On Sun, Dec 09, 2018 at 04:57:43PM -0800, Nadav Amit wrote: > > On Dec 8, 2018, at 2:52 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > My patch proposed here: > > > > https://marc.info/?l=linux-mm&m=154409548410209 > > > > would actually fix that one I think, preempt_count() uses > > raw_cpu_read_4() which will loose the volatile with that patch. > I tested the patch you referenced, and it certainly improves the situation > for reads, but there are still small and big issues lying around. I'm sure :-(, this has been 'festering' for a long while it seems. And esp. on x86 specific code, where for a long time we all assumed the various per-cpu APIs were in fact the same (which turns out to very much not be true). > The biggest one is that (I think) smp_processor_id() should apparently use > __this_cpu_read(). Agreed, and note that this will also improve code generation on !x86. However, I'm not sure the current !debug definition: #define smp_processor_id() raw_smp_processor_id() is actually correct. Where raw_smp_processor_id() must be this_cpu_read() to avoid CSE, we actually want to allow CSE on smp_processor_id() etc.. > There are all kind of other smaller issues, such as set_irq_regs() and > get_irq_regs(), which should run with disabled interrupts. They affect the > generated code in do_IRQ() and others. > > But beyond that, there are so many places in the code that use > this_cpu_read() while IRQs are guaranteed to be disabled. For example > arch/x86/mm/tlb.c is full with this_cpu_read/write() and almost(?) all > should be running with interrupts disabled. Having said that, in my build > only flush_tlb_func_common() was affected. This all feels like something static analysis could help with; such tools would also make sense for !x86 where the difference between the various per-cpu accessors is even bigger.
