xref: /openbmc/linux/arch/x86/kernel/cpu/mce/amd.c (revision 2455f0e1)
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(unsigned int, bank_map);
239 
240 /* Map of banks that have more than MCA_MISC0 available. */
241 static DEFINE_PER_CPU(u32, 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(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 		wrmsr(smca_config, low, high);
310 	}
311 
312 	smca_set_misc_banks_map(bank, cpu);
313 
314 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
315 		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
316 		return;
317 	}
318 
319 	hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
320 				    (high & MCI_IPID_MCATYPE) >> 16);
321 
322 	for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
323 		s_hwid = &smca_hwid_mcatypes[i];
324 
325 		if (hwid_mcatype == s_hwid->hwid_mcatype) {
326 			this_cpu_ptr(smca_banks)[bank].hwid = s_hwid;
327 			this_cpu_ptr(smca_banks)[bank].id = low;
328 			this_cpu_ptr(smca_banks)[bank].sysfs_id = bank_counts[s_hwid->bank_type]++;
329 			break;
330 		}
331 	}
332 }
333 
334 struct thresh_restart {
335 	struct threshold_block	*b;
336 	int			reset;
337 	int			set_lvt_off;
338 	int			lvt_off;
339 	u16			old_limit;
340 };
341 
342 static inline bool is_shared_bank(int bank)
343 {
344 	/*
345 	 * Scalable MCA provides for only one core to have access to the MSRs of
346 	 * a shared bank.
347 	 */
348 	if (mce_flags.smca)
349 		return false;
350 
351 	/* Bank 4 is for northbridge reporting and is thus shared */
352 	return (bank == 4);
353 }
354 
355 static const char *bank4_names(const struct threshold_block *b)
356 {
357 	switch (b->address) {
358 	/* MSR4_MISC0 */
359 	case 0x00000413:
360 		return "dram";
361 
362 	case 0xc0000408:
363 		return "ht_links";
364 
365 	case 0xc0000409:
366 		return "l3_cache";
367 
368 	default:
369 		WARN(1, "Funny MSR: 0x%08x\n", b->address);
370 		return "";
371 	}
372 };
373 
374 
375 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
376 {
377 	/*
378 	 * bank 4 supports APIC LVT interrupts implicitly since forever.
379 	 */
380 	if (bank == 4)
381 		return true;
382 
383 	/*
384 	 * IntP: interrupt present; if this bit is set, the thresholding
385 	 * bank can generate APIC LVT interrupts
386 	 */
387 	return msr_high_bits & BIT(28);
388 }
389 
390 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
391 {
392 	int msr = (hi & MASK_LVTOFF_HI) >> 20;
393 
394 	if (apic < 0) {
395 		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
396 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
397 		       b->bank, b->block, b->address, hi, lo);
398 		return 0;
399 	}
400 
401 	if (apic != msr) {
402 		/*
403 		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
404 		 * the BIOS provides the value. The original field where LVT offset
405 		 * was set is reserved. Return early here:
406 		 */
407 		if (mce_flags.smca)
408 			return 0;
409 
410 		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
411 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
412 		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
413 		return 0;
414 	}
415 
416 	return 1;
417 };
418 
419 /* Reprogram MCx_MISC MSR behind this threshold bank. */
420 static void threshold_restart_bank(void *_tr)
421 {
422 	struct thresh_restart *tr = _tr;
423 	u32 hi, lo;
424 
425 	/* sysfs write might race against an offline operation */
426 	if (!this_cpu_read(threshold_banks) && !tr->set_lvt_off)
427 		return;
428 
429 	rdmsr(tr->b->address, lo, hi);
430 
431 	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
432 		tr->reset = 1;	/* limit cannot be lower than err count */
433 
434 	if (tr->reset) {		/* reset err count and overflow bit */
435 		hi =
436 		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
437 		    (THRESHOLD_MAX - tr->b->threshold_limit);
438 	} else if (tr->old_limit) {	/* change limit w/o reset */
439 		int new_count = (hi & THRESHOLD_MAX) +
440 		    (tr->old_limit - tr->b->threshold_limit);
441 
442 		hi = (hi & ~MASK_ERR_COUNT_HI) |
443 		    (new_count & THRESHOLD_MAX);
444 	}
445 
446 	/* clear IntType */
447 	hi &= ~MASK_INT_TYPE_HI;
448 
449 	if (!tr->b->interrupt_capable)
450 		goto done;
451 
452 	if (tr->set_lvt_off) {
453 		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
454 			/* set new lvt offset */
455 			hi &= ~MASK_LVTOFF_HI;
456 			hi |= tr->lvt_off << 20;
457 		}
458 	}
459 
460 	if (tr->b->interrupt_enable)
461 		hi |= INT_TYPE_APIC;
462 
463  done:
464 
465 	hi |= MASK_COUNT_EN_HI;
466 	wrmsr(tr->b->address, lo, hi);
467 }
468 
469 static void mce_threshold_block_init(struct threshold_block *b, int offset)
470 {
471 	struct thresh_restart tr = {
472 		.b			= b,
473 		.set_lvt_off		= 1,
474 		.lvt_off		= offset,
475 	};
476 
477 	b->threshold_limit		= THRESHOLD_MAX;
478 	threshold_restart_bank(&tr);
479 };
480 
481 static int setup_APIC_mce_threshold(int reserved, int new)
482 {
483 	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
484 					      APIC_EILVT_MSG_FIX, 0))
485 		return new;
486 
487 	return reserved;
488 }
489 
490 static int setup_APIC_deferred_error(int reserved, int new)
491 {
492 	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
493 					      APIC_EILVT_MSG_FIX, 0))
494 		return new;
495 
496 	return reserved;
497 }
498 
499 static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
500 {
501 	u32 low = 0, high = 0;
502 	int def_offset = -1, def_new;
503 
504 	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
505 		return;
506 
507 	def_new = (low & MASK_DEF_LVTOFF) >> 4;
508 	if (!(low & MASK_DEF_LVTOFF)) {
509 		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
510 		def_new = DEF_LVT_OFF;
511 		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
512 	}
513 
514 	def_offset = setup_APIC_deferred_error(def_offset, def_new);
515 	if ((def_offset == def_new) &&
516 	    (deferred_error_int_vector != amd_deferred_error_interrupt))
517 		deferred_error_int_vector = amd_deferred_error_interrupt;
518 
519 	if (!mce_flags.smca)
520 		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
521 
522 	wrmsr(MSR_CU_DEF_ERR, low, high);
523 }
524 
525 static u32 smca_get_block_address(unsigned int bank, unsigned int block,
526 				  unsigned int cpu)
527 {
528 	if (!block)
529 		return MSR_AMD64_SMCA_MCx_MISC(bank);
530 
531 	if (!(per_cpu(smca_misc_banks_map, cpu) & BIT(bank)))
532 		return 0;
533 
534 	return MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
535 }
536 
537 static u32 get_block_address(u32 current_addr, u32 low, u32 high,
538 			     unsigned int bank, unsigned int block,
539 			     unsigned int cpu)
540 {
541 	u32 addr = 0, offset = 0;
542 
543 	if ((bank >= per_cpu(mce_num_banks, cpu)) || (block >= NR_BLOCKS))
544 		return addr;
545 
546 	if (mce_flags.smca)
547 		return smca_get_block_address(bank, block, cpu);
548 
549 	/* Fall back to method we used for older processors: */
550 	switch (block) {
551 	case 0:
552 		addr = mca_msr_reg(bank, MCA_MISC);
553 		break;
554 	case 1:
555 		offset = ((low & MASK_BLKPTR_LO) >> 21);
556 		if (offset)
557 			addr = MCG_XBLK_ADDR + offset;
558 		break;
559 	default:
560 		addr = ++current_addr;
561 	}
562 	return addr;
563 }
564 
565 static int
566 prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
567 			int offset, u32 misc_high)
568 {
569 	unsigned int cpu = smp_processor_id();
570 	u32 smca_low, smca_high;
571 	struct threshold_block b;
572 	int new;
573 
574 	if (!block)
575 		per_cpu(bank_map, cpu) |= (1 << bank);
576 
577 	memset(&b, 0, sizeof(b));
578 	b.cpu			= cpu;
579 	b.bank			= bank;
580 	b.block			= block;
581 	b.address		= addr;
582 	b.interrupt_capable	= lvt_interrupt_supported(bank, misc_high);
583 
584 	if (!b.interrupt_capable)
585 		goto done;
586 
587 	b.interrupt_enable = 1;
588 
589 	if (!mce_flags.smca) {
590 		new = (misc_high & MASK_LVTOFF_HI) >> 20;
591 		goto set_offset;
592 	}
593 
594 	/* Gather LVT offset for thresholding: */
595 	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
596 		goto out;
597 
598 	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
599 
600 set_offset:
601 	offset = setup_APIC_mce_threshold(offset, new);
602 	if (offset == new)
603 		thresholding_irq_en = true;
604 
605 done:
606 	mce_threshold_block_init(&b, offset);
607 
608 out:
609 	return offset;
610 }
611 
612 bool amd_filter_mce(struct mce *m)
613 {
614 	enum smca_bank_types bank_type = smca_get_bank_type(m->extcpu, m->bank);
615 	struct cpuinfo_x86 *c = &boot_cpu_data;
616 
617 	/* See Family 17h Models 10h-2Fh Erratum #1114. */
618 	if (c->x86 == 0x17 &&
619 	    c->x86_model >= 0x10 && c->x86_model <= 0x2F &&
620 	    bank_type == SMCA_IF && XEC(m->status, 0x3f) == 10)
621 		return true;
622 
623 	/* NB GART TLB error reporting is disabled by default. */
624 	if (c->x86 < 0x17) {
625 		if (m->bank == 4 && XEC(m->status, 0x1f) == 0x5)
626 			return true;
627 	}
628 
629 	return false;
630 }
631 
632 /*
633  * Turn off thresholding banks for the following conditions:
634  * - MC4_MISC thresholding is not supported on Family 0x15.
635  * - Prevent possible spurious interrupts from the IF bank on Family 0x17
636  *   Models 0x10-0x2F due to Erratum #1114.
637  */
638 static void disable_err_thresholding(struct cpuinfo_x86 *c, unsigned int bank)
639 {
640 	int i, num_msrs;
641 	u64 hwcr;
642 	bool need_toggle;
643 	u32 msrs[NR_BLOCKS];
644 
645 	if (c->x86 == 0x15 && bank == 4) {
646 		msrs[0] = 0x00000413; /* MC4_MISC0 */
647 		msrs[1] = 0xc0000408; /* MC4_MISC1 */
648 		num_msrs = 2;
649 	} else if (c->x86 == 0x17 &&
650 		   (c->x86_model >= 0x10 && c->x86_model <= 0x2F)) {
651 
652 		if (smca_get_bank_type(smp_processor_id(), bank) != SMCA_IF)
653 			return;
654 
655 		msrs[0] = MSR_AMD64_SMCA_MCx_MISC(bank);
656 		num_msrs = 1;
657 	} else {
658 		return;
659 	}
660 
661 	rdmsrl(MSR_K7_HWCR, hwcr);
662 
663 	/* McStatusWrEn has to be set */
664 	need_toggle = !(hwcr & BIT(18));
665 	if (need_toggle)
666 		wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
667 
668 	/* Clear CntP bit safely */
669 	for (i = 0; i < num_msrs; i++)
670 		msr_clear_bit(msrs[i], 62);
671 
672 	/* restore old settings */
673 	if (need_toggle)
674 		wrmsrl(MSR_K7_HWCR, hwcr);
675 }
676 
677 /* cpu init entry point, called from mce.c with preempt off */
678 void mce_amd_feature_init(struct cpuinfo_x86 *c)
679 {
680 	unsigned int bank, block, cpu = smp_processor_id();
681 	u32 low = 0, high = 0, address = 0;
682 	int offset = -1;
683 
684 
685 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
686 		if (mce_flags.smca)
687 			smca_configure(bank, cpu);
688 
689 		disable_err_thresholding(c, bank);
690 
691 		for (block = 0; block < NR_BLOCKS; ++block) {
692 			address = get_block_address(address, low, high, bank, block, cpu);
693 			if (!address)
694 				break;
695 
696 			if (rdmsr_safe(address, &low, &high))
697 				break;
698 
699 			if (!(high & MASK_VALID_HI))
700 				continue;
701 
702 			if (!(high & MASK_CNTP_HI)  ||
703 			     (high & MASK_LOCKED_HI))
704 				continue;
705 
706 			offset = prepare_threshold_block(bank, block, address, offset, high);
707 		}
708 	}
709 
710 	if (mce_flags.succor)
711 		deferred_error_interrupt_enable(c);
712 }
713 
714 bool amd_mce_is_memory_error(struct mce *m)
715 {
716 	/* ErrCodeExt[20:16] */
717 	u8 xec = (m->status >> 16) & 0x1f;
718 
719 	if (mce_flags.smca)
720 		return smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC && xec == 0x0;
721 
722 	return m->bank == 4 && xec == 0x8;
723 }
724 
725 static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
726 {
727 	struct mce m;
728 
729 	mce_setup(&m);
730 
731 	m.status = status;
732 	m.misc   = misc;
733 	m.bank   = bank;
734 	m.tsc	 = rdtsc();
735 
736 	if (m.status & MCI_STATUS_ADDRV) {
737 		m.addr = addr;
738 
739 		/*
740 		 * Extract [55:<lsb>] where lsb is the least significant
741 		 * *valid* bit of the address bits.
742 		 */
743 		if (mce_flags.smca) {
744 			u8 lsb = (m.addr >> 56) & 0x3f;
745 
746 			m.addr &= GENMASK_ULL(55, lsb);
747 		}
748 	}
749 
750 	if (mce_flags.smca) {
751 		rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
752 
753 		if (m.status & MCI_STATUS_SYNDV)
754 			rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
755 	}
756 
757 	mce_log(&m);
758 }
759 
760 DEFINE_IDTENTRY_SYSVEC(sysvec_deferred_error)
761 {
762 	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
763 	inc_irq_stat(irq_deferred_error_count);
764 	deferred_error_int_vector();
765 	trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
766 	ack_APIC_irq();
767 }
768 
769 /*
770  * Returns true if the logged error is deferred. False, otherwise.
771  */
772 static inline bool
773 _log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
774 {
775 	u64 status, addr = 0;
776 
777 	rdmsrl(msr_stat, status);
778 	if (!(status & MCI_STATUS_VAL))
779 		return false;
780 
781 	if (status & MCI_STATUS_ADDRV)
782 		rdmsrl(msr_addr, addr);
783 
784 	__log_error(bank, status, addr, misc);
785 
786 	wrmsrl(msr_stat, 0);
787 
788 	return status & MCI_STATUS_DEFERRED;
789 }
790 
791 static bool _log_error_deferred(unsigned int bank, u32 misc)
792 {
793 	if (!_log_error_bank(bank, mca_msr_reg(bank, MCA_STATUS),
794 			     mca_msr_reg(bank, MCA_ADDR), misc))
795 		return false;
796 
797 	/*
798 	 * Non-SMCA systems don't have MCA_DESTAT/MCA_DEADDR registers.
799 	 * Return true here to avoid accessing these registers.
800 	 */
801 	if (!mce_flags.smca)
802 		return true;
803 
804 	/* Clear MCA_DESTAT if the deferred error was logged from MCA_STATUS. */
805 	wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
806 	return true;
807 }
808 
809 /*
810  * We have three scenarios for checking for Deferred errors:
811  *
812  * 1) Non-SMCA systems check MCA_STATUS and log error if found.
813  * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
814  *    clear MCA_DESTAT.
815  * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
816  *    log it.
817  */
818 static void log_error_deferred(unsigned int bank)
819 {
820 	if (_log_error_deferred(bank, 0))
821 		return;
822 
823 	/*
824 	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
825 	 * for a valid error.
826 	 */
827 	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
828 			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
829 }
830 
831 /* APIC interrupt handler for deferred errors */
832 static void amd_deferred_error_interrupt(void)
833 {
834 	unsigned int bank;
835 
836 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank)
837 		log_error_deferred(bank);
838 }
839 
840 static void log_error_thresholding(unsigned int bank, u64 misc)
841 {
842 	_log_error_deferred(bank, misc);
843 }
844 
845 static void log_and_reset_block(struct threshold_block *block)
846 {
847 	struct thresh_restart tr;
848 	u32 low = 0, high = 0;
849 
850 	if (!block)
851 		return;
852 
853 	if (rdmsr_safe(block->address, &low, &high))
854 		return;
855 
856 	if (!(high & MASK_OVERFLOW_HI))
857 		return;
858 
859 	/* Log the MCE which caused the threshold event. */
860 	log_error_thresholding(block->bank, ((u64)high << 32) | low);
861 
862 	/* Reset threshold block after logging error. */
863 	memset(&tr, 0, sizeof(tr));
864 	tr.b = block;
865 	threshold_restart_bank(&tr);
866 }
867 
868 /*
869  * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
870  * goes off when error_count reaches threshold_limit.
871  */
872 static void amd_threshold_interrupt(void)
873 {
874 	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
875 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
876 	unsigned int bank, cpu = smp_processor_id();
877 
878 	/*
879 	 * Validate that the threshold bank has been initialized already. The
880 	 * handler is installed at boot time, but on a hotplug event the
881 	 * interrupt might fire before the data has been initialized.
882 	 */
883 	if (!bp)
884 		return;
885 
886 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
887 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
888 			continue;
889 
890 		first_block = bp[bank]->blocks;
891 		if (!first_block)
892 			continue;
893 
894 		/*
895 		 * The first block is also the head of the list. Check it first
896 		 * before iterating over the rest.
897 		 */
898 		log_and_reset_block(first_block);
899 		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
900 			log_and_reset_block(block);
901 	}
902 }
903 
904 /*
905  * Sysfs Interface
906  */
907 
908 struct threshold_attr {
909 	struct attribute attr;
910 	ssize_t (*show) (struct threshold_block *, char *);
911 	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
912 };
913 
914 #define SHOW_FIELDS(name)						\
915 static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
916 {									\
917 	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
918 }
919 SHOW_FIELDS(interrupt_enable)
920 SHOW_FIELDS(threshold_limit)
921 
922 static ssize_t
923 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
924 {
925 	struct thresh_restart tr;
926 	unsigned long new;
927 
928 	if (!b->interrupt_capable)
929 		return -EINVAL;
930 
931 	if (kstrtoul(buf, 0, &new) < 0)
932 		return -EINVAL;
933 
934 	b->interrupt_enable = !!new;
935 
936 	memset(&tr, 0, sizeof(tr));
937 	tr.b		= b;
938 
939 	if (smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1))
940 		return -ENODEV;
941 
942 	return size;
943 }
944 
945 static ssize_t
946 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
947 {
948 	struct thresh_restart tr;
949 	unsigned long new;
950 
951 	if (kstrtoul(buf, 0, &new) < 0)
952 		return -EINVAL;
953 
954 	if (new > THRESHOLD_MAX)
955 		new = THRESHOLD_MAX;
956 	if (new < 1)
957 		new = 1;
958 
959 	memset(&tr, 0, sizeof(tr));
960 	tr.old_limit = b->threshold_limit;
961 	b->threshold_limit = new;
962 	tr.b = b;
963 
964 	if (smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1))
965 		return -ENODEV;
966 
967 	return size;
968 }
969 
970 static ssize_t show_error_count(struct threshold_block *b, char *buf)
971 {
972 	u32 lo, hi;
973 
974 	/* CPU might be offline by now */
975 	if (rdmsr_on_cpu(b->cpu, b->address, &lo, &hi))
976 		return -ENODEV;
977 
978 	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
979 				     (THRESHOLD_MAX - b->threshold_limit)));
980 }
981 
982 static struct threshold_attr error_count = {
983 	.attr = {.name = __stringify(error_count), .mode = 0444 },
984 	.show = show_error_count,
985 };
986 
987 #define RW_ATTR(val)							\
988 static struct threshold_attr val = {					\
989 	.attr	= {.name = __stringify(val), .mode = 0644 },		\
990 	.show	= show_## val,						\
991 	.store	= store_## val,						\
992 };
993 
994 RW_ATTR(interrupt_enable);
995 RW_ATTR(threshold_limit);
996 
997 static struct attribute *default_attrs[] = {
998 	&threshold_limit.attr,
999 	&error_count.attr,
1000 	NULL,	/* possibly interrupt_enable if supported, see below */
1001 	NULL,
1002 };
1003 ATTRIBUTE_GROUPS(default);
1004 
1005 #define to_block(k)	container_of(k, struct threshold_block, kobj)
1006 #define to_attr(a)	container_of(a, struct threshold_attr, attr)
1007 
1008 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1009 {
1010 	struct threshold_block *b = to_block(kobj);
1011 	struct threshold_attr *a = to_attr(attr);
1012 	ssize_t ret;
1013 
1014 	ret = a->show ? a->show(b, buf) : -EIO;
1015 
1016 	return ret;
1017 }
1018 
1019 static ssize_t store(struct kobject *kobj, struct attribute *attr,
1020 		     const char *buf, size_t count)
1021 {
1022 	struct threshold_block *b = to_block(kobj);
1023 	struct threshold_attr *a = to_attr(attr);
1024 	ssize_t ret;
1025 
1026 	ret = a->store ? a->store(b, buf, count) : -EIO;
1027 
1028 	return ret;
1029 }
1030 
1031 static const struct sysfs_ops threshold_ops = {
1032 	.show			= show,
1033 	.store			= store,
1034 };
1035 
1036 static void threshold_block_release(struct kobject *kobj);
1037 
1038 static struct kobj_type threshold_ktype = {
1039 	.sysfs_ops		= &threshold_ops,
1040 	.default_groups		= default_groups,
1041 	.release		= threshold_block_release,
1042 };
1043 
1044 static const char *get_name(unsigned int cpu, unsigned int bank, struct threshold_block *b)
1045 {
1046 	enum smca_bank_types bank_type;
1047 
1048 	if (!mce_flags.smca) {
1049 		if (b && bank == 4)
1050 			return bank4_names(b);
1051 
1052 		return th_names[bank];
1053 	}
1054 
1055 	bank_type = smca_get_bank_type(cpu, bank);
1056 	if (bank_type >= N_SMCA_BANK_TYPES)
1057 		return NULL;
1058 
1059 	if (b && bank_type == SMCA_UMC) {
1060 		if (b->block < ARRAY_SIZE(smca_umc_block_names))
1061 			return smca_umc_block_names[b->block];
1062 		return NULL;
1063 	}
1064 
1065 	if (per_cpu(smca_bank_counts, cpu)[bank_type] == 1)
1066 		return smca_get_name(bank_type);
1067 
1068 	snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1069 		 "%s_%u", smca_get_name(bank_type),
1070 			  per_cpu(smca_banks, cpu)[bank].sysfs_id);
1071 	return buf_mcatype;
1072 }
1073 
1074 static int allocate_threshold_blocks(unsigned int cpu, struct threshold_bank *tb,
1075 				     unsigned int bank, unsigned int block,
1076 				     u32 address)
1077 {
1078 	struct threshold_block *b = NULL;
1079 	u32 low, high;
1080 	int err;
1081 
1082 	if ((bank >= this_cpu_read(mce_num_banks)) || (block >= NR_BLOCKS))
1083 		return 0;
1084 
1085 	if (rdmsr_safe(address, &low, &high))
1086 		return 0;
1087 
1088 	if (!(high & MASK_VALID_HI)) {
1089 		if (block)
1090 			goto recurse;
1091 		else
1092 			return 0;
1093 	}
1094 
1095 	if (!(high & MASK_CNTP_HI)  ||
1096 	     (high & MASK_LOCKED_HI))
1097 		goto recurse;
1098 
1099 	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1100 	if (!b)
1101 		return -ENOMEM;
1102 
1103 	b->block		= block;
1104 	b->bank			= bank;
1105 	b->cpu			= cpu;
1106 	b->address		= address;
1107 	b->interrupt_enable	= 0;
1108 	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
1109 	b->threshold_limit	= THRESHOLD_MAX;
1110 
1111 	if (b->interrupt_capable) {
1112 		default_attrs[2] = &interrupt_enable.attr;
1113 		b->interrupt_enable = 1;
1114 	} else {
1115 		default_attrs[2] = NULL;
1116 	}
1117 
1118 	INIT_LIST_HEAD(&b->miscj);
1119 
1120 	/* This is safe as @tb is not visible yet */
1121 	if (tb->blocks)
1122 		list_add(&b->miscj, &tb->blocks->miscj);
1123 	else
1124 		tb->blocks = b;
1125 
1126 	err = kobject_init_and_add(&b->kobj, &threshold_ktype, tb->kobj, get_name(cpu, bank, b));
1127 	if (err)
1128 		goto out_free;
1129 recurse:
1130 	address = get_block_address(address, low, high, bank, ++block, cpu);
1131 	if (!address)
1132 		return 0;
1133 
1134 	err = allocate_threshold_blocks(cpu, tb, bank, block, address);
1135 	if (err)
1136 		goto out_free;
1137 
1138 	if (b)
1139 		kobject_uevent(&b->kobj, KOBJ_ADD);
1140 
1141 	return 0;
1142 
1143 out_free:
1144 	if (b) {
1145 		list_del(&b->miscj);
1146 		kobject_put(&b->kobj);
1147 	}
1148 	return err;
1149 }
1150 
1151 static int __threshold_add_blocks(struct threshold_bank *b)
1152 {
1153 	struct list_head *head = &b->blocks->miscj;
1154 	struct threshold_block *pos = NULL;
1155 	struct threshold_block *tmp = NULL;
1156 	int err = 0;
1157 
1158 	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1159 	if (err)
1160 		return err;
1161 
1162 	list_for_each_entry_safe(pos, tmp, head, miscj) {
1163 
1164 		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1165 		if (err) {
1166 			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1167 				kobject_del(&pos->kobj);
1168 
1169 			return err;
1170 		}
1171 	}
1172 	return err;
1173 }
1174 
1175 static int threshold_create_bank(struct threshold_bank **bp, unsigned int cpu,
1176 				 unsigned int bank)
1177 {
1178 	struct device *dev = this_cpu_read(mce_device);
1179 	struct amd_northbridge *nb = NULL;
1180 	struct threshold_bank *b = NULL;
1181 	const char *name = get_name(cpu, bank, NULL);
1182 	int err = 0;
1183 
1184 	if (!dev)
1185 		return -ENODEV;
1186 
1187 	if (is_shared_bank(bank)) {
1188 		nb = node_to_amd_nb(topology_die_id(cpu));
1189 
1190 		/* threshold descriptor already initialized on this node? */
1191 		if (nb && nb->bank4) {
1192 			/* yes, use it */
1193 			b = nb->bank4;
1194 			err = kobject_add(b->kobj, &dev->kobj, name);
1195 			if (err)
1196 				goto out;
1197 
1198 			bp[bank] = b;
1199 			refcount_inc(&b->cpus);
1200 
1201 			err = __threshold_add_blocks(b);
1202 
1203 			goto out;
1204 		}
1205 	}
1206 
1207 	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1208 	if (!b) {
1209 		err = -ENOMEM;
1210 		goto out;
1211 	}
1212 
1213 	/* Associate the bank with the per-CPU MCE device */
1214 	b->kobj = kobject_create_and_add(name, &dev->kobj);
1215 	if (!b->kobj) {
1216 		err = -EINVAL;
1217 		goto out_free;
1218 	}
1219 
1220 	if (is_shared_bank(bank)) {
1221 		b->shared = 1;
1222 		refcount_set(&b->cpus, 1);
1223 
1224 		/* nb is already initialized, see above */
1225 		if (nb) {
1226 			WARN_ON(nb->bank4);
1227 			nb->bank4 = b;
1228 		}
1229 	}
1230 
1231 	err = allocate_threshold_blocks(cpu, b, bank, 0, mca_msr_reg(bank, MCA_MISC));
1232 	if (err)
1233 		goto out_kobj;
1234 
1235 	bp[bank] = b;
1236 	return 0;
1237 
1238 out_kobj:
1239 	kobject_put(b->kobj);
1240 out_free:
1241 	kfree(b);
1242 out:
1243 	return err;
1244 }
1245 
1246 static void threshold_block_release(struct kobject *kobj)
1247 {
1248 	kfree(to_block(kobj));
1249 }
1250 
1251 static void deallocate_threshold_blocks(struct threshold_bank *bank)
1252 {
1253 	struct threshold_block *pos, *tmp;
1254 
1255 	list_for_each_entry_safe(pos, tmp, &bank->blocks->miscj, miscj) {
1256 		list_del(&pos->miscj);
1257 		kobject_put(&pos->kobj);
1258 	}
1259 
1260 	kobject_put(&bank->blocks->kobj);
1261 }
1262 
1263 static void __threshold_remove_blocks(struct threshold_bank *b)
1264 {
1265 	struct threshold_block *pos = NULL;
1266 	struct threshold_block *tmp = NULL;
1267 
1268 	kobject_del(b->kobj);
1269 
1270 	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1271 		kobject_del(&pos->kobj);
1272 }
1273 
1274 static void threshold_remove_bank(struct threshold_bank *bank)
1275 {
1276 	struct amd_northbridge *nb;
1277 
1278 	if (!bank->blocks)
1279 		goto out_free;
1280 
1281 	if (!bank->shared)
1282 		goto out_dealloc;
1283 
1284 	if (!refcount_dec_and_test(&bank->cpus)) {
1285 		__threshold_remove_blocks(bank);
1286 		return;
1287 	} else {
1288 		/*
1289 		 * The last CPU on this node using the shared bank is going
1290 		 * away, remove that bank now.
1291 		 */
1292 		nb = node_to_amd_nb(topology_die_id(smp_processor_id()));
1293 		nb->bank4 = NULL;
1294 	}
1295 
1296 out_dealloc:
1297 	deallocate_threshold_blocks(bank);
1298 
1299 out_free:
1300 	kobject_put(bank->kobj);
1301 	kfree(bank);
1302 }
1303 
1304 static void __threshold_remove_device(struct threshold_bank **bp)
1305 {
1306 	unsigned int bank, numbanks = this_cpu_read(mce_num_banks);
1307 
1308 	for (bank = 0; bank < numbanks; bank++) {
1309 		if (!bp[bank])
1310 			continue;
1311 
1312 		threshold_remove_bank(bp[bank]);
1313 		bp[bank] = NULL;
1314 	}
1315 	kfree(bp);
1316 }
1317 
1318 int mce_threshold_remove_device(unsigned int cpu)
1319 {
1320 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
1321 
1322 	if (!bp)
1323 		return 0;
1324 
1325 	/*
1326 	 * Clear the pointer before cleaning up, so that the interrupt won't
1327 	 * touch anything of this.
1328 	 */
1329 	this_cpu_write(threshold_banks, NULL);
1330 
1331 	__threshold_remove_device(bp);
1332 	return 0;
1333 }
1334 
1335 /**
1336  * mce_threshold_create_device - Create the per-CPU MCE threshold device
1337  * @cpu:	The plugged in CPU
1338  *
1339  * Create directories and files for all valid threshold banks.
1340  *
1341  * This is invoked from the CPU hotplug callback which was installed in
1342  * mcheck_init_device(). The invocation happens in context of the hotplug
1343  * thread running on @cpu.  The callback is invoked on all CPUs which are
1344  * online when the callback is installed or during a real hotplug event.
1345  */
1346 int mce_threshold_create_device(unsigned int cpu)
1347 {
1348 	unsigned int numbanks, bank;
1349 	struct threshold_bank **bp;
1350 	int err;
1351 
1352 	if (!mce_flags.amd_threshold)
1353 		return 0;
1354 
1355 	bp = this_cpu_read(threshold_banks);
1356 	if (bp)
1357 		return 0;
1358 
1359 	numbanks = this_cpu_read(mce_num_banks);
1360 	bp = kcalloc(numbanks, sizeof(*bp), GFP_KERNEL);
1361 	if (!bp)
1362 		return -ENOMEM;
1363 
1364 	for (bank = 0; bank < numbanks; ++bank) {
1365 		if (!(this_cpu_read(bank_map) & (1 << bank)))
1366 			continue;
1367 		err = threshold_create_bank(bp, cpu, bank);
1368 		if (err) {
1369 			__threshold_remove_device(bp);
1370 			return err;
1371 		}
1372 	}
1373 	this_cpu_write(threshold_banks, bp);
1374 
1375 	if (thresholding_irq_en)
1376 		mce_threshold_vector = amd_threshold_interrupt;
1377 	return 0;
1378 }
1379