On Tue, Aug 11, 2020 at 06:41:20PM +0200, Tomasz Figa wrote: > 2020年8月11日(火) 18:34 Krzysztof Kozlowski <krzk@xxxxxxxxxx>: > > > > On Tue, Aug 11, 2020 at 06:28:18PM +0200, Tomasz Figa wrote: > > > 2020年8月11日(火) 18:24 Krzysztof Kozlowski <krzk@xxxxxxxxxx>: > > > > > > > > On Tue, Aug 11, 2020 at 02:59:07PM +0200, Tomasz Figa wrote: > > > > > Hi Sylwester, > > > > > > > > > > 2020年8月11日(火) 13:25 Sylwester Nawrocki <s.nawrocki@xxxxxxxxxxx>: > > > > > > > > > > > > In the .set_rate callback for some PLLs there is a loop polling state > > > > > > of the PLL lock bit and it may become an endless loop when something > > > > > > goes wrong with the PLL. For some PLLs there is already (a duplicated) > > > > > > code for polling with timeout. This patch replaces that code with > > > > > > the readl_relaxed_poll_timeout_atomic() macro and moves it to a common > > > > > > helper function, which is then used for all the PLLs. The downside > > > > > > of switching to the common macro is that we drop the cpu_relax() call. > > > > > > > > > > Tbh. I'm not sure what effect was exactly expected from cpu_relax() in > > > > > the functions which already had timeout handling. Could someone shed > > > > > some light on this? > > > > > > > > For us, it should not matter much, except: > > > > 1. when on A9 with ARM_ERRATA_754327, but we do not enable it on our > > > > platforms, > > > > 2. it is a generic pattern for busy loops. > > > > > > > > On other architectures it could mean something (e.g. yield to other > > > > hyper-threading CPU). > > > > > > Okay, thanks for confirming that it doesn't matter for us. > > > > > > Now, I wonder if the readx_poll_*() helpers are supposed to take all > > > of those into account or on systems which would benefit from such > > > operations, it would be the caller's responsibility. > > > > That's a very good point. In case of ARM_ERRATA_754327, busy waiting > > should have a barrier thus cpu_relax() is desired. I guess the generic > > macro for busy waiting therefore should use them. > > Is there yet another macro available somewhere or you mean > read_poll_timeout_atomic()? The latter doesn't include cpu_relax(). Yes, I meant the generic read_poll_timeout_atomic(). > Given that udelay() likely already does this kind of an idle call, > perhaps it could be as simple as this? > > if (__delay_us) \ > udelay(__delay_us); \ > + else \ > + cpu_relax(); \ > I think udelay does not have it. Delaying by some simple operations (e.g. multiplication) does not require IO barriers. > On the other hand, I wonder if there are cases where a call to > cpu_relax() is not desirable. Hmmm, it is really a generic pattern all over the kernel, so I doubt that generic macros should target such case. Best regards, Krzysztof
