Re: [PATCH 0/2] perf probe fixes for ppc64le

["Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com>]

On 2016/04/08 04:57PM, Balbir Singh wrote:
> On Thu, 2016-04-07 at 14:56 +0530, Naveen N. Rao wrote:
> > On 2016/04/07 06:19PM, Balbir Singh wrote:
> > > 
> > > 
> > > On 06/04/16 22:32, Naveen N. Rao wrote:
> > > > 
> > > > This patchset fixes three issues found with perf probe on ppc64le:
> > > > 1. 'perf test kallsyms' failure on ppc64le (reported by Michael
> > > > Ellerman). This was due to the symbols being fixed up during symbol
> > > > table load. This is fixed in patch 2 by delaying symbol fixup until
> > > > later.
> > > > 2. perf probe function offset was being calculated from the local entry
> > > > point (LEP), which does not match user expectation when trying to look
> > > > at function disassembly output (reported by Ananth N). This is fixed for
> > > > kallsyms in patch 1 and for symbol table in patch 2.
> > > I think the bit where the offset is w.r.t LEP when using a name, but w.r.t
> > > GEP when using function+offset can be confusing.
> > Thanks for your review!
> > 
> > The rationale for this is actually from the end-user perspective. The 
> > two use cases we are considering are:
> > 1. User just wants to probe at function entry point:
> > 	# perf probe _do_fork
> > 
> > In this case, the user most definitely needs the local entry point, 
> > without which the probe won't be hit. So, for this case, we 
> > automatically insert the probe at the LEP.
> > 
> > [We really only want to alter perf probe behavior in this case only, but 
> > we were incorrectly changing the behavior of perf with the below 
> > scenario as well.]
> > 
> > 2. User wants to probe at a specific location. In this case, the user 
> > most likely starts by looking at the function disassembly. For instance:
> > 	# objdump -S -d vmlinux.bak | grep -A100 \<_do_fork\>:
> > 	c0000000000b6a00 <_do_fork>:
> > 		      unsigned long stack_start,
> > 		      unsigned long stack_size,
> > 		      int __user *parent_tidptr,
> > 		      int __user *child_tidptr,
> > 		      unsigned long tls)
> > 	{
> > 	c0000000000b6a00:	f7 00 4c 3c 	addis   r2,r12,247
> > 	c0000000000b6a04:	00 86 42 38 	addi    r2,r2,-31232
> > 	c0000000000b6a08:	a6 02 08 7c 	mflr    r0
> > 	c0000000000b6a0c:	d0 ff 41 fb 	std     r26,-48(r1)
> > 	c0000000000b6a10:	26 80 90 7d 	mfocrf  r12,8
> > 	...<snip>...
> > 		if (!(clone_flags & CLONE_UNTRACED)) {
> > 	c0000000000b6a54:	e3 4f c7 7b 	rldicl. r7,r30,41,63
> > 	c0000000000b6a58:	2c 00 82 40 	bne     c0000000000b6a84 <_do_fork+0x84>
> > 			if (clone_flags & CLONE_VFORK)
> > 	c0000000000b6a5c:	e3 97 c8 7b 	rldicl. r8,r30,50,63
> > 	c0000000000b6a60:	a0 01 82 41 	beq     c0000000000b6c00 <_do_fork+0x200>
> > 	c0000000000b6a64:	20 00 20 39 	li      r9,32
> > 				trace = PTRACE_EVENT_VFORK;
> > 	c0000000000b6a68:	02 00 80 3b 	li      r28,2
> > 	c0000000000b6a6c:	10 02 4d e9 	ld      r10,528(r13)
> > 
> > If the user wants to probe at _do_fork+0x54, he'd do:
> > 	# perf probe _do_fork+0x54
> > 
> > With the earlier approach, we would insert the probe at _do_fork+0x5c 
> > (0x54 from the LEP) instead, which is incorrect.
> > 
> > In reality, user would probably just use debuginfo:
> > 	# perf probe -L _do_fork
> > 	<_do_fork@/root/linus/kernel/fork.c:0>
> > 	      0  long _do_fork(unsigned long clone_flags,
> > 			      unsigned long stack_start,
> > 			      unsigned long stack_size,
> > 			      int __user *parent_tidptr,
> > 			      int __user *child_tidptr,
> > 			      unsigned long tls)
> > 	      6  {
> > 			struct task_struct *p;
> > 	      8         int trace = 0;
> > 			long nr;
> > 		 
> > 			/*
> > 			 * Determine whether and which event to report to ptracer.  When
> > 			 * called from kernel_thread or CLONE_UNTRACED is explicitly
> > 			 * requested, no event is reported; otherwise, report if the event
> > 			 * for the type of forking is enabled.
> > 			 */
> > 	     17         if (!(clone_flags & CLONE_UNTRACED)) {
> > 	     18                 if (clone_flags & CLONE_VFORK)
> > 	     19                         trace = PTRACE_EVENT_VFORK;
> > 	     20                 else if ((clone_flags & CSIGNAL) != SIGCHLD)
> > 	     21                         trace = PTRACE_EVENT_CLONE;
> > 
> > 	# perf probe _do_fork:17
> > 
> > In this case, perf chooses the right address based on DWARF. The current 
> > patchset matches the behavior of perf without debuginfo with this.
> 
> 
> I agree what I worry is that perf probe _do_fork sets a breakpoint after
> perf probe _do_fork+0x4. I am not sure if there is an easy solution to
> the problem. 

I suppose this boils down to the quirkiness of ABIv2. Though, in 
reality, I don't think most users will notice. As I stated above, users 
will most likely start with the disassembly or debuginfo and this patch 
ensures there are actually no surprises there.

- Naveen