Re: [PATCH] kernel-shark: Add check for return of ksmodel_get_cpu_front()

Yordan Karadzhov <y.karadz@xxxxxxxxx> · Tue, 11 May 2021 16:25:35 +0300

Hi Steven,

On 7.05.21 г. 16:38, Steven Rostedt wrote:

From: Steven Rostedt (VMware) <rostedt@xxxxxxxxxxx>

When I loaded two trace.dat files (host and guest), but the mapping of
which host thread is associated to which guest vCPU was missing from the
file, it caused a SEGFAULT. That's because in fillTaskGraph(), the
lamGetPidCPU() calls ksmodule_get_cpu_front() which returns a negative
number and does not set index. But the next line checks data[index] where
index is some random number, and the application crashes.

By checking the return of ksmodule_get_cpu_front(), and if it is negative
do not reference data[index] and just let eFront be nullptr, the
application shows no mapping, but at least it does not crash.

Signed-off-by: Steven Rostedt (VMware) <rostedt@xxxxxxxxxxx>
---
You can reproduce this with:

    http://rostedt.org/private/trace-no-map.tar.bz2

diff --git a/src/KsPlotTools.cpp b/src/KsPlotTools.cpp
index 225dc34..abef5f8 100644
--- a/src/KsPlotTools.cpp
+++ b/src/KsPlotTools.cpp
@@ -1280,7 +1280,7 @@ void Graph::fillTaskGraph(int sd, int pid)
  						 false,
  						 _collectionPtr,
  						 &index);



I wonder why this fails? Is it because "pid" is negative? The fix below 

is appropriate, however we should detect negative PIDs and abort 

plotting much earlier.


Thanks!
Yordan


-		if (index >= 0)
+		if (cpuFront >= 0 && index >= 0)
  			eFront = _histoPtr->data[index];

  

  		cpuBack = ksmodel_get_cpu_back(_histoPtr, bin,





