Previously it might read an address multiple times if an abstract
command took longer to execute than expected.
The new implementations reads from the target how far it has gotten
along reading memory, and resumes from there if cmderr=busy.
This ended up being a bigger change than I envisioned, but in the end it
deleted more lines than it added, so I'm happy. :-)
The downloaded program now post-increments, and there's no longer an
attempt to read the current address from the target. This made it easier
to fix the problem where at the start of the loop the current address
was already read (in regular entry) or has not yet been read (when the
first round through the loop encountered busy more than once, or busy
was encountered at least once later on).
The interesting new code concerns ignore_prev_addr and
this_is_last_read.
Additionally, I tweaked some debug output, and optimized
riscv_batch_run() when the batch is empty.
The actual implementation of triggers didn't change between those two
versions, so there's no need to duplicate the code.
In the process, I also fixed a minor multicore bug where tselect didn't
always get written on all harts.
This is the only way the spec guarantees that GPRs are accessible, and
depending on the implementation this might be the only way that CSRs are
accessible.
Also changed the debug code that parses out DMI fields to be simpler to
maintain (albeit a little slower).
riscv013_execute_debug_buffer() now automatically clears cmderr if the
command fails. That feels like the right behavior. (It does return the
error to its caller.)
When first connecting to a target, have the debugger disable any
hardware triggers that are set by a previously connected debugger.
The 0.11 code already did this, but 0.13 did not.
To achieve this I decided to share the code to enumerate triggers
between 0.11 and 0.13, which required me to implement get_register() and
set_register() for 0.11, which made the whole change a lot larger than
you might have guessed.
Hopefully this sets us up to in the future share the code to set/remove
triggers as well.
They were implemented, and people want to keep using them.
Also make OpenOCD tolerate cores that have $misa at 0xf10 instead of the
current address of 0x301.
Actually return an error when we fail to read a CSR.
Tweak cache_set32() debug output.
Now that we're supporting non-RTOS multi-hart mode there's some more
assertions that you're running on the right hart. Those assertions
aren't sane very early in examine, so I avoid them.
Rather than having a bunch of "if rtos" stuff, I now just check "if
hart_enabled". This makes some code paths cleaner, all of which were
buggy in the non-RTOS multi-hart mode.
When I disappeared the polls everywhere I forgot to sanitize the hartid
after halting. This is an invariant that GDB expects: when you return
from a halt whatever thread is marked as currently selected is the
thread that the next register accesses reference.
Main change is to make riscv_addr_t be unsigned. The rest is mechanical
fixing of types, print statements, and a few signed/unsigned compares.
Smoketest indicates everything is working more or less as before.
This means I don't know what hart to look at, so I might as well
invalidate the register cache. Without this, you might get stale
registers the first time you ask for them.
I thought OpenOCD did this, but it looks like that doesn't happen when
runningi in RTOS mode. With this I can get to the end of most of the
RTOS tests, but they SIGINT instead of exiting.
I'm not sure why this is necessary, but for some reason GDB is asking
for registers before OpenOCD thinks there's been a halt. This is really
just a workaround, but I need to refactor the v0.11 stuff anyway so I
don't want to figure it out.
This is a major rewrite of the RISC-V v0.13 OpenOCD port. This
shouldn't have any meaningful effect on the v0.11 support, but it does
add generic versions of many functions that will allow me to later
refactor the v0.11 support so it's easier to maintain both ports. This
started as an emergency feature branch and went on for a long time, so
it's all been squashed down into one commit so there isn't a big set of
broken commits lying around. The changes are:
* You can pass "-rtos riscv" to the target in OpenOCD's configuration
file, which enables multi-hart mode. This uses OpenOCD's RTOS
support to control all the harts from the debug module using commands
like "info threads" in GDB. This support is still expermental.
* There is support for RV64I, but due to OpenOCD limitations we only
support 32-bit physical addresses. I hope to remedy this by rebasing
onto the latest OpenOCD release, which I've heard should fix this.
* This matches the latest draft version of the RISC-V debug spec, as of
April 26th. This version fixes a number of spec bugs and should be
close to the final debug spec.
read_memory was doing autoexec-initialized commands, but wasn't using
ac_busy_delay to add delay or wait between the commands. This isn't the optimal
solution, but the whole read_memory sequence needs to be made more efficient
anyway, so this was the quick and dirty solution.
Replace the constants with the ones from the ISA spec, since those are
not updated as often.
Also delete a bunch of old code.
Change-Id: I9201b1455d64a9d2d203bb362fefaa68cbf35aeb
The remote bitbang implementation is really slow when reading back data.
During many operations, like writing a block of memory, we don't need to
see the return data. By communicating this to the lower layers, we get
huge speedups. Downloads to spike now are 295KB/s.
That means the gdbserver.py download test now runs to completion,
unfortunately it fails. Everything else is still passing, though, so I'm
committing this change.
Change-Id: I44cc9db1ade0908c1a12d09b23fc8e529a802d88