platform/kernel/linux-starfive.git
2 years agopowerpc/64s: Get LPID bit width from device tree
Nicholas Piggin [Mon, 29 Nov 2021 03:09:15 +0000 (13:09 +1000)]
powerpc/64s: Get LPID bit width from device tree

Allow the LPID bit width and partition table size to be set at runtime
from the device tree.

Move the PID bit width detection into the same place.

KVM does not support using the extra bits yet, this is mainly required
to get the PTCR register values correct (so KVM will run but it will
not allocate > 4096 LPIDs).

OPAL firmware provides this property for POWER10 CPUs since skiboot
commit 9b85f7d961f2 ("hdata: add mmu-pid-bits and mmu-lpid-bits for
POWER10 CPUs").

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211129030915.1888332-1-npiggin@gmail.com
2 years agopowerpc/perf: Fix PMU callbacks to clear pending PMI before resetting an overflown PMC
Athira Rajeev [Wed, 21 Jul 2021 05:48:29 +0000 (01:48 -0400)]
powerpc/perf: Fix PMU callbacks to clear pending PMI before resetting an overflown PMC

Running perf fuzzer showed below in dmesg logs:
  "Can't find PMC that caused IRQ"

This means a PMU exception happened, but none of the PMC's (Performance
Monitor Counter) were found to be overflown. There are some corner cases
that clears the PMCs after PMI gets masked. In such cases, the perf
interrupt handler will not find the active PMC values that had caused
the overflow and thus leads to this message while replaying.

Case 1: PMU Interrupt happens during replay of other interrupts and
counter values gets cleared by PMU callbacks before replay:

During replay of interrupts like timer, __do_irq() and doorbell
exception, we conditionally enable interrupts via may_hard_irq_enable().
This could potentially create a window to generate a PMI. Since irq soft
mask is set to ALL_DISABLED, the PMI will get masked here. We could get
IPIs run before perf interrupt is replayed and the PMU events could
be deleted or stopped. This will change the PMU SPR values and resets
the counters. Snippet of ftrace log showing PMU callbacks invoked in
__do_irq():

  <idle>-0 [051] dns. 132025441306354: __do_irq <-call_do_irq
  <idle>-0 [051] dns. 132025441306430: irq_enter <-__do_irq
  <idle>-0 [051] dns. 132025441306503: irq_enter_rcu <-__do_irq
  <idle>-0 [051] dnH. 132025441306599: xive_get_irq <-__do_irq
  <<>>
  <idle>-0 [051] dnH. 132025441307770: generic_smp_call_function_single_interrupt <-smp_ipi_demux_relaxed
  <idle>-0 [051] dnH. 132025441307839: flush_smp_call_function_queue <-smp_ipi_demux_relaxed
  <idle>-0 [051] dnH. 132025441308057: _raw_spin_lock <-event_function
  <idle>-0 [051] dnH. 132025441308206: power_pmu_disable <-perf_pmu_disable
  <idle>-0 [051] dnH. 132025441308337: power_pmu_del <-event_sched_out
  <idle>-0 [051] dnH. 132025441308407: power_pmu_read <-power_pmu_del
  <idle>-0 [051] dnH. 132025441308477: read_pmc <-power_pmu_read
  <idle>-0 [051] dnH. 132025441308590: isa207_disable_pmc <-power_pmu_del
  <idle>-0 [051] dnH. 132025441308663: write_pmc <-power_pmu_del
  <idle>-0 [051] dnH. 132025441308787: power_pmu_event_idx <-perf_event_update_userpage
  <idle>-0 [051] dnH. 132025441308859: rcu_read_unlock_strict <-perf_event_update_userpage
  <idle>-0 [051] dnH. 132025441308975: power_pmu_enable <-perf_pmu_enable
  <<>>
  <idle>-0 [051] dnH. 132025441311108: irq_exit <-__do_irq
  <idle>-0 [051] dns. 132025441311319: performance_monitor_exception <-replay_soft_interrupts

Case 2: PMI's masked during local_* operations, example local_add(). If
the local_add() operation happens within a local_irq_save(), replay of
PMI will be during local_irq_restore(). Similar to case 1, this could
also create a window before replay where PMU events gets deleted or
stopped.

Fix it by updating the PMU callback function power_pmu_disable() to
check for pending perf interrupt. If there is an overflown PMC and
pending perf interrupt indicated in paca, clear the PMI bit in paca to
drop that sample. Clearing of PMI bit is done in power_pmu_disable()
since disable is invoked before any event gets deleted/stopped. With
this fix, if there are more than one event running in the PMU, there is
a chance that we clear the PMI bit for the event which is not getting
deleted/stopped. The other events may still remain active. Hence to make
sure we don't drop valid sample in such cases, another check is added in
power_pmu_enable. This checks if there is an overflown PMC found among
the active events and if so enable back the PMI bit. Two new helper
functions are introduced to clear/set the PMI, ie
clear_pmi_irq_pending() and set_pmi_irq_pending(). Helper function
pmi_irq_pending() is introduced to give a warning if there is pending
PMI bit in paca, but no PMC is overflown.

Also there are corner cases which result in performance monitor
interrupts being triggered during power_pmu_disable(). This happens
since PMXE bit is not cleared along with disabling of other MMCR0 bits
in the pmu_disable. Such PMI's could leave the PMU running and could
trigger PMI again which will set MMCR0 PMAO bit. This could lead to
spurious interrupts in some corner cases. Example, a timer after
power_pmu_del() which will re-enable interrupts and triggers a PMI again
since PMAO bit is still set. But fails to find valid overflow since PMC
was cleared in power_pmu_del(). Fix that by disabling PMXE along with
disabling of other MMCR0 bits in power_pmu_disable().

We can't just replay PMI any time. Hence this approach is preferred
rather than replaying PMI before resetting overflown PMC. Patch also
documents core-book3s on a race condition which can trigger these PMC
messages during idle path in PowerNV.

Fixes: f442d004806e ("powerpc/64s: Add support to mask perf interrupts and replay them")
Reported-by: Nageswara R Sastry <nasastry@in.ibm.com>
Suggested-by: Nicholas Piggin <npiggin@gmail.com>
Suggested-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Tested-by: Nageswara R Sastry <rnsastry@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Make pmi_irq_pending() return bool, reflow/reword some comments]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1626846509-1350-2-git-send-email-atrajeev@linux.vnet.ibm.com
2 years agopowerpc/atomics: Remove atomic_inc()/atomic_dec() and friends
Christophe Leroy [Tue, 21 Sep 2021 15:09:49 +0000 (17:09 +0200)]
powerpc/atomics: Remove atomic_inc()/atomic_dec() and friends

Now that atomic_add() and atomic_sub() handle immediate operands,
atomic_inc() and atomic_dec() have no added value compared to the
generic fallback which calls atomic_add(1) and atomic_sub(1).

Also remove atomic_inc_not_zero() which fallsback to
atomic_add_unless() which itself fallsback to
atomic_fetch_add_unless() which now handles immediate operands.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/0bc64a2f18726055093dbb2e479cefc60a409cfd.1632236981.git.christophe.leroy@csgroup.eu
2 years agopowerpc/atomics: Use immediate operand when possible
Christophe Leroy [Tue, 21 Sep 2021 15:09:48 +0000 (17:09 +0200)]
powerpc/atomics: Use immediate operand when possible

Today we get the following code generation for atomic operations:

c001bb2c: 39 20 00 01  li      r9,1
c001bb30: 7d 40 18 28  lwarx   r10,0,r3
c001bb34: 7d 09 50 50  subf    r8,r9,r10
c001bb38: 7d 00 19 2d  stwcx.  r8,0,r3

c001c7a8: 39 40 00 01  li      r10,1
c001c7ac: 7d 00 18 28  lwarx   r8,0,r3
c001c7b0: 7c ea 42 14  add     r7,r10,r8
c001c7b4: 7c e0 19 2d  stwcx.  r7,0,r3

By allowing GCC to choose between immediate or regular operation,
we get:

c001bb2c: 7d 20 18 28  lwarx   r9,0,r3
c001bb30: 39 49 ff ff  addi    r10,r9,-1
c001bb34: 7d 40 19 2d  stwcx.  r10,0,r3
--
c001c7a4: 7d 40 18 28  lwarx   r10,0,r3
c001c7a8: 39 0a 00 01  addi    r8,r10,1
c001c7ac: 7d 00 19 2d  stwcx.  r8,0,r3

For "and", the dot form has to be used because "andi" doesn't exist.

For logical operations we use unsigned 16 bits immediate.
For arithmetic operations we use signed 16 bits immediate.

On pmac32_defconfig, it reduces the text by approx another 8 kbytes.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Acked-by: Segher Boessenkool <segher@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/2ec558d44db8045752fe9dbd29c9ba84bab6030b.1632236981.git.christophe.leroy@csgroup.eu
2 years agopowerpc/bitops: Use immediate operand when possible
Christophe Leroy [Tue, 21 Sep 2021 15:09:47 +0000 (17:09 +0200)]
powerpc/bitops: Use immediate operand when possible

Today we get the following code generation for bitops like
set or clear bit:

c0009fe0: 39 40 08 00  li      r10,2048
c0009fe4: 7c e0 40 28  lwarx   r7,0,r8
c0009fe8: 7c e7 53 78  or      r7,r7,r10
c0009fec: 7c e0 41 2d  stwcx.  r7,0,r8

c000d568: 39 00 18 00  li      r8,6144
c000d56c: 7c c0 38 28  lwarx   r6,0,r7
c000d570: 7c c6 40 78  andc    r6,r6,r8
c000d574: 7c c0 39 2d  stwcx.  r6,0,r7

Most set bits are constant on lower 16 bits, so it can easily
be replaced by the "immediate" version of the operation. Allow
GCC to choose between the normal or immediate form.

For clear bits, on 32 bits 'rlwinm' can be used instead of 'andc' for
when all bits to be cleared are consecutive.

On 64 bits we don't have any equivalent single operation for clearing,
single bits or a few bits, we'd need two 'rldicl' so it is not
worth it, the li/andc sequence is doing the same.

With this patch we get:

c0009fe0: 7d 00 50 28  lwarx   r8,0,r10
c0009fe4: 61 08 08 00  ori     r8,r8,2048
c0009fe8: 7d 00 51 2d  stwcx.  r8,0,r10

c000d558: 7c e0 40 28  lwarx   r7,0,r8
c000d55c: 54 e7 05 64  rlwinm  r7,r7,0,21,18
c000d560: 7c e0 41 2d  stwcx.  r7,0,r8

On pmac32_defconfig, it reduces the text by approx 10 kbytes.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Reviewed-by: Segher Boessenkool <segher@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/e6f815d9181bab09df3b350af51149437863e9f9.1632236981.git.christophe.leroy@csgroup.eu
2 years agopowerpc: flexible GPR range save/restore macros
Nicholas Piggin [Fri, 22 Oct 2021 06:13:22 +0000 (16:13 +1000)]
powerpc: flexible GPR range save/restore macros

Introduce macros that operate on a (start, end) range of GPRs, which
reduces lines of code and need to do mental arithmetic while reading the
code.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Segher Boessenkool <segher@kernel.crashing.org>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211022061322.2671178-1-npiggin@gmail.com
2 years agopowerpc/watchdog: help remote CPUs to flush NMI printk output
Nicholas Piggin [Fri, 19 Nov 2021 11:31:46 +0000 (21:31 +1000)]
powerpc/watchdog: help remote CPUs to flush NMI printk output

The printk layer at the moment does not seem to have a good way to force
flush printk messages that are created in NMI context, except in the
panic path.

NMI-context printk messages normally get to the console with irq_work,
but that won't help if the CPU is stuck with irqs disabled, as can be
the case for hard lockup watchdog messages.

The watchdog currently flushes the printk buffers after detecting a
lockup on remote CPUs, but they may not have processed their NMI IPI
yet by that stage, or they may have self-detected a lockup in which
case they won't go via this NMI IPI path.

Improve the situation by having NMI-context mark a flag if it called
printk, and have watchdog timer interrupts check if that flag was set
and try to flush if it was. Latency is not a big problem because we
were already stuck for a while, just need to try to make sure the
messages eventually make it out.

Depends-on: 5d5e4522a7f4 ("printk: restore flushing of NMI buffers on remote CPUs after NMI backtraces")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211119113146.752759-6-npiggin@gmail.com
2 years agopowerpc: Don't bother about .data..Lubsan sections
Christophe Leroy [Thu, 25 Nov 2021 11:43:33 +0000 (12:43 +0100)]
powerpc: Don't bother about .data..Lubsan sections

Since commit 9a427556fb8e ("vmlinux.lds.h: catch compound literals
into data and BSS") .data..Lubsan sections are taken into account
in DATA_MAIN which is included in DATA_DATA macro.

No need to take care of them anymore in powerpc vmlinux.lds.S

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/3eb14570612eef17e01bb67f14a4450136001794.1637840601.git.christophe.leroy@csgroup.eu
2 years agopowerpc/ptdump: Fix display a BAT's size unit
Christophe Leroy [Fri, 26 Nov 2021 10:30:03 +0000 (11:30 +0100)]
powerpc/ptdump: Fix display a BAT's size unit

We have wrong units on BAT's sizes (G instead of M, M instead of ...)

---[ Instruction Block Address Translation ]---
0: 0xc0000000-0xc03fffff 0x00000000         4G Kernel   x     m
1: 0xc0400000-0xc05fffff 0x00400000         2G Kernel   x     m
2: 0xc0600000-0xc06fffff 0x00600000         1G Kernel   x     m
3: 0xc0700000-0xc077ffff 0x00700000       512M Kernel   x     m
4: 0xc0780000-0xc079ffff 0x00780000       128M Kernel   x     m
5: 0xc07a0000-0xc07bffff 0x007a0000       128M Kernel   x     m
6:         -
7:         -

This is because pt_dump_size() expects a size in Kbytes but
bat_show_603() gives the size in bytes.

To avoid risk of confusion, change pt_dump_size() to take bytes.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/f16c30f5c9185a63335322cf1a8b22f189d335ef.1637922595.git.christophe.leroy@csgroup.eu
2 years agopowerpc/ftrace: Activate HAVE_DYNAMIC_FTRACE_WITH_REGS on PPC32
Christophe Leroy [Thu, 28 Oct 2021 12:24:04 +0000 (14:24 +0200)]
powerpc/ftrace: Activate HAVE_DYNAMIC_FTRACE_WITH_REGS on PPC32

Unlike PPC64, PPC32 doesn't require any special compiler option
to get _mcount() call not clobbering registers.

Provide ftrace_regs_caller() and ftrace_regs_call() and activate
HAVE_DYNAMIC_FTRACE_WITH_REGS.

That's heavily copied from ftrace_64_mprofile.S

For the time being leave livepatching aside, it will come with
following patch.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1862dc7719855cc2a4eec80920d94c955877557e.1635423081.git.christophe.leroy@csgroup.eu
2 years agopowerpc/ftrace: Add module_trampoline_target() for PPC32
Christophe Leroy [Thu, 28 Oct 2021 12:24:03 +0000 (14:24 +0200)]
powerpc/ftrace: Add module_trampoline_target() for PPC32

module_trampoline_target() is used by __ftrace_modify_call().

Implement it for PPC32 so that CONFIG_DYNAMIC_FTRACE_WITH_REGS
can be activated on PPC32 as well.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/42345f464fb465f0fc76f3090e250be8fc1729f0.1635423081.git.christophe.leroy@csgroup.eu
2 years agopowerpc/ftrace: No need to read LR from stack in _mcount()
Christophe Leroy [Thu, 28 Oct 2021 12:24:02 +0000 (14:24 +0200)]
powerpc/ftrace: No need to read LR from stack in _mcount()

All functions calling _mcount do it exactly the same way, with the
following sequence of instructions:

c07de788:       7c 08 02 a6     mflr    r0
c07de78c:       90 01 00 04     stw     r0,4(r1)
c07de790:       4b 84 13 65     bl      c001faf4 <_mcount>

Allthough LR is pushed on stack, it is still in r0 while entering
_mcount().

Function arguments are in r3-r10, so r11 and r12 are still available
at that point.

Do like PPC64 and use r12 to move LR into CTR, so that r0 is preserved
and doesn't need to be restored from the stack.

While at it, bring back the EXPORT_SYMBOL at the end of _mcount.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/24a3ba7db388537c44a038026f926d885372e6d3.1635423081.git.christophe.leroy@csgroup.eu
2 years agopowerpc: Mark probe_machine() __init and static
Michael Ellerman [Wed, 24 Nov 2021 09:32:54 +0000 (20:32 +1100)]
powerpc: Mark probe_machine() __init and static

Prior to commit b1923caa6e64 ("powerpc: Merge 32-bit and 64-bit
setup_arch()") probe_machine() was called from setup_32/64.c and lived
in setup-common.c. But now it's only called from setup-common.c so it
can be static and __init, and we don't need the declaration in
machdep.h either.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211124093254.1054750-6-mpe@ellerman.id.au
2 years agopowerpc/smp: Move setup_profiling_timer() under CONFIG_PROFILING
Michael Ellerman [Wed, 24 Nov 2021 09:32:53 +0000 (20:32 +1100)]
powerpc/smp: Move setup_profiling_timer() under CONFIG_PROFILING

setup_profiling_timer() is only needed when CONFIG_PROFILING is enabled.

Fixes the following W=1 warning when CONFIG_PROFILING=n:
  linux/arch/powerpc/kernel/smp.c:1638:5: error: no previous prototype for ‘setup_profiling_timer’

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211124093254.1054750-5-mpe@ellerman.id.au
2 years agopowerpc/mm: Move tlbcam_sz() and make it static
Michael Ellerman [Wed, 24 Nov 2021 09:32:52 +0000 (20:32 +1100)]
powerpc/mm: Move tlbcam_sz() and make it static

Building with W=1 we see a warning:
  linux/arch/powerpc/mm/nohash/fsl_book3e.c:63:15: error: no previous prototype for ‘tlbcam_sz’

tlbcam_sz() is not used outside this file, so we can make it static.
However it's only used inside #ifdef CONFIG_PPC32, so move it within
that ifdef, otherwise we would get a defined but not used error.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211124093254.1054750-4-mpe@ellerman.id.au
2 years agopowerpc/85xx: Make c293_pcie_pic_init() static
Michael Ellerman [Wed, 24 Nov 2021 09:32:51 +0000 (20:32 +1100)]
powerpc/85xx: Make c293_pcie_pic_init() static

To fix the W=1 warning:
  linux/arch/powerpc/platforms/85xx/c293pcie.c:22:13: error: no previous prototype for ‘c293_pcie_pic_init’

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211124093254.1054750-3-mpe@ellerman.id.au
2 years agopowerpc/85xx: Make mpc85xx_smp_kexec_cpu_down() static
Michael Ellerman [Wed, 24 Nov 2021 09:32:50 +0000 (20:32 +1100)]
powerpc/85xx: Make mpc85xx_smp_kexec_cpu_down() static

To fix the W=1 warning:
  arch/powerpc/platforms/85xx/smp.c:369:6: error: no previous prototype for ‘mpc85xx_smp_kexec_cpu_down’

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211124093254.1054750-2-mpe@ellerman.id.au
2 years agopowerpc/85xx: Fix no previous prototype warning for mpc85xx_setup_pmc()
Michael Ellerman [Wed, 24 Nov 2021 09:32:49 +0000 (20:32 +1100)]
powerpc/85xx: Fix no previous prototype warning for mpc85xx_setup_pmc()

Fixes the following W=1 warning:
  arch/powerpc/platforms/85xx/mpc85xx_pm_ops.c:89:12: warning: no previous prototype for 'mpc85xx_setup_pmc'

Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211124093254.1054750-1-mpe@ellerman.id.au
2 years agopowerpc: select CPUMASK_OFFSTACK if NR_CPUS >= 8192
Nicholas Piggin [Fri, 5 Nov 2021 03:50:42 +0000 (13:50 +1000)]
powerpc: select CPUMASK_OFFSTACK if NR_CPUS >= 8192

Some core kernel code starts to go beyond the 2048 byte stack size
warning at NR_CPUS=8192, so select CPUMASK_OFFSTACK in that case.
x86 does similarly for very large NR_CPUS.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105035042.1398309-2-npiggin@gmail.com
2 years agopowerpc: remove cpu_online_cores_map function
Nicholas Piggin [Fri, 5 Nov 2021 03:50:41 +0000 (13:50 +1000)]
powerpc: remove cpu_online_cores_map function

This function builds the cores online map with on-stack cpumasks which
can cause high stack usage with large NR_CPUS.

It is not used in any performance sensitive paths, so instead just check
for first thread sibling.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105035042.1398309-1-npiggin@gmail.com
2 years agoRevert "powerpc/code-patching: Improve verification of patchability"
Michael Ellerman [Mon, 29 Nov 2021 06:41:52 +0000 (17:41 +1100)]
Revert "powerpc/code-patching: Improve verification of patchability"

This reverts commit 8b8a8f0ab3f5519e45c526f826a655817486c5bb.

As reported[1] by Sachin this causes problems with ftrace, and it also
causes the code patching selftests to fail as reported[2] by Stephen.

So revert it for now.

1: https://lore.kernel.org/linuxppc-dev/3668743C-09DF-4673-B15C-2FFE2A57F7D7@linux.vnet.ibm.com/
2: https://lore.kernel.org/linuxppc-dev/20211126161747.1f7795b0@canb.auug.org.au/

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2 years agopowerpc/watchdog: Fix wd_smp_last_reset_tb reporting
Nicholas Piggin [Thu, 25 Nov 2021 10:33:46 +0000 (20:33 +1000)]
powerpc/watchdog: Fix wd_smp_last_reset_tb reporting

wd_smp_last_reset_tb now gets reset by watchdog_smp_panic() as part of
marking CPUs stuck and removing them from the pending mask before it
begins any printing. This causes last reset times reported to be off.

Fix this by reading it into a local variable before it gets reset.

Fixes: 76521c4b0291 ("powerpc/watchdog: Avoid holding wd_smp_lock over printk and smp_send_nmi_ipi")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211125103346.1188958-1-npiggin@gmail.com
2 years agopowerpc/microwatt: Make microwatt_get_random_darn() static
Michael Ellerman [Thu, 18 Nov 2021 00:44:15 +0000 (11:44 +1100)]
powerpc/microwatt: Make microwatt_get_random_darn() static

Make microwatt_get_random_darn() static, because it can be.

Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211118004415.1706863-1-mpe@ellerman.id.au
2 years agopowerpc/watchdog: read TB close to where it is used
Nicholas Piggin [Wed, 10 Nov 2021 02:50:56 +0000 (12:50 +1000)]
powerpc/watchdog: read TB close to where it is used

When taking watchdog actions, printing messages, comparing and
re-setting wd_smp_last_reset_tb, etc., read TB close to the point of use
and under wd_smp_lock or printing lock (if applicable).

This should keep timebase mostly monotonic with kernel log messages, and
could prevent (in theory) a laggy CPU updating wd_smp_last_reset_tb to
something a long way in the past, and causing other CPUs to appear to be
stuck.

These additional TB reads are all slowpath (lockup has been detected),
so performance does not matter.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211110025056.2084347-5-npiggin@gmail.com
2 years agopowerpc/watchdog: Avoid holding wd_smp_lock over printk and smp_send_nmi_ipi
Nicholas Piggin [Wed, 10 Nov 2021 02:50:55 +0000 (12:50 +1000)]
powerpc/watchdog: Avoid holding wd_smp_lock over printk and smp_send_nmi_ipi

There is a deadlock with the console_owner lock and the wd_smp_lock:

CPU x takes the console_owner lock
CPU y takes a watchdog timer interrupt and takes __wd_smp_lock
CPU x takes a soft-NMI interrupt, detects deadlock, spins on __wd_smp_lock
CPU y detects deadlock, tries to print something and spins on console_owner
-> deadlock

Change the watchdog locking scheme so wd_smp_lock protects the watchdog
internal data, but "reporting" (printing, issuing NMI IPIs, taking any
action outside of watchdog) uses a non-waiting exclusion. If a CPU detects
a problem but can not take the reporting lock, it just returns because
something else is already reporting. It will try again at some point.

Typically hard lockup watchdog report usefulness is not impacted due to
failure to spewing a large enough amount of data in as short a time as
possible, but by messages getting garbled.

Laurent debugged this and found the deadlock, and this patch is based on
his general approach to avoid expensive operations while holding the lock.
With the addition of the reporting exclusion.

Signed-off-by: Laurent Dufour <ldufour@linux.ibm.com>
[np: rework to add reporting exclusion update changelog]
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211110025056.2084347-4-npiggin@gmail.com
2 years agopowerpc/watchdog: tighten non-atomic read-modify-write access
Nicholas Piggin [Wed, 10 Nov 2021 02:50:54 +0000 (12:50 +1000)]
powerpc/watchdog: tighten non-atomic read-modify-write access

Most updates to wd_smp_cpus_pending are under lock except the watchdog
interrupt bit clear.

This can race with non-atomic RMW updates to the mask under lock, which
can happen in two instances:

Firstly, if another CPU detects this one is stuck, removes it from the
mask, mask becomes empty and is re-filled with non-atomic stores. This
is okay because it would re-fill the mask with this CPU's bit clear
anyway (because this CPU is now stuck), so it doesn't matter that the
bit clear update got "lost". Add a comment for this.

Secondly, if another CPU detects a different CPU is stuck and removes it
from the pending mask with a non-atomic store to bytes which also
include the bit of this CPU. This case can result in the bit clear being
lost and the end result being the bit is set. This should be so rare it
hardly matters, but to make things simpler to reason about just avoid
the non-atomic access for that case.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211110025056.2084347-3-npiggin@gmail.com
2 years agopowerpc/watchdog: Fix missed watchdog reset due to memory ordering race
Nicholas Piggin [Wed, 10 Nov 2021 02:50:53 +0000 (12:50 +1000)]
powerpc/watchdog: Fix missed watchdog reset due to memory ordering race

It is possible for all CPUs to miss the pending cpumask becoming clear,
and then nobody resetting it, which will cause the lockup detector to
stop working. It will eventually expire, but watchdog_smp_panic will
avoid doing anything if the pending mask is clear and it will never be
reset.

Order the cpumask clear vs the subsequent test to close this race.

Add an extra check for an empty pending mask when the watchdog fires and
finds its bit still clear, to try to catch any other possible races or
bugs here and keep the watchdog working. The extra test in
arch_touch_nmi_watchdog is required to prevent the new warning from
firing off.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com>
Debugged-by: Laurent Dufour <ldufour@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211110025056.2084347-2-npiggin@gmail.com
2 years agopowerpc/prom_init: Fix improper check of prom_getprop()
Peiwei Hu [Fri, 19 Nov 2021 09:12:18 +0000 (17:12 +0800)]
powerpc/prom_init: Fix improper check of prom_getprop()

prom_getprop() can return PROM_ERROR. Binary operator can not identify
it.

Fixes: 94d2dde738a5 ("[POWERPC] Efika: prune fixups and make them more carefull")
Signed-off-by: Peiwei Hu <jlu.hpw@foxmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/tencent_BA28CC6897B7C95A92EB8C580B5D18589105@qq.com
2 years agopowerpc/rtas: rtas_busy_delay_time() kernel-doc
Nathan Lynch [Wed, 17 Nov 2021 06:02:59 +0000 (00:02 -0600)]
powerpc/rtas: rtas_busy_delay_time() kernel-doc

Provide API documentation for rtas_busy_delay_time(), explaining why we
return the same value for 9900 and -2.

Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211117060259.957178-3-nathanl@linux.ibm.com
2 years agopowerpc/rtas: rtas_busy_delay() improvements
Nathan Lynch [Wed, 17 Nov 2021 06:02:58 +0000 (00:02 -0600)]
powerpc/rtas: rtas_busy_delay() improvements

Generally RTAS cannot block, and in PAPR it is required to return control
to the OS within a few tens of microseconds. In order to support operations
which may take longer to complete, many RTAS primitives can return
intermediate -2 ("busy") or 990x ("extended delay") values, which indicate
that the OS should reattempt the same call with the same arguments at some
point in the future.

Current versions of PAPR are less than clear about this, but the intended
meanings of these values in more detail are:

RTAS_BUSY (-2): RTAS has suspended a potentially long-running operation in
order to meet its latency obligation and give the OS the opportunity to
perform other work. RTAS can resume making progress as soon as the OS
reattempts the call.

RTAS_EXTENDED_DELAY_{MIN...MAX} (9900-9905): RTAS must wait for an external
event to occur or for internal contention to resolve before it can complete
the requested operation. The value encodes a non-binding hint as to roughly
how long the OS should wait before calling again, but the OS is allowed to
reattempt the call sooner or even immediately.

Linux of course must take its own CPU scheduling obligations into account
when handling these statuses; e.g. a task which receives an RTAS_BUSY
status should check whether to reschedule before it attempts the RTAS call
again to avoid starving other tasks.

rtas_busy_delay() is a helper function that "consumes" a busy or extended
delay status. Common usage:

    int rc;

    do {
        rc = rtas_call(rtas_token("some-function"), ...);
    } while (rtas_busy_delay(rc));

    /* convert rc to Linux error value, etc */

If rc is a busy or extended delay status, the caller can rely on
rtas_busy_delay() to perform an appropriate sleep or reschedule and return
nonzero. Other statuses are handled normally by the caller.

The current implementation of rtas_busy_delay() both oversleeps and
overuses the CPU:

*  It performs msleep() for all 990x and even when no delay is
   suggested (-2), but this is understood to actually sleep for two jiffies
   minimum in practice (20ms with HZ=100). 9900 (1ms) and 9901 (10ms)
   appear to be the most common extended delay statuses, and the
   oversleeping measurably lengthens DLPAR operations, which perform
   many RTAS calls.

*  It does not sleep on 990x unless need_resched() is true, causing code
   like the loop above to needlessly retry, wasting CPU time.

Alter the logic to align better with the intended meanings:

*  When passed RTAS_BUSY, perform cond_resched() and return without
   sleeping. The caller should reattempt immediately

*  Always sleep when passed an extended delay status, using usleep_range()
   for precise shorter sleeps. Limit the sleep time to one second even
   though there are higher architected values.

Change rtas_busy_delay()'s return type to bool to better reflect its usage,
and add kernel-doc.

rtas_busy_delay_time() is unchanged, even though it "incorrectly" returns 1
for RTAS_BUSY. There are users of that API with open-coded delay loops in
sensitive contexts that will have to be taken on an individual basis.

Brief results for addition and removal of 5GB memory on a small P9 PowerVM
partition follow. Load was generated with stress-ng --cpu N. For add,
elapsed time is greatly reduced without significant change in the number of
RTAS calls or time spent on CPU. For remove, elapsed time is modestly
reduced, with significant reductions in RTAS calls and time spent on CPU.

With no competing workload (- before, + after):

  Performance counter stats for 'bash -c echo "memory add count 20" > /sys/kernel/dlpar' (10 runs):

-             1,935      probe:rtas_call           #    0.003 M/sec                    ( +-  0.22% )
-            609.99 msec task-clock                #    0.183 CPUs utilized            ( +-  0.19% )
+             1,956      probe:rtas_call           #    0.003 M/sec                    ( +-  0.17% )
+            618.56 msec task-clock                #    0.278 CPUs utilized            ( +-  0.11% )

-            3.3322 +- 0.0670 seconds time elapsed  ( +-  2.01% )
+            2.2222 +- 0.0416 seconds time elapsed  ( +-  1.87% )

  Performance counter stats for 'bash -c echo "memory remove count 20" > /sys/kernel/dlpar' (10 runs):

-             6,224      probe:rtas_call           #    0.008 M/sec                    ( +-  2.57% )
-            750.36 msec task-clock                #    0.190 CPUs utilized            ( +-  2.01% )
+               843      probe:rtas_call           #    0.003 M/sec                    ( +-  0.12% )
+            250.66 msec task-clock                #    0.068 CPUs utilized            ( +-  0.17% )

-            3.9394 +- 0.0890 seconds time elapsed  ( +-  2.26% )
+             3.678 +- 0.113 seconds time elapsed  ( +-  3.07% )

With all CPUs 100% busy (- before, + after):

  Performance counter stats for 'bash -c echo "memory add count 20" > /sys/kernel/dlpar' (10 runs):

-             2,979      probe:rtas_call           #    0.003 M/sec                    ( +-  0.12% )
-          1,096.62 msec task-clock                #    0.105 CPUs utilized            ( +-  0.10% )
+             2,981      probe:rtas_call           #    0.003 M/sec                    ( +-  0.22% )
+          1,095.26 msec task-clock                #    0.154 CPUs utilized            ( +-  0.21% )

-            10.476 +- 0.104 seconds time elapsed  ( +-  1.00% )
+            7.1124 +- 0.0865 seconds time elapsed  ( +-  1.22% )

  Performance counter stats for 'bash -c echo "memory remove count 20" > /sys/kernel/dlpar' (10 runs):

-             2,702      probe:rtas_call           #    0.004 M/sec                    ( +-  4.00% )
-            722.71 msec task-clock                #    0.067 CPUs utilized            ( +-  2.41% )
+             1,246      probe:rtas_call           #    0.003 M/sec                    ( +-  0.25% )
+            487.73 msec task-clock                #    0.049 CPUs utilized            ( +-  0.20% )

-            10.829 +- 0.163 seconds time elapsed  ( +-  1.51% )
+            9.9887 +- 0.0866 seconds time elapsed  ( +-  0.87% )

Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211117060259.957178-2-nathanl@linux.ibm.com
2 years agopowerpc/pseries: delete scanlog
Nathan Lynch [Mon, 20 Sep 2021 17:32:03 +0000 (12:32 -0500)]
powerpc/pseries: delete scanlog

Remove the pseries scanlog driver.

This code supports functions from Power4-era servers that are not present
on targets currently supported by arch/powerpc. System manuals from this
time have this description:

  Scan Dump data is a set of chip data that the service processor gathers
  after a system malfunction. It consists of chip scan rings, chip trace
  arrays, and Scan COM (SCOM) registers. This data is stored in the
  scan-log partition of the system’s Nonvolatile Random Access
  Memory (NVRAM).

PowerVM partition firmware development doesn't recognize the associated
function call or property, and they don't see any references to them in
their codebase. It seems to have been specific to non-virtualized pseries.

References:

Historical Linux commit from February 2003 (interesting to note this seems
to be the source of non-GPL exports for rtas_call etc):
https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git/commit/?id=f92e361842d5251e50562b09664082dcbd0548bb

IntelliStation and pSeries docs which refer to the feature:
http://ps-2.retropc.se/basil.holloway/ALL%20PDF/380635.pdf
http://ps-2.kev009.com/rs6000/manuals/p/p615-6C3-6E3/6C3_and_6E3_Users_Guide_SA38-0629.pdf

Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Reviewed-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210920173203.1800475-1-nathanl@linux.ibm.com
2 years agopowerpc/rtas: kernel-doc fixes
Nathan Lynch [Tue, 16 Nov 2021 21:58:06 +0000 (15:58 -0600)]
powerpc/rtas: kernel-doc fixes

Fix the following issues reported by kernel-doc:

$ scripts/kernel-doc -v -none arch/powerpc/kernel/rtas.c
arch/powerpc/kernel/rtas.c:810: info: Scanning doc for function rtas_activate_firmware
arch/powerpc/kernel/rtas.c:818: warning: contents before sections
arch/powerpc/kernel/rtas.c:841: info: Scanning doc for function rtas_call_reentrant
arch/powerpc/kernel/rtas.c:893: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
 * Find a specific pseries error log in an RTAS extended event log.

Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211116215806.928235-1-nathanl@linux.ibm.com
2 years agopowerpc/code-patching: Improve verification of patchability
Christophe Leroy [Mon, 15 Nov 2021 10:12:22 +0000 (11:12 +0100)]
powerpc/code-patching: Improve verification of patchability

Today, patch_instruction() assumes that it is called exclusively on
valid addresses, and only checks that it is not called on an init
address after init section has been freed.

Improve verification by calling kernel_text_address() instead.

kernel_text_address() already includes a verification of
initmem release.

Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/bc683d499a411730504b132a924de0ccc2ef1f79.1636971137.git.christophe.leroy@csgroup.eu
2 years agopowerpc/tsi108: make EXPORT_SYMBOL follow its function immediately
Jason Wang [Sun, 14 Nov 2021 11:56:16 +0000 (19:56 +0800)]
powerpc/tsi108: make EXPORT_SYMBOL follow its function immediately

EXPORT_SYMBOL(foo); should immediately follow its function/variable.

Signed-off-by: Jason Wang <wangborong@cdjrlc.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211114115616.493815-1-wangborong@cdjrlc.com
2 years agobpf ppc32: Access only if addr is kernel address
Hari Bathini [Tue, 12 Oct 2021 12:30:56 +0000 (18:00 +0530)]
bpf ppc32: Access only if addr is kernel address

With KUAP enabled, any kernel code which wants to access userspace
needs to be surrounded by disable-enable KUAP. But that is not
happening for BPF_PROBE_MEM load instruction. Though PPC32 does not
support read protection, considering the fact that PTR_TO_BTF_ID
(which uses BPF_PROBE_MEM mode) could either be a valid kernel pointer
or NULL but should never be a pointer to userspace address, execute
BPF_PROBE_MEM load only if addr is kernel address, otherwise set
dst_reg=0 and move on.

This will catch NULL, valid or invalid userspace pointers. Only bad
kernel pointer will be handled by BPF exception table.

[Alexei suggested for x86]

Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-9-hbathini@linux.ibm.com
2 years agobpf ppc32: Add BPF_PROBE_MEM support for JIT
Hari Bathini [Tue, 12 Oct 2021 12:30:55 +0000 (18:00 +0530)]
bpf ppc32: Add BPF_PROBE_MEM support for JIT

BPF load instruction with BPF_PROBE_MEM mode can cause a fault
inside kernel. Append exception table for such instructions
within BPF program.

Unlike other archs which uses extable 'fixup' field to pass dest_reg
and nip, BPF exception table on PowerPC follows the generic PowerPC
exception table design, where it populates both fixup and extable
sections within BPF program. fixup section contains 3 instructions,
first 2 instructions clear dest_reg (lower & higher 32-bit registers)
and last instruction jumps to next instruction in the BPF code.
extable 'insn' field contains relative offset of the instruction and
'fixup' field contains relative offset of the fixup entry. Example
layout of BPF program with extable present:

             +------------------+
             |                  |
             |                  |
   0x4020 -->| lwz   r28,4(r4)  |
             |                  |
             |                  |
   0x40ac -->| lwz  r3,0(r24)   |
             | lwz  r4,4(r24)   |
             |                  |
             |                  |
             |------------------|
   0x4278 -->| li  r28,0        |  \
             | li  r27,0        |  | fixup entry
             | b   0x4024       |  /
   0x4284 -->| li  r4,0         |
             | li  r3,0         |
             | b   0x40b4       |
             |------------------|
   0x4290 -->| insn=0xfffffd90  |  \ extable entry
             | fixup=0xffffffe4 |  /
   0x4298 -->| insn=0xfffffe14  |
             | fixup=0xffffffe8 |
             +------------------+

   (Addresses shown here are chosen random, not real)

Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-8-hbathini@linux.ibm.com
2 years agobpf ppc64: Access only if addr is kernel address
Ravi Bangoria [Tue, 12 Oct 2021 12:30:54 +0000 (18:00 +0530)]
bpf ppc64: Access only if addr is kernel address

On PPC64 with KUAP enabled, any kernel code which wants to
access userspace needs to be surrounded by disable-enable KUAP.
But that is not happening for BPF_PROBE_MEM load instruction.
So, when BPF program tries to access invalid userspace address,
page-fault handler considers it as bad KUAP fault:

  Kernel attempted to read user page (d0000000) - exploit attempt? (uid: 0)

Considering the fact that PTR_TO_BTF_ID (which uses BPF_PROBE_MEM
mode) could either be a valid kernel pointer or NULL but should
never be a pointer to userspace address, execute BPF_PROBE_MEM load
only if addr is kernel address, otherwise set dst_reg=0 and move on.

This will catch NULL, valid or invalid userspace pointers. Only bad
kernel pointer will be handled by BPF exception table.

[Alexei suggested for x86]

Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-7-hbathini@linux.ibm.com
2 years agobpf ppc64: Add BPF_PROBE_MEM support for JIT
Ravi Bangoria [Tue, 12 Oct 2021 12:30:53 +0000 (18:00 +0530)]
bpf ppc64: Add BPF_PROBE_MEM support for JIT

BPF load instruction with BPF_PROBE_MEM mode can cause a fault
inside kernel. Append exception table for such instructions
within BPF program.

Unlike other archs which uses extable 'fixup' field to pass dest_reg
and nip, BPF exception table on PowerPC follows the generic PowerPC
exception table design, where it populates both fixup and extable
sections within BPF program. fixup section contains two instructions,
first instruction clears dest_reg and 2nd jumps to next instruction
in the BPF code. extable 'insn' field contains relative offset of
the instruction and 'fixup' field contains relative offset of the
fixup entry. Example layout of BPF program with extable present:

             +------------------+
             |                  |
             |                  |
   0x4020 -->| ld   r27,4(r3)   |
             |                  |
             |                  |
   0x40ac -->| lwz  r3,0(r4)    |
             |                  |
             |                  |
             |------------------|
   0x4280 -->| li  r27,0        |  \ fixup entry
             | b   0x4024       |  /
   0x4288 -->| li  r3,0         |
             | b   0x40b0       |
             |------------------|
   0x4290 -->| insn=0xfffffd90  |  \ extable entry
             | fixup=0xffffffec |  /
   0x4298 -->| insn=0xfffffe14  |
             | fixup=0xffffffec |
             +------------------+

   (Addresses shown here are chosen random, not real)

Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-6-hbathini@linux.ibm.com
2 years agopowerpc/ppc-opcode: introduce PPC_RAW_BRANCH() macro
Hari Bathini [Tue, 12 Oct 2021 12:30:52 +0000 (18:00 +0530)]
powerpc/ppc-opcode: introduce PPC_RAW_BRANCH() macro

Define and use PPC_RAW_BRANCH() macro instead of open coding it. This
macro is used while adding BPF_PROBE_MEM support.

Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-5-hbathini@linux.ibm.com
2 years agobpf powerpc: refactor JIT compiler code
Hari Bathini [Tue, 12 Oct 2021 12:30:51 +0000 (18:00 +0530)]
bpf powerpc: refactor JIT compiler code

Refactor powerpc LDX JITing code to simplify adding BPF_PROBE_MEM
support.

Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-4-hbathini@linux.ibm.com
2 years agobpf powerpc: Remove extra_pass from bpf_jit_build_body()
Ravi Bangoria [Tue, 12 Oct 2021 12:30:50 +0000 (18:00 +0530)]
bpf powerpc: Remove extra_pass from bpf_jit_build_body()

In case of extra_pass, usual JIT passes are always skipped. So,
extra_pass is always false while calling bpf_jit_build_body() and
can be removed.

Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-3-hbathini@linux.ibm.com
2 years agobpf powerpc: Remove unused SEEN_STACK
Ravi Bangoria [Tue, 12 Oct 2021 12:30:49 +0000 (18:00 +0530)]
bpf powerpc: Remove unused SEEN_STACK

SEEN_STACK is unused on PowerPC. Remove it. Also, have
SEEN_TAILCALL use 0x40000000.

Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211012123056.485795-2-hbathini@linux.ibm.com
2 years agopowerpc/eeh: Use a goto for recovery failures
Oliver O'Halloran [Fri, 15 Oct 2021 07:06:28 +0000 (18:06 +1100)]
powerpc/eeh: Use a goto for recovery failures

The EEH recovery logic in eeh_handle_normal_event() has some pretty strange
flow control. If we remove all the actual recovery logic we're left with
the following skeleton:

if (result != PCI_ERS_RESULT_DISCONNECT) {
...
}

if (result != PCI_ERS_RESULT_DISCONNECT) {
...
}

if (result == PCI_ERS_RESULT_NONE) {
...
}

if (result == PCI_ERS_RESULT_CAN_RECOVER) {
...
}

if (result == PCI_ERS_RESULT_CAN_RECOVER) {
...
}

if (result == PCI_ERS_RESULT_NEED_RESET) {
...
}

if ((result == PCI_ERS_RESULT_RECOVERED) ||
    (result == PCI_ERS_RESULT_NONE)) {
...
goto out;
}

/*
 * unsuccessful recovery / PCI_ERS_RESULT_DISCONECTED
 * handling is here.
 */
...

out:
...

Most of the "if () { ... }" blocks above change "result" to
PCI_ERS_RESULT_DISCONNECTED if an error occurs in that recovery step. This
makes the control flow a bit confusing since it breaks the early-exit
pattern that is generally used in Linux. In any case we end up handling the
error in the final else block so why not just jump there directly? Doing so
also allows us to de-indent a bunch of code.

No functional changes.

[dja: rebase on top of linux-next + my preceeding refactor,
      move clearing the EEH_DEV_NO_HANDLER bit above the first goto so that
      it is always clear in the error handler code as it was before.]

Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211015070628.1331635-2-dja@axtens.net
2 years agopowerpc/eeh: Small refactor of eeh_handle_normal_event()
Daniel Axtens [Fri, 15 Oct 2021 07:06:27 +0000 (18:06 +1100)]
powerpc/eeh: Small refactor of eeh_handle_normal_event()

The control flow of eeh_handle_normal_event() is a bit tricky.

Break out one of the error handling paths - rather than be in an else
block, we'll make it part of the regular body of the function and put a
'goto out;' in the true limb of the if.

Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211015070628.1331635-1-dja@axtens.net
2 years agopowerpc/xive: Add a debugfs toggle for save-restore
Cédric Le Goater [Fri, 5 Nov 2021 10:26:35 +0000 (11:26 +0100)]
powerpc/xive: Add a debugfs toggle for save-restore

On POWER10, the automatic "save & restore" of interrupt context is
always available. Provide a way to deactivate it for tests or
performance.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-11-clg@kaod.org
2 years agopowerpc/xive: Add a kernel parameter for StoreEOI
Cédric Le Goater [Fri, 5 Nov 2021 10:26:34 +0000 (11:26 +0100)]
powerpc/xive: Add a kernel parameter for StoreEOI

StoreEOI is activated by default on platforms supporting the feature
(POWER10) and will be used as soon as firmware advertises its
availability. The kernel parameter provides a way to deactivate its
use. It can be still be reactivated through debugfs.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-10-clg@kaod.org
2 years agopowerpc/xive: Add a debugfs toggle for StoreEOI
Cédric Le Goater [Fri, 5 Nov 2021 10:26:33 +0000 (11:26 +0100)]
powerpc/xive: Add a debugfs toggle for StoreEOI

It can be used to deactivate temporarily StoreEOI for tests or
performance on platforms supporting the feature (POWER10)

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-9-clg@kaod.org
2 years agopowerpc/xive: Add a debugfs file to dump EQs
Cédric Le Goater [Fri, 5 Nov 2021 10:26:32 +0000 (11:26 +0100)]
powerpc/xive: Add a debugfs file to dump EQs

The XIVE driver under Linux uses a single interrupt priority and only
one event queue is configured per CPU. Expose the contents under
a 'xive/eqs/cpuX' debugfs file.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-8-clg@kaod.org
2 years agopowerpc/xive: Rename the 'cpus' debugfs file to 'ipis'
Cédric Le Goater [Fri, 5 Nov 2021 10:26:31 +0000 (11:26 +0100)]
powerpc/xive: Rename the 'cpus' debugfs file to 'ipis'

and remove the EQ entries output which is not very useful since only
the next two events of the queue are taken into account. We will
improve the dump of the EQ in the next patches.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-7-clg@kaod.org
2 years agopowerpc/xive: Change the debugfs file 'xive' into a directory
Cédric Le Goater [Fri, 5 Nov 2021 10:26:30 +0000 (11:26 +0100)]
powerpc/xive: Change the debugfs file 'xive' into a directory

Use a 'cpus' file to dump CPU states and 'interrupts' to dump IRQ states.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-6-clg@kaod.org
2 years agopowerpc/xive: Introduce xive_core_debugfs_create()
Cédric Le Goater [Fri, 5 Nov 2021 10:26:29 +0000 (11:26 +0100)]
powerpc/xive: Introduce xive_core_debugfs_create()

and fix some compile issues when !CONFIG_DEBUG_FS.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
[mpe: Add empty stub to fix !CONFIG_DEBUG_FS build]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-5-clg@kaod.org
2 years agopowerpc/xive: Activate StoreEOI on P10
Cédric Le Goater [Fri, 5 Nov 2021 10:26:28 +0000 (11:26 +0100)]
powerpc/xive: Activate StoreEOI on P10

StoreEOI (the capability to EOI with a store) requires load-after-store
ordering in some cases to be reliable. P10 introduced a new offset for
load operations to enforce correct ordering and the XIVE driver has
the required support since kernel 5.8, commit b1f9be9392f0
("powerpc/xive: Enforce load-after-store ordering when StoreEOI is active")

Since skiboot v7, StoreEOI support is advertised on P10 with a new flag
on the PowerNV platform. See skiboot commit 4bd7d84afe46 ("xive/p10:
Introduce a new OPAL_XIVE_IRQ_STORE_EOI2 flag"). When detected,
activate the feature.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-4-clg@kaod.org
2 years agopowerpc/xive: Introduce an helper to print out interrupt characteristics
Cédric Le Goater [Fri, 5 Nov 2021 10:26:27 +0000 (11:26 +0100)]
powerpc/xive: Introduce an helper to print out interrupt characteristics

and extend output of debugfs and xmon with addresses of the ESB
management and trigger pages.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-3-clg@kaod.org
2 years agopowerpc/xive: Replace pr_devel() by pr_debug() to ease debug
Cédric Le Goater [Fri, 5 Nov 2021 10:26:26 +0000 (11:26 +0100)]
powerpc/xive: Replace pr_devel() by pr_debug() to ease debug

These routines are not on hot code paths and pr_debug() is easier to
activate. Also add a '0x' prefix to hex printed values (HW IRQ number).

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211105102636.1016378-2-clg@kaod.org
2 years agopowerpc/powernv: Remove POWER9 PVR version check for entry and uaccess flushes
Nicholas Piggin [Mon, 3 May 2021 13:02:43 +0000 (23:02 +1000)]
powerpc/powernv: Remove POWER9 PVR version check for entry and uaccess flushes

These aren't necessarily POWER9 only, and it's not to say some new
vulnerability may not get discovered on other processors for which
we would like the flexibility of having the workaround enabled by
firmware.

Remove the restriction that the workarounds only apply to POWER9.

However POWER7 and POWER8 are not affected, and they may not have
older firmware that does not advertise this, so clear these workarounds
manually.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
[mpe: Incorporate changes from Nick, reword comment slightly.]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210503130243.891868-5-npiggin@gmail.com
2 years agopowerpc/btext: add missing of_node_put
Julia Lawall [Fri, 20 Nov 2015 20:33:23 +0000 (20:33 +0000)]
powerpc/btext: add missing of_node_put

for_each_node_by_type performs an of_node_get on each iteration, so
a break out of the loop requires an of_node_put.

A simplified version of the semantic patch that fixes this problem is as
follows (http://coccinelle.lip6.fr):

// <smpl>
@@
local idexpression n;
expression e;
@@

 for_each_node_by_type(n,...) {
   ...
(
   of_node_put(n);
|
   e = n
|
+  of_node_put(n);
?  break;
)
   ...
 }
... when != n
// </smpl>

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1448051604-25256-6-git-send-email-Julia.Lawall@lip6.fr
2 years agopowerpc/cell: add missing of_node_put
Julia Lawall [Fri, 20 Nov 2015 20:33:24 +0000 (21:33 +0100)]
powerpc/cell: add missing of_node_put

for_each_node_by_name performs an of_node_get on each iteration, so
a break out of the loop requires an of_node_put.

A simplified version of the semantic patch that fixes this problem is as
follows (http://coccinelle.lip6.fr):

// <smpl>
@@
expression e,e1;
local idexpression n;
@@

 for_each_node_by_name(n, e1) {
   ... when != of_node_put(n)
       when != e = n
(
   return n;
|
+  of_node_put(n);
?  return ...;
)
   ...
 }
// </smpl>

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1448051604-25256-7-git-send-email-Julia.Lawall@lip6.fr
2 years agopowerpc/powernv: add missing of_node_put
Julia Lawall [Fri, 20 Nov 2015 20:33:21 +0000 (20:33 +0000)]
powerpc/powernv: add missing of_node_put

for_each_compatible_node performs an of_node_get on each iteration, so
a break out of the loop requires an of_node_put.

A simplified version of the semantic patch that fixes this problem is as
follows (http://coccinelle.lip6.fr):

// <smpl>
@@
local idexpression n;
expression e;
@@

 for_each_compatible_node(n,...) {
   ...
(
   of_node_put(n);
|
   e = n
|
+  of_node_put(n);
?  break;
)
   ...
 }
... when != n
// </smpl>

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1448051604-25256-4-git-send-email-Julia.Lawall@lip6.fr
2 years agopowerpc/6xx: add missing of_node_put
Julia Lawall [Fri, 20 Nov 2015 20:33:19 +0000 (20:33 +0000)]
powerpc/6xx: add missing of_node_put

for_each_compatible_node performs an of_node_get on each iteration, so
a break out of the loop requires an of_node_put.

A simplified version of the semantic patch that fixes this problem is as
follows (http://coccinelle.lip6.fr):

// <smpl>
@@
expression e;
local idexpression n;
@@

@@
local idexpression n;
expression e;
@@

 for_each_compatible_node(n,...) {
   ...
(
   of_node_put(n);
|
   e = n
|
+  of_node_put(n);
?  break;
)
   ...
 }
... when != n
// </smpl>

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1448051604-25256-2-git-send-email-Julia.Lawall@lip6.fr
2 years agoMerge branch 'topic/ppc-kvm' into next
Michael Ellerman [Thu, 25 Nov 2021 00:23:59 +0000 (11:23 +1100)]
Merge branch 'topic/ppc-kvm' into next

This merge's Nick's big P9 KVM series, original cover letter follows:

KVM: PPC: Book3S HV P9: entry/exit optimisations

This reduces radix guest full entry/exit latency on POWER9 and POWER10
by 2x.

Nested HV guests should see smaller improvements in their L1 entry/exit,
but this is also combined with most L0 speedups also applying to nested
entry. nginx localhost throughput test in a SMP nested guest is improved
about 10% (in a direct guest it doesn't change much because it uses XIVE
for IPIs) when L0 and L1 are patched.

It does this in several main ways:

- Rearrange code to optimise SPR accesses. Mainly, avoid scoreboard
  stalls.

- Test SPR values to avoid mtSPRs where possible. mtSPRs are expensive.

- Reduce mftb. mftb is expensive.

- Demand fault certain facilities to avoid saving and/or restoring them
  (at the cost of fault when they are used, but this is mitigated over
  a number of entries, like the facilities when context switching
  processes). PM, TM, and EBB so far.

- Defer some sequences that are made just in case a guest is interrupted
  in the middle of a critical section to the case where the guest is
  scheduled on a different CPU, rather than every time (at the cost of
  an extra IPI in this case). Namely the tlbsync sequence for radix with
  GTSE, which is very expensive.

- Reduce locking, barriers, atomics related to the vcpus-per-vcore > 1
  handling that the P9 path does not require.

2 years agoKVM: PPC: Book3S HV P9: Remove subcore HMI handling
Nicholas Piggin [Tue, 23 Nov 2021 09:52:31 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Remove subcore HMI handling

On POWER9 and newer, rather than the complex HMI synchronisation and
subcore state, have each thread un-apply the guest TB offset before
calling into the early HMI handler.

This allows the subcore state to be avoided, including subcore enter
/ exit guest, which includes an expensive divide that shows up
slightly in profiles.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-54-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Stop using vc->dpdes
Nicholas Piggin [Tue, 23 Nov 2021 09:52:30 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Stop using vc->dpdes

The P9 path uses vc->dpdes only for msgsndp / SMT emulation. This adds
an ordering requirement between vcpu->doorbell_request and vc->dpdes for
no real benefit. Use vcpu->doorbell_request directly.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-53-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Tidy kvmppc_create_dtl_entry
Nicholas Piggin [Tue, 23 Nov 2021 09:52:29 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Tidy kvmppc_create_dtl_entry

This goes further to removing vcores from the P9 path. Also avoid the
memset in favour of explicitly initialising all fields.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-52-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Remove most of the vcore logic
Nicholas Piggin [Tue, 23 Nov 2021 09:52:28 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Remove most of the vcore logic

The P9 path always uses one vcpu per vcore, so none of the vcore, locks,
stolen time, blocking logic, shared waitq, etc., is required.

Remove most of it.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-51-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Avoid cpu_in_guest atomics on entry and exit
Nicholas Piggin [Tue, 23 Nov 2021 09:52:27 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Avoid cpu_in_guest atomics on entry and exit

cpu_in_guest is set to determine if a CPU needs to be IPI'ed to exit
the guest and notice the need_tlb_flush bit.

This can be implemented as a global per-CPU pointer to the currently
running guest instead of per-guest cpumasks, saving 2 atomics per
entry/exit. P7/8 doesn't require cpu_in_guest, nor does a nested HV
(only the L0 does), so move it to the P9 HV path.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-50-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Add unlikely annotation for !mmu_ready
Nicholas Piggin [Tue, 23 Nov 2021 09:52:26 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Add unlikely annotation for !mmu_ready

The mmu will almost always be ready.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-49-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Avoid changing MSR[RI] in entry and exit
Nicholas Piggin [Tue, 23 Nov 2021 09:52:25 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Avoid changing MSR[RI] in entry and exit

kvm_hstate.in_guest provides the equivalent of MSR[RI]=0 protection,
and it covers the existing MSR[RI]=0 section in late entry and early
exit, so clearing and setting MSR[RI] in those cases does not
actually do anything useful.

Remove the RI manipulation and replace it with comments. Make the
in_guest memory accesses a bit closer to a proper critical section
pattern. This speeds up guest entry/exit performance.

This also removes the MSR[RI] warnings which aren't very interesting
and would cause crashes if they hit due to causing an interrupt in
non-recoverable code.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-48-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Optimise hash guest SLB saving
Nicholas Piggin [Tue, 23 Nov 2021 09:52:24 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Optimise hash guest SLB saving

slbmfee/slbmfev instructions are very expensive, moreso than a regular
mfspr instruction, so minimising them significantly improves hash guest
exit performance. The slbmfev is only required if slbmfee found a valid
SLB entry.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-47-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Improve mfmsr performance on entry
Nicholas Piggin [Tue, 23 Nov 2021 09:52:23 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Improve mfmsr performance on entry

Rearrange the MSR saving on entry so it does not follow the mtmsrd to
disable interrupts, avoiding a possible RAW scoreboard stall.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-46-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV Nested: Avoid extra mftb() in nested entry
Nicholas Piggin [Tue, 23 Nov 2021 09:52:22 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV Nested: Avoid extra mftb() in nested entry

mftb() is expensive and one can be avoided on nested guest dispatch.

If the time checking code distinguishes between the L0 timer and the
nested HV timer, then both can be tested in the same place with the
same mftb() value.

This also nicely illustrates the relationship between the L0 and nested
HV timers.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-45-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Avoid tlbsync sequence on radix guest exit
Nicholas Piggin [Tue, 23 Nov 2021 09:52:21 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Avoid tlbsync sequence on radix guest exit

Use the existing TLB flushing logic to IPI the previous CPU and run the
necessary barriers before running a guest vCPU on a new physical CPU,
to do the necessary radix GTSE barriers for handling the case of an
interrupted guest tlbie sequence.

This requires the vCPU TLB flush sequence that is currently just done
on one thread, to be expanded to ensure the other threads execute a
ptesync, because causing them to exit the guest will no longer cause a
ptesync by itself.

This results in more IPIs than the TLB flush logic requires, but it's
a significant win for common case scheduling when the vCPU remains on
the same physical CPU.

This saves about 520 cycles (nearly 10%) on a guest entry+exit micro
benchmark on a POWER9.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-44-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV: Split P8 from P9 path guest vCPU TLB flushing
Nicholas Piggin [Tue, 23 Nov 2021 09:52:20 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV: Split P8 from P9 path guest vCPU TLB flushing

This creates separate functions for old and new paths for vCPU TLB
flushing, which will reduce complexity of the next change.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-43-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Don't restore PSSCR if not needed
Nicholas Piggin [Tue, 23 Nov 2021 09:52:19 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Don't restore PSSCR if not needed

This also moves the PSSCR update in nested entry to avoid a SPR
scoreboard stall.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-42-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Test dawr_enabled() before saving host DAWR SPRs
Nicholas Piggin [Tue, 23 Nov 2021 09:52:18 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Test dawr_enabled() before saving host DAWR SPRs

Some of the DAWR SPR access is already predicated on dawr_enabled(),
apply this to the remainder of the accesses.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-41-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Comment and fix MMU context switching code
Nicholas Piggin [Tue, 23 Nov 2021 09:52:17 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Comment and fix MMU context switching code

Tighten up partition switching code synchronisation and comments.

In particular, hwsync ; isync is required after the last access that is
performed in the context of a partition, before the partition is
switched away from.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-40-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Use Linux SPR save/restore to manage some host SPRs
Nicholas Piggin [Tue, 23 Nov 2021 09:52:16 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Use Linux SPR save/restore to manage some host SPRs

Linux implements SPR save/restore including storage space for registers
in the task struct for process context switching. Make use of this
similarly to the way we make use of the context switching fp/vec save
restore.

This improves code reuse, allows some stack space to be saved, and helps
with avoiding VRSAVE updates if they are not required.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-39-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Demand fault TM facility registers
Nicholas Piggin [Tue, 23 Nov 2021 09:52:15 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Demand fault TM facility registers

Use HFSCR facility disabling to implement demand faulting for TM, with
a hysteresis counter similar to the load_fp etc counters in context
switching that implement the equivalent demand faulting for userspace
facilities.

This speeds up guest entry/exit by avoiding the register save/restore
when a guest is not frequently using them. When a guest does use them
often, there will be some additional demand fault overhead, but these
are not commonly used facilities.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-38-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Demand fault EBB facility registers
Nicholas Piggin [Tue, 23 Nov 2021 09:52:14 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Demand fault EBB facility registers

Use HFSCR facility disabling to implement demand faulting for EBB, with
a hysteresis counter similar to the load_fp etc counters in context
switching that implement the equivalent demand faulting for userspace
facilities.

This speeds up guest entry/exit by avoiding the register save/restore
when a guest is not frequently using them. When a guest does use them
often, there will be some additional demand fault overhead, but these
are not commonly used facilities.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-37-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: More SPR speed improvements
Nicholas Piggin [Tue, 23 Nov 2021 09:52:13 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: More SPR speed improvements

This avoids more scoreboard stalls and reduces mtSPRs.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-36-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Restrict DSISR canary workaround to processors that require it
Nicholas Piggin [Tue, 23 Nov 2021 09:52:12 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Restrict DSISR canary workaround to processors that require it

Use CPU_FTR_P9_RADIX_PREFETCH_BUG to apply the workaround, to test for
DD2.1 and below processors. This saves a mtSPR in guest entry.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-35-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Switch PMU to guest as late as possible
Nicholas Piggin [Tue, 23 Nov 2021 09:52:11 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Switch PMU to guest as late as possible

This moves PMU switch to guest as late as possible in entry, and switch
back to host as early as possible at exit. This helps the host get the
most perf coverage of KVM entry/exit code as possible.

This is slightly suboptimal for SPR scheduling point of view when the
PMU is enabled, but when perf is disabled there is no real difference.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-34-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Implement TM fastpath for guest entry/exit
Nicholas Piggin [Tue, 23 Nov 2021 09:52:10 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Implement TM fastpath for guest entry/exit

If TM is not active, only TM register state needs to be saved and
restored, avoiding several mfmsr/mtmsrd instructions and improving
performance.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-33-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Move remaining SPR and MSR access into low level entry
Nicholas Piggin [Tue, 23 Nov 2021 09:52:09 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Move remaining SPR and MSR access into low level entry

Move register saving and loading from kvmhv_p9_guest_entry() into the HV
and nested entry handlers.

Accesses are scheduled to reduce mtSPR / mfSPR interleaving which
reduces SPR scoreboard stalls.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-32-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Move nested guest entry into its own function
Nicholas Piggin [Tue, 23 Nov 2021 09:52:08 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Move nested guest entry into its own function

Move the part of the guest entry which is specific to nested HV into its
own function. This is just refactoring.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-31-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Move host OS save/restore functions to built-in
Nicholas Piggin [Tue, 23 Nov 2021 09:52:07 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Move host OS save/restore functions to built-in

Move the P9 guest/host register switching functions to the built-in
P9 entry code, and export it for nested to use as well.

This allows more flexibility in scheduling these supervisor privileged
SPR accesses with the HV privileged and PR SPR accesses in the low level
entry code.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-30-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Move vcpu register save/restore into functions
Nicholas Piggin [Tue, 23 Nov 2021 09:52:06 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Move vcpu register save/restore into functions

This should be no functional difference but makes the caller easier
to read.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-29-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Juggle SPR switching around
Nicholas Piggin [Tue, 23 Nov 2021 09:52:05 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Juggle SPR switching around

This juggles SPR switching on the entry and exit sides to be more
symmetric, which makes the next refactoring patch possible with no
functional change.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-28-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Only execute mtSPR if the value changed
Nicholas Piggin [Tue, 23 Nov 2021 09:52:04 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Only execute mtSPR if the value changed

Keep better track of the current SPR value in places where
they are to be loaded with a new context, to reduce expensive
mtSPR operations.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-27-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Avoid SPR scoreboard stalls
Nicholas Piggin [Tue, 23 Nov 2021 09:52:03 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Avoid SPR scoreboard stalls

Avoid interleaving mfSPR and mtSPR to reduce SPR scoreboard stalls.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-26-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Optimise timebase reads
Nicholas Piggin [Tue, 23 Nov 2021 09:52:02 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Optimise timebase reads

Reduce the number of mfTB executed by passing the current timebase
around entry and exit code rather than read it multiple times.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-25-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Move TB updates
Nicholas Piggin [Tue, 23 Nov 2021 09:52:01 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV P9: Move TB updates

Move the TB updates between saving and loading guest and host SPRs,
to improve scheduling by keeping issue-NTC operations together as
much as possible.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-24-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV: Change dec_expires to be relative to guest timebase
Nicholas Piggin [Tue, 23 Nov 2021 09:52:00 +0000 (19:52 +1000)]
KVM: PPC: Book3S HV: Change dec_expires to be relative to guest timebase

Change dec_expires to be relative to the guest timebase, and allow
it to be moved into low level P9 guest entry functions, to improve
SPR access scheduling.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-23-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Add kvmppc_stop_thread to match kvmppc_start_thread
Nicholas Piggin [Tue, 23 Nov 2021 09:51:59 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Add kvmppc_stop_thread to match kvmppc_start_thread

Small cleanup makes it a bit easier to match up entry and exit
operations.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-22-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Improve mtmsrd scheduling by delaying MSR[EE] disable
Nicholas Piggin [Tue, 23 Nov 2021 09:51:58 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Improve mtmsrd scheduling by delaying MSR[EE] disable

Moving the mtmsrd after the host SPRs are saved and before the guest
SPRs start to be loaded can prevent an SPR scoreboard stall (because
the mtmsrd is L=1 type which does not cause context synchronisation.

This is also now more convenient to combined with the mtmsrd L=0
instruction to enable facilities just below, but that is not done yet.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-21-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Reduce mtmsrd instructions required to save host SPRs
Nicholas Piggin [Tue, 23 Nov 2021 09:51:57 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Reduce mtmsrd instructions required to save host SPRs

This reduces the number of mtmsrd required to enable facility bits when
saving/restoring registers, by having the KVM code set all bits up front
rather than using individual facility functions that set their particular
MSR bits.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-20-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Move SPRG restore to restore_p9_host_os_sprs
Nicholas Piggin [Tue, 23 Nov 2021 09:51:56 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Move SPRG restore to restore_p9_host_os_sprs

Move the SPR update into its relevant helper function. This will
help with SPR scheduling improvements in later changes.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-19-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV: CTRL SPR does not require read-modify-write
Nicholas Piggin [Tue, 23 Nov 2021 09:51:55 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV: CTRL SPR does not require read-modify-write

Processors that support KVM HV do not require read-modify-write of
the CTRL SPR to set/clear their thread's runlatch. Just write 1 or 0
to it.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-18-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Factor out yield_count increment
Nicholas Piggin [Tue, 23 Nov 2021 09:51:54 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Factor out yield_count increment

Factor duplicated code into a helper function.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-17-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Demand fault PMU SPRs when marked not inuse
Nicholas Piggin [Tue, 23 Nov 2021 09:51:53 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Demand fault PMU SPRs when marked not inuse

The pmcregs_in_use field in the guest VPA can not be trusted to reflect
what the guest is doing with PMU SPRs, so the PMU must always be managed
(stopped) when exiting the guest, and SPR values set when entering the
guest to ensure it can't cause a covert channel or otherwise cause other
guests or the host to misbehave.

So prevent guest access to the PMU with HFSCR[PM] if pmcregs_in_use is
clear, and avoid the PMU SPR access on every partition switch. Guests
that set pmcregs_in_use incorrectly or when first setting it and using
the PMU will take a hypervisor facility unavailable interrupt that will
bring in the PMU SPRs.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Athira Jajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-16-npiggin@gmail.com
2 years agoKVM: PPC: Book3S HV P9: Factor PMU save/load into context switch functions
Nicholas Piggin [Tue, 23 Nov 2021 09:51:52 +0000 (19:51 +1000)]
KVM: PPC: Book3S HV P9: Factor PMU save/load into context switch functions

Rather than guest/host save/retsore functions, implement context switch
functions that take care of details like the VPA update for nested.

The reason to split these kind of helpers into explicit save/load
functions is mainly to schedule SPR access nicely, but PMU is a special
case where the load requires mtSPR (to stop counters) and other
difficulties, so there's less possibility to schedule those nicely. The
SPR accesses also have side-effects if the PMU is running, and in later
changes we keep the host PMU running as long as possible so this code
can be better profiled, which also complicates scheduling.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Athira Jajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211123095231.1036501-15-npiggin@gmail.com