Re: [PATCH v5 2/2] x86/mm: Improve TLB flush documentation

[Nadav Amit]

Andy Lutomirski <luto@xxxxxxxxxx> wrote:

> Improve comments as requested by PeterZ and also add some
> documentation at the top of the file.
> 
> Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
> ---
> arch/x86/mm/tlb.c | 43 +++++++++++++++++++++++++++++++++----------
> 1 file changed, 33 insertions(+), 10 deletions(-)
> 
> diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
> index ce104b962a17..d4ee781ca656 100644
> --- a/arch/x86/mm/tlb.c
> +++ b/arch/x86/mm/tlb.c
> @@ -15,17 +15,24 @@
> #include <linux/debugfs.h>
> 
> /*
> - *	TLB flushing, formerly SMP-only
> - *		c/o Linus Torvalds.
> + * The code in this file handles mm switches and TLB flushes.
>  *
> - *	These mean you can really definitely utterly forget about
> - *	writing to user space from interrupts. (Its not allowed anyway).
> + * An mm's TLB state is logically represented by a totally ordered sequence
> + * of TLB flushes.  Each flush increments the mm's tlb_gen.
>  *
> - *	Optimizations Manfred Spraul <manfred@xxxxxxxxxxxxxxxx>
> + * Each CPU that might have an mm in its TLB (and that might ever use
> + * those TLB entries) will have an entry for it in its cpu_tlbstate.ctxs
> + * array.  The kernel maintains the following invariant: for each CPU and
> + * for each mm in its cpu_tlbstate.ctxs array, the CPU has performed all
> + * flushes in that mms history up to the tlb_gen in cpu_tlbstate.ctxs
> + * or the CPU has performed an equivalent set of flushes.
>  *
> - *	More scalable flush, from Andi Kleen
> - *
> - *	Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
> + * For this purpose, an equivalent set is a set that is at least as strong.
> + * So, for example, if the flush history is a full flush at time 1,
> + * a full flush after time 1 is sufficient, but a full flush before time 1
> + * is not.  Similarly, any number of flushes can be replaced by a single
> + * full flush so long as that replacement flush is after all the flushes
> + * that it's replacing.
>  */
> 
> atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1);
> @@ -138,7 +145,16 @@ void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
> 			return;
> 		}
> 
> -		/* Resume remote flushes and then read tlb_gen. */
> +		/*
> +		 * Resume remote flushes and then read tlb_gen.  The
> +		 * implied barrier in atomic64_read() synchronizes
> +		 * with inc_mm_tlb_gen() like this:

You mean the implied memory barrier in cpumask_set_cpu(), no?