From: Tero Kristo Subject: Re: [PATCH 02/28] crypto: omap-sham: Don't idle/start SHA device between Encrypt operations Date: Wed, 22 Jun 2016 12:17:09 +0300 Message-ID: <576A5795.5030300@ti.com> References: <1464771389-10640-1-git-send-email-t-kristo@ti.com> <1464771389-10640-3-git-send-email-t-kristo@ti.com> <574EB091.6040001@ti.com> <574F69D8.1010809@ti.com> <20160607100837.GA30875@gondor.apana.org.au> <5756B584.2000900@ti.com> <5756BD0B.4030506@ti.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii"; Format="flowed" Content-Transfer-Encoding: 7bit Cc: tony@atomide.com, lokeshvutla@ti.com, linux-crypto@vger.kernel.org, linux-omap@vger.kernel.org, davem@davemloft.net, linux-arm-kernel@lists.infradead.org To: Grygorii Strashko , Herbert Xu , Dave Gerlach Return-path: In-Reply-To: <5756BD0B.4030506@ti.com> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: "linux-arm-kernel" Errors-To: linux-arm-kernel-bounces+linux-arm-kernel=m.gmane.org@lists.infradead.org List-Id: linux-crypto.vger.kernel.org On 07/06/16 15:24, Grygorii Strashko wrote: > On 06/07/2016 02:52 PM, Tero Kristo wrote: >> On 07/06/16 13:08, Herbert Xu wrote: >>> On Wed, Jun 01, 2016 at 06:03:52PM -0500, Dave Gerlach wrote: >>>> On 06/01/2016 04:53 AM, Grygorii Strashko wrote: >>>>> On 06/01/2016 11:56 AM, Tero Kristo wrote: >>>>>> From: Lokesh Vutla >>>>>> >>>>>> Calling runtime PM API for every block causes serious perf hit to >>>>>> crypto operations that are done on a long buffer. >>>>>> As crypto is performed on a page boundary, encrypting large buffers >>>>>> can >>>>>> cause a series of crypto operations divided by page. The runtime PM >>>>>> API >>>>>> is also called those many times. >>>>>> >>>>>> We call runtime_pm_get_sync only at beginning on the session >>>>>> (cra_init) >>>>>> and runtime_pm_put at the end. This result in upto a 50% speedup. >>>>>> This doesn't make the driver to keep the system awake as runtime >>>>>> get/put >>>>>> is only called during a crypto session which completes usually >>>>>> quickly. >>>>>> >>>>>> Signed-off-by: Lokesh Vutla >>>>>> Signed-off-by: Tero Kristo >>>>>> --- >>>>>> drivers/crypto/omap-sham.c | 27 +++++++++++++++++---------- >>>>>> 1 file changed, 17 insertions(+), 10 deletions(-) >>>>>> >>>>>> diff --git a/drivers/crypto/omap-sham.c b/drivers/crypto/omap-sham.c >>>>>> index 6eefaa2..bd0258f 100644 >>>>>> --- a/drivers/crypto/omap-sham.c >>>>>> +++ b/drivers/crypto/omap-sham.c >>>>>> @@ -360,14 +360,6 @@ static void omap_sham_copy_ready_hash(struct >>>>>> ahash_request *req) >>>>>> >>>>>> static int omap_sham_hw_init(struct omap_sham_dev *dd) >>>>>> { >>>>>> - int err; >>>>>> - >>>>>> - err = pm_runtime_get_sync(dd->dev); >>>>>> - if (err < 0) { >>>>>> - dev_err(dd->dev, "failed to get sync: %d\n", err); >>>>>> - return err; >>>>>> - } >>>>>> - >>>> >>>> Would it be worth it to investigate a pm_runtime autosuspend >>>> approach rather than knocking runtime PM out here completely? I am >>>> not clear if the overhead is coming from the pm_runtime calls >>>> themselves or the actual idling of the IP, but if it's the idling of >>>> the IP causing the slowdown, with a large enough autosuspend_delay >>>> we don't actually sleep between each block but after a long enough >>>> period of idle time we would actually suspend. >>> >>> Indeed, I think this patch is bogus. cra_init is associated >>> with the tfm object which is usually long-lived. So doing power >>> management there makes no sense. >>> >>> Cheers, >>> >> >> I can investigate this further, but I believe this patch itself gave a >> noticeable performance boost. >> >> This is an optimization anyway, and not critical for functionality. >> > > It is not critical only if below code would not introduce races I don't get your point here. This patch is an optimization, and the driver works fine without it. > + spin_lock_bh(&sham.lock); > + list_for_each_entry(dd, &sham.dev_list, list) { > + break; > + } > + spin_unlock_bh(&sham.lock); > > Is it guaranteed that dd will alive always at this moment? Typically yes, but I think there might be a race condition here if the driver is removed during operation. Anyway, I'll drop this patch and change the optimization to use autosuspend as Dave suggested; that gives almost the same performance boost as this one (I miss a couple of percent in the overall performance, but I can live with that.) -Tero > > + > + pm_runtime_get_sync(dd->dev); > > >