linux-kernel - Re: [PATCH v2] sched/debug: Add new tracepoint to track cpu

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 7 Sep 2020 15:51:55 +0100
From:   Qais Yousef <qais.yousef@....com>
To:     peterz@...radead.org
Cc:     Dietmar Eggemann <dietmar.eggemann@....com>,
        vincent.donnefort@....com, mingo@...hat.com,
        vincent.guittot@...aro.org, linux-kernel@...r.kernel.org,
        valentin.schneider@....com, Phil Auld <pauld@...hat.com>
Subject: Re: [PATCH v2] sched/debug: Add new tracepoint to track cpu_capacity

On 09/07/20 13:13, peterz@...radead.org wrote:
> On Mon, Sep 07, 2020 at 11:48:45AM +0100, Qais Yousef wrote:
> > IMHO the above is a hack. Out-of-tree modules should rely on public headers and
> > exported functions only. What you propose means that people who want to use
> > these tracepoints in meaningful way must have a prebuilt kernel handy. Which is
> > maybe true for us who work in the embedded world. But users who run normal
> > distro kernels (desktop/servers) will fail to build against
> 
> But this isn't really aimed at regular users. We're aiming this at
> developers (IIUC) so I dont really see this as a problem.
> 
> > FWIW, I did raise this concern with Peter in 2019 OSPM and he was okay with the
> > exports as it's still not a contract and they can disappear anytime we want.
> > Migrating to using BTF is the right way forward IMO. I don't think what we have
> > here is out-of-control yet. Though I agree they're annoying.
> 
> Right, we're hiding behind the explicit lack of ABI for modules.
> 
> Anyway, CTF/BTF/random other crap that isn't DWARFs should work fine to
> replace all this muck. Just no idea what the state of any of that is.

So I was thinking of having a function that allows a module to read member of
struct rq (or any struct for that matters), but I think that's the harder
(though neater) way around.

Just compiled a kernel with CONFIG_DEBUG_INFO_BTF_INFO; and doing

	$ pahole rq
	struct rq {
        raw_spinlock_t             lock;                 /*     0     4 */
        unsigned int               nr_running;           /*     4     4 */
        long unsigned int          last_blocked_load_update_tick; /*     8     8 */
        unsigned int               has_blocked_load;     /*    16     4 */

        /* XXX 12 bytes hole, try to pack */

        call_single_data_t         nohz_csd;             /*    32    32 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        unsigned int               nohz_tick_stopped;    /*    64     4 */
        atomic_t                   nohz_flags;           /*    68     4 */
        unsigned int               ttwu_pending;         /*    72     4 */

        /* XXX 4 bytes hole, try to pack */

        u64                        nr_switches;          /*    80     8 */
	.
	.
	.
	}

dumps the struct rq {...}; which means one can easily use that to autogenerate
a header containing the structs they care about accessing for their running
kernel.

pahole automatically knows how to find /sys/kernel/btf/vmlinux to parse the
debug info btw.

The only caveat is that one has to recompile the module for each running
kernel; but that's acceptable I think. Not sure how many allow loading a module
that's not compiled for that particular kernel version anyway.

Note to try this you'll need pahole v1.16 or newer. And compiling pahole on
Ubuntu is a pain. I had to create a fedora docker image to compile it in.

So I think we have this already solved. Though not sure how to document it..

Thanks

--
Qais Yousef