Usability and Mac improvements #14
Conversation
Adds an exit() overload and a library destructor routine that call libc_sync() when syncs have been eaten and the EATMYDATA_END_SYNC env. var is set. They also print how many syncs were swallowed, if EATMYDATA_VERBOSE is set. Committed from host : Portia.local
Committed from host : Portia.local
let the OS deallocate progName and cache nosyncs before printing, all to prevent race conditions and print something that's less likely to be wrong. Committed from host : Bola
Set EATMYDATA_END_SYNC=1 to sync only when the 1st eatmydata instance exits (libeatmydata reads and unsets the variable internally). Use EATMYDATA_END_SYNC=2 to present the variable and cause an end-sync in each child process. Also, reformat the source. Committed from host : Portia.local
EATMYDATA_END_SYNC=-1 will disable eating syncs after exit() or _exit() have been called, allowing the libc runtime to do any of the explicit syncs it wants to do. A new close() overload also starts doing explicit fsync's on the descriptors it has to close. Potential because I haven't yet seen it do any syncs or NOT eat any. Committed from host : Bola
| exe="`which "${cmd}"`" | ||
| if [ -x "$exe" ] ;then | ||
| ok="yes" | ||
| else |
There was a problem hiding this comment.
Just FYI, the whole point of the path-searching logic here is that you can't use which to look up the command in question — which will almost certainly return your own location, in the case where some other executable's name is symlinked to eatmydata.
i.e.: If you have $HOME/bin/git symlinked to /usr/bin/eatmydata, because you want to use libeatmydata on /usr/bin/git, then assuming $HOME/bin/ is earlier in your $PATH you can just run git clone https://github.com/stewartsmith/libeatmydata and it'll launch /usr/bin/eatmydata with $1 set to git and $2 set to https://github.com/stewartsmith/libeatmydata.
But a which git will ALWAYS RETURN $HOME/bin/git in that scenario, which is the exact thing the "avoid loops with self" code is trying to prevent.
In order to find /usr/bin/git as the real executable to launch, you have to search the $PATH yourself instead of using which.
I don't know how universally supported it is, but if the available which implementation supports the -a flag, using which -a "$cmd" (to get all commands matching $cmd on the $PATH) could be a useful way to avoid walking the $PATH manually. But each of the returned executables would still have to be checked against "`readlink -f "$exe"`" != "$self", because the first item would be expected to fail that test.
|
On Wednesday February 05 2025 20:46:43 Frank Dana wrote:
Talk about a blast from the past...
Just FYI, the whole point of the path-searching logic here is that you _can't_ use `which` to look up the command in question — `which` will almost certainly return **your own location**, in the case where some other executable's name is symlinked to `eatmydata`.
Doh, yes.
IIRC the idea here was to speed up the process. I don't know if there's a universal and robust way to determine if "$exe = $0" but there has to be one to check if "$exe" is in fact a link to `eatmydata`.
|
2 participants
This is the PR announced in issue #13 . It's become a bit of a catch-all for a number of more or less (un)related changes:
EATMYDATA_END_SYNC=1to do this only in the application launched by the user, set itEATMYDATA_END_SYNC=2to do it in every child process too (that does slow things down if the child processes each do only tiny bits of work).EATMYDATA_VERBOSEto print out the number of syncs and sync modes that were swallowed. Uses a simple counter without any thread-safeties because there's nothing crucial to the feature and I don't want to introduce locks in overloads of functions that have to remain signal-safe.And, a bit stupidly: a source code reformat.