lkml.org 
[lkml]   [2019]   [Jan]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: uprobes: bug in comm/string output?
On Thu, Jan 17, 2019 at 02:44:41PM +0100, Andreas Ziegler wrote:
> Hi,
>
> On 1/17/19 10:47 AM, Masami Hiramatsu wrote:
> > On Thu, 17 Jan 2019 09:08:41 +0100
> > Andreas Ziegler <andreas.ziegler@fau.de> wrote:
> >
> > > On 17.01.19 09:00, Masami Hiramatsu wrote:
> > > > On Thu, 17 Jan 2019 15:13:09 +0900
> > > > Masami Hiramatsu <mhiramat@kernel.org> wrote:
> > > >
> > > > > On Wed, 16 Jan 2019 11:16:07 +0100
> > > > > Andreas Ziegler <andreas.ziegler@fau.de> wrote:
> > > > >
> > > > > >
> > > > > > I went into this a bit deeper today, and right now it is simply failing
> > > > > > to parse the code because there is no FETCH_OP_COMM case in
> > > > > > process_fetch_insn() for uprobes so that will return -EILSEQ, leading to
> > > > > > a make_data_loc(0, ...) in store_trace_args(). If we just add
> > > > > > FETCH_OP_COMM and let val point to current->comm (that's what
> > > > > > trace_kprobe.c does), we get an -EFAULT return value from
> > > > > > fetch_store_string because strncpy_from_user() checks if the argument is
> > > > > > in user space.
> > > > >
> > > > > Correct. I missed to add OP_COMM support. And uprobe's fetch_store_string
> > > > > is only for user space strings.
> > > > >
> > > > > > So I think we might need a special case for that, something like
> > > > > > FETCH_OP_ST_COMM_STRING which is only used for FETCH_OP_COMM and copies
> > > > > > current->comm over to the dynamic area. The implementation could be
> > > > > > similar to the old fetch_comm_string implementation before your rewrite.
> > > > >
> > > > > Hmm, instead, I would like to add current->comm checker and only allows
> > > > > to copy that. That would be simpler and enough.
> > > > >
> > > > > Could you test below patch?
> > > > >
> > > > >
> > > > > tracing: uprobes: Re-enable $comm support for uprobe events
> > > > >
> > > > > From: Masami Hiramatsu <mhiramat@kernel.org>
> > > > >
> > > > > Since commit 533059281ee5 ("tracing: probeevent: Introduce new
> > > > > argument fetching code") dropped the $comm support from uprobe
> > > > > events, this re-enable it.
> > >
> > > this should read 're-enables'.
> > >
> > > > >
> > > > > For $comm support, use strncpy() instead of strncpy_from_user()
> > > ^
> > > we're using strlcpy(), not strncpy().
> > >
> > > > > to copy current task's comm. Because it is in the kernel space,
> > > > > strncpy_from_user() always fails to copy the comm.
> > > > > This also use strlen() instead of strlen_user() to measure the
> > > ^ ^
> > > 'uses', and the function should be 'strnlen_user()'.
> > >
> > > > > length of the comm.
> > > > >
> > > > > Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
> > > > > Reported-by: Andreas Ziegler <andreas.ziegler@fau.de>
> > > > > ---
> > > > > kernel/trace/trace_uprobe.c | 13 +++++++++++--
> > > > > 1 file changed, 11 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/kernel/trace/trace_uprobe.c b/kernel/trace/trace_uprobe.c
> > > > > index e335576b9411..97d134e83e0f 100644
> > > > > --- a/kernel/trace/trace_uprobe.c
> > > > > +++ b/kernel/trace/trace_uprobe.c
> > > > > @@ -156,7 +156,10 @@ fetch_store_string(unsigned long addr, void *dest, void *base)
> > > > > if (unlikely(!maxlen))
> > > > > return -ENOMEM;
> > > > > - ret = strncpy_from_user(dst, src, maxlen);
> > > > > + if (addr == (unsigned long)current->comm)
> > > > > + ret = strlcpy(dst, current->comm, maxlen);
> > > > > + else
> > > > > + ret = strncpy_from_user(dst, src, maxlen);
> > > > > if (ret >= 0) {
> > > > > if (ret == maxlen)
> > > > > dst[ret - 1] = '\0';
> > > > > @@ -173,7 +176,10 @@ fetch_store_strlen(unsigned long addr)
> > > > > int len;
> > > > > void __user *vaddr = (void __force __user *) addr;
> > > > > - len = strnlen_user(vaddr, MAX_STRING_SIZE);
> > > > > + if (addr == (unsigned long)current->comm)
> > > > > + len = strlen(current->comm);
> > > >
> > > > To balance with the strnlen_user, we must increse the len in this block.
> > > > (strlen doesn't count the final '\0', but strnlen_user counts it)
> > > >
> > >
> > > yes, we need to add a '+ 1' here.
> > >
> > > With the typos and this one fixed, this is
> > >
> > > Acked-by: Andreas Ziegler <andreas.ziegler@fau.de>
> >
> > Thank you for fixing typo and Ack :)
> >
> > Thanks you,
> >
> > >
> > > > Thank you,
> > > >
> > > > > + else
> > > > > + len = strnlen_user(vaddr, MAX_STRING_SIZE);
> > > > > return (len > MAX_STRING_SIZE) ? 0 : len;
> > > > > }
> > > > > @@ -213,6 +219,9 @@ process_fetch_insn(struct fetch_insn *code, struct pt_regs *regs, void *dest,
> > > > > case FETCH_OP_IMM:
> > > > > val = code->immediate;
> > > > > break;
> > > > > + case FETCH_OP_COMM:
> > > > > + val = (unsigned long)current->comm;
> > > > > + break;
> > > > > case FETCH_OP_FOFFS:
> > > > > val = translate_user_vaddr(code->immediate);
> > > > > break;
> > > >
> > > >
> > >
> >
> >
> as the original commit breaking $comm support was merged for v4.20 (which is
> a stable kernel) and the wrong behaviour with multiple strings exists in all
> longterm/stable releases (tested back to v4.4), do you think this should be
> going into a stable release once it's merged?
>
> I added Greg as he might know the answer to that.

Sounds like it would be a patch to be backported to the stable kernels
to me.

greg k-h

\
 
 \ /
  Last update: 2019-01-17 14:49    [W:0.071 / U:14.336 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site