[PATCH] gdb: Disable address space randomization for a test

[PATCH] gdb: Disable address space randomization for a test

[Andrew Burgess]

The gdb.opt/inline-break.exp test script relies on capturing a $pc
value in one run of GDB, then restarting GDB and placing a breakpoint
at the previously captured $pc and checking we get the expected
behaviour.

With address space randomization on then the captured $pc value is
almost never valid in the second run of GDB, and so the test fails.

Disabling address space randomization where appropriate fixes this
issue.

gdb/testsuite/ChangeLog:

	* gdb.opt/inline-break.exp: Disable address space randomization
	for the two runs where we care about addresses being consistent.
---
 gdb/testsuite/ChangeLog                |  5 +++++
 gdb/testsuite/gdb.opt/inline-break.exp | 12 ++++++++++++
 2 files changed, 17 insertions(+)

diff --git a/gdb/testsuite/gdb.opt/inline-break.exp b/gdb/testsuite/gdb.opt/inline-break.exp
index aed38ed631f..00a5467d617 100644
--- a/gdb/testsuite/gdb.opt/inline-break.exp
+++ b/gdb/testsuite/gdb.opt/inline-break.exp
@@ -264,6 +264,12 @@ set line [gdb_get_line_number "break here"]
 with_test_prefix "line number" {
     clean_restart $binfile
 
+    # At the end of this test we store a $pc value, then restart GDB,
+    # and use this $pc value to place a breakpoint.  If address space
+    # randomization is on, then the chance of this $pc value being
+    # correct in a second GDB instance is pretty small.
+    gdb_test "set disable-randomization off"
+
     if {![runto main]} {
 	untested "could not run to main"
 	continue
@@ -289,6 +295,12 @@ with_test_prefix "address" {
 
     clean_restart $binfile
 
+    # We're about to reuse the $pc value from a prevous run of GDB.
+    # Ensure address space randomization is off so that the inferior
+    # is placed at the same address as previous, and the stored $pc
+    # will still be valid.
+    gdb_test "set disable-randomization off"
+
     if {![runto main]} {
 	untested "could not run to main"
 	continue
-- 
2.14.4

Re: [PATCH] gdb: Disable address space randomization for a test

[Paul Koning]

> On Jul 30, 2018, at 2:50 PM, Andrew Burgess <andrew.burgess@embecosm.com> wrote:
> 
> The gdb.opt/inline-break.exp test script relies on capturing a $pc
> value in one run of GDB, then restarting GDB and placing a breakpoint
> at the previously captured $pc and checking we get the expected
> behaviour.
> 
> With address space randomization on then the captured $pc value is
> almost never valid in the second run of GDB, and so the test fails.
> 
> Disabling address space randomization where appropriate fixes this
> issue.
> 
> ...
> 
> +    # At the end of this test we store a $pc value, then restart GDB,
> +    # and use this $pc value to place a breakpoint.  If address space
> +    # randomization is on, then the chance of this $pc value being
> +    # correct in a second GDB instance is pretty small.
> +    gdb_test "set disable-randomization off"

Does that turn off address randomization?  The keyword implies the opposite (double negative, so "disable... off" means "turn on").

Also, does this work on every target OS?

	paul

Re: [PATCH] gdb: Disable address space randomization for a test

[Andrew Burgess]

* Paul Koning <paulkoning@comcast.net> [2018-07-30 14:56:32 -0400]:

> 
> 
> > On Jul 30, 2018, at 2:50 PM, Andrew Burgess <andrew.burgess@embecosm.com> wrote:
> > 
> > The gdb.opt/inline-break.exp test script relies on capturing a $pc
> > value in one run of GDB, then restarting GDB and placing a breakpoint
> > at the previously captured $pc and checking we get the expected
> > behaviour.
> > 
> > With address space randomization on then the captured $pc value is
> > almost never valid in the second run of GDB, and so the test fails.
> > 
> > Disabling address space randomization where appropriate fixes this
> > issue.
> > 
> > ...
> > 
> > +    # At the end of this test we store a $pc value, then restart GDB,
> > +    # and use this $pc value to place a breakpoint.  If address space
> > +    # randomization is on, then the chance of this $pc value being
> > +    # correct in a second GDB instance is pretty small.
> > +    gdb_test "set disable-randomization off"
> 
> Does that turn off address randomization?  The keyword implies the
> opposite (double negative, so "disable... off" means "turn on").

That's a very good question... now I just have to figure out why my
incorrect patch fixed the problem I was seeing.... watch this space...

> Also, does this work on every target OS?

That's a good question, I don't have an answer for you.

Thanks,
Andrew

Re: [PATCH] gdb: Disable address space randomization for a test

[Simon Marchi]

On 2018-07-30 15:01, Andrew Burgess wrote:
>> Does that turn off address randomization?  The keyword implies the
>> opposite (double negative, so "disable... off" means "turn on").
> 
> That's a very good question... now I just have to figure out why my
> incorrect patch fixed the problem I was seeing.... watch this space...

GDB usually disables address randomization by default, so I was also 
wondering why this would be needed.  Do you have a scenario where GDB 
doesn't disable it by default?

Simon

Re: [PATCH] gdb: Disable address space randomization for a test

[Andrew Burgess]

* Simon Marchi <simon.marchi@polymtl.ca> [2018-07-30 15:49:01 -0400]:

> On 2018-07-30 15:01, Andrew Burgess wrote:
> > > Does that turn off address randomization?  The keyword implies the
> > > opposite (double negative, so "disable... off" means "turn on").
> > 
> > That's a very good question... now I just have to figure out why my
> > incorrect patch fixed the problem I was seeing.... watch this space...
> 
> GDB usually disables address randomization by default, so I was also
> wondering why this would be needed.  Do you have a scenario where GDB
> doesn't disable it by default?

Thank you to you both for pointing me in the right direction.

It turned out the mistake was user error.  I've recently switched to a
jenkins/docker combo to try and streamline testing my patches.

When running GDB under docker I'm having to supply the '--privileged'
flag in order to allow GDB to function correctly.

The problem was that I use a separate docker instance to build GDB and
then a set of instances to test GDB in different modes (native,
gdbserver, etc).  When testing GDB I supplied the '--privileged', but
not when building GDB.  As a result configure decided that my system
didn't support the personality API, and so all the code to disable
address space randomization was not compiled into GDB.

As for why I failed to spot this, again user error, but far less
interesting, I forgot I pushed a change to the scripts I used for
comparing test results.

Anyway, sorry for the noise, and thanks for the help figuring out my
mistakes.

Thanks,
Andrew

Re: [PATCH] gdb: Disable address space randomization for a test

[Simon Marchi]

On 2018-07-31 08:47, Andrew Burgess wrote:
> * Simon Marchi <simon.marchi@polymtl.ca> [2018-07-30 15:49:01 -0400]:
> 
>> On 2018-07-30 15:01, Andrew Burgess wrote:
>> > > Does that turn off address randomization?  The keyword implies the
>> > > opposite (double negative, so "disable... off" means "turn on").
>> >
>> > That's a very good question... now I just have to figure out why my
>> > incorrect patch fixed the problem I was seeing.... watch this space...
>> 
>> GDB usually disables address randomization by default, so I was also
>> wondering why this would be needed.  Do you have a scenario where GDB
>> doesn't disable it by default?
> 
> Thank you to you both for pointing me in the right direction.
> 
> It turned out the mistake was user error.  I've recently switched to a
> jenkins/docker combo to try and streamline testing my patches.
> 
> When running GDB under docker I'm having to supply the '--privileged'
> flag in order to allow GDB to function correctly.
> 
> The problem was that I use a separate docker instance to build GDB and
> then a set of instances to test GDB in different modes (native,
> gdbserver, etc).  When testing GDB I supplied the '--privileged', but
> not when building GDB.  As a result configure decided that my system
> didn't support the personality API, and so all the code to disable
> address space randomization was not compiled into GDB.
> 
> As for why I failed to spot this, again user error, but far less
> interesting, I forgot I pushed a change to the scripts I used for
> comparing test results.
> 
> Anyway, sorry for the noise, and thanks for the help figuring out my
> mistakes.
> 
> Thanks,
> Andrew

Well, that's good information to know, thanks!

Simon

Re: [PATCH] gdb: Disable address space randomization for a test

[Tom Tromey]

Paul> Also, does this work on every target OS?

It doesn't.  The target has to implement the
supports_disable_randomization method, and then actually implement the
disabling.  At least macOS doesn't do this (see
https://sourceware.org/bugzilla/show_bug.cgi?id=23365), and probably
others as well.

Tom