From mboxrd@z Thu Jan 1 00:00:00 1970 From: Sergei Shtylyov Subject: Re: [PATCH] (pata-2.6 fix queue) cmd64x: remove broken SW/MW DMA support Date: Fri, 16 Feb 2007 18:00:22 +0300 Message-ID: <45D5C706.6060303@ru.mvista.com> References: <200702080858.l188wjmG008432@harpo.it.uu.se> <17877.33048.23816.692749@alkaid.it.uu.se> <45D5AD8B.1060702@ru.mvista.com> <200702161537.37691.bzolnier@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii; format=flowed Content-Transfer-Encoding: 7bit Return-path: Received: from h155.mvista.com ([63.81.120.155]:51225 "EHLO imap.sh.mvista.com" rhost-flags-OK-FAIL-OK-FAIL) by vger.kernel.org with ESMTP id S932377AbXBPPAa (ORCPT ); Fri, 16 Feb 2007 10:00:30 -0500 In-Reply-To: <200702161537.37691.bzolnier@gmail.com> Sender: linux-ide-owner@vger.kernel.org List-Id: linux-ide@vger.kernel.org To: Bartlomiej Zolnierkiewicz Cc: Mikael Pettersson , linux-ide@vger.kernel.org Hello. Bartlomiej Zolnierkiewicz wrote: >>Mikael Pettersson wrote: >>>>>Remove the bogus code pretending to set SW/MW DMA timings -- I wonder whether >>>>>its author really thought that he could achieve that wrtiting to BMIDE status >>>>>registers? Stop fiddling with the DMA capable bits in the speedproc() -- they >>>>>do not enable DMA, and are properly dealt with by the dma_host_{on,off} methods; >>>>>also, get rid of the duplicate reads/writes of UDIDETCRx registers, and do some >>>>>coding style and whitespace changes while at it... >>>>>Unfortunately, fixing the SW/MW DMA support would requre a major driver rewrite >>>>>along with some more fixing, so I'm putting it off... >>>>>Warning: this has been compile-tested only. >>>>Worked fine on my SPARC Ultra5. >>>Correction: I was only looking for absence of errors when testing >>>this patch. However, later I found that this patch (version 1.42 >>>of cmd64x.c) disabled DMA on my CMD646, dropping performance to >>>1/4th (from about 13MB/s to about 3.5MB/s according to hdparm -Tt). >> That was expected behavior. >>>Here's the relevant kernel messages from before this patch: >>>ide: Assuming 33MHz system bus speed for PIO modes; override with idebus=xx >>>CMD646: IDE controller at PCI slot 0000:01:03.0 >>>CMD646: chipset revision 3 >>>CMD646: chipset revision 0x03, MultiWord DMA Force Limited >> The driver never supported UltraDMA on this revision, as indicated by that >>message. >>>CMD646: 100% native mode on irq 14 >>> ide0: BM-DMA at 0x1fe02c00020-0x1fe02c00027, BIOS settings: hda:pio, hdb:pio >>> ide1: BM-DMA at 0x1fe02c00028-0x1fe02c0002f, BIOS settings: hdc:pio, hdd:pio >>>Probing IDE interface ide0... >>>hda: ST320420A, ATA DISK drive >>>ide0 at 0x1fe02c00000-0x1fe02c00007,0x1fe02c0000a on irq 14 >>>Probing IDE interface ide1... >>>hdc: CRD-8483B, ATAPI CD/DVD-ROM drive >>>ide1 at 0x1fe02c00010-0x1fe02c00017,0x1fe02c0001a on irq 14 (shared with ide0) >>>hda: max request size: 128KiB >>>hda: 39851760 sectors (20404 MB) w/2048KiB Cache, CHS=39535/16/63, (U)DMA >>>hda: cache flushes not supported >>> hda: hda1 hda2 hda3 hda4 hda5 >>>With the patch the kernel messages are the same, except for the >>>3rd last line which becomes: >>>hda: 39851760 sectors (20404 MB) w/2048KiB Cache, CHS=39535/16/63 >>>i.e., the (U)DMA indicator is gone. >>>Please revert this until the regression is fixed. >> The intent of the patch was exactly to *remove* broken DMA support until >>it's fixed (which requires more work). It only worked by chance -- because >>MWDMA2 timings are the same as of PIO4. Have patience please. > It only worked by chance but it _worked_, especially for the usual case, > MWDMA2/PIO4 == all newer drives (despite writing 0x10 reserved bit of > BMIDESR0/1 for the master devices). I think I'll remove the patch for now. Then I will protest. :-) This was *not* fixable all at once. And removing it just because some users happen to have the *known buggy* (and so reduced to MWDMA only) chips is a wrong thing to do. > To fix SWDMA/MWDMA properly isn't it enough to just call cmd64x_tune_pio() > from cmd64x_tune_chipset() to tune the corresponding PIO mode? No, that wouldn't be a clean fix, just a kind of workaround -- it will also change the address setup times (which is not quite desirable). I can compose such quickie in five minutes though. > Bart WBR, Sergei