[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <9f778311-6d4d-cc57-00db-db2141803416@imgtec.com>
Date: Wed, 7 Dec 2016 09:30:23 +0000
From: Matt Redfearn <matt.redfearn@...tec.com>
To: "Jason A. Donenfeld" <Jason@...c4.com>
CC: Ralf Baechle <ralf@...ux-mips.org>, <linux-mips@...ux-mips.org>,
Thomas Gleixner <tglx@...utronix.de>,
Paolo Bonzini <pbonzini@...hat.com>,
Marcin Nowakowski <marcin.nowakowski@...tec.com>,
Chris Metcalf <cmetcalf@...lanox.com>,
Petr Mladek <pmladek@...e.com>,
LKML <linux-kernel@...r.kernel.org>,
Andrea Gelmini <andrea.gelmini@...ma.net>,
James Hogan <james.hogan@...tec.com>,
Paul Burton <paul.burton@...tec.com>,
Jiri Slaby <jslaby@...e.cz>,
"Maciej W. Rozycki" <macro@...tec.com>,
Aaron Tomlin <atomlin@...hat.com>,
Andrew Morton <akpm@...ux-foundation.org>
Subject: Re: [PATCH 0/5] MIPS: Add per-cpu IRQ stack
Hi Jason,
On 06/12/16 22:09, Jason A. Donenfeld wrote:
> Hi Matt,
>
> Thanks for submitting this. A happy OpenWRT/WireGuard user has
> reported to me that this patch set frees ~1300 bytes of stack on a
> small MIPS router. This kind of savings should allow me to reintroduce
> my crypto operations to be on the stack, rather than the conditional
> MIPS kmallocing, which will be a performance win.
Excellent.
>
> By the way, it looks like x86 and SPARC have separately allocated hard
> IRQ and soft IRQ stacks. ARM has only one for both, similar to this
> MIPS patchset.
The separate stack for soft-IRQs could be added as well, but of course
there is a tradeoff against the additional memory consumed by it. I
believe that most soft IRQs will now be handled on the new irq stack,
because HAVE_IRQ_EXIT_ON_IRQ_STACK causes them to be invoked within
irq_exit, while still executing on the irq stack
Thanks,
Matt
>
> Jason
>
> On Fri, Dec 2, 2016 at 2:39 PM, Matt Redfearn <matt.redfearn@...tec.com> wrote:
>> This series adds a separate stack for each CPU wihin the system to use
>> when handling IRQs. Previously IRQs were handled on the kernel stack of
>> the current task. If that task was deep down a call stack at the point
>> of the interrupt, and handling the interrupt required a deep IRQ stack,
>> then there was a likelihood of stack overflow. Since the kernel stack
>> is in normal unmapped memory, overflowing it can lead to silent
>> corruption of other kernel data, with weird and wonderful results.
>>
>> Before this patch series, ftracing the maximum stack size of a v4.9-rc6
>> kernel running on a Ci40 board gave:
>> 4996
>>
>> And with this series:
>> 4084
>>
>> Handling interrupts on a separate stack reduces the maximum kernel stack
>> usage in this configuration by ~900 bytes.
>>
>> Since do_IRQ is now invoked on a separate stack, we select
>> HAVE_IRQ_EXIT_ON_IRQ_STACK so that softirqs will also be executed on the
>> irq stack rather than attempting to switch with do_softirq_own_stack().
>>
>> This series has been tested on MIPS Boston, Malta and SEAD3 platforms,
>> Pistachio on the Creator Ci40 board and Cavium Octeon III.
>>
>>
>>
>> Matt Redfearn (5):
>> MIPS: Introduce irq_stack
>> MIPS: Stack unwinding while on IRQ stack
>> MIPS: Only change $28 to thread_info if coming from user mode
>> MIPS: Switch to the irq_stack in interrupts
>> MIPS: Select HAVE_IRQ_EXIT_ON_IRQ_STACK
>>
>> arch/mips/Kconfig | 1 +
>> arch/mips/include/asm/irq.h | 12 ++++++
>> arch/mips/include/asm/stackframe.h | 10 +++++
>> arch/mips/kernel/asm-offsets.c | 1 +
>> arch/mips/kernel/genex.S | 81 +++++++++++++++++++++++++++++++++++---
>> arch/mips/kernel/irq.c | 11 ++++++
>> arch/mips/kernel/process.c | 15 ++++++-
>> 7 files changed, 125 insertions(+), 6 deletions(-)
>>
>> --
>> 2.7.4
>>
Powered by blists - more mailing lists