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 #include <linux/atomic.h>
  12
  13 #include <asm/paravirt.h>
  14
  15 #include <xen/interface/xen.h>
  16 #include <xen/events.h>
  17
  18 #include "xen-ops.h"
  19 #include "debugfs.h"
  20
  21 static DEFINE_PER_CPU(int, lock_kicker_irq) = -1;
  22 static DEFINE_PER_CPU(char *, irq_name);
  23 static DEFINE_PER_CPU(atomic_t, xen_qlock_wait_nest);
  24 static bool xen_pvspin = true;
  25
  26 #ifdef CONFIG_QUEUED_SPINLOCKS
  27
  28 #include <asm/qspinlock.h>
  29
  30 static void xen_qlock_kick(int cpu)
  31 {
  32         int irq = per_cpu(lock_kicker_irq, cpu);
  33
  34         /* Don't kick if the target's kicker interrupt is not initialized. */
  35         if (irq == -1)
  36                 return;
  37
  38         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
  39 }
  40
  41 /*
  42  * Halt the current CPU & release it back to the host
  43  */
  44 static void xen_qlock_wait(u8 *byte, u8 val)
  45 {
  46         int irq = __this_cpu_read(lock_kicker_irq);
  47         atomic_t *nest_cnt = this_cpu_ptr(&xen_qlock_wait_nest);
  48
  49         /* If kicker interrupts not initialized yet, just spin */
  50         if (irq == -1 || in_nmi())
  51                 return;
  52
  53         /* Detect reentry. */
  54         atomic_inc(nest_cnt);
  55
  56         /* If irq pending already and no nested call clear it. */
  57         if (atomic_read(nest_cnt) == 1 && xen_test_irq_pending(irq)) {
  58                 xen_clear_irq_pending(irq);
  59         } else if (READ_ONCE(*byte) == val) {
  60                 /* Block until irq becomes pending (or a spurious wakeup) */
  61                 xen_poll_irq(irq);
  62         }
  63
  64         atomic_dec(nest_cnt);
  65 }
  66
  67 #else /* CONFIG_QUEUED_SPINLOCKS */
  68
  69 enum xen_contention_stat {
  70         TAKEN_SLOW,
  71         TAKEN_SLOW_PICKUP,
  72         TAKEN_SLOW_SPURIOUS,
  73         RELEASED_SLOW,
  74         RELEASED_SLOW_KICKED,
  75         NR_CONTENTION_STATS
  76 };
  77
  78
  79 #ifdef CONFIG_XEN_DEBUG_FS
  80 #define HISTO_BUCKETS   30
  81 static struct xen_spinlock_stats
  82 {
  83         u32 contention_stats[NR_CONTENTION_STATS];
  84         u32 histo_spin_blocked[HISTO_BUCKETS+1];
  85         u64 time_blocked;
  86 } spinlock_stats;
  87
  88 static u8 zero_stats;
  89
  90 static inline void check_zero(void)
  91 {
  92         u8 ret;
  93         u8 old = READ_ONCE(zero_stats);
  94         if (unlikely(old)) {
  95                 ret = cmpxchg(&zero_stats, old, 0);
  96                 /* This ensures only one fellow resets the stat */
  97                 if (ret == old)
  98                         memset(&spinlock_stats, 0, sizeof(spinlock_stats));
  99         }
 100 }
 101
 102 static inline void add_stats(enum xen_contention_stat var, u32 val)
 103 {
 104         check_zero();
 105         spinlock_stats.contention_stats[var] += val;
 106 }
 107
 108 static inline u64 spin_time_start(void)
 109 {
 110         return xen_clocksource_read();
 111 }
 112
 113 static void __spin_time_accum(u64 delta, u32 *array)
 114 {
 115         unsigned index = ilog2(delta);
 116
 117         check_zero();
 118
 119         if (index < HISTO_BUCKETS)
 120                 array[index]++;
 121         else
 122                 array[HISTO_BUCKETS]++;
 123 }
 124
 125 static inline void spin_time_accum_blocked(u64 start)
 126 {
 127         u32 delta = xen_clocksource_read() - start;
 128
 129         __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
 130         spinlock_stats.time_blocked += delta;
 131 }
 132 #else  /* !CONFIG_XEN_DEBUG_FS */
 133 static inline void add_stats(enum xen_contention_stat var, u32 val)
 134 {
 135 }
 136
 137 static inline u64 spin_time_start(void)
 138 {
 139         return 0;
 140 }
 141
 142 static inline void spin_time_accum_blocked(u64 start)
 143 {
 144 }
 145 #endif  /* CONFIG_XEN_DEBUG_FS */
 146
 147 struct xen_lock_waiting {
 148         struct arch_spinlock *lock;
 149         __ticket_t want;
 150 };
 151
 152 static DEFINE_PER_CPU(struct xen_lock_waiting, lock_waiting);
 153 static cpumask_t waiting_cpus;
 154
 155 __visible void xen_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
 156 {
 157         int irq = __this_cpu_read(lock_kicker_irq);
 158         struct xen_lock_waiting *w = this_cpu_ptr(&lock_waiting);
 159         int cpu = smp_processor_id();
 160         u64 start;
 161         __ticket_t head;
 162         unsigned long flags;
 163
 164         /* If kicker interrupts not initialized yet, just spin */
 165         if (irq == -1)
 166                 return;
 167
 168         start = spin_time_start();
 169
 170         /*
 171          * Make sure an interrupt handler can't upset things in a
 172          * partially setup state.
 173          */
 174         local_irq_save(flags);
 175         /*
 176          * We don't really care if we're overwriting some other
 177          * (lock,want) pair, as that would mean that we're currently
 178          * in an interrupt context, and the outer context had
 179          * interrupts enabled.  That has already kicked the VCPU out
 180          * of xen_poll_irq(), so it will just return spuriously and
 181          * retry with newly setup (lock,want).
 182          *
 183          * The ordering protocol on this is that the "lock" pointer
 184          * may only be set non-NULL if the "want" ticket is correct.
 185          * If we're updating "want", we must first clear "lock".
 186          */
 187         w->lock = NULL;
 188         smp_wmb();
 189         w->want = want;
 190         smp_wmb();
 191         w->lock = lock;
 192
 193         /* This uses set_bit, which atomic and therefore a barrier */
 194         cpumask_set_cpu(cpu, &waiting_cpus);
 195         add_stats(TAKEN_SLOW, 1);
 196
 197         /* clear pending */
 198         xen_clear_irq_pending(irq);
 199
 200         /* Only check lock once pending cleared */
 201         barrier();
 202
 203         /*
 204          * Mark entry to slowpath before doing the pickup test to make
 205          * sure we don't deadlock with an unlocker.
 206          */
 207         __ticket_enter_slowpath(lock);
 208
 209         /* make sure enter_slowpath, which is atomic does not cross the read */
 210         smp_mb__after_atomic();
 211
 212         /*
 213          * check again make sure it didn't become free while
 214          * we weren't looking
 215          */
 216         head = READ_ONCE(lock->tickets.head);
 217         if (__tickets_equal(head, want)) {
 218                 add_stats(TAKEN_SLOW_PICKUP, 1);
 219                 goto out;
 220         }
 221
 222         /* Allow interrupts while blocked */
 223         local_irq_restore(flags);
 224
 225         /*
 226          * If an interrupt happens here, it will leave the wakeup irq
 227          * pending, which will cause xen_poll_irq() to return
 228          * immediately.
 229          */
 230
 231         /* Block until irq becomes pending (or perhaps a spurious wakeup) */
 232         xen_poll_irq(irq);
 233         add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq));
 234
 235         local_irq_save(flags);
 236
 237         kstat_incr_irq_this_cpu(irq);
 238 out:
 239         cpumask_clear_cpu(cpu, &waiting_cpus);
 240         w->lock = NULL;
 241
 242         local_irq_restore(flags);
 243
 244         spin_time_accum_blocked(start);
 245 }
 246 PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning);
 247
 248 static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next)
 249 {
 250         int cpu;
 251
 252         add_stats(RELEASED_SLOW, 1);
 253
 254         for_each_cpu(cpu, &waiting_cpus) {
 255                 const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu);
 256
 257                 /* Make sure we read lock before want */
 258                 if (READ_ONCE(w->lock) == lock &&
 259                     READ_ONCE(w->want) == next) {
 260                         add_stats(RELEASED_SLOW_KICKED, 1);
 261                         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
 262                         break;
 263                 }
 264         }
 265 }
 266 #endif /* CONFIG_QUEUED_SPINLOCKS */
 267
 268 static irqreturn_t dummy_handler(int irq, void *dev_id)
 269 {
 270         BUG();
 271         return IRQ_HANDLED;
 272 }
 273
 274 void xen_init_lock_cpu(int cpu)
 275 {
 276         int irq;
 277         char *name;
 278
 279         if (!xen_pvspin)
 280                 return;
 281
 282         WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
 283              cpu, per_cpu(lock_kicker_irq, cpu));
 284
 285         name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
 286         irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
 287                                      cpu,
 288                                      dummy_handler,
 289                                      IRQF_PERCPU|IRQF_NOBALANCING,
 290                                      name,
 291                                      NULL);
 292
 293         if (irq >= 0) {
 294                 disable_irq(irq); /* make sure it's never delivered */
 295                 per_cpu(lock_kicker_irq, cpu) = irq;
 296                 per_cpu(irq_name, cpu) = name;
 297         }
 298
 299         printk("cpu %d spinlock event irq %d\n", cpu, irq);
 300 }
 301
 302 void xen_uninit_lock_cpu(int cpu)
 303 {
 304         int irq;
 305
 306         if (!xen_pvspin)
 307                 return;
 308
 309         /*
 310          * When booting the kernel with 'mitigations=auto,nosmt', the secondary
 311          * CPUs are not activated, and lock_kicker_irq is not initialized.
 312          */
 313         irq = per_cpu(lock_kicker_irq, cpu);
 314         if (irq == -1)
 315                 return;
 316
 317         unbind_from_irqhandler(irq, NULL);
 318         per_cpu(lock_kicker_irq, cpu) = -1;
 319         kfree(per_cpu(irq_name, cpu));
 320         per_cpu(irq_name, cpu) = NULL;
 321 }
 322
 323
 324 /*
 325  * Our init of PV spinlocks is split in two init functions due to us
 326  * using paravirt patching and jump labels patching and having to do
 327  * all of this before SMP code is invoked.
 328  *
 329  * The paravirt patching needs to be done _before_ the alternative asm code
 330  * is started, otherwise we would not patch the core kernel code.
 331  */
 332 void __init xen_init_spinlocks(void)
 333 {
 334
 335         if (!xen_pvspin) {
 336                 printk(KERN_DEBUG "xen: PV spinlocks disabled\n");
 337                 return;
 338         }
 339         printk(KERN_DEBUG "xen: PV spinlocks enabled\n");
 340 #ifdef CONFIG_QUEUED_SPINLOCKS
 341         __pv_init_lock_hash();
 342         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 343         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 344         pv_lock_ops.wait = xen_qlock_wait;
 345         pv_lock_ops.kick = xen_qlock_kick;
 346 #else
 347         pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
 348         pv_lock_ops.unlock_kick = xen_unlock_kick;
 349 #endif
 350 }
 351
 352 /*
 353  * While the jump_label init code needs to happend _after_ the jump labels are
 354  * enabled and before SMP is started. Hence we use pre-SMP initcall level
 355  * init. We cannot do it in xen_init_spinlocks as that is done before
 356  * jump labels are activated.
 357  */
 358 static __init int xen_init_spinlocks_jump(void)
 359 {
 360         if (!xen_pvspin)
 361                 return 0;
 362
 363         if (!xen_domain())
 364                 return 0;
 365
 366         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 367         return 0;
 368 }
 369 early_initcall(xen_init_spinlocks_jump);
 370
 371 static __init int xen_parse_nopvspin(char *arg)
 372 {
 373         xen_pvspin = false;
 374         return 0;
 375 }
 376 early_param("xen_nopvspin", xen_parse_nopvspin);
 377
 378 #if defined(CONFIG_XEN_DEBUG_FS) && !defined(CONFIG_QUEUED_SPINLOCKS)
 379
 380 static struct dentry *d_spin_debug;
 381
 382 static int __init xen_spinlock_debugfs(void)
 383 {
 384         struct dentry *d_xen = xen_init_debugfs();
 385
 386         if (d_xen == NULL)
 387                 return -ENOMEM;
 388
 389         if (!xen_pvspin)
 390                 return 0;
 391
 392         d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
 393
 394         debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
 395
 396         debugfs_create_u32("taken_slow", 0444, d_spin_debug,
 397                            &spinlock_stats.contention_stats[TAKEN_SLOW]);
 398         debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
 399                            &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
 400         debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
 401                            &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]);
 402
 403         debugfs_create_u32("released_slow", 0444, d_spin_debug,
 404                            &spinlock_stats.contention_stats[RELEASED_SLOW]);
 405         debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
 406                            &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
 407
 408         debugfs_create_u64("time_blocked", 0444, d_spin_debug,
 409                            &spinlock_stats.time_blocked);
 410
 411         debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
 412                                 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
 413
 414         return 0;
 415 }
 416 fs_initcall(xen_spinlock_debugfs);
 417
 418 #endif  /* CONFIG_XEN_DEBUG_FS */