On 3/26/20 4:08 PM, Zbigniew Jędrzejewski-Szmek wrote:
On Thu, Mar 26, 2020 at 08:44:47AM -0400, Neal Gompa wrote:
On Thu, Mar 26, 2020 at 8:22 AM Zbigniew Jędrzejewski-Szmek zbyszek@in.waw.pl wrote:
On Thu, Mar 26, 2020 at 07:38:33AM -0400, Neal Gompa wrote:
On Thu, Mar 26, 2020 at 7:33 AM Zbigniew Jędrzejewski-Szmek zbyszek@in.waw.pl wrote:
- a one-shot service: this is easier to implement, it just needs to happen in one place. The hard part is making sure that the machine does not get reboot while the upgrade is happening. This is in particular a problem with VMs and containers. The rebuild should be wrapped with systemd-inhibit and other guards to make it hard to interrupt.
Wouldn't the systemd-inhibit plugin automatically ensure that a rebuild action would block sleep/poweroff?
Unfortunately... not. From the man page: inhibitors "may be used to block or delay system sleep and shutdown requests from the user, as well as automatic idle handling of the OS." Explicit non-interactive privileged requests override inhibitors [1,2]. This has been discussed, and I think there's general sentiment that we should have an ability to inhibit "everything", but so far nobody has pushed for a solution. A solution could be proritized if it turns out to be required in Fedora.
[1] https://github.com/systemd/systemd/issues/2680 [2] https://github.com/systemd/systemd/issues/6644
I'm not sure it's needed with rpm 4.14+, since worst case is that you have to trigger the rebuild some other time if it's interrupted.
No matter how it wrapped, is the upgrade itself atomic? Having the new db built under a temporary file name and then atomically rename(2)d into place would be ideal.
Since RPM 4.14, RPM creates a new directory, writes the database content there, then renames the directory when it's done.
Does it use renameat2(RENAME_EXCHANGE)?
No, but I don't think that matters with the way it's implemented?
See: https://github.com/rpm-software-management/rpm/commit/fffd652c56eaef8fc41d23...
I think otherwise it'd be hard to do an atomic replacement when the database consists of more than one file. But looking at the code:
xx = rename(dest, old); if (xx) { goto exit; } xx = rename(src, dest);(dest, src, old are all single-file paths) if I'm reading this correctly, it doesn't even to atomic replacement of individual files. If the machine is rebooted between the two renames above, no database ;(
It *used to* replace files one by one before https://github.com/rpm-software-management/rpm/commit/fffd652c56eaef8fc41d23...
Now it's just two rename() calls on directories, so it's worlds better and the best we can do with portable system calls. So yes if you win the unlucky lottery to have system reboot between those two rename() calls, you can end up with having to manually rename it into place.
So to put that into perspective, --rebuilddb was *much* worse for over twenty years and I've never heard anybodys database getting nuked because of *that*.
On Thu, Mar 26, 2020 at 03:29:50PM +0200, Panu Matilainen wrote:
No, rpm doesn't use many Linux-specific calls and this is no exception. In fact it doesn't use any of the *at() family calls directly either.
But why?! It's not like rpm is massive on Windows Server... Isn't good support for Linux absolutely the most important thing?
The places where rpm turns up never ceases to astonish me, starting from OS X to FreeBSD to the old proprietory unixen (AIX was already mentioned). There's even an actively maintained OS/2 port of rpm. Seriously. To that, I had to mostly say no.
Of course, these things can be abstracted out in portable wrappers to take advantage of newer and/or OS-specific features, but guess what, that takes time and effort. As does replacing any code.
The *at() family was added in POSIX.1-2008, and time may finally be ripe for rpm to start requiring them without having to paper over with wrapper calls, but even then, somebody needs to do the work. We haven't recently had any spare cycles to go chase looking for stuff that could be maybe replaced with something newer just because its ... newer. I don't recall a single rpm bug report that would've been avoided by use of the *at() stuff, so there doesn't seem to be a whole lot of benefit.
Rewriting rpm's core file management operations for more robust error handling has been in the todo for years. When that finally happens, it'll be a good opportunity to take advantage of the newer interfaces.
This is actually the first I hear about that system call which indeed seems highly useful for rpm, so thanks for the tip :)
So... does this mean that we can get #61413 a.k.a. #447156 resolved?
No.
- Panu -
Zbyszek _______________________________________________ devel mailing list -- devel@lists.fedoraproject.org To unsubscribe send an email to devel-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/devel@lists.fedoraproject.org