powerpc/powernv: Return to cpu offline loop when finished in KVM guest
authorPaul Mackerras <paulus@samba.org>
Wed, 3 Dec 2014 03:48:40 +0000 (14:48 +1100)
committerMichael Ellerman <mpe@ellerman.id.au>
Mon, 8 Dec 2014 02:16:31 +0000 (13:16 +1100)
When a secondary hardware thread has finished running a KVM guest, we
currently put that thread into nap mode using a nap instruction in
the KVM code.  This changes the code so that instead of doing a nap
instruction directly, we instead cause the call to power7_nap() that
put the thread into nap mode to return.  The reason for doing this is
to avoid having the KVM code having to know what low-power mode to
put the thread into.

In the case of a secondary thread used to run a KVM guest, the thread
will be offline from the point of view of the host kernel, and the
relevant power7_nap() call is the one in pnv_smp_cpu_disable().
In this case we don't want to clear pending IPIs in the offline loop
in that function, since that might cause us to miss the wakeup for
the next time the thread needs to run a guest.  To tell whether or
not to clear the interrupt, we use the SRR1 value returned from
power7_nap(), and check if it indicates an external interrupt.  We
arrange that the return from power7_nap() when we have finished running
a guest returns 0, so pending interrupts don't get flushed in that
case.

Note that it is important a secondary thread that has finished
executing in the guest, or that didn't have a guest to run, should
not return to power7_nap's caller while the kvm_hstate.hwthread_req
flag in the PACA is non-zero, because the return from power7_nap
will reenable the MMU, and the MMU might still be in guest context.
In this situation we spin at low priority in real mode waiting for
hwthread_req to become zero.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
arch/powerpc/include/asm/processor.h
arch/powerpc/kernel/exceptions-64s.S
arch/powerpc/kernel/idle_power7.S
arch/powerpc/kvm/book3s_hv_rmhandlers.S
arch/powerpc/platforms/powernv/smp.c

index dda7ac4c80bdd807c76cb9af3e6e684900b61f60..29c3798cf8000fad0c5124c5aa09577aa28d8f71 100644 (file)
@@ -451,7 +451,7 @@ extern unsigned long cpuidle_disable;
 enum idle_boot_override {IDLE_NO_OVERRIDE = 0, IDLE_POWERSAVE_OFF};
 
 extern int powersave_nap;      /* set if nap mode can be used in idle loop */
-extern void power7_nap(int check_irq);
+extern unsigned long power7_nap(int check_irq);
 extern void power7_sleep(void);
 extern void flush_instruction_cache(void);
 extern void hard_reset_now(void);
index 6213f494f40bea87736994d53c981c1f563853f8..db08382e19f1ab89592dbebe4a97af1d9d8857bb 100644 (file)
@@ -131,6 +131,8 @@ BEGIN_FTR_SECTION
 1:
 #endif
 
+       /* Return SRR1 from power7_nap() */
+       mfspr   r3,SPRN_SRR1
        beq     cr1,2f
        b       power7_wakeup_noloss
 2:     b       power7_wakeup_loss
index c0754bbf81181c93c36cd8e1b12776b5ab729463..18c0687e5ab3d2b9c4dc13587af411c755eef776 100644 (file)
@@ -212,6 +212,10 @@ END_FTR_SECTION_IFSET(CPU_FTR_HVMODE)
        mtspr   SPRN_SRR0,r5
        rfid
 
+/*
+ * R3 here contains the value that will be returned to the caller
+ * of power7_nap.
+ */
 _GLOBAL(power7_wakeup_loss)
        ld      r1,PACAR1(r13)
 BEGIN_FTR_SECTION
@@ -219,15 +223,19 @@ BEGIN_FTR_SECTION
 END_FTR_SECTION_IFSET(CPU_FTR_HVMODE)
        REST_NVGPRS(r1)
        REST_GPR(2, r1)
-       ld      r3,_CCR(r1)
+       ld      r6,_CCR(r1)
        ld      r4,_MSR(r1)
        ld      r5,_NIP(r1)
        addi    r1,r1,INT_FRAME_SIZE
-       mtcr    r3
+       mtcr    r6
        mtspr   SPRN_SRR1,r4
        mtspr   SPRN_SRR0,r5
        rfid
 
+/*
+ * R3 here contains the value that will be returned to the caller
+ * of power7_nap.
+ */
 _GLOBAL(power7_wakeup_noloss)
        lbz     r0,PACA_NAPSTATELOST(r13)
        cmpwi   r0,0
index edb2ccdbb2ba068e6ce4e3df0ebfe80243ed34c4..65c105b17a2558a2ee41a1479475431d899c805f 100644 (file)
@@ -201,8 +201,6 @@ kvmppc_primary_no_guest:
        bge     kvm_novcpu_exit /* another thread already exiting */
        li      r3, NAPPING_NOVCPU
        stb     r3, HSTATE_NAPPING(r13)
-       li      r3, 1
-       stb     r3, HSTATE_HWTHREAD_REQ(r13)
 
        b       kvm_do_nap
 
@@ -293,6 +291,8 @@ kvm_start_guest:
        /* if we have no vcpu to run, go back to sleep */
        beq     kvm_no_guest
 
+kvm_secondary_got_guest:
+
        /* Set HSTATE_DSCR(r13) to something sensible */
        ld      r6, PACA_DSCR(r13)
        std     r6, HSTATE_DSCR(r13)
@@ -318,27 +318,46 @@ kvm_start_guest:
        stwcx.  r3, 0, r4
        bne     51b
 
+/*
+ * At this point we have finished executing in the guest.
+ * We need to wait for hwthread_req to become zero, since
+ * we may not turn on the MMU while hwthread_req is non-zero.
+ * While waiting we also need to check if we get given a vcpu to run.
+ */
 kvm_no_guest:
-       li      r0, KVM_HWTHREAD_IN_NAP
+       lbz     r3, HSTATE_HWTHREAD_REQ(r13)
+       cmpwi   r3, 0
+       bne     53f
+       HMT_MEDIUM
+       li      r0, KVM_HWTHREAD_IN_KERNEL
        stb     r0, HSTATE_HWTHREAD_STATE(r13)
-kvm_do_nap:
-       /* Clear the runlatch bit before napping */
-       mfspr   r2, SPRN_CTRLF
-       clrrdi  r2, r2, 1
-       mtspr   SPRN_CTRLT, r2
-
+       /* need to recheck hwthread_req after a barrier, to avoid race */
+       sync
+       lbz     r3, HSTATE_HWTHREAD_REQ(r13)
+       cmpwi   r3, 0
+       bne     54f
+/*
+ * We jump to power7_wakeup_loss, which will return to the caller
+ * of power7_nap in the powernv cpu offline loop.  The value we
+ * put in r3 becomes the return value for power7_nap.
+ */
        li      r3, LPCR_PECE0
        mfspr   r4, SPRN_LPCR
        rlwimi  r4, r3, 0, LPCR_PECE0 | LPCR_PECE1
        mtspr   SPRN_LPCR, r4
-       isync
-       std     r0, HSTATE_SCRATCH0(r13)
-       ptesync
-       ld      r0, HSTATE_SCRATCH0(r13)
-1:     cmpd    r0, r0
-       bne     1b
-       nap
-       b       .
+       li      r3, 0
+       b       power7_wakeup_loss
+
+53:    HMT_LOW
+       ld      r4, HSTATE_KVM_VCPU(r13)
+       cmpdi   r4, 0
+       beq     kvm_no_guest
+       HMT_MEDIUM
+       b       kvm_secondary_got_guest
+
+54:    li      r0, KVM_HWTHREAD_IN_KVM
+       stb     r0, HSTATE_HWTHREAD_STATE(r13)
+       b       kvm_no_guest
 
 /******************************************************************************
  *                                                                            *
@@ -2172,6 +2191,7 @@ END_FTR_SECTION_IFCLR(CPU_FTR_ARCH_206)
         * occurs, with PECE1, PECE0 and PECEDP set in LPCR. Also clear the
         * runlatch bit before napping.
         */
+kvm_do_nap:
        mfspr   r2, SPRN_CTRLF
        clrrdi  r2, r2, 1
        mtspr   SPRN_CTRLT, r2
index 4753958cd509bdfeec1be8bf088eb69e32077569..b716f666e48ae96c974dd01d0bb5eb40c0197b7b 100644 (file)
@@ -149,6 +149,7 @@ static int pnv_smp_cpu_disable(void)
 static void pnv_smp_cpu_kill_self(void)
 {
        unsigned int cpu;
+       unsigned long srr1;
 
        /* Standard hot unplug procedure */
        local_irq_disable();
@@ -165,13 +166,25 @@ static void pnv_smp_cpu_kill_self(void)
        mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) & ~(u64)LPCR_PECE1);
        while (!generic_check_cpu_restart(cpu)) {
                ppc64_runlatch_off();
-               power7_nap(1);
+               srr1 = power7_nap(1);
                ppc64_runlatch_on();
 
-               /* Clear the IPI that woke us up */
-               icp_native_flush_interrupt();
-               local_paca->irq_happened &= PACA_IRQ_HARD_DIS;
-               mb();
+               /*
+                * If the SRR1 value indicates that we woke up due to
+                * an external interrupt, then clear the interrupt.
+                * We clear the interrupt before checking for the
+                * reason, so as to avoid a race where we wake up for
+                * some other reason, find nothing and clear the interrupt
+                * just as some other cpu is sending us an interrupt.
+                * If we returned from power7_nap as a result of
+                * having finished executing in a KVM guest, then srr1
+                * contains 0.
+                */
+               if ((srr1 & SRR1_WAKEMASK) == SRR1_WAKEEE) {
+                       icp_native_flush_interrupt();
+                       local_paca->irq_happened &= PACA_IRQ_HARD_DIS;
+                       smp_mb();
+               }
 
                if (cpu_core_split_required())
                        continue;