On Fri, August 30, 2013, Doug Anderson wrote: > Seungwon, > > On Thu, Aug 29, 2013 at 12:04 AM, Seungwon Jeon <tgih.jun@xxxxxxxxxxx> wrote: > >> I'd really still rather honor the MMC subsystem's request. It > >> shouldn't _hurt_ to turn the clock off when the subsystem requests it, > > Even though turning off by clock programming doesn't hurt, > > it is costly behavior when considering low power mode of host's own support. > > It is costly? We are talking about these two commands, right? > > mci_writel(host, CLKENA, 0); > mci_send_cmd(slot, > SDMMC_CMD_UPD_CLK | SDMMC_CMD_PRV_DAT_WAIT, 0); I mean that because host supports auto clock gating, we don't need to clock programming with setting '0'. If MMC_CLKGATE is enabled, clock programming will be executed with between '0' clock and working clock frequently. Actually the result is same. Of course, if host didn't support this feature, we would have considered that manually to save the power consumption. > > Do you have a reason to believe that these are more costly than all of > the rest of the code that's executed when the user defines > CONFIG_MMC_CLKGATE? You're still proposing doing all of the updates > of the clock when slot->clock is non-zero, right? ...so at best No, origin condition should be remained; Required clock should be different with current clock. > skipping this code will be 33% faster since the re-enable code > disables and then reenables the clock. If it's the > "SDMMC_CMD_PRV_DAT_WAIT" that you're worried about then skipping this > code will only be 25% faster since there are already three calls with > SDMMC_CMD_PRV_DAT_WAIT in the enable code. > > > > Just now, how about focusing on the problem clock isn't updated properly after suspend/resume? > > I tried to do that in the original patches, but you pointed out > (correctly) that we should do the correct fix rather than a hackier > fix. IMHO the most correct fix is to honor the MMC core's request to > turn the clock off. Partially honoring the MMC core (as you suggest) > is certainly less hacky that my original proposal but I still think > turning the clock off is better. > > > >> right? One reason to honor the mmc core is that it will make things > >> cleaner if/when we support a voltage change operation. The MMC core > >> has the logic for the voltage change, and part of that involves > >> turning off the clock. We'll already need a bunch of special case > >> code in dw_mmc for voltage change, but it would be nice to avoid one > >> extra bit. > > Turning off clock during voltage switching would be another procedure. > > I guess it could be discussed later. > > Agreed that we're not trying to get voltage switching done here, but > forward thinking is nice. If there's no reason _not_ to turn the > clock off and it will help us later, let's do it. Also, we've already > agreed that MMC_CLKGATE isn't so useful for dw_mmc, so trying to do > something awkward to make MMC_CLKGATE slightly faster doesn't seem > worth it. > > > > I want to fix some minor change to prevent frequent message that Jaehoon pointed. > > As far as I can tell, the frequent messages and whether or not to > actually turn the clock off are unrelated. I will send up a patch > that fixes the frequent messages by caching the last value printed and > only printing if it changed. I have verified that this works and that > the system still functions OK (can boot to prompt) with > CONFIG_MMC_CLKGATE. > > > Note: re-reading over some of the previous messages, it sounds like > you're proposing using the patch from your email directly, AKA: > > http://article.gmane.org/gmane.linux.kernel/1542482 > > Did you test that patch? Did it work for you? It doesn't actually > compile cleanly for me (you removed the "force_clkinit" param in the > function but not the callers). That's easy to fix, but implies that > this patch was just a proposal and not a tested solution. > > ...but aside from not compiling cleanly, I don't think it will work > for the same reasons that the original code didn't work. Specifically > it doesn't address the core problem that we need to update > host->current_speed when the clock is 0. Otherwise we won't re-init > and we run into the original problem, right? To be certain I took > your patch and applied it, then fixed the callers of > dw_mci_setup_bus() not to pass a second parameter. I did a > suspend/resume with no card in and then plugged a card in. It didn't > work. Some change proposed from me are mixed with both current existing part and your new patch. It's not whole code to replace your patch. But if it makes you confused, sorry about that. Important thing I intended is that if required clock(slot->clock) is '0', not try to update clock. with considering automatic clock gating. Please check first comment from me on v6. Thanks, Seungwon Jeon > > > As I said above, new patch coming shortly. As always: feel free to > point out any glaring mistakes I made in the above. ;) Note that I > will be out of communication for the next week or so and buried > beneath email for a few days after that, so my response might be a > little delayed. > > > > -Doug > -- > To unsubscribe from this list: send the line "unsubscribe linux-mmc" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-mmc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
