Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Fri, Nov 10, 2017 at 10:44:30AM -0600, Kim Phillips wrote: > On Fri, 10 Nov 2017 08:02:01 + > Radu Andrei Alexe wrote: > > > On 11/9/2017 6:34 PM, Kim Phillips wrote: > > > On Thu, 9 Nov 2017 11:54:13 + > > > Radu Andrei Alexe wrote: > > >> The next patch version will create the platform device dynamically at > > >> run time. > > > > > > Why create a new device when that h/w already has one? > > > > > > Why doesn't the existing crypto driver register dma capabilities with > > > the dma driver subsystem? > > > > > I can think of two reasons: > > > > 1. The code that this driver introduces has nothing to do with crypto > > and everything to do with dma. > > I would think that at least a crypto "null" algorithm implementation > would share code. > > > Placing the code in the same directory as > > the caam subsystem would only create confusion for the reader of an > > already complex driver. > > this different directory argument seems to be identical to your 2 below: > > > 2. I wanted this driver to be tracked by the dma engine team. They have > > the right expertise to provide adequate feedback. If all the code was in > > the crypto directory they wouldn't know about this driver or any > > subsequent changes to it. > > dma subsystem bits could still be put in the dma area if deemed > necessary but I don't think it is: I see > drivers/crypto/ccp/ccp-dmaengine.c calls dma_async_device_register for > example. which is a shame as it was sneaked past the dmaengine community!! This is the *only* example and there and other examples where people use dmaengine: $ grep -rl dmaengine_prep* drivers/crypto/* |uniq drivers/crypto/atmel-aes.c drivers/crypto/atmel-sha.c drivers/crypto/atmel-tdes.c drivers/crypto/img-hash.c drivers/crypto/omap-aes.c drivers/crypto/omap-des.c drivers/crypto/omap-sham.c drivers/crypto/qce/dma.c drivers/crypto/stm32/stm32-hash.c drivers/crypto/ux500/cryp/cryp_core.c drivers/crypto/ux500/hash/hash_core.c > > I also don't see how that complicates things much further. > > What is the rationale for using the crypto h/w as a dma engine anyway? > Are there supporting performance figures? that is a very good question, perf does matter. Given that we have many folks using it, I think it would help, but yes nothing better than numbers speak for themselves. -- ~Vinod
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Fri, Nov 10, 2017 at 08:02:01AM +, Radu Andrei Alexe wrote: > On 11/9/2017 6:34 PM, Kim Phillips wrote: > > On Thu, 9 Nov 2017 11:54:13 + > > Radu Andrei Alexe wrote: > > > >> On 10/30/2017 4:24 PM, Kim Phillips wrote: > >>> On Mon, 30 Oct 2017 15:46:51 +0200 > >>> Horia Geantă wrote: > >>> > += > +CAAM DMA Node > + > +Child node of the crypto node that enables the use of the DMA > capabilities > +of the CAAM by a stand-alone driver. The only required property is > the > +"compatible" property. All the other properties are determined from > +the job rings on which the CAAM DMA driver depends (ex: the number > of > +dma-channels is equal to the number of defined job rings). > + > + - compatible > + Usage: required > + Value type: > + Definition: Must include "fsl,sec-v4.0-dma" > + > +EXAMPLE > + caam-dma { > +compatible = "fsl,sec-v5.4-dma", > + "fsl,sec-v5.0-dma", > + "fsl,sec-v4.0-dma"; > + } > >>> > >>> If this isn't describing an aspect of the hardware, then what is it > >>> doing in the device tree? > >>> > >>> Kim > >>> > >> > >> Because the caam_dma driver is a platform driver I needed to create a > >> platform device to activate the driver. My thought was that it was > >> simpler to implement it using device tree. > >> The next patch version will create the platform device dynamically at > >> run time. > > > > Why create a new device when that h/w already has one? > > > > Why doesn't the existing crypto driver register dma capabilities with > > the dma driver subsystem? > > > > Kim > > > > > I can think of two reasons: > > 1. The code that this driver introduces has nothing to do with crypto > and everything to do with dma. Placing the code in the same directory as > the caam subsystem would only create confusion for the reader of an > already complex driver. > > 2. I wanted this driver to be tracked by the dma engine team. They have > the right expertise to provide adequate feedback. If all the code was in > the crypto directory they wouldn't know about this driver or any > subsequent changes to it. These are very good reasons. We already have one DMA implementation drivers/crypto/ccp/ccp-dmaengine.c which was sneaked past us and put into kernel with proper review! On the other hand we have *bunch* of dmaengine users in crypto subsystem which use dmaengine drivers and APIs and are good citizens. It allows folks to share code with other usages of dmaengine and the usual arguments for common code and frameworks... If there are enough users we can add up a crypto-dmaengine lib which programs dma controller for crypto users and avoid open coding for everyone. for example sound-dmaengine layers does that... HTH -- ~Vinod
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Mon, 13 Nov 2017 09:44:06 + Radu Andrei Alexe wrote: > On 11/10/2017 6:44 PM, Kim Phillips wrote: > > On Fri, 10 Nov 2017 08:02:01 + > > Radu Andrei Alexe wrote: > > > >> On 11/9/2017 6:34 PM, Kim Phillips wrote: > >>> On Thu, 9 Nov 2017 11:54:13 + > >>> Radu Andrei Alexe wrote: > The next patch version will create the platform device dynamically at > run time. > >>> > >>> Why create a new device when that h/w already has one? > >>> > >>> Why doesn't the existing crypto driver register dma capabilities with > >>> the dma driver subsystem? > >>> > >> I can think of two reasons: > >> > >> 1. The code that this driver introduces has nothing to do with crypto > >> and everything to do with dma. > > > > I would think that at least a crypto "null" algorithm implementation > > would share code. > > > >> Placing the code in the same directory as > >> the caam subsystem would only create confusion for the reader of an > >> already complex driver. > > > > this different directory argument seems to be identical to your 2 below: > > > >> 2. I wanted this driver to be tracked by the dma engine team. They have > >> the right expertise to provide adequate feedback. If all the code was in > >> the crypto directory they wouldn't know about this driver or any > >> subsequent changes to it. > > > > dma subsystem bits could still be put in the dma area if deemed > > necessary but I don't think it is: I see > > drivers/crypto/ccp/ccp-dmaengine.c calls dma_async_device_register for > > example. > > > > I also don't see how that complicates things much further. > > > > So who made their review? The guys from crypto? Don't see how that's relevant here, but people applying patches should solicit acks from the appropriate sources, esp. if a patch is across multiple subsystems. > If someone wants to enable only the DMA functionality of the CCP and not > the crypto part how do they do it? Look for it in the crypto submenu? Why would they want to do that? In any case, I suspect you're thinking about cross-subsystem Kconfig entries, which is common, but something like that can be a module parameter, too. I would say that maybe CRYPTO_DEV_FSL_CAAM should be made to not depend on CRYPTO_HW, but I think that's overkill for the addition of this minor feature. > > What is the rationale for using the crypto h/w as a dma engine anyway? > > Are there supporting performance figures? > > We have a platform that doesn't have a dedicated DMA controller but has > the CAAM hardware block that can perform dma transfers. We have a OK, please mention that next time. > use-case where we need to issue large transfers (hundred of MBs) > asynchronously, without using the core. Curious: what subsystem does that? Thanks, Kim
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Mon, 13 Nov 2017 08:32:24 + Horia Geantă wrote: > On 11/10/2017 6:44 PM, Kim Phillips wrote: > > On Fri, 10 Nov 2017 08:02:01 + > > Radu Andrei Alexe wrote: > [snip]>> 2. I wanted this driver to be tracked by the dma engine team. > They have > >> the right expertise to provide adequate feedback. If all the code was in > >> the crypto directory they wouldn't know about this driver or any > >> subsequent changes to it. > > > > dma subsystem bits could still be put in the dma area if deemed > > necessary but I don't think it is: I see > > drivers/crypto/ccp/ccp-dmaengine.c calls dma_async_device_register for > > example. > > > Please see previous discussion with Vinod: > https://www.mail-archive.com/linux-crypto@vger.kernel.org/msg21468.html Vinod says: "If the dma controller is internal to crypto, then it might be okay to be inside the crypto driver." Is that the case for the CCP driver? Isn't it the case here? In any case, I don't care that much about that, this all begat from new *devices* coming out of nowhere. > > What is the rationale for using the crypto h/w as a dma engine anyway? > SoCs that don't have a system DMA, for e.g. LS1012A. OK. Kim
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On 11/10/2017 6:44 PM, Kim Phillips wrote: > On Fri, 10 Nov 2017 08:02:01 + > Radu Andrei Alexe wrote: > >> On 11/9/2017 6:34 PM, Kim Phillips wrote: >>> On Thu, 9 Nov 2017 11:54:13 + >>> Radu Andrei Alexe wrote: The next patch version will create the platform device dynamically at run time. >>> >>> Why create a new device when that h/w already has one? >>> >>> Why doesn't the existing crypto driver register dma capabilities with >>> the dma driver subsystem? >>> >> I can think of two reasons: >> >> 1. The code that this driver introduces has nothing to do with crypto >> and everything to do with dma. > > I would think that at least a crypto "null" algorithm implementation > would share code. > >> Placing the code in the same directory as >> the caam subsystem would only create confusion for the reader of an >> already complex driver. > > this different directory argument seems to be identical to your 2 below: > >> 2. I wanted this driver to be tracked by the dma engine team. They have >> the right expertise to provide adequate feedback. If all the code was in >> the crypto directory they wouldn't know about this driver or any >> subsequent changes to it. > > dma subsystem bits could still be put in the dma area if deemed > necessary but I don't think it is: I see > drivers/crypto/ccp/ccp-dmaengine.c calls dma_async_device_register for > example. > > I also don't see how that complicates things much further. > So who made their review? The guys from crypto? If someone wants to enable only the DMA functionality of the CCP and not the crypto part how do they do it? Look for it in the crypto submenu? > What is the rationale for using the crypto h/w as a dma engine anyway? > Are there supporting performance figures? We have a platform that doesn't have a dedicated DMA controller but has the CAAM hardware block that can perform dma transfers. We have a use-case where we need to issue large transfers (hundred of MBs) asynchronously, without using the core. > > Kim > BR, Radu
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On 11/10/2017 6:44 PM, Kim Phillips wrote: > On Fri, 10 Nov 2017 08:02:01 + > Radu Andrei Alexe wrote: [snip]>> 2. I wanted this driver to be tracked by the dma engine team. They have >> the right expertise to provide adequate feedback. If all the code was in >> the crypto directory they wouldn't know about this driver or any >> subsequent changes to it. > > dma subsystem bits could still be put in the dma area if deemed > necessary but I don't think it is: I see > drivers/crypto/ccp/ccp-dmaengine.c calls dma_async_device_register for > example. > Please see previous discussion with Vinod: https://www.mail-archive.com/linux-crypto@vger.kernel.org/msg21468.html > What is the rationale for using the crypto h/w as a dma engine anyway? SoCs that don't have a system DMA, for e.g. LS1012A. Horia
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Fri, 10 Nov 2017 08:02:01 + Radu Andrei Alexe wrote: > On 11/9/2017 6:34 PM, Kim Phillips wrote: > > On Thu, 9 Nov 2017 11:54:13 + > > Radu Andrei Alexe wrote: > >> The next patch version will create the platform device dynamically at > >> run time. > > > > Why create a new device when that h/w already has one? > > > > Why doesn't the existing crypto driver register dma capabilities with > > the dma driver subsystem? > > > I can think of two reasons: > > 1. The code that this driver introduces has nothing to do with crypto > and everything to do with dma. I would think that at least a crypto "null" algorithm implementation would share code. > Placing the code in the same directory as > the caam subsystem would only create confusion for the reader of an > already complex driver. this different directory argument seems to be identical to your 2 below: > 2. I wanted this driver to be tracked by the dma engine team. They have > the right expertise to provide adequate feedback. If all the code was in > the crypto directory they wouldn't know about this driver or any > subsequent changes to it. dma subsystem bits could still be put in the dma area if deemed necessary but I don't think it is: I see drivers/crypto/ccp/ccp-dmaengine.c calls dma_async_device_register for example. I also don't see how that complicates things much further. What is the rationale for using the crypto h/w as a dma engine anyway? Are there supporting performance figures? Kim
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On 11/9/2017 6:34 PM, Kim Phillips wrote: > On Thu, 9 Nov 2017 11:54:13 + > Radu Andrei Alexe wrote: > >> On 10/30/2017 4:24 PM, Kim Phillips wrote: >>> On Mon, 30 Oct 2017 15:46:51 +0200 >>> Horia Geantă wrote: >>> += +CAAM DMA Node + +Child node of the crypto node that enables the use of the DMA capabilities +of the CAAM by a stand-alone driver. The only required property is the +"compatible" property. All the other properties are determined from +the job rings on which the CAAM DMA driver depends (ex: the number of +dma-channels is equal to the number of defined job rings). + + - compatible + Usage: required + Value type: + Definition: Must include "fsl,sec-v4.0-dma" + +EXAMPLE + caam-dma { +compatible = "fsl,sec-v5.4-dma", + "fsl,sec-v5.0-dma", + "fsl,sec-v4.0-dma"; + } >>> >>> If this isn't describing an aspect of the hardware, then what is it >>> doing in the device tree? >>> >>> Kim >>> >> >> Because the caam_dma driver is a platform driver I needed to create a >> platform device to activate the driver. My thought was that it was >> simpler to implement it using device tree. >> The next patch version will create the platform device dynamically at >> run time. > > Why create a new device when that h/w already has one? > > Why doesn't the existing crypto driver register dma capabilities with > the dma driver subsystem? > > Kim > I can think of two reasons: 1. The code that this driver introduces has nothing to do with crypto and everything to do with dma. Placing the code in the same directory as the caam subsystem would only create confusion for the reader of an already complex driver. 2. I wanted this driver to be tracked by the dma engine team. They have the right expertise to provide adequate feedback. If all the code was in the crypto directory they wouldn't know about this driver or any subsequent changes to it. BR, Radu
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Thu, 9 Nov 2017 11:54:13 + Radu Andrei Alexe wrote: > On 10/30/2017 4:24 PM, Kim Phillips wrote: > > On Mon, 30 Oct 2017 15:46:51 +0200 > > Horia Geantă wrote: > > > >> += > >> +CAAM DMA Node > >> + > >> +Child node of the crypto node that enables the use of the DMA > >> capabilities > >> +of the CAAM by a stand-alone driver. The only required property is the > >> +"compatible" property. All the other properties are determined from > >> +the job rings on which the CAAM DMA driver depends (ex: the number of > >> +dma-channels is equal to the number of defined job rings). > >> + > >> + - compatible > >> + Usage: required > >> + Value type: > >> + Definition: Must include "fsl,sec-v4.0-dma" > >> + > >> +EXAMPLE > >> + caam-dma { > >> +compatible = "fsl,sec-v5.4-dma", > >> + "fsl,sec-v5.0-dma", > >> + "fsl,sec-v4.0-dma"; > >> + } > > > > If this isn't describing an aspect of the hardware, then what is it > > doing in the device tree? > > > > Kim > > > > Because the caam_dma driver is a platform driver I needed to create a > platform device to activate the driver. My thought was that it was > simpler to implement it using device tree. > The next patch version will create the platform device dynamically at > run time. Why create a new device when that h/w already has one? Why doesn't the existing crypto driver register dma capabilities with the dma driver subsystem? Kim
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On 10/30/2017 4:24 PM, Kim Phillips wrote: > On Mon, 30 Oct 2017 15:46:51 +0200 > Horia Geantă wrote: > >> += >> +CAAM DMA Node >> + >> +Child node of the crypto node that enables the use of the DMA >> capabilities >> +of the CAAM by a stand-alone driver. The only required property is the >> +"compatible" property. All the other properties are determined from >> +the job rings on which the CAAM DMA driver depends (ex: the number of >> +dma-channels is equal to the number of defined job rings). >> + >> + - compatible >> + Usage: required >> + Value type: >> + Definition: Must include "fsl,sec-v4.0-dma" >> + >> +EXAMPLE >> + caam-dma { >> +compatible = "fsl,sec-v5.4-dma", >> + "fsl,sec-v5.0-dma", >> + "fsl,sec-v4.0-dma"; >> + } > > If this isn't describing an aspect of the hardware, then what is it > doing in the device tree? > > Kim > Because the caam_dma driver is a platform driver I needed to create a platform device to activate the driver. My thought was that it was simpler to implement it using device tree. The next patch version will create the platform device dynamically at run time. BR, Radu
Re: [PATCH RESEND 1/4] crypto: caam: add caam-dma node to SEC4.0 device tree binding
On Mon, 30 Oct 2017 15:46:51 +0200 Horia Geantă wrote: > += > +CAAM DMA Node > + > +Child node of the crypto node that enables the use of the DMA > capabilities > +of the CAAM by a stand-alone driver. The only required property is the > +"compatible" property. All the other properties are determined from > +the job rings on which the CAAM DMA driver depends (ex: the number of > +dma-channels is equal to the number of defined job rings). > + > + - compatible > + Usage: required > + Value type: > + Definition: Must include "fsl,sec-v4.0-dma" > + > +EXAMPLE > + caam-dma { > +compatible = "fsl,sec-v5.4-dma", > + "fsl,sec-v5.0-dma", > + "fsl,sec-v4.0-dma"; > + } If this isn't describing an aspect of the hardware, then what is it doing in the device tree? Kim