[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1303296040.8345.156.camel@twins>
Date: Wed, 20 Apr 2011 12:40:40 +0200
From: Peter Zijlstra <a.p.zijlstra@...llo.nl>
To: Andrew Morton <akpm@...ux-foundation.org>
Cc: Andrea Arcangeli <aarcange@...hat.com>,
Avi Kivity <avi@...hat.com>,
Thomas Gleixner <tglx@...utronix.de>,
Rik van Riel <riel@...hat.com>, Ingo Molnar <mingo@...e.hu>,
Linus Torvalds <torvalds@...ux-foundation.org>,
linux-kernel@...r.kernel.org, linux-arch@...r.kernel.org,
linux-mm@...ck.org,
Benjamin Herrenschmidt <benh@...nel.crashing.org>,
David Miller <davem@...emloft.net>,
Hugh Dickins <hugh.dickins@...cali.co.uk>,
Mel Gorman <mel@....ul.ie>, Nick Piggin <npiggin@...nel.dk>,
Paul McKenney <paulmck@...ux.vnet.ibm.com>,
Yanmin Zhang <yanmin_zhang@...ux.intel.com>,
Hugh Dickins <hughd@...gle.com>
Subject: Re: [PATCH 12/20] mm: Extended batches for generic mmu_gather
On Tue, 2011-04-19 at 13:06 -0700, Andrew Morton wrote:
> On Fri, 01 Apr 2011 14:13:10 +0200
> Peter Zijlstra <a.p.zijlstra@...llo.nl> wrote:
>
> > Instead of using a single batch (the small on-stack, or an allocated
> > page), try and extend the batch every time it runs out and only flush
> > once either the extend fails or we're done.
>
> why?
To avoid sending extra TLB invalidates.
> > @@ -86,22 +86,48 @@ struct mmu_gather {
> > #ifdef CONFIG_HAVE_RCU_TABLE_FREE
> > struct mmu_table_batch *batch;
> > #endif
> > + unsigned int need_flush : 1, /* Did free PTEs */
> > + fast_mode : 1; /* No batching */
>
> mmu_gather.fast_mode gets modified in several places apparently without
> locking to protect itself. I don't think that these modifications will
> accidentally trash need_flush, mainly by luck.
The other way around I'd think.
> Please review the concurrency issues here and document them clearly.
Its an on-stack structure, there is no concurrency. /me shall add a
comment.
> > +#ifdef CONFIG_SMP
> > + #define tlb_fast_mode(tlb) (tlb->fast_mode)
> > +#else
> > + #define tlb_fast_mode(tlb) 1
> > +#endif
>
> Mutter.
>
> Could have been written in C.
Fixed in my last patch uninlining bits
> Will cause a compile error with, for example, tlb_fast_mode(tlb + 1).
Well, that'd actually be a good reason to keep the macro ;-)
> > +static inline int tlb_next_batch(struct mmu_gather *tlb)
> > {
> > + struct mmu_gather_batch *batch;
> >
> > + batch = tlb->active;
> > + if (batch->next) {
> > + tlb->active = batch->next;
> > + return 1;
> > }
> > +
> > + batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
>
> A comment explaining the gfp_t decision would be useful.
Done
> > + if (!batch)
> > + return 0;
> > +
> > + batch->next = NULL;
> > + batch->nr = 0;
> > + batch->max = MAX_GATHER_BATCH;
> > +
> > + tlb->active->next = batch;
> > + tlb->active = batch;
> > +
> > + return 1;
> > }
> >
> > /* tlb_gather_mmu
> > @@ -114,16 +140,13 @@ tlb_gather_mmu(struct mmu_gather *tlb, s
> > {
> > tlb->mm = mm;
> >
> > + tlb->fullmm = fullmm;
> > + tlb->need_flush = 0;
> > + tlb->fast_mode = (num_possible_cpus() == 1);
>
> The changelog didn't tell us why we switched from num_online_cpus() to
> num_possible_cpus().
And that is a very good question... somehow I remember a conversation
with BenH about this, but on second thought that might have been about
his pgtable_free_tlb() optimization (which is somewhat similar).
Let me restore that to num_online_cpus() and maybe do a later patch
removing fast_mode all together as Hugh suggested, since even UP might
have benefit from the batching due to less zone-lock activity on bulk
frees.
> > + tlb->local.next = NULL;
> > + tlb->local.nr = 0;
> > + tlb->local.max = ARRAY_SIZE(tlb->__pages);
> > + tlb->active = &tlb->local;
> >
> > #ifdef CONFIG_HAVE_RCU_TABLE_FREE
> > tlb->batch = NULL;
> >
> > ...
> >
> > @@ -177,15 +205,24 @@ tlb_finish_mmu(struct mmu_gather *tlb, u
> > + batch = tlb->active;
> > + batch->pages[batch->nr++] = page;
> > + VM_BUG_ON(batch->nr > batch->max);
> > + if (batch->nr == batch->max) {
> > + if (!tlb_next_batch(tlb))
> > + return 0;
> > + }
>
> Moving the VM_BUG_ON() down to after the if() would save a few cycles.
Done.
> > + return batch->max - batch->nr;
> > }
> >
> > /* tlb_remove_page
> >
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists