1/*
2 *  (c) 2005-2012 Advanced Micro Devices, Inc.
3 *  Your use of this code is subject to the terms and conditions of the
4 *  GNU general public license version 2. See "COPYING" or
5 *  http://www.gnu.org/licenses/gpl.html
6 *
7 *  Written by Jacob Shin - AMD, Inc.
8 *
9 *  Maintained by: Borislav Petkov <bp@alien8.de>
10 *
11 *  April 2006
12 *     - added support for AMD Family 0x10 processors
13 *  May 2012
14 *     - major scrubbing
15 *
16 *  All MC4_MISCi registers are shared between multi-cores
17 */
18#include <linux/interrupt.h>
19#include <linux/notifier.h>
20#include <linux/kobject.h>
21#include <linux/percpu.h>
22#include <linux/errno.h>
23#include <linux/sched.h>
24#include <linux/sysfs.h>
25#include <linux/slab.h>
26#include <linux/init.h>
27#include <linux/cpu.h>
28#include <linux/smp.h>
29
30#include <asm/amd_nb.h>
31#include <asm/apic.h>
32#include <asm/idle.h>
33#include <asm/mce.h>
34#include <asm/msr.h>
35
36#define NR_BLOCKS         9
37#define THRESHOLD_MAX     0xFFF
38#define INT_TYPE_APIC     0x00020000
39#define MASK_VALID_HI     0x80000000
40#define MASK_CNTP_HI      0x40000000
41#define MASK_LOCKED_HI    0x20000000
42#define MASK_LVTOFF_HI    0x00F00000
43#define MASK_COUNT_EN_HI  0x00080000
44#define MASK_INT_TYPE_HI  0x00060000
45#define MASK_OVERFLOW_HI  0x00010000
46#define MASK_ERR_COUNT_HI 0x00000FFF
47#define MASK_BLKPTR_LO    0xFF000000
48#define MCG_XBLK_ADDR     0xC0000400
49
50static const char * const th_names[] = {
51	"load_store",
52	"insn_fetch",
53	"combined_unit",
54	"",
55	"northbridge",
56	"execution_unit",
57};
58
59static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
60static DEFINE_PER_CPU(unsigned char, bank_map);	/* see which banks are on */
61
62static void amd_threshold_interrupt(void);
63
64/*
65 * CPU Initialization
66 */
67
68struct thresh_restart {
69	struct threshold_block	*b;
70	int			reset;
71	int			set_lvt_off;
72	int			lvt_off;
73	u16			old_limit;
74};
75
76static inline bool is_shared_bank(int bank)
77{
78	/* Bank 4 is for northbridge reporting and is thus shared */
79	return (bank == 4);
80}
81
82static const char *bank4_names(const struct threshold_block *b)
83{
84	switch (b->address) {
85	/* MSR4_MISC0 */
86	case 0x00000413:
87		return "dram";
88
89	case 0xc0000408:
90		return "ht_links";
91
92	case 0xc0000409:
93		return "l3_cache";
94
95	default:
96		WARN(1, "Funny MSR: 0x%08x\n", b->address);
97		return "";
98	}
99};
100
101
102static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
103{
104	/*
105	 * bank 4 supports APIC LVT interrupts implicitly since forever.
106	 */
107	if (bank == 4)
108		return true;
109
110	/*
111	 * IntP: interrupt present; if this bit is set, the thresholding
112	 * bank can generate APIC LVT interrupts
113	 */
114	return msr_high_bits & BIT(28);
115}
116
117static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
118{
119	int msr = (hi & MASK_LVTOFF_HI) >> 20;
120
121	if (apic < 0) {
122		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
123		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
124		       b->bank, b->block, b->address, hi, lo);
125		return 0;
126	}
127
128	if (apic != msr) {
129		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
130		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
131		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
132		return 0;
133	}
134
135	return 1;
136};
137
138/*
139 * Called via smp_call_function_single(), must be called with correct
140 * cpu affinity.
141 */
142static void threshold_restart_bank(void *_tr)
143{
144	struct thresh_restart *tr = _tr;
145	u32 hi, lo;
146
147	rdmsr(tr->b->address, lo, hi);
148
149	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
150		tr->reset = 1;	/* limit cannot be lower than err count */
151
152	if (tr->reset) {		/* reset err count and overflow bit */
153		hi =
154		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
155		    (THRESHOLD_MAX - tr->b->threshold_limit);
156	} else if (tr->old_limit) {	/* change limit w/o reset */
157		int new_count = (hi & THRESHOLD_MAX) +
158		    (tr->old_limit - tr->b->threshold_limit);
159
160		hi = (hi & ~MASK_ERR_COUNT_HI) |
161		    (new_count & THRESHOLD_MAX);
162	}
163
164	/* clear IntType */
165	hi &= ~MASK_INT_TYPE_HI;
166
167	if (!tr->b->interrupt_capable)
168		goto done;
169
170	if (tr->set_lvt_off) {
171		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
172			/* set new lvt offset */
173			hi &= ~MASK_LVTOFF_HI;
174			hi |= tr->lvt_off << 20;
175		}
176	}
177
178	if (tr->b->interrupt_enable)
179		hi |= INT_TYPE_APIC;
180
181 done:
182
183	hi |= MASK_COUNT_EN_HI;
184	wrmsr(tr->b->address, lo, hi);
185}
186
187static void mce_threshold_block_init(struct threshold_block *b, int offset)
188{
189	struct thresh_restart tr = {
190		.b			= b,
191		.set_lvt_off		= 1,
192		.lvt_off		= offset,
193	};
194
195	b->threshold_limit		= THRESHOLD_MAX;
196	threshold_restart_bank(&tr);
197};
198
199static int setup_APIC_mce(int reserved, int new)
200{
201	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
202					      APIC_EILVT_MSG_FIX, 0))
203		return new;
204
205	return reserved;
206}
207
208/* cpu init entry point, called from mce.c with preempt off */
209void mce_amd_feature_init(struct cpuinfo_x86 *c)
210{
211	struct threshold_block b;
212	unsigned int cpu = smp_processor_id();
213	u32 low = 0, high = 0, address = 0;
214	unsigned int bank, block;
215	int offset = -1, new;
216
217	for (bank = 0; bank < mca_cfg.banks; ++bank) {
218		for (block = 0; block < NR_BLOCKS; ++block) {
219			if (block == 0)
220				address = MSR_IA32_MCx_MISC(bank);
221			else if (block == 1) {
222				address = (low & MASK_BLKPTR_LO) >> 21;
223				if (!address)
224					break;
225
226				address += MCG_XBLK_ADDR;
227			} else
228				++address;
229
230			if (rdmsr_safe(address, &low, &high))
231				break;
232
233			if (!(high & MASK_VALID_HI))
234				continue;
235
236			if (!(high & MASK_CNTP_HI)  ||
237			     (high & MASK_LOCKED_HI))
238				continue;
239
240			if (!block)
241				per_cpu(bank_map, cpu) |= (1 << bank);
242
243			memset(&b, 0, sizeof(b));
244			b.cpu			= cpu;
245			b.bank			= bank;
246			b.block			= block;
247			b.address		= address;
248			b.interrupt_capable	= lvt_interrupt_supported(bank, high);
249
250			if (!b.interrupt_capable)
251				goto init;
252
253			b.interrupt_enable = 1;
254			new	= (high & MASK_LVTOFF_HI) >> 20;
255			offset  = setup_APIC_mce(offset, new);
256
257			if ((offset == new) &&
258			    (mce_threshold_vector != amd_threshold_interrupt))
259				mce_threshold_vector = amd_threshold_interrupt;
260
261init:
262			mce_threshold_block_init(&b, offset);
263		}
264	}
265}
266
267/*
268 * APIC Interrupt Handler
269 */
270
271/*
272 * threshold interrupt handler will service THRESHOLD_APIC_VECTOR.
273 * the interrupt goes off when error_count reaches threshold_limit.
274 * the handler will simply log mcelog w/ software defined bank number.
275 */
276static void amd_threshold_interrupt(void)
277{
278	u32 low = 0, high = 0, address = 0;
279	int cpu = smp_processor_id();
280	unsigned int bank, block;
281	struct mce m;
282
283	/* assume first bank caused it */
284	for (bank = 0; bank < mca_cfg.banks; ++bank) {
285		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
286			continue;
287		for (block = 0; block < NR_BLOCKS; ++block) {
288			if (block == 0) {
289				address = MSR_IA32_MCx_MISC(bank);
290			} else if (block == 1) {
291				address = (low & MASK_BLKPTR_LO) >> 21;
292				if (!address)
293					break;
294				address += MCG_XBLK_ADDR;
295			} else {
296				++address;
297			}
298
299			if (rdmsr_safe(address, &low, &high))
300				break;
301
302			if (!(high & MASK_VALID_HI)) {
303				if (block)
304					continue;
305				else
306					break;
307			}
308
309			if (!(high & MASK_CNTP_HI)  ||
310			     (high & MASK_LOCKED_HI))
311				continue;
312
313			/*
314			 * Log the machine check that caused the threshold
315			 * event.
316			 */
317			if (high & MASK_OVERFLOW_HI)
318				goto log;
319		}
320	}
321	return;
322
323log:
324	mce_setup(&m);
325	rdmsrl(MSR_IA32_MCx_STATUS(bank), m.status);
326	if (!(m.status & MCI_STATUS_VAL))
327		return;
328	m.misc = ((u64)high << 32) | low;
329	m.bank = bank;
330	mce_log(&m);
331
332	wrmsrl(MSR_IA32_MCx_STATUS(bank), 0);
333}
334
335/*
336 * Sysfs Interface
337 */
338
339struct threshold_attr {
340	struct attribute attr;
341	ssize_t (*show) (struct threshold_block *, char *);
342	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
343};
344
345#define SHOW_FIELDS(name)						\
346static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
347{									\
348	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
349}
350SHOW_FIELDS(interrupt_enable)
351SHOW_FIELDS(threshold_limit)
352
353static ssize_t
354store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
355{
356	struct thresh_restart tr;
357	unsigned long new;
358
359	if (!b->interrupt_capable)
360		return -EINVAL;
361
362	if (kstrtoul(buf, 0, &new) < 0)
363		return -EINVAL;
364
365	b->interrupt_enable = !!new;
366
367	memset(&tr, 0, sizeof(tr));
368	tr.b		= b;
369
370	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
371
372	return size;
373}
374
375static ssize_t
376store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
377{
378	struct thresh_restart tr;
379	unsigned long new;
380
381	if (kstrtoul(buf, 0, &new) < 0)
382		return -EINVAL;
383
384	if (new > THRESHOLD_MAX)
385		new = THRESHOLD_MAX;
386	if (new < 1)
387		new = 1;
388
389	memset(&tr, 0, sizeof(tr));
390	tr.old_limit = b->threshold_limit;
391	b->threshold_limit = new;
392	tr.b = b;
393
394	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
395
396	return size;
397}
398
399static ssize_t show_error_count(struct threshold_block *b, char *buf)
400{
401	u32 lo, hi;
402
403	rdmsr_on_cpu(b->cpu, b->address, &lo, &hi);
404
405	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
406				     (THRESHOLD_MAX - b->threshold_limit)));
407}
408
409static struct threshold_attr error_count = {
410	.attr = {.name = __stringify(error_count), .mode = 0444 },
411	.show = show_error_count,
412};
413
414#define RW_ATTR(val)							\
415static struct threshold_attr val = {					\
416	.attr	= {.name = __stringify(val), .mode = 0644 },		\
417	.show	= show_## val,						\
418	.store	= store_## val,						\
419};
420
421RW_ATTR(interrupt_enable);
422RW_ATTR(threshold_limit);
423
424static struct attribute *default_attrs[] = {
425	&threshold_limit.attr,
426	&error_count.attr,
427	NULL,	/* possibly interrupt_enable if supported, see below */
428	NULL,
429};
430
431#define to_block(k)	container_of(k, struct threshold_block, kobj)
432#define to_attr(a)	container_of(a, struct threshold_attr, attr)
433
434static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
435{
436	struct threshold_block *b = to_block(kobj);
437	struct threshold_attr *a = to_attr(attr);
438	ssize_t ret;
439
440	ret = a->show ? a->show(b, buf) : -EIO;
441
442	return ret;
443}
444
445static ssize_t store(struct kobject *kobj, struct attribute *attr,
446		     const char *buf, size_t count)
447{
448	struct threshold_block *b = to_block(kobj);
449	struct threshold_attr *a = to_attr(attr);
450	ssize_t ret;
451
452	ret = a->store ? a->store(b, buf, count) : -EIO;
453
454	return ret;
455}
456
457static const struct sysfs_ops threshold_ops = {
458	.show			= show,
459	.store			= store,
460};
461
462static struct kobj_type threshold_ktype = {
463	.sysfs_ops		= &threshold_ops,
464	.default_attrs		= default_attrs,
465};
466
467static int allocate_threshold_blocks(unsigned int cpu, unsigned int bank,
468				     unsigned int block, u32 address)
469{
470	struct threshold_block *b = NULL;
471	u32 low, high;
472	int err;
473
474	if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
475		return 0;
476
477	if (rdmsr_safe_on_cpu(cpu, address, &low, &high))
478		return 0;
479
480	if (!(high & MASK_VALID_HI)) {
481		if (block)
482			goto recurse;
483		else
484			return 0;
485	}
486
487	if (!(high & MASK_CNTP_HI)  ||
488	     (high & MASK_LOCKED_HI))
489		goto recurse;
490
491	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
492	if (!b)
493		return -ENOMEM;
494
495	b->block		= block;
496	b->bank			= bank;
497	b->cpu			= cpu;
498	b->address		= address;
499	b->interrupt_enable	= 0;
500	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
501	b->threshold_limit	= THRESHOLD_MAX;
502
503	if (b->interrupt_capable) {
504		threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
505		b->interrupt_enable = 1;
506	} else {
507		threshold_ktype.default_attrs[2] = NULL;
508	}
509
510	INIT_LIST_HEAD(&b->miscj);
511
512	if (per_cpu(threshold_banks, cpu)[bank]->blocks) {
513		list_add(&b->miscj,
514			 &per_cpu(threshold_banks, cpu)[bank]->blocks->miscj);
515	} else {
516		per_cpu(threshold_banks, cpu)[bank]->blocks = b;
517	}
518
519	err = kobject_init_and_add(&b->kobj, &threshold_ktype,
520				   per_cpu(threshold_banks, cpu)[bank]->kobj,
521				   (bank == 4 ? bank4_names(b) : th_names[bank]));
522	if (err)
523		goto out_free;
524recurse:
525	if (!block) {
526		address = (low & MASK_BLKPTR_LO) >> 21;
527		if (!address)
528			return 0;
529		address += MCG_XBLK_ADDR;
530	} else {
531		++address;
532	}
533
534	err = allocate_threshold_blocks(cpu, bank, ++block, address);
535	if (err)
536		goto out_free;
537
538	if (b)
539		kobject_uevent(&b->kobj, KOBJ_ADD);
540
541	return err;
542
543out_free:
544	if (b) {
545		kobject_put(&b->kobj);
546		list_del(&b->miscj);
547		kfree(b);
548	}
549	return err;
550}
551
552static int __threshold_add_blocks(struct threshold_bank *b)
553{
554	struct list_head *head = &b->blocks->miscj;
555	struct threshold_block *pos = NULL;
556	struct threshold_block *tmp = NULL;
557	int err = 0;
558
559	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
560	if (err)
561		return err;
562
563	list_for_each_entry_safe(pos, tmp, head, miscj) {
564
565		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
566		if (err) {
567			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
568				kobject_del(&pos->kobj);
569
570			return err;
571		}
572	}
573	return err;
574}
575
576static int threshold_create_bank(unsigned int cpu, unsigned int bank)
577{
578	struct device *dev = per_cpu(mce_device, cpu);
579	struct amd_northbridge *nb = NULL;
580	struct threshold_bank *b = NULL;
581	const char *name = th_names[bank];
582	int err = 0;
583
584	if (is_shared_bank(bank)) {
585		nb = node_to_amd_nb(amd_get_nb_id(cpu));
586
587		/* threshold descriptor already initialized on this node? */
588		if (nb && nb->bank4) {
589			/* yes, use it */
590			b = nb->bank4;
591			err = kobject_add(b->kobj, &dev->kobj, name);
592			if (err)
593				goto out;
594
595			per_cpu(threshold_banks, cpu)[bank] = b;
596			atomic_inc(&b->cpus);
597
598			err = __threshold_add_blocks(b);
599
600			goto out;
601		}
602	}
603
604	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
605	if (!b) {
606		err = -ENOMEM;
607		goto out;
608	}
609
610	b->kobj = kobject_create_and_add(name, &dev->kobj);
611	if (!b->kobj) {
612		err = -EINVAL;
613		goto out_free;
614	}
615
616	per_cpu(threshold_banks, cpu)[bank] = b;
617
618	if (is_shared_bank(bank)) {
619		atomic_set(&b->cpus, 1);
620
621		/* nb is already initialized, see above */
622		if (nb) {
623			WARN_ON(nb->bank4);
624			nb->bank4 = b;
625		}
626	}
627
628	err = allocate_threshold_blocks(cpu, bank, 0, MSR_IA32_MCx_MISC(bank));
629	if (!err)
630		goto out;
631
632 out_free:
633	kfree(b);
634
635 out:
636	return err;
637}
638
639/* create dir/files for all valid threshold banks */
640static int threshold_create_device(unsigned int cpu)
641{
642	unsigned int bank;
643	struct threshold_bank **bp;
644	int err = 0;
645
646	bp = kzalloc(sizeof(struct threshold_bank *) * mca_cfg.banks,
647		     GFP_KERNEL);
648	if (!bp)
649		return -ENOMEM;
650
651	per_cpu(threshold_banks, cpu) = bp;
652
653	for (bank = 0; bank < mca_cfg.banks; ++bank) {
654		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
655			continue;
656		err = threshold_create_bank(cpu, bank);
657		if (err)
658			return err;
659	}
660
661	return err;
662}
663
664static void deallocate_threshold_block(unsigned int cpu,
665						 unsigned int bank)
666{
667	struct threshold_block *pos = NULL;
668	struct threshold_block *tmp = NULL;
669	struct threshold_bank *head = per_cpu(threshold_banks, cpu)[bank];
670
671	if (!head)
672		return;
673
674	list_for_each_entry_safe(pos, tmp, &head->blocks->miscj, miscj) {
675		kobject_put(&pos->kobj);
676		list_del(&pos->miscj);
677		kfree(pos);
678	}
679
680	kfree(per_cpu(threshold_banks, cpu)[bank]->blocks);
681	per_cpu(threshold_banks, cpu)[bank]->blocks = NULL;
682}
683
684static void __threshold_remove_blocks(struct threshold_bank *b)
685{
686	struct threshold_block *pos = NULL;
687	struct threshold_block *tmp = NULL;
688
689	kobject_del(b->kobj);
690
691	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
692		kobject_del(&pos->kobj);
693}
694
695static void threshold_remove_bank(unsigned int cpu, int bank)
696{
697	struct amd_northbridge *nb;
698	struct threshold_bank *b;
699
700	b = per_cpu(threshold_banks, cpu)[bank];
701	if (!b)
702		return;
703
704	if (!b->blocks)
705		goto free_out;
706
707	if (is_shared_bank(bank)) {
708		if (!atomic_dec_and_test(&b->cpus)) {
709			__threshold_remove_blocks(b);
710			per_cpu(threshold_banks, cpu)[bank] = NULL;
711			return;
712		} else {
713			/*
714			 * the last CPU on this node using the shared bank is
715			 * going away, remove that bank now.
716			 */
717			nb = node_to_amd_nb(amd_get_nb_id(cpu));
718			nb->bank4 = NULL;
719		}
720	}
721
722	deallocate_threshold_block(cpu, bank);
723
724free_out:
725	kobject_del(b->kobj);
726	kobject_put(b->kobj);
727	kfree(b);
728	per_cpu(threshold_banks, cpu)[bank] = NULL;
729}
730
731static void threshold_remove_device(unsigned int cpu)
732{
733	unsigned int bank;
734
735	for (bank = 0; bank < mca_cfg.banks; ++bank) {
736		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
737			continue;
738		threshold_remove_bank(cpu, bank);
739	}
740	kfree(per_cpu(threshold_banks, cpu));
741}
742
743/* get notified when a cpu comes on/off */
744static void
745amd_64_threshold_cpu_callback(unsigned long action, unsigned int cpu)
746{
747	switch (action) {
748	case CPU_ONLINE:
749	case CPU_ONLINE_FROZEN:
750		threshold_create_device(cpu);
751		break;
752	case CPU_DEAD:
753	case CPU_DEAD_FROZEN:
754		threshold_remove_device(cpu);
755		break;
756	default:
757		break;
758	}
759}
760
761static __init int threshold_init_device(void)
762{
763	unsigned lcpu = 0;
764
765	/* to hit CPUs online before the notifier is up */
766	for_each_online_cpu(lcpu) {
767		int err = threshold_create_device(lcpu);
768
769		if (err)
770			return err;
771	}
772	threshold_cpu_callback = amd_64_threshold_cpu_callback;
773
774	return 0;
775}
776/*
777 * there are 3 funcs which need to be _initcalled in a logic sequence:
778 * 1. xen_late_init_mcelog
779 * 2. mcheck_init_device
780 * 3. threshold_init_device
781 *
782 * xen_late_init_mcelog must register xen_mce_chrdev_device before
783 * native mce_chrdev_device registration if running under xen platform;
784 *
785 * mcheck_init_device should be inited before threshold_init_device to
786 * initialize mce_device, otherwise a NULL ptr dereference will cause panic.
787 *
788 * so we use following _initcalls
789 * 1. device_initcall(xen_late_init_mcelog);
790 * 2. device_initcall_sync(mcheck_init_device);
791 * 3. late_initcall(threshold_init_device);
792 *
793 * when running under xen, the initcall order is 1,2,3;
794 * on baremetal, we skip 1 and we do only 2 and 3.
795 */
796late_initcall(threshold_init_device);
797