HAMMER2 Freemap Design Notes Overview HAMMER2 Media is broken down into 2 GByte zones. Each 2 GByte zone contains a 4 MByte header (64 x 64K blocks). The blocks in this header are reserved for various purposes. For example, block #0 is used for the volume header or for a volume header backup. * It is very important to remember that the Freemap only uses blocks from these reserved areas. Freemap blocks are NOT dynamically allocated. * On-mount, the synchronization TID for the main H2 filesystem is compared against the synchronization TID of the freemap and the H2 topology is incrementally iterated using mirror_tid to update the freemap with any missing information. This way the freemap flush does not need to be synchronized with the normal H2 flush. This can be done very quickly on-mount. * The freemap is flushed in a manner similar to the normal H2 filesystem, but as mentioned above it can be synchronized independently of the data it represents. One freemap flush could cover several H2 flushes. A freemap flush is not necessary for e.g. a fsync() or sync() to complete successfully. * The freemap granularity is 64KB (radix of 16) but the minimum allocation radix for code is 1KB (radix of 10). 1KB inodes can hold up to 512 bytes of direct data, so small files eat exactly 1KB of media storage inclusive of the inode. * Representation of storage is block-oriented with ~1KB granularity in the filesystem topology. However, H2 also stores freemap locality hints in the inode at all levels which specifies which freemap zones all storage allocations made by the sub-tree are allocated from. Up to four zones may be listed in each inode. The zones are power-of-2 sized and aligned the same and use a base/radix representation (same as used for blockref->data_off). During updates higher level inodes may not have a sufficient number of entries to represent the storage used on a fine-grain. In this situation the representations back-off to larger radix values. Ultimately these representations will be optimized by background scans. That is, ultimately storage localization can be optimized bottom-up such that it winds up being fairly optimal. This includes the ability to detect when a writable snapshot has differentiated sufficiently to warrant a split. This optimization should NOT attempt to dup common data blocks. XXX * The zone oriented forward storage references in the inode (the four entries) is used by the bulk free-scan to reduce the amount of meta-data which must be duplicatively scanned. More specifically, when the sysadmin deletes storage and/or files or even whole directory subhierachies, it is possible for a bulk free-scan to incrementally scan the meta-data topology that covers ONLY those areas to determine if it is possible to free up any actual blocks. XXX * H2 does not require that a rm -rf or snapshot destruction, truncation, or any other operation actually mark freemap blocks as being almost-free. That is, the freemap elements can remain set to ALLOCATED (11). In fact, it is possible to just delete the directory inode itself and not even recursively scan or delete sub-directories or files. The related storage will eventually be freed by an exhaustive bulk free-scan anyway. Freemap Topology The freemap topology contains 4 levels of meta-data (blockref arrays), one of which is embedded in the volume header (so only three real meta-data levels), plus one level of leaf-data. Level 1 - (radix 10) 64KB blockmap representing 2GB. There are 1024 entries representing ~2MB worth of media storage per entry. Each entry maps 32 x 64KB allocations @ 2 bits per allocation, plus contains additional meta-data which allows H2 to cluster I/O operations. Each entry locks the allocation granularity (e.g. to 1KB = radix 10 for inodes). Level 2 - (radix 10) 64KB blockmap representing 2TB (~2GB per entry) Level 3 - (radix 10) 64KB blockmap representing 2PB (~2TB per entry) Level 4 - (radix 10) 64KB blockmap representing 2EB (~2PB per entry) Level 5 - (radix 3) blockref x 8 in volume header representing 16EB (2^64) (this conveniently eats one 512-byte 'sector' of the 64KB volume header). Each level is assign reserved blocks in the 4MB header per 2GB zone. Since we use block 0 for the volume header / volume header backup, our level names above can simply also represent the relative block number. Level 1 uses block 1 through level 4 using block 4. Level 5 is stored in the volume header. In addition there are FOUR SETS, A, B, C, and D, each containing blocks for level 1-4. Hammer2 alternates between sets on a block-by-block basis in order to maintain consistency when updating the freemap. Leaf Substructure * radix - Clustering radix. All allocations for any given ~2MB zone are always the same size, allowing the filesystem code to cluster buffer cache I/O. * bitmap - four 32 bit words representing ~2MB in 64KB allocation chunks at 2 bits per chunk. The filesystem allocation granularity can be smaller (currently ~1KB minimum), and the live filesystem keeps caches iterations when allocating multiple chunks. However, on remount any partial allocations out of a 64KB allocation block causes the entire 64KB to be considered allocated. Fragmented space can potentially be reclaimed and/or relocated by the bulk block free scan. The 2-bit bitmap fields are assigned as follows: 00 FREE 01 ARMED for free stage (future use) 10 ARMED for free stage (future use) 11 ALLOCATED It should be noted that in some cases, such as snapshot destruction, H2 does not bother to actually ARM the related blocks (which would take a long time). Instead, the bulk free-scan may have to do a more exhaustive scan. Blockref Substructure The blockref substructure at each level steals some space from the check code area (a 24-byte area). We only need 4 bytes for the check code icrc. We use some of the remaining space to store information that allows the block allocator to do its work more efficiently. * bigmask - A mask of radixes available for allocation under this blockref. Typically initialized to -1. * avail - Total available space in bytes. The freemap allocator uses a cylinder-group-like abstraction using the localized allocation concept first implemented by UFS. In HAMMER2 there is no such thing as a real cylinder group, but we do the next best thing by implementing our layer 1 blockmap representing 2GB. The layer 1 blockmap is an array of 1024 blockrefs, so each blockref covers 2MB worth of media storage. HAMMER2's 'cylinder group' concept thus has a minimum granularity of 2MB. A typical setting might be e.g. 10MB. By localizing allocations to cylinder groups based on various bits of information, HAMMER2 tries to allocate space on the disk and still leave some left over for localized expansion and to reduce fragmentation at the same time. Not an easy task, especially considering the copy-on-write nature of the filesystem. This part of the algorithm likely needs a lot of work but I hope I've laid down a media format that will not have to be changed down the line to accomodate better allocation strategies. Initial Conditions The freemap is a multi-indirect block structure but there is no real reason to pre-format it in newfs_hammer2. Instead, newfs_hammer2 simply leaves the associated top-level indirect blocks empty and uses the (voldata->allocator_beg) field to allocate space linearly, then leaves it to the live filesystem to initialize the freemap as more space gets allocated. How blocks are freed The freemap bit patterns for each 64KB block are as follows: 00 FREE 01 ARMED (for free) (future use) 10 ARMED (for free) (future use) 11 ALLOCATED Currently H2 only implements 00 and 11. When a file, topology, or snapshot is deleted H2 simply leaves the blocks marked allocated but records the related freezone/radix(s) in memory. At some point a background bulk free-scan will run. This code must scan meta-data and has a limited cache to detect duplicative sub-trees (due to snapshots). It uses the freezone/radix information recorded in memory to reduce the complexity of the scan, find all references to the related blocks in the meta-data, and determines what can actually be freed. Once this determination is made the bulk free-scan sets the related freemap bits to FREE (00). An exhaustive free-scan is not usually required during normal operation but is typically run incrementally by cron every so often to ensure, over time, that all freeable blocks are actually freed. This is most useful when maintaining multiple snapshots. Use of Generic indirect-block API I decided to use the same indirect-block allocation model for the freemap that normal files use, with a few special cases added to force specific radix values and to 'allocate' the freemap-related blocks and indirect blocks via a reserved-block calculation and (obviously) not via a recursive call to the allocator. The Freemap is defined above as a fixed 5-level scheme (level 1-5), but in actual operation the radix tree can be shortcut just as it is with normal files. However, shorcuts are forced into the radix values of this specification and reserved blocks are calculated based on the radix level and offset, so as the freemap becomes more fleshed out the tree looks more and more like the specification. One advantage of doing things this way is that smaller filesystems won't actually use a 6-level scheme. A 16GB filesystem can use 8 blockrefs at layer 5 (in the volume header) that point directly to layer 1. A 16TB filesystem can use 8 blockrefs at layer5 that point to layer 2. And so forth. At the moment we have no plans to return any of the unused 4MB zone header space (per 2GB of storage) back to the filesystem for general use. There are lots of things we may want to use the reserved areas for in the future.