On 03/25/2015 11:29 AM, Dave Chinner wrote: > On Wed, Mar 25, 2015 at 10:10:31AM +0200, Boaz Harrosh wrote: >> On 03/25/2015 04:22 AM, Dave Chinner wrote: >>> On Tue, Mar 24, 2015 at 08:14:59AM +0200, Boaz Harrosh wrote: >> <> <> >> The sync does happen, .fsync of the FS is called on each >> file just as if the user called it. If this is broken it just >> needs to be fixed there at the .fsync vector. POSIX mandate >> persistence at .fsync so at the vfs layer we rely on that. > > right now, the filesystems will see that there are no dirty pages > on the inode, and then just sync the inode metadata. They will do > nothing else as filesystems are not aware of CPU cachelines at all. > Sigh yes. There is this bug. And I am sitting on a wide fix for this. The strategy is. All Kernel writes are done with a new copy_user_nt. NT stands for none-temporal. This shows 20% improvements since cachelines need not be fetched when written too. The arches that do not have NT instructions, will use a generic copy_user_nt that does a copy_user and then flush cashes. Same flush cashes we do before DMA IO. (effectively every 4k) [Its more complicated with the edges and all, by I have solved all this. Will post in a week or two] So what is left is the mmaped inodes. The logic here is that at .fsync vector dax inodes will do a cl_flush only if mapping_mapped() is true. Also .msync is the same as .fsync And one last thing we also call .fsync at vm_operations_struct->close because it is allowed for an app to do mmap, munmap, .fsync so we just call dax .fsync at munmap always. So by now we should be covered for fsync guaranty. >> So everything at this stage should be synced to real media. > > Actually no. This is what intel are introducing new CPU instructions > for - so fsync can flush the cpu caches and commit them to th > persistence domain correctly. > The new intel instructions are for an optimization, and they will fit in the picture for the CPUs that have it. But there are already NT instructions for existing CPUs. (Just not as fast and precise) Every ARCH will do its best under a small API copy_user_nt - data is at media memset_nt - data is at media cl_flush - partial written cachelines flushed to media sfence - New data seen by all CPUs >> What does not happen is writeback. since dax does not have >> any writeback. > > Which is precisely the problem we need to address - we don't need > writeback to a block device, but we do need the dirty CPU cachelines > flushed and the mappings cleaned. > I see what you mean. Since nothing dirtied the inode then above .fsync will not be called and we have not pushed mmap data to media. Again here we only need to do this for mmaped inodes, because Kernel written data is (will be) written NT style. >> And because of that nothing turned the >> user mappings to read only. This is what I do here but >> instead of write-protecting I just unmap because it is >> easier for me to code it. > > That doesn't mean it is the correct solution. Please note that even if we properly .fsync cachlines the page-faults are orthogonal to this. There is no point in making mmapped dax pages read-only after every .fsync and pay a page-fault. We should leave them mapped has is. The only place that we need page protection is at freeze time. But I see that we might have a problem with .fsync not being called. I see that you sent a second mail. I'll try to answer there. > Cheers, > Dave. Thanks Boaz -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
