Some kernel hackers believe that the code responsible for write barriers has been overly restrictive in various places. A major restructuring effort is to put an end to this and transfer the main responsibility for determining the order of write operations to the file system code, a solution said to considerably improve data throughput with certain tasks. Further background information is available in the article "The end of block barriers" on LWN.net, in a discussion on the LKML, and in the comments to some of the relevant commits (1, 2, 3, 4).
Control groups (cgroups) now allow the maximum throughput of storage devices to be limited / restricted; details are available in various commit comments (1, 2, 3, 4) and in the updated documentation. Several changes to the CFQ I/O scheduler are designed to improve the scheduler's performance when processing an fsync ("file synchronize") for small files. The code for mounting the root partition can now identify the correct partition via a UUID (Universally Unique Identifier), provided the kernel has been instructed accordingly via a parameter such as "root=PARTUUID=hex-uuid".
The Libata subsystem now cooperates with hard disks which use logical, as well as physical, sectors that are not 512 bytes in size; the developer has tested the code with a native 4K "Engineering Sample" drive provided by Hitachi GST. The new Libata transport class now allows the kernel code for ATA adapters to export much more ATA information via Sysfs. The kernel now also supports Intel IDE-R devices that are part of the Intel AMT (Active Management Technology) set up; to fix various problems, the Libata developers have re-implemented the link power management code.
In the SCSI subsystem, and in the Libata code based on it, the developers have made various major changes that were only integrated with the third RC of 2.6.37 after being discussed on the LKML. These changes restructure the locking mechanisms to speed up certain drivers (for instance 1, 2).
The new cxgb4i driver offers iSCSI connection acceleration for T4 products by Chelsio. The InfiniBand stack and the mlx4_ib driver now also support InfiniBand-over-Ethernet / IBoE (for instance 1, 2). While the technology is actually called "RDMA over Converged Ethernet" (RoCE), the maintainer of the InfiniBand code explains in two related blog postings why he prefers "InfiniBand over Ethernet" (1, 2).