bluestore: latest and greatest #6896

Signed-off-by: Sage Weil <sage@redhat.com>

Instead of a single, variable-length fragment for each object, set a fixed size (newstore_min_frag_size = 1 MB) and stripe the object over these. The last fragment will be smaller than 1 MB if the object is not a multiple of 1 MB. On write, this is basically free: we can just as cheaply write 4 inodes created together and fsync them than we can one. On overwrite, it allows us to replace individual fragments and avoid write-ahead many cases. On read it is a bit slower because of inode lookups and disk seeks. In the common case (big object written sequentially) we hope that fs prefetching will hide most of it (e.g., all inodes will be loaded together in the same metadata btree node, and the files' data is written sequentially on disk). Allowing for a singe large fragment in the case of a sequentially written large object may save us something, but it complicates the code significantly. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com> fix wal_oP_t

Signed-off-by: Sage Weil <sage@redhat.com>

- pack u32 and u64 in binary (instead of in hex) - avoid duplicating the object name while making things still sort by (key,name). Use < when key < name, = when key == name, > when key > name) as a prefix. And in the = case (which is basically always) include the name just once. Note that this breaks on-disk compatibility. Signed-off-by: Sage Weil <sage@redhat.com>

Because of this change (ceph#6076), the hobject_t will contain pool id, hence the ghobject_t having this hobject_t will be not equal to ghobject_t(). In newstore, this will cause assertion failure: FAILED assert(k >= start_key && k < end_key) The fix is to make compatible with previous change to create a ghobject_t object with pool id and shard id in newstore. Fixes: ceph#13801 Reported-by: Zhi Zhang <zhangz.david@outlook.com> Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

FreelistManager perists our freelist. Allocator is a policy that allocates it. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Can't use invalid iterator; fix init_rm_free. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

BlueFS is a simple file system that will back rocksdb. BlueRocksEnv is the rocksdb::Env implementation that glues them together. Signed-off-by: Sage Weil <sage@redhat.com>

Just for now, while we get the rest of this working. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Currently we just do a dump. We'll add more functionality later. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

There are FileWriters that exist when the file is deleted. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Merge contiguous extents. Signed-off-by: Sage Weil <sage@redhat.com>

Rewrite only the current metadata in a fresh log periodically to free log space. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

use atomics, do not track in-flight extents or magically cope with racing ios (that is the users responsibility). Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Use this device for the bluefs log. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

If a file has dirty metadata (but no dirty data), we still need to flush the log when it is flushed. Signed-off-by: Sage Weil <sage@redhat.com>

Label all of our block devices with a simple label that includes the osd_uuid. Wire this into the ObjectStore and OSD probe mechanism. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

rocksdb will occasionally overwrite an existing file if it is not present/valid in the manifest. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Support a mid- to fast device that will preferentially store the rocksdb data (and wal, if block.wal is not present). Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Record kv backend at mkfs time instead of relying on current value of config option. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

This will be used to refcount extents for some subset of the store (objects with same name or hash value?). Signed-off-by: Sage Weil <sage@redhat.com>

Prefix with bluestore_ Signed-off-by: Sage Weil <sage@redhat.com>

This is based on BlueStore, but with all of the block-related code and complexity ripped out, and a simple striping strategy added in its place. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Reported-by: David Zafman <dzafman@redhat.com> Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Any overlap is an error. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Allocate aligned buffer. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

We submit each operation's transaction individually to rocksdb, and then since a final transction to flush them all. However, they may not commit atomically (all together), which means we need to leave the individual freelist updates within each transaction. Signed-off-by: Sage Weil <sage@redhat.com>

Rocksdb does a flush after every append, each of which is often less than a full block. This is very inefficient when our _flush() will send that to disk (and block). Avoid this most of the time by ignoring small flush requests entirely, unless the force flag is set (e.g., by fsync). Signed-off-by: Sage Weil <sage@redhat.com>

The rocksdb log recycle option allows us to overwrite previously allocated space in an old log file to avoid updating the file metadata on normal file systems. Take advantage of that here to by implementing what is effectively O_NOCMTIME semantics: we do not dirty the file metadata just because mtime is updated. Instead, we dirty the file only if we allocate new space or if the size has to be increased. Note that on my NVME drive a single-thread rados bench test, we jump from 30MB/sec to 50MB/sec 128KB writes as soon as we start recycling previous logs (about 40 second into the run). Signed-off-by: Sage Weil <sage@redhat.com>

Return (and log) usage for all bdevs. Signed-off-by: Sage Weil <sage@redhat.com>

...just so we log bdev utilization in the log. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Prefer aio unless explicitly directed otherwise. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

We may have multiple WAL ops that do read/modify/write covering the same blocks. To avoid the complexity of identifying those situations and ensuring that we, say, wait for writes to complete before reading them back again, just make the IO buffered and let the page cache handle that for us. This fixes the failure of LibRadosAio.RoundTripWriteFull. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

5 helpful (read/write offsets) 10 more, with aio completions 20 everything 30 fire hose 40 data hexdumps Signed-off-by: Sage Weil <sage@redhat.com>

We install negative entries instead. Signed-off-by: Sage Weil <sage@redhat.com>

Install a negative onode entry at the old name position. Otherwise, a simple transaction like rename a -> b touch b will re-read the old b onode key on the second op, and chaos will ensue (e.g., because it'll reference the same extents from a different object). Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Include these in used_blocks (they are about to be released but not reflected in the onode). Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

This condition was backwards. Signed-off-by: Sage Weil <sage@redhat.com>

The long names don't exercise useful code paths, and having consistent naming makes it easier to grep through logs. Signed-off-by: Sage Weil <sage@redhat.com>

Copy the buffer, in case other threads modify it in place. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

…en them We only guarantee support for clone between objects with the same hash. Signed-off-by: Sage Weil <sage@redhat.com>

400k for objects. Signed-off-by: Sage Weil <sage@redhat.com>

This makes debugging a bit easier because we never use the same extent of the disk twice, leaving useful evidence behind. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Assume block-aligned. Signed-off-by: Sage Weil <sage@redhat.com>

…ent_ref_map Signed-off-by: Sage Weil <sage@redhat.com>

These should run during make check. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Enodes will track extent ref counts for any extent that is marked shared. There will be an enode for any unique hash value that has any refs. We will keep in-memory copies of only those Enodes that are referenced by in-memory Onodes, and only if the enode is requested (e.g., the enode won't be loaded as a result of an object read because we never need to call get_enode.). Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

…itten Signed-off-by: Sage Weil <sage@redhat.com>

Force small allocations for debugging purposes. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

That's what fsync is for. Moreover, this can lead to some squirreliness if we trigger this from _flush_log(). Signed-off-by: Sage Weil <sage@redhat.com>

We've already forced everyone to upgrade through hammer, so everyone supports this. Just unconditionally set the feature if it is not set (for consistency's sake). Signed-off-by: Sage Weil <sage@redhat.com>

Print the *omap* digest, not *data* digest. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Only bluestore and kstore implement this currently. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Use up the amount we allocated, not the amount we asked for. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

This is an ancient obsolete test that is never used. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Make sure the extent we return is aligned to alloc_unit. If the first extent isn't suitable, try the next one. Keep bin 0 for things that are smaller than min_alloc_size. Signed-off-by: Sage Weil <sage@redhat.com>

We either want to take a clean error path, or dump core. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

...the better to diff them with. Signed-off-by: Sage Weil <sage@redhat.com>

The write process may do a read/modify/write on a stripe. In order to allow multiple writes to coexist within the same transaction, we need to be able to "see" our writes. Clear the "cached" stripe values when the last TransContext touching an onode is finished. In theory we could pin memory with a constant stream of updates to an object; we may need to address that later. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

In particular, we may get a shard specified along with hobject_t min or max (from PGBackend::objects_list_range()). Signed-off-by: Sage Weil <sage@redhat.com>

It didn't like k=1 with the default profile. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

This lets us do leak checking. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

It's possible for the IO to be in flight when the caller closes the writer handle (although dangerous of them). Queue the IOContext for async cleanup when we sync everything to disk. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

This makes all the IOs effectively synchronous (while holding the lock), which isn't great, but at least it's correct. We can contemplate async later.. Signed-off-by: Sage Weil <sage@redhat.com>

Just like ceph-osd. Signed-off-by: Sage Weil <sage@redhat.com>

Assume journal symlink is present. Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Signed-off-by: Jianpeng Ma <jianpeng.ma@intel.com>

Signed-off-by: Sage Weil <sage@redhat.com>

New backends don't work if it's off. Signed-off-by: Sage Weil <sage@redhat.com>

Fixes: ceph#14210 Signed-off-by: xie.xingguo <xie.xingguo@zte.com.cn>

Signed-off-by: YiQiang Chen <cyqsign@163.com>

Signed-off-by: Ning Yao <zay11022@gmail.com>

Signed-off-by: Sage Weil <sage@redhat.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

bluestore: latest and greatest #6896

bluestore: latest and greatest #6896

Commits on Jan 1, 2016

Commits on Jan 3, 2016