xref: /openbmc/linux/arch/x86/kernel/cpu/mce/amd.c (revision 1cac4f26)
1 /*
2  *  (c) 2005-2016 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  *  Maintained by: Borislav Petkov <bp@alien8.de>
9  *
10  *  All MC4_MISCi registers are shared between cores on a node.
11  */
12 #include <linux/interrupt.h>
13 #include <linux/notifier.h>
14 #include <linux/kobject.h>
15 #include <linux/percpu.h>
16 #include <linux/errno.h>
17 #include <linux/sched.h>
18 #include <linux/sysfs.h>
19 #include <linux/slab.h>
20 #include <linux/init.h>
21 #include <linux/cpu.h>
22 #include <linux/smp.h>
23 #include <linux/string.h>
24 
25 #include <asm/amd_nb.h>
26 #include <asm/traps.h>
27 #include <asm/apic.h>
28 #include <asm/mce.h>
29 #include <asm/msr.h>
30 #include <asm/trace/irq_vectors.h>
31 
32 #include "internal.h"
33 
34 #define NR_BLOCKS         5
35 #define THRESHOLD_MAX     0xFFF
36 #define INT_TYPE_APIC     0x00020000
37 #define MASK_VALID_HI     0x80000000
38 #define MASK_CNTP_HI      0x40000000
39 #define MASK_LOCKED_HI    0x20000000
40 #define MASK_LVTOFF_HI    0x00F00000
41 #define MASK_COUNT_EN_HI  0x00080000
42 #define MASK_INT_TYPE_HI  0x00060000
43 #define MASK_OVERFLOW_HI  0x00010000
44 #define MASK_ERR_COUNT_HI 0x00000FFF
45 #define MASK_BLKPTR_LO    0xFF000000
46 #define MCG_XBLK_ADDR     0xC0000400
47 
48 /* Deferred error settings */
49 #define MSR_CU_DEF_ERR		0xC0000410
50 #define MASK_DEF_LVTOFF		0x000000F0
51 #define MASK_DEF_INT_TYPE	0x00000006
52 #define DEF_LVT_OFF		0x2
53 #define DEF_INT_TYPE_APIC	0x2
54 
55 /* Scalable MCA: */
56 
57 /* Threshold LVT offset is at MSR0xC0000410[15:12] */
58 #define SMCA_THR_LVT_OFF	0xF000
59 
60 static bool thresholding_irq_en;
61 
62 static const char * const th_names[] = {
63 	"load_store",
64 	"insn_fetch",
65 	"combined_unit",
66 	"decode_unit",
67 	"northbridge",
68 	"execution_unit",
69 };
70 
71 static const char * const smca_umc_block_names[] = {
72 	"dram_ecc",
73 	"misc_umc"
74 };
75 
76 struct smca_bank_name {
77 	const char *name;	/* Short name for sysfs */
78 	const char *long_name;	/* Long name for pretty-printing */
79 };
80 
81 static struct smca_bank_name smca_names[] = {
82 	[SMCA_LS]	= { "load_store",	"Load Store Unit" },
83 	[SMCA_IF]	= { "insn_fetch",	"Instruction Fetch Unit" },
84 	[SMCA_L2_CACHE]	= { "l2_cache",		"L2 Cache" },
85 	[SMCA_DE]	= { "decode_unit",	"Decode Unit" },
86 	[SMCA_RESERVED]	= { "reserved",		"Reserved" },
87 	[SMCA_EX]	= { "execution_unit",	"Execution Unit" },
88 	[SMCA_FP]	= { "floating_point",	"Floating Point Unit" },
89 	[SMCA_L3_CACHE]	= { "l3_cache",		"L3 Cache" },
90 	[SMCA_CS]	= { "coherent_slave",	"Coherent Slave" },
91 	[SMCA_PIE]	= { "pie",		"Power, Interrupts, etc." },
92 	[SMCA_UMC]	= { "umc",		"Unified Memory Controller" },
93 	[SMCA_PB]	= { "param_block",	"Parameter Block" },
94 	[SMCA_PSP]	= { "psp",		"Platform Security Processor" },
95 	[SMCA_SMU]	= { "smu",		"System Management Unit" },
96 };
97 
98 static u32 smca_bank_addrs[MAX_NR_BANKS][NR_BLOCKS] __ro_after_init =
99 {
100 	[0 ... MAX_NR_BANKS - 1] = { [0 ... NR_BLOCKS - 1] = -1 }
101 };
102 
103 static const char *smca_get_name(enum smca_bank_types t)
104 {
105 	if (t >= N_SMCA_BANK_TYPES)
106 		return NULL;
107 
108 	return smca_names[t].name;
109 }
110 
111 const char *smca_get_long_name(enum smca_bank_types t)
112 {
113 	if (t >= N_SMCA_BANK_TYPES)
114 		return NULL;
115 
116 	return smca_names[t].long_name;
117 }
118 EXPORT_SYMBOL_GPL(smca_get_long_name);
119 
120 static enum smca_bank_types smca_get_bank_type(unsigned int bank)
121 {
122 	struct smca_bank *b;
123 
124 	if (bank >= MAX_NR_BANKS)
125 		return N_SMCA_BANK_TYPES;
126 
127 	b = &smca_banks[bank];
128 	if (!b->hwid)
129 		return N_SMCA_BANK_TYPES;
130 
131 	return b->hwid->bank_type;
132 }
133 
134 static struct smca_hwid smca_hwid_mcatypes[] = {
135 	/* { bank_type, hwid_mcatype, xec_bitmap } */
136 
137 	/* Reserved type */
138 	{ SMCA_RESERVED, HWID_MCATYPE(0x00, 0x0), 0x0 },
139 
140 	/* ZN Core (HWID=0xB0) MCA types */
141 	{ SMCA_LS,	 HWID_MCATYPE(0xB0, 0x0), 0x1FFFEF },
142 	{ SMCA_IF,	 HWID_MCATYPE(0xB0, 0x1), 0x3FFF },
143 	{ SMCA_L2_CACHE, HWID_MCATYPE(0xB0, 0x2), 0xF },
144 	{ SMCA_DE,	 HWID_MCATYPE(0xB0, 0x3), 0x1FF },
145 	/* HWID 0xB0 MCATYPE 0x4 is Reserved */
146 	{ SMCA_EX,	 HWID_MCATYPE(0xB0, 0x5), 0x7FF },
147 	{ SMCA_FP,	 HWID_MCATYPE(0xB0, 0x6), 0x7F },
148 	{ SMCA_L3_CACHE, HWID_MCATYPE(0xB0, 0x7), 0xFF },
149 
150 	/* Data Fabric MCA types */
151 	{ SMCA_CS,	 HWID_MCATYPE(0x2E, 0x0), 0x1FF },
152 	{ SMCA_PIE,	 HWID_MCATYPE(0x2E, 0x1), 0xF },
153 
154 	/* Unified Memory Controller MCA type */
155 	{ SMCA_UMC,	 HWID_MCATYPE(0x96, 0x0), 0x3F },
156 
157 	/* Parameter Block MCA type */
158 	{ SMCA_PB,	 HWID_MCATYPE(0x05, 0x0), 0x1 },
159 
160 	/* Platform Security Processor MCA type */
161 	{ SMCA_PSP,	 HWID_MCATYPE(0xFF, 0x0), 0x1 },
162 
163 	/* System Management Unit MCA type */
164 	{ SMCA_SMU,	 HWID_MCATYPE(0x01, 0x0), 0x1 },
165 };
166 
167 struct smca_bank smca_banks[MAX_NR_BANKS];
168 EXPORT_SYMBOL_GPL(smca_banks);
169 
170 /*
171  * In SMCA enabled processors, we can have multiple banks for a given IP type.
172  * So to define a unique name for each bank, we use a temp c-string to append
173  * the MCA_IPID[InstanceId] to type's name in get_name().
174  *
175  * InstanceId is 32 bits which is 8 characters. Make sure MAX_MCATYPE_NAME_LEN
176  * is greater than 8 plus 1 (for underscore) plus length of longest type name.
177  */
178 #define MAX_MCATYPE_NAME_LEN	30
179 static char buf_mcatype[MAX_MCATYPE_NAME_LEN];
180 
181 static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
182 static DEFINE_PER_CPU(unsigned int, bank_map);	/* see which banks are on */
183 
184 static void amd_threshold_interrupt(void);
185 static void amd_deferred_error_interrupt(void);
186 
187 static void default_deferred_error_interrupt(void)
188 {
189 	pr_err("Unexpected deferred interrupt at vector %x\n", DEFERRED_ERROR_VECTOR);
190 }
191 void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
192 
193 static void smca_configure(unsigned int bank, unsigned int cpu)
194 {
195 	unsigned int i, hwid_mcatype;
196 	struct smca_hwid *s_hwid;
197 	u32 high, low;
198 	u32 smca_config = MSR_AMD64_SMCA_MCx_CONFIG(bank);
199 
200 	/* Set appropriate bits in MCA_CONFIG */
201 	if (!rdmsr_safe(smca_config, &low, &high)) {
202 		/*
203 		 * OS is required to set the MCAX bit to acknowledge that it is
204 		 * now using the new MSR ranges and new registers under each
205 		 * bank. It also means that the OS will configure deferred
206 		 * errors in the new MCx_CONFIG register. If the bit is not set,
207 		 * uncorrectable errors will cause a system panic.
208 		 *
209 		 * MCA_CONFIG[MCAX] is bit 32 (0 in the high portion of the MSR.)
210 		 */
211 		high |= BIT(0);
212 
213 		/*
214 		 * SMCA sets the Deferred Error Interrupt type per bank.
215 		 *
216 		 * MCA_CONFIG[DeferredIntTypeSupported] is bit 5, and tells us
217 		 * if the DeferredIntType bit field is available.
218 		 *
219 		 * MCA_CONFIG[DeferredIntType] is bits [38:37] ([6:5] in the
220 		 * high portion of the MSR). OS should set this to 0x1 to enable
221 		 * APIC based interrupt. First, check that no interrupt has been
222 		 * set.
223 		 */
224 		if ((low & BIT(5)) && !((high >> 5) & 0x3))
225 			high |= BIT(5);
226 
227 		wrmsr(smca_config, low, high);
228 	}
229 
230 	/* Return early if this bank was already initialized. */
231 	if (smca_banks[bank].hwid)
232 		return;
233 
234 	if (rdmsr_safe_on_cpu(cpu, MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
235 		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
236 		return;
237 	}
238 
239 	hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
240 				    (high & MCI_IPID_MCATYPE) >> 16);
241 
242 	for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
243 		s_hwid = &smca_hwid_mcatypes[i];
244 		if (hwid_mcatype == s_hwid->hwid_mcatype) {
245 			smca_banks[bank].hwid = s_hwid;
246 			smca_banks[bank].id = low;
247 			smca_banks[bank].sysfs_id = s_hwid->count++;
248 			break;
249 		}
250 	}
251 }
252 
253 struct thresh_restart {
254 	struct threshold_block	*b;
255 	int			reset;
256 	int			set_lvt_off;
257 	int			lvt_off;
258 	u16			old_limit;
259 };
260 
261 static inline bool is_shared_bank(int bank)
262 {
263 	/*
264 	 * Scalable MCA provides for only one core to have access to the MSRs of
265 	 * a shared bank.
266 	 */
267 	if (mce_flags.smca)
268 		return false;
269 
270 	/* Bank 4 is for northbridge reporting and is thus shared */
271 	return (bank == 4);
272 }
273 
274 static const char *bank4_names(const struct threshold_block *b)
275 {
276 	switch (b->address) {
277 	/* MSR4_MISC0 */
278 	case 0x00000413:
279 		return "dram";
280 
281 	case 0xc0000408:
282 		return "ht_links";
283 
284 	case 0xc0000409:
285 		return "l3_cache";
286 
287 	default:
288 		WARN(1, "Funny MSR: 0x%08x\n", b->address);
289 		return "";
290 	}
291 };
292 
293 
294 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
295 {
296 	/*
297 	 * bank 4 supports APIC LVT interrupts implicitly since forever.
298 	 */
299 	if (bank == 4)
300 		return true;
301 
302 	/*
303 	 * IntP: interrupt present; if this bit is set, the thresholding
304 	 * bank can generate APIC LVT interrupts
305 	 */
306 	return msr_high_bits & BIT(28);
307 }
308 
309 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
310 {
311 	int msr = (hi & MASK_LVTOFF_HI) >> 20;
312 
313 	if (apic < 0) {
314 		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
315 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
316 		       b->bank, b->block, b->address, hi, lo);
317 		return 0;
318 	}
319 
320 	if (apic != msr) {
321 		/*
322 		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
323 		 * the BIOS provides the value. The original field where LVT offset
324 		 * was set is reserved. Return early here:
325 		 */
326 		if (mce_flags.smca)
327 			return 0;
328 
329 		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
330 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
331 		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
332 		return 0;
333 	}
334 
335 	return 1;
336 };
337 
338 /* Reprogram MCx_MISC MSR behind this threshold bank. */
339 static void threshold_restart_bank(void *_tr)
340 {
341 	struct thresh_restart *tr = _tr;
342 	u32 hi, lo;
343 
344 	rdmsr(tr->b->address, lo, hi);
345 
346 	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
347 		tr->reset = 1;	/* limit cannot be lower than err count */
348 
349 	if (tr->reset) {		/* reset err count and overflow bit */
350 		hi =
351 		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
352 		    (THRESHOLD_MAX - tr->b->threshold_limit);
353 	} else if (tr->old_limit) {	/* change limit w/o reset */
354 		int new_count = (hi & THRESHOLD_MAX) +
355 		    (tr->old_limit - tr->b->threshold_limit);
356 
357 		hi = (hi & ~MASK_ERR_COUNT_HI) |
358 		    (new_count & THRESHOLD_MAX);
359 	}
360 
361 	/* clear IntType */
362 	hi &= ~MASK_INT_TYPE_HI;
363 
364 	if (!tr->b->interrupt_capable)
365 		goto done;
366 
367 	if (tr->set_lvt_off) {
368 		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
369 			/* set new lvt offset */
370 			hi &= ~MASK_LVTOFF_HI;
371 			hi |= tr->lvt_off << 20;
372 		}
373 	}
374 
375 	if (tr->b->interrupt_enable)
376 		hi |= INT_TYPE_APIC;
377 
378  done:
379 
380 	hi |= MASK_COUNT_EN_HI;
381 	wrmsr(tr->b->address, lo, hi);
382 }
383 
384 static void mce_threshold_block_init(struct threshold_block *b, int offset)
385 {
386 	struct thresh_restart tr = {
387 		.b			= b,
388 		.set_lvt_off		= 1,
389 		.lvt_off		= offset,
390 	};
391 
392 	b->threshold_limit		= THRESHOLD_MAX;
393 	threshold_restart_bank(&tr);
394 };
395 
396 static int setup_APIC_mce_threshold(int reserved, int new)
397 {
398 	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
399 					      APIC_EILVT_MSG_FIX, 0))
400 		return new;
401 
402 	return reserved;
403 }
404 
405 static int setup_APIC_deferred_error(int reserved, int new)
406 {
407 	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
408 					      APIC_EILVT_MSG_FIX, 0))
409 		return new;
410 
411 	return reserved;
412 }
413 
414 static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
415 {
416 	u32 low = 0, high = 0;
417 	int def_offset = -1, def_new;
418 
419 	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
420 		return;
421 
422 	def_new = (low & MASK_DEF_LVTOFF) >> 4;
423 	if (!(low & MASK_DEF_LVTOFF)) {
424 		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
425 		def_new = DEF_LVT_OFF;
426 		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
427 	}
428 
429 	def_offset = setup_APIC_deferred_error(def_offset, def_new);
430 	if ((def_offset == def_new) &&
431 	    (deferred_error_int_vector != amd_deferred_error_interrupt))
432 		deferred_error_int_vector = amd_deferred_error_interrupt;
433 
434 	if (!mce_flags.smca)
435 		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
436 
437 	wrmsr(MSR_CU_DEF_ERR, low, high);
438 }
439 
440 static u32 smca_get_block_address(unsigned int bank, unsigned int block)
441 {
442 	u32 low, high;
443 	u32 addr = 0;
444 
445 	if (smca_get_bank_type(bank) == SMCA_RESERVED)
446 		return addr;
447 
448 	if (!block)
449 		return MSR_AMD64_SMCA_MCx_MISC(bank);
450 
451 	/* Check our cache first: */
452 	if (smca_bank_addrs[bank][block] != -1)
453 		return smca_bank_addrs[bank][block];
454 
455 	/*
456 	 * For SMCA enabled processors, BLKPTR field of the first MISC register
457 	 * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
458 	 */
459 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
460 		goto out;
461 
462 	if (!(low & MCI_CONFIG_MCAX))
463 		goto out;
464 
465 	if (!rdmsr_safe(MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high) &&
466 	    (low & MASK_BLKPTR_LO))
467 		addr = MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
468 
469 out:
470 	smca_bank_addrs[bank][block] = addr;
471 	return addr;
472 }
473 
474 static u32 get_block_address(u32 current_addr, u32 low, u32 high,
475 			     unsigned int bank, unsigned int block)
476 {
477 	u32 addr = 0, offset = 0;
478 
479 	if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
480 		return addr;
481 
482 	if (mce_flags.smca)
483 		return smca_get_block_address(bank, block);
484 
485 	/* Fall back to method we used for older processors: */
486 	switch (block) {
487 	case 0:
488 		addr = msr_ops.misc(bank);
489 		break;
490 	case 1:
491 		offset = ((low & MASK_BLKPTR_LO) >> 21);
492 		if (offset)
493 			addr = MCG_XBLK_ADDR + offset;
494 		break;
495 	default:
496 		addr = ++current_addr;
497 	}
498 	return addr;
499 }
500 
501 static int
502 prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
503 			int offset, u32 misc_high)
504 {
505 	unsigned int cpu = smp_processor_id();
506 	u32 smca_low, smca_high;
507 	struct threshold_block b;
508 	int new;
509 
510 	if (!block)
511 		per_cpu(bank_map, cpu) |= (1 << bank);
512 
513 	memset(&b, 0, sizeof(b));
514 	b.cpu			= cpu;
515 	b.bank			= bank;
516 	b.block			= block;
517 	b.address		= addr;
518 	b.interrupt_capable	= lvt_interrupt_supported(bank, misc_high);
519 
520 	if (!b.interrupt_capable)
521 		goto done;
522 
523 	b.interrupt_enable = 1;
524 
525 	if (!mce_flags.smca) {
526 		new = (misc_high & MASK_LVTOFF_HI) >> 20;
527 		goto set_offset;
528 	}
529 
530 	/* Gather LVT offset for thresholding: */
531 	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
532 		goto out;
533 
534 	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
535 
536 set_offset:
537 	offset = setup_APIC_mce_threshold(offset, new);
538 	if (offset == new)
539 		thresholding_irq_en = true;
540 
541 done:
542 	mce_threshold_block_init(&b, offset);
543 
544 out:
545 	return offset;
546 }
547 
548 /* cpu init entry point, called from mce.c with preempt off */
549 void mce_amd_feature_init(struct cpuinfo_x86 *c)
550 {
551 	u32 low = 0, high = 0, address = 0;
552 	unsigned int bank, block, cpu = smp_processor_id();
553 	int offset = -1;
554 
555 	for (bank = 0; bank < mca_cfg.banks; ++bank) {
556 		if (mce_flags.smca)
557 			smca_configure(bank, cpu);
558 
559 		for (block = 0; block < NR_BLOCKS; ++block) {
560 			address = get_block_address(address, low, high, bank, block);
561 			if (!address)
562 				break;
563 
564 			if (rdmsr_safe(address, &low, &high))
565 				break;
566 
567 			if (!(high & MASK_VALID_HI))
568 				continue;
569 
570 			if (!(high & MASK_CNTP_HI)  ||
571 			     (high & MASK_LOCKED_HI))
572 				continue;
573 
574 			offset = prepare_threshold_block(bank, block, address, offset, high);
575 		}
576 	}
577 
578 	if (mce_flags.succor)
579 		deferred_error_interrupt_enable(c);
580 }
581 
582 int umc_normaddr_to_sysaddr(u64 norm_addr, u16 nid, u8 umc, u64 *sys_addr)
583 {
584 	u64 dram_base_addr, dram_limit_addr, dram_hole_base;
585 	/* We start from the normalized address */
586 	u64 ret_addr = norm_addr;
587 
588 	u32 tmp;
589 
590 	u8 die_id_shift, die_id_mask, socket_id_shift, socket_id_mask;
591 	u8 intlv_num_dies, intlv_num_chan, intlv_num_sockets;
592 	u8 intlv_addr_sel, intlv_addr_bit;
593 	u8 num_intlv_bits, hashed_bit;
594 	u8 lgcy_mmio_hole_en, base = 0;
595 	u8 cs_mask, cs_id = 0;
596 	bool hash_enabled = false;
597 
598 	/* Read D18F0x1B4 (DramOffset), check if base 1 is used. */
599 	if (amd_df_indirect_read(nid, 0, 0x1B4, umc, &tmp))
600 		goto out_err;
601 
602 	/* Remove HiAddrOffset from normalized address, if enabled: */
603 	if (tmp & BIT(0)) {
604 		u64 hi_addr_offset = (tmp & GENMASK_ULL(31, 20)) << 8;
605 
606 		if (norm_addr >= hi_addr_offset) {
607 			ret_addr -= hi_addr_offset;
608 			base = 1;
609 		}
610 	}
611 
612 	/* Read D18F0x110 (DramBaseAddress). */
613 	if (amd_df_indirect_read(nid, 0, 0x110 + (8 * base), umc, &tmp))
614 		goto out_err;
615 
616 	/* Check if address range is valid. */
617 	if (!(tmp & BIT(0))) {
618 		pr_err("%s: Invalid DramBaseAddress range: 0x%x.\n",
619 			__func__, tmp);
620 		goto out_err;
621 	}
622 
623 	lgcy_mmio_hole_en = tmp & BIT(1);
624 	intlv_num_chan	  = (tmp >> 4) & 0xF;
625 	intlv_addr_sel	  = (tmp >> 8) & 0x7;
626 	dram_base_addr	  = (tmp & GENMASK_ULL(31, 12)) << 16;
627 
628 	/* {0, 1, 2, 3} map to address bits {8, 9, 10, 11} respectively */
629 	if (intlv_addr_sel > 3) {
630 		pr_err("%s: Invalid interleave address select %d.\n",
631 			__func__, intlv_addr_sel);
632 		goto out_err;
633 	}
634 
635 	/* Read D18F0x114 (DramLimitAddress). */
636 	if (amd_df_indirect_read(nid, 0, 0x114 + (8 * base), umc, &tmp))
637 		goto out_err;
638 
639 	intlv_num_sockets = (tmp >> 8) & 0x1;
640 	intlv_num_dies	  = (tmp >> 10) & 0x3;
641 	dram_limit_addr	  = ((tmp & GENMASK_ULL(31, 12)) << 16) | GENMASK_ULL(27, 0);
642 
643 	intlv_addr_bit = intlv_addr_sel + 8;
644 
645 	/* Re-use intlv_num_chan by setting it equal to log2(#channels) */
646 	switch (intlv_num_chan) {
647 	case 0:	intlv_num_chan = 0; break;
648 	case 1: intlv_num_chan = 1; break;
649 	case 3: intlv_num_chan = 2; break;
650 	case 5:	intlv_num_chan = 3; break;
651 	case 7:	intlv_num_chan = 4; break;
652 
653 	case 8: intlv_num_chan = 1;
654 		hash_enabled = true;
655 		break;
656 	default:
657 		pr_err("%s: Invalid number of interleaved channels %d.\n",
658 			__func__, intlv_num_chan);
659 		goto out_err;
660 	}
661 
662 	num_intlv_bits = intlv_num_chan;
663 
664 	if (intlv_num_dies > 2) {
665 		pr_err("%s: Invalid number of interleaved nodes/dies %d.\n",
666 			__func__, intlv_num_dies);
667 		goto out_err;
668 	}
669 
670 	num_intlv_bits += intlv_num_dies;
671 
672 	/* Add a bit if sockets are interleaved. */
673 	num_intlv_bits += intlv_num_sockets;
674 
675 	/* Assert num_intlv_bits <= 4 */
676 	if (num_intlv_bits > 4) {
677 		pr_err("%s: Invalid interleave bits %d.\n",
678 			__func__, num_intlv_bits);
679 		goto out_err;
680 	}
681 
682 	if (num_intlv_bits > 0) {
683 		u64 temp_addr_x, temp_addr_i, temp_addr_y;
684 		u8 die_id_bit, sock_id_bit, cs_fabric_id;
685 
686 		/*
687 		 * Read FabricBlockInstanceInformation3_CS[BlockFabricID].
688 		 * This is the fabric id for this coherent slave. Use
689 		 * umc/channel# as instance id of the coherent slave
690 		 * for FICAA.
691 		 */
692 		if (amd_df_indirect_read(nid, 0, 0x50, umc, &tmp))
693 			goto out_err;
694 
695 		cs_fabric_id = (tmp >> 8) & 0xFF;
696 		die_id_bit   = 0;
697 
698 		/* If interleaved over more than 1 channel: */
699 		if (intlv_num_chan) {
700 			die_id_bit = intlv_num_chan;
701 			cs_mask	   = (1 << die_id_bit) - 1;
702 			cs_id	   = cs_fabric_id & cs_mask;
703 		}
704 
705 		sock_id_bit = die_id_bit;
706 
707 		/* Read D18F1x208 (SystemFabricIdMask). */
708 		if (intlv_num_dies || intlv_num_sockets)
709 			if (amd_df_indirect_read(nid, 1, 0x208, umc, &tmp))
710 				goto out_err;
711 
712 		/* If interleaved over more than 1 die. */
713 		if (intlv_num_dies) {
714 			sock_id_bit  = die_id_bit + intlv_num_dies;
715 			die_id_shift = (tmp >> 24) & 0xF;
716 			die_id_mask  = (tmp >> 8) & 0xFF;
717 
718 			cs_id |= ((cs_fabric_id & die_id_mask) >> die_id_shift) << die_id_bit;
719 		}
720 
721 		/* If interleaved over more than 1 socket. */
722 		if (intlv_num_sockets) {
723 			socket_id_shift	= (tmp >> 28) & 0xF;
724 			socket_id_mask	= (tmp >> 16) & 0xFF;
725 
726 			cs_id |= ((cs_fabric_id & socket_id_mask) >> socket_id_shift) << sock_id_bit;
727 		}
728 
729 		/*
730 		 * The pre-interleaved address consists of XXXXXXIIIYYYYY
731 		 * where III is the ID for this CS, and XXXXXXYYYYY are the
732 		 * address bits from the post-interleaved address.
733 		 * "num_intlv_bits" has been calculated to tell us how many "I"
734 		 * bits there are. "intlv_addr_bit" tells us how many "Y" bits
735 		 * there are (where "I" starts).
736 		 */
737 		temp_addr_y = ret_addr & GENMASK_ULL(intlv_addr_bit-1, 0);
738 		temp_addr_i = (cs_id << intlv_addr_bit);
739 		temp_addr_x = (ret_addr & GENMASK_ULL(63, intlv_addr_bit)) << num_intlv_bits;
740 		ret_addr    = temp_addr_x | temp_addr_i | temp_addr_y;
741 	}
742 
743 	/* Add dram base address */
744 	ret_addr += dram_base_addr;
745 
746 	/* If legacy MMIO hole enabled */
747 	if (lgcy_mmio_hole_en) {
748 		if (amd_df_indirect_read(nid, 0, 0x104, umc, &tmp))
749 			goto out_err;
750 
751 		dram_hole_base = tmp & GENMASK(31, 24);
752 		if (ret_addr >= dram_hole_base)
753 			ret_addr += (BIT_ULL(32) - dram_hole_base);
754 	}
755 
756 	if (hash_enabled) {
757 		/* Save some parentheses and grab ls-bit at the end. */
758 		hashed_bit =	(ret_addr >> 12) ^
759 				(ret_addr >> 18) ^
760 				(ret_addr >> 21) ^
761 				(ret_addr >> 30) ^
762 				cs_id;
763 
764 		hashed_bit &= BIT(0);
765 
766 		if (hashed_bit != ((ret_addr >> intlv_addr_bit) & BIT(0)))
767 			ret_addr ^= BIT(intlv_addr_bit);
768 	}
769 
770 	/* Is calculated system address is above DRAM limit address? */
771 	if (ret_addr > dram_limit_addr)
772 		goto out_err;
773 
774 	*sys_addr = ret_addr;
775 	return 0;
776 
777 out_err:
778 	return -EINVAL;
779 }
780 EXPORT_SYMBOL_GPL(umc_normaddr_to_sysaddr);
781 
782 bool amd_mce_is_memory_error(struct mce *m)
783 {
784 	/* ErrCodeExt[20:16] */
785 	u8 xec = (m->status >> 16) & 0x1f;
786 
787 	if (mce_flags.smca)
788 		return smca_get_bank_type(m->bank) == SMCA_UMC && xec == 0x0;
789 
790 	return m->bank == 4 && xec == 0x8;
791 }
792 
793 static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
794 {
795 	struct mce m;
796 
797 	mce_setup(&m);
798 
799 	m.status = status;
800 	m.misc   = misc;
801 	m.bank   = bank;
802 	m.tsc	 = rdtsc();
803 
804 	if (m.status & MCI_STATUS_ADDRV) {
805 		m.addr = addr;
806 
807 		/*
808 		 * Extract [55:<lsb>] where lsb is the least significant
809 		 * *valid* bit of the address bits.
810 		 */
811 		if (mce_flags.smca) {
812 			u8 lsb = (m.addr >> 56) & 0x3f;
813 
814 			m.addr &= GENMASK_ULL(55, lsb);
815 		}
816 	}
817 
818 	if (mce_flags.smca) {
819 		rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
820 
821 		if (m.status & MCI_STATUS_SYNDV)
822 			rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
823 	}
824 
825 	mce_log(&m);
826 }
827 
828 asmlinkage __visible void __irq_entry smp_deferred_error_interrupt(struct pt_regs *regs)
829 {
830 	entering_irq();
831 	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
832 	inc_irq_stat(irq_deferred_error_count);
833 	deferred_error_int_vector();
834 	trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
835 	exiting_ack_irq();
836 }
837 
838 /*
839  * Returns true if the logged error is deferred. False, otherwise.
840  */
841 static inline bool
842 _log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
843 {
844 	u64 status, addr = 0;
845 
846 	rdmsrl(msr_stat, status);
847 	if (!(status & MCI_STATUS_VAL))
848 		return false;
849 
850 	if (status & MCI_STATUS_ADDRV)
851 		rdmsrl(msr_addr, addr);
852 
853 	__log_error(bank, status, addr, misc);
854 
855 	wrmsrl(msr_stat, 0);
856 
857 	return status & MCI_STATUS_DEFERRED;
858 }
859 
860 /*
861  * We have three scenarios for checking for Deferred errors:
862  *
863  * 1) Non-SMCA systems check MCA_STATUS and log error if found.
864  * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
865  *    clear MCA_DESTAT.
866  * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
867  *    log it.
868  */
869 static void log_error_deferred(unsigned int bank)
870 {
871 	bool defrd;
872 
873 	defrd = _log_error_bank(bank, msr_ops.status(bank),
874 					msr_ops.addr(bank), 0);
875 
876 	if (!mce_flags.smca)
877 		return;
878 
879 	/* Clear MCA_DESTAT if we logged the deferred error from MCA_STATUS. */
880 	if (defrd) {
881 		wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
882 		return;
883 	}
884 
885 	/*
886 	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
887 	 * for a valid error.
888 	 */
889 	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
890 			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
891 }
892 
893 /* APIC interrupt handler for deferred errors */
894 static void amd_deferred_error_interrupt(void)
895 {
896 	unsigned int bank;
897 
898 	for (bank = 0; bank < mca_cfg.banks; ++bank)
899 		log_error_deferred(bank);
900 }
901 
902 static void log_error_thresholding(unsigned int bank, u64 misc)
903 {
904 	_log_error_bank(bank, msr_ops.status(bank), msr_ops.addr(bank), misc);
905 }
906 
907 static void log_and_reset_block(struct threshold_block *block)
908 {
909 	struct thresh_restart tr;
910 	u32 low = 0, high = 0;
911 
912 	if (!block)
913 		return;
914 
915 	if (rdmsr_safe(block->address, &low, &high))
916 		return;
917 
918 	if (!(high & MASK_OVERFLOW_HI))
919 		return;
920 
921 	/* Log the MCE which caused the threshold event. */
922 	log_error_thresholding(block->bank, ((u64)high << 32) | low);
923 
924 	/* Reset threshold block after logging error. */
925 	memset(&tr, 0, sizeof(tr));
926 	tr.b = block;
927 	threshold_restart_bank(&tr);
928 }
929 
930 /*
931  * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
932  * goes off when error_count reaches threshold_limit.
933  */
934 static void amd_threshold_interrupt(void)
935 {
936 	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
937 	unsigned int bank, cpu = smp_processor_id();
938 
939 	for (bank = 0; bank < mca_cfg.banks; ++bank) {
940 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
941 			continue;
942 
943 		first_block = per_cpu(threshold_banks, cpu)[bank]->blocks;
944 		if (!first_block)
945 			continue;
946 
947 		/*
948 		 * The first block is also the head of the list. Check it first
949 		 * before iterating over the rest.
950 		 */
951 		log_and_reset_block(first_block);
952 		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
953 			log_and_reset_block(block);
954 	}
955 }
956 
957 /*
958  * Sysfs Interface
959  */
960 
961 struct threshold_attr {
962 	struct attribute attr;
963 	ssize_t (*show) (struct threshold_block *, char *);
964 	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
965 };
966 
967 #define SHOW_FIELDS(name)						\
968 static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
969 {									\
970 	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
971 }
972 SHOW_FIELDS(interrupt_enable)
973 SHOW_FIELDS(threshold_limit)
974 
975 static ssize_t
976 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
977 {
978 	struct thresh_restart tr;
979 	unsigned long new;
980 
981 	if (!b->interrupt_capable)
982 		return -EINVAL;
983 
984 	if (kstrtoul(buf, 0, &new) < 0)
985 		return -EINVAL;
986 
987 	b->interrupt_enable = !!new;
988 
989 	memset(&tr, 0, sizeof(tr));
990 	tr.b		= b;
991 
992 	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
993 
994 	return size;
995 }
996 
997 static ssize_t
998 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
999 {
1000 	struct thresh_restart tr;
1001 	unsigned long new;
1002 
1003 	if (kstrtoul(buf, 0, &new) < 0)
1004 		return -EINVAL;
1005 
1006 	if (new > THRESHOLD_MAX)
1007 		new = THRESHOLD_MAX;
1008 	if (new < 1)
1009 		new = 1;
1010 
1011 	memset(&tr, 0, sizeof(tr));
1012 	tr.old_limit = b->threshold_limit;
1013 	b->threshold_limit = new;
1014 	tr.b = b;
1015 
1016 	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
1017 
1018 	return size;
1019 }
1020 
1021 static ssize_t show_error_count(struct threshold_block *b, char *buf)
1022 {
1023 	u32 lo, hi;
1024 
1025 	rdmsr_on_cpu(b->cpu, b->address, &lo, &hi);
1026 
1027 	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
1028 				     (THRESHOLD_MAX - b->threshold_limit)));
1029 }
1030 
1031 static struct threshold_attr error_count = {
1032 	.attr = {.name = __stringify(error_count), .mode = 0444 },
1033 	.show = show_error_count,
1034 };
1035 
1036 #define RW_ATTR(val)							\
1037 static struct threshold_attr val = {					\
1038 	.attr	= {.name = __stringify(val), .mode = 0644 },		\
1039 	.show	= show_## val,						\
1040 	.store	= store_## val,						\
1041 };
1042 
1043 RW_ATTR(interrupt_enable);
1044 RW_ATTR(threshold_limit);
1045 
1046 static struct attribute *default_attrs[] = {
1047 	&threshold_limit.attr,
1048 	&error_count.attr,
1049 	NULL,	/* possibly interrupt_enable if supported, see below */
1050 	NULL,
1051 };
1052 
1053 #define to_block(k)	container_of(k, struct threshold_block, kobj)
1054 #define to_attr(a)	container_of(a, struct threshold_attr, attr)
1055 
1056 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1057 {
1058 	struct threshold_block *b = to_block(kobj);
1059 	struct threshold_attr *a = to_attr(attr);
1060 	ssize_t ret;
1061 
1062 	ret = a->show ? a->show(b, buf) : -EIO;
1063 
1064 	return ret;
1065 }
1066 
1067 static ssize_t store(struct kobject *kobj, struct attribute *attr,
1068 		     const char *buf, size_t count)
1069 {
1070 	struct threshold_block *b = to_block(kobj);
1071 	struct threshold_attr *a = to_attr(attr);
1072 	ssize_t ret;
1073 
1074 	ret = a->store ? a->store(b, buf, count) : -EIO;
1075 
1076 	return ret;
1077 }
1078 
1079 static const struct sysfs_ops threshold_ops = {
1080 	.show			= show,
1081 	.store			= store,
1082 };
1083 
1084 static struct kobj_type threshold_ktype = {
1085 	.sysfs_ops		= &threshold_ops,
1086 	.default_attrs		= default_attrs,
1087 };
1088 
1089 static const char *get_name(unsigned int bank, struct threshold_block *b)
1090 {
1091 	enum smca_bank_types bank_type;
1092 
1093 	if (!mce_flags.smca) {
1094 		if (b && bank == 4)
1095 			return bank4_names(b);
1096 
1097 		return th_names[bank];
1098 	}
1099 
1100 	bank_type = smca_get_bank_type(bank);
1101 	if (bank_type >= N_SMCA_BANK_TYPES)
1102 		return NULL;
1103 
1104 	if (b && bank_type == SMCA_UMC) {
1105 		if (b->block < ARRAY_SIZE(smca_umc_block_names))
1106 			return smca_umc_block_names[b->block];
1107 		return NULL;
1108 	}
1109 
1110 	if (smca_banks[bank].hwid->count == 1)
1111 		return smca_get_name(bank_type);
1112 
1113 	snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1114 		 "%s_%x", smca_get_name(bank_type),
1115 			  smca_banks[bank].sysfs_id);
1116 	return buf_mcatype;
1117 }
1118 
1119 static int allocate_threshold_blocks(unsigned int cpu, unsigned int bank,
1120 				     unsigned int block, u32 address)
1121 {
1122 	struct threshold_block *b = NULL;
1123 	u32 low, high;
1124 	int err;
1125 
1126 	if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
1127 		return 0;
1128 
1129 	if (rdmsr_safe_on_cpu(cpu, address, &low, &high))
1130 		return 0;
1131 
1132 	if (!(high & MASK_VALID_HI)) {
1133 		if (block)
1134 			goto recurse;
1135 		else
1136 			return 0;
1137 	}
1138 
1139 	if (!(high & MASK_CNTP_HI)  ||
1140 	     (high & MASK_LOCKED_HI))
1141 		goto recurse;
1142 
1143 	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1144 	if (!b)
1145 		return -ENOMEM;
1146 
1147 	b->block		= block;
1148 	b->bank			= bank;
1149 	b->cpu			= cpu;
1150 	b->address		= address;
1151 	b->interrupt_enable	= 0;
1152 	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
1153 	b->threshold_limit	= THRESHOLD_MAX;
1154 
1155 	if (b->interrupt_capable) {
1156 		threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
1157 		b->interrupt_enable = 1;
1158 	} else {
1159 		threshold_ktype.default_attrs[2] = NULL;
1160 	}
1161 
1162 	INIT_LIST_HEAD(&b->miscj);
1163 
1164 	if (per_cpu(threshold_banks, cpu)[bank]->blocks) {
1165 		list_add(&b->miscj,
1166 			 &per_cpu(threshold_banks, cpu)[bank]->blocks->miscj);
1167 	} else {
1168 		per_cpu(threshold_banks, cpu)[bank]->blocks = b;
1169 	}
1170 
1171 	err = kobject_init_and_add(&b->kobj, &threshold_ktype,
1172 				   per_cpu(threshold_banks, cpu)[bank]->kobj,
1173 				   get_name(bank, b));
1174 	if (err)
1175 		goto out_free;
1176 recurse:
1177 	address = get_block_address(address, low, high, bank, ++block);
1178 	if (!address)
1179 		return 0;
1180 
1181 	err = allocate_threshold_blocks(cpu, bank, block, address);
1182 	if (err)
1183 		goto out_free;
1184 
1185 	if (b)
1186 		kobject_uevent(&b->kobj, KOBJ_ADD);
1187 
1188 	return err;
1189 
1190 out_free:
1191 	if (b) {
1192 		kobject_put(&b->kobj);
1193 		list_del(&b->miscj);
1194 		kfree(b);
1195 	}
1196 	return err;
1197 }
1198 
1199 static int __threshold_add_blocks(struct threshold_bank *b)
1200 {
1201 	struct list_head *head = &b->blocks->miscj;
1202 	struct threshold_block *pos = NULL;
1203 	struct threshold_block *tmp = NULL;
1204 	int err = 0;
1205 
1206 	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1207 	if (err)
1208 		return err;
1209 
1210 	list_for_each_entry_safe(pos, tmp, head, miscj) {
1211 
1212 		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1213 		if (err) {
1214 			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1215 				kobject_del(&pos->kobj);
1216 
1217 			return err;
1218 		}
1219 	}
1220 	return err;
1221 }
1222 
1223 static int threshold_create_bank(unsigned int cpu, unsigned int bank)
1224 {
1225 	struct device *dev = per_cpu(mce_device, cpu);
1226 	struct amd_northbridge *nb = NULL;
1227 	struct threshold_bank *b = NULL;
1228 	const char *name = get_name(bank, NULL);
1229 	int err = 0;
1230 
1231 	if (!dev)
1232 		return -ENODEV;
1233 
1234 	if (is_shared_bank(bank)) {
1235 		nb = node_to_amd_nb(amd_get_nb_id(cpu));
1236 
1237 		/* threshold descriptor already initialized on this node? */
1238 		if (nb && nb->bank4) {
1239 			/* yes, use it */
1240 			b = nb->bank4;
1241 			err = kobject_add(b->kobj, &dev->kobj, name);
1242 			if (err)
1243 				goto out;
1244 
1245 			per_cpu(threshold_banks, cpu)[bank] = b;
1246 			refcount_inc(&b->cpus);
1247 
1248 			err = __threshold_add_blocks(b);
1249 
1250 			goto out;
1251 		}
1252 	}
1253 
1254 	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1255 	if (!b) {
1256 		err = -ENOMEM;
1257 		goto out;
1258 	}
1259 
1260 	b->kobj = kobject_create_and_add(name, &dev->kobj);
1261 	if (!b->kobj) {
1262 		err = -EINVAL;
1263 		goto out_free;
1264 	}
1265 
1266 	per_cpu(threshold_banks, cpu)[bank] = b;
1267 
1268 	if (is_shared_bank(bank)) {
1269 		refcount_set(&b->cpus, 1);
1270 
1271 		/* nb is already initialized, see above */
1272 		if (nb) {
1273 			WARN_ON(nb->bank4);
1274 			nb->bank4 = b;
1275 		}
1276 	}
1277 
1278 	err = allocate_threshold_blocks(cpu, bank, 0, msr_ops.misc(bank));
1279 	if (!err)
1280 		goto out;
1281 
1282  out_free:
1283 	kfree(b);
1284 
1285  out:
1286 	return err;
1287 }
1288 
1289 static void deallocate_threshold_block(unsigned int cpu,
1290 						 unsigned int bank)
1291 {
1292 	struct threshold_block *pos = NULL;
1293 	struct threshold_block *tmp = NULL;
1294 	struct threshold_bank *head = per_cpu(threshold_banks, cpu)[bank];
1295 
1296 	if (!head)
1297 		return;
1298 
1299 	list_for_each_entry_safe(pos, tmp, &head->blocks->miscj, miscj) {
1300 		kobject_put(&pos->kobj);
1301 		list_del(&pos->miscj);
1302 		kfree(pos);
1303 	}
1304 
1305 	kfree(per_cpu(threshold_banks, cpu)[bank]->blocks);
1306 	per_cpu(threshold_banks, cpu)[bank]->blocks = NULL;
1307 }
1308 
1309 static void __threshold_remove_blocks(struct threshold_bank *b)
1310 {
1311 	struct threshold_block *pos = NULL;
1312 	struct threshold_block *tmp = NULL;
1313 
1314 	kobject_del(b->kobj);
1315 
1316 	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1317 		kobject_del(&pos->kobj);
1318 }
1319 
1320 static void threshold_remove_bank(unsigned int cpu, int bank)
1321 {
1322 	struct amd_northbridge *nb;
1323 	struct threshold_bank *b;
1324 
1325 	b = per_cpu(threshold_banks, cpu)[bank];
1326 	if (!b)
1327 		return;
1328 
1329 	if (!b->blocks)
1330 		goto free_out;
1331 
1332 	if (is_shared_bank(bank)) {
1333 		if (!refcount_dec_and_test(&b->cpus)) {
1334 			__threshold_remove_blocks(b);
1335 			per_cpu(threshold_banks, cpu)[bank] = NULL;
1336 			return;
1337 		} else {
1338 			/*
1339 			 * the last CPU on this node using the shared bank is
1340 			 * going away, remove that bank now.
1341 			 */
1342 			nb = node_to_amd_nb(amd_get_nb_id(cpu));
1343 			nb->bank4 = NULL;
1344 		}
1345 	}
1346 
1347 	deallocate_threshold_block(cpu, bank);
1348 
1349 free_out:
1350 	kobject_del(b->kobj);
1351 	kobject_put(b->kobj);
1352 	kfree(b);
1353 	per_cpu(threshold_banks, cpu)[bank] = NULL;
1354 }
1355 
1356 int mce_threshold_remove_device(unsigned int cpu)
1357 {
1358 	unsigned int bank;
1359 
1360 	for (bank = 0; bank < mca_cfg.banks; ++bank) {
1361 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1362 			continue;
1363 		threshold_remove_bank(cpu, bank);
1364 	}
1365 	kfree(per_cpu(threshold_banks, cpu));
1366 	per_cpu(threshold_banks, cpu) = NULL;
1367 	return 0;
1368 }
1369 
1370 /* create dir/files for all valid threshold banks */
1371 int mce_threshold_create_device(unsigned int cpu)
1372 {
1373 	unsigned int bank;
1374 	struct threshold_bank **bp;
1375 	int err = 0;
1376 
1377 	bp = per_cpu(threshold_banks, cpu);
1378 	if (bp)
1379 		return 0;
1380 
1381 	bp = kcalloc(mca_cfg.banks, sizeof(struct threshold_bank *),
1382 		     GFP_KERNEL);
1383 	if (!bp)
1384 		return -ENOMEM;
1385 
1386 	per_cpu(threshold_banks, cpu) = bp;
1387 
1388 	for (bank = 0; bank < mca_cfg.banks; ++bank) {
1389 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1390 			continue;
1391 		err = threshold_create_bank(cpu, bank);
1392 		if (err)
1393 			goto err;
1394 	}
1395 	return err;
1396 err:
1397 	mce_threshold_remove_device(cpu);
1398 	return err;
1399 }
1400 
1401 static __init int threshold_init_device(void)
1402 {
1403 	unsigned lcpu = 0;
1404 
1405 	/* to hit CPUs online before the notifier is up */
1406 	for_each_online_cpu(lcpu) {
1407 		int err = mce_threshold_create_device(lcpu);
1408 
1409 		if (err)
1410 			return err;
1411 	}
1412 
1413 	if (thresholding_irq_en)
1414 		mce_threshold_vector = amd_threshold_interrupt;
1415 
1416 	return 0;
1417 }
1418 /*
1419  * there are 3 funcs which need to be _initcalled in a logic sequence:
1420  * 1. xen_late_init_mcelog
1421  * 2. mcheck_init_device
1422  * 3. threshold_init_device
1423  *
1424  * xen_late_init_mcelog must register xen_mce_chrdev_device before
1425  * native mce_chrdev_device registration if running under xen platform;
1426  *
1427  * mcheck_init_device should be inited before threshold_init_device to
1428  * initialize mce_device, otherwise a NULL ptr dereference will cause panic.
1429  *
1430  * so we use following _initcalls
1431  * 1. device_initcall(xen_late_init_mcelog);
1432  * 2. device_initcall_sync(mcheck_init_device);
1433  * 3. late_initcall(threshold_init_device);
1434  *
1435  * when running under xen, the initcall order is 1,2,3;
1436  * on baremetal, we skip 1 and we do only 2 and 3.
1437  */
1438 late_initcall(threshold_init_device);
1439