Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 18 Aug 2013 16:05:11 -0600
From:      Chris Torek <torek@torek.net>
To:        Carlos Jacobo Puga Medina <cjpugmed@gmail.com>
Cc:        freebsd-hackers@freebsd.org
Subject:   Re: ps_strings
Message-ID:  <201308182205.r7IM5BQR062114@elf.torek.net>
In-Reply-To: Your message of "Sat, 17 Aug 2013 20:00:55 %2B0200." <CAHHLbRNz5TMdz67fXfFk9pLD8UFiRg%2B0G_oD1GXP23Fo4B72Bw@mail.gmail.com>

next in thread | previous in thread | raw e-mail | index | archive | help
>Despite I made a request not long ago[1], I'm looking for
>documentation to create the ps_strings structure man page because
>isn't covered in other man page such e.g.  execve(2). So, I'm
>interested to know for what it's currently used.

Nothing.  (Well, backwards compatibility, depending on how far
backwards you go.)

I invented the "ps_strings" struct a long time ago, about the same
time libkvm was new.

Some background: There was code in "ps" that would grub around in
the top stack page of each user process and extract the argv
strings.  This code knew how execve() worked inside the kernel
(copying the argv and environment strings into the user stack,
just below the signal trampoline code, and then setting up argv
and envp pointers and invoking the libc/csu "start program" code
at the entry point).

We moved this grub-around-in-process-stack code to libkvm, but it
was still rather horrible code.

Meanwhile, we had programs like sendmail that would "set their
process title" by saving, in some secret global variable, the
space where the "argv" array itself and its strings lived, and
then -- in setproctitle() -- carefully overwrite it.  Of course
there was only as much room there as the kernel provided, based on
the actual strings at execve() time.

I figured this would all be much cleaner if we created a small
data structure, namely "ps_strings", to hold the information that
libkvm would extract (and hence, ps would show).  It would be
simpler than the original code, because the ps_strings structure
lived at a fixed, well known user-space virtual address (the same
VA in every process).  Moreover, a user program (like sendmail)
could modify the ps_strings data to point to any other user-space
area: libkvm was smart enough to extract arbitrary data (and
verify the validity of the address too).  This removed the limit
on how large a "process title" could be.

FreeBSD now, however, uses a per-process p_args field in the
"proc" structure, with sysctl()s to set and get p_args.  (I had
nothing to do with this new code, but I approve, as if anyone
cares. :-) )  This removes the fixed-virtual-address limitation.
The cost is a bit more kernel code (for the sysctl()s) and this
per-process data, but there is no more messing-about with "where
is ps_strings in this memory-layout / emulation" etc.  (Meanwhile
libkvm still retrieves the arguments.  It just does it now with
sysctl().)

Chris



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201308182205.r7IM5BQR062114>