Function armv4_5_mmu_translate_va() now properly signals
errors in the return value.
Remove former error handling by setting variable "type" to
value "-1".
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Function arm920t_write_memory() default return value
should be ERROR_OK.
All cases of local errors are handled immediately and
not further propagated.
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Commit 0538081246
introduces a compile time warning:
arm920t.c: In function ‘arm920t_write_memory’:
arm920t.c:567: warning: ‘retval’ may be used uninitialized in this function
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
ETM analyze produced no output when the trace buffer was empty.
This patch provides users with a clue.
Signed-off-by: Jon Povey <jon.povey@racelogic.co.uk>
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
We request a id register read at the end of ahbap_debugport_init
but we never actually run the queue. In some cases this causes a
segfault.
Signed-off-by: Spencer Oliver <ntfreak@users.sourceforge.net>
This second half of the patch is proposed to clean up some GDB keep alive
issues on arm7_9 targets that start up with very slow clocks. If an attempt
is made to write to key registers on the processor with a slow jtag speed,
GDB timeout warnings appear on the console (at least mine) when "reset halt"
or "reset init" commands are issued from the gdb client:
*** BEFORE PATCH ***
(gdb) monitor reset init
fast memory access is disabled
2 kHz
keep_alive() was not invoked in the 1000ms timelimit. GDB alive packet not
sent! (1026). Workaround: increase "set remotetimeout" in GDB
JTAG tap: at91sam9g20.cpu tap/device found: 0x0792603f (mfg: 0x01f, part:
0x7926, ver: 0x0)
target state: halted
target halted in ARM state due to breakpoint, current mode: Supervisor
cpsr: 0x000000d3 pc: 0x00000000
MMU: disabled, D-Cache: disabled, I-Cache: disabled
keep_alive() was not invoked in the 1000ms timelimit. GDB alive packet not
sent! (1027). Workaround: increase "set remotetimeout" in GDB
keep_alive() was not invoked in the 1000ms timelimit. GDB alive packet not
sent! (1006). Workaround: increase "set remotetimeout" in GDB
keep_alive() was not invoked in the 1000ms timelimit. GDB alive packet not
sent! (1006). Workaround: increase "set remotetimeout" in GDB
keep_alive() was not invoked in the 1000ms timelimit. GDB alive packet not
sent! (1006). Workaround: increase "set remotetimeout" in GDB
keep_alive() was not invoked in the 1000ms timelimit. GDB alive packet not
sent! (1004). Workaround: increase "set remotetimeout" in GDB
RCLK - adaptive
dcc downloads are enabled
fast memory access is enabled
NAND flash device 'NAND 256MiB 3,3V 8-bit' found
(gdb)
I added additional keep alive steps in areas that troubleshooting revealed
were causing problems. I only did this however for non-fast write memory
accesses. I don't think most people would be using fast memory accesses to
write to memory when the jtag and system clocks are slow anyway.
If you disagree with my feeling, think there is a more elegant way to handle
the problem, or think the patch will cause other unforeseen problems with
other targets, let me know. As you can see below, the patch does eliminate
the problem on my development station and I suspect that it will benefit
others.
*** AFTER PATCH ***
(gdb) monitor reset init
fast memory access is disabled
2 kHz
JTAG tap: at91sam9g20.cpu tap/device found: 0x0792603f (mfg: 0x01f, part:
0x7926, ver: 0x0)
target state: halted
target halted in ARM state due to breakpoint, current mode: Supervisor
cpsr: 0x000000d3 pc: 0x00000000
MMU: disabled, D-Cache: disabled, I-Cache: disabled
RCLK - adaptive
dcc downloads are enabled
fast memory access is enabled
NAND flash device 'NAND 256MiB 3,3V 8-bit' found
(gdb)
Gary Carlson
Gary Carlson, MSEE
Principal Engineer
Carlson-Minot Inc.
There are a million reasons why cached protection state might
be stale: power cycling of target, reset, code executing on
the target, etc.
The "flash protect_check" command is now gone. This is *always*
executed when running a "flash info".
As a bonus for more a more robust approach, lots of code could
be deleted.
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
Various commands, e.g. "arm mcr xxxx" would fail if invoked upon startup
since it there was no command context defined for the jim interpreter
in that case.
A Jim interpreter is now associated with a command context(telnet,
gdb server's) or the default global command context.
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
target memory allocation can be implemented not to show
bogus error messages.
E.g. when trying a big allocation first and then a
smaller one if that fails.
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
This patch adds support for the length argument to the xscale implementation of
the wp command. Per discussion with David, the length argument specifies the
range of addresses over which a memory access should generate a debug exception.
This patch utilizes the "mask" feature of the xscale debug hardware to implement
the correct functionality of the length argument. Some limitations imposed by
the hardware are:
- The length must be a power of two, with a minumum of 4.
- Two data breakpoint registers are available, allowing for two watchpoints.
However, if the length of a watchpoint is greater than four, both registers
are used (the second for a mask value), limiting the number of watchpoints
to one.
This patch also removes a useless call to xscale_get_reg(dbcon) in
xscale_set_watchpoint() (value had already been read from the register cache,
and the same previously read value is then modified and written back).
I have been using and testing this patch for a couple days.
Questions, corrections, criticisms of course gratefully received.
This patch fixes the xscale_analyze_trace() function. This function was
defective for a trace collected in 'fill' mode (hiccups with repeated
instructions) and completely broken when buffer overflowed in 'wrap' mode. The
reason for the latter case is that the checkpoint registers were interpreted
incorrectly when two checkpoints are present in the trace (which will be true in
'wrap' mode once the buffer fills). In this case, checkpoint1 register will
contain the older entry, and checkpoint0 the newer. The original code assumed
the opposite. I eventually gave up trying to understand all the logic of the
function, and rewrote it. I think it's much cleaner and understandable now. I
have been using and testing this for a few weeks now. I'm confident it hasn't
regressed in any way.
Also added capability to handle (as best as possible) the case where an
instruction can not be read from the loaded trace image; e.g., partial image.
This was a 'TODO' comment in the original xscale_analyze_trace().
Outside of xcsale_analyze_trace(), these (related) changes were made:
- Remove pc_ok and current_pc elements from struct xscale_trace. These elements
and associated logic are useless clutter because the very first entry placed
in the trace buffer is always an indirect jump to the address at which
execution resumed. This type of trace entry includes the literal address in
the trace buffer, so the initial address of the trace is immediately
determined from the trace buffer contents and does not need to be recorded
when trace is enabled.
- Added num_checkpoints to struct xscale_trace_data, which is necessary in order
to correctly interpret the checkpoint register contents.
- In xscale_read_trace()
- Fix potential array out-of-bounds condition.
- Eliminate partial address entries when parsing trace (can occur in wrap mode).
- Count and record number of checkpoints in trace.
- Added small, inlined utility function xscale_display_instruction() to help
make the code more concise and clear.
TODO:
- Save processor state (arm or thumb) in struct xscale_trace when trace is
enabled so that trace can be analyzed correctly (currently assumes arm mode).
- Add element to struct xscale_trace that records (when trace is enabled)
whether vector table is relocated high (to 0xffff0000) or not, so that a
branch to an exception vector is traced correctly (curently assumes vectors
at 0x0).
+ virt2phys() can now convert virtual address to real
+ read_memory() and write_memory() are renamed to read_phys_memory()
and write_phys_memory()
+ new read_memory() and write_memory() try to resolve real address if
mmu is enambled than perform real address reading/writing
+ if address is bellow 0xc000000 than TTB0 is used for page table
dereference, if above - than TTB1. Linux style of user/kernel address
separation
+ if above fails (i.e address is unspecified) than mode is checked
whether it is Supervisor (than TTB1) or User (than TTB0)
- Software breakpoints doesn't work. You should invoke
"gdb_breakpoint_override hard" before you start debugging
+ cortex_a8_mmu(), cortex_a8_enable_mmu_caches(),
cortex_a8_disable_mmu_caches() are implemented
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
I'm not sure what caused this significant character to get deleted.
it may be related to intermittent Editor or terminal flakes I've
been seeing lately (sigh). This fix is trivial.
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Fixing one bug can easily uncover another .... in this case,
making sure that we properly invalidate some cached NOR state when
resuming arbitrary target code turned up an issue when the code
wasn't quite arbitrary (and we couldn't know that, but some parts
of OpenOCD assumed the cache would not be invalidated.
Specifically: some flash drivers (like CFI) update that state in loops
with downloaded algorithms, thus invalidating the state as it's probed.
+ Add a new target state flag, to record whether the target is
running downloaded algorithm code.
+ Use that flag to add a special case: "trust" downloaded algorithms
not to corrupt that cached state, bypassing cache invalidation.
Also update some of the documentation to stipulate that this flavor of
trustworthiness is now *required* ... not just a fortuitous acident.
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
For some reason there are *two* schemes for interposing logic into
the run_algorithm() code path... One is a standard procedural wapper
around the target method invocation.
the other (superfluous) one hacked the method table by splicing
a second procedural wrapper into the method table. Remove it:
* Rename its slightly-more-featureful wrapper so it becomes
the standard procedural wrapper, leaving its added logic
(where it should have been in the first place.
Also add a paranoia check, to report targets that don't
support algorithms without traversing a NULL pointer, and
tweak its code structure a bit so it's easier to modify.
* Get rid of the superfluous/conusing method table hacks.
This is a net simplification, making it simpler to analyse what's
going on, and then interpose logic . ... by ensuring there's only one
natural place for it to live.
------------
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Problem: halt at a breakpoint, enable trace buffer ('xscale trace_buffer enable
fill'), then resume. Wait for debug exception when trace buffer fills (if not
sooner due to another breakpoint, vector catch, etc). Instead, never halts.
When halted explicitly from OpenOCD and trace buffer dumped, it contains only
one entry; a branch to the address of the original breakpoint. If the above
steps are repeated, except that the breakpoint is removed before resuming, the
trace buffer fills and the debug exception is generated, as expected.
Cause: related to how a breakpoint is stepped over on resume. The breakpoint is
temporarily removed, and a hardware breakpoint is set on the next instruction
that will execute. xscale_debug_entry() is called when that breakpoint hits.
This function checks if the trace buffer is enabled, and if so reads the trace
buffer from the target and then disables the trace (unless multiple trace
buffers are specified by the user when trace is enabled). Thus you only trace
one instruction before it is disabled.
Solution: kind of a hack on top of a hack, but it's simple. Anything better
would involve some refactoring. This has been tested and trace now works as
intended, except that the very first instruction is not part of the trace when
resuming from a breakpoint.
TODO: still many issues with trace: doesn't work during single-stepping (trace
buffer is flushed each step), 'xscale analyze_trace' works only marginally for
a trace captured in 'fill' mode, and not at all for a trace captured in 'wrap'
mode.
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
I don't know when "poll off" broke, but "poll off" didn't
stop background polling of target. The polling status flag
simply wasn't checked in the handle_target timer callback.
All target polling(including power/reset state) is now stopped
upon "poll off".
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
This patch fixes xscale software breakpoints by cleaning the dcache and
invalidating the icache after the bkpt instruction is inserted or removed. The
icache operation is necessary in order to flush the fetch buffers, even if the
icache is disabled (see section 4.2.7 of the xscale core developer's manual).
The dcache is presumed to be enabled; no harm done if not. The dcache is also
invalidated after cleaning in order to safeguard against a future load of
invalid data, in the event that cache_clean_address points to memory that is
valid and in use.
Also corrected a confusing typo I noticed in a comment.
TODO (or not TODO...?): the xscale's 2K "mini dcache" is not cleaned. This
cache is not used unless the 'X' bit in the page table entry is set. This is a
proprietary xscale extension to the ARM architecture. If a target's OS or
executive makes use of this for memory regions holding code, the breakpoint
problem will persist. Flushing the mini dcache requires that 2K of valid
cacheable memory (mapped with 'X' bit set) be designated by the user for this
purpose. The debug handler that gets downloaded to the target will also need to
be extended.
the handling of caches, should be moved into the breakpoint
specific callbacks rather than being plonked into generic
memory write fn's.
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
By a bit of code inspection it seems like all of these
instances of jtag_get_end_state() can be unambigously
replaced by constants.
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
Fix problem with the xscale icache and dcache commands. Both commands were
enabling or disabling the mmu, not the caches
I didn't look any further after my earlier patch fixed the trivial problem
with command argument parsing. Turns out the underlying code was broken.
The resolution is straightforward when you look at the arguments to
xscale_enable_mmu_caches() and xscale_disable_mmu_caches(). I finally
took a deeper look after dumping the cp15 control register (XSCALE_CTRL)
and seeing that the cache bits weren't changing, but the mmu bit was
(which caused all manner of grief, as you can imagine). This has been
tested and works OK now.
src/target/xscale.c | 17 +++++++++++------
1 files changed, 11 insertions(+), 6 deletions(-)
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
