Nowadays, if user requests HW watchpoint to monitor a large memory area
or unaligned area, aarch64 GDB will split into multiple aligned areas,
and use multiple debugging registers to watch them. However, the
registers are not updated in a transaction way. GDBserver doesn't revert
updates in previous iterations if some debugging registers fail to update
due to some reason, like no free debugging registers available, in the
latter iteration. For example, if we have a char buf[34], and watch buf
in gdb,
(gdb) watch buf
Hardware watchpoint 2: buf
(gdb) c
Continuing.
infrun: clear_proceed_status_thread (Thread 13466)
infrun: proceed (addr=0xffffffffffffffff, signal=GDB_SIGNAL_DEFAULT)
infrun: step-over queue now empty
infrun: resuming [Thread 13466] for step-over
Sending packet: $m410838,22#35...Packet received: 00000000000000000000000000000000000000000000000000000000000000000000
infrun: skipping breakpoint: stepping past insn at: 0x400524
infrun: skipping breakpoint: stepping past insn at: 0x400524
Sending packet: $Z2,410838,22#80...Packet received: E01 <----- [1]
Packet Z2 (write-watchpoint) is supported
Sending packet: $Z0,
7fb7fe0a8c,4#43...Packet received: OK
Warning:
Could not insert hardware watchpoint 2.
Could not insert hardware breakpoints:
You may have requested too many hardware breakpoints/watchpoints.
GDB receives E01 for Z2 packet [1] but GDBserver updates the debugging
register status,
insert_point (addr=0x00410838, len=34, type=hw-write-watchpoint):
BREAKPOINTs:
BP0: addr=0x0, ctrl=0x00000000, ref.count=0
BP1: addr=0x0, ctrl=0x00000000, ref.count=0
BP2: addr=0x0, ctrl=0x00000000, ref.count=0
BP3: addr=0x0, ctrl=0x00000000, ref.count=0
BP4: addr=0x0, ctrl=0x00000000, ref.count=0
BP5: addr=0x0, ctrl=0x00000000, ref.count=0
WATCHPOINTs:
WP0: addr=0x410850, ctrl=0x00001ff5, ref.count=1
WP1: addr=0x410848, ctrl=0x00001ff5, ref.count=1
WP2: addr=0x410840, ctrl=0x00001ff5, ref.count=1
WP3: addr=0x410838, ctrl=0x00001ff5, ref.count=1
four debugging registers can not monitor 34-byte long area, so the last
iteration of updating debugging register state fails but previous
iterations succeed. This makes GDB think no HW watchpoint is inserted
but some debugging registers are used.
This problem was exposed by "watch buf" gdb.base/watchpoint.exp with
aarch64 GDBserver debugging arm 32-bit program. The buf is 30-byte long
but 4-byte aligned, and four debugging registers can't cover 34-byte
(extend 4 bytes to be 8-byte aligned) area. However, this problem
does exist on non-multi-arch debugging scenario as well.
This patch moves code in aarch64_linux_region_ok_for_hw_watchpoint to
aarch64_linux_region_ok_for_watchpoint in nat/aarch64-linux-hw-point.c.
Then, checks with aarch64_linux_region_ok_for_watchpoint, like what we
are doing in GDB. If the region is OK, call aarch64_handle_watchpoint.
Regression tested on aarch64 with both 64-bit program and 32-bit
program. Some fails in gdb.base/watchpoint.exp are fixed.
gdb:
2015-09-03 Yao Qi <yao.qi@linaro.org>
* aarch64-linux-nat.c (aarch64_linux_region_ok_for_hw_watchpoint):
Move code to aarch64_linux_region_ok_for_watchpoint. Call
aarch64_linux_region_ok_for_watchpoint.
* nat/aarch64-linux-hw-point.c (aarch64_linux_region_ok_for_watchpoint):
New function.
* nat/aarch64-linux-hw-point.h (aarch64_linux_region_ok_for_watchpoint):
Declare it.
gdb/gdbserver:
2015-09-03 Yao Qi <yao.qi@linaro.org>
* linux-aarch64-low.c (aarch64_insert_point): Call
aarch64_handle_watchpoint if aarch64_linux_region_ok_for_watchpoint
returns true.
+2015-09-03 Yao Qi <yao.qi@linaro.org>
+
+ * aarch64-linux-nat.c (aarch64_linux_region_ok_for_hw_watchpoint):
+ Move code to aarch64_linux_region_ok_for_watchpoint. Call
+ aarch64_linux_region_ok_for_watchpoint.
+ * nat/aarch64-linux-hw-point.c (aarch64_linux_region_ok_for_watchpoint):
+ New function.
+ * nat/aarch64-linux-hw-point.h (aarch64_linux_region_ok_for_watchpoint):
+ Declare it.
+
2015-09-02 Patrick Palka <patrick@parcs.ath.cx>
* gdb_obstack.h (obstack_strdup): Declare.
aarch64_linux_region_ok_for_hw_watchpoint (struct target_ops *self,
CORE_ADDR addr, int len)
{
- CORE_ADDR aligned_addr;
-
- /* Can not set watchpoints for zero or negative lengths. */
- if (len <= 0)
- return 0;
-
- /* Must have hardware watchpoint debug register(s). */
- if (aarch64_num_wp_regs == 0)
- return 0;
-
- /* We support unaligned watchpoint address and arbitrary length,
- as long as the size of the whole watched area after alignment
- doesn't exceed size of the total area that all watchpoint debug
- registers can watch cooperatively.
-
- This is a very relaxed rule, but unfortunately there are
- limitations, e.g. false-positive hits, due to limited support of
- hardware debug registers in the kernel. See comment above
- aarch64_align_watchpoint for more information. */
-
- aligned_addr = addr & ~(AARCH64_HWP_MAX_LEN_PER_REG - 1);
- if (aligned_addr + aarch64_num_wp_regs * AARCH64_HWP_MAX_LEN_PER_REG
- < addr + len)
- return 0;
-
- /* All tests passed so we are likely to be able to set the watchpoint.
- The reason that it is 'likely' rather than 'must' is because
- we don't check the current usage of the watchpoint registers, and
- there may not be enough registers available for this watchpoint.
- Ideally we should check the cached debug register state, however
- the checking is costly. */
- return 1;
+ return aarch64_linux_region_ok_for_watchpoint (addr, len);
}
/* Implement the "to_stopped_data_address" target_ops method. */
+2015-09-03 Yao Qi <yao.qi@linaro.org>
+
+ * linux-aarch64-low.c (aarch64_insert_point): Call
+ aarch64_handle_watchpoint if aarch64_linux_region_ok_for_watchpoint
+ returns true.
+
2015-08-27 Ulrich Weigand <Ulrich.Weigand@de.ibm.com>
* linux-low.c (check_stopped_by_breakpoint): Use
targ_type = raw_bkpt_type_to_target_hw_bp_type (type);
if (targ_type != hw_execute)
- ret =
- aarch64_handle_watchpoint (targ_type, addr, len, 1 /* is_insert */,
- state);
+ {
+ if (aarch64_linux_region_ok_for_watchpoint (addr, len))
+ ret = aarch64_handle_watchpoint (targ_type, addr, len,
+ 1 /* is_insert */, state);
+ else
+ ret = -1;
+ }
else
ret =
aarch64_handle_breakpoint (targ_type, addr, len, 1 /* is_insert */,
aarch64_num_bp_regs = 0;
}
}
+
+/* Return true if we can watch a memory region that starts address
+ ADDR and whose length is LEN in bytes. */
+
+int
+aarch64_linux_region_ok_for_watchpoint (CORE_ADDR addr, int len)
+{
+ CORE_ADDR aligned_addr;
+
+ /* Can not set watchpoints for zero or negative lengths. */
+ if (len <= 0)
+ return 0;
+
+ /* Must have hardware watchpoint debug register(s). */
+ if (aarch64_num_wp_regs == 0)
+ return 0;
+
+ /* We support unaligned watchpoint address and arbitrary length,
+ as long as the size of the whole watched area after alignment
+ doesn't exceed size of the total area that all watchpoint debug
+ registers can watch cooperatively.
+
+ This is a very relaxed rule, but unfortunately there are
+ limitations, e.g. false-positive hits, due to limited support of
+ hardware debug registers in the kernel. See comment above
+ aarch64_align_watchpoint for more information. */
+
+ aligned_addr = addr & ~(AARCH64_HWP_MAX_LEN_PER_REG - 1);
+ if (aligned_addr + aarch64_num_wp_regs * AARCH64_HWP_MAX_LEN_PER_REG
+ < addr + len)
+ return 0;
+
+ /* All tests passed so we are likely to be able to set the watchpoint.
+ The reason that it is 'likely' rather than 'must' is because
+ we don't check the current usage of the watchpoint registers, and
+ there may not be enough registers available for this watchpoint.
+ Ideally we should check the cached debug register state, however
+ the checking is costly. */
+ return 1;
+}
struct aarch64_debug_reg_state *aarch64_get_debug_reg_state (pid_t pid);
+int aarch64_linux_region_ok_for_watchpoint (CORE_ADDR addr, int len);
+
#endif /* AARCH64_LINUX_HW_POINT_H */
projects / binutils-gdb.git / commitdiff