        if (force_flush)
                tlb_flush_mmu_tlbonly(tlb);
        pte_unmap_unlock(start_pte, ptl);

But, both call-sites are still keeping 'force_flush' to store the
information about whether we ever saw a dirty pte.  If we moved _that_
logic into the x86 mmu_gather code, we could get rid of all the
'force_flush' tracking in both call sites.  It also makes us a bit more
future-proof against these page_mkclean() races if we ever grow a third
site for clearing ptes.

Instead of exporting and calling tlb_flush_mmu_tlbonly(), we'd need
something like tlb_flush_mmu_before_ptl_release() (but with a better
name, of course :).