arch/powerpc/include/asm/paravirt.h

   1 /* SPDX-License-Identifier: GPL-2.0-or-later */
   2 #ifndef _ASM_POWERPC_PARAVIRT_H
   3 #define _ASM_POWERPC_PARAVIRT_H
   4
   5 #include <linux/jump_label.h>
   6 #include <asm/smp.h>
   7 #ifdef CONFIG_PPC64
   8 #include <asm/paca.h>
   9 #include <asm/lppaca.h>
  10 #include <asm/hvcall.h>
  11 #endif
  12
  13 #ifdef CONFIG_PPC_SPLPAR
  14 #include <linux/smp.h>
  15 #include <asm/kvm_guest.h>
  16 #include <asm/cputhreads.h>
  17
  18 DECLARE_STATIC_KEY_FALSE(shared_processor);
  19
  20 static inline bool is_shared_processor(void)
  21 {
  22         return static_branch_unlikely(&shared_processor);
  23 }
  24
  25 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
  26 extern struct static_key paravirt_steal_enabled;
  27 extern struct static_key paravirt_steal_rq_enabled;
  28
  29 u64 pseries_paravirt_steal_clock(int cpu);
  30
  31 static inline u64 paravirt_steal_clock(int cpu)
  32 {
  33         return pseries_paravirt_steal_clock(cpu);
  34 }
  35 #endif
  36
  37 /* If bit 0 is set, the cpu has been ceded, conferred, or preempted */
  38 static inline u32 yield_count_of(int cpu)
  39 {
  40         __be32 yield_count = READ_ONCE(lppaca_of(cpu).yield_count);
  41         return be32_to_cpu(yield_count);
  42 }
  43
  44 /*
  45  * Spinlock code confers and prods, so don't trace the hcalls because the
  46  * tracing code takes spinlocks which can cause recursion deadlocks.
  47  *
  48  * These calls are made while the lock is not held: the lock slowpath yields if
  49  * it can not acquire the lock, and unlock slow path might prod if a waiter has
  50  * yielded). So this may not be a problem for simple spin locks because the
  51  * tracing does not technically recurse on the lock, but we avoid it anyway.
  52  *
  53  * However the queued spin lock contended path is more strictly ordered: the
  54  * H_CONFER hcall is made after the task has queued itself on the lock, so then
  55  * recursing on that lock will cause the task to then queue up again behind the
  56  * first instance (or worse: queued spinlocks use tricks that assume a context
  57  * never waits on more than one spinlock, so such recursion may cause random
  58  * corruption in the lock code).
  59  */
  60 static inline void yield_to_preempted(int cpu, u32 yield_count)
  61 {
  62         plpar_hcall_norets_notrace(H_CONFER, get_hard_smp_processor_id(cpu), yield_count);
  63 }
  64
  65 static inline void prod_cpu(int cpu)
  66 {
  67         plpar_hcall_norets_notrace(H_PROD, get_hard_smp_processor_id(cpu));
  68 }
  69
  70 static inline void yield_to_any(void)
  71 {
  72         plpar_hcall_norets_notrace(H_CONFER, -1, 0);
  73 }
  74
  75 static inline bool is_vcpu_idle(int vcpu)
  76 {
  77         return lppaca_of(vcpu).idle;
  78 }
  79
  80 static inline bool vcpu_is_dispatched(int vcpu)
  81 {
  82         /*
  83          * This is the yield_count.  An "odd" value (low bit on) means that
  84          * the processor is yielded (either because of an OS yield or a
  85          * hypervisor preempt).  An even value implies that the processor is
  86          * currently executing.
  87          */
  88         return (!(yield_count_of(vcpu) & 1));
  89 }
  90 #else
  91 static inline bool is_shared_processor(void)
  92 {
  93         return false;
  94 }
  95
  96 static inline u32 yield_count_of(int cpu)
  97 {
  98         return 0;
  99 }
 100
 101 extern void ___bad_yield_to_preempted(void);
 102 static inline void yield_to_preempted(int cpu, u32 yield_count)
 103 {
 104         ___bad_yield_to_preempted(); /* This would be a bug */
 105 }
 106
 107 extern void ___bad_yield_to_any(void);
 108 static inline void yield_to_any(void)
 109 {
 110         ___bad_yield_to_any(); /* This would be a bug */
 111 }
 112
 113 extern void ___bad_prod_cpu(void);
 114 static inline void prod_cpu(int cpu)
 115 {
 116         ___bad_prod_cpu(); /* This would be a bug */
 117 }
 118
 119 static inline bool is_vcpu_idle(int vcpu)
 120 {
 121         return false;
 122 }
 123 static inline bool vcpu_is_dispatched(int vcpu)
 124 {
 125         return true;
 126 }
 127 #endif
 128
 129 #define vcpu_is_preempted vcpu_is_preempted
 130 static inline bool vcpu_is_preempted(int cpu)
 131 {
 132         /*
 133          * The dispatch/yield bit alone is an imperfect indicator of
 134          * whether the hypervisor has dispatched @cpu to run on a physical
 135          * processor. When it is clear, @cpu is definitely not preempted.
 136          * But when it is set, it means only that it *might* be, subject to
 137          * other conditions. So we check other properties of the VM and
 138          * @cpu first, resorting to the yield count last.
 139          */
 140
 141         /*
 142          * Hypervisor preemption isn't possible in dedicated processor
 143          * mode by definition.
 144          */
 145         if (!is_shared_processor())
 146                 return false;
 147
 148         /*
 149          * If the hypervisor has dispatched the target CPU on a physical
 150          * processor, then the target CPU is definitely not preempted.
 151          */
 152         if (vcpu_is_dispatched(cpu))
 153                 return false;
 154
 155         /*
 156          * if the target CPU is not dispatched and the guest OS
 157          * has not marked the CPU idle, then it is hypervisor preempted.
 158          */
 159         if (!is_vcpu_idle(cpu))
 160                 return true;
 161
 162 #ifdef CONFIG_PPC_SPLPAR
 163         if (!is_kvm_guest()) {
 164                 int first_cpu, i;
 165
 166                 /*
 167                  * The result of vcpu_is_preempted() is used in a
 168                  * speculative way, and is always subject to invalidation
 169                  * by events internal and external to Linux. While we can
 170                  * be called in preemptable context (in the Linux sense),
 171                  * we're not accessing per-cpu resources in a way that can
 172                  * race destructively with Linux scheduler preemption and
 173                  * migration, and callers can tolerate the potential for
 174                  * error introduced by sampling the CPU index without
 175                  * pinning the task to it. So it is permissible to use
 176                  * raw_smp_processor_id() here to defeat the preempt debug
 177                  * warnings that can arise from using smp_processor_id()
 178                  * in arbitrary contexts.
 179                  */
 180                 first_cpu = cpu_first_thread_sibling(raw_smp_processor_id());
 181
 182                 /*
 183                  * The PowerVM hypervisor dispatches VMs on a whole core
 184                  * basis. So we know that a thread sibling of the executing CPU
 185                  * cannot have been preempted by the hypervisor, even if it
 186                  * has called H_CONFER, which will set the yield bit.
 187                  */
 188                 if (cpu_first_thread_sibling(cpu) == first_cpu)
 189                         return false;
 190
 191                 /*
 192                  * The specific target CPU was marked by guest OS as idle, but
 193                  * then also check all other cpus in the core for PowerVM
 194                  * because it does core scheduling and one of the vcpu
 195                  * of the core getting preempted by hypervisor implies
 196                  * other vcpus can also be considered preempted.
 197                  */
 198                 first_cpu = cpu_first_thread_sibling(cpu);
 199                 for (i = first_cpu; i < first_cpu + threads_per_core; i++) {
 200                         if (i == cpu)
 201                                 continue;
 202                         if (vcpu_is_dispatched(i))
 203                                 return false;
 204                         if (!is_vcpu_idle(i))
 205                                 return true;
 206                 }
 207         }
 208 #endif
 209
 210         /*
 211          * None of the threads in target CPU's core are running but none of
 212          * them were preempted too. Hence assume the target CPU to be
 213          * non-preempted.
 214          */
 215         return false;
 216 }
 217
 218 static inline bool pv_is_native_spin_unlock(void)
 219 {
 220         return !is_shared_processor();
 221 }
 222
 223 #endif /* _ASM_POWERPC_PARAVIRT_H */