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 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
  21 static DEFINE_PER_CPU(char *, irq_name);
  22 static bool xen_pvspin = true;
  23
  24 #ifdef CONFIG_QUEUED_SPINLOCKS
  25
  26 #include <asm/qspinlock.h>
  27
  28 static void xen_qlock_kick(int cpu)
  29 {
  30         int irq = per_cpu(lock_kicker_irq, cpu);
  31
  32         /* Don't kick if the target's kicker interrupt is not initialized. */
  33         if (irq == -1)
  34                 return;
  35
  36         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
  37 }
  38
  39 /*
  40  * Halt the current CPU & release it back to the host
  41  */
  42 static void xen_qlock_wait(u8 *byte, u8 val)
  43 {
  44         int irq = __this_cpu_read(lock_kicker_irq);
  45
  46         /* If kicker interrupts not initialized yet, just spin */
  47         if (irq == -1)
  48                 return;
  49
  50         /* clear pending */
  51         xen_clear_irq_pending(irq);
  52         barrier();
  53
  54         /*
  55          * We check the byte value after clearing pending IRQ to make sure
  56          * that we won't miss a wakeup event because of the clearing.
  57          *
  58          * The sync_clear_bit() call in xen_clear_irq_pending() is atomic.
  59          * So it is effectively a memory barrier for x86.
  60          */
  61         if (READ_ONCE(*byte) != val)
  62                 return;
  63
  64         /*
  65          * If an interrupt happens here, it will leave the wakeup irq
  66          * pending, which will cause xen_poll_irq() to return
  67          * immediately.
  68          */
  69
  70         /* Block until irq becomes pending (or perhaps a spurious wakeup) */
  71         xen_poll_irq(irq);
  72 }
  73
  74 #else /* CONFIG_QUEUED_SPINLOCKS */
  75
  76 enum xen_contention_stat {
  77         TAKEN_SLOW,
  78         TAKEN_SLOW_PICKUP,
  79         TAKEN_SLOW_SPURIOUS,
  80         RELEASED_SLOW,
  81         RELEASED_SLOW_KICKED,
  82         NR_CONTENTION_STATS
  83 };
  84
  85
  86 #ifdef CONFIG_XEN_DEBUG_FS
  87 #define HISTO_BUCKETS   30
  88 static struct xen_spinlock_stats
  89 {
  90         u32 contention_stats[NR_CONTENTION_STATS];
  91         u32 histo_spin_blocked[HISTO_BUCKETS+1];
  92         u64 time_blocked;
  93 } spinlock_stats;
  94
  95 static u8 zero_stats;
  96
  97 static inline void check_zero(void)
  98 {
  99         u8 ret;
 100         u8 old = READ_ONCE(zero_stats);
 101         if (unlikely(old)) {
 102                 ret = cmpxchg(&zero_stats, old, 0);
 103                 /* This ensures only one fellow resets the stat */
 104                 if (ret == old)
 105                         memset(&spinlock_stats, 0, sizeof(spinlock_stats));
 106         }
 107 }
 108
 109 static inline void add_stats(enum xen_contention_stat var, u32 val)
 110 {
 111         check_zero();
 112         spinlock_stats.contention_stats[var] += val;
 113 }
 114
 115 static inline u64 spin_time_start(void)
 116 {
 117         return xen_clocksource_read();
 118 }
 119
 120 static void __spin_time_accum(u64 delta, u32 *array)
 121 {
 122         unsigned index = ilog2(delta);
 123
 124         check_zero();
 125
 126         if (index < HISTO_BUCKETS)
 127                 array[index]++;
 128         else
 129                 array[HISTO_BUCKETS]++;
 130 }
 131
 132 static inline void spin_time_accum_blocked(u64 start)
 133 {
 134         u32 delta = xen_clocksource_read() - start;
 135
 136         __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
 137         spinlock_stats.time_blocked += delta;
 138 }
 139 #else  /* !CONFIG_XEN_DEBUG_FS */
 140 static inline void add_stats(enum xen_contention_stat var, u32 val)
 141 {
 142 }
 143
 144 static inline u64 spin_time_start(void)
 145 {
 146         return 0;
 147 }
 148
 149 static inline void spin_time_accum_blocked(u64 start)
 150 {
 151 }
 152 #endif  /* CONFIG_XEN_DEBUG_FS */
 153
 154 struct xen_lock_waiting {
 155         struct arch_spinlock *lock;
 156         __ticket_t want;
 157 };
 158
 159 static DEFINE_PER_CPU(struct xen_lock_waiting, lock_waiting);
 160 static cpumask_t waiting_cpus;
 161
 162 __visible void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
 163 {
 164         int irq = __this_cpu_read(lock_kicker_irq);
 165         struct xen_lock_waiting *w = this_cpu_ptr(&lock_waiting);
 166         int cpu = smp_processor_id();
 167         u64 start;
 168         __ticket_t head;
 169         unsigned long flags;
 170
 171         /* If kicker interrupts not initialized yet, just spin */
 172         if (irq == -1)
 173                 return;
 174
 175         start = spin_time_start();
 176
 177         /*
 178          * Make sure an interrupt handler can't upset things in a
 179          * partially setup state.
 180          */
 181         local_irq_save(flags);
 182         /*
 183          * We don't really care if we're overwriting some other
 184          * (lock,want) pair, as that would mean that we're currently
 185          * in an interrupt context, and the outer context had
 186          * interrupts enabled.  That has already kicked the VCPU out
 187          * of xen_poll_irq(), so it will just return spuriously and
 188          * retry with newly setup (lock,want).
 189          *
 190          * The ordering protocol on this is that the "lock" pointer
 191          * may only be set non-NULL if the "want" ticket is correct.
 192          * If we're updating "want", we must first clear "lock".
 193          */
 194         w->lock = NULL;
 195         smp_wmb();
 196         w->want = want;
 197         smp_wmb();
 198         w->lock = lock;
 199
 200         /* This uses set_bit, which atomic and therefore a barrier */
 201         cpumask_set_cpu(cpu, &waiting_cpus);
 202         add_stats(TAKEN_SLOW, 1);
 203
 204         /* clear pending */
 205         xen_clear_irq_pending(irq);
 206
 207         /* Only check lock once pending cleared */
 208         barrier();
 209
 210         /*
 211          * Mark entry to slowpath before doing the pickup test to make
 212          * sure we don't deadlock with an unlocker.
 213          */
 214         __ticket_enter_slowpath(lock);
 215
 216         /* make sure enter_slowpath, which is atomic does not cross the read */
 217         smp_mb__after_atomic();
 218
 219         /*
 220          * check again make sure it didn't become free while
 221          * we weren't looking
 222          */
 223         head = READ_ONCE(lock->tickets.head);
 224         if (__tickets_equal(head, want)) {
 225                 add_stats(TAKEN_SLOW_PICKUP, 1);
 226                 goto out;
 227         }
 228
 229         /* Allow interrupts while blocked */
 230         local_irq_restore(flags);
 231
 232         /*
 233          * If an interrupt happens here, it will leave the wakeup irq
 234          * pending, which will cause xen_poll_irq() to return
 235          * immediately.
 236          */
 237
 238         /* Block until irq becomes pending (or perhaps a spurious wakeup) */
 239         xen_poll_irq(irq);
 240         add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq));
 241
 242         local_irq_save(flags);
 243
 244         kstat_incr_irq_this_cpu(irq);
 245 out:
 246         cpumask_clear_cpu(cpu, &waiting_cpus);
 247         w->lock = NULL;
 248
 249         local_irq_restore(flags);
 250
 251         spin_time_accum_blocked(start);
 252 }
 253 PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning);
 254
 255 static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next)
 256 {
 257         int cpu;
 258
 259         add_stats(RELEASED_SLOW, 1);
 260
 261         for_each_cpu(cpu, &waiting_cpus) {
 262                 const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu);
 263
 264                 /* Make sure we read lock before want */
 265                 if (READ_ONCE(w->lock) == lock &&
 266                     READ_ONCE(w->want) == next) {
 267                         add_stats(RELEASED_SLOW_KICKED, 1);
 268                         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
 269                         break;
 270                 }
 271         }
 272 }
 273 #endif /* CONFIG_QUEUED_SPINLOCKS */
 274
 275 static irqreturn_t dummy_handler(int irq, void *dev_id)
 276 {
 277         BUG();
 278         return IRQ_HANDLED;
 279 }
 280
 281 void xen_init_lock_cpu(int cpu)
 282 {
 283         int irq;
 284         char *name;
 285
 286         if (!xen_pvspin)
 287                 return;
 288
 289         WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
 290              cpu, per_cpu(lock_kicker_irq, cpu));
 291
 292         name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
 293         irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
 294                                      cpu,
 295                                      dummy_handler,
 296                                      IRQF_PERCPU|IRQF_NOBALANCING,
 297                                      name,
 298                                      NULL);
 299
 300         if (irq >= 0) {
 301                 disable_irq(irq); /* make sure it's never delivered */
 302                 per_cpu(lock_kicker_irq, cpu) = irq;
 303                 per_cpu(irq_name, cpu) = name;
 304         }
 305
 306         printk("cpu %d spinlock event irq %d\n", cpu, irq);
 307 }
 308
 309 void xen_uninit_lock_cpu(int cpu)
 310 {
 311         if (!xen_pvspin)
 312                 return;
 313
 314         unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
 315         per_cpu(lock_kicker_irq, cpu) = -1;
 316         kfree(per_cpu(irq_name, cpu));
 317         per_cpu(irq_name, cpu) = NULL;
 318 }
 319
 320
 321 /*
 322  * Our init of PV spinlocks is split in two init functions due to us
 323  * using paravirt patching and jump labels patching and having to do
 324  * all of this before SMP code is invoked.
 325  *
 326  * The paravirt patching needs to be done _before_ the alternative asm code
 327  * is started, otherwise we would not patch the core kernel code.
 328  */
 329 void __init xen_init_spinlocks(void)
 330 {
 331
 332         if (!xen_pvspin) {
 333                 printk(KERN_DEBUG "xen: PV spinlocks disabled\n");
 334                 return;
 335         }
 336         printk(KERN_DEBUG "xen: PV spinlocks enabled\n");
 337 #ifdef CONFIG_QUEUED_SPINLOCKS
 338         __pv_init_lock_hash();
 339         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 340         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 341         pv_lock_ops.wait = xen_qlock_wait;
 342         pv_lock_ops.kick = xen_qlock_kick;
 343 #else
 344         pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
 345         pv_lock_ops.unlock_kick = xen_unlock_kick;
 346 #endif
 347 }
 348
 349 /*
 350  * While the jump_label init code needs to happend _after_ the jump labels are
 351  * enabled and before SMP is started. Hence we use pre-SMP initcall level
 352  * init. We cannot do it in xen_init_spinlocks as that is done before
 353  * jump labels are activated.
 354  */
 355 static __init int xen_init_spinlocks_jump(void)
 356 {
 357         if (!xen_pvspin)
 358                 return 0;
 359
 360         if (!xen_domain())
 361                 return 0;
 362
 363         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 364         return 0;
 365 }
 366 early_initcall(xen_init_spinlocks_jump);
 367
 368 static __init int xen_parse_nopvspin(char *arg)
 369 {
 370         xen_pvspin = false;
 371         return 0;
 372 }
 373 early_param("xen_nopvspin", xen_parse_nopvspin);
 374
 375 #if defined(CONFIG_XEN_DEBUG_FS) && !defined(CONFIG_QUEUED_SPINLOCKS)
 376
 377 static struct dentry *d_spin_debug;
 378
 379 static int __init xen_spinlock_debugfs(void)
 380 {
 381         struct dentry *d_xen = xen_init_debugfs();
 382
 383         if (d_xen == NULL)
 384                 return -ENOMEM;
 385
 386         if (!xen_pvspin)
 387                 return 0;
 388
 389         d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
 390
 391         debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
 392
 393         debugfs_create_u32("taken_slow", 0444, d_spin_debug,
 394                            &spinlock_stats.contention_stats[TAKEN_SLOW]);
 395         debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
 396                            &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
 397         debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
 398                            &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]);
 399
 400         debugfs_create_u32("released_slow", 0444, d_spin_debug,
 401                            &spinlock_stats.contention_stats[RELEASED_SLOW]);
 402         debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
 403                            &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
 404
 405         debugfs_create_u64("time_blocked", 0444, d_spin_debug,
 406                            &spinlock_stats.time_blocked);
 407
 408         debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
 409                                 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
 410
 411         return 0;
 412 }
 413 fs_initcall(xen_spinlock_debugfs);
 414
 415 #endif  /* CONFIG_XEN_DEBUG_FS */