dawr-power9.rst

  1=====================
  2DAWR issues on POWER9
  3=====================
  4
  5On older POWER9 processors, the Data Address Watchpoint Register (DAWR) can
  6cause a checkstop if it points to cache inhibited (CI) memory. Currently Linux
  7has no way to distinguish CI memory when configuring the DAWR, so on affected
  8systems, the DAWR is disabled.
  9
 10Affected processor revisions
 11============================
 12
 13This issue is only present on processors prior to v2.3. The revision can be
 14found in /proc/cpuinfo::
 15
 16    processor       : 0
 17    cpu             : POWER9, altivec supported
 18    clock           : 3800.000000MHz
 19    revision        : 2.3 (pvr 004e 1203)
 20
 21On a system with the issue, the DAWR is disabled as detailed below.
 22
 23Technical Details:
 24==================
 25
 26DAWR has 6 different ways of being set.
 271) ptrace
 282) h_set_mode(DAWR)
 293) h_set_dabr()
 304) kvmppc_set_one_reg()
 315) xmon
 32
 33For ptrace, we now advertise zero breakpoints on POWER9 via the
 34PPC_PTRACE_GETHWDBGINFO call. This results in GDB falling back to
 35software emulation of the watchpoint (which is slow).
 36
 37h_set_mode(DAWR) and h_set_dabr() will now return an error to the
 38guest on a POWER9 host. Current Linux guests ignore this error, so
 39they will silently not get the DAWR.
 40
 41kvmppc_set_one_reg() will store the value in the vcpu but won't
 42actually set it on POWER9 hardware. This is done so we don't break
 43migration from POWER8 to POWER9, at the cost of silently losing the
 44DAWR on the migration.
 45
 46For xmon, the 'bd' command will return an error on P9.
 47
 48Consequences for users
 49======================
 50
 51For GDB watchpoints (ie 'watch' command) on POWER9 bare metal , GDB
 52will accept the command. Unfortunately since there is no hardware
 53support for the watchpoint, GDB will software emulate the watchpoint
 54making it run very slowly.
 55
 56The same will also be true for any guests started on a POWER9
 57host. The watchpoint will fail and GDB will fall back to software
 58emulation.
 59
 60If a guest is started on a POWER8 host, GDB will accept the watchpoint
 61and configure the hardware to use the DAWR. This will run at full
 62speed since it can use the hardware emulation. Unfortunately if this
 63guest is migrated to a POWER9 host, the watchpoint will be lost on the
 64POWER9. Loads and stores to the watchpoint locations will not be
 65trapped in GDB. The watchpoint is remembered, so if the guest is
 66migrated back to the POWER8 host, it will start working again.
 67
 68Force enabling the DAWR
 69=======================
 70Kernels (since ~v5.2) have an option to force enable the DAWR via::
 71
 72  echo Y > /sys/kernel/debug/powerpc/dawr_enable_dangerous
 73
 74This enables the DAWR even on POWER9.
 75
 76This is a dangerous setting, USE AT YOUR OWN RISK.
 77
 78Some users may not care about a bad user crashing their box
 79(ie. single user/desktop systems) and really want the DAWR.  This
 80allows them to force enable DAWR.
 81
 82This flag can also be used to disable DAWR access. Once this is
 83cleared, all DAWR access should be cleared immediately and your
 84machine once again safe from crashing.
 85
 86Userspace may get confused by toggling this. If DAWR is force
 87enabled/disabled between getting the number of breakpoints (via
 88PTRACE_GETHWDBGINFO) and setting the breakpoint, userspace will get an
 89inconsistent view of what's available. Similarly for guests.
 90
 91For the DAWR to be enabled in a KVM guest, the DAWR needs to be force
 92enabled in the host AND the guest. For this reason, this won't work on
 93POWERVM as it doesn't allow the HCALL to work. Writes of 'Y' to the
 94dawr_enable_dangerous file will fail if the hypervisor doesn't support
 95writing the DAWR.
 96
 97To double check the DAWR is working, run this kernel selftest:
 98
 99  tools/testing/selftests/powerpc/ptrace/ptrace-hwbreak.c
100
101Any errors/failures/skips mean something is wrong.