On 11/8/18 5:09 PM, Philippe Mathieu-Daudé wrote: > On 6/11/18 12:05, Luc Michel wrote: >> changes since v3: >> - patch 1 cpu_cluster.h: remove QEMU_ from the multiple includes >> guard #ifdef/#define [Alistair] >> >> - patch 1 cpu_cluster.c: include osdep.h first [Alistair] >> >> - patch 1 use uint64_t for cluster ID for prosperity :) [Philippe] > > Actually my comment about 32-bit had the opposite meaning, I suppose > having 256 kind of cpu cores in the same SoC is probably enough for a > long time IMHO. Oops sorry about that :) I thought you were referring to an old quote from a guy at Redmond about 640KiB of RAM ;)
A case I was thinking about that would require more than 8- or 16-bit ID is if for whatever reason we need to model a SoC with sparse (non-continuous) cluster IDs. Anyway, I can put back the uint32_t for the next re-roll, I think it's a good compromise. Thanks. Luc > >> >> - patch 1 auto-assign a cluster ID to newly created clusters >> [Philippe] >> >> - patch 2 fix mid-code variable declaration [Alistair] >> >> - patch 3 add a comment in gdb_get_cpu_pid() when retrieving CPU >> parent canonical path [Alistair] >> >> - patch 14 fix a typo in the commit message [Alistair] >> >> changes since v2: >> - patch 1 introducing the cpu-cluster type. I didn't opt for an >> Interface, but I can add one if you think it's necessary. >> For now this class inherits from Device and has a >> cluster-id property, used by the GDB stub to compute a >> PID. >> >> - patch 2 removed GDB group related code as it has been replaced >> with CPU clusters >> >> - patch 2/8 moved GDBProcess target_xml field introduction into patch >> 8 [Philippe] >> >> - patch 3 gdb_get_cpu_pid() now search for CPU being a child of a >> cpu-cluster object. Use the cluster-id to compute the >> PID. >> >> - patch 4 gdb_get_process() does not rely on s->processes array >> indices anymore as PIDs can now be sparse. Instead, >> iterate >> over the array to find the process. >> >> - patch 3/4 removed Reviewed-by tags because of substantial changes. >> >> - patch 4/7 read_thread_id() hardening [Philippe] >> >> - patch 12 safer vAttach packet parsing [Phillipe] >> >> - patch 16 put APUs and RPUs in different clusters instead of GDB >> groups >> >> changes since v1: >> - rename qemu_get_thread_id() to gdb_fmt_thread_id() [Philippe] >> - check qemu_strtoul() return value for 'D' packets [Philippe] >> >> >> This series adds support for the multiprocess extension of the GDB >> remote protocol in the QEMU GDB stub. >> >> This extension is useful to split QEMU emulated CPUs in different >> processes from the point of view of the GDB client. It adds the >> possibility to debug different kind of processors (e.g. an AArch64 and >> an ARMv7 CPU) at the same time (it is not possible otherwise since GDB >> expects an SMP view at the thread granularity. >> >> CPUs are grouped using specially named QOM containers. CPUs that are >> children of such a container are grouped under the same GDB process. >> >> The last patch groups the CPUs of different model in the zynqmp machines >> into separate processes. >> >> To test this patchset, you can use the following commands: >> >> (Note that this requires a recent enough GDB, I think GDB 7.2 is OK. >> Also, it must be compiled to support both ARM and AArch64 architectures) >> >> Run QEMU: (-smp 6 in xlnx-zcu102 enables both cortex-a53 and cortex-r5 >> CPUs) >> >> qemu-system-aarch64 -M xlnx-zcu102 -gdb tcp::1234 -S -smp 6 >> >> Run the following commands in GDB: >> >> target extended :1234 >> add-inferior >> inferior 2 >> attach 2 >> info threads >> >> I want to thanks the Xilinx's QEMU team who sponsored this work for >> their collaboration and their prototype implementation. >> >> Luc Michel (16): >> hw/cpu: introduce CPU clusters >> gdbstub: introduce GDB processes >> gdbstub: add multiprocess support to '?' packets >> gdbstub: add multiprocess support to 'H' and 'T' packets >> gdbstub: add multiprocess support to vCont packets >> gdbstub: add multiprocess support to 'sC' packets >> gdbstub: add multiprocess support to (f|s)ThreadInfo and >> ThreadExtraInfo >> gdbstub: add multiprocess support to Xfer:features:read: >> gdbstub: add multiprocess support to gdb_vm_state_change() >> gdbstub: add multiprocess support to 'D' packets >> gdbstub: add support for extended mode packet >> gdbstub: add support for vAttach packets >> gdbstub: processes initialization on new peer connection >> gdbstub: gdb_set_stop_cpu: ignore request when process is not attached >> gdbstub: add multiprocess extension support >> arm/xlnx-zynqmp: put APUs and RPUs in separate CPU clusters >> >> include/hw/arm/xlnx-zynqmp.h | 3 + >> include/hw/cpu/cluster.h | 38 +++ >> gdbstub.c | 632 ++++++++++++++++++++++++++++++----- >> hw/arm/xlnx-zynqmp.c | 21 +- >> hw/cpu/cluster.c | 59 ++++ >> hw/cpu/Makefile.objs | 2 +- >> 6 files changed, 674 insertions(+), 81 deletions(-) >> create mode 100644 include/hw/cpu/cluster.h >> create mode 100644 hw/cpu/cluster.c >>
