xref: /linux/arch/x86/kernel/cpu/mce/amd.c (revision e724e7aaf9ca794670a4d4931af7a7e24e37fec3)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  (c) 2005-2016 Advanced Micro Devices, Inc.
4  *
5  *  Written by Jacob Shin - AMD, Inc.
6  *  Maintained by: Borislav Petkov <bp@alien8.de>
7  *
8  *  All MC4_MISCi registers are shared between cores on a node.
9  */
10 #include <linux/interrupt.h>
11 #include <linux/notifier.h>
12 #include <linux/kobject.h>
13 #include <linux/percpu.h>
14 #include <linux/errno.h>
15 #include <linux/sched.h>
16 #include <linux/sysfs.h>
17 #include <linux/slab.h>
18 #include <linux/init.h>
19 #include <linux/cpu.h>
20 #include <linux/smp.h>
21 #include <linux/string.h>
22 
23 #include <asm/amd_nb.h>
24 #include <asm/traps.h>
25 #include <asm/apic.h>
26 #include <asm/mce.h>
27 #include <asm/msr.h>
28 #include <asm/trace/irq_vectors.h>
29 
30 #include "internal.h"
31 
32 #define NR_BLOCKS         5
33 #define THRESHOLD_MAX     0xFFF
34 #define INT_TYPE_APIC     0x00020000
35 #define MASK_VALID_HI     0x80000000
36 #define MASK_CNTP_HI      0x40000000
37 #define MASK_LOCKED_HI    0x20000000
38 #define MASK_LVTOFF_HI    0x00F00000
39 #define MASK_COUNT_EN_HI  0x00080000
40 #define MASK_INT_TYPE_HI  0x00060000
41 #define MASK_OVERFLOW_HI  0x00010000
42 #define MASK_ERR_COUNT_HI 0x00000FFF
43 #define MASK_BLKPTR_LO    0xFF000000
44 #define MCG_XBLK_ADDR     0xC0000400
45 
46 /* Deferred error settings */
47 #define MSR_CU_DEF_ERR		0xC0000410
48 #define MASK_DEF_LVTOFF		0x000000F0
49 #define MASK_DEF_INT_TYPE	0x00000006
50 #define DEF_LVT_OFF		0x2
51 #define DEF_INT_TYPE_APIC	0x2
52 
53 /* Scalable MCA: */
54 
55 /* Threshold LVT offset is at MSR0xC0000410[15:12] */
56 #define SMCA_THR_LVT_OFF	0xF000
57 
58 static bool thresholding_irq_en;
59 
60 static const char * const th_names[] = {
61 	"load_store",
62 	"insn_fetch",
63 	"combined_unit",
64 	"decode_unit",
65 	"northbridge",
66 	"execution_unit",
67 };
68 
69 static const char * const smca_umc_block_names[] = {
70 	"dram_ecc",
71 	"misc_umc"
72 };
73 
74 #define HWID_MCATYPE(hwid, mcatype) (((hwid) << 16) | (mcatype))
75 
76 struct smca_hwid {
77 	unsigned int bank_type;	/* Use with smca_bank_types for easy indexing. */
78 	u32 hwid_mcatype;	/* (hwid,mcatype) tuple */
79 };
80 
81 struct smca_bank {
82 	const struct smca_hwid *hwid;
83 	u32 id;			/* Value of MCA_IPID[InstanceId]. */
84 	u8 sysfs_id;		/* Value used for sysfs name. */
85 };
86 
87 static DEFINE_PER_CPU_READ_MOSTLY(struct smca_bank[MAX_NR_BANKS], smca_banks);
88 static DEFINE_PER_CPU_READ_MOSTLY(u8[N_SMCA_BANK_TYPES], smca_bank_counts);
89 
90 struct smca_bank_name {
91 	const char *name;	/* Short name for sysfs */
92 	const char *long_name;	/* Long name for pretty-printing */
93 };
94 
95 static struct smca_bank_name smca_names[] = {
96 	[SMCA_LS ... SMCA_LS_V2]	= { "load_store",	"Load Store Unit" },
97 	[SMCA_IF]			= { "insn_fetch",	"Instruction Fetch Unit" },
98 	[SMCA_L2_CACHE]			= { "l2_cache",		"L2 Cache" },
99 	[SMCA_DE]			= { "decode_unit",	"Decode Unit" },
100 	[SMCA_RESERVED]			= { "reserved",		"Reserved" },
101 	[SMCA_EX]			= { "execution_unit",	"Execution Unit" },
102 	[SMCA_FP]			= { "floating_point",	"Floating Point Unit" },
103 	[SMCA_L3_CACHE]			= { "l3_cache",		"L3 Cache" },
104 	[SMCA_CS ... SMCA_CS_V2]	= { "coherent_slave",	"Coherent Slave" },
105 	[SMCA_PIE]			= { "pie",		"Power, Interrupts, etc." },
106 
107 	/* UMC v2 is separate because both of them can exist in a single system. */
108 	[SMCA_UMC]			= { "umc",		"Unified Memory Controller" },
109 	[SMCA_UMC_V2]			= { "umc_v2",		"Unified Memory Controller v2" },
110 	[SMCA_PB]			= { "param_block",	"Parameter Block" },
111 	[SMCA_PSP ... SMCA_PSP_V2]	= { "psp",		"Platform Security Processor" },
112 	[SMCA_SMU ... SMCA_SMU_V2]	= { "smu",		"System Management Unit" },
113 	[SMCA_MP5]			= { "mp5",		"Microprocessor 5 Unit" },
114 	[SMCA_MPDMA]			= { "mpdma",		"MPDMA Unit" },
115 	[SMCA_NBIO]			= { "nbio",		"Northbridge IO Unit" },
116 	[SMCA_PCIE ... SMCA_PCIE_V2]	= { "pcie",		"PCI Express Unit" },
117 	[SMCA_XGMI_PCS]			= { "xgmi_pcs",		"Ext Global Memory Interconnect PCS Unit" },
118 	[SMCA_NBIF]			= { "nbif",		"NBIF Unit" },
119 	[SMCA_SHUB]			= { "shub",		"System Hub Unit" },
120 	[SMCA_SATA]			= { "sata",		"SATA Unit" },
121 	[SMCA_USB]			= { "usb",		"USB Unit" },
122 	[SMCA_GMI_PCS]			= { "gmi_pcs",		"Global Memory Interconnect PCS Unit" },
123 	[SMCA_XGMI_PHY]			= { "xgmi_phy",		"Ext Global Memory Interconnect PHY Unit" },
124 	[SMCA_WAFL_PHY]			= { "wafl_phy",		"WAFL PHY Unit" },
125 	[SMCA_GMI_PHY]			= { "gmi_phy",		"Global Memory Interconnect PHY Unit" },
126 };
127 
128 static const char *smca_get_name(enum smca_bank_types t)
129 {
130 	if (t >= N_SMCA_BANK_TYPES)
131 		return NULL;
132 
133 	return smca_names[t].name;
134 }
135 
136 const char *smca_get_long_name(enum smca_bank_types t)
137 {
138 	if (t >= N_SMCA_BANK_TYPES)
139 		return NULL;
140 
141 	return smca_names[t].long_name;
142 }
143 EXPORT_SYMBOL_GPL(smca_get_long_name);
144 
145 enum smca_bank_types smca_get_bank_type(unsigned int cpu, unsigned int bank)
146 {
147 	struct smca_bank *b;
148 
149 	if (bank >= MAX_NR_BANKS)
150 		return N_SMCA_BANK_TYPES;
151 
152 	b = &per_cpu(smca_banks, cpu)[bank];
153 	if (!b->hwid)
154 		return N_SMCA_BANK_TYPES;
155 
156 	return b->hwid->bank_type;
157 }
158 EXPORT_SYMBOL_GPL(smca_get_bank_type);
159 
160 static const struct smca_hwid smca_hwid_mcatypes[] = {
161 	/* { bank_type, hwid_mcatype } */
162 
163 	/* Reserved type */
164 	{ SMCA_RESERVED, HWID_MCATYPE(0x00, 0x0)	},
165 
166 	/* ZN Core (HWID=0xB0) MCA types */
167 	{ SMCA_LS,	 HWID_MCATYPE(0xB0, 0x0)	},
168 	{ SMCA_LS_V2,	 HWID_MCATYPE(0xB0, 0x10)	},
169 	{ SMCA_IF,	 HWID_MCATYPE(0xB0, 0x1)	},
170 	{ SMCA_L2_CACHE, HWID_MCATYPE(0xB0, 0x2)	},
171 	{ SMCA_DE,	 HWID_MCATYPE(0xB0, 0x3)	},
172 	/* HWID 0xB0 MCATYPE 0x4 is Reserved */
173 	{ SMCA_EX,	 HWID_MCATYPE(0xB0, 0x5)	},
174 	{ SMCA_FP,	 HWID_MCATYPE(0xB0, 0x6)	},
175 	{ SMCA_L3_CACHE, HWID_MCATYPE(0xB0, 0x7)	},
176 
177 	/* Data Fabric MCA types */
178 	{ SMCA_CS,	 HWID_MCATYPE(0x2E, 0x0)	},
179 	{ SMCA_PIE,	 HWID_MCATYPE(0x2E, 0x1)	},
180 	{ SMCA_CS_V2,	 HWID_MCATYPE(0x2E, 0x2)	},
181 
182 	/* Unified Memory Controller MCA type */
183 	{ SMCA_UMC,	 HWID_MCATYPE(0x96, 0x0)	},
184 	{ SMCA_UMC_V2,	 HWID_MCATYPE(0x96, 0x1)	},
185 
186 	/* Parameter Block MCA type */
187 	{ SMCA_PB,	 HWID_MCATYPE(0x05, 0x0)	},
188 
189 	/* Platform Security Processor MCA type */
190 	{ SMCA_PSP,	 HWID_MCATYPE(0xFF, 0x0)	},
191 	{ SMCA_PSP_V2,	 HWID_MCATYPE(0xFF, 0x1)	},
192 
193 	/* System Management Unit MCA type */
194 	{ SMCA_SMU,	 HWID_MCATYPE(0x01, 0x0)	},
195 	{ SMCA_SMU_V2,	 HWID_MCATYPE(0x01, 0x1)	},
196 
197 	/* Microprocessor 5 Unit MCA type */
198 	{ SMCA_MP5,	 HWID_MCATYPE(0x01, 0x2)	},
199 
200 	/* MPDMA MCA type */
201 	{ SMCA_MPDMA,	 HWID_MCATYPE(0x01, 0x3)	},
202 
203 	/* Northbridge IO Unit MCA type */
204 	{ SMCA_NBIO,	 HWID_MCATYPE(0x18, 0x0)	},
205 
206 	/* PCI Express Unit MCA type */
207 	{ SMCA_PCIE,	 HWID_MCATYPE(0x46, 0x0)	},
208 	{ SMCA_PCIE_V2,	 HWID_MCATYPE(0x46, 0x1)	},
209 
210 	{ SMCA_XGMI_PCS, HWID_MCATYPE(0x50, 0x0)	},
211 	{ SMCA_NBIF,	 HWID_MCATYPE(0x6C, 0x0)	},
212 	{ SMCA_SHUB,	 HWID_MCATYPE(0x80, 0x0)	},
213 	{ SMCA_SATA,	 HWID_MCATYPE(0xA8, 0x0)	},
214 	{ SMCA_USB,	 HWID_MCATYPE(0xAA, 0x0)	},
215 	{ SMCA_GMI_PCS,  HWID_MCATYPE(0x241, 0x0)	},
216 	{ SMCA_XGMI_PHY, HWID_MCATYPE(0x259, 0x0)	},
217 	{ SMCA_WAFL_PHY, HWID_MCATYPE(0x267, 0x0)	},
218 	{ SMCA_GMI_PHY,	 HWID_MCATYPE(0x269, 0x0)	},
219 };
220 
221 /*
222  * In SMCA enabled processors, we can have multiple banks for a given IP type.
223  * So to define a unique name for each bank, we use a temp c-string to append
224  * the MCA_IPID[InstanceId] to type's name in get_name().
225  *
226  * InstanceId is 32 bits which is 8 characters. Make sure MAX_MCATYPE_NAME_LEN
227  * is greater than 8 plus 1 (for underscore) plus length of longest type name.
228  */
229 #define MAX_MCATYPE_NAME_LEN	30
230 static char buf_mcatype[MAX_MCATYPE_NAME_LEN];
231 
232 static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
233 
234 /*
235  * A list of the banks enabled on each logical CPU. Controls which respective
236  * descriptors to initialize later in mce_threshold_create_device().
237  */
238 static DEFINE_PER_CPU(u64, bank_map);
239 
240 /* Map of banks that have more than MCA_MISC0 available. */
241 static DEFINE_PER_CPU(u64, smca_misc_banks_map);
242 
243 static void amd_threshold_interrupt(void);
244 static void amd_deferred_error_interrupt(void);
245 
246 static void default_deferred_error_interrupt(void)
247 {
248 	pr_err("Unexpected deferred interrupt at vector %x\n", DEFERRED_ERROR_VECTOR);
249 }
250 void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
251 
252 static void smca_set_misc_banks_map(unsigned int bank, unsigned int cpu)
253 {
254 	u32 low, high;
255 
256 	/*
257 	 * For SMCA enabled processors, BLKPTR field of the first MISC register
258 	 * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
259 	 */
260 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
261 		return;
262 
263 	if (!(low & MCI_CONFIG_MCAX))
264 		return;
265 
266 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high))
267 		return;
268 
269 	if (low & MASK_BLKPTR_LO)
270 		per_cpu(smca_misc_banks_map, cpu) |= BIT_ULL(bank);
271 
272 }
273 
274 static void smca_configure(unsigned int bank, unsigned int cpu)
275 {
276 	u8 *bank_counts = this_cpu_ptr(smca_bank_counts);
277 	const struct smca_hwid *s_hwid;
278 	unsigned int i, hwid_mcatype;
279 	u32 high, low;
280 	u32 smca_config = MSR_AMD64_SMCA_MCx_CONFIG(bank);
281 
282 	/* Set appropriate bits in MCA_CONFIG */
283 	if (!rdmsr_safe(smca_config, &low, &high)) {
284 		/*
285 		 * OS is required to set the MCAX bit to acknowledge that it is
286 		 * now using the new MSR ranges and new registers under each
287 		 * bank. It also means that the OS will configure deferred
288 		 * errors in the new MCx_CONFIG register. If the bit is not set,
289 		 * uncorrectable errors will cause a system panic.
290 		 *
291 		 * MCA_CONFIG[MCAX] is bit 32 (0 in the high portion of the MSR.)
292 		 */
293 		high |= BIT(0);
294 
295 		/*
296 		 * SMCA sets the Deferred Error Interrupt type per bank.
297 		 *
298 		 * MCA_CONFIG[DeferredIntTypeSupported] is bit 5, and tells us
299 		 * if the DeferredIntType bit field is available.
300 		 *
301 		 * MCA_CONFIG[DeferredIntType] is bits [38:37] ([6:5] in the
302 		 * high portion of the MSR). OS should set this to 0x1 to enable
303 		 * APIC based interrupt. First, check that no interrupt has been
304 		 * set.
305 		 */
306 		if ((low & BIT(5)) && !((high >> 5) & 0x3))
307 			high |= BIT(5);
308 
309 		this_cpu_ptr(mce_banks_array)[bank].lsb_in_status = !!(low & BIT(8));
310 
311 		wrmsr(smca_config, low, high);
312 	}
313 
314 	smca_set_misc_banks_map(bank, cpu);
315 
316 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
317 		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
318 		return;
319 	}
320 
321 	hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
322 				    (high & MCI_IPID_MCATYPE) >> 16);
323 
324 	for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
325 		s_hwid = &smca_hwid_mcatypes[i];
326 
327 		if (hwid_mcatype == s_hwid->hwid_mcatype) {
328 			this_cpu_ptr(smca_banks)[bank].hwid = s_hwid;
329 			this_cpu_ptr(smca_banks)[bank].id = low;
330 			this_cpu_ptr(smca_banks)[bank].sysfs_id = bank_counts[s_hwid->bank_type]++;
331 			break;
332 		}
333 	}
334 }
335 
336 struct thresh_restart {
337 	struct threshold_block	*b;
338 	int			reset;
339 	int			set_lvt_off;
340 	int			lvt_off;
341 	u16			old_limit;
342 };
343 
344 static inline bool is_shared_bank(int bank)
345 {
346 	/*
347 	 * Scalable MCA provides for only one core to have access to the MSRs of
348 	 * a shared bank.
349 	 */
350 	if (mce_flags.smca)
351 		return false;
352 
353 	/* Bank 4 is for northbridge reporting and is thus shared */
354 	return (bank == 4);
355 }
356 
357 static const char *bank4_names(const struct threshold_block *b)
358 {
359 	switch (b->address) {
360 	/* MSR4_MISC0 */
361 	case 0x00000413:
362 		return "dram";
363 
364 	case 0xc0000408:
365 		return "ht_links";
366 
367 	case 0xc0000409:
368 		return "l3_cache";
369 
370 	default:
371 		WARN(1, "Funny MSR: 0x%08x\n", b->address);
372 		return "";
373 	}
374 };
375 
376 
377 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
378 {
379 	/*
380 	 * bank 4 supports APIC LVT interrupts implicitly since forever.
381 	 */
382 	if (bank == 4)
383 		return true;
384 
385 	/*
386 	 * IntP: interrupt present; if this bit is set, the thresholding
387 	 * bank can generate APIC LVT interrupts
388 	 */
389 	return msr_high_bits & BIT(28);
390 }
391 
392 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
393 {
394 	int msr = (hi & MASK_LVTOFF_HI) >> 20;
395 
396 	if (apic < 0) {
397 		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
398 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
399 		       b->bank, b->block, b->address, hi, lo);
400 		return 0;
401 	}
402 
403 	if (apic != msr) {
404 		/*
405 		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
406 		 * the BIOS provides the value. The original field where LVT offset
407 		 * was set is reserved. Return early here:
408 		 */
409 		if (mce_flags.smca)
410 			return 0;
411 
412 		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
413 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
414 		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
415 		return 0;
416 	}
417 
418 	return 1;
419 };
420 
421 /* Reprogram MCx_MISC MSR behind this threshold bank. */
422 static void threshold_restart_bank(void *_tr)
423 {
424 	struct thresh_restart *tr = _tr;
425 	u32 hi, lo;
426 
427 	/* sysfs write might race against an offline operation */
428 	if (!this_cpu_read(threshold_banks) && !tr->set_lvt_off)
429 		return;
430 
431 	rdmsr(tr->b->address, lo, hi);
432 
433 	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
434 		tr->reset = 1;	/* limit cannot be lower than err count */
435 
436 	if (tr->reset) {		/* reset err count and overflow bit */
437 		hi =
438 		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
439 		    (THRESHOLD_MAX - tr->b->threshold_limit);
440 	} else if (tr->old_limit) {	/* change limit w/o reset */
441 		int new_count = (hi & THRESHOLD_MAX) +
442 		    (tr->old_limit - tr->b->threshold_limit);
443 
444 		hi = (hi & ~MASK_ERR_COUNT_HI) |
445 		    (new_count & THRESHOLD_MAX);
446 	}
447 
448 	/* clear IntType */
449 	hi &= ~MASK_INT_TYPE_HI;
450 
451 	if (!tr->b->interrupt_capable)
452 		goto done;
453 
454 	if (tr->set_lvt_off) {
455 		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
456 			/* set new lvt offset */
457 			hi &= ~MASK_LVTOFF_HI;
458 			hi |= tr->lvt_off << 20;
459 		}
460 	}
461 
462 	if (tr->b->interrupt_enable)
463 		hi |= INT_TYPE_APIC;
464 
465  done:
466 
467 	hi |= MASK_COUNT_EN_HI;
468 	wrmsr(tr->b->address, lo, hi);
469 }
470 
471 static void mce_threshold_block_init(struct threshold_block *b, int offset)
472 {
473 	struct thresh_restart tr = {
474 		.b			= b,
475 		.set_lvt_off		= 1,
476 		.lvt_off		= offset,
477 	};
478 
479 	b->threshold_limit		= THRESHOLD_MAX;
480 	threshold_restart_bank(&tr);
481 };
482 
483 static int setup_APIC_mce_threshold(int reserved, int new)
484 {
485 	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
486 					      APIC_EILVT_MSG_FIX, 0))
487 		return new;
488 
489 	return reserved;
490 }
491 
492 static int setup_APIC_deferred_error(int reserved, int new)
493 {
494 	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
495 					      APIC_EILVT_MSG_FIX, 0))
496 		return new;
497 
498 	return reserved;
499 }
500 
501 static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
502 {
503 	u32 low = 0, high = 0;
504 	int def_offset = -1, def_new;
505 
506 	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
507 		return;
508 
509 	def_new = (low & MASK_DEF_LVTOFF) >> 4;
510 	if (!(low & MASK_DEF_LVTOFF)) {
511 		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
512 		def_new = DEF_LVT_OFF;
513 		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
514 	}
515 
516 	def_offset = setup_APIC_deferred_error(def_offset, def_new);
517 	if ((def_offset == def_new) &&
518 	    (deferred_error_int_vector != amd_deferred_error_interrupt))
519 		deferred_error_int_vector = amd_deferred_error_interrupt;
520 
521 	if (!mce_flags.smca)
522 		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
523 
524 	wrmsr(MSR_CU_DEF_ERR, low, high);
525 }
526 
527 static u32 smca_get_block_address(unsigned int bank, unsigned int block,
528 				  unsigned int cpu)
529 {
530 	if (!block)
531 		return MSR_AMD64_SMCA_MCx_MISC(bank);
532 
533 	if (!(per_cpu(smca_misc_banks_map, cpu) & BIT_ULL(bank)))
534 		return 0;
535 
536 	return MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
537 }
538 
539 static u32 get_block_address(u32 current_addr, u32 low, u32 high,
540 			     unsigned int bank, unsigned int block,
541 			     unsigned int cpu)
542 {
543 	u32 addr = 0, offset = 0;
544 
545 	if ((bank >= per_cpu(mce_num_banks, cpu)) || (block >= NR_BLOCKS))
546 		return addr;
547 
548 	if (mce_flags.smca)
549 		return smca_get_block_address(bank, block, cpu);
550 
551 	/* Fall back to method we used for older processors: */
552 	switch (block) {
553 	case 0:
554 		addr = mca_msr_reg(bank, MCA_MISC);
555 		break;
556 	case 1:
557 		offset = ((low & MASK_BLKPTR_LO) >> 21);
558 		if (offset)
559 			addr = MCG_XBLK_ADDR + offset;
560 		break;
561 	default:
562 		addr = ++current_addr;
563 	}
564 	return addr;
565 }
566 
567 static int
568 prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
569 			int offset, u32 misc_high)
570 {
571 	unsigned int cpu = smp_processor_id();
572 	u32 smca_low, smca_high;
573 	struct threshold_block b;
574 	int new;
575 
576 	if (!block)
577 		per_cpu(bank_map, cpu) |= BIT_ULL(bank);
578 
579 	memset(&b, 0, sizeof(b));
580 	b.cpu			= cpu;
581 	b.bank			= bank;
582 	b.block			= block;
583 	b.address		= addr;
584 	b.interrupt_capable	= lvt_interrupt_supported(bank, misc_high);
585 
586 	if (!b.interrupt_capable)
587 		goto done;
588 
589 	b.interrupt_enable = 1;
590 
591 	if (!mce_flags.smca) {
592 		new = (misc_high & MASK_LVTOFF_HI) >> 20;
593 		goto set_offset;
594 	}
595 
596 	/* Gather LVT offset for thresholding: */
597 	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
598 		goto out;
599 
600 	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
601 
602 set_offset:
603 	offset = setup_APIC_mce_threshold(offset, new);
604 	if (offset == new)
605 		thresholding_irq_en = true;
606 
607 done:
608 	mce_threshold_block_init(&b, offset);
609 
610 out:
611 	return offset;
612 }
613 
614 bool amd_filter_mce(struct mce *m)
615 {
616 	enum smca_bank_types bank_type = smca_get_bank_type(m->extcpu, m->bank);
617 	struct cpuinfo_x86 *c = &boot_cpu_data;
618 
619 	/* See Family 17h Models 10h-2Fh Erratum #1114. */
620 	if (c->x86 == 0x17 &&
621 	    c->x86_model >= 0x10 && c->x86_model <= 0x2F &&
622 	    bank_type == SMCA_IF && XEC(m->status, 0x3f) == 10)
623 		return true;
624 
625 	/* NB GART TLB error reporting is disabled by default. */
626 	if (c->x86 < 0x17) {
627 		if (m->bank == 4 && XEC(m->status, 0x1f) == 0x5)
628 			return true;
629 	}
630 
631 	return false;
632 }
633 
634 /*
635  * Turn off thresholding banks for the following conditions:
636  * - MC4_MISC thresholding is not supported on Family 0x15.
637  * - Prevent possible spurious interrupts from the IF bank on Family 0x17
638  *   Models 0x10-0x2F due to Erratum #1114.
639  */
640 static void disable_err_thresholding(struct cpuinfo_x86 *c, unsigned int bank)
641 {
642 	int i, num_msrs;
643 	u64 hwcr;
644 	bool need_toggle;
645 	u32 msrs[NR_BLOCKS];
646 
647 	if (c->x86 == 0x15 && bank == 4) {
648 		msrs[0] = 0x00000413; /* MC4_MISC0 */
649 		msrs[1] = 0xc0000408; /* MC4_MISC1 */
650 		num_msrs = 2;
651 	} else if (c->x86 == 0x17 &&
652 		   (c->x86_model >= 0x10 && c->x86_model <= 0x2F)) {
653 
654 		if (smca_get_bank_type(smp_processor_id(), bank) != SMCA_IF)
655 			return;
656 
657 		msrs[0] = MSR_AMD64_SMCA_MCx_MISC(bank);
658 		num_msrs = 1;
659 	} else {
660 		return;
661 	}
662 
663 	rdmsrl(MSR_K7_HWCR, hwcr);
664 
665 	/* McStatusWrEn has to be set */
666 	need_toggle = !(hwcr & BIT(18));
667 	if (need_toggle)
668 		wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
669 
670 	/* Clear CntP bit safely */
671 	for (i = 0; i < num_msrs; i++)
672 		msr_clear_bit(msrs[i], 62);
673 
674 	/* restore old settings */
675 	if (need_toggle)
676 		wrmsrl(MSR_K7_HWCR, hwcr);
677 }
678 
679 /* cpu init entry point, called from mce.c with preempt off */
680 void mce_amd_feature_init(struct cpuinfo_x86 *c)
681 {
682 	unsigned int bank, block, cpu = smp_processor_id();
683 	u32 low = 0, high = 0, address = 0;
684 	int offset = -1;
685 
686 
687 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
688 		if (mce_flags.smca)
689 			smca_configure(bank, cpu);
690 
691 		disable_err_thresholding(c, bank);
692 
693 		for (block = 0; block < NR_BLOCKS; ++block) {
694 			address = get_block_address(address, low, high, bank, block, cpu);
695 			if (!address)
696 				break;
697 
698 			if (rdmsr_safe(address, &low, &high))
699 				break;
700 
701 			if (!(high & MASK_VALID_HI))
702 				continue;
703 
704 			if (!(high & MASK_CNTP_HI)  ||
705 			     (high & MASK_LOCKED_HI))
706 				continue;
707 
708 			offset = prepare_threshold_block(bank, block, address, offset, high);
709 		}
710 	}
711 
712 	if (mce_flags.succor)
713 		deferred_error_interrupt_enable(c);
714 }
715 
716 bool amd_mce_is_memory_error(struct mce *m)
717 {
718 	enum smca_bank_types bank_type;
719 	/* ErrCodeExt[20:16] */
720 	u8 xec = (m->status >> 16) & 0x1f;
721 
722 	bank_type = smca_get_bank_type(m->extcpu, m->bank);
723 	if (mce_flags.smca)
724 		return (bank_type == SMCA_UMC || bank_type == SMCA_UMC_V2) && xec == 0x0;
725 
726 	return m->bank == 4 && xec == 0x8;
727 }
728 
729 static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
730 {
731 	struct mce m;
732 
733 	mce_setup(&m);
734 
735 	m.status = status;
736 	m.misc   = misc;
737 	m.bank   = bank;
738 	m.tsc	 = rdtsc();
739 
740 	if (m.status & MCI_STATUS_ADDRV) {
741 		m.addr = addr;
742 
743 		smca_extract_err_addr(&m);
744 	}
745 
746 	if (mce_flags.smca) {
747 		rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
748 
749 		if (m.status & MCI_STATUS_SYNDV)
750 			rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
751 	}
752 
753 	mce_log(&m);
754 }
755 
756 DEFINE_IDTENTRY_SYSVEC(sysvec_deferred_error)
757 {
758 	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
759 	inc_irq_stat(irq_deferred_error_count);
760 	deferred_error_int_vector();
761 	trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
762 	ack_APIC_irq();
763 }
764 
765 /*
766  * Returns true if the logged error is deferred. False, otherwise.
767  */
768 static inline bool
769 _log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
770 {
771 	u64 status, addr = 0;
772 
773 	rdmsrl(msr_stat, status);
774 	if (!(status & MCI_STATUS_VAL))
775 		return false;
776 
777 	if (status & MCI_STATUS_ADDRV)
778 		rdmsrl(msr_addr, addr);
779 
780 	__log_error(bank, status, addr, misc);
781 
782 	wrmsrl(msr_stat, 0);
783 
784 	return status & MCI_STATUS_DEFERRED;
785 }
786 
787 static bool _log_error_deferred(unsigned int bank, u32 misc)
788 {
789 	if (!_log_error_bank(bank, mca_msr_reg(bank, MCA_STATUS),
790 			     mca_msr_reg(bank, MCA_ADDR), misc))
791 		return false;
792 
793 	/*
794 	 * Non-SMCA systems don't have MCA_DESTAT/MCA_DEADDR registers.
795 	 * Return true here to avoid accessing these registers.
796 	 */
797 	if (!mce_flags.smca)
798 		return true;
799 
800 	/* Clear MCA_DESTAT if the deferred error was logged from MCA_STATUS. */
801 	wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
802 	return true;
803 }
804 
805 /*
806  * We have three scenarios for checking for Deferred errors:
807  *
808  * 1) Non-SMCA systems check MCA_STATUS and log error if found.
809  * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
810  *    clear MCA_DESTAT.
811  * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
812  *    log it.
813  */
814 static void log_error_deferred(unsigned int bank)
815 {
816 	if (_log_error_deferred(bank, 0))
817 		return;
818 
819 	/*
820 	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
821 	 * for a valid error.
822 	 */
823 	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
824 			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
825 }
826 
827 /* APIC interrupt handler for deferred errors */
828 static void amd_deferred_error_interrupt(void)
829 {
830 	unsigned int bank;
831 
832 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank)
833 		log_error_deferred(bank);
834 }
835 
836 static void log_error_thresholding(unsigned int bank, u64 misc)
837 {
838 	_log_error_deferred(bank, misc);
839 }
840 
841 static void log_and_reset_block(struct threshold_block *block)
842 {
843 	struct thresh_restart tr;
844 	u32 low = 0, high = 0;
845 
846 	if (!block)
847 		return;
848 
849 	if (rdmsr_safe(block->address, &low, &high))
850 		return;
851 
852 	if (!(high & MASK_OVERFLOW_HI))
853 		return;
854 
855 	/* Log the MCE which caused the threshold event. */
856 	log_error_thresholding(block->bank, ((u64)high << 32) | low);
857 
858 	/* Reset threshold block after logging error. */
859 	memset(&tr, 0, sizeof(tr));
860 	tr.b = block;
861 	threshold_restart_bank(&tr);
862 }
863 
864 /*
865  * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
866  * goes off when error_count reaches threshold_limit.
867  */
868 static void amd_threshold_interrupt(void)
869 {
870 	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
871 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
872 	unsigned int bank, cpu = smp_processor_id();
873 
874 	/*
875 	 * Validate that the threshold bank has been initialized already. The
876 	 * handler is installed at boot time, but on a hotplug event the
877 	 * interrupt might fire before the data has been initialized.
878 	 */
879 	if (!bp)
880 		return;
881 
882 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
883 		if (!(per_cpu(bank_map, cpu) & BIT_ULL(bank)))
884 			continue;
885 
886 		first_block = bp[bank]->blocks;
887 		if (!first_block)
888 			continue;
889 
890 		/*
891 		 * The first block is also the head of the list. Check it first
892 		 * before iterating over the rest.
893 		 */
894 		log_and_reset_block(first_block);
895 		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
896 			log_and_reset_block(block);
897 	}
898 }
899 
900 /*
901  * Sysfs Interface
902  */
903 
904 struct threshold_attr {
905 	struct attribute attr;
906 	ssize_t (*show) (struct threshold_block *, char *);
907 	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
908 };
909 
910 #define SHOW_FIELDS(name)						\
911 static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
912 {									\
913 	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
914 }
915 SHOW_FIELDS(interrupt_enable)
916 SHOW_FIELDS(threshold_limit)
917 
918 static ssize_t
919 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
920 {
921 	struct thresh_restart tr;
922 	unsigned long new;
923 
924 	if (!b->interrupt_capable)
925 		return -EINVAL;
926 
927 	if (kstrtoul(buf, 0, &new) < 0)
928 		return -EINVAL;
929 
930 	b->interrupt_enable = !!new;
931 
932 	memset(&tr, 0, sizeof(tr));
933 	tr.b		= b;
934 
935 	if (smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1))
936 		return -ENODEV;
937 
938 	return size;
939 }
940 
941 static ssize_t
942 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
943 {
944 	struct thresh_restart tr;
945 	unsigned long new;
946 
947 	if (kstrtoul(buf, 0, &new) < 0)
948 		return -EINVAL;
949 
950 	if (new > THRESHOLD_MAX)
951 		new = THRESHOLD_MAX;
952 	if (new < 1)
953 		new = 1;
954 
955 	memset(&tr, 0, sizeof(tr));
956 	tr.old_limit = b->threshold_limit;
957 	b->threshold_limit = new;
958 	tr.b = b;
959 
960 	if (smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1))
961 		return -ENODEV;
962 
963 	return size;
964 }
965 
966 static ssize_t show_error_count(struct threshold_block *b, char *buf)
967 {
968 	u32 lo, hi;
969 
970 	/* CPU might be offline by now */
971 	if (rdmsr_on_cpu(b->cpu, b->address, &lo, &hi))
972 		return -ENODEV;
973 
974 	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
975 				     (THRESHOLD_MAX - b->threshold_limit)));
976 }
977 
978 static struct threshold_attr error_count = {
979 	.attr = {.name = __stringify(error_count), .mode = 0444 },
980 	.show = show_error_count,
981 };
982 
983 #define RW_ATTR(val)							\
984 static struct threshold_attr val = {					\
985 	.attr	= {.name = __stringify(val), .mode = 0644 },		\
986 	.show	= show_## val,						\
987 	.store	= store_## val,						\
988 };
989 
990 RW_ATTR(interrupt_enable);
991 RW_ATTR(threshold_limit);
992 
993 static struct attribute *default_attrs[] = {
994 	&threshold_limit.attr,
995 	&error_count.attr,
996 	NULL,	/* possibly interrupt_enable if supported, see below */
997 	NULL,
998 };
999 ATTRIBUTE_GROUPS(default);
1000 
1001 #define to_block(k)	container_of(k, struct threshold_block, kobj)
1002 #define to_attr(a)	container_of(a, struct threshold_attr, attr)
1003 
1004 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1005 {
1006 	struct threshold_block *b = to_block(kobj);
1007 	struct threshold_attr *a = to_attr(attr);
1008 	ssize_t ret;
1009 
1010 	ret = a->show ? a->show(b, buf) : -EIO;
1011 
1012 	return ret;
1013 }
1014 
1015 static ssize_t store(struct kobject *kobj, struct attribute *attr,
1016 		     const char *buf, size_t count)
1017 {
1018 	struct threshold_block *b = to_block(kobj);
1019 	struct threshold_attr *a = to_attr(attr);
1020 	ssize_t ret;
1021 
1022 	ret = a->store ? a->store(b, buf, count) : -EIO;
1023 
1024 	return ret;
1025 }
1026 
1027 static const struct sysfs_ops threshold_ops = {
1028 	.show			= show,
1029 	.store			= store,
1030 };
1031 
1032 static void threshold_block_release(struct kobject *kobj);
1033 
1034 static const struct kobj_type threshold_ktype = {
1035 	.sysfs_ops		= &threshold_ops,
1036 	.default_groups		= default_groups,
1037 	.release		= threshold_block_release,
1038 };
1039 
1040 static const char *get_name(unsigned int cpu, unsigned int bank, struct threshold_block *b)
1041 {
1042 	enum smca_bank_types bank_type;
1043 
1044 	if (!mce_flags.smca) {
1045 		if (b && bank == 4)
1046 			return bank4_names(b);
1047 
1048 		return th_names[bank];
1049 	}
1050 
1051 	bank_type = smca_get_bank_type(cpu, bank);
1052 	if (bank_type >= N_SMCA_BANK_TYPES)
1053 		return NULL;
1054 
1055 	if (b && (bank_type == SMCA_UMC || bank_type == SMCA_UMC_V2)) {
1056 		if (b->block < ARRAY_SIZE(smca_umc_block_names))
1057 			return smca_umc_block_names[b->block];
1058 		return NULL;
1059 	}
1060 
1061 	if (per_cpu(smca_bank_counts, cpu)[bank_type] == 1)
1062 		return smca_get_name(bank_type);
1063 
1064 	snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1065 		 "%s_%u", smca_get_name(bank_type),
1066 			  per_cpu(smca_banks, cpu)[bank].sysfs_id);
1067 	return buf_mcatype;
1068 }
1069 
1070 static int allocate_threshold_blocks(unsigned int cpu, struct threshold_bank *tb,
1071 				     unsigned int bank, unsigned int block,
1072 				     u32 address)
1073 {
1074 	struct threshold_block *b = NULL;
1075 	u32 low, high;
1076 	int err;
1077 
1078 	if ((bank >= this_cpu_read(mce_num_banks)) || (block >= NR_BLOCKS))
1079 		return 0;
1080 
1081 	if (rdmsr_safe(address, &low, &high))
1082 		return 0;
1083 
1084 	if (!(high & MASK_VALID_HI)) {
1085 		if (block)
1086 			goto recurse;
1087 		else
1088 			return 0;
1089 	}
1090 
1091 	if (!(high & MASK_CNTP_HI)  ||
1092 	     (high & MASK_LOCKED_HI))
1093 		goto recurse;
1094 
1095 	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1096 	if (!b)
1097 		return -ENOMEM;
1098 
1099 	b->block		= block;
1100 	b->bank			= bank;
1101 	b->cpu			= cpu;
1102 	b->address		= address;
1103 	b->interrupt_enable	= 0;
1104 	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
1105 	b->threshold_limit	= THRESHOLD_MAX;
1106 
1107 	if (b->interrupt_capable) {
1108 		default_attrs[2] = &interrupt_enable.attr;
1109 		b->interrupt_enable = 1;
1110 	} else {
1111 		default_attrs[2] = NULL;
1112 	}
1113 
1114 	INIT_LIST_HEAD(&b->miscj);
1115 
1116 	/* This is safe as @tb is not visible yet */
1117 	if (tb->blocks)
1118 		list_add(&b->miscj, &tb->blocks->miscj);
1119 	else
1120 		tb->blocks = b;
1121 
1122 	err = kobject_init_and_add(&b->kobj, &threshold_ktype, tb->kobj, get_name(cpu, bank, b));
1123 	if (err)
1124 		goto out_free;
1125 recurse:
1126 	address = get_block_address(address, low, high, bank, ++block, cpu);
1127 	if (!address)
1128 		return 0;
1129 
1130 	err = allocate_threshold_blocks(cpu, tb, bank, block, address);
1131 	if (err)
1132 		goto out_free;
1133 
1134 	if (b)
1135 		kobject_uevent(&b->kobj, KOBJ_ADD);
1136 
1137 	return 0;
1138 
1139 out_free:
1140 	if (b) {
1141 		list_del(&b->miscj);
1142 		kobject_put(&b->kobj);
1143 	}
1144 	return err;
1145 }
1146 
1147 static int __threshold_add_blocks(struct threshold_bank *b)
1148 {
1149 	struct list_head *head = &b->blocks->miscj;
1150 	struct threshold_block *pos = NULL;
1151 	struct threshold_block *tmp = NULL;
1152 	int err = 0;
1153 
1154 	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1155 	if (err)
1156 		return err;
1157 
1158 	list_for_each_entry_safe(pos, tmp, head, miscj) {
1159 
1160 		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1161 		if (err) {
1162 			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1163 				kobject_del(&pos->kobj);
1164 
1165 			return err;
1166 		}
1167 	}
1168 	return err;
1169 }
1170 
1171 static int threshold_create_bank(struct threshold_bank **bp, unsigned int cpu,
1172 				 unsigned int bank)
1173 {
1174 	struct device *dev = this_cpu_read(mce_device);
1175 	struct amd_northbridge *nb = NULL;
1176 	struct threshold_bank *b = NULL;
1177 	const char *name = get_name(cpu, bank, NULL);
1178 	int err = 0;
1179 
1180 	if (!dev)
1181 		return -ENODEV;
1182 
1183 	if (is_shared_bank(bank)) {
1184 		nb = node_to_amd_nb(topology_die_id(cpu));
1185 
1186 		/* threshold descriptor already initialized on this node? */
1187 		if (nb && nb->bank4) {
1188 			/* yes, use it */
1189 			b = nb->bank4;
1190 			err = kobject_add(b->kobj, &dev->kobj, name);
1191 			if (err)
1192 				goto out;
1193 
1194 			bp[bank] = b;
1195 			refcount_inc(&b->cpus);
1196 
1197 			err = __threshold_add_blocks(b);
1198 
1199 			goto out;
1200 		}
1201 	}
1202 
1203 	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1204 	if (!b) {
1205 		err = -ENOMEM;
1206 		goto out;
1207 	}
1208 
1209 	/* Associate the bank with the per-CPU MCE device */
1210 	b->kobj = kobject_create_and_add(name, &dev->kobj);
1211 	if (!b->kobj) {
1212 		err = -EINVAL;
1213 		goto out_free;
1214 	}
1215 
1216 	if (is_shared_bank(bank)) {
1217 		b->shared = 1;
1218 		refcount_set(&b->cpus, 1);
1219 
1220 		/* nb is already initialized, see above */
1221 		if (nb) {
1222 			WARN_ON(nb->bank4);
1223 			nb->bank4 = b;
1224 		}
1225 	}
1226 
1227 	err = allocate_threshold_blocks(cpu, b, bank, 0, mca_msr_reg(bank, MCA_MISC));
1228 	if (err)
1229 		goto out_kobj;
1230 
1231 	bp[bank] = b;
1232 	return 0;
1233 
1234 out_kobj:
1235 	kobject_put(b->kobj);
1236 out_free:
1237 	kfree(b);
1238 out:
1239 	return err;
1240 }
1241 
1242 static void threshold_block_release(struct kobject *kobj)
1243 {
1244 	kfree(to_block(kobj));
1245 }
1246 
1247 static void deallocate_threshold_blocks(struct threshold_bank *bank)
1248 {
1249 	struct threshold_block *pos, *tmp;
1250 
1251 	list_for_each_entry_safe(pos, tmp, &bank->blocks->miscj, miscj) {
1252 		list_del(&pos->miscj);
1253 		kobject_put(&pos->kobj);
1254 	}
1255 
1256 	kobject_put(&bank->blocks->kobj);
1257 }
1258 
1259 static void __threshold_remove_blocks(struct threshold_bank *b)
1260 {
1261 	struct threshold_block *pos = NULL;
1262 	struct threshold_block *tmp = NULL;
1263 
1264 	kobject_put(b->kobj);
1265 
1266 	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1267 		kobject_put(b->kobj);
1268 }
1269 
1270 static void threshold_remove_bank(struct threshold_bank *bank)
1271 {
1272 	struct amd_northbridge *nb;
1273 
1274 	if (!bank->blocks)
1275 		goto out_free;
1276 
1277 	if (!bank->shared)
1278 		goto out_dealloc;
1279 
1280 	if (!refcount_dec_and_test(&bank->cpus)) {
1281 		__threshold_remove_blocks(bank);
1282 		return;
1283 	} else {
1284 		/*
1285 		 * The last CPU on this node using the shared bank is going
1286 		 * away, remove that bank now.
1287 		 */
1288 		nb = node_to_amd_nb(topology_die_id(smp_processor_id()));
1289 		nb->bank4 = NULL;
1290 	}
1291 
1292 out_dealloc:
1293 	deallocate_threshold_blocks(bank);
1294 
1295 out_free:
1296 	kobject_put(bank->kobj);
1297 	kfree(bank);
1298 }
1299 
1300 static void __threshold_remove_device(struct threshold_bank **bp)
1301 {
1302 	unsigned int bank, numbanks = this_cpu_read(mce_num_banks);
1303 
1304 	for (bank = 0; bank < numbanks; bank++) {
1305 		if (!bp[bank])
1306 			continue;
1307 
1308 		threshold_remove_bank(bp[bank]);
1309 		bp[bank] = NULL;
1310 	}
1311 	kfree(bp);
1312 }
1313 
1314 int mce_threshold_remove_device(unsigned int cpu)
1315 {
1316 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
1317 
1318 	if (!bp)
1319 		return 0;
1320 
1321 	/*
1322 	 * Clear the pointer before cleaning up, so that the interrupt won't
1323 	 * touch anything of this.
1324 	 */
1325 	this_cpu_write(threshold_banks, NULL);
1326 
1327 	__threshold_remove_device(bp);
1328 	return 0;
1329 }
1330 
1331 /**
1332  * mce_threshold_create_device - Create the per-CPU MCE threshold device
1333  * @cpu:	The plugged in CPU
1334  *
1335  * Create directories and files for all valid threshold banks.
1336  *
1337  * This is invoked from the CPU hotplug callback which was installed in
1338  * mcheck_init_device(). The invocation happens in context of the hotplug
1339  * thread running on @cpu.  The callback is invoked on all CPUs which are
1340  * online when the callback is installed or during a real hotplug event.
1341  */
1342 int mce_threshold_create_device(unsigned int cpu)
1343 {
1344 	unsigned int numbanks, bank;
1345 	struct threshold_bank **bp;
1346 	int err;
1347 
1348 	if (!mce_flags.amd_threshold)
1349 		return 0;
1350 
1351 	bp = this_cpu_read(threshold_banks);
1352 	if (bp)
1353 		return 0;
1354 
1355 	numbanks = this_cpu_read(mce_num_banks);
1356 	bp = kcalloc(numbanks, sizeof(*bp), GFP_KERNEL);
1357 	if (!bp)
1358 		return -ENOMEM;
1359 
1360 	for (bank = 0; bank < numbanks; ++bank) {
1361 		if (!(this_cpu_read(bank_map) & BIT_ULL(bank)))
1362 			continue;
1363 		err = threshold_create_bank(bp, cpu, bank);
1364 		if (err) {
1365 			__threshold_remove_device(bp);
1366 			return err;
1367 		}
1368 	}
1369 	this_cpu_write(threshold_banks, bp);
1370 
1371 	if (thresholding_irq_en)
1372 		mce_threshold_vector = amd_threshold_interrupt;
1373 	return 0;
1374 }
1375