gdb/aarch64: restore in-order watchpoint matchingusers/aburgess/try-watchpoint-api-changes

At Red Hat we have an out of tree AArch64 watchpoint test which broke
after this commit:

  commit cf16ab724a41e4cbaf723b5633d4e7b29f61372b
  Date:   Tue Mar 12 17:08:18 2024 +0100

      [gdb/tdep] Fix gdb.base/watch-bitfields.exp on aarch64

The problem with AArch64 hardware watchpoints is that they (as I
understand it) are restricted to a minimum of 8 bytes.  This means
that, if the thing you are watching is less than 8-bytes, then there
is always scope for invalid watchpoint triggers caused by activity in
the part of the 8-bytes that are not being watched.

Or, as is the case in this RH test, multiple watchpoint are created
within an 8-byte region, and GDB can miss-identify which watchpoint
actually triggered.

Prior to the above commit the RH test was passing.  However, the test
was relying on, in the case of ambiguity, GDB selecting the first
created watchpoint.  That behaviour changed with the above commit.
Now GDB favours reporting non write breakpoints, and will only report
a write breakpoint if no non-write breakpoint exists in the same
region.

I originally posted a patch to try and tweak the existing logic to
restore enough of the original behaviour that the RH test would pass,
this can be found here (2 iterations):

  https://inbox.sourceware.org/gdb-patches/65e746b6394f04faa027e778f733eda95d20f368.1753115072.git.aburgess@redhat.com
  https://inbox.sourceware.org/gdb-patches/638cbe9b738c0c529f6370f90ba4a395711f63ae.1753971315.git.aburgess@redhat.com

Neither of these really resolved the problem, they fixed some cases,
but broke others.

Ultimately, the problem on AArch64 is that for a single watchpoint
trap, there could be multiple watchpoints that are potentially
responsible.  The existing API defined by the target_ops methods
stopped_by_watchpoint() and stopped_data_address() only allow for two
possible options:

  1. If stopped_by_watchpoint() is true then stopped_data_address()
     can return true and a single address which identifies all
     watchpoints at that single address, or

  2. If stopped_by_watchpoint() is true then stopped_data_address()
     can return false, in which case GDB will check all write
     watchpoints to see if any have changed, if they have, then GDB
     tells the user that that was the triggering watchpoint.

If we are in a situation where we have to choose between multiple
write and read watchpoints then the current API doesn't allow the
architecture specific code to tell GDB core about this case.

In this commit I propose that we change the target_ops API,
specifically, the method:

  bool target_ops::stopped_data_address (CORE_ADDR *);

will change to:

  std::vector<CORE_ADDR> target_ops::stopped_data_addresses ();

The architecture specific code can now return a set of watchpoint
addresses, allowing GDB to identify a set of watchpoints that might
have triggered.  GDB core can then select the most likely watchpoint,
and present that to the user.

As with the old API, target_ops::stopped_data_addresses should only be
called when target_ops::stopped_by_watchpoint is true, in which case
it's return values can be interpreted like this:

  a. An empty vector; this replaces the old case where false was
     returned.  GDB should check all the write watchpoints and select
     the one that changed as the responsible watchpoint.

  b. A single entry vector; all targets except AArch64 currently
     return at most a single entry vector.  The single address
     indicates the watchpoint(s) that triggered.

  c. A multi-entry vector; currently AArch64 only.  These addresses
     indicate the set of watchpoints that might have triggered.  GDB
     will check the write watchpoints to see which (if any) changed,
     and if no write watchpoints changed, GDB will present the first
     access watchpoint.

In the future, we might want to improve the handling of (c) so that
GDB tells the user that multiple access watchpoints might have
triggered, and then list all of them.  This might clear up some
confusion.  But I think that can be done in the future (I don't have
an immediate plan to work on this).  I think this change is already a
good improvement.

The changes for this are pretty extensive, but here's a basic summary:

  * Within gdb/ changing the API name from stopped_data_address to
    stopped_data_addresses throughout.  Comments are updated too where
    needed.

  * For targets other than AArch64, the existing code is retained with
    as few changes as possible, we only allow for a single address to
    be returned, the address is now wrapped in a vector.  Where we
    used to return false, we now return the empty vector.

  * For AArch64, the return a vector logic is pushed through to
    gdb/nat/aarch64-hw-point.{c,h}, and aarch64_stopped_data_address
    changes to aarch64_stopped_data_addresses, and is updated to
    return a vector of addresses.

  * In infrun.c there's some updates to some debug output.

  * In breakpoint.c the interesting changes are in
    watchpoints_triggered.  The existing code has three cases to
    handle:

    (i) target_stopped_by_watchpoint returns false.  This case is
        unchanged.

    (ii) target_stopped_data_address returns false.  This case is now
         calling target_stopped_data_addresses, and checks for the
	 empty vector, but otherwise is unchanged.

    (iii) target_stopped_data_address returns true, and a single
          address.  This code calls target_stopped_data_addresses, and
	  now handles the possibility of a vector containing multiple
	  entries.  We need to first loop over every watchpoint
	  setting its triggered status to 'no', then we check every
	  address in the vector setting matching watchpoint's
	  triggered status to 'yes'.  But the actual logic for if a
	  watchpoint matches an address or not is unchanged.

    The important thing to notice here is that in case (iii), before
    this patch, GDB could already set _multiple_ watchpoints to
    triggered.  For example, setting a read and write watchpoint on
    the same address would result in multiple watchpoints being marked
    as triggered.  This patch just extends this so that multiple
    watchpoints, at multiple addresses, can now be marked as
    triggered.

  * In remote.c there is an interesting change.  We need to allow
    gdbserver to pass the multiple addresses back to GDB.  To achieve
    this, I now allow multiple 'watch', 'rwatch', and 'awatch' tokens
    in a 'T' stop reply packet.  This change is largely backward
    compatible.  For old versions of GDB, GDB will just use the last
    such token as the watchpoint stop address.  For new GDBs, all of
    the addresses are collected and returned from the
    target_ops::stopped_data_addresses call.  If a new GDB connects to
    an old gdbserver then it'll only get a single watchpoint address
    in the 'T' packet, but that's no worse than we are now, and will
    not cause a GDB crash, GDB will just end up checking a restricted
    set of watchpoints (which is where we are right now).

  * In gdbserver/ the changes are pretty similar.  The API is renamed
    from ::stopped_data_address to ::stopped_data_addresses, and
    ::low_stopped_data_address to ::low_stopped_data_addresses.

  * For all targets except AArch64, the existing code is retained, we
    just wrap the single address into a vector.

  * For AArch64, we call aarch64_stopped_data_addresses, which returns
    the required vector.

For testing, I've built GDB on GNU/Linux for i386, x86-64, PPC64le,
ARM, and AArch64.  That still leaves a lot of targets possibly
impacted by this change as untested.  Which is a risk.  I certainly
wouldn't want to push this patch until after GDB 17 branches so we
have time to find and fix any regressions that are introduced.

Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=33240
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=33252

1 files changed, 18 insertions, 8 deletions

diff --git a/gdb/infrun.c b/gdb/infrun.c
index 9d3e1b7..32c6a34 100644
--- a/gdb/infrun.c
+++ b/gdb/infrun.c
@@ -6889,16 +6889,26 @@ handle_signal_stop (struct execution_control_state *ecs)
 	("stop_pc=%s", paddress (reg_gdbarch, ecs->event_thread->stop_pc ()));
       if (target_stopped_by_watchpoint ())
 	{
-	  CORE_ADDR addr;
+	  auto inf_target = current_inferior ()->top_target ();
+	  std::vector<CORE_ADDR> addr_list
+	    = target_stopped_data_addresses (inf_target);
 
-	  infrun_debug_printf ("stopped by watchpoint");
-
-	  if (target_stopped_data_address (current_inferior ()->top_target (),
-					   &addr))
-	    infrun_debug_printf ("stopped data address=%s",
-				 paddress (reg_gdbarch, addr));
+	  std::string addr_str;
+	  if (addr_list.empty ())
+	    addr_str = "(no data addressses available)";
 	  else
-	    infrun_debug_printf ("(no data address available)");
+	    {
+	      for (const CORE_ADDR addr : addr_list)
+		{
+		  if (addr_str.length () > 0)
+		    addr_str += ", ";
+
+		  addr_str += paddress (reg_gdbarch, addr);
+		}
+	    }
+
+	  infrun_debug_printf ("stopped by watchpoint, data addresses = %s",
+			       addr_str.c_str ());
 	}
     }