On Mon, Jun 09, 2014 at 02:24:29PM +0100, Leif Lindholm wrote: > On Mon, Jun 09, 2014 at 12:03:56PM +0100, Catalin Marinas wrote: > > So I'm proposing an alternative patch (which needs some benchmarking as > > well to see if anything is affected, maybe application startup time).
I don't like the proposed patch at all -- keeping the dsb out of set_pte is worthwhile if we can manage it. That said, it would be interesting to know how often we get a subsequent page fault after mapping invalid -> valid because of the missing dsb. It could be that the cost of the benign fault is hitting us more than we think. > I'm happy for any fix which can be included in 3.16. > > But is the dsb(ishst) sufficient? We need to also prevent reads from > overtaking the set_pte(). i.e.: > > ptr = early_ioremap(phys_addr, size); > if (ptr && strcmp(ptr, "magic") == 0) > ... > > Does it not require a dsb(ish)? I don't think so. Crudely, the sequence above would look like: STR x0, [PTEP] DSB ISHST LDR x0, [MAGIC] The DSB can't complete until the STR is globally observed within the inner-shareable domain, so the LDR cannot execute until the page table update is visible to the walker. If it was a DMB, we'd have a problem. Interestingly, the asm-generic page table allocators (e.g. __pmd_alloc) *do* use dmb for ordering observability of page-table updates via smp_wmb. I'm struggling to decide whether that's broken or not. Will -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
