Re: [RFC v2 bpf-next 0/9] mm/bpf/perf: Store build id in inode object

[Jiri Olsa]

On Wed, Mar 01, 2023 at 12:41:20PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Wed, Mar 01, 2023 at 09:07:14AM +1100, Dave Chinner escreveu:
> > On Tue, Feb 28, 2023 at 10:31:57AM +0100, Jiri Olsa wrote:
> > > this is RFC patchset for adding build id under inode's object.
> 
> > > The main change to previous post [1] is to use inode object instead of file
> > > object for build id data.
> > 
> > Please explain what a "build id" is, the use case for it, why we
> > need to store it in VFS objects, what threat model it is protecting
> > the system against, etc.
> 
> [root@quaco ~]# 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]=160df51238a38ca27d03290f3ad5f7df75560ae0, for GNU/Linux 3.2.0, stripped
> [root@quaco ~]# file /lib64/libc.so.6
> /lib64/libc.so.6: ELF 64-bit LSB shared object, x86-64, version 1 (GNU/Linux), dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2, BuildID[sha1]=8257ee907646e9b057197533d1e4ac8ede7a9c5c, for GNU/Linux 3.2.0, not stripped
> [root@quaco ~]#
> 
> Those BuildID[sha1]= bits, that is present in all binaries I think in
> all distros for quite a while.
> 
> This page, from when this was initially designed, has a discussion about
> it, why it is needed, etc:
> 
>   https://fedoraproject.org/wiki/RolandMcGrath/BuildID
> 
> 'perf record' will receive MMAP records, initially without build-ids,
> now we have one that has, but collecting it when the mmap is executed
> (and thus a PERF_RECORD_MMAP* record is emitted) may not work, thus this
> work from Jiri.

thanks for the pointers

build id is unique id for binary that's been used to identify
correct binary version for related stuff.. like binary's debuginfo
in perf or match binary with stack trace entries in bpf stackmap

jirka

> 
> - Arnaldo
>  
> > > 
> > > However.. ;-) while using inode as build id storage place saves some memory
> > > by keeping just one copy of the build id for all file instances, there seems
> > > to be another problem.
>  
> > Yes, the problem being that we can cache hundreds of millions of
> > inodes in memory, and only a very small subset of them are going to
> > have open files associated with them. And an even smaller subset are
> > going to be mmapped.
>  
> > So, in reality, this proposal won't save any memory at all - it
> > costs memory for every inode that is not currently being used as
> > a mmapped elf executable, right?
> > 
> > > The problem is that we read the build id when the file is mmap-ed.
> > 
> > Why? I'm completely clueless as to what this thing does or how it's
> > used....
> > 
> > > Which is fine for our use case,
> > 
> > Which is?
> > 
> > -Dave.
> > -- 
> > Dave Chinner
> > david@xxxxxxxxxxxxx