Re: [PATCH 1/1] mm: change inlined allocation helpers to account at the call site

[Suren Baghdasaryan]

On Fri, Apr 5, 2024 at 5:44 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Apr 04, 2024 at 07:00:51PM -0400, Kent Overstreet wrote:
> > On Thu, Apr 04, 2024 at 03:41:50PM -0700, Andrew Morton wrote:
> > > On Thu, 4 Apr 2024 18:38:39 -0400 Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote:
> > >
> > > > On Thu, Apr 04, 2024 at 11:33:22PM +0100, Matthew Wilcox wrote:
> > > > > On Thu, Apr 04, 2024 at 03:17:43PM -0700, Suren Baghdasaryan wrote:
> > > > > > Ironically, checkpatch generates warnings for these type casts:
> > > > > >
> > > > > > WARNING: unnecessary cast may hide bugs, see
> > > > > > http://c-faq.com/malloc/mallocnocast.html
> > > > > > #425: FILE: include/linux/dma-fence-chain.h:90:
> > > > > > + ((struct dma_fence_chain *)kmalloc(sizeof(struct dma_fence_chain),
> > > > > > GFP_KERNEL))
> > > > > >
> > > > > > I guess I can safely ignore them in this case (since we cast to the
> > > > > > expected type)?
> > > > >
> > > > > I find ignoring checkpatch to be a solid move 99% of the time.
> > > > >
> > > > > I really don't like the codetags.  This is so much churn, and it could
> > > > > all be avoided by just passing in _RET_IP_ or _THIS_IP_ depending on
> > > > > whether we wanted to profile this function or its caller.  vmalloc
> > > > > has done it this way since 2008 (OK, using __builtin_return_address())
> > > > > and lockdep has used _THIS_IP_ / _RET_IP_ since 2006.
> > > >
> > > > Except you can't. We've been over this; using that approach for tracing
> > > > is one thing, using it for actual accounting isn't workable.
> > >
> > > I missed that.  There have been many emails.  Please remind us of the
> > > reasoning here.
> >
> > I think it's on the other people claiming 'oh this would be so easy if
> > you just do it this other way' to put up some code - or at least more
> > than hot takes.
>
> Well, /proc/vmallocinfo exists, and has existed since 2008, so this is
> slightly more than a "hot take".
>
> > But, since you asked - one of the main goals of this patchset was to be
> > fast enough to run in production, and if you do it by return address
> > then you've added at minimum a hash table lookup to every allocate and
> > free; if you do that, running it in production is completely out of the
> > question.
>
> And yet vmalloc doesn't do that.
>
> > Besides that - the issues with annotating and tracking the correct
> > callsite really don't go away, they just shift around a bit. It's true
> > that the return address approach would be easier initially, but that's
> > not all we're concerned with; we're concerned with making sure
> > allocations get accounted to the _correct_ callsite so that we're giving
> > numbers that you can trust, and by making things less explicit you make
> > that harder.
>
> I'm not convinced that _THIS_IP_ is less precise than a codetag.  They
> do essentially the same thing, except that codetags embed the source
> location in the file while _THIS_IP_ requires a tool like faddr2line
> to decode kernel_clone+0xc0/0x430 into a file + line number.
>
> > This is all stuff that I've explained before; let's please dial back on
> > the whining - or I'll just bookmark this for next time...
>
> Please stop mischaracterising serious thoughtful criticism as whining.
> I don't understand what value codetags bring over using _THIS_IP_ and
> _RET_IP_ and you need to explain that.

The conceptual difference between codetag and _THIS_IP_/_RET_IP_ is
that codetag injects counters at the call site, so you don't need to
spend time finding the appropriate counter to operate on during
allocation.