From: Guennadi Liakhovetski <g.liakhovetski@gmx.de> To: Arnd Bergmann <arnd-r2nGTMty4D4@public.gmane.org> Cc: Russell King - ARM Linux <linux-lFZ/pmaqli7XmaaqVzeoHQ@public.gmane.org>, linux-sh-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-mmc-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, devicetree-discuss-uLR06cmDAlY/bJ5BZ2RsiQ@public.gmane.org, Magnus Damm <magnus.damm-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> Subject: Re: [PATCH v4 13/13] mmc: tmio: add barriers to IO operations Date: Tue, 19 Feb 2013 21:59:50 +0000 [thread overview] Message-ID: <Pine.LNX.4.64.1302192257240.22058@axis700.grange> (raw) In-Reply-To: <201302182211.46697.arnd-r2nGTMty4D4@public.gmane.org> On Mon, 18 Feb 2013, Arnd Bergmann wrote: > On Monday 18 February 2013, Guennadi Liakhovetski wrote: > > On Mon, 18 Feb 2013, Arnd Bergmann wrote: > > > > Sorry, I'm not sure how best to describe it, and I don't have sufficient > > information myself. In any case on a block-diagram of sh73a0 SDHI devices > > aren't directly connected to a common super-highway bus, instead they are > > on a bus-splitter. One more thing I forgot to mention - this error has > > been observed on SMP. > > Ah, I guess that explains it. If you have one CPU filling the buffer > and another CPU reading it, you need an smp_wmb/wmp_rmb pair. > > That case can easily happen with tmio interrupt handler in PIO > mode. Ok, I also managed to reproduce it on UP, and it definitely is related to MMC_CAP_POWER_OFF_CARD, but I'll have to do some more work on it. I asked Chris to hold back on this patch for now. Thanks Guennadi > > I checked clock code under drivers/sh/clk, I don't see any relevant > > __raw_* operations there. One I do see, however, GIC code does use > > *_relaxed() accessors, but under raw spinlocks... Do those provide > > sufficient barriers? > > A spinlock by definition contains an SMP barrier, which seems to be > what is missing here. I don't know if that is enough here, because > the spinlock only needs barriers to protect against other users > behind the same spinlock. > > In architecture independent code, any smp_wmb() on the writer side > must be paired with an smp_rmb on the reader. You probably have the > sufficient barriers on the side that initiates the I/O because > that uses a dmb() in the MMIO instructions, but that is only > by accident and not portable. > > Also, in the uniprocessor case, you have more barriers than you > need, which limits the performance. The best solution for > both performance and readability would be to use readw_relaxed() > and writew_relaxed() and manually add the barriers you actually > need, i.e. two smp_rmb()/smp_wmb() in the PIO case, and a > wmb() before you start a real DMA to the device and an rmb() > after a DMA from the device is complete. > > Arnd > --- Guennadi Liakhovetski, Ph.D. Freelance Open-Source Software Developer http://www.open-technology.de/
WARNING: multiple messages have this Message-ID (diff)
From: Guennadi Liakhovetski <g.liakhovetski-Mmb7MZpHnFY@public.gmane.org> To: Arnd Bergmann <arnd-r2nGTMty4D4@public.gmane.org> Cc: Russell King - ARM Linux <linux-lFZ/pmaqli7XmaaqVzeoHQ@public.gmane.org>, linux-sh-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-mmc-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, devicetree-discuss-uLR06cmDAlY/bJ5BZ2RsiQ@public.gmane.org, Magnus Damm <magnus.damm-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> Subject: Re: [PATCH v4 13/13] mmc: tmio: add barriers to IO operations Date: Tue, 19 Feb 2013 22:59:50 +0100 (CET) [thread overview] Message-ID: <Pine.LNX.4.64.1302192257240.22058@axis700.grange> (raw) In-Reply-To: <201302182211.46697.arnd-r2nGTMty4D4@public.gmane.org> On Mon, 18 Feb 2013, Arnd Bergmann wrote: > On Monday 18 February 2013, Guennadi Liakhovetski wrote: > > On Mon, 18 Feb 2013, Arnd Bergmann wrote: > > > > Sorry, I'm not sure how best to describe it, and I don't have sufficient > > information myself. In any case on a block-diagram of sh73a0 SDHI devices > > aren't directly connected to a common super-highway bus, instead they are > > on a bus-splitter. One more thing I forgot to mention - this error has > > been observed on SMP. > > Ah, I guess that explains it. If you have one CPU filling the buffer > and another CPU reading it, you need an smp_wmb/wmp_rmb pair. > > That case can easily happen with tmio interrupt handler in PIO > mode. Ok, I also managed to reproduce it on UP, and it definitely is related to MMC_CAP_POWER_OFF_CARD, but I'll have to do some more work on it. I asked Chris to hold back on this patch for now. Thanks Guennadi > > I checked clock code under drivers/sh/clk, I don't see any relevant > > __raw_* operations there. One I do see, however, GIC code does use > > *_relaxed() accessors, but under raw spinlocks... Do those provide > > sufficient barriers? > > A spinlock by definition contains an SMP barrier, which seems to be > what is missing here. I don't know if that is enough here, because > the spinlock only needs barriers to protect against other users > behind the same spinlock. > > In architecture independent code, any smp_wmb() on the writer side > must be paired with an smp_rmb on the reader. You probably have the > sufficient barriers on the side that initiates the I/O because > that uses a dmb() in the MMIO instructions, but that is only > by accident and not portable. > > Also, in the uniprocessor case, you have more barriers than you > need, which limits the performance. The best solution for > both performance and readability would be to use readw_relaxed() > and writew_relaxed() and manually add the barriers you actually > need, i.e. two smp_rmb()/smp_wmb() in the PIO case, and a > wmb() before you start a real DMA to the device and an rmb() > after a DMA from the device is complete. > > Arnd > --- Guennadi Liakhovetski, Ph.D. Freelance Open-Source Software Developer http://www.open-technology.de/
next prev parent reply other threads:[~2013-02-19 21:59 UTC|newest] Thread overview: 71+ messages / expand[flat|nested] mbox.gz Atom feed top 2013-02-15 15:13 [PATCH v4 00/13] mmc: core and driver DT and related development Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 01/13] mmc: sdhi, tmio: only check flags in tmio-mmc driver proper Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 02/13] mmc: detailed definition of CD and WP MMC line polarities in DT Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 03/13] mmc: provide a standard MMC device-tree binding parser centrally Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-16 6:05 ` Simon Horman 2013-02-16 6:05 ` Simon Horman 2013-02-16 10:54 ` Arnd Bergmann 2013-02-16 11:41 ` Simon Horman 2013-02-16 11:41 ` Simon Horman 2013-02-16 14:02 ` Guennadi Liakhovetski 2013-02-16 14:02 ` Guennadi Liakhovetski 2013-02-16 15:21 ` [PATCH v5 " Guennadi Liakhovetski 2013-02-16 15:21 ` Guennadi Liakhovetski 2013-02-16 16:58 ` Sascha Hauer 2013-02-16 16:58 ` Sascha Hauer 2013-02-17 7:52 ` Simon Horman 2013-02-17 7:52 ` Simon Horman 2013-02-17 7:58 ` Simon Horman 2013-02-17 7:58 ` Simon Horman 2013-02-18 8:54 ` Guennadi Liakhovetski 2013-02-18 8:54 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 04/13] mmc: (cosmetic) remove "extern" from function declarations Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 05/13] mmc: sh-mmcif: use mmc_of_parse() to parse standard MMC DT bindings Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 06/13] mmc: tmio-mmc: define device-tree bindings Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 07/13] mmc: tmio-mmc: parse " Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 08/13] mmc: sh_mobile_sdhi: remove unused .pdata field Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 09/13] mmc: sh_mobile_sdhi: use managed resource allocations Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:13 ` [PATCH v4 10/13] mmc: tmio: remove unused and deprecated symbols Guennadi Liakhovetski 2013-02-15 15:13 ` Guennadi Liakhovetski 2013-02-15 15:14 ` [PATCH v4 11/13] mmc: tmio: add support for the VccQ regulator Guennadi Liakhovetski 2013-02-15 15:14 ` Guennadi Liakhovetski 2013-02-15 15:14 ` [PATCH v4 12/13] mmc: add DT bindings for more MMC capability flags Guennadi Liakhovetski 2013-02-15 15:14 ` Guennadi Liakhovetski 2013-02-16 22:58 ` Sergei Shtylyov 2013-02-16 23:58 ` Sergei Shtylyov [not found] ` <51201D32.20006-M4DtvfQ/ZS1MRgGoP+s0PdBPR1lH4CV8@public.gmane.org> 2013-02-18 8:52 ` Guennadi Liakhovetski 2013-02-18 8:52 ` Guennadi Liakhovetski 2013-06-06 1:55 ` Olof Johansson 2013-06-06 1:55 ` Olof Johansson 2013-02-15 15:14 ` [PATCH v4 13/13] mmc: tmio: add barriers to IO operations Guennadi Liakhovetski 2013-02-15 15:14 ` Guennadi Liakhovetski 2013-02-18 15:05 ` Arnd Bergmann 2013-02-18 15:56 ` Guennadi Liakhovetski 2013-02-18 15:56 ` Guennadi Liakhovetski 2013-02-18 16:34 ` Arnd Bergmann 2013-02-18 17:20 ` Guennadi Liakhovetski 2013-02-18 17:20 ` Guennadi Liakhovetski 2013-02-18 22:11 ` Arnd Bergmann [not found] ` <201302182211.46697.arnd-r2nGTMty4D4@public.gmane.org> 2013-02-19 21:59 ` Guennadi Liakhovetski [this message] 2013-02-19 21:59 ` Guennadi Liakhovetski 2013-02-16 2:26 ` [PATCH v4 00/13] mmc: core and driver DT and related development Simon Horman 2013-02-16 2:26 ` Simon Horman 2013-02-18 13:15 ` Chris Ball 2013-02-18 13:15 ` Chris Ball 2013-02-19 21:57 ` Guennadi Liakhovetski 2013-02-19 21:57 ` Guennadi Liakhovetski 2013-02-19 22:00 ` Chris Ball 2013-02-19 22:00 ` Chris Ball 2013-02-18 14:52 ` Arnd Bergmann 2013-02-19 19:20 ` Stephen Warren 2013-02-19 19:20 ` Stephen Warren
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=Pine.LNX.4.64.1302192257240.22058@axis700.grange \ --to=g.liakhovetski@gmx.de \ --cc=arnd-r2nGTMty4D4@public.gmane.org \ --cc=devicetree-discuss-uLR06cmDAlY/bJ5BZ2RsiQ@public.gmane.org \ --cc=horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org \ --cc=linux-lFZ/pmaqli7XmaaqVzeoHQ@public.gmane.org \ --cc=linux-mmc-u79uwXL29TY76Z2rM5mHXA@public.gmane.org \ --cc=linux-sh-u79uwXL29TY76Z2rM5mHXA@public.gmane.org \ --cc=magnus.damm-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.