| commit | 630e3a334375c671bcf4a9e07e80277d839c36db | |
| author | Matthew Dillon <dillon@apollo.backplane.com> | |
| Sat, 13 Feb 2010 06:45:23 +0000 (12 22:45 -0800) | ||
| committer | Matthew Dillon <dillon@apollo.backplane.com> | |
| Sat, 13 Feb 2010 06:45:23 +0000 (12 22:45 -0800) | ||
| tree | c5a6631012ff76dbb47581f4faaa830f33cd95f7 | treesnapshot (tar.gz zip) |
| parent | 0786baf1fc52cfb8dfbb4703b677a498aa96bfe0 | commitdiff |
kernel - TMPFS - Bug fixing pass - fsync, vnode locks
* tmpfs_fsync() needs to have a couple of special cases to properly deal
with the backing store.
(1) When unmounting any vnode or recycling a dead vnode the backing
store can be destroyed.
(2) When recycling a live vnode the OS is trying to free the vnode,
meaning also free up the buffer cache for that vnode. tmpfs must
sync the data to swap backing store.
(3) Otherwise fsync can be ignored. We don't want to force data into
swap.
* The DragonFly kernel handles nearly all vnode and namespace locking
for VFSs. If tmpfs tries to do its own it will cause deadlocks.
At the moment we are not running MPSAFE but ultimately we will have
to do a pass to properly lock tmpfs_nodes during operations.
* tmpfs_fsync() needs to have a couple of special cases to properly deal
with the backing store.
(1) When unmounting any vnode or recycling a dead vnode the backing
store can be destroyed.
(2) When recycling a live vnode the OS is trying to free the vnode,
meaning also free up the buffer cache for that vnode. tmpfs must
sync the data to swap backing store.
(3) Otherwise fsync can be ignored. We don't want to force data into
swap.
* The DragonFly kernel handles nearly all vnode and namespace locking
for VFSs. If tmpfs tries to do its own it will cause deadlocks.
At the moment we are not running MPSAFE but ultimately we will have
to do a pass to properly lock tmpfs_nodes during operations.
| sys/vfs/tmpfs/tmpfs_vnops.c | diffblobblamehistory |