x86/spinlocks/paravirt: Fix memory corruption on unlock
[sfrench/cifs-2.6.git] / 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 = READ_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 __visible 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 = this_cpu_ptr(&lock_waiting);
113         int cpu = smp_processor_id();
114         u64 start;
115         __ticket_t head;
116         unsigned long flags;
117
118         /* If kicker interrupts not initialized yet, just spin */
119         if (irq == -1)
120                 return;
121
122         start = spin_time_start();
123
124         /*
125          * Make sure an interrupt handler can't upset things in a
126          * partially setup state.
127          */
128         local_irq_save(flags);
129         /*
130          * We don't really care if we're overwriting some other
131          * (lock,want) pair, as that would mean that we're currently
132          * in an interrupt context, and the outer context had
133          * interrupts enabled.  That has already kicked the VCPU out
134          * of xen_poll_irq(), so it will just return spuriously and
135          * retry with newly setup (lock,want).
136          *
137          * The ordering protocol on this is that the "lock" pointer
138          * may only be set non-NULL if the "want" ticket is correct.
139          * If we're updating "want", we must first clear "lock".
140          */
141         w->lock = NULL;
142         smp_wmb();
143         w->want = want;
144         smp_wmb();
145         w->lock = lock;
146
147         /* This uses set_bit, which atomic and therefore a barrier */
148         cpumask_set_cpu(cpu, &waiting_cpus);
149         add_stats(TAKEN_SLOW, 1);
150
151         /* clear pending */
152         xen_clear_irq_pending(irq);
153
154         /* Only check lock once pending cleared */
155         barrier();
156
157         /*
158          * Mark entry to slowpath before doing the pickup test to make
159          * sure we don't deadlock with an unlocker.
160          */
161         __ticket_enter_slowpath(lock);
162
163         /* make sure enter_slowpath, which is atomic does not cross the read */
164         smp_mb__after_atomic();
165
166         /*
167          * check again make sure it didn't become free while
168          * we weren't looking
169          */
170         head = READ_ONCE(lock->tickets.head);
171         if (__tickets_equal(head, want)) {
172                 add_stats(TAKEN_SLOW_PICKUP, 1);
173                 goto out;
174         }
175
176         /* Allow interrupts while blocked */
177         local_irq_restore(flags);
178
179         /*
180          * If an interrupt happens here, it will leave the wakeup irq
181          * pending, which will cause xen_poll_irq() to return
182          * immediately.
183          */
184
185         /* Block until irq becomes pending (or perhaps a spurious wakeup) */
186         xen_poll_irq(irq);
187         add_stats(TAKEN_SLOW_SPURIOUS, !xen_test_irq_pending(irq));
188
189         local_irq_save(flags);
190
191         kstat_incr_irq_this_cpu(irq);
192 out:
193         cpumask_clear_cpu(cpu, &waiting_cpus);
194         w->lock = NULL;
195
196         local_irq_restore(flags);
197
198         spin_time_accum_blocked(start);
199 }
200 PV_CALLEE_SAVE_REGS_THUNK(xen_lock_spinning);
201
202 static void xen_unlock_kick(struct arch_spinlock *lock, __ticket_t next)
203 {
204         int cpu;
205
206         add_stats(RELEASED_SLOW, 1);
207
208         for_each_cpu(cpu, &waiting_cpus) {
209                 const struct xen_lock_waiting *w = &per_cpu(lock_waiting, cpu);
210
211                 /* Make sure we read lock before want */
212                 if (READ_ONCE(w->lock) == lock &&
213                     READ_ONCE(w->want) == next) {
214                         add_stats(RELEASED_SLOW_KICKED, 1);
215                         xen_send_IPI_one(cpu, XEN_SPIN_UNLOCK_VECTOR);
216                         break;
217                 }
218         }
219 }
220
221 static irqreturn_t dummy_handler(int irq, void *dev_id)
222 {
223         BUG();
224         return IRQ_HANDLED;
225 }
226
227 void xen_init_lock_cpu(int cpu)
228 {
229         int irq;
230         char *name;
231
232         if (!xen_pvspin)
233                 return;
234
235         WARN(per_cpu(lock_kicker_irq, cpu) >= 0, "spinlock on CPU%d exists on IRQ%d!\n",
236              cpu, per_cpu(lock_kicker_irq, cpu));
237
238         name = kasprintf(GFP_KERNEL, "spinlock%d", cpu);
239         irq = bind_ipi_to_irqhandler(XEN_SPIN_UNLOCK_VECTOR,
240                                      cpu,
241                                      dummy_handler,
242                                      IRQF_PERCPU|IRQF_NOBALANCING,
243                                      name,
244                                      NULL);
245
246         if (irq >= 0) {
247                 disable_irq(irq); /* make sure it's never delivered */
248                 per_cpu(lock_kicker_irq, cpu) = irq;
249                 per_cpu(irq_name, cpu) = name;
250         }
251
252         printk("cpu %d spinlock event irq %d\n", cpu, irq);
253 }
254
255 void xen_uninit_lock_cpu(int cpu)
256 {
257         if (!xen_pvspin)
258                 return;
259
260         unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
261         per_cpu(lock_kicker_irq, cpu) = -1;
262         kfree(per_cpu(irq_name, cpu));
263         per_cpu(irq_name, cpu) = NULL;
264 }
265
266
267 /*
268  * Our init of PV spinlocks is split in two init functions due to us
269  * using paravirt patching and jump labels patching and having to do
270  * all of this before SMP code is invoked.
271  *
272  * The paravirt patching needs to be done _before_ the alternative asm code
273  * is started, otherwise we would not patch the core kernel code.
274  */
275 void __init xen_init_spinlocks(void)
276 {
277
278         if (!xen_pvspin) {
279                 printk(KERN_DEBUG "xen: PV spinlocks disabled\n");
280                 return;
281         }
282         printk(KERN_DEBUG "xen: PV spinlocks enabled\n");
283         pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
284         pv_lock_ops.unlock_kick = xen_unlock_kick;
285 }
286
287 /*
288  * While the jump_label init code needs to happend _after_ the jump labels are
289  * enabled and before SMP is started. Hence we use pre-SMP initcall level
290  * init. We cannot do it in xen_init_spinlocks as that is done before
291  * jump labels are activated.
292  */
293 static __init int xen_init_spinlocks_jump(void)
294 {
295         if (!xen_pvspin)
296                 return 0;
297
298         if (!xen_domain())
299                 return 0;
300
301         static_key_slow_inc(&paravirt_ticketlocks_enabled);
302         return 0;
303 }
304 early_initcall(xen_init_spinlocks_jump);
305
306 static __init int xen_parse_nopvspin(char *arg)
307 {
308         xen_pvspin = false;
309         return 0;
310 }
311 early_param("xen_nopvspin", xen_parse_nopvspin);
312
313 #ifdef CONFIG_XEN_DEBUG_FS
314
315 static struct dentry *d_spin_debug;
316
317 static int __init xen_spinlock_debugfs(void)
318 {
319         struct dentry *d_xen = xen_init_debugfs();
320
321         if (d_xen == NULL)
322                 return -ENOMEM;
323
324         if (!xen_pvspin)
325                 return 0;
326
327         d_spin_debug = debugfs_create_dir("spinlocks", d_xen);
328
329         debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
330
331         debugfs_create_u32("taken_slow", 0444, d_spin_debug,
332                            &spinlock_stats.contention_stats[TAKEN_SLOW]);
333         debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
334                            &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
335         debugfs_create_u32("taken_slow_spurious", 0444, d_spin_debug,
336                            &spinlock_stats.contention_stats[TAKEN_SLOW_SPURIOUS]);
337
338         debugfs_create_u32("released_slow", 0444, d_spin_debug,
339                            &spinlock_stats.contention_stats[RELEASED_SLOW]);
340         debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
341                            &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
342
343         debugfs_create_u64("time_blocked", 0444, d_spin_debug,
344                            &spinlock_stats.time_blocked);
345
346         debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
347                                 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
348
349         return 0;
350 }
351 fs_initcall(xen_spinlock_debugfs);
352
353 #endif  /* CONFIG_XEN_DEBUG_FS */