Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S934701AbcKKRG2 (ORCPT ); Fri, 11 Nov 2016 12:06:28 -0500 Received: from mail-pf0-f195.google.com ([209.85.192.195]:33211 "EHLO mail-pf0-f195.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S934529AbcKKRGX (ORCPT ); Fri, 11 Nov 2016 12:06:23 -0500 Subject: Re: [PATCH V5 3/3] ARM64 LPC: LPC driver implementation on Hip06 To: "liviu.dudau@arm.com" , Gabriele Paoloni References: <1478576829-112707-1-git-send-email-yuanzhichang@hisilicon.com> <17821285.aIcTyCGn5n@wuerfel> <10334260.ztLXZ2Oynd@wuerfel> <20161111144539.GL10219@e106497-lin.cambridge.arm.com> Cc: Arnd Bergmann , "linux-arm-kernel@lists.infradead.org" , Yuanzhichang , "mark.rutland@arm.com" , "devicetree@vger.kernel.org" , "lorenzo.pieralisi@arm.com" , "minyard@acm.org" , "linux-pci@vger.kernel.org" , "benh@kernel.crashing.org" , John Garry , "will.deacon@arm.com" , "linux-kernel@vger.kernel.org" , "xuwei (O)" , Linuxarm , "zourongrong@gmail.com" , "robh+dt@kernel.org" , "kantyzc@163.com" , "linux-serial@vger.kernel.org" , "catalin.marinas@arm.com" , "olof@lixom.net" , "bhelgaas@googl e.com" From: "zhichang.yuan" Message-ID: <1eae3951-f1d9-c1c5-f70c-33150bc514ea@gmail.com> Date: Sat, 12 Nov 2016 00:54:05 +0800 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.4.0 MIME-Version: 1.0 In-Reply-To: <20161111144539.GL10219@e106497-lin.cambridge.arm.com> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 7926 Lines: 185 Hi, Liviu, On 11/11/2016 10:45 PM, liviu.dudau@arm.com wrote: > On Fri, Nov 11, 2016 at 01:39:35PM +0000, Gabriele Paoloni wrote: >> Hi Arnd >> >>> -----Original Message----- >>> From: Arnd Bergmann [mailto:arnd@arndb.de] >>> Sent: 10 November 2016 16:07 >>> To: Gabriele Paoloni >>> Cc: linux-arm-kernel@lists.infradead.org; Yuanzhichang; >>> mark.rutland@arm.com; devicetree@vger.kernel.org; >>> lorenzo.pieralisi@arm.com; minyard@acm.org; linux-pci@vger.kernel.org; >>> benh@kernel.crashing.org; John Garry; will.deacon@arm.com; linux- >>> kernel@vger.kernel.org; xuwei (O); Linuxarm; zourongrong@gmail.com; >>> robh+dt@kernel.org; kantyzc@163.com; linux-serial@vger.kernel.org; >>> catalin.marinas@arm.com; olof@lixom.net; liviu.dudau@arm.com; >>> bhelgaas@googl e.com; zhichang.yuan02@gmail.com >>> Subject: Re: [PATCH V5 3/3] ARM64 LPC: LPC driver implementation on >>> Hip06 >>> >>> On Thursday, November 10, 2016 3:36:49 PM CET Gabriele Paoloni wrote: >>>> >>>> Where should we get the range from? For LPC we know that it is going >>>> Work on anything that is not used by PCI I/O space, and this is >>>> why we use [0, PCIBIOS_MIN_IO] >>> >>> It should be allocated the same way we allocate PCI config space >>> segments. This is currently done with the io_range list in >>> drivers/pci/pci.c, which isn't perfect but could be extended >>> if necessary. Based on what others commented here, I'd rather >>> make the differences between ISA/LPC and PCI I/O ranges smaller >>> than larger. > > Gabriele, > >> >> I am not sure this would make sense... >> >> IMHO all the mechanism around io_range_list is needed to provide the >> "mapping" between I/O tokens and physical CPU addresses. >> >> Currently the available tokens range from 0 to IO_SPACE_LIMIT. >> >> As you know the I/O memory accessors operate on whatever >> __of_address_to_resource sets into the resource (start, end). >> >> With this special device in place we cannot know if a resource is >> assigned with an I/O token or a physical address, unless we forbid >> the I/O tokens to be in a specific range. >> >> So this is why we are changing the offsets of all the functions >> handling io_range_list (to make sure that a range is forbidden to >> the tokens and is available to the physical addresses). >> >> We have chosen this forbidden range to be [0, PCIBIOS_MIN_IO) >> because this is the maximum physical I/O range that a non PCI device >> can operate on and because we believe this does not impose much >> restriction on the available I/O token range; that now is >> [PCIBIOS_MIN_IO, IO_SPACE_LIMIT]. >> So we believe that the chosen forbidden range can accommodate >> any special ISA bus device with no much constraint on the rest >> of I/O tokens... > > Your idea is a good one, however you are abusing PCIBIOS_MIN_IO and you > actually need another variable for "reserving" an area in the I/O space > that can be used for physical addresses rather than I/O tokens. > I think selecting PCIBIOS_MIN_IO as the separator of mapped and non-mapped I/O range probably is not so reasonable. PCIBIOS_MIN_IN is specific to PCI devices, it seems as the recommended minimal start I/O address when assigning the pci device I/O region. It is probably not defined in some platforms/architectures when no PCI is needed there. That is why my patch caused some compile error on some archs; But more important thing is that the PCIBIOS_MIN_IO has different value on different platforms/architectures. On Arm64, it is 4K currently, but in other archs, it is not true. And the maximum LPC I/O address should be 64K theoretically, although for compatible ISA, 2K is enough. So, It means using PCIBIOS_MIN_IO on arm64 can match our I/O reservation require. But we can not make this indirectIO work well on other architectures. I am thinking Arnd's suggestion. But I worry about I haven't completely understood his idea. What about create a new bus host for LPC/ISA whose I/O range can be 64KB? This LPC/ISA I/O range works similar to PCI host bridge's I/O window, all the downstream devices under LPC/ISA should request I/O from that root resource. But it seems Arnd want this root resource registered dynamically, I am not sure how to do... Anyway, if we have this root I/O resource, we don't need any new macro or variable for the LPC/ISA I/O reservation. Hope my thought is right. Best, Zhichang > The one good example for using PCIBIOS_MIN_IO is when your platform/architecture > does not support legacy ISA operations *at all*. In that case someone > sets the PCIBIOS_MIN_IO to a non-zero value to reserve that I/O range > so that it doesn't get used. With Zhichang's patch you now start forcing > those platforms to have a valid address below PCIBIOS_MIN_IO. > > For the general case you also have to bear in mind that PCIBIOS_MIN_IO could > be zero. In that case, what is your "forbidden" range? [0, 0) ? So it makes > sense to add a new #define that should only be defined by those architectures/ > platforms that want to reserve on top of PCIBIOS_MIN_IO another region > where I/O tokens can't be generated for. > > Best regards, > Liviu > >> >>> >>>>> Your current version has >>>>> >>>>> if (arm64_extio_ops->pfout) \ >>>>> arm64_extio_ops->pfout(arm64_extio_ops->devpara,\ >>>>> addr, value, sizeof(type)); \ >>>>> >>>>> Instead, just subtract the start of the range from the logical >>>>> port number to transform it back into a bus-local port number: >>>> >>>> These accessors do not operate on IO tokens: >>>> >>>> If (arm64_extio_ops->start > addr || arm64_extio_ops->end < addr) >>>> addr is not going to be an I/O token; in fact patch 2/3 imposes that >>>> the I/O tokens will start at PCIBIOS_MIN_IO. So from 0 to >>> PCIBIOS_MIN_IO >>>> we have free physical addresses that the accessors can operate on. >>> >>> Ah, I missed that part. I'd rather not use PCIBIOS_MIN_IO to refer to >>> the logical I/O tokens, the purpose of that macro is really meant >>> for allocating PCI I/O port numbers within the address space of >>> one bus. >> >> As I mentioned above, special devices operate on CPU addresses directly, >> not I/O tokens. For them there is no way to distinguish.... >> >>> >>> Note that it's equally likely that whichever next platform needs >>> non-mapped I/O access like this actually needs them for PCI I/O space, >>> and that will use it on addresses registered to a PCI host bridge. >> >> Ok so here you are talking about a platform that has got an I/O range >> under the PCI host controller, right? >> And this I/O range cannot be directly memory mapped but needs special >> redirections for the I/O tokens, right? >> >> In this scenario registering the I/O ranges with the forbidden range >> implemented by the current patch would still allow to redirect I/O >> tokens as long as arm64_extio_ops->start >= PCIBIOS_MIN_IO >> >> So effectively the special PCI host controller >> 1) knows the physical range that needs special redirection >> 2) register such range >> 3) uses pci_pio_to_address() to retrieve the IO tokens for the >> special accessors >> 4) sets arm64_extio_ops->start/end to the IO tokens retrieved in 3) >> >> So to be honest I think this patch can fit well both with >> special PCI controllers that need I/O tokens redirection and with >> special non-PCI controllers that need non-PCI I/O physical >> address redirection... >> >> Thanks (and sorry for the long reply but I didn't know how >> to make the explanation shorter :) ) >> >> Gab >> >>> >>> If we separate the two steps: >>> >>> a) assign a range of logical I/O port numbers to a bus >>> b) register a set of helpers for redirecting logical I/O >>> port to a helper function >>> >>> then I think the code will get cleaner and more flexible. >>> It should actually then be able to replace the powerpc >>> specific implementation. >>> >>> Arnd >