Fix empty registry file on machine crash #3668
Conversation
ruflin
added
discuss
| syscall.CREATE_ALWAYS, // Create Mode | ||
| uint32(syscall.FILE_ATTRIBUTE_NORMAL|_FILE_FLAG_WRITE_THROUGH), // Flags and Attributes | ||
| syscall.CREATE_ALWAYS, // Create Mode | ||
| uint32(syscall.FILE_ATTRIBUTE_NORMAL|_FILE_FLAG_WRITE_THROUGH|_FILE_FLAG_NO_BUFFERING), // Flags and Attributes |
There was a problem hiding this comment.
@andrewkroh I tried to find O_SYNC for windows and it seems to be the NO_BUFFERING. Looking into filebeat, we do not treat creating the registry for different OS differently. Should we?
In any case, we should still unify registrar handling in libbeat.
There was a problem hiding this comment.
Yeah, we should. We could move this file creation code up to libbeat so that we could share it. That would be simpler than trying to unify the registrar (which is still a good idea, but more time consuming).
There was a problem hiding this comment.
From what I understand on Windows you need to use FILE_FLAG_WRITE_THROUGH to get writes to bypass the cache. FILE_FLAG_NO_BUFFERING implies FILE_FLAG_WRITE_THROUGH too, but it affects reads and adds the requirement that your reads must be aligned to disk sector boundaries. So I don't think we need FILE_FLAG_NO_BUFFERING here.
There was a problem hiding this comment.
This is what originally pointed me to FILE_FLAG_WRITE_THROUGH. https://stackoverflow.com/questions/6525943/cause-of-corrupted-file-contents/24302389#24302389
ruflin
added
bug
needs_backport
|
I removed |
ruflin
removed
discuss
ruflin
force-pushed
the
fix-registry-osync
branch
from
7ba38f5
to
f717a9e
Compare
Closes elastic#3537 (cherry picked from commit e2e58c0)
Closes #3537