Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 1991, 1992 Linus Torvalds |
| 3 | * Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs |
Don Zickus | 9c48f1c | 2011-09-30 15:06:21 -0400 | [diff] [blame] | 4 | * Copyright (C) 2011 Don Zickus Red Hat, Inc. |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 5 | * |
| 6 | * Pentium III FXSR, SSE support |
| 7 | * Gareth Hughes <gareth@valinux.com>, May 2000 |
| 8 | */ |
| 9 | |
| 10 | /* |
| 11 | * Handle hardware traps and faults. |
| 12 | */ |
| 13 | #include <linux/spinlock.h> |
| 14 | #include <linux/kprobes.h> |
| 15 | #include <linux/kdebug.h> |
| 16 | #include <linux/nmi.h> |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 17 | #include <linux/debugfs.h> |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 18 | #include <linux/delay.h> |
| 19 | #include <linux/hardirq.h> |
Arnd Bergmann | c361db5 | 2016-06-06 16:09:52 +0200 | [diff] [blame] | 20 | #include <linux/ratelimit.h> |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 21 | #include <linux/slab.h> |
Paul Gortmaker | 69c60c8 | 2011-05-26 12:22:53 -0400 | [diff] [blame] | 22 | #include <linux/export.h> |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 23 | |
| 24 | #if defined(CONFIG_EDAC) |
| 25 | #include <linux/edac.h> |
| 26 | #endif |
| 27 | |
| 28 | #include <linux/atomic.h> |
| 29 | #include <asm/traps.h> |
| 30 | #include <asm/mach_traps.h> |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 31 | #include <asm/nmi.h> |
Mathias Nyman | 6fd36ba | 2011-11-10 13:45:24 +0000 | [diff] [blame] | 32 | #include <asm/x86_init.h> |
Hidehiro Kawai | b279d67 | 2015-12-14 11:19:13 +0100 | [diff] [blame] | 33 | #include <asm/reboot.h> |
Kostenzer Felix | 8e2a7f5 | 2016-03-06 23:20:06 +0100 | [diff] [blame] | 34 | #include <asm/cache.h> |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 35 | |
Dave Hansen | 0c4df02 | 2013-06-21 08:51:38 -0700 | [diff] [blame] | 36 | #define CREATE_TRACE_POINTS |
| 37 | #include <trace/events/nmi.h> |
| 38 | |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 39 | struct nmi_desc { |
| 40 | spinlock_t lock; |
| 41 | struct list_head head; |
| 42 | }; |
| 43 | |
| 44 | static struct nmi_desc nmi_desc[NMI_MAX] = |
| 45 | { |
| 46 | { |
| 47 | .lock = __SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock), |
| 48 | .head = LIST_HEAD_INIT(nmi_desc[0].head), |
| 49 | }, |
| 50 | { |
| 51 | .lock = __SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock), |
| 52 | .head = LIST_HEAD_INIT(nmi_desc[1].head), |
| 53 | }, |
Don Zickus | 553222f | 2012-03-29 16:11:16 -0400 | [diff] [blame] | 54 | { |
| 55 | .lock = __SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock), |
| 56 | .head = LIST_HEAD_INIT(nmi_desc[2].head), |
| 57 | }, |
| 58 | { |
| 59 | .lock = __SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock), |
| 60 | .head = LIST_HEAD_INIT(nmi_desc[3].head), |
| 61 | }, |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 62 | |
| 63 | }; |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 64 | |
Don Zickus | efc3aac | 2011-09-30 15:06:23 -0400 | [diff] [blame] | 65 | struct nmi_stats { |
| 66 | unsigned int normal; |
| 67 | unsigned int unknown; |
| 68 | unsigned int external; |
| 69 | unsigned int swallow; |
| 70 | }; |
| 71 | |
| 72 | static DEFINE_PER_CPU(struct nmi_stats, nmi_stats); |
| 73 | |
Kostenzer Felix | 8e2a7f5 | 2016-03-06 23:20:06 +0100 | [diff] [blame] | 74 | static int ignore_nmis __read_mostly; |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 75 | |
| 76 | int unknown_nmi_panic; |
| 77 | /* |
| 78 | * Prevent NMI reason port (0x61) being accessed simultaneously, can |
| 79 | * only be used in NMI handler. |
| 80 | */ |
| 81 | static DEFINE_RAW_SPINLOCK(nmi_reason_lock); |
| 82 | |
| 83 | static int __init setup_unknown_nmi_panic(char *str) |
| 84 | { |
| 85 | unknown_nmi_panic = 1; |
| 86 | return 1; |
| 87 | } |
| 88 | __setup("unknown_nmi_panic", setup_unknown_nmi_panic); |
| 89 | |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 90 | #define nmi_to_desc(type) (&nmi_desc[type]) |
| 91 | |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 92 | static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC; |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 93 | |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 94 | static int __init nmi_warning_debugfs(void) |
| 95 | { |
| 96 | debugfs_create_u64("nmi_longest_ns", 0644, |
| 97 | arch_debugfs_dir, &nmi_longest_ns); |
| 98 | return 0; |
| 99 | } |
| 100 | fs_initcall(nmi_warning_debugfs); |
| 101 | |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 102 | static void nmi_max_handler(struct irq_work *w) |
| 103 | { |
| 104 | struct nmiaction *a = container_of(w, struct nmiaction, irq_work); |
| 105 | int remainder_ns, decimal_msecs; |
| 106 | u64 whole_msecs = ACCESS_ONCE(a->max_duration); |
| 107 | |
| 108 | remainder_ns = do_div(whole_msecs, (1000 * 1000)); |
| 109 | decimal_msecs = remainder_ns / 1000; |
| 110 | |
| 111 | printk_ratelimited(KERN_INFO |
| 112 | "INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n", |
| 113 | a->handler, whole_msecs, decimal_msecs); |
| 114 | } |
| 115 | |
Andy Lutomirski | bf9f2ee | 2015-07-20 11:52:23 -0700 | [diff] [blame] | 116 | static int nmi_handle(unsigned int type, struct pt_regs *regs) |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 117 | { |
| 118 | struct nmi_desc *desc = nmi_to_desc(type); |
| 119 | struct nmiaction *a; |
| 120 | int handled=0; |
| 121 | |
| 122 | rcu_read_lock(); |
| 123 | |
| 124 | /* |
| 125 | * NMIs are edge-triggered, which means if you have enough |
| 126 | * of them concurrently, you can lose some because only one |
| 127 | * can be latched at any given time. Walk the whole list |
| 128 | * to handle those situations. |
| 129 | */ |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 130 | list_for_each_entry_rcu(a, &desc->head, list) { |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 131 | int thishandled; |
| 132 | u64 delta; |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 133 | |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 134 | delta = sched_clock(); |
Dave Hansen | 0c4df02 | 2013-06-21 08:51:38 -0700 | [diff] [blame] | 135 | thishandled = a->handler(type, regs); |
| 136 | handled += thishandled; |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 137 | delta = sched_clock() - delta; |
Dave Hansen | 0c4df02 | 2013-06-21 08:51:38 -0700 | [diff] [blame] | 138 | trace_nmi_handler(a->handler, (int)delta, thishandled); |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 139 | |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 140 | if (delta < nmi_longest_ns || delta < a->max_duration) |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 141 | continue; |
| 142 | |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 143 | a->max_duration = delta; |
| 144 | irq_work_queue(&a->irq_work); |
Dave Hansen | 2ab00456 | 2013-06-21 08:51:35 -0700 | [diff] [blame] | 145 | } |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 146 | |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 147 | rcu_read_unlock(); |
| 148 | |
| 149 | /* return total number of NMI events handled */ |
| 150 | return handled; |
| 151 | } |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 152 | NOKPROBE_SYMBOL(nmi_handle); |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 153 | |
Li Zhong | 72b3fb2 | 2012-03-29 16:11:17 -0400 | [diff] [blame] | 154 | int __register_nmi_handler(unsigned int type, struct nmiaction *action) |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 155 | { |
| 156 | struct nmi_desc *desc = nmi_to_desc(type); |
| 157 | unsigned long flags; |
| 158 | |
Li Zhong | 72b3fb2 | 2012-03-29 16:11:17 -0400 | [diff] [blame] | 159 | if (!action->handler) |
| 160 | return -EINVAL; |
| 161 | |
Peter Zijlstra | e90c785 | 2014-02-03 18:02:09 +0100 | [diff] [blame] | 162 | init_irq_work(&action->irq_work, nmi_max_handler); |
| 163 | |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 164 | spin_lock_irqsave(&desc->lock, flags); |
| 165 | |
| 166 | /* |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 167 | * most handlers of type NMI_UNKNOWN never return because |
| 168 | * they just assume the NMI is theirs. Just a sanity check |
| 169 | * to manage expectations |
| 170 | */ |
| 171 | WARN_ON_ONCE(type == NMI_UNKNOWN && !list_empty(&desc->head)); |
Don Zickus | 553222f | 2012-03-29 16:11:16 -0400 | [diff] [blame] | 172 | WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head)); |
| 173 | WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head)); |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 174 | |
| 175 | /* |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 176 | * some handlers need to be executed first otherwise a fake |
| 177 | * event confuses some handlers (kdump uses this flag) |
| 178 | */ |
| 179 | if (action->flags & NMI_FLAG_FIRST) |
| 180 | list_add_rcu(&action->list, &desc->head); |
| 181 | else |
| 182 | list_add_tail_rcu(&action->list, &desc->head); |
| 183 | |
| 184 | spin_unlock_irqrestore(&desc->lock, flags); |
| 185 | return 0; |
| 186 | } |
Li Zhong | 72b3fb2 | 2012-03-29 16:11:17 -0400 | [diff] [blame] | 187 | EXPORT_SYMBOL(__register_nmi_handler); |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 188 | |
Li Zhong | 72b3fb2 | 2012-03-29 16:11:17 -0400 | [diff] [blame] | 189 | void unregister_nmi_handler(unsigned int type, const char *name) |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 190 | { |
| 191 | struct nmi_desc *desc = nmi_to_desc(type); |
| 192 | struct nmiaction *n; |
| 193 | unsigned long flags; |
| 194 | |
| 195 | spin_lock_irqsave(&desc->lock, flags); |
| 196 | |
| 197 | list_for_each_entry_rcu(n, &desc->head, list) { |
| 198 | /* |
| 199 | * the name passed in to describe the nmi handler |
| 200 | * is used as the lookup key |
| 201 | */ |
| 202 | if (!strcmp(n->name, name)) { |
| 203 | WARN(in_nmi(), |
| 204 | "Trying to free NMI (%s) from NMI context!\n", n->name); |
| 205 | list_del_rcu(&n->list); |
| 206 | break; |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | spin_unlock_irqrestore(&desc->lock, flags); |
| 211 | synchronize_rcu(); |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 212 | } |
Don Zickus | c9126b2 | 2011-09-30 15:06:20 -0400 | [diff] [blame] | 213 | EXPORT_SYMBOL_GPL(unregister_nmi_handler); |
| 214 | |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 215 | static void |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 216 | pci_serr_error(unsigned char reason, struct pt_regs *regs) |
| 217 | { |
Don Zickus | 553222f | 2012-03-29 16:11:16 -0400 | [diff] [blame] | 218 | /* check to see if anyone registered against these types of errors */ |
Andy Lutomirski | bf9f2ee | 2015-07-20 11:52:23 -0700 | [diff] [blame] | 219 | if (nmi_handle(NMI_SERR, regs)) |
Don Zickus | 553222f | 2012-03-29 16:11:16 -0400 | [diff] [blame] | 220 | return; |
| 221 | |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 222 | pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n", |
| 223 | reason, smp_processor_id()); |
| 224 | |
| 225 | /* |
| 226 | * On some machines, PCI SERR line is used to report memory |
| 227 | * errors. EDAC makes use of it. |
| 228 | */ |
| 229 | #if defined(CONFIG_EDAC) |
| 230 | if (edac_handler_set()) { |
| 231 | edac_atomic_assert_error(); |
| 232 | return; |
| 233 | } |
| 234 | #endif |
| 235 | |
| 236 | if (panic_on_unrecovered_nmi) |
Hidehiro Kawai | 58c5661 | 2015-12-14 11:19:10 +0100 | [diff] [blame] | 237 | nmi_panic(regs, "NMI: Not continuing"); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 238 | |
| 239 | pr_emerg("Dazed and confused, but trying to continue\n"); |
| 240 | |
| 241 | /* Clear and disable the PCI SERR error line. */ |
| 242 | reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR; |
| 243 | outb(reason, NMI_REASON_PORT); |
| 244 | } |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 245 | NOKPROBE_SYMBOL(pci_serr_error); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 246 | |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 247 | static void |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 248 | io_check_error(unsigned char reason, struct pt_regs *regs) |
| 249 | { |
| 250 | unsigned long i; |
| 251 | |
Don Zickus | 553222f | 2012-03-29 16:11:16 -0400 | [diff] [blame] | 252 | /* check to see if anyone registered against these types of errors */ |
Andy Lutomirski | bf9f2ee | 2015-07-20 11:52:23 -0700 | [diff] [blame] | 253 | if (nmi_handle(NMI_IO_CHECK, regs)) |
Don Zickus | 553222f | 2012-03-29 16:11:16 -0400 | [diff] [blame] | 254 | return; |
| 255 | |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 256 | pr_emerg( |
| 257 | "NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n", |
| 258 | reason, smp_processor_id()); |
Jan Beulich | 57da8b9 | 2012-05-09 08:47:37 +0100 | [diff] [blame] | 259 | show_regs(regs); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 260 | |
Hidehiro Kawai | 1717f20 | 2015-12-14 11:19:09 +0100 | [diff] [blame] | 261 | if (panic_on_io_nmi) { |
Hidehiro Kawai | 58c5661 | 2015-12-14 11:19:10 +0100 | [diff] [blame] | 262 | nmi_panic(regs, "NMI IOCK error: Not continuing"); |
Hidehiro Kawai | 1717f20 | 2015-12-14 11:19:09 +0100 | [diff] [blame] | 263 | |
| 264 | /* |
| 265 | * If we end up here, it means we have received an NMI while |
| 266 | * processing panic(). Simply return without delaying and |
| 267 | * re-enabling NMIs. |
| 268 | */ |
| 269 | return; |
| 270 | } |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 271 | |
| 272 | /* Re-enable the IOCK line, wait for a few seconds */ |
| 273 | reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK; |
| 274 | outb(reason, NMI_REASON_PORT); |
| 275 | |
| 276 | i = 20000; |
| 277 | while (--i) { |
| 278 | touch_nmi_watchdog(); |
| 279 | udelay(100); |
| 280 | } |
| 281 | |
| 282 | reason &= ~NMI_REASON_CLEAR_IOCHK; |
| 283 | outb(reason, NMI_REASON_PORT); |
| 284 | } |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 285 | NOKPROBE_SYMBOL(io_check_error); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 286 | |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 287 | static void |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 288 | unknown_nmi_error(unsigned char reason, struct pt_regs *regs) |
| 289 | { |
Don Zickus | 9c48f1c | 2011-09-30 15:06:21 -0400 | [diff] [blame] | 290 | int handled; |
| 291 | |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 292 | /* |
| 293 | * Use 'false' as back-to-back NMIs are dealt with one level up. |
| 294 | * Of course this makes having multiple 'unknown' handlers useless |
| 295 | * as only the first one is ever run (unless it can actually determine |
| 296 | * if it caused the NMI) |
| 297 | */ |
Andy Lutomirski | bf9f2ee | 2015-07-20 11:52:23 -0700 | [diff] [blame] | 298 | handled = nmi_handle(NMI_UNKNOWN, regs); |
Don Zickus | efc3aac | 2011-09-30 15:06:23 -0400 | [diff] [blame] | 299 | if (handled) { |
| 300 | __this_cpu_add(nmi_stats.unknown, handled); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 301 | return; |
Don Zickus | efc3aac | 2011-09-30 15:06:23 -0400 | [diff] [blame] | 302 | } |
| 303 | |
| 304 | __this_cpu_add(nmi_stats.unknown, 1); |
| 305 | |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 306 | pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n", |
| 307 | reason, smp_processor_id()); |
| 308 | |
| 309 | pr_emerg("Do you have a strange power saving mode enabled?\n"); |
| 310 | if (unknown_nmi_panic || panic_on_unrecovered_nmi) |
Hidehiro Kawai | 58c5661 | 2015-12-14 11:19:10 +0100 | [diff] [blame] | 311 | nmi_panic(regs, "NMI: Not continuing"); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 312 | |
| 313 | pr_emerg("Dazed and confused, but trying to continue\n"); |
| 314 | } |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 315 | NOKPROBE_SYMBOL(unknown_nmi_error); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 316 | |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 317 | static DEFINE_PER_CPU(bool, swallow_nmi); |
| 318 | static DEFINE_PER_CPU(unsigned long, last_nmi_rip); |
| 319 | |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 320 | static void default_do_nmi(struct pt_regs *regs) |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 321 | { |
| 322 | unsigned char reason = 0; |
Don Zickus | 9c48f1c | 2011-09-30 15:06:21 -0400 | [diff] [blame] | 323 | int handled; |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 324 | bool b2b = false; |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 325 | |
| 326 | /* |
| 327 | * CPU-specific NMI must be processed before non-CPU-specific |
| 328 | * NMI, otherwise we may lose it, because the CPU-specific |
| 329 | * NMI can not be detected/processed on other CPUs. |
| 330 | */ |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 331 | |
| 332 | /* |
| 333 | * Back-to-back NMIs are interesting because they can either |
| 334 | * be two NMI or more than two NMIs (any thing over two is dropped |
| 335 | * due to NMI being edge-triggered). If this is the second half |
| 336 | * of the back-to-back NMI, assume we dropped things and process |
| 337 | * more handlers. Otherwise reset the 'swallow' NMI behaviour |
| 338 | */ |
| 339 | if (regs->ip == __this_cpu_read(last_nmi_rip)) |
| 340 | b2b = true; |
| 341 | else |
| 342 | __this_cpu_write(swallow_nmi, false); |
| 343 | |
| 344 | __this_cpu_write(last_nmi_rip, regs->ip); |
| 345 | |
Andy Lutomirski | bf9f2ee | 2015-07-20 11:52:23 -0700 | [diff] [blame] | 346 | handled = nmi_handle(NMI_LOCAL, regs); |
Don Zickus | efc3aac | 2011-09-30 15:06:23 -0400 | [diff] [blame] | 347 | __this_cpu_add(nmi_stats.normal, handled); |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 348 | if (handled) { |
| 349 | /* |
| 350 | * There are cases when a NMI handler handles multiple |
| 351 | * events in the current NMI. One of these events may |
| 352 | * be queued for in the next NMI. Because the event is |
| 353 | * already handled, the next NMI will result in an unknown |
| 354 | * NMI. Instead lets flag this for a potential NMI to |
| 355 | * swallow. |
| 356 | */ |
| 357 | if (handled > 1) |
| 358 | __this_cpu_write(swallow_nmi, true); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 359 | return; |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 360 | } |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 361 | |
Hidehiro Kawai | b279d67 | 2015-12-14 11:19:13 +0100 | [diff] [blame] | 362 | /* |
| 363 | * Non-CPU-specific NMI: NMI sources can be processed on any CPU. |
| 364 | * |
| 365 | * Another CPU may be processing panic routines while holding |
| 366 | * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping, |
| 367 | * and if so, call its callback directly. If there is no CPU preparing |
| 368 | * crash dump, we simply loop here. |
| 369 | */ |
| 370 | while (!raw_spin_trylock(&nmi_reason_lock)) { |
| 371 | run_crash_ipi_callback(regs); |
| 372 | cpu_relax(); |
| 373 | } |
| 374 | |
Jacob Pan | 064a59b | 2011-11-10 13:43:05 +0000 | [diff] [blame] | 375 | reason = x86_platform.get_nmi_reason(); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 376 | |
| 377 | if (reason & NMI_REASON_MASK) { |
| 378 | if (reason & NMI_REASON_SERR) |
| 379 | pci_serr_error(reason, regs); |
| 380 | else if (reason & NMI_REASON_IOCHK) |
| 381 | io_check_error(reason, regs); |
| 382 | #ifdef CONFIG_X86_32 |
| 383 | /* |
| 384 | * Reassert NMI in case it became active |
| 385 | * meanwhile as it's edge-triggered: |
| 386 | */ |
| 387 | reassert_nmi(); |
| 388 | #endif |
Don Zickus | efc3aac | 2011-09-30 15:06:23 -0400 | [diff] [blame] | 389 | __this_cpu_add(nmi_stats.external, 1); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 390 | raw_spin_unlock(&nmi_reason_lock); |
| 391 | return; |
| 392 | } |
| 393 | raw_spin_unlock(&nmi_reason_lock); |
| 394 | |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 395 | /* |
| 396 | * Only one NMI can be latched at a time. To handle |
| 397 | * this we may process multiple nmi handlers at once to |
| 398 | * cover the case where an NMI is dropped. The downside |
| 399 | * to this approach is we may process an NMI prematurely, |
| 400 | * while its real NMI is sitting latched. This will cause |
| 401 | * an unknown NMI on the next run of the NMI processing. |
| 402 | * |
| 403 | * We tried to flag that condition above, by setting the |
| 404 | * swallow_nmi flag when we process more than one event. |
| 405 | * This condition is also only present on the second half |
| 406 | * of a back-to-back NMI, so we flag that condition too. |
| 407 | * |
| 408 | * If both are true, we assume we already processed this |
| 409 | * NMI previously and we swallow it. Otherwise we reset |
| 410 | * the logic. |
| 411 | * |
| 412 | * There are scenarios where we may accidentally swallow |
| 413 | * a 'real' unknown NMI. For example, while processing |
| 414 | * a perf NMI another perf NMI comes in along with a |
| 415 | * 'real' unknown NMI. These two NMIs get combined into |
| 416 | * one (as descibed above). When the next NMI gets |
| 417 | * processed, it will be flagged by perf as handled, but |
| 418 | * noone will know that there was a 'real' unknown NMI sent |
| 419 | * also. As a result it gets swallowed. Or if the first |
| 420 | * perf NMI returns two events handled then the second |
| 421 | * NMI will get eaten by the logic below, again losing a |
| 422 | * 'real' unknown NMI. But this is the best we can do |
| 423 | * for now. |
| 424 | */ |
| 425 | if (b2b && __this_cpu_read(swallow_nmi)) |
Don Zickus | efc3aac | 2011-09-30 15:06:23 -0400 | [diff] [blame] | 426 | __this_cpu_add(nmi_stats.swallow, 1); |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 427 | else |
| 428 | unknown_nmi_error(reason, regs); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 429 | } |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 430 | NOKPROBE_SYMBOL(default_do_nmi); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 431 | |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 432 | /* |
Andy Lutomirski | 0b22930 | 2015-07-15 10:29:36 -0700 | [diff] [blame] | 433 | * NMIs can page fault or hit breakpoints which will cause it to lose |
| 434 | * its NMI context with the CPU when the breakpoint or page fault does an IRET. |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 435 | * |
| 436 | * As a result, NMIs can nest if NMIs get unmasked due an IRET during |
| 437 | * NMI processing. On x86_64, the asm glue protects us from nested NMIs |
| 438 | * if the outer NMI came from kernel mode, but we can still nest if the |
| 439 | * outer NMI came from user mode. |
| 440 | * |
| 441 | * To handle these nested NMIs, we have three states: |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 442 | * |
| 443 | * 1) not running |
| 444 | * 2) executing |
| 445 | * 3) latched |
| 446 | * |
| 447 | * When no NMI is in progress, it is in the "not running" state. |
| 448 | * When an NMI comes in, it goes into the "executing" state. |
| 449 | * Normally, if another NMI is triggered, it does not interrupt |
| 450 | * the running NMI and the HW will simply latch it so that when |
| 451 | * the first NMI finishes, it will restart the second NMI. |
| 452 | * (Note, the latch is binary, thus multiple NMIs triggering, |
| 453 | * when one is running, are ignored. Only one NMI is restarted.) |
| 454 | * |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 455 | * If an NMI executes an iret, another NMI can preempt it. We do not |
| 456 | * want to allow this new NMI to run, but we want to execute it when the |
| 457 | * first one finishes. We set the state to "latched", and the exit of |
| 458 | * the first NMI will perform a dec_return, if the result is zero |
| 459 | * (NOT_RUNNING), then it will simply exit the NMI handler. If not, the |
| 460 | * dec_return would have set the state to NMI_EXECUTING (what we want it |
| 461 | * to be when we are running). In this case, we simply jump back to |
| 462 | * rerun the NMI handler again, and restart the 'latched' NMI. |
Steven Rostedt | c7d65a7 | 2012-06-07 11:03:00 -0400 | [diff] [blame] | 463 | * |
| 464 | * No trap (breakpoint or page fault) should be hit before nmi_restart, |
| 465 | * thus there is no race between the first check of state for NOT_RUNNING |
| 466 | * and setting it to NMI_EXECUTING. The HW will prevent nested NMIs |
| 467 | * at this point. |
Steven Rostedt | 70fb74a | 2012-06-07 11:54:37 -0400 | [diff] [blame] | 468 | * |
| 469 | * In case the NMI takes a page fault, we need to save off the CR2 |
| 470 | * because the NMI could have preempted another page fault and corrupt |
| 471 | * the CR2 that is about to be read. As nested NMIs must be restarted |
| 472 | * and they can not take breakpoints or page faults, the update of the |
| 473 | * CR2 must be done before converting the nmi state back to NOT_RUNNING. |
| 474 | * Otherwise, there would be a race of another nested NMI coming in |
| 475 | * after setting state to NOT_RUNNING but before updating the nmi_cr2. |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 476 | */ |
| 477 | enum nmi_states { |
Steven Rostedt | c7d65a7 | 2012-06-07 11:03:00 -0400 | [diff] [blame] | 478 | NMI_NOT_RUNNING = 0, |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 479 | NMI_EXECUTING, |
| 480 | NMI_LATCHED, |
| 481 | }; |
| 482 | static DEFINE_PER_CPU(enum nmi_states, nmi_state); |
Steven Rostedt | 70fb74a | 2012-06-07 11:54:37 -0400 | [diff] [blame] | 483 | static DEFINE_PER_CPU(unsigned long, nmi_cr2); |
Steven Rostedt | 228bdaa | 2011-12-09 03:02:19 -0500 | [diff] [blame] | 484 | |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 485 | #ifdef CONFIG_X86_64 |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 486 | /* |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 487 | * In x86_64, we need to handle breakpoint -> NMI -> breakpoint. Without |
| 488 | * some care, the inner breakpoint will clobber the outer breakpoint's |
| 489 | * stack. |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 490 | * |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 491 | * If a breakpoint is being processed, and the debug stack is being |
| 492 | * used, if an NMI comes in and also hits a breakpoint, the stack |
| 493 | * pointer will be set to the same fixed address as the breakpoint that |
| 494 | * was interrupted, causing that stack to be corrupted. To handle this |
| 495 | * case, check if the stack that was interrupted is the debug stack, and |
| 496 | * if so, change the IDT so that new breakpoints will use the current |
| 497 | * stack and not switch to the fixed address. On return of the NMI, |
| 498 | * switch back to the original IDT. |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 499 | */ |
| 500 | static DEFINE_PER_CPU(int, update_debug_stack); |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 501 | #endif |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 502 | |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 503 | dotraplinkage notrace void |
| 504 | do_nmi(struct pt_regs *regs, long error_code) |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 505 | { |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 506 | if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) { |
| 507 | this_cpu_write(nmi_state, NMI_LATCHED); |
| 508 | return; |
| 509 | } |
| 510 | this_cpu_write(nmi_state, NMI_EXECUTING); |
| 511 | this_cpu_write(nmi_cr2, read_cr2()); |
| 512 | nmi_restart: |
| 513 | |
| 514 | #ifdef CONFIG_X86_64 |
Steven Rostedt | 228bdaa | 2011-12-09 03:02:19 -0500 | [diff] [blame] | 515 | /* |
| 516 | * If we interrupted a breakpoint, it is possible that |
| 517 | * the nmi handler will have breakpoints too. We need to |
| 518 | * change the IDT such that breakpoints that happen here |
| 519 | * continue to use the NMI stack. |
| 520 | */ |
| 521 | if (unlikely(is_debug_stack(regs->sp))) { |
| 522 | debug_stack_set_zero(); |
Steven Rostedt | c0525a6 | 2012-05-30 11:43:19 -0400 | [diff] [blame] | 523 | this_cpu_write(update_debug_stack, 1); |
Steven Rostedt | 228bdaa | 2011-12-09 03:02:19 -0500 | [diff] [blame] | 524 | } |
Steven Rostedt | ccd49c2 | 2011-12-13 16:44:16 -0500 | [diff] [blame] | 525 | #endif |
| 526 | |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 527 | nmi_enter(); |
| 528 | |
| 529 | inc_irq_stat(__nmi_count); |
| 530 | |
| 531 | if (!ignore_nmis) |
| 532 | default_do_nmi(regs); |
| 533 | |
| 534 | nmi_exit(); |
Steven Rostedt | 228bdaa | 2011-12-09 03:02:19 -0500 | [diff] [blame] | 535 | |
Andy Lutomirski | 9d05041 | 2015-07-15 10:29:33 -0700 | [diff] [blame] | 536 | #ifdef CONFIG_X86_64 |
| 537 | if (unlikely(this_cpu_read(update_debug_stack))) { |
| 538 | debug_stack_reset(); |
| 539 | this_cpu_write(update_debug_stack, 0); |
| 540 | } |
| 541 | #endif |
| 542 | |
| 543 | if (unlikely(this_cpu_read(nmi_cr2) != read_cr2())) |
| 544 | write_cr2(this_cpu_read(nmi_cr2)); |
| 545 | if (this_cpu_dec_return(nmi_state)) |
| 546 | goto nmi_restart; |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 547 | } |
Masami Hiramatsu | 9326638 | 2014-04-17 17:18:14 +0900 | [diff] [blame] | 548 | NOKPROBE_SYMBOL(do_nmi); |
Don Zickus | 1d48922 | 2011-09-30 15:06:19 -0400 | [diff] [blame] | 549 | |
| 550 | void stop_nmi(void) |
| 551 | { |
| 552 | ignore_nmis++; |
| 553 | } |
| 554 | |
| 555 | void restart_nmi(void) |
| 556 | { |
| 557 | ignore_nmis--; |
| 558 | } |
Don Zickus | b227e23 | 2011-09-30 15:06:22 -0400 | [diff] [blame] | 559 | |
| 560 | /* reset the back-to-back NMI logic */ |
| 561 | void local_touch_nmi(void) |
| 562 | { |
| 563 | __this_cpu_write(last_nmi_rip, 0); |
| 564 | } |
Jacob Pan | 29c6fb7 | 2013-01-04 11:12:44 +0000 | [diff] [blame] | 565 | EXPORT_SYMBOL_GPL(local_touch_nmi); |