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 bool xen_pvspin = true;
 109 static void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
 110 {
 111         int irq = __this_cpu_read(lock_kicker_irq);
 112         struct xen_lock_waiting *w = &__get_cpu_var(lock_waiting);
 113         int cpu = smp_processor_id();
 114         u64 start;
 115         unsigned long flags;
 116
 117         /* If kicker interrupts not initialized yet, just spin */
 118         if (irq == -1)
 119                 return;
 120
 121         start = spin_time_start();
 122
 123         /*
 124          * Make sure an interrupt handler can't upset things in a
 125          * partially setup state.
 126          */
 127         local_irq_save(flags);
 128         /*
 129          * We don't really care if we're overwriting some other
 130          * (lock,want) pair, as that would mean that we're currently
 131          * in an interrupt context, and the outer context had
 132          * interrupts enabled.  That has already kicked the VCPU out
 133          * of xen_poll_irq(), so it will just return spuriously and
 134          * retry with newly setup (lock,want).
 135          *
 136          * The ordering protocol on this is that the "lock" pointer
 137          * may only be set non-NULL if the "want" ticket is correct.
 138          * If we're updating "want", we must first clear "lock".
 139          */
 140         w->lock = NULL;
 141         smp_wmb();
 142         w->want = want;
 143         smp_wmb();
 144         w->lock = lock;
 145
 146         /* This uses set_bit, which atomic and therefore a barrier */
 147         cpumask_set_cpu(cpu, &waiting_cpus);
 148         add_stats(TAKEN_SLOW, 1);
 149
 150         /* clear pending */
 151         xen_clear_irq_pending(irq);
 152
 153         /* Only check lock once pending cleared */
 154         barrier();
 155
 156         /*
 157          * Mark entry to slowpath before doing the pickup test to make
 158          * sure we don't deadlock with an unlocker.
 159          */
 160         __ticket_enter_slowpath(lock);
 161
 162         /*
 163          * check again make sure it didn't become free while
 164          * we weren't looking
 165          */
 166         if (ACCESS_ONCE(lock->tickets.head) == want) {
 167                 add_stats(TAKEN_SLOW_PICKUP, 1);
 168                 goto out;
 169         }
 170
 171         /* Allow interrupts while blocked */
 172         local_irq_restore(flags);
 173
 174         /*
 175          * If an interrupt happens here, it will leave the wakeup irq
 176          * pending, which will cause xen_poll_irq() to return
 177          * immediately.
 178          */
 179
 180         /* Block until irq becomes pending (or perhaps a spurious wakeup) */
 181         xen_poll_irq(irq);
 182         add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq));
 183
 184         local_irq_save(flags);
 185
 186         kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
 187 out:
 188         cpumask_clear_cpu(cpu, &waiting_cpus);
 189         w->lock = NULL;
 190
 191         local_irq_restore(flags);
 192
 193         spin_time_accum_blocked(start);
 194 }
 195 PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning);
 196
 197 static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next)
 198 {
 199         int cpu;
 200
 201         add_stats(RELEASED_SLOW, 1);
 202
 203         for_each_cpu(cpu, &waiting_cpus) {
 204                 const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu);
 205
 206                 /* Make sure we read lock before want */
 207                 if (ACCESS_ONCE(w->lock) == lock &&
 208                     ACCESS_ONCE(w->want) == next) {
 209                         add_stats(RELEASED_SLOW_KICKED, 1);
 210                         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
 211                         break;
 212                 }
 213         }
 214 }
 215
 216 static irqreturn_t dummy_handler(int irq, void *dev_id)
 217 {
 218         BUG();
 219         return IRQ_HANDLED;
 220 }
 221
 222 void xen_init_lock_cpu(int cpu)
 223 {
 224         int irq;
 225         char *name;
 226
 227         if (!xen_pvspin)
 228                 return;
 229
 230         WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
 231              cpu, per_cpu(lock_kicker_irq, cpu));
 232
 233         name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
 234         irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
 235                                      cpu,
 236                                      dummy_handler,
 237                                      IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 238                                      name,
 239                                      NULL);
 240
 241         if (irq >= 0) {
 242                 disable_irq(irq); /* make sure it's never delivered */
 243                 per_cpu(lock_kicker_irq, cpu) = irq;
 244                 per_cpu(irq_name, cpu) = name;
 245         }
 246
 247         printk("cpu %d spinlock event irq %d\n", cpu, irq);
 248 }
 249
 250 void xen_uninit_lock_cpu(int cpu)
 251 {
 252         if (!xen_pvspin)
 253                 return;
 254
 255         unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
 256         per_cpu(lock_kicker_irq, cpu) = -1;
 257         kfree(per_cpu(irq_name, cpu));
 258         per_cpu(irq_name, cpu) = NULL;
 259 }
 260
 261
 262 /*
 263  * Our init of PV spinlocks is split in two init functions due to us
 264  * using paravirt patching and jump labels patching and having to do
 265  * all of this before SMP code is invoked.
 266  *
 267  * The paravirt patching needs to be done _before_ the alternative asm code
 268  * is started, otherwise we would not patch the core kernel code.
 269  */
 270 void __init xen_init_spinlocks(void)
 271 {
 272
 273         if (!xen_pvspin) {
 274                 printk(KERN_DEBUG "xen: PV spinlocks disabled\n");
 275                 return;
 276         }
 277         printk(KERN_DEBUG "xen: PV spinlocks enabled\n");
 278         pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
 279         pv_lock_ops.unlock_kick = xen_unlock_kick;
 280 }
 281
 282 /*
 283  * While the jump_label init code needs to happend _after_ the jump labels are
 284  * enabled and before SMP is started. Hence we use pre-SMP initcall level
 285  * init. We cannot do it in xen_init_spinlocks as that is done before
 286  * jump labels are activated.
 287  */
 288 static __init int xen_init_spinlocks_jump(void)
 289 {
 290         if (!xen_pvspin)
 291                 return 0;
 292
 293         if (!xen_domain())
 294                 return 0;
 295
 296         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 297         return 0;
 298 }
 299 early_initcall(xen_init_spinlocks_jump);
 300
 301 static __init int xen_parse_nopvspin(char *arg)
 302 {
 303         xen_pvspin = false;
 304         return 0;
 305 }
 306 early_param("xen_nopvspin", xen_parse_nopvspin);
 307
 308 #ifdef CONFIG_XEN_DEBUG_FS
 309
 310 static struct dentry *d_spin_debug;
 311
 312 static int __init xen_spinlock_debugfs(void)
 313 {
 314         struct dentry *d_xen = xen_init_debugfs();
 315
 316         if (d_xen == NULL)
 317                 return -ENOMEM;
 318
 319         if (!xen_pvspin)
 320                 return 0;
 321
 322         d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
 323
 324         debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
 325
 326         debugfs_create_u32("taken_slow", 0444, d_spin_debug,
 327                            &spinlock_stats.contention_stats[TAKEN_SLOW]);
 328         debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
 329                            &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
 330         debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
 331                            &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]);
 332
 333         debugfs_create_u32("released_slow", 0444, d_spin_debug,
 334                            &spinlock_stats.contention_stats[RELEASED_SLOW]);
 335         debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
 336                            &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
 337
 338         debugfs_create_u64("time_blocked", 0444, d_spin_debug,
 339                            &spinlock_stats.time_blocked);
 340
 341         debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
 342                                 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
 343
 344         return 0;
 345 }
 346 fs_initcall(xen_spinlock_debugfs);
 347
 348 #endif  /* CONFIG_XEN_DEBUG_FS */