Now that common suspend helpers are in linux-next, I'm going back to trying to use them with the L2 cache on, which is required when implementing cpuidle states that put a cpu in reset in an SMP system. There are a few obvious requirements - the stack pointer, stack contents, and page tables must be flushed in the L2 during suspend. There is one trickier problem. The current resume sequence starts from reset, with cache and MMU off: Restore cpu registers, including TTBR0, but not SCTLR Rewrite the page table to map the MMU trampoline VA=PA Restore SCTLR with dcache bit off (enables MMU) Jump to virtual address Restore SCTLR with original dcache bit The page table is rewritten with the temporary mapping with the cache off, so the write goes directly to memory. When the MMU is enabled, the TTBR0 RGN and IRGN bits determine the cachability attributes of the page table walk memory accesses. These bits are normally set to cachable, so when the MMU is turned on, the page table walks can hit the L2 instead of memory. If a stale entry for the temporary mapping is in the L2, which can happen if it is prefetched by the other cpu during suspend, the page table walk will not see the new mapping. This could be fixed by restoring the TTBR0 register with caching disabled, and re-enabling caching after the page table entry is restored, or by switching to a custom page table prepared during init to avoid any writes while the cache is off. -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
