[An update before the Kernel Summit because of the numerous requests that I have had for this patchset. Please speak up if you feel that we need something like this.] This patchset modifies the Linux kernel so that larger block sizes than page size can be supported. Larger block sizes are handled by using compound pages of an arbitrary order for the page cache instead of single pages with order 0. Support is added in a way that limits the changes to existing code. As a result filesystems can support I/O using large buffers with minimal changes. The page cache functions are mostly unchanged. Instead of a page struct representing a single page they take a head page struct (which looks the same as a regular page struct apart from the compound flags) and operate on those. Most page cache functions can stay as they are. No locking protocols are added or modified. The support is also fully transparent at the level of the OS. No specialized heuristics are added to switch to larger pages. Large page support is enabled by filesystems or device drivers when a device or volume is mounted. Larger block sizes are usually set during volume creation although the patchset supports setting these sizes per file. The formattted partition will then always be accessed with the configured blocksize. Some of the changes are: - Replace the use of PAGE_CACHE_XXX constants to calculate offsets into pages with functions that do the the same and allow the constants to be parameterized. - Extend the capabilities of compound pages so that they can be put onto the LRU and reclaimed. - Allow setting a larger blocksize via set_blocksize() Rationales: ----------- 1. The ability to handle memory of an arbitrarily large size using a singe page struct "handle" is essential for scaling memory handling and reducing overhead in multiple kernel subsystems. This patchset is a strategic move that allows performance gains throughout the kernel. 2. Reduce fsck times. Larger block sizes mean faster file system checking. Using 64k block size will reduce the number of blocks to be managed by a factor of 16 and produce much denser and contiguous metadata. 3. Performance. If we look at IA64 vs. x86_64 then it seems that the faster interrupt handling on x86_64 compensate for the speed loss due to a smaller page size (4k vs 16k on IA64). Supporting larger block sizes sizes on all allows a significant reduction in I/O overhead and increases the size of I/O that can be performed by hardware in a single request since the number of scatter gather entries are typically limited for one request. This is going to become increasingly important to support the ever growing memory sizes since we may have to handle excessively large amounts of 4k requests for data sizes that may become common soon. For example to write a 1 terabyte file the kernel would have to handle 256 million 4k chunks. 4. Cross arch compatibility: It is currently not possible to mount an 16k blocksize ext2 filesystem created on IA64 on an x86_64 system. With this patch this becomes possible. Note that this also means that some filesystems are already capable of working with blocksizes of up to 64k (ext2, XFS) which is currently only available on a select few arches. This patchset enables that functionality on all arches. There are no special modifications needed to the filesystems. The set_blocksize() function call will simply support a larger blocksize. 5. VM scalability Large block sizes mean less state keeping for the information being transferred. For a 1TB file one needs to handle 256 million page structs in the VM if one uses 4k page size. A 64k page size reduces that amount to 16 million. If the limitation in existing filesystems are removed then even higher reductions become possible. For very large files like that a page size of 2 MB may be beneficial which will reduce the number of page struct to handle to 512k. The variable nature of the block size means that the size can be tuned at file system creation time for the anticipated needs on a volume. 6. IO scalability The IO layer will receive large blocks of contiguious memory with this patchset. This means that less scatter gather elements are needed and the memory used is guaranteed to be contiguous. Instead of having to handle 4k chunks we can f.e. handle 64k chunks in one go. 7. Limited scatter gather support restricts I/O sizes. A lot of I/O controllers are limited in the number of scatter gather elements that they support. For example a controller that support 128 entries in the scatter gather lists can only perform I/O of 128*4k = 512k in one go. If the blocksize is larger (f.e. 64k) then we can perform larger I/O transfers. If we support 128 entries then 128*64k = 8M can be transferred in one transaction. Dave Chinner measured a performance increase of 50% when going to 64k blocksize with XFS with an earlier version of this patchset. 8. We have problems supporting devices with a higher blocksize than page size. This is for example important to support CD and DVDs that can only read and write 32k or 64k blocks. We currently have a shim layer in there to deal with this situation which limits the speed of I/O. The developers are currently looking for ways to completely bypass the page cache because of this deficiency. 9. 32/64k blocksize is also used in flash devices. Same issues. 10. Future harddisks will support bigger block sizes that Linux cannot support since we are limited to PAGE_SIZE. Ok the on board cache may buffer this for us but what is the point of handling smaller page sizes than what the drive supports? Acceptance issues: ------------------ The patchset is a pretty significant change to the way that the Linux kernel operates. I have tried to keep the changes as minimal as possible and I believe that this is a reasonable start to introduce large block I/O capabilities. The Linux VM is gradually acquiring abilities to defragment memory. These capabilities are partially present for 2.6.23. Later versions may merge more of the defragmentation work. The use of large pages may cause significant fragmentation to memory. Without proper defragmentation support these patches cannot work reliably and may cause OOMs (although I have rarely seen those and I have seen none with 2.6.23 when testing with 16k blocksize. Possibly the effect of the limited defragmentation capabilities in 2.6.23 is already sufficient. Beware: Larger blocksize can cause more defragmentation). A number of key developers are hesitant about this functionality given the problems that we have had in the past with memory defragmentation and the invasiveness of the patchset. It is to be expected that it will take awhile until confidence in the defragmentation logic builds up. So the patchset may still have to exist for a long time unmerged. However, it is necessary that work on this patchset continue even outside of the kernel tree in order to mature this patchset until it can be merged at some point in the future. The most serious shortcoming of this patchset is the lack of mmap support. This means f.e. that it is not possible to execute binaries. Adding mmap support would mean more changes to the VM. In order to do that some support by multiple developers is likely needed. The current idea is to allow the mapping of 4K segments of larger pages to preserve mmap semantics. This means that application programs could still map pages in 4k chunks and the VM would provide a mapping into a subsection of a larger page. How to make this patchset work: ------------------------------- 1. Apply this patchset or do a git pull git://git.kernel.org/pub/scm/linux/kernel/git/christoph/largeblocksize.git largeblock (The git archive is used to keep the patchset up to date. Please send patches against the git tree) 2. Enable LARGE_BLOCKSIZE Support 3. Compile kernel In order to use a filesystem with a larger blocksize it needs to be formatted for that larger blocksize. This is done using the mkfs.xxx tool for each filesystem. Surprisingly the existing tools work without modification. These formatting tools may warn you that the blocksize you specify is not supported on your particular architecture. Ignore that warning since this is no longer true after you have applied this patchset. Tested file systems: Filesystem Max Blocksize Changes Reiserfs 8k Page size functions Ext2 64k Page size functions XFS 64k Page size functions / Remove PAGE_SIZE check Ramfs MAX_ORDER Parameter to specify order Todo/Issues: - There are certainly numerous issues with this patch. I have only tested copying files back and forth, volume creation etc. Others have run fsxlinux on the volumes. The missing mmap support limits what can be done for now. - ZONE_MOVABLE is available in 2.6.23. Using the kernelcore=xxx as a kernel parameter enables an area where defragmentation can work. This may be necessary to avoid OOMs although I have seen no problems with up to 32k blocksize even without that measure. - The antifragmentation patches in Andrew's tree address more fragmentation issues. However, large orders may still lead to fragmentation of the movable sections. Memory compaction is still not merged and will likely be needed to reliably support even larger orders of 256k or more. How memory compaction impacts performance still has to be determined. - Support for bouncing pages. - Remove PAGE_CACHE_xxx constants after using page_cache_xxx functions everywhere. But that will have to wait until merging becomes possible. For now certain subsystems (shmem f.e.) are not using these functions. They will only use order 0 pages. - Support for non harddisk based filesystems. Remove the pktdvd etc layers needed because the VM current does not support sufficiently large blocksizes for these devices. Look for other places in the kernel where we have similar issues. - Mmap support V5->V6: - Rediff against 2.6.23-rc4 - Fix breakage introduced by updates to reiserfs - Readahead fixes by Fengguang Wu <fengguang.wu@xxxxxxxxx> - Provide a git tree that is kept up to date V4->V5: - Diff against 2.6.22-rc6-mm1 - provide test tree on ftp.kernel.org:/pub/linux V3->V4 - It is possible to transparently make filesystems support larger blocksizes by simply allowing larger blocksizes in set_blocksize. Remove all special modifications for mmap etc from the filesystems. This now makes 3 disk based filesystems that can use larger blocks (reiser, ext2, xfs). Are there any other useful ones to make work? - Patch against 2.6.22-rc4-mm2 which allows the use of Mel's antifrag logic to avoid fragmentation. - More page cache cleanup by applying the functions to filesystems. - Disable bouncing when the gfp mask is setup. - Disable mmap directly in mm/filemap.c to avoid filesystem changes while we have no mmap support for higher order pages. RFC V2->V3 - More restructuring - It actually works! - Add XFS support - Fix up UP support - Work out the direct I/O issues - Add CONFIG_LARGE_BLOCKSIZE. Off by default which makes the inlines revert back to constants. Disabled for 32bit and HIGHMEM configurations. This also allows a gradual migration to the new page cache inline functions. LARGE_BLOCKSIZE capabilities can be added gradually and if there is a problem then we can disable a subsystem. RFC V1->V2 - Some ext2 support - Some block layer, fs layer support etc. - Better page cache macros - Use macros to clean up code. -- - 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