On Tue, Feb 02, 2016 at 04:33:16PM -0800, Jared Hulbert wrote: > On Tue, Feb 2, 2016 at 3:41 PM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > > On Tue, Feb 2, 2016 at 3:36 PM, Jared Hulbert <jaredeh@xxxxxxxxx> wrote: > >> On Tue, Feb 2, 2016 at 3:19 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > >>> > >>> On Tue, Feb 02, 2016 at 04:11:42PM -0700, Ross Zwisler wrote: > >>> > >>> > However, for raw block devices and for XFS with a real-time device, the > >>> > value in inode->i_sb->s_bdev is not correct. With the code as it is > >>> > currently written, an fsync or msync to a DAX enabled raw block device will > >>> > cause a NULL pointer dereference kernel BUG. For this to work correctly we > >>> > need to ask the block device or filesystem what struct block_device is > >>> > appropriate for our inode. > >>> > > >>> > To that end, add a get_bdev(struct inode *) entry point to struct > >>> > super_operations. If this function pointer is non-NULL, this notifies DAX > >>> > that it needs to use it to look up the correct block_device. If > >>> > i_sb->get_bdev() is NULL DAX will default to inode->i_sb->s_bdev. > >>> > >>> Umm... It assumes that bdev will stay pinned for as long as inode is > >>> referenced, presumably? If so, that needs to be documented (and verified > >>> for existing fs instances). In principle, multi-disk fs might want to > >>> support things like "silently move the inodes backed by that disk to other > >>> ones"... > >> > >> Dan, This is exactly the kind of thing I'm taking about WRT the > >> weirder device models and directly calling bdev_direct_access(). > >> Filesystems don't have the monogamous relationship with a device that > >> is implicitly assumed in DAX, you have to ask the filesystem what the > >> relationship is and is migrating to, and allow the filesystem to > >> update DAX when the relationship is changing. > > > > That's precisely what ->get_bdev() does. When the answer > > inode->i_sb->s_bdev lookup is invalid, use ->get_bdev(). > > > >> As we start to see many > >> DIMM's and 10s TiB pmem systems this is going be an even bigger deal > >> as load balancing, wear leveling, and fault tolerance concerned are > >> inevitably driven by the filesystem. > > > > No, there are no plans on the horizon for an fs to manage these media > > specific concerns for persistent memory. > > So the filesystem is now directly in charge of mapping user pages to > physical memory. The filesystem is effectively bypassing NUMA and > zones and all that stuff that tries to balance memory bus and QPI > traffic etc. No, it's isn't bypassing NUMA, zones, etc. The pmem block device can linearise a typical NUMA layout quite sanely. i.e. if there is 10GB of pmem per node, the pmem device would need to map that as: node block device offsets 0 0..10GB 1 10..20GB 2 20..30GB .... n N..(N+1)GB i.e. present a *linear concatentation* of discrete nodes in a linear address space. Then, we can use the fact that XFS has a piecewise address space architecture that can map linear chunks of the block device address space to different logical domains. Each piece of an XFS filesystem is an allocation group. Hence we tell mkfs.xfs to set the allocation group size to 10GB, thereby mapping each individual allocation group to a different physical node of pmem. Suddenly all the filesystem allocation control algorithms become physical device locality control algorithms. Then we simply map process locality control (cpusets or memcgs or whatever is being used for that now) to the allocator - instead of selecting AGs for allocation based on inode/parent inode locality, we select AGs based on the allowed CPU/numa node mask of the process that is running... An even better architecture would be to present a pmem device per discrete node and then use DM to build the concat as required. Or enable us to stripe across nodes for higher performance in large concurrent applications, or configure RAID mirrors in physically separate parts of the NUMA topology for redundancy (e.g a water leak that physically destroys a rack doesn't cause data loss because the copies are in different racks (i.e. located in different failure domains)) then we can concat/stripe those mirrors together, etc. IOWs, we've already got all the pieces in place that we need to handle pmem in just about any way you can imagine in NUMA machines; the filesystem is just one of the pieces. This is just another example of how yet another new-fangled storage technology maps precisely to a well known, long serving storage architecture that we already have many, many experts out there that know to build reliable, performant storage from... :) > You don't think the filesystem will therefore be in > charge of memory bus hotspots? Filesystems and DM are already in charge of avoiding hotspots on disks, RAID arrays or in storage fabrics that can sustain tens of GB/s throughput. This really is a solved problem - pmem on NUMA systems is not very different to having tens of GB/s available on a multi-pathed SAN. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- 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