File book3s_hv_p9_entry.c in powerpc/kvm folder consists of functions
like freeze_pmu, switch_pmu_to_guest and switch_pmu_to_host which are
specific to Performance Monitoring Unit(PMU) for power9 and later
platforms.
For better maintenance, moving pmu related code from
book3s_hv_p9_entry.c to a new file called book3s_hv_p9_perf.c,
without any logic change.
Also make corresponding changes in the Makefile to include
book3s_hv_p9_perf.c during compilation.
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220711034927.213192-1-kjain@linux.ibm.com
Alter the data collection points for the debug timing code in the P9
path to be more in line with what the code does. The points where we
accumulate time are now the following:
vcpu_entry: From vcpu_run_hv entry until the start of the inner loop;
guest_entry: From the start of the inner loop until the guest entry in
asm;
in_guest: From the guest entry in asm until the return to KVM C code;
guest_exit: From the return into KVM C code until the corresponding
hypercall/page fault handling or re-entry into the guest;
hypercall: Time spent handling hcalls in the kernel (hcalls can go to
QEMU, not accounted here);
page_fault: Time spent handling page faults;
vcpu_exit: vcpu_run_hv exit (almost no code here currently).
Like before, these are exposed in debugfs in a file called
"timings". There are four values:
- number of occurrences of the accumulation point;
- total time the vcpu spent in the phase in ns;
- shortest time the vcpu spent in the phase in ns;
- longest time the vcpu spent in the phase in ns;
===
Before:
rm_entry: 53132 16793518 256 4060
rm_intr: 53132 2125914 22 340
rm_exit: 53132 24108344 374 2180
guest: 53132 40980507996 404 9997650
cede: 0 0 0 0
After:
vcpu_entry: 34637 7716108 178 4416
guest_entry: 52414 49365608 324 747542
in_guest: 52411 40828715840 258 9997480
guest_exit: 52410 19681717182 826 102496674
vcpu_exit: 34636 1744462 38 182
hypercall: 45712 22878288 38 1307962
page_fault: 992 111104034 568 168688
With just one instruction (hcall):
vcpu_entry: 1 942 942 942
guest_entry: 1 4044 4044 4044
in_guest: 1 1540 1540 1540
guest_exit: 1 3542 3542 3542
vcpu_exit: 1 80 80 80
hypercall: 0 0 0 0
page_fault: 0 0 0 0
===
Signed-off-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220525130554.2614394-6-farosas@linux.ibm.com
We are currently doing the timing for debug purposes of the P9 entry
path using the accumulators and terminology defined by the old entry
path for P8 machines.
Not only the "real-mode" and "napping" mentions are out of place for
the P9 Radix entry path but also we cannot change them because the
timing code is coupled to the structures defined in struct
kvm_vcpu_arch.
Add a new CONFIG_KVM_BOOK3S_HV_P9_TIMING to enable the timing code for
the P9 entry path. For now, just add the new CONFIG and duplicate the
structures. A subsequent patch will add the P9 changes.
Signed-off-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220525130554.2614394-4-farosas@linux.ibm.com
At debugfs/kvm/<pid>/vcpu0/timings we show how long each part of the
code takes to run:
$ cat /sys/kernel/debug/kvm/*-*/vcpu0/timings
rm_entry: 123785 49398892 118 4898
rm_intr: 123780 6075890 22 390
rm_exit: 0 0 0 0 <-- NOK
guest: 123780 46732919988 402 9997638
cede: 0 0 0 0 <-- OK, no cede napping in P9
The "rm_exit" is always showing zero because it is the last one and
end_timing does not increment the counter of the previous entry.
We can fix it by calling accumulate_time again instead of
end_timing. That way the counter gets incremented. The rest of the
arithmetic can be ignored because there are no timing points after
this and the accumulators are reset before the next round.
Signed-off-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220525130554.2614394-2-farosas@linux.ibm.com
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
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
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
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
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
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
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
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
The conversion to C introduced several bugs in TM handling that can
cause host crashes with TM bad thing interrupts. Mostly just simple
typos or missed logic in the conversion that got through due to my
not testing TM in the guest sufficiently.
- Early TM emulation for the softpatch interrupt should be done if fake
suspend mode is _not_ active.
- Early TM emulation wants to return immediately to the guest so as to
not doom transactions unnecessarily.
- And if exiting from the guest, the host MSR should include the TM[S]
bit if the guest was T/S, before it is treclaimed.
After this fix, all the TM selftests pass when running on a P9 processor
that implements TM with softpatch interrupt.
Fixes: 89d35b2391 ("KVM: PPC: Book3S HV P9: Implement the rest of the P9 path in C")
Reported-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210712013650.376325-1-npiggin@gmail.com
Implement support for hash guests under hash host. This has to save and
restore the host SLB, and ensure that the MMU is off while switching
into the guest SLB.
POWER9 and later CPUs now always go via the P9 path. The "fast" guest
mode is now renamed to the P9 mode, which is consistent with its
functionality and the rest of the naming.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210528090752.3542186-32-npiggin@gmail.com
Implement hash guest support. Guest entry/exit has to restore and
save/clear the SLB, plus several other bits to accommodate hash guests
in the P9 path. Radix host, hash guest support is removed from the P7/8
path.
The HPT hcalls and faults are not handled in real mode, which is a
performance regression. A worst-case fork/exit microbenchmark takes 3x
longer after this patch. kbuild benchmark performance is in the noise,
but the slowdown is likely to be noticed somewhere.
For now, accept this penalty for the benefit of simplifying the P7/8
paths and unifying P9 hash with the new code, because hash is a less
important configuration than radix on processors that support it. Hash
will benefit from future optimisations to this path, including possibly
a faster path to handle such hcalls and interrupts without doing a full
exit.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210528090752.3542186-31-npiggin@gmail.com
Rather than partition the guest PID space + flush a rogue guest PID to
work around this problem, instead fix it by always disabling the MMU when
switching in or out of guest MMU context in HV mode.
This may be a bit less efficient, but it is a lot less complicated and
allows the P9 path to trivally implement the workaround too. Newer CPUs
are not subject to this issue.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210528090752.3542186-22-npiggin@gmail.com
Move MMU context switch as late as reasonably possible to minimise code
running with guest context switched in. This becomes more important when
this code may run in real-mode, with later changes.
Move WARN_ON as early as possible so program check interrupts are less
likely to tangle everything up.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210528090752.3542186-21-npiggin@gmail.com
Almost all logic is moved to C, by introducing a new in_guest mode for
the P9 path that branches very early in the KVM interrupt handler to P9
exit code.
The main P9 entry and exit assembly is now only about 160 lines of low
level stack setup and register save/restore, plus a bad-interrupt
handler.
There are two motivations for this, the first is just make the code more
maintainable being in C. The second is to reduce the amount of code
running in a special KVM mode, "realmode". In quotes because with radix
it is no longer necessarily real-mode in the MMU, but it still has to be
treated specially because it may be in real-mode, and has various
important registers like PID, DEC, TB, etc set to guest. This is hostile
to the rest of Linux and can't use arbitrary kernel functionality or be
instrumented well.
This initial patch is a reasonably faithful conversion of the asm code,
but it does lack any loop to return quickly back into the guest without
switching out of realmode in the case of unimportant or easily handled
interrupts. As explained in previous changes, handling HV interrupts
very quickly in this low level realmode is not so important for P9
performance, and are important to avoid for security, observability,
debugability reasons.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210528090752.3542186-15-npiggin@gmail.com
