Patchwork mem-break: Fix breakpoint insertion location

login
register
mail settings
Submitter Maciej W. Rozycki
Date Aug. 1, 2017, 4:36 p.m.
Message ID <alpine.DEB.2.00.1708011707230.29991@tp.orcam.me.uk>
Download mbox | patch
Permalink /patch/21840/
State Superseded
Headers show

Comments

Maciej W. Rozycki - Aug. 1, 2017, 4:36 p.m.
Fix a commit cd6c3b4ffc4e ("New gdbarch methods breakpoint_kind_from_pc 
and sw_breakpoint_from_kind") regression and restore the use of 
->placed_size rather than ->reqstd_address as the location for a memory 
breakpoint to be inserted at.  Previously `gdbarch_breakpoint_from_pc' 
was used that made that adjustment in `default_memory_insert_breakpoint' 
from the preinitialized value, however with the said commit that call is 
gone, so the passed ->placed_size has to be used for the initialization.

The regression manifests itself as the inability to debug any MIPS/Linux 
compressed ISA dynamic executable as GDB corrupts the dynamic loader 
with one of its implicit breakpoints, causing the program to crash, as 
seen for example with the `mips-linux-gnu' target, o32 ABI, MIPS16 code, 
and the gdb.base/advance.exp test case:

(gdb) continue
Continuing.

Program received signal SIGBUS, Bus error.
_dl_debug_initialize (ldbase=0, ns=0) at dl-debug.c:51
51	    r = &_r_debug;
(gdb) FAIL: gdb.base/advance.exp: Can't run to main

	gdb/
	* mem-break.c (default_memory_insert_breakpoint): Use 
	`->placed_address' rather than `->reqstd_address' for the 
	breakpoint location.
---
Hi,

 No regressions between plain commit cd6c3b4ffc4e^ and commit cd6c3b4ffc4e 
with this change applied in `mips-linux-gnu', o32, MIPS16 testing.  This 
brings that configuration back to sanity.

 OK for master and (as a grave regression) for 8.0?

  Maciej

---
 gdb/mem-break.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

gdb-mem-break-placed-address.diff
Maciej W. Rozycki - Aug. 2, 2017, 5:53 p.m.
On Tue, 1 Aug 2017, Maciej W. Rozycki wrote:

>  OK for master and (as a grave regression) for 8.0?

 This is now PR breakpoints/21886 as per release branch requirements.

  Maciej
Simon Marchi - Aug. 4, 2017, 1:17 p.m.
On 2017-08-01 18:36, Maciej W. Rozycki wrote:
> Fix a commit cd6c3b4ffc4e ("New gdbarch methods breakpoint_kind_from_pc
> and sw_breakpoint_from_kind") regression and restore the use of
> ->placed_size rather than ->reqstd_address as the location for a memory
> breakpoint to be inserted at.  Previously `gdbarch_breakpoint_from_pc'
> was used that made that adjustment in 
> `default_memory_insert_breakpoint'
> from the preinitialized value, however with the said commit that call 
> is
> gone, so the passed ->placed_size has to be used for the 
> initialization.
> 
> The regression manifests itself as the inability to debug any 
> MIPS/Linux
> compressed ISA dynamic executable as GDB corrupts the dynamic loader
> with one of its implicit breakpoints, causing the program to crash, as
> seen for example with the `mips-linux-gnu' target, o32 ABI, MIPS16 
> code,
> and the gdb.base/advance.exp test case:
> 
> (gdb) continue
> Continuing.
> 
> Program received signal SIGBUS, Bus error.
> _dl_debug_initialize (ldbase=0, ns=0) at dl-debug.c:51
> 51	    r = &_r_debug;
> (gdb) FAIL: gdb.base/advance.exp: Can't run to main
> 
> 	gdb/
> 	* mem-break.c (default_memory_insert_breakpoint): Use
> 	`->placed_address' rather than `->reqstd_address' for the
> 	breakpoint location.
> ---
> Hi,
> 
>  No regressions between plain commit cd6c3b4ffc4e^ and commit 
> cd6c3b4ffc4e
> with this change applied in `mips-linux-gnu', o32, MIPS16 testing.  
> This
> brings that configuration back to sanity.
> 
>  OK for master and (as a grave regression) for 8.0?
> 
>   Maciej
> 
> ---
>  gdb/mem-break.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> gdb-mem-break-placed-address.diff
> Index: binutils/gdb/mem-break.c
> ===================================================================
> --- binutils.orig/gdb/mem-break.c	2017-07-30 22:45:34.000000000 +0100
> +++ binutils/gdb/mem-break.c	2017-07-30 23:41:28.595612206 +0100
> @@ -37,7 +37,7 @@ int
>  default_memory_insert_breakpoint (struct gdbarch *gdbarch,
>  				  struct bp_target_info *bp_tgt)
>  {
> -  CORE_ADDR addr = bp_tgt->reqstd_address;
> +  CORE_ADDR addr = bp_tgt->placed_address;
>    const unsigned char *bp;
>    gdb_byte *readbuf;
>    int bplen;

IIUC, we end up writing the good breakpoint kind, but at the wrong 
address?  For example, if the requested address is 0x1001, it means that 
there should be a micro/compressed MIPS breakpoint at address 0x1000, 
but that bug caused the breakpoint to be written at address 0x1001 
instead.  Is that right?

If so, I think the patch makes sense, I think Yao should have the final 
say.

Thanks,

Simon
Yao Qi - Aug. 4, 2017, 1:24 p.m.
"Maciej W. Rozycki" <macro@imgtec.com> writes:

> Fix a commit cd6c3b4ffc4e ("New gdbarch methods breakpoint_kind_from_pc 
> and sw_breakpoint_from_kind") regression and restore the use of 
> ->placed_size rather than ->reqstd_address as the location for a memory 

s/placed_size/placed_address/

The patch looks good to me, but please give me two or three days to run
the tests on an armv7 board.  The board is being used for other tests,
and I'll start the regression test on Monday next week.
Maciej W. Rozycki - Aug. 4, 2017, 1:57 p.m.
On Fri, 4 Aug 2017, Simon Marchi wrote:

> On 2017-08-01 18:36, Maciej W. Rozycki wrote:
> > Fix a commit cd6c3b4ffc4e ("New gdbarch methods breakpoint_kind_from_pc
> > and sw_breakpoint_from_kind") regression and restore the use of
> > ->placed_size rather than ->reqstd_address as the location for a memory
> > breakpoint to be inserted at.  Previously `gdbarch_breakpoint_from_pc'
> > was used that made that adjustment in `default_memory_insert_breakpoint'
> > from the preinitialized value, however with the said commit that call is
> > gone, so the passed ->placed_size has to be used for the initialization.
[...]
> IIUC, we end up writing the good breakpoint kind, but at the wrong address?
> For example, if the requested address is 0x1001, it means that there should be
> a micro/compressed MIPS breakpoint at address 0x1000, but that bug caused the
> breakpoint to be written at address 0x1001 instead.  Is that right?

 Exactly!

 Moreover, as the breakpoint is removed the original instruction bytes 
will be written back to 0x1000, further corrupting the executable, as 
`default_memory_remove_breakpoint' already correctly uses 
`->placed_address'.

 I can see now that I incorrectly wrote `->placed_size' across the patch 
description where I meant `->placed_address'.  I'll correct that and 
repost the patch with PR annotation additionally included.

  Maciej

Patch

Index: binutils/gdb/mem-break.c
===================================================================
--- binutils.orig/gdb/mem-break.c	2017-07-30 22:45:34.000000000 +0100
+++ binutils/gdb/mem-break.c	2017-07-30 23:41:28.595612206 +0100
@@ -37,7 +37,7 @@  int
 default_memory_insert_breakpoint (struct gdbarch *gdbarch,
 				  struct bp_target_info *bp_tgt)
 {
-  CORE_ADDR addr = bp_tgt->reqstd_address;
+  CORE_ADDR addr = bp_tgt->placed_address;
   const unsigned char *bp;
   gdb_byte *readbuf;
   int bplen;