Hi Hucai, On Wed, Jul 30, 2014 at 04:35:26PM +0800, Huacai Chen wrote: > Hi, Aurelien, > > After some days debugging, we found the root cause: If we revert the > commit 21b40200cfe961 (aio: use flush_dcache_page()), everything is > OK. This commit add two flush_dcache_page() in irq disabled context, > but in MIPS, flush_dcache_page() is implemented via call_function IPI. > Unfortunately, call_function IPI shouldn't be called in irq disabled > context, otherwise there will be deadlock. Thanks a lot for digging into the problem. I will try to revert this patch to confirm it fixes the problem for us, and I'll keep you updated. > I don't know how to solve this problem, since commit 21b40200 > shouldn't be reverted (Loongson can revert it because of > hardware-maintained cache, but other MIPS need this). May be the > original author (Kent Overstreet) have good ideas? Maybe we should look if it's possible to reduce the window where interrupts are disabled in this function, but I guess we'll have to wait for Kent about that. Do we know if other MIPS systems or other architectures also implement the flush_dcache_page() function via IPI? Thanks, Aurelien -- Aurelien Jarno GPG: 4096R/1DDD8C9B aurelien@xxxxxxxxxxx http://www.aurel32.net
