Re: [PATCH resend mm,net-next 2/3] mm: Add vm_insert_pages().

From: Arjun Roy <arjunroy@google.com>
To: Andrew Morton <akpm@linux-foundation.org>
Cc: Arjun Roy <arjunroy.kdev@gmail.com>,
	davem@davemloft.net, netdev@vger.kernel.org,  linux-mm@kvack.org,
	Eric Dumazet <edumazet@google.com>,
	 Soheil Hassas Yeganeh <soheil@google.com>,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [PATCH resend mm,net-next 2/3] mm: Add vm_insert_pages().
Date: Thu, 13 Feb 2020 09:09:13 -0800	[thread overview]
Message-ID: <CAOFY-A16-eba-aNO+=062jy67cRs7-YSGoctp4GVApBnhYFPhA@mail.gmail.com> (raw)
In-Reply-To: <20200212184101.b8551710bd19c8216d62290d@linux-foundation.org>

I think at least to start it probably makes sense to keep regular
vm_insert_page() around - smaller stack used, less branches, if you
know you just need one page - not sure if gcc would err towards
smaller binary or not when compiling.

Regarding the page_count() check - as far as I can tell that's just
checking to make sure that at least *someone* has a reference to the
page before inserting it; in the zerocopy case we most definitely do
but I guess a bad caller could call it with a bad page argument and
this would guard against that.

Actually, I appear to have fat fingered it - I intended for this check
to be in there but seem to have forgotten (per the comment "/* Defer
page refcount checking till we're about to map that page. */" but with
no actual check). So that check should go inside
insert_page_in_batch_locked(), right before the
validate_page_before_insert() check. I'll send an updated fixup diff
shortly.

-Arjun

On Wed, Feb 12, 2020 at 6:41 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Mon, 27 Jan 2020 18:59:57 -0800 Arjun Roy <arjunroy.kdev@gmail.com> wrote:
>
> > Add the ability to insert multiple pages at once to a user VM with
> > lower PTE spinlock operations.
> >
> > The intention of this patch-set is to reduce atomic ops for
> > tcp zerocopy receives, which normally hits the same spinlock multiple
> > times consecutively.
>
> Seems sensible, thanks.  Some other vm_insert_page() callers might want
> to know about this, but I can't immediately spot any which appear to be
> high bandwidth.
>
> Is there much point in keeping the vm_insert_page() implementation
> around?  Replace it with
>
> static inline int
> vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
>                struct page *page)
> {
>         return vm_insert_pages(vma, addr, &page, 1);
> }
>
> ?
>
> Also, vm_insert_page() does
>
>         if (!page_count(page))
>                 return -EINVAL;
>
> and this was not carried over into vm_insert_pages().  How come?
>
> I don't know what that test does - it was added by Linus in the
> original commit a145dd411eb28c83.  It's only been 15 years so I'm sure
> he remembers ;)