1 * Need backend synchronization / serialization when the frontend detaches
2 a XOP. modify_tid tests won't be enough, the backend may wind up executing
3 the XOP out of order after the detach.
5 * xop_start - only start synchronized elements
7 * See if we can remove hammer2_inode_repoint()
9 * FIXME - logical buffer associated with write-in-progress on backend
10 disappears once the cluster validates, even if more backend nodes
13 * FIXME - backend ops need per-node transactions using spmp to protect
16 * FIXME - modifying backend ops are not currently validating the cluster.
17 That probably needs to be done by the frontend in hammer2_xop_start()
19 * modify_tid handling probably broken w/ the XOP code for the moment.
21 * embedded transactions in XOPs - interlock early completion
23 * remove current incarnation of EAGAIN
25 * mtx locks should not track td_locks count?. They can be acquired by one
26 thread and released by another. Need API function for exclusive locks.
28 * Convert xops and hammer2_update_spans() from cluster back into chain calls
30 * syncthr leaves inode locks for entire sync, which is wrong.
32 * recovery scan vs unmount. At the moment an unmount does its flushes,
33 and if successful the freemap will be fully up-to-date, but the mount
34 code doesn't know that and the last flush batch will probably match
35 the PFS root mirror_tid. If it was a large cpdup the (unnecessary)
36 recovery pass at mount time can be extensive. Add a CLEAN flag to the
37 volume header to optimize out the unnecessary recovery pass.
39 * More complex transaction sequencing and flush merging. Right now it is
40 all serialized against flushes.
42 * adding new pfs - freeze and force remaster
44 * removing a pfs - freeze and force remaster
46 * bulkfree - sync between passes and enforce serialization of operation
48 * bulkfree - signal check, allow interrupt
50 * bulkfree - sub-passes when kernel memory block isn't large enough
52 * bulkfree - limit kernel memory allocation for bmap space
54 * bulkfree - must include any detached vnodes in scan so open unlinked files
55 are not ripped out from under the system.
57 * bulkfree - must include all volume headers in scan so they can be used
58 for recovery or automatic snapshot retrieval.
60 * bulkfree - snapshot duplicate sub-tree cache and tests needed to reduce
63 * Currently the check code (bref.methods / crc, sha, etc) is being checked
64 every single blasted time a chain is locked, even if the underlying buffer
65 was previously checked for that chain. This needs an optimization to
66 (significantly) improve performance.
68 * flush synchronization boundary crossing check and current flush chain
71 * snapshot creation must allocate and separately pass a new pmp for the pfs
72 degenerate 'cluster' representing the snapshot. This theoretically will
73 also allow a snapshot to be generated inside a cluster of more than one
76 * snapshot copy currently also copies uuids and can confuse cluster code
78 * hidden dir or other dirs/files/modifications made to PFS before
79 additional cluster entries added.
81 * transaction on cluster - multiple trans structures, subtrans
83 * inode always contains target cluster/chain, not hardlink
85 * chain refs in cluster, cluster refs
87 * check inode shared lock ... can end up in endless loop if following
88 hardlink because ip->chain is not updated in the exclusive lock cycle
89 when following hardlink.
91 cpdup /build/boomdata/jails/bleeding-edge/usr/share/man/man4 /mnt/x3
94 * The block freeing code. At the very least a bulk scan is needed
95 to implement freeing blocks.
97 * Crash stability. Right now the allocation table on-media is not
98 properly synchronized with the flush. This needs to be adjusted
99 such that H2 can do an incremental scan on mount to fixup
100 allocations on mount as part of its crash recovery mechanism.
102 * We actually have to start checking and acting upon the CRCs being
105 * Remaining known hardlink issues need to be addressed.
107 * Core 'copies' mechanism needs to be implemented to support multiple
108 copies on the same media.
110 * Core clustering mechanism needs to be implemented to support
111 mirroring and basic multi-master operation from a single host
112 (multi-host requires additional network protocols and won't
115 * make sure we aren't using a shared lock during RB_SCAN's?
117 * overwrite in write_file case w/compression - if device block size changes
118 the block has to be deleted and reallocated. See hammer2_assign_physical()
121 * freemap / clustering. Set block size on 2MB boundary so the cluster code
122 can be used for reading.
124 * need API layer for shared buffers (unfortunately).
126 * add magic number to inode header, add parent inode number too, to
127 help with brute-force recovery.
129 * modifications past our flush point do not adjust vchain.
130 need to make vchain dynamic so we can (see flush_scan2).??
132 * MINIOSIZE/RADIX set to 1KB for now to avoid buffer cache deadlocks
133 on multiple locked inodes. Fix so we can use LBUFSIZE! Or,
134 alternatively, allow a smaller I/O size based on the sector size
135 (not optimal though).
137 * When making a snapshot, do not allow the snapshot to be mounted until
138 the in-memory chain has been freed in order to break the shared core.
140 * Snapshotting a sub-directory does not snapshot any
141 parent-directory-spanning hardlinks.
143 * Snapshot / flush-synchronization point. remodified data that crosses
144 the synchronization boundary is not currently reallocated. see
145 hammer2_chain_modify(), explicit check (requires logical buffer cache
148 * on fresh mount with multiple hardlinks present separate lookups will
149 result in separate vnodes pointing to separate inodes pointing to a
150 common chain (the hardlink target).
152 When the hardlink target consolidates upward only one vp/ip will be
153 adjusted. We need code to fixup the other chains (probably put in
154 inode_lock_*()) which will be pointing to an older deleted hardlink
157 * Filesystem must ensure that modify_tid is not too large relative to
158 the iterator in the volume header, on load, or flush sequencing will
159 not work properly. We should be able to just override it, but we
160 should complain if it happens.
162 * Kernel-side needs to clean up transaction queues and make appropriate
165 * Userland side needs to do the same for any initiated transactions.
167 * Nesting problems in the flusher.
169 * Inefficient vfsync due to thousands of file buffers, one per-vnode.
170 (need to aggregate using a device buffer?)
172 * Use bp->b_dep to interlock the buffer with the chain structure so the
173 strategy code can calculate the crc and assert that the chain is marked
174 modified (not yet flushed).
176 * Deleted inode not reachable via tree for volume flush but still reachable
177 via fsync/inactive/reclaim. Its tree can be destroyed at that point.
179 * The direct write code needs to invalidate any underlying physical buffers.
180 Direct write needs to be implemented.
182 * Make sure a resized block (hammer2_chain_resize()) calculates a new
183 hash code in the parent bref
185 * The freemap allocator needs to getblk/clrbuf/bdwrite any partial
186 block allocations (less than 64KB) that allocate out of a new 64K
187 block, to avoid causing a read-before-write I/O.
189 * Check flush race upward recursion setting SUBMODIFIED vs downward
190 recursion checking SUBMODIFIED then locking (must clear before the
191 recursion and might need additional synchronization)
193 * There is definitely a flush race in the hardlink implementation between
194 the forwarding entries and the actual (hidden) hardlink inode.
196 This will require us to associate a small hard-link-adjust structure
197 with the chain whenever we create or delete hardlinks, on top of
198 adjusting the hardlink inode itself. Any actual flush to the media
199 has to synchronize the correct nlinks value based on whether related
200 created or deleted hardlinks were also flushed.
202 * When a directory entry is created and also if an indirect block is
203 created and entries moved into it, the directory seek position can
204 potentially become incorrect during a scan.
206 * When a directory entry is deleted a directory seek position depending
207 on that key can cause readdir to skip entries.
209 * TWO PHASE COMMIT - store two data offsets in the chain, and
210 hammer2_chain_delete() needs to leave the chain intact if MODIFIED2 is
211 set on its buffer until the flusher gets to it?
216 * If a file is unlinked buts its descriptors is left open and used, we
217 should allow data blocks on-media to be reused since there is no
218 topology left to point at them.