arch/x86/xen/spinlock.c

   1 /*
   2  * Split spinlock implementation out into its own file, so it can be
   3  * compiled in a FTRACE-compatible way.
   4  */
   5 #include <linux/kernel_stat.h>
   6 #include <linux/spinlock.h>
   7 #include <linux/debugfs.h>
   8 #include <linux/log2.h>
   9 #include <linux/gfp.h>
  10 #include <linux/slab.h>
  11
  12 #include <asm/paravirt.h>
  13
  14 #include <xen/interface/xen.h>
  15 #include <xen/events.h>
  16
  17 #include "xen-ops.h"
  18 #include "debugfs.h"
  19
  20 enum xen_contention_stat {
  21         TAKEN_SLOW,
  22         TAKEN_SLOW_PICKUP,
  23         TAKEN_SLOW_SPURIOUS,
  24         RELEASED_SLOW,
  25         RELEASED_SLOW_KICKED,
  26         NR_CONTENTION_STATS
  27 };
  28
  29
  30 #ifdef CONFIG_XEN_DEBUG_FS
  31 #define HISTO_BUCKETS   30
  32 static struct xen_spinlock_stats
  33 {
  34         u32 contention_stats[NR_CONTENTION_STATS];
  35         u32 histo_spin_blocked[HISTO_BUCKETS+1];
  36         u64 time_blocked;
  37 } spinlock_stats;
  38
  39 static u8 zero_stats;
  40
  41 static inline void check_zero(void)
  42 {
  43         u8 ret;
  44         u8 old = ACCESS_ONCE(zero_stats);
  45         if (unlikely(old)) {
  46                 ret = cmpxchg(&zero_stats, old, 0);
  47                 /* This ensures only one fellow resets the stat */
  48                 if (ret == old)
  49                         memset(&spinlock_stats, 0, sizeof(spinlock_stats));
  50         }
  51 }
  52
  53 static inline void add_stats(enum xen_contention_stat var, u32 val)
  54 {
  55         check_zero();
  56         spinlock_stats.contention_stats[var] += val;
  57 }
  58
  59 static inline u64 spin_time_start(void)
  60 {
  61         return xen_clocksource_read();
  62 }
  63
  64 static void __spin_time_accum(u64 delta, u32 *array)
  65 {
  66         unsigned index = ilog2(delta);
  67
  68         check_zero();
  69
  70         if (index < HISTO_BUCKETS)
  71                 array[index]++;
  72         else
  73                 array[HISTO_BUCKETS]++;
  74 }
  75
  76 static inline void spin_time_accum_blocked(u64 start)
  77 {
  78         u32 delta = xen_clocksource_read() - start;
  79
  80         __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
  81         spinlock_stats.time_blocked += delta;
  82 }
  83 #else  /* !CONFIG_XEN_DEBUG_FS */
  84 static inline void add_stats(enum xen_contention_stat var, u32 val)
  85 {
  86 }
  87
  88 static inline u64 spin_time_start(void)
  89 {
  90         return 0;
  91 }
  92
  93 static inline void spin_time_accum_blocked(u64 start)
  94 {
  95 }
  96 #endif  /* CONFIG_XEN_DEBUG_FS */
  97
  98 struct xen_lock_waiting {
  99         struct arch_spinlock *lock;
 100         __ticket_t want;
 101 };
 102
 103 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
 104 static DEFINE_PER_CPU(char *, irq_name);
 105 static DEFINE_PER_CPU(struct xen_lock_waiting, lock_waiting);
 106 static cpumask_t waiting_cpus;
 107
 108 static void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
 109 {
 110         int irq = __this_cpu_read(lock_kicker_irq);
 111         struct xen_lock_waiting *w = &__get_cpu_var(lock_waiting);
 112         int cpu = smp_processor_id();
 113         u64 start;
 114         unsigned long flags;
 115
 116         /* If kicker interrupts not initialized yet, just spin */
 117         if (irq == -1)
 118                 return;
 119
 120         start = spin_time_start();
 121
 122         /*
 123          * Make sure an interrupt handler can't upset things in a
 124          * partially setup state.
 125          */
 126         local_irq_save(flags);
 127         /*
 128          * We don't really care if we're overwriting some other
 129          * (lock,want) pair, as that would mean that we're currently
 130          * in an interrupt context, and the outer context had
 131          * interrupts enabled.  That has already kicked the VCPU out
 132          * of xen_poll_irq(), so it will just return spuriously and
 133          * retry with newly setup (lock,want).
 134          *
 135          * The ordering protocol on this is that the "lock" pointer
 136          * may only be set non-NULL if the "want" ticket is correct.
 137          * If we're updating "want", we must first clear "lock".
 138          */
 139         w->lock = NULL;
 140         smp_wmb();
 141         w->want = want;
 142         smp_wmb();
 143         w->lock = lock;
 144
 145         /* This uses set_bit, which atomic and therefore a barrier */
 146         cpumask_set_cpu(cpu, &waiting_cpus);
 147         add_stats(TAKEN_SLOW, 1);
 148
 149         /* clear pending */
 150         xen_clear_irq_pending(irq);
 151
 152         /* Only check lock once pending cleared */
 153         barrier();
 154
 155         /*
 156          * Mark entry to slowpath before doing the pickup test to make
 157          * sure we don't deadlock with an unlocker.
 158          */
 159         __ticket_enter_slowpath(lock);
 160
 161         /*
 162          * check again make sure it didn't become free while
 163          * we weren't looking
 164          */
 165         if (ACCESS_ONCE(lock->tickets.head) == want) {
 166                 add_stats(TAKEN_SLOW_PICKUP, 1);
 167                 goto out;
 168         }
 169
 170         /* Allow interrupts while blocked */
 171         local_irq_restore(flags);
 172
 173         /*
 174          * If an interrupt happens here, it will leave the wakeup irq
 175          * pending, which will cause xen_poll_irq() to return
 176          * immediately.
 177          */
 178
 179         /* Block until irq becomes pending (or perhaps a spurious wakeup) */
 180         xen_poll_irq(irq);
 181         add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq));
 182
 183         local_irq_save(flags);
 184
 185         kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
 186 out:
 187         cpumask_clear_cpu(cpu, &waiting_cpus);
 188         w->lock = NULL;
 189
 190         local_irq_restore(flags);
 191
 192         spin_time_accum_blocked(start);
 193 }
 194 PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning);
 195
 196 static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next)
 197 {
 198         int cpu;
 199
 200         add_stats(RELEASED_SLOW, 1);
 201
 202         for_each_cpu(cpu, &waiting_cpus) {
 203                 const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu);
 204
 205                 /* Make sure we read lock before want */
 206                 if (ACCESS_ONCE(w->lock) == lock &&
 207                     ACCESS_ONCE(w->want) == next) {
 208                         add_stats(RELEASED_SLOW_KICKED, 1);
 209                         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
 210                         break;
 211                 }
 212         }
 213 }
 214
 215 static irqreturn_t dummy_handler(int irq, void *dev_id)
 216 {
 217         BUG();
 218         return IRQ_HANDLED;
 219 }
 220
 221 void xen_init_lock_cpu(int cpu)
 222 {
 223         int irq;
 224         char *name;
 225
 226         WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
 227              cpu, per_cpu(lock_kicker_irq, cpu));
 228
 229         /*
 230          * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
 231          * (xen: disable PV spinlocks on HVM)
 232          */
 233         if (xen_hvm_domain())
 234                 return;
 235
 236         name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
 237         irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
 238                                      cpu,
 239                                      dummy_handler,
 240                                      IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 241                                      name,
 242                                      NULL);
 243
 244         if (irq >= 0) {
 245                 disable_irq(irq); /* make sure it's never delivered */
 246                 per_cpu(lock_kicker_irq, cpu) = irq;
 247                 per_cpu(irq_name, cpu) = name;
 248         }
 249
 250         printk("cpu %d spinlock event irq %d\n", cpu, irq);
 251 }
 252
 253 void xen_uninit_lock_cpu(int cpu)
 254 {
 255         /*
 256          * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
 257          * (xen: disable PV spinlocks on HVM)
 258          */
 259         if (xen_hvm_domain())
 260                 return;
 261
 262         unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
 263         per_cpu(lock_kicker_irq, cpu) = -1;
 264         kfree(per_cpu(irq_name, cpu));
 265         per_cpu(irq_name, cpu) = NULL;
 266 }
 267
 268 static bool xen_pvspin __initdata = true;
 269
 270 void __init xen_init_spinlocks(void)
 271 {
 272         /*
 273          * See git commit f10cd522c5fbfec9ae3cc01967868c9c2401ed23
 274          * (xen: disable PV spinlocks on HVM)
 275          */
 276         if (xen_hvm_domain())
 277                 return;
 278
 279         if (!xen_pvspin) {
 280                 printk(KERN_DEBUG "xen: PV spinlocks disabled\n");
 281                 return;
 282         }
 283
 284         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 285
 286         pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
 287         pv_lock_ops.unlock_kick = xen_unlock_kick;
 288 }
 289
 290 static __init int xen_parse_nopvspin(char *arg)
 291 {
 292         xen_pvspin = false;
 293         return 0;
 294 }
 295 early_param("xen_nopvspin", xen_parse_nopvspin);
 296
 297 #ifdef CONFIG_XEN_DEBUG_FS
 298
 299 static struct dentry *d_spin_debug;
 300
 301 static int __init xen_spinlock_debugfs(void)
 302 {
 303         struct dentry *d_xen = xen_init_debugfs();
 304
 305         if (d_xen == NULL)
 306                 return -ENOMEM;
 307
 308         d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
 309
 310         debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
 311
 312         debugfs_create_u32("taken_slow", 0444, d_spin_debug,
 313                            &spinlock_stats.contention_stats[TAKEN_SLOW]);
 314         debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
 315                            &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
 316         debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
 317                            &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]);
 318
 319         debugfs_create_u32("released_slow", 0444, d_spin_debug,
 320                            &spinlock_stats.contention_stats[RELEASED_SLOW]);
 321         debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
 322                            &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
 323
 324         debugfs_create_u64("time_blocked", 0444, d_spin_debug,
 325                            &spinlock_stats.time_blocked);
 326
 327         debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
 328                                 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
 329
 330         return 0;
 331 }
 332 fs_initcall(xen_spinlock_debugfs);
 333
 334 #endif  /* CONFIG_XEN_DEBUG_FS */