As discussed with David earlier this week, the current implementation of watchpoints cannot work, at least reliably. We are raising an exception out of the middle of the i/o access path which does not even attempt to unwind the guest cpu state, nor does it have the information required to do so.
This moves the implementation to the cputlb helpers. This is a point at which we can and do raise exceptions properly. In addition, this fixes a bug in that unaligned stores were detecting watchpoints in the middle of the byte-by-byte operation, which means that we didn't signal the watchpoint early enough to avoid state change. r~ David Hildenbrand (2): exec: Factor out core logic of check_watchpoint() tcg: Check for watchpoints in probe_write() Richard Henderson (4): exec: Move user-only watchpoint stubs inline cputlb: Fold TLB_RECHECK into TLB_INVALID_MASK exec: Factor out cpu_watchpoint_address_matches cputlb: Handle watchpoints via TLB_WATCHPOINT include/exec/cpu-all.h | 8 +- include/hw/core/cpu.h | 37 +++++++++ accel/tcg/cputlb.c | 156 ++++++++++++++++++++++++-------------- exec.c | 167 +++++++++-------------------------------- 4 files changed, 173 insertions(+), 195 deletions(-) -- 2.17.1