[v5,0/3] target/ppc: single step for KVM HV

Message

Fabiano Rosas Dec. 11, 2019, 7:10 p.m. UTC

I'm resending this series now that the KVM capability got merged:

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=1a9167a214f

---
Original cover-letter:

Single stepping via GDB/gdbstub is currently not working with KVM
HV. When asking for a single step (stepi), KVM simply ignores the
request and execution continues.

This has the direct effect of breaking GDB's 'step', 'stepi', 'next',
'nexti' commands. The 'continue' command is also affected since
continuing right after a breakpoint requires that GDB first perform a
single step so that the breakpoint can be re-inserted before
continuing - in this case the breakpoint is not re-inserted and it
won't hit again.

The issue here is that single stepping in POWER makes use of an
interrupt (Trace Interrupt [1]) that does not reach the hypervisor, so
while the single step would happen if properly triggered, it would not
cause an exit to KVM so there would be no way of handing control back
to GDB. Aside from that, the guest kernel is not prepared to deal with
such an interrupt in kernel mode (when not using KGDB, or some other
debugging facility) and it causes an Oops.

This series implements a "software single step" approach that makes
use of: i) the Trace Interrupt to perform the step inside the guest
and ii) a breakpoint at the Trace Interrupt handler address to cause a
vm exit (Emulation Assist) so that we can return control to QEMU.

With (i), we basically get the single step for free, without having to
discover what are the possible targets of instructions that divert
execution.

With (ii), we hide the single step from the guest and keep all of the
step logic in QEMU.

Supported scenarios:

- Stepping of multiple vcpus;
- GDB scheduler locking on and off [2];
- single stepping of kernel code with QEMU while stepping with GDB
  inside the guest (user space, KGDB).

1- PowerISA Section 6.5.15 - Trace Interrupt
2- https://sourceware.org/gdb/onlinedocs/gdb/All_002dStop-Mode.html

v4 -> v5:
 - rebase to v4.2.0-rc5;

 - use KVM_CAP_PPC_GUEST_DEBUG_SSTEP (#176) which is now in kernel
   v5.5-rc1:
   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=1a9167a214f

v3 -> v4:
 - patch 1: fix uninitialized 'offset' variable;

 - patch 2: squash with patch 7/7 (now 5/5);
            fix exception vector offset calculation when AIL == 0;

 - patch 3: squash with 4/7 (now 2/5);

 - patch 7: introduce ppc_gdb_get_{op,xop,spr} functions;

            introduce ppc_gdb_read_insn function;

	    define constants for mfmsr, rfid, mtspr extended opcodes;

            fix bug where instructions would not be properly
	    recognized in SLOF due to guest endianness being different
	    from host's;

	    pass arch_info->address directly into functions that only
	    need the address;

	    improve indentation by returning early when possible.

 https://lists.nongnu.org/archive/html/qemu-devel/2019-02/msg07994.html

v2 -> v3:
 - take Alternate Interrupt Location (AIL) into consideration when
   calculating the Trace Interrupt handler address (this allows single
   stepping in SLOF code);

 - check for a new KVM_GUEST_DEBUG_SSTEP capability (still to be
   submitted to kernel ml);

 - handle other vcpus (not currently stepping) hitting the single step
   breakpoint - by ignoring the breakpoint;

 - handle simultaneous single step by GDB inside guest - by first
   performing our step into the trace interrupt handler itself and
   returning to the guest afterwards;

 - handle single stepping when at the first trace interrupt handler
   instruction - by displacing the breakpoint to the next instruction;

 - restore MSR, SRR0, SRR1 after the step, taking into consideration
   possible mtspr, mtmsr instructions;

 - use stubs for arch-specific code that will not be implemented by
   other architectures at this point;

 https://lists.gnu.org/archive/html/qemu-devel/2019-01/msg04627.html

v1 -> v2:
 - split in more patches to facilitate review
 - use extract32 for decoding instruction instead of open-coding
 - add more people to CC

 https://lists.gnu.org/archive/html/qemu-devel/2018-11/msg04269.html

v1:

 https://lists.gnu.org/archive/html/qemu-devel/2018-11/msg03738.html


Fabiano Rosas (3):
  linux-headers: Update kvm.h for ppc single step capability
  kvm-all: Introduce kvm_set_singlestep
  target/ppc: support single stepping with KVM HV

 accel/kvm/kvm-all.c         |  14 +++
 accel/stubs/kvm-stub.c      |   4 +
 exec.c                      |   2 +-
 include/sysemu/kvm.h        |   4 +
 linux-headers/linux/kvm.h   |   1 +
 stubs/Makefile.objs         |   1 +
 stubs/kvm-arch-singlestep.c |  14 +++
 target/ppc/cpu.h            |  16 +++
 target/ppc/excp_helper.c    |  13 +++
 target/ppc/gdbstub.c        |  35 ++++++
 target/ppc/kvm.c            | 209 ++++++++++++++++++++++++++++++++++--
 11 files changed, 305 insertions(+), 8 deletions(-)
 create mode 100644 stubs/kvm-arch-singlestep.c