| Message ID | 1467713872-4051-3-git-send-email-p.zabel@pengutronix.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tuesday, July 5, 2016 12:17:52 PM CEST Philipp Zabel wrote: > This just removes the rmb()/wmb() pair between register read and > write. Since no relevant reads follow the rmb and no relevant writes > precede the wmb, they should be safe to remove. > > Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> We should only do this if you are fixing a bug (which you don't mention in the changelog), or if you can show a relevant performance improvement. Is this code ever used in a fast path? If it is, wouldn't that indicate a problem in some driver? Arnd
Am Dienstag, den 05.07.2016, 13:20 +0200 schrieb Arnd Bergmann: > On Tuesday, July 5, 2016 12:17:52 PM CEST Philipp Zabel wrote: > > This just removes the rmb()/wmb() pair between register read and > > write. Since no relevant reads follow the rmb and no relevant writes > > precede the wmb, they should be safe to remove. > > > > Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> > > We should only do this if you are fixing a bug (which you don't mention > in the changelog), or if you can show a relevant performance > improvement. Is this code ever used in a fast path? If it is, > wouldn't that indicate a problem in some driver? It does not fix a bug, and it's not about performance either. I'd like to align code with the recently posted stm32 driver, to unify them in a future patch. Of course we can change the stm32 driver to use readl/writel instead of the relaxed variants, it just seemed useless to have those barriers between the read and write. If anything, we'd need to try to make sure that the writel in assert hits the hardware before the function returns, so that a assert-delay-deassert doesn't accidentally spend half its delay with the writel still in the store buffer, and we'd need a full barrier after the writel in deassert so that there can be no successive reads from still disabled IP cores. regards Philipp
On Tuesday, July 5, 2016 1:40:16 PM CEST Philipp Zabel wrote: > Am Dienstag, den 05.07.2016, 13:20 +0200 schrieb Arnd Bergmann: > > On Tuesday, July 5, 2016 12:17:52 PM CEST Philipp Zabel wrote: > > > This just removes the rmb()/wmb() pair between register read and > > > write. Since no relevant reads follow the rmb and no relevant writes > > > precede the wmb, they should be safe to remove. > > > > > > Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> > > > > We should only do this if you are fixing a bug (which you don't mention > > in the changelog), or if you can show a relevant performance > > improvement. Is this code ever used in a fast path? If it is, > > wouldn't that indicate a problem in some driver? > > It does not fix a bug, and it's not about performance either. I'd like > to align code with the recently posted stm32 driver, to unify them in a > future patch. > Of course we can change the stm32 driver to use readl/writel instead of > the relaxed variants, it just seemed useless to have those barriers > between the read and write. On stm32, there is no barrier because ARM_DMA_MEM_BUFFERABLE is not set. I'd really prefer to just have readl/writel everywhere except in the few places that are performance critical and have a comment explaining why it's safe there, mainly to avoid having new developers blindly add the relaxed accessors in drivers because they think it's the normal coding style. > If anything, we'd need to try to make sure that the writel in assert > hits the hardware before the function returns, so that a > assert-delay-deassert doesn't accidentally spend half its delay with the > writel still in the store buffer, and we'd need a full barrier after the > writel in deassert so that there can be no successive reads from still > disabled IP cores. In general, I think you need a readl() following the writel() to guarantee that it has actually hit the hardware. On ARM you often have just the CPU write buffer that needs to be flushed, but if you have a PCI device or a more complex SoC, then a barriers doesn't wait for a write to arrive at the device, it only ensures that a subsequent write cannot arrive any earlier. Arnd
Am Dienstag, den 05.07.2016, 14:59 +0200 schrieb Arnd Bergmann: > On Tuesday, July 5, 2016 1:40:16 PM CEST Philipp Zabel wrote: > > Am Dienstag, den 05.07.2016, 13:20 +0200 schrieb Arnd Bergmann: > > > On Tuesday, July 5, 2016 12:17:52 PM CEST Philipp Zabel wrote: > > > > This just removes the rmb()/wmb() pair between register read and > > > > write. Since no relevant reads follow the rmb and no relevant writes > > > > precede the wmb, they should be safe to remove. > > > > > > > > Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> > > > > > > We should only do this if you are fixing a bug (which you don't mention > > > in the changelog), or if you can show a relevant performance > > > improvement. Is this code ever used in a fast path? If it is, > > > wouldn't that indicate a problem in some driver? > > > > It does not fix a bug, and it's not about performance either. I'd like > > to align code with the recently posted stm32 driver, to unify them in a > > future patch. > > Of course we can change the stm32 driver to use readl/writel instead of > > the relaxed variants, it just seemed useless to have those barriers > > between the read and write. > > On stm32, there is no barrier because ARM_DMA_MEM_BUFFERABLE is not set. > > I'd really prefer to just have readl/writel everywhere except in the > few places that are performance critical and have a comment explaining > why it's safe there, mainly to avoid having new developers blindly > add the relaxed accessors in drivers because they think it's the > normal coding style. I get your point. I'll ask the stm32 developers to use non-relaxed readl/writel then. > > If anything, we'd need to try to make sure that the writel in assert > > hits the hardware before the function returns, so that a > > assert-delay-deassert doesn't accidentally spend half its delay with the > > writel still in the store buffer, and we'd need a full barrier after the > > writel in deassert so that there can be no successive reads from still > > disabled IP cores. > > In general, I think you need a readl() following the writel() to guarantee > that it has actually hit the hardware. > > On ARM you often have just the CPU write buffer that needs to be flushed, > but if you have a PCI device or a more complex SoC, then a barriers doesn't > wait for a write to arrive at the device, it only ensures that a subsequent > write cannot arrive any earlier. Yes, exactly. Until now I have not considered this at all. regards Philipp
diff --git a/drivers/reset/reset-socfpga.c b/drivers/reset/reset-socfpga.c index 78ebf84..4a65128 100644 --- a/drivers/reset/reset-socfpga.c +++ b/drivers/reset/reset-socfpga.c @@ -44,8 +44,8 @@ static int socfpga_reset_assert(struct reset_controller_dev *rcdev, spin_lock_irqsave(&data->lock, flags); - reg = readl(data->membase + (bank * NR_BANKS)); - writel(reg | BIT(offset), data->membase + (bank * NR_BANKS)); + reg = readl_relaxed(data->membase + (bank * NR_BANKS)); + writel_relaxed(reg | BIT(offset), data->membase + (bank * NR_BANKS)); spin_unlock_irqrestore(&data->lock, flags); return 0; @@ -65,8 +65,8 @@ static int socfpga_reset_deassert(struct reset_controller_dev *rcdev, spin_lock_irqsave(&data->lock, flags); - reg = readl(data->membase + (bank * NR_BANKS)); - writel(reg & ~BIT(offset), data->membase + (bank * NR_BANKS)); + reg = readl_relaxed(data->membase + (bank * NR_BANKS)); + writel_relaxed(reg & ~BIT(offset), data->membase + (bank * NR_BANKS)); spin_unlock_irqrestore(&data->lock, flags);
This just removes the rmb()/wmb() pair between register read and write. Since no relevant reads follow the rmb and no relevant writes precede the wmb, they should be safe to remove. Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de> --- drivers/reset/reset-socfpga.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)