On Thu, Aug 28, 2025 at 12:18:39PM -0700, Linus Torvalds wrote: > On Thu, 28 Aug 2025 at 11:58, Arnaldo Carvalho de Melo > <arnaldo.m...@gmail.com> wrote:
> > >Give the damn thing an actual filename or something *useful*, not a > > >number that user space can't even necessarily match up to anything. > > A build ID? > I think that's a better thing than the disgusting inode number, yes. > That said, I think they are problematic too, in that I don't think > they are universally available, so if you want to trace some > executable without build ids - and there are good reasons to do that - > you might hate being limited that way. Right, but these days gdb (and other traditional tools) supports it and downloads it (perf should do it with a one-time sticky question too, does it already in some cases, unconditionally, that should be fixed as well), most distros have it: ⬢ [acme@toolbx perf-tools-next]$ file /bin/bash /bin/bash: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2, BuildID[sha1]=707a1c670cd72f8e55ffedfbe94ea98901b7ce3a, for GNU/Linux 3.2.0, stripped ⬢ [acme@toolbx perf-tools-next]$ We have debuginfod-servers that brings ELF images with debug keyed by that build id and finally build-ids come together with pathnames, so if one is null, fallback to the other. Default in fedora: ⬢ [acme@toolbx perf-tools-next]$ echo $DEBUGINFOD_ $DEBUGINFOD_IMA_CERT_PATH $DEBUGINFOD_URLS ⬢ [acme@toolbx perf-tools-next]$ echo $DEBUGINFOD_ $DEBUGINFOD_IMA_CERT_PATH $DEBUGINFOD_URLS ⬢ [acme@toolbx perf-tools-next]$ echo $DEBUGINFOD_IMA_CERT_PATH /etc/keys/ima: ⬢ [acme@toolbx perf-tools-next]$ echo $DEBUGINFOD_URLS https://debuginfod.fedoraproject.org/ ⬢ [acme@toolbx perf-tools-next]$ I wasn't aware of that IMA stuff. So even without the mandate and with sometimes not being able to get that build-id, most of the time they are there and deterministically allows tooling to fetch it in most cases, I guess that is as far as we can pragmatically get. - Arnaldo > So I think you'd be much better off with just actual pathnames. > > Are there no trace events for "mmap this path"? Create a good u64 hash > from the contents of a 'struct path' (which is just two pointers: the > dentry and the mnt) when mmap'ing the file, and then you can just > associate the stack trace entry with that hash. > > That should be simple and straightforward, and hashing two pointers > should be simple and straightforward. > > And then matching that hash against the mmap event where the actual > path was saved off gives you an actual *pathname*. Which is *so* much > better than those horrific inode numbers. > > And yes, yes, obviously filenames can go away and aren't some kind of > long-term stable thing. But inode numbers can be re-used too, so > that's no different. > > With the "create a hash of 'struct path' contents" you basically have > an ID that can be associated with whatever the file name was at the > time it was mmap'ed into the thing you are tracing, which is I think > what you really want anyway. > > Now, what would be even simpler is to not create a hash at all, but > simply just create the whole pathname when the stack trace entry is > created. But it would probably waste too much space, since you'd > probably want to have at least 32 bytes (as opposed to just 64 bits) > for a (truncated) pathname. > > And it would be more expensive than just hashing the dentry/mnt > pointers, although '%pD' isn't actually *that* expensive. But probably > expensive enough to not really be acceptable. I'm just throwing it out > as a stupid idea that at least generates much more usable output than > the inode numbers do. > > Linus
