blob: 858fe603722361f4292052952a738f3f9cc4a03e [file] [log] [blame]
Doug Thompson2bc65412009-05-04 20:11:14 +02001#include "amd64_edac.h"
Doug Thompson7d6034d2009-04-27 20:01:01 +02002#include <asm/k8.h>
Doug Thompson2bc65412009-05-04 20:11:14 +02003
4static struct edac_pci_ctl_info *amd64_ctl_pci;
5
6static int report_gart_errors;
7module_param(report_gart_errors, int, 0644);
8
9/*
10 * Set by command line parameter. If BIOS has enabled the ECC, this override is
11 * cleared to prevent re-enabling the hardware by this driver.
12 */
13static int ecc_enable_override;
14module_param(ecc_enable_override, int, 0644);
15
16/* Lookup table for all possible MC control instances */
17struct amd64_pvt;
18static struct mem_ctl_info *mci_lookup[MAX_NUMNODES];
19static struct amd64_pvt *pvt_lookup[MAX_NUMNODES];
20
21/*
22 * Memory scrubber control interface. For K8, memory scrubbing is handled by
23 * hardware and can involve L2 cache, dcache as well as the main memory. With
24 * F10, this is extended to L3 cache scrubbing on CPU models sporting that
25 * functionality.
26 *
27 * This causes the "units" for the scrubbing speed to vary from 64 byte blocks
28 * (dram) over to cache lines. This is nasty, so we will use bandwidth in
29 * bytes/sec for the setting.
30 *
31 * Currently, we only do dram scrubbing. If the scrubbing is done in software on
32 * other archs, we might not have access to the caches directly.
33 */
34
35/*
36 * scan the scrub rate mapping table for a close or matching bandwidth value to
37 * issue. If requested is too big, then use last maximum value found.
38 */
39static int amd64_search_set_scrub_rate(struct pci_dev *ctl, u32 new_bw,
40 u32 min_scrubrate)
41{
42 u32 scrubval;
43 int i;
44
45 /*
46 * map the configured rate (new_bw) to a value specific to the AMD64
47 * memory controller and apply to register. Search for the first
48 * bandwidth entry that is greater or equal than the setting requested
49 * and program that. If at last entry, turn off DRAM scrubbing.
50 */
51 for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
52 /*
53 * skip scrub rates which aren't recommended
54 * (see F10 BKDG, F3x58)
55 */
56 if (scrubrates[i].scrubval < min_scrubrate)
57 continue;
58
59 if (scrubrates[i].bandwidth <= new_bw)
60 break;
61
62 /*
63 * if no suitable bandwidth found, turn off DRAM scrubbing
64 * entirely by falling back to the last element in the
65 * scrubrates array.
66 */
67 }
68
69 scrubval = scrubrates[i].scrubval;
70 if (scrubval)
71 edac_printk(KERN_DEBUG, EDAC_MC,
72 "Setting scrub rate bandwidth: %u\n",
73 scrubrates[i].bandwidth);
74 else
75 edac_printk(KERN_DEBUG, EDAC_MC, "Turning scrubbing off.\n");
76
77 pci_write_bits32(ctl, K8_SCRCTRL, scrubval, 0x001F);
78
79 return 0;
80}
81
82static int amd64_set_scrub_rate(struct mem_ctl_info *mci, u32 *bandwidth)
83{
84 struct amd64_pvt *pvt = mci->pvt_info;
85 u32 min_scrubrate = 0x0;
86
87 switch (boot_cpu_data.x86) {
88 case 0xf:
89 min_scrubrate = K8_MIN_SCRUB_RATE_BITS;
90 break;
91 case 0x10:
92 min_scrubrate = F10_MIN_SCRUB_RATE_BITS;
93 break;
94 case 0x11:
95 min_scrubrate = F11_MIN_SCRUB_RATE_BITS;
96 break;
97
98 default:
99 amd64_printk(KERN_ERR, "Unsupported family!\n");
100 break;
101 }
102 return amd64_search_set_scrub_rate(pvt->misc_f3_ctl, *bandwidth,
103 min_scrubrate);
104}
105
106static int amd64_get_scrub_rate(struct mem_ctl_info *mci, u32 *bw)
107{
108 struct amd64_pvt *pvt = mci->pvt_info;
109 u32 scrubval = 0;
110 int status = -1, i, ret = 0;
111
112 ret = pci_read_config_dword(pvt->misc_f3_ctl, K8_SCRCTRL, &scrubval);
113 if (ret)
114 debugf0("Reading K8_SCRCTRL failed\n");
115
116 scrubval = scrubval & 0x001F;
117
118 edac_printk(KERN_DEBUG, EDAC_MC,
119 "pci-read, sdram scrub control value: %d \n", scrubval);
120
121 for (i = 0; ARRAY_SIZE(scrubrates); i++) {
122 if (scrubrates[i].scrubval == scrubval) {
123 *bw = scrubrates[i].bandwidth;
124 status = 0;
125 break;
126 }
127 }
128
129 return status;
130}
131
Doug Thompson67757632009-04-27 15:53:22 +0200132/* Map from a CSROW entry to the mask entry that operates on it */
133static inline u32 amd64_map_to_dcs_mask(struct amd64_pvt *pvt, int csrow)
134{
135 return csrow >> (pvt->num_dcsm >> 3);
136}
137
138/* return the 'base' address the i'th CS entry of the 'dct' DRAM controller */
139static u32 amd64_get_dct_base(struct amd64_pvt *pvt, int dct, int csrow)
140{
141 if (dct == 0)
142 return pvt->dcsb0[csrow];
143 else
144 return pvt->dcsb1[csrow];
145}
146
147/*
148 * Return the 'mask' address the i'th CS entry. This function is needed because
149 * there number of DCSM registers on Rev E and prior vs Rev F and later is
150 * different.
151 */
152static u32 amd64_get_dct_mask(struct amd64_pvt *pvt, int dct, int csrow)
153{
154 if (dct == 0)
155 return pvt->dcsm0[amd64_map_to_dcs_mask(pvt, csrow)];
156 else
157 return pvt->dcsm1[amd64_map_to_dcs_mask(pvt, csrow)];
158}
159
160
161/*
162 * In *base and *limit, pass back the full 40-bit base and limit physical
163 * addresses for the node given by node_id. This information is obtained from
164 * DRAM Base (section 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers. The
165 * base and limit addresses are of type SysAddr, as defined at the start of
166 * section 3.4.4 (p. 70). They are the lowest and highest physical addresses
167 * in the address range they represent.
168 */
169static void amd64_get_base_and_limit(struct amd64_pvt *pvt, int node_id,
170 u64 *base, u64 *limit)
171{
172 *base = pvt->dram_base[node_id];
173 *limit = pvt->dram_limit[node_id];
174}
175
176/*
177 * Return 1 if the SysAddr given by sys_addr matches the base/limit associated
178 * with node_id
179 */
180static int amd64_base_limit_match(struct amd64_pvt *pvt,
181 u64 sys_addr, int node_id)
182{
183 u64 base, limit, addr;
184
185 amd64_get_base_and_limit(pvt, node_id, &base, &limit);
186
187 /* The K8 treats this as a 40-bit value. However, bits 63-40 will be
188 * all ones if the most significant implemented address bit is 1.
189 * Here we discard bits 63-40. See section 3.4.2 of AMD publication
190 * 24592: AMD x86-64 Architecture Programmer's Manual Volume 1
191 * Application Programming.
192 */
193 addr = sys_addr & 0x000000ffffffffffull;
194
195 return (addr >= base) && (addr <= limit);
196}
197
198/*
199 * Attempt to map a SysAddr to a node. On success, return a pointer to the
200 * mem_ctl_info structure for the node that the SysAddr maps to.
201 *
202 * On failure, return NULL.
203 */
204static struct mem_ctl_info *find_mc_by_sys_addr(struct mem_ctl_info *mci,
205 u64 sys_addr)
206{
207 struct amd64_pvt *pvt;
208 int node_id;
209 u32 intlv_en, bits;
210
211 /*
212 * Here we use the DRAM Base (section 3.4.4.1) and DRAM Limit (section
213 * 3.4.4.2) registers to map the SysAddr to a node ID.
214 */
215 pvt = mci->pvt_info;
216
217 /*
218 * The value of this field should be the same for all DRAM Base
219 * registers. Therefore we arbitrarily choose to read it from the
220 * register for node 0.
221 */
222 intlv_en = pvt->dram_IntlvEn[0];
223
224 if (intlv_en == 0) {
225 for (node_id = 0; ; ) {
226 if (amd64_base_limit_match(pvt, sys_addr, node_id))
227 break;
228
229 if (++node_id >= DRAM_REG_COUNT)
230 goto err_no_match;
231 }
232 goto found;
233 }
234
235 if (unlikely((intlv_en != (0x01 << 8)) &&
236 (intlv_en != (0x03 << 8)) &&
237 (intlv_en != (0x07 << 8)))) {
238 amd64_printk(KERN_WARNING, "junk value of 0x%x extracted from "
239 "IntlvEn field of DRAM Base Register for node 0: "
240 "This probably indicates a BIOS bug.\n", intlv_en);
241 return NULL;
242 }
243
244 bits = (((u32) sys_addr) >> 12) & intlv_en;
245
246 for (node_id = 0; ; ) {
247 if ((pvt->dram_limit[node_id] & intlv_en) == bits)
248 break; /* intlv_sel field matches */
249
250 if (++node_id >= DRAM_REG_COUNT)
251 goto err_no_match;
252 }
253
254 /* sanity test for sys_addr */
255 if (unlikely(!amd64_base_limit_match(pvt, sys_addr, node_id))) {
256 amd64_printk(KERN_WARNING,
257 "%s(): sys_addr 0x%lx falls outside base/limit "
258 "address range for node %d with node interleaving "
259 "enabled.\n", __func__, (unsigned long)sys_addr,
260 node_id);
261 return NULL;
262 }
263
264found:
265 return edac_mc_find(node_id);
266
267err_no_match:
268 debugf2("sys_addr 0x%lx doesn't match any node\n",
269 (unsigned long)sys_addr);
270
271 return NULL;
272}
Doug Thompsone2ce7252009-04-27 15:57:12 +0200273
274/*
275 * Extract the DRAM CS base address from selected csrow register.
276 */
277static u64 base_from_dct_base(struct amd64_pvt *pvt, int csrow)
278{
279 return ((u64) (amd64_get_dct_base(pvt, 0, csrow) & pvt->dcsb_base)) <<
280 pvt->dcs_shift;
281}
282
283/*
284 * Extract the mask from the dcsb0[csrow] entry in a CPU revision-specific way.
285 */
286static u64 mask_from_dct_mask(struct amd64_pvt *pvt, int csrow)
287{
288 u64 dcsm_bits, other_bits;
289 u64 mask;
290
291 /* Extract bits from DRAM CS Mask. */
292 dcsm_bits = amd64_get_dct_mask(pvt, 0, csrow) & pvt->dcsm_mask;
293
294 other_bits = pvt->dcsm_mask;
295 other_bits = ~(other_bits << pvt->dcs_shift);
296
297 /*
298 * The extracted bits from DCSM belong in the spaces represented by
299 * the cleared bits in other_bits.
300 */
301 mask = (dcsm_bits << pvt->dcs_shift) | other_bits;
302
303 return mask;
304}
305
306/*
307 * @input_addr is an InputAddr associated with the node given by mci. Return the
308 * csrow that input_addr maps to, or -1 on failure (no csrow claims input_addr).
309 */
310static int input_addr_to_csrow(struct mem_ctl_info *mci, u64 input_addr)
311{
312 struct amd64_pvt *pvt;
313 int csrow;
314 u64 base, mask;
315
316 pvt = mci->pvt_info;
317
318 /*
319 * Here we use the DRAM CS Base and DRAM CS Mask registers. For each CS
320 * base/mask register pair, test the condition shown near the start of
321 * section 3.5.4 (p. 84, BKDG #26094, K8, revA-E).
322 */
323 for (csrow = 0; csrow < CHIPSELECT_COUNT; csrow++) {
324
325 /* This DRAM chip select is disabled on this node */
326 if ((pvt->dcsb0[csrow] & K8_DCSB_CS_ENABLE) == 0)
327 continue;
328
329 base = base_from_dct_base(pvt, csrow);
330 mask = ~mask_from_dct_mask(pvt, csrow);
331
332 if ((input_addr & mask) == (base & mask)) {
333 debugf2("InputAddr 0x%lx matches csrow %d (node %d)\n",
334 (unsigned long)input_addr, csrow,
335 pvt->mc_node_id);
336
337 return csrow;
338 }
339 }
340
341 debugf2("no matching csrow for InputAddr 0x%lx (MC node %d)\n",
342 (unsigned long)input_addr, pvt->mc_node_id);
343
344 return -1;
345}
346
347/*
348 * Return the base value defined by the DRAM Base register for the node
349 * represented by mci. This function returns the full 40-bit value despite the
350 * fact that the register only stores bits 39-24 of the value. See section
351 * 3.4.4.1 (BKDG #26094, K8, revA-E)
352 */
353static inline u64 get_dram_base(struct mem_ctl_info *mci)
354{
355 struct amd64_pvt *pvt = mci->pvt_info;
356
357 return pvt->dram_base[pvt->mc_node_id];
358}
359
360/*
361 * Obtain info from the DRAM Hole Address Register (section 3.4.8, pub #26094)
362 * for the node represented by mci. Info is passed back in *hole_base,
363 * *hole_offset, and *hole_size. Function returns 0 if info is valid or 1 if
364 * info is invalid. Info may be invalid for either of the following reasons:
365 *
366 * - The revision of the node is not E or greater. In this case, the DRAM Hole
367 * Address Register does not exist.
368 *
369 * - The DramHoleValid bit is cleared in the DRAM Hole Address Register,
370 * indicating that its contents are not valid.
371 *
372 * The values passed back in *hole_base, *hole_offset, and *hole_size are
373 * complete 32-bit values despite the fact that the bitfields in the DHAR
374 * only represent bits 31-24 of the base and offset values.
375 */
376int amd64_get_dram_hole_info(struct mem_ctl_info *mci, u64 *hole_base,
377 u64 *hole_offset, u64 *hole_size)
378{
379 struct amd64_pvt *pvt = mci->pvt_info;
380 u64 base;
381
382 /* only revE and later have the DRAM Hole Address Register */
383 if (boot_cpu_data.x86 == 0xf && pvt->ext_model < OPTERON_CPU_REV_E) {
384 debugf1(" revision %d for node %d does not support DHAR\n",
385 pvt->ext_model, pvt->mc_node_id);
386 return 1;
387 }
388
389 /* only valid for Fam10h */
390 if (boot_cpu_data.x86 == 0x10 &&
391 (pvt->dhar & F10_DRAM_MEM_HOIST_VALID) == 0) {
392 debugf1(" Dram Memory Hoisting is DISABLED on this system\n");
393 return 1;
394 }
395
396 if ((pvt->dhar & DHAR_VALID) == 0) {
397 debugf1(" Dram Memory Hoisting is DISABLED on this node %d\n",
398 pvt->mc_node_id);
399 return 1;
400 }
401
402 /* This node has Memory Hoisting */
403
404 /* +------------------+--------------------+--------------------+-----
405 * | memory | DRAM hole | relocated |
406 * | [0, (x - 1)] | [x, 0xffffffff] | addresses from |
407 * | | | DRAM hole |
408 * | | | [0x100000000, |
409 * | | | (0x100000000+ |
410 * | | | (0xffffffff-x))] |
411 * +------------------+--------------------+--------------------+-----
412 *
413 * Above is a diagram of physical memory showing the DRAM hole and the
414 * relocated addresses from the DRAM hole. As shown, the DRAM hole
415 * starts at address x (the base address) and extends through address
416 * 0xffffffff. The DRAM Hole Address Register (DHAR) relocates the
417 * addresses in the hole so that they start at 0x100000000.
418 */
419
420 base = dhar_base(pvt->dhar);
421
422 *hole_base = base;
423 *hole_size = (0x1ull << 32) - base;
424
425 if (boot_cpu_data.x86 > 0xf)
426 *hole_offset = f10_dhar_offset(pvt->dhar);
427 else
428 *hole_offset = k8_dhar_offset(pvt->dhar);
429
430 debugf1(" DHAR info for node %d base 0x%lx offset 0x%lx size 0x%lx\n",
431 pvt->mc_node_id, (unsigned long)*hole_base,
432 (unsigned long)*hole_offset, (unsigned long)*hole_size);
433
434 return 0;
435}
436EXPORT_SYMBOL_GPL(amd64_get_dram_hole_info);
437
Doug Thompson93c2df52009-05-04 20:46:50 +0200438/*
439 * Return the DramAddr that the SysAddr given by @sys_addr maps to. It is
440 * assumed that sys_addr maps to the node given by mci.
441 *
442 * The first part of section 3.4.4 (p. 70) shows how the DRAM Base (section
443 * 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers are used to translate a
444 * SysAddr to a DramAddr. If the DRAM Hole Address Register (DHAR) is enabled,
445 * then it is also involved in translating a SysAddr to a DramAddr. Sections
446 * 3.4.8 and 3.5.8.2 describe the DHAR and how it is used for memory hoisting.
447 * These parts of the documentation are unclear. I interpret them as follows:
448 *
449 * When node n receives a SysAddr, it processes the SysAddr as follows:
450 *
451 * 1. It extracts the DRAMBase and DRAMLimit values from the DRAM Base and DRAM
452 * Limit registers for node n. If the SysAddr is not within the range
453 * specified by the base and limit values, then node n ignores the Sysaddr
454 * (since it does not map to node n). Otherwise continue to step 2 below.
455 *
456 * 2. If the DramHoleValid bit of the DHAR for node n is clear, the DHAR is
457 * disabled so skip to step 3 below. Otherwise see if the SysAddr is within
458 * the range of relocated addresses (starting at 0x100000000) from the DRAM
459 * hole. If not, skip to step 3 below. Else get the value of the
460 * DramHoleOffset field from the DHAR. To obtain the DramAddr, subtract the
461 * offset defined by this value from the SysAddr.
462 *
463 * 3. Obtain the base address for node n from the DRAMBase field of the DRAM
464 * Base register for node n. To obtain the DramAddr, subtract the base
465 * address from the SysAddr, as shown near the start of section 3.4.4 (p.70).
466 */
467static u64 sys_addr_to_dram_addr(struct mem_ctl_info *mci, u64 sys_addr)
468{
469 u64 dram_base, hole_base, hole_offset, hole_size, dram_addr;
470 int ret = 0;
471
472 dram_base = get_dram_base(mci);
473
474 ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
475 &hole_size);
476 if (!ret) {
477 if ((sys_addr >= (1ull << 32)) &&
478 (sys_addr < ((1ull << 32) + hole_size))) {
479 /* use DHAR to translate SysAddr to DramAddr */
480 dram_addr = sys_addr - hole_offset;
481
482 debugf2("using DHAR to translate SysAddr 0x%lx to "
483 "DramAddr 0x%lx\n",
484 (unsigned long)sys_addr,
485 (unsigned long)dram_addr);
486
487 return dram_addr;
488 }
489 }
490
491 /*
492 * Translate the SysAddr to a DramAddr as shown near the start of
493 * section 3.4.4 (p. 70). Although sys_addr is a 64-bit value, the k8
494 * only deals with 40-bit values. Therefore we discard bits 63-40 of
495 * sys_addr below. If bit 39 of sys_addr is 1 then the bits we
496 * discard are all 1s. Otherwise the bits we discard are all 0s. See
497 * section 3.4.2 of AMD publication 24592: AMD x86-64 Architecture
498 * Programmer's Manual Volume 1 Application Programming.
499 */
500 dram_addr = (sys_addr & 0xffffffffffull) - dram_base;
501
502 debugf2("using DRAM Base register to translate SysAddr 0x%lx to "
503 "DramAddr 0x%lx\n", (unsigned long)sys_addr,
504 (unsigned long)dram_addr);
505 return dram_addr;
506}
507
508/*
509 * @intlv_en is the value of the IntlvEn field from a DRAM Base register
510 * (section 3.4.4.1). Return the number of bits from a SysAddr that are used
511 * for node interleaving.
512 */
513static int num_node_interleave_bits(unsigned intlv_en)
514{
515 static const int intlv_shift_table[] = { 0, 1, 0, 2, 0, 0, 0, 3 };
516 int n;
517
518 BUG_ON(intlv_en > 7);
519 n = intlv_shift_table[intlv_en];
520 return n;
521}
522
523/* Translate the DramAddr given by @dram_addr to an InputAddr. */
524static u64 dram_addr_to_input_addr(struct mem_ctl_info *mci, u64 dram_addr)
525{
526 struct amd64_pvt *pvt;
527 int intlv_shift;
528 u64 input_addr;
529
530 pvt = mci->pvt_info;
531
532 /*
533 * See the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
534 * concerning translating a DramAddr to an InputAddr.
535 */
536 intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
537 input_addr = ((dram_addr >> intlv_shift) & 0xffffff000ull) +
538 (dram_addr & 0xfff);
539
540 debugf2(" Intlv Shift=%d DramAddr=0x%lx maps to InputAddr=0x%lx\n",
541 intlv_shift, (unsigned long)dram_addr,
542 (unsigned long)input_addr);
543
544 return input_addr;
545}
546
547/*
548 * Translate the SysAddr represented by @sys_addr to an InputAddr. It is
549 * assumed that @sys_addr maps to the node given by mci.
550 */
551static u64 sys_addr_to_input_addr(struct mem_ctl_info *mci, u64 sys_addr)
552{
553 u64 input_addr;
554
555 input_addr =
556 dram_addr_to_input_addr(mci, sys_addr_to_dram_addr(mci, sys_addr));
557
558 debugf2("SysAdddr 0x%lx translates to InputAddr 0x%lx\n",
559 (unsigned long)sys_addr, (unsigned long)input_addr);
560
561 return input_addr;
562}
563
564
565/*
566 * @input_addr is an InputAddr associated with the node represented by mci.
567 * Translate @input_addr to a DramAddr and return the result.
568 */
569static u64 input_addr_to_dram_addr(struct mem_ctl_info *mci, u64 input_addr)
570{
571 struct amd64_pvt *pvt;
572 int node_id, intlv_shift;
573 u64 bits, dram_addr;
574 u32 intlv_sel;
575
576 /*
577 * Near the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
578 * shows how to translate a DramAddr to an InputAddr. Here we reverse
579 * this procedure. When translating from a DramAddr to an InputAddr, the
580 * bits used for node interleaving are discarded. Here we recover these
581 * bits from the IntlvSel field of the DRAM Limit register (section
582 * 3.4.4.2) for the node that input_addr is associated with.
583 */
584 pvt = mci->pvt_info;
585 node_id = pvt->mc_node_id;
586 BUG_ON((node_id < 0) || (node_id > 7));
587
588 intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
589
590 if (intlv_shift == 0) {
591 debugf1(" InputAddr 0x%lx translates to DramAddr of "
592 "same value\n", (unsigned long)input_addr);
593
594 return input_addr;
595 }
596
597 bits = ((input_addr & 0xffffff000ull) << intlv_shift) +
598 (input_addr & 0xfff);
599
600 intlv_sel = pvt->dram_IntlvSel[node_id] & ((1 << intlv_shift) - 1);
601 dram_addr = bits + (intlv_sel << 12);
602
603 debugf1("InputAddr 0x%lx translates to DramAddr 0x%lx "
604 "(%d node interleave bits)\n", (unsigned long)input_addr,
605 (unsigned long)dram_addr, intlv_shift);
606
607 return dram_addr;
608}
609
610/*
611 * @dram_addr is a DramAddr that maps to the node represented by mci. Convert
612 * @dram_addr to a SysAddr.
613 */
614static u64 dram_addr_to_sys_addr(struct mem_ctl_info *mci, u64 dram_addr)
615{
616 struct amd64_pvt *pvt = mci->pvt_info;
617 u64 hole_base, hole_offset, hole_size, base, limit, sys_addr;
618 int ret = 0;
619
620 ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
621 &hole_size);
622 if (!ret) {
623 if ((dram_addr >= hole_base) &&
624 (dram_addr < (hole_base + hole_size))) {
625 sys_addr = dram_addr + hole_offset;
626
627 debugf1("using DHAR to translate DramAddr 0x%lx to "
628 "SysAddr 0x%lx\n", (unsigned long)dram_addr,
629 (unsigned long)sys_addr);
630
631 return sys_addr;
632 }
633 }
634
635 amd64_get_base_and_limit(pvt, pvt->mc_node_id, &base, &limit);
636 sys_addr = dram_addr + base;
637
638 /*
639 * The sys_addr we have computed up to this point is a 40-bit value
640 * because the k8 deals with 40-bit values. However, the value we are
641 * supposed to return is a full 64-bit physical address. The AMD
642 * x86-64 architecture specifies that the most significant implemented
643 * address bit through bit 63 of a physical address must be either all
644 * 0s or all 1s. Therefore we sign-extend the 40-bit sys_addr to a
645 * 64-bit value below. See section 3.4.2 of AMD publication 24592:
646 * AMD x86-64 Architecture Programmer's Manual Volume 1 Application
647 * Programming.
648 */
649 sys_addr |= ~((sys_addr & (1ull << 39)) - 1);
650
651 debugf1(" Node %d, DramAddr 0x%lx to SysAddr 0x%lx\n",
652 pvt->mc_node_id, (unsigned long)dram_addr,
653 (unsigned long)sys_addr);
654
655 return sys_addr;
656}
657
658/*
659 * @input_addr is an InputAddr associated with the node given by mci. Translate
660 * @input_addr to a SysAddr.
661 */
662static inline u64 input_addr_to_sys_addr(struct mem_ctl_info *mci,
663 u64 input_addr)
664{
665 return dram_addr_to_sys_addr(mci,
666 input_addr_to_dram_addr(mci, input_addr));
667}
668
669/*
670 * Find the minimum and maximum InputAddr values that map to the given @csrow.
671 * Pass back these values in *input_addr_min and *input_addr_max.
672 */
673static void find_csrow_limits(struct mem_ctl_info *mci, int csrow,
674 u64 *input_addr_min, u64 *input_addr_max)
675{
676 struct amd64_pvt *pvt;
677 u64 base, mask;
678
679 pvt = mci->pvt_info;
680 BUG_ON((csrow < 0) || (csrow >= CHIPSELECT_COUNT));
681
682 base = base_from_dct_base(pvt, csrow);
683 mask = mask_from_dct_mask(pvt, csrow);
684
685 *input_addr_min = base & ~mask;
686 *input_addr_max = base | mask | pvt->dcs_mask_notused;
687}
688
689/*
690 * Extract error address from MCA NB Address Low (section 3.6.4.5) and MCA NB
691 * Address High (section 3.6.4.6) register values and return the result. Address
692 * is located in the info structure (nbeah and nbeal), the encoding is device
693 * specific.
694 */
695static u64 extract_error_address(struct mem_ctl_info *mci,
696 struct amd64_error_info_regs *info)
697{
698 struct amd64_pvt *pvt = mci->pvt_info;
699
700 return pvt->ops->get_error_address(mci, info);
701}
702
703
704/* Map the Error address to a PAGE and PAGE OFFSET. */
705static inline void error_address_to_page_and_offset(u64 error_address,
706 u32 *page, u32 *offset)
707{
708 *page = (u32) (error_address >> PAGE_SHIFT);
709 *offset = ((u32) error_address) & ~PAGE_MASK;
710}
711
712/*
713 * @sys_addr is an error address (a SysAddr) extracted from the MCA NB Address
714 * Low (section 3.6.4.5) and MCA NB Address High (section 3.6.4.6) registers
715 * of a node that detected an ECC memory error. mci represents the node that
716 * the error address maps to (possibly different from the node that detected
717 * the error). Return the number of the csrow that sys_addr maps to, or -1 on
718 * error.
719 */
720static int sys_addr_to_csrow(struct mem_ctl_info *mci, u64 sys_addr)
721{
722 int csrow;
723
724 csrow = input_addr_to_csrow(mci, sys_addr_to_input_addr(mci, sys_addr));
725
726 if (csrow == -1)
727 amd64_mc_printk(mci, KERN_ERR,
728 "Failed to translate InputAddr to csrow for "
729 "address 0x%lx\n", (unsigned long)sys_addr);
730 return csrow;
731}
Doug Thompsone2ce7252009-04-27 15:57:12 +0200732
Doug Thompson2da11652009-04-27 16:09:09 +0200733static int get_channel_from_ecc_syndrome(unsigned short syndrome);
734
735static void amd64_cpu_display_info(struct amd64_pvt *pvt)
736{
737 if (boot_cpu_data.x86 == 0x11)
738 edac_printk(KERN_DEBUG, EDAC_MC, "F11h CPU detected\n");
739 else if (boot_cpu_data.x86 == 0x10)
740 edac_printk(KERN_DEBUG, EDAC_MC, "F10h CPU detected\n");
741 else if (boot_cpu_data.x86 == 0xf)
742 edac_printk(KERN_DEBUG, EDAC_MC, "%s detected\n",
743 (pvt->ext_model >= OPTERON_CPU_REV_F) ?
744 "Rev F or later" : "Rev E or earlier");
745 else
746 /* we'll hardly ever ever get here */
747 edac_printk(KERN_ERR, EDAC_MC, "Unknown cpu!\n");
748}
749
750/*
751 * Determine if the DIMMs have ECC enabled. ECC is enabled ONLY if all the DIMMs
752 * are ECC capable.
753 */
754static enum edac_type amd64_determine_edac_cap(struct amd64_pvt *pvt)
755{
756 int bit;
Borislav Petkov584fcff2009-06-10 18:29:54 +0200757 enum dev_type edac_cap = EDAC_FLAG_NONE;
Doug Thompson2da11652009-04-27 16:09:09 +0200758
759 bit = (boot_cpu_data.x86 > 0xf || pvt->ext_model >= OPTERON_CPU_REV_F)
760 ? 19
761 : 17;
762
Borislav Petkov584fcff2009-06-10 18:29:54 +0200763 if (pvt->dclr0 & BIT(bit))
Doug Thompson2da11652009-04-27 16:09:09 +0200764 edac_cap = EDAC_FLAG_SECDED;
765
766 return edac_cap;
767}
768
769
770static void f10_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt,
771 int ganged);
772
773/* Display and decode various NB registers for debug purposes. */
774static void amd64_dump_misc_regs(struct amd64_pvt *pvt)
775{
776 int ganged;
777
778 debugf1(" nbcap:0x%8.08x DctDualCap=%s DualNode=%s 8-Node=%s\n",
779 pvt->nbcap,
780 (pvt->nbcap & K8_NBCAP_DCT_DUAL) ? "True" : "False",
781 (pvt->nbcap & K8_NBCAP_DUAL_NODE) ? "True" : "False",
782 (pvt->nbcap & K8_NBCAP_8_NODE) ? "True" : "False");
783 debugf1(" ECC Capable=%s ChipKill Capable=%s\n",
784 (pvt->nbcap & K8_NBCAP_SECDED) ? "True" : "False",
785 (pvt->nbcap & K8_NBCAP_CHIPKILL) ? "True" : "False");
786 debugf1(" DramCfg0-low=0x%08x DIMM-ECC=%s Parity=%s Width=%s\n",
787 pvt->dclr0,
788 (pvt->dclr0 & BIT(19)) ? "Enabled" : "Disabled",
789 (pvt->dclr0 & BIT(8)) ? "Enabled" : "Disabled",
790 (pvt->dclr0 & BIT(11)) ? "128b" : "64b");
791 debugf1(" DIMM x4 Present: L0=%s L1=%s L2=%s L3=%s DIMM Type=%s\n",
792 (pvt->dclr0 & BIT(12)) ? "Y" : "N",
793 (pvt->dclr0 & BIT(13)) ? "Y" : "N",
794 (pvt->dclr0 & BIT(14)) ? "Y" : "N",
795 (pvt->dclr0 & BIT(15)) ? "Y" : "N",
796 (pvt->dclr0 & BIT(16)) ? "UN-Buffered" : "Buffered");
797
798
799 debugf1(" online-spare: 0x%8.08x\n", pvt->online_spare);
800
801 if (boot_cpu_data.x86 == 0xf) {
802 debugf1(" dhar: 0x%8.08x Base=0x%08x Offset=0x%08x\n",
803 pvt->dhar, dhar_base(pvt->dhar),
804 k8_dhar_offset(pvt->dhar));
805 debugf1(" DramHoleValid=%s\n",
806 (pvt->dhar & DHAR_VALID) ? "True" : "False");
807
808 debugf1(" dbam-dkt: 0x%8.08x\n", pvt->dbam0);
809
810 /* everything below this point is Fam10h and above */
811 return;
812
813 } else {
814 debugf1(" dhar: 0x%8.08x Base=0x%08x Offset=0x%08x\n",
815 pvt->dhar, dhar_base(pvt->dhar),
816 f10_dhar_offset(pvt->dhar));
817 debugf1(" DramMemHoistValid=%s DramHoleValid=%s\n",
818 (pvt->dhar & F10_DRAM_MEM_HOIST_VALID) ?
819 "True" : "False",
820 (pvt->dhar & DHAR_VALID) ?
821 "True" : "False");
822 }
823
824 /* Only if NOT ganged does dcl1 have valid info */
825 if (!dct_ganging_enabled(pvt)) {
826 debugf1(" DramCfg1-low=0x%08x DIMM-ECC=%s Parity=%s "
827 "Width=%s\n", pvt->dclr1,
828 (pvt->dclr1 & BIT(19)) ? "Enabled" : "Disabled",
829 (pvt->dclr1 & BIT(8)) ? "Enabled" : "Disabled",
830 (pvt->dclr1 & BIT(11)) ? "128b" : "64b");
831 debugf1(" DIMM x4 Present: L0=%s L1=%s L2=%s L3=%s "
832 "DIMM Type=%s\n",
833 (pvt->dclr1 & BIT(12)) ? "Y" : "N",
834 (pvt->dclr1 & BIT(13)) ? "Y" : "N",
835 (pvt->dclr1 & BIT(14)) ? "Y" : "N",
836 (pvt->dclr1 & BIT(15)) ? "Y" : "N",
837 (pvt->dclr1 & BIT(16)) ? "UN-Buffered" : "Buffered");
838 }
839
840 /*
841 * Determine if ganged and then dump memory sizes for first controller,
842 * and if NOT ganged dump info for 2nd controller.
843 */
844 ganged = dct_ganging_enabled(pvt);
845
846 f10_debug_display_dimm_sizes(0, pvt, ganged);
847
848 if (!ganged)
849 f10_debug_display_dimm_sizes(1, pvt, ganged);
850}
851
852/* Read in both of DBAM registers */
853static void amd64_read_dbam_reg(struct amd64_pvt *pvt)
854{
855 int err = 0;
856 unsigned int reg;
857
858 reg = DBAM0;
859 err = pci_read_config_dword(pvt->dram_f2_ctl, reg, &pvt->dbam0);
860 if (err)
861 goto err_reg;
862
863 if (boot_cpu_data.x86 >= 0x10) {
864 reg = DBAM1;
865 err = pci_read_config_dword(pvt->dram_f2_ctl, reg, &pvt->dbam1);
866
867 if (err)
868 goto err_reg;
869 }
870
871err_reg:
872 debugf0("Error reading F2x%03x.\n", reg);
873}
874
Doug Thompson94be4bf2009-04-27 16:12:00 +0200875/*
876 * NOTE: CPU Revision Dependent code: Rev E and Rev F
877 *
878 * Set the DCSB and DCSM mask values depending on the CPU revision value. Also
879 * set the shift factor for the DCSB and DCSM values.
880 *
881 * ->dcs_mask_notused, RevE:
882 *
883 * To find the max InputAddr for the csrow, start with the base address and set
884 * all bits that are "don't care" bits in the test at the start of section
885 * 3.5.4 (p. 84).
886 *
887 * The "don't care" bits are all set bits in the mask and all bits in the gaps
888 * between bit ranges [35:25] and [19:13]. The value REV_E_DCS_NOTUSED_BITS
889 * represents bits [24:20] and [12:0], which are all bits in the above-mentioned
890 * gaps.
891 *
892 * ->dcs_mask_notused, RevF and later:
893 *
894 * To find the max InputAddr for the csrow, start with the base address and set
895 * all bits that are "don't care" bits in the test at the start of NPT section
896 * 4.5.4 (p. 87).
897 *
898 * The "don't care" bits are all set bits in the mask and all bits in the gaps
899 * between bit ranges [36:27] and [21:13].
900 *
901 * The value REV_F_F1Xh_DCS_NOTUSED_BITS represents bits [26:22] and [12:0],
902 * which are all bits in the above-mentioned gaps.
903 */
904static void amd64_set_dct_base_and_mask(struct amd64_pvt *pvt)
905{
906 if (pvt->ext_model >= OPTERON_CPU_REV_F) {
907 pvt->dcsb_base = REV_F_F1Xh_DCSB_BASE_BITS;
908 pvt->dcsm_mask = REV_F_F1Xh_DCSM_MASK_BITS;
909 pvt->dcs_mask_notused = REV_F_F1Xh_DCS_NOTUSED_BITS;
910 pvt->dcs_shift = REV_F_F1Xh_DCS_SHIFT;
911
912 switch (boot_cpu_data.x86) {
913 case 0xf:
914 pvt->num_dcsm = REV_F_DCSM_COUNT;
915 break;
916
917 case 0x10:
918 pvt->num_dcsm = F10_DCSM_COUNT;
919 break;
920
921 case 0x11:
922 pvt->num_dcsm = F11_DCSM_COUNT;
923 break;
924
925 default:
926 amd64_printk(KERN_ERR, "Unsupported family!\n");
927 break;
928 }
929 } else {
930 pvt->dcsb_base = REV_E_DCSB_BASE_BITS;
931 pvt->dcsm_mask = REV_E_DCSM_MASK_BITS;
932 pvt->dcs_mask_notused = REV_E_DCS_NOTUSED_BITS;
933 pvt->dcs_shift = REV_E_DCS_SHIFT;
934 pvt->num_dcsm = REV_E_DCSM_COUNT;
935 }
936}
937
938/*
939 * Function 2 Offset F10_DCSB0; read in the DCS Base and DCS Mask hw registers
940 */
941static void amd64_read_dct_base_mask(struct amd64_pvt *pvt)
942{
943 int cs, reg, err = 0;
944
945 amd64_set_dct_base_and_mask(pvt);
946
947 for (cs = 0; cs < CHIPSELECT_COUNT; cs++) {
948 reg = K8_DCSB0 + (cs * 4);
949 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
950 &pvt->dcsb0[cs]);
951 if (unlikely(err))
952 debugf0("Reading K8_DCSB0[%d] failed\n", cs);
953 else
954 debugf0(" DCSB0[%d]=0x%08x reg: F2x%x\n",
955 cs, pvt->dcsb0[cs], reg);
956
957 /* If DCT are NOT ganged, then read in DCT1's base */
958 if (boot_cpu_data.x86 >= 0x10 && !dct_ganging_enabled(pvt)) {
959 reg = F10_DCSB1 + (cs * 4);
960 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
961 &pvt->dcsb1[cs]);
962 if (unlikely(err))
963 debugf0("Reading F10_DCSB1[%d] failed\n", cs);
964 else
965 debugf0(" DCSB1[%d]=0x%08x reg: F2x%x\n",
966 cs, pvt->dcsb1[cs], reg);
967 } else {
968 pvt->dcsb1[cs] = 0;
969 }
970 }
971
972 for (cs = 0; cs < pvt->num_dcsm; cs++) {
973 reg = K8_DCSB0 + (cs * 4);
974 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
975 &pvt->dcsm0[cs]);
976 if (unlikely(err))
977 debugf0("Reading K8_DCSM0 failed\n");
978 else
979 debugf0(" DCSM0[%d]=0x%08x reg: F2x%x\n",
980 cs, pvt->dcsm0[cs], reg);
981
982 /* If DCT are NOT ganged, then read in DCT1's mask */
983 if (boot_cpu_data.x86 >= 0x10 && !dct_ganging_enabled(pvt)) {
984 reg = F10_DCSM1 + (cs * 4);
985 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
986 &pvt->dcsm1[cs]);
987 if (unlikely(err))
988 debugf0("Reading F10_DCSM1[%d] failed\n", cs);
989 else
990 debugf0(" DCSM1[%d]=0x%08x reg: F2x%x\n",
991 cs, pvt->dcsm1[cs], reg);
992 } else
993 pvt->dcsm1[cs] = 0;
994 }
995}
996
997static enum mem_type amd64_determine_memory_type(struct amd64_pvt *pvt)
998{
999 enum mem_type type;
1000
1001 if (boot_cpu_data.x86 >= 0x10 || pvt->ext_model >= OPTERON_CPU_REV_F) {
1002 /* Rev F and later */
1003 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR2 : MEM_RDDR2;
1004 } else {
1005 /* Rev E and earlier */
1006 type = (pvt->dclr0 & BIT(18)) ? MEM_DDR : MEM_RDDR;
1007 }
1008
1009 debugf1(" Memory type is: %s\n",
1010 (type == MEM_DDR2) ? "MEM_DDR2" :
1011 (type == MEM_RDDR2) ? "MEM_RDDR2" :
1012 (type == MEM_DDR) ? "MEM_DDR" : "MEM_RDDR");
1013
1014 return type;
1015}
1016
Doug Thompsonddff8762009-04-27 16:14:52 +02001017/*
1018 * Read the DRAM Configuration Low register. It differs between CG, D & E revs
1019 * and the later RevF memory controllers (DDR vs DDR2)
1020 *
1021 * Return:
1022 * number of memory channels in operation
1023 * Pass back:
1024 * contents of the DCL0_LOW register
1025 */
1026static int k8_early_channel_count(struct amd64_pvt *pvt)
1027{
1028 int flag, err = 0;
1029
1030 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_0, &pvt->dclr0);
1031 if (err)
1032 return err;
1033
1034 if ((boot_cpu_data.x86_model >> 4) >= OPTERON_CPU_REV_F) {
1035 /* RevF (NPT) and later */
1036 flag = pvt->dclr0 & F10_WIDTH_128;
1037 } else {
1038 /* RevE and earlier */
1039 flag = pvt->dclr0 & REVE_WIDTH_128;
1040 }
1041
1042 /* not used */
1043 pvt->dclr1 = 0;
1044
1045 return (flag) ? 2 : 1;
1046}
1047
1048/* extract the ERROR ADDRESS for the K8 CPUs */
1049static u64 k8_get_error_address(struct mem_ctl_info *mci,
1050 struct amd64_error_info_regs *info)
1051{
1052 return (((u64) (info->nbeah & 0xff)) << 32) +
1053 (info->nbeal & ~0x03);
1054}
1055
1056/*
1057 * Read the Base and Limit registers for K8 based Memory controllers; extract
1058 * fields from the 'raw' reg into separate data fields
1059 *
1060 * Isolates: BASE, LIMIT, IntlvEn, IntlvSel, RW_EN
1061 */
1062static void k8_read_dram_base_limit(struct amd64_pvt *pvt, int dram)
1063{
1064 u32 low;
1065 u32 off = dram << 3; /* 8 bytes between DRAM entries */
1066 int err;
1067
1068 err = pci_read_config_dword(pvt->addr_f1_ctl,
1069 K8_DRAM_BASE_LOW + off, &low);
1070 if (err)
1071 debugf0("Reading K8_DRAM_BASE_LOW failed\n");
1072
1073 /* Extract parts into separate data entries */
1074 pvt->dram_base[dram] = ((u64) low & 0xFFFF0000) << 8;
1075 pvt->dram_IntlvEn[dram] = (low >> 8) & 0x7;
1076 pvt->dram_rw_en[dram] = (low & 0x3);
1077
1078 err = pci_read_config_dword(pvt->addr_f1_ctl,
1079 K8_DRAM_LIMIT_LOW + off, &low);
1080 if (err)
1081 debugf0("Reading K8_DRAM_LIMIT_LOW failed\n");
1082
1083 /*
1084 * Extract parts into separate data entries. Limit is the HIGHEST memory
1085 * location of the region, so lower 24 bits need to be all ones
1086 */
1087 pvt->dram_limit[dram] = (((u64) low & 0xFFFF0000) << 8) | 0x00FFFFFF;
1088 pvt->dram_IntlvSel[dram] = (low >> 8) & 0x7;
1089 pvt->dram_DstNode[dram] = (low & 0x7);
1090}
1091
1092static void k8_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
1093 struct amd64_error_info_regs *info,
1094 u64 SystemAddress)
1095{
1096 struct mem_ctl_info *src_mci;
1097 unsigned short syndrome;
1098 int channel, csrow;
1099 u32 page, offset;
1100
1101 /* Extract the syndrome parts and form a 16-bit syndrome */
1102 syndrome = EXTRACT_HIGH_SYNDROME(info->nbsl) << 8;
1103 syndrome |= EXTRACT_LOW_SYNDROME(info->nbsh);
1104
1105 /* CHIPKILL enabled */
1106 if (info->nbcfg & K8_NBCFG_CHIPKILL) {
1107 channel = get_channel_from_ecc_syndrome(syndrome);
1108 if (channel < 0) {
1109 /*
1110 * Syndrome didn't map, so we don't know which of the
1111 * 2 DIMMs is in error. So we need to ID 'both' of them
1112 * as suspect.
1113 */
1114 amd64_mc_printk(mci, KERN_WARNING,
1115 "unknown syndrome 0x%x - possible error "
1116 "reporting race\n", syndrome);
1117 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1118 return;
1119 }
1120 } else {
1121 /*
1122 * non-chipkill ecc mode
1123 *
1124 * The k8 documentation is unclear about how to determine the
1125 * channel number when using non-chipkill memory. This method
1126 * was obtained from email communication with someone at AMD.
1127 * (Wish the email was placed in this comment - norsk)
1128 */
1129 channel = ((SystemAddress & BIT(3)) != 0);
1130 }
1131
1132 /*
1133 * Find out which node the error address belongs to. This may be
1134 * different from the node that detected the error.
1135 */
1136 src_mci = find_mc_by_sys_addr(mci, SystemAddress);
1137 if (src_mci) {
1138 amd64_mc_printk(mci, KERN_ERR,
1139 "failed to map error address 0x%lx to a node\n",
1140 (unsigned long)SystemAddress);
1141 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1142 return;
1143 }
1144
1145 /* Now map the SystemAddress to a CSROW */
1146 csrow = sys_addr_to_csrow(src_mci, SystemAddress);
1147 if (csrow < 0) {
1148 edac_mc_handle_ce_no_info(src_mci, EDAC_MOD_STR);
1149 } else {
1150 error_address_to_page_and_offset(SystemAddress, &page, &offset);
1151
1152 edac_mc_handle_ce(src_mci, page, offset, syndrome, csrow,
1153 channel, EDAC_MOD_STR);
1154 }
1155}
1156
1157/*
1158 * determrine the number of PAGES in for this DIMM's size based on its DRAM
1159 * Address Mapping.
1160 *
1161 * First step is to calc the number of bits to shift a value of 1 left to
1162 * indicate show many pages. Start with the DBAM value as the starting bits,
1163 * then proceed to adjust those shift bits, based on CPU rev and the table.
1164 * See BKDG on the DBAM
1165 */
1166static int k8_dbam_map_to_pages(struct amd64_pvt *pvt, int dram_map)
1167{
1168 int nr_pages;
1169
1170 if (pvt->ext_model >= OPTERON_CPU_REV_F) {
1171 nr_pages = 1 << (revf_quad_ddr2_shift[dram_map] - PAGE_SHIFT);
1172 } else {
1173 /*
1174 * RevE and less section; this line is tricky. It collapses the
1175 * table used by RevD and later to one that matches revisions CG
1176 * and earlier.
1177 */
1178 dram_map -= (pvt->ext_model >= OPTERON_CPU_REV_D) ?
1179 (dram_map > 8 ? 4 : (dram_map > 5 ?
1180 3 : (dram_map > 2 ? 1 : 0))) : 0;
1181
1182 /* 25 shift is 32MiB minimum DIMM size in RevE and prior */
1183 nr_pages = 1 << (dram_map + 25 - PAGE_SHIFT);
1184 }
1185
1186 return nr_pages;
1187}
1188
Doug Thompson1afd3c92009-04-27 16:16:50 +02001189/*
1190 * Get the number of DCT channels in use.
1191 *
1192 * Return:
1193 * number of Memory Channels in operation
1194 * Pass back:
1195 * contents of the DCL0_LOW register
1196 */
1197static int f10_early_channel_count(struct amd64_pvt *pvt)
1198{
1199 int err = 0, channels = 0;
1200 u32 dbam;
Doug Thompsonddff8762009-04-27 16:14:52 +02001201
Doug Thompson1afd3c92009-04-27 16:16:50 +02001202 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_0, &pvt->dclr0);
1203 if (err)
1204 goto err_reg;
1205
1206 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_1, &pvt->dclr1);
1207 if (err)
1208 goto err_reg;
1209
1210 /* If we are in 128 bit mode, then we are using 2 channels */
1211 if (pvt->dclr0 & F10_WIDTH_128) {
1212 debugf0("Data WIDTH is 128 bits - 2 channels\n");
1213 channels = 2;
1214 return channels;
1215 }
1216
1217 /*
1218 * Need to check if in UN-ganged mode: In such, there are 2 channels,
1219 * but they are NOT in 128 bit mode and thus the above 'dcl0' status bit
1220 * will be OFF.
1221 *
1222 * Need to check DCT0[0] and DCT1[0] to see if only one of them has
1223 * their CSEnable bit on. If so, then SINGLE DIMM case.
1224 */
1225 debugf0("Data WIDTH is NOT 128 bits - need more decoding\n");
1226
1227 /*
1228 * Check DRAM Bank Address Mapping values for each DIMM to see if there
1229 * is more than just one DIMM present in unganged mode. Need to check
1230 * both controllers since DIMMs can be placed in either one.
1231 */
1232 channels = 0;
1233 err = pci_read_config_dword(pvt->dram_f2_ctl, DBAM0, &dbam);
1234 if (err)
1235 goto err_reg;
1236
1237 if (DBAM_DIMM(0, dbam) > 0)
1238 channels++;
1239 if (DBAM_DIMM(1, dbam) > 0)
1240 channels++;
1241 if (DBAM_DIMM(2, dbam) > 0)
1242 channels++;
1243 if (DBAM_DIMM(3, dbam) > 0)
1244 channels++;
1245
1246 /* If more than 2 DIMMs are present, then we have 2 channels */
1247 if (channels > 2)
1248 channels = 2;
1249 else if (channels == 0) {
1250 /* No DIMMs on DCT0, so look at DCT1 */
1251 err = pci_read_config_dword(pvt->dram_f2_ctl, DBAM1, &dbam);
1252 if (err)
1253 goto err_reg;
1254
1255 if (DBAM_DIMM(0, dbam) > 0)
1256 channels++;
1257 if (DBAM_DIMM(1, dbam) > 0)
1258 channels++;
1259 if (DBAM_DIMM(2, dbam) > 0)
1260 channels++;
1261 if (DBAM_DIMM(3, dbam) > 0)
1262 channels++;
1263
1264 if (channels > 2)
1265 channels = 2;
1266 }
1267
1268 /* If we found ALL 0 values, then assume just ONE DIMM-ONE Channel */
1269 if (channels == 0)
1270 channels = 1;
1271
Borislav Petkov37da0452009-06-10 17:36:57 +02001272 debugf0("MCT channel count: %d\n", channels);
Doug Thompson1afd3c92009-04-27 16:16:50 +02001273
1274 return channels;
1275
1276err_reg:
1277 return -1;
1278
1279}
1280
1281static int f10_dbam_map_to_pages(struct amd64_pvt *pvt, int dram_map)
1282{
1283 return 1 << (revf_quad_ddr2_shift[dram_map] - PAGE_SHIFT);
1284}
1285
1286/* Enable extended configuration access via 0xCF8 feature */
1287static void amd64_setup(struct amd64_pvt *pvt)
1288{
1289 u32 reg;
1290
1291 pci_read_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, &reg);
1292
1293 pvt->flags.cf8_extcfg = !!(reg & F10_NB_CFG_LOW_ENABLE_EXT_CFG);
1294 reg |= F10_NB_CFG_LOW_ENABLE_EXT_CFG;
1295 pci_write_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, reg);
1296}
1297
1298/* Restore the extended configuration access via 0xCF8 feature */
1299static void amd64_teardown(struct amd64_pvt *pvt)
1300{
1301 u32 reg;
1302
1303 pci_read_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, &reg);
1304
1305 reg &= ~F10_NB_CFG_LOW_ENABLE_EXT_CFG;
1306 if (pvt->flags.cf8_extcfg)
1307 reg |= F10_NB_CFG_LOW_ENABLE_EXT_CFG;
1308 pci_write_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, reg);
1309}
1310
1311static u64 f10_get_error_address(struct mem_ctl_info *mci,
1312 struct amd64_error_info_regs *info)
1313{
1314 return (((u64) (info->nbeah & 0xffff)) << 32) +
1315 (info->nbeal & ~0x01);
1316}
1317
1318/*
1319 * Read the Base and Limit registers for F10 based Memory controllers. Extract
1320 * fields from the 'raw' reg into separate data fields.
1321 *
1322 * Isolates: BASE, LIMIT, IntlvEn, IntlvSel, RW_EN.
1323 */
1324static void f10_read_dram_base_limit(struct amd64_pvt *pvt, int dram)
1325{
1326 u32 high_offset, low_offset, high_base, low_base, high_limit, low_limit;
1327
1328 low_offset = K8_DRAM_BASE_LOW + (dram << 3);
1329 high_offset = F10_DRAM_BASE_HIGH + (dram << 3);
1330
1331 /* read the 'raw' DRAM BASE Address register */
1332 pci_read_config_dword(pvt->addr_f1_ctl, low_offset, &low_base);
1333
1334 /* Read from the ECS data register */
1335 pci_read_config_dword(pvt->addr_f1_ctl, high_offset, &high_base);
1336
1337 /* Extract parts into separate data entries */
1338 pvt->dram_rw_en[dram] = (low_base & 0x3);
1339
1340 if (pvt->dram_rw_en[dram] == 0)
1341 return;
1342
1343 pvt->dram_IntlvEn[dram] = (low_base >> 8) & 0x7;
1344
1345 pvt->dram_base[dram] = (((((u64) high_base & 0x000000FF) << 32) |
1346 ((u64) low_base & 0xFFFF0000))) << 8;
1347
1348 low_offset = K8_DRAM_LIMIT_LOW + (dram << 3);
1349 high_offset = F10_DRAM_LIMIT_HIGH + (dram << 3);
1350
1351 /* read the 'raw' LIMIT registers */
1352 pci_read_config_dword(pvt->addr_f1_ctl, low_offset, &low_limit);
1353
1354 /* Read from the ECS data register for the HIGH portion */
1355 pci_read_config_dword(pvt->addr_f1_ctl, high_offset, &high_limit);
1356
1357 debugf0(" HW Regs: BASE=0x%08x-%08x LIMIT= 0x%08x-%08x\n",
1358 high_base, low_base, high_limit, low_limit);
1359
1360 pvt->dram_DstNode[dram] = (low_limit & 0x7);
1361 pvt->dram_IntlvSel[dram] = (low_limit >> 8) & 0x7;
1362
1363 /*
1364 * Extract address values and form a LIMIT address. Limit is the HIGHEST
1365 * memory location of the region, so low 24 bits need to be all ones.
1366 */
1367 low_limit |= 0x0000FFFF;
1368 pvt->dram_limit[dram] =
1369 ((((u64) high_limit << 32) + (u64) low_limit) << 8) | (0xFF);
1370}
Doug Thompson6163b5d2009-04-27 16:20:17 +02001371
1372static void f10_read_dram_ctl_register(struct amd64_pvt *pvt)
1373{
1374 int err = 0;
1375
1376 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCTL_SEL_LOW,
1377 &pvt->dram_ctl_select_low);
1378 if (err) {
1379 debugf0("Reading F10_DCTL_SEL_LOW failed\n");
1380 } else {
1381 debugf0("DRAM_DCTL_SEL_LOW=0x%x DctSelBaseAddr=0x%x\n",
1382 pvt->dram_ctl_select_low, dct_sel_baseaddr(pvt));
1383
1384 debugf0(" DRAM DCTs are=%s DRAM Is=%s DRAM-Ctl-"
1385 "sel-hi-range=%s\n",
1386 (dct_ganging_enabled(pvt) ? "GANGED" : "NOT GANGED"),
1387 (dct_dram_enabled(pvt) ? "Enabled" : "Disabled"),
1388 (dct_high_range_enabled(pvt) ? "Enabled" : "Disabled"));
1389
1390 debugf0(" DctDatIntLv=%s MemCleared=%s DctSelIntLvAddr=0x%x\n",
1391 (dct_data_intlv_enabled(pvt) ? "Enabled" : "Disabled"),
1392 (dct_memory_cleared(pvt) ? "True " : "False "),
1393 dct_sel_interleave_addr(pvt));
1394 }
1395
1396 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCTL_SEL_HIGH,
1397 &pvt->dram_ctl_select_high);
1398 if (err)
1399 debugf0("Reading F10_DCTL_SEL_HIGH failed\n");
1400}
1401
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001402/*
1403 * determine channel based on the interleaving mode: F10h BKDG, 2.8.9 Memory
1404 * Interleaving Modes.
1405 */
Doug Thompson6163b5d2009-04-27 16:20:17 +02001406static u32 f10_determine_channel(struct amd64_pvt *pvt, u64 sys_addr,
1407 int hi_range_sel, u32 intlv_en)
1408{
1409 u32 cs, temp, dct_sel_high = (pvt->dram_ctl_select_low >> 1) & 1;
1410
1411 if (dct_ganging_enabled(pvt))
1412 cs = 0;
1413 else if (hi_range_sel)
1414 cs = dct_sel_high;
1415 else if (dct_interleave_enabled(pvt)) {
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001416 /*
1417 * see F2x110[DctSelIntLvAddr] - channel interleave mode
1418 */
Doug Thompson6163b5d2009-04-27 16:20:17 +02001419 if (dct_sel_interleave_addr(pvt) == 0)
1420 cs = sys_addr >> 6 & 1;
1421 else if ((dct_sel_interleave_addr(pvt) >> 1) & 1) {
1422 temp = hweight_long((u32) ((sys_addr >> 16) & 0x1F)) % 2;
1423
1424 if (dct_sel_interleave_addr(pvt) & 1)
1425 cs = (sys_addr >> 9 & 1) ^ temp;
1426 else
1427 cs = (sys_addr >> 6 & 1) ^ temp;
1428 } else if (intlv_en & 4)
1429 cs = sys_addr >> 15 & 1;
1430 else if (intlv_en & 2)
1431 cs = sys_addr >> 14 & 1;
1432 else if (intlv_en & 1)
1433 cs = sys_addr >> 13 & 1;
1434 else
1435 cs = sys_addr >> 12 & 1;
1436 } else if (dct_high_range_enabled(pvt) && !dct_ganging_enabled(pvt))
1437 cs = ~dct_sel_high & 1;
1438 else
1439 cs = 0;
1440
1441 return cs;
1442}
1443
1444static inline u32 f10_map_intlv_en_to_shift(u32 intlv_en)
1445{
1446 if (intlv_en == 1)
1447 return 1;
1448 else if (intlv_en == 3)
1449 return 2;
1450 else if (intlv_en == 7)
1451 return 3;
1452
1453 return 0;
1454}
1455
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001456/* See F10h BKDG, 2.8.10.2 DctSelBaseOffset Programming */
1457static inline u64 f10_get_base_addr_offset(u64 sys_addr, int hi_range_sel,
Doug Thompson6163b5d2009-04-27 16:20:17 +02001458 u32 dct_sel_base_addr,
1459 u64 dct_sel_base_off,
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001460 u32 hole_valid, u32 hole_off,
Doug Thompson6163b5d2009-04-27 16:20:17 +02001461 u64 dram_base)
1462{
1463 u64 chan_off;
1464
1465 if (hi_range_sel) {
1466 if (!(dct_sel_base_addr & 0xFFFFF800) &&
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001467 hole_valid && (sys_addr >= 0x100000000ULL))
Doug Thompson6163b5d2009-04-27 16:20:17 +02001468 chan_off = hole_off << 16;
1469 else
1470 chan_off = dct_sel_base_off;
1471 } else {
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001472 if (hole_valid && (sys_addr >= 0x100000000ULL))
Doug Thompson6163b5d2009-04-27 16:20:17 +02001473 chan_off = hole_off << 16;
1474 else
1475 chan_off = dram_base & 0xFFFFF8000000ULL;
1476 }
1477
1478 return (sys_addr & 0x0000FFFFFFFFFFC0ULL) -
1479 (chan_off & 0x0000FFFFFF800000ULL);
1480}
1481
1482/* Hack for the time being - Can we get this from BIOS?? */
1483#define CH0SPARE_RANK 0
1484#define CH1SPARE_RANK 1
1485
1486/*
1487 * checks if the csrow passed in is marked as SPARED, if so returns the new
1488 * spare row
1489 */
1490static inline int f10_process_possible_spare(int csrow,
1491 u32 cs, struct amd64_pvt *pvt)
1492{
1493 u32 swap_done;
1494 u32 bad_dram_cs;
1495
1496 /* Depending on channel, isolate respective SPARING info */
1497 if (cs) {
1498 swap_done = F10_ONLINE_SPARE_SWAPDONE1(pvt->online_spare);
1499 bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS1(pvt->online_spare);
1500 if (swap_done && (csrow == bad_dram_cs))
1501 csrow = CH1SPARE_RANK;
1502 } else {
1503 swap_done = F10_ONLINE_SPARE_SWAPDONE0(pvt->online_spare);
1504 bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS0(pvt->online_spare);
1505 if (swap_done && (csrow == bad_dram_cs))
1506 csrow = CH0SPARE_RANK;
1507 }
1508 return csrow;
1509}
1510
1511/*
1512 * Iterate over the DRAM DCT "base" and "mask" registers looking for a
1513 * SystemAddr match on the specified 'ChannelSelect' and 'NodeID'
1514 *
1515 * Return:
1516 * -EINVAL: NOT FOUND
1517 * 0..csrow = Chip-Select Row
1518 */
1519static int f10_lookup_addr_in_dct(u32 in_addr, u32 nid, u32 cs)
1520{
1521 struct mem_ctl_info *mci;
1522 struct amd64_pvt *pvt;
1523 u32 cs_base, cs_mask;
1524 int cs_found = -EINVAL;
1525 int csrow;
1526
1527 mci = mci_lookup[nid];
1528 if (!mci)
1529 return cs_found;
1530
1531 pvt = mci->pvt_info;
1532
1533 debugf1("InputAddr=0x%x channelselect=%d\n", in_addr, cs);
1534
1535 for (csrow = 0; csrow < CHIPSELECT_COUNT; csrow++) {
1536
1537 cs_base = amd64_get_dct_base(pvt, cs, csrow);
1538 if (!(cs_base & K8_DCSB_CS_ENABLE))
1539 continue;
1540
1541 /*
1542 * We have an ENABLED CSROW, Isolate just the MASK bits of the
1543 * target: [28:19] and [13:5], which map to [36:27] and [21:13]
1544 * of the actual address.
1545 */
1546 cs_base &= REV_F_F1Xh_DCSB_BASE_BITS;
1547
1548 /*
1549 * Get the DCT Mask, and ENABLE the reserved bits: [18:16] and
1550 * [4:0] to become ON. Then mask off bits [28:0] ([36:8])
1551 */
1552 cs_mask = amd64_get_dct_mask(pvt, cs, csrow);
1553
1554 debugf1(" CSROW=%d CSBase=0x%x RAW CSMask=0x%x\n",
1555 csrow, cs_base, cs_mask);
1556
1557 cs_mask = (cs_mask | 0x0007C01F) & 0x1FFFFFFF;
1558
1559 debugf1(" Final CSMask=0x%x\n", cs_mask);
1560 debugf1(" (InputAddr & ~CSMask)=0x%x "
1561 "(CSBase & ~CSMask)=0x%x\n",
1562 (in_addr & ~cs_mask), (cs_base & ~cs_mask));
1563
1564 if ((in_addr & ~cs_mask) == (cs_base & ~cs_mask)) {
1565 cs_found = f10_process_possible_spare(csrow, cs, pvt);
1566
1567 debugf1(" MATCH csrow=%d\n", cs_found);
1568 break;
1569 }
1570 }
1571 return cs_found;
1572}
1573
Doug Thompsonf71d0a02009-04-27 16:22:43 +02001574/* For a given @dram_range, check if @sys_addr falls within it. */
1575static int f10_match_to_this_node(struct amd64_pvt *pvt, int dram_range,
1576 u64 sys_addr, int *nid, int *chan_sel)
1577{
1578 int node_id, cs_found = -EINVAL, high_range = 0;
1579 u32 intlv_en, intlv_sel, intlv_shift, hole_off;
1580 u32 hole_valid, tmp, dct_sel_base, channel;
1581 u64 dram_base, chan_addr, dct_sel_base_off;
1582
1583 dram_base = pvt->dram_base[dram_range];
1584 intlv_en = pvt->dram_IntlvEn[dram_range];
1585
1586 node_id = pvt->dram_DstNode[dram_range];
1587 intlv_sel = pvt->dram_IntlvSel[dram_range];
1588
1589 debugf1("(dram=%d) Base=0x%llx SystemAddr= 0x%llx Limit=0x%llx\n",
1590 dram_range, dram_base, sys_addr, pvt->dram_limit[dram_range]);
1591
1592 /*
1593 * This assumes that one node's DHAR is the same as all the other
1594 * nodes' DHAR.
1595 */
1596 hole_off = (pvt->dhar & 0x0000FF80);
1597 hole_valid = (pvt->dhar & 0x1);
1598 dct_sel_base_off = (pvt->dram_ctl_select_high & 0xFFFFFC00) << 16;
1599
1600 debugf1(" HoleOffset=0x%x HoleValid=0x%x IntlvSel=0x%x\n",
1601 hole_off, hole_valid, intlv_sel);
1602
1603 if (intlv_en ||
1604 (intlv_sel != ((sys_addr >> 12) & intlv_en)))
1605 return -EINVAL;
1606
1607 dct_sel_base = dct_sel_baseaddr(pvt);
1608
1609 /*
1610 * check whether addresses >= DctSelBaseAddr[47:27] are to be used to
1611 * select between DCT0 and DCT1.
1612 */
1613 if (dct_high_range_enabled(pvt) &&
1614 !dct_ganging_enabled(pvt) &&
1615 ((sys_addr >> 27) >= (dct_sel_base >> 11)))
1616 high_range = 1;
1617
1618 channel = f10_determine_channel(pvt, sys_addr, high_range, intlv_en);
1619
1620 chan_addr = f10_get_base_addr_offset(sys_addr, high_range, dct_sel_base,
1621 dct_sel_base_off, hole_valid,
1622 hole_off, dram_base);
1623
1624 intlv_shift = f10_map_intlv_en_to_shift(intlv_en);
1625
1626 /* remove Node ID (in case of memory interleaving) */
1627 tmp = chan_addr & 0xFC0;
1628
1629 chan_addr = ((chan_addr >> intlv_shift) & 0xFFFFFFFFF000ULL) | tmp;
1630
1631 /* remove channel interleave and hash */
1632 if (dct_interleave_enabled(pvt) &&
1633 !dct_high_range_enabled(pvt) &&
1634 !dct_ganging_enabled(pvt)) {
1635 if (dct_sel_interleave_addr(pvt) != 1)
1636 chan_addr = (chan_addr >> 1) & 0xFFFFFFFFFFFFFFC0ULL;
1637 else {
1638 tmp = chan_addr & 0xFC0;
1639 chan_addr = ((chan_addr & 0xFFFFFFFFFFFFC000ULL) >> 1)
1640 | tmp;
1641 }
1642 }
1643
1644 debugf1(" (ChannelAddrLong=0x%llx) >> 8 becomes InputAddr=0x%x\n",
1645 chan_addr, (u32)(chan_addr >> 8));
1646
1647 cs_found = f10_lookup_addr_in_dct(chan_addr >> 8, node_id, channel);
1648
1649 if (cs_found >= 0) {
1650 *nid = node_id;
1651 *chan_sel = channel;
1652 }
1653 return cs_found;
1654}
1655
1656static int f10_translate_sysaddr_to_cs(struct amd64_pvt *pvt, u64 sys_addr,
1657 int *node, int *chan_sel)
1658{
1659 int dram_range, cs_found = -EINVAL;
1660 u64 dram_base, dram_limit;
1661
1662 for (dram_range = 0; dram_range < DRAM_REG_COUNT; dram_range++) {
1663
1664 if (!pvt->dram_rw_en[dram_range])
1665 continue;
1666
1667 dram_base = pvt->dram_base[dram_range];
1668 dram_limit = pvt->dram_limit[dram_range];
1669
1670 if ((dram_base <= sys_addr) && (sys_addr <= dram_limit)) {
1671
1672 cs_found = f10_match_to_this_node(pvt, dram_range,
1673 sys_addr, node,
1674 chan_sel);
1675 if (cs_found >= 0)
1676 break;
1677 }
1678 }
1679 return cs_found;
1680}
1681
1682/*
1683 * This the F10h reference code from AMD to map a @sys_addr to NodeID,
1684 * CSROW, Channel.
1685 *
1686 * The @sys_addr is usually an error address received from the hardware.
1687 */
1688static void f10_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
1689 struct amd64_error_info_regs *info,
1690 u64 sys_addr)
1691{
1692 struct amd64_pvt *pvt = mci->pvt_info;
1693 u32 page, offset;
1694 unsigned short syndrome;
1695 int nid, csrow, chan = 0;
1696
1697 csrow = f10_translate_sysaddr_to_cs(pvt, sys_addr, &nid, &chan);
1698
1699 if (csrow >= 0) {
1700 error_address_to_page_and_offset(sys_addr, &page, &offset);
1701
1702 syndrome = EXTRACT_HIGH_SYNDROME(info->nbsl) << 8;
1703 syndrome |= EXTRACT_LOW_SYNDROME(info->nbsh);
1704
1705 /*
1706 * Is CHIPKILL on? If so, then we can attempt to use the
1707 * syndrome to isolate which channel the error was on.
1708 */
1709 if (pvt->nbcfg & K8_NBCFG_CHIPKILL)
1710 chan = get_channel_from_ecc_syndrome(syndrome);
1711
1712 if (chan >= 0) {
1713 edac_mc_handle_ce(mci, page, offset, syndrome,
1714 csrow, chan, EDAC_MOD_STR);
1715 } else {
1716 /*
1717 * Channel unknown, report all channels on this
1718 * CSROW as failed.
1719 */
1720 for (chan = 0; chan < mci->csrows[csrow].nr_channels;
1721 chan++) {
1722 edac_mc_handle_ce(mci, page, offset,
1723 syndrome,
1724 csrow, chan,
1725 EDAC_MOD_STR);
1726 }
1727 }
1728
1729 } else {
1730 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1731 }
1732}
1733
1734/*
1735 * Input (@index) is the DBAM DIMM value (1 of 4) used as an index into a shift
1736 * table (revf_quad_ddr2_shift) which starts at 128MB DIMM size. Index of 0
1737 * indicates an empty DIMM slot, as reported by Hardware on empty slots.
1738 *
1739 * Normalize to 128MB by subracting 27 bit shift.
1740 */
1741static int map_dbam_to_csrow_size(int index)
1742{
1743 int mega_bytes = 0;
1744
1745 if (index > 0 && index <= DBAM_MAX_VALUE)
1746 mega_bytes = ((128 << (revf_quad_ddr2_shift[index]-27)));
1747
1748 return mega_bytes;
1749}
1750
1751/*
1752 * debug routine to display the memory sizes of a DIMM (ganged or not) and it
1753 * CSROWs as well
1754 */
1755static void f10_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt,
1756 int ganged)
1757{
1758 int dimm, size0, size1;
1759 u32 dbam;
1760 u32 *dcsb;
1761
1762 debugf1(" dbam%d: 0x%8.08x CSROW is %s\n", ctrl,
1763 ctrl ? pvt->dbam1 : pvt->dbam0,
1764 ganged ? "GANGED - dbam1 not used" : "NON-GANGED");
1765
1766 dbam = ctrl ? pvt->dbam1 : pvt->dbam0;
1767 dcsb = ctrl ? pvt->dcsb1 : pvt->dcsb0;
1768
1769 /* Dump memory sizes for DIMM and its CSROWs */
1770 for (dimm = 0; dimm < 4; dimm++) {
1771
1772 size0 = 0;
1773 if (dcsb[dimm*2] & K8_DCSB_CS_ENABLE)
1774 size0 = map_dbam_to_csrow_size(DBAM_DIMM(dimm, dbam));
1775
1776 size1 = 0;
1777 if (dcsb[dimm*2 + 1] & K8_DCSB_CS_ENABLE)
1778 size1 = map_dbam_to_csrow_size(DBAM_DIMM(dimm, dbam));
1779
1780 debugf1(" CTRL-%d DIMM-%d=%5dMB CSROW-%d=%5dMB "
1781 "CSROW-%d=%5dMB\n",
1782 ctrl,
1783 dimm,
1784 size0 + size1,
1785 dimm * 2,
1786 size0,
1787 dimm * 2 + 1,
1788 size1);
1789 }
1790}
1791
1792/*
1793 * Very early hardware probe on pci_probe thread to determine if this module
1794 * supports the hardware.
1795 *
1796 * Return:
1797 * 0 for OK
1798 * 1 for error
1799 */
1800static int f10_probe_valid_hardware(struct amd64_pvt *pvt)
1801{
1802 int ret = 0;
1803
1804 /*
1805 * If we are on a DDR3 machine, we don't know yet if
1806 * we support that properly at this time
1807 */
1808 if ((pvt->dchr0 & F10_DCHR_Ddr3Mode) ||
1809 (pvt->dchr1 & F10_DCHR_Ddr3Mode)) {
1810
1811 amd64_printk(KERN_WARNING,
1812 "%s() This machine is running with DDR3 memory. "
1813 "This is not currently supported. "
1814 "DCHR0=0x%x DCHR1=0x%x\n",
1815 __func__, pvt->dchr0, pvt->dchr1);
1816
1817 amd64_printk(KERN_WARNING,
1818 " Contact '%s' module MAINTAINER to help add"
1819 " support.\n",
1820 EDAC_MOD_STR);
1821
1822 ret = 1;
1823
1824 }
1825 return ret;
1826}
Doug Thompson6163b5d2009-04-27 16:20:17 +02001827
Doug Thompson4d376072009-04-27 16:25:05 +02001828/*
1829 * There currently are 3 types type of MC devices for AMD Athlon/Opterons
1830 * (as per PCI DEVICE_IDs):
1831 *
1832 * Family K8: That is the Athlon64 and Opteron CPUs. They all have the same PCI
1833 * DEVICE ID, even though there is differences between the different Revisions
1834 * (CG,D,E,F).
1835 *
1836 * Family F10h and F11h.
1837 *
1838 */
1839static struct amd64_family_type amd64_family_types[] = {
1840 [K8_CPUS] = {
1841 .ctl_name = "RevF",
1842 .addr_f1_ctl = PCI_DEVICE_ID_AMD_K8_NB_ADDRMAP,
1843 .misc_f3_ctl = PCI_DEVICE_ID_AMD_K8_NB_MISC,
1844 .ops = {
1845 .early_channel_count = k8_early_channel_count,
1846 .get_error_address = k8_get_error_address,
1847 .read_dram_base_limit = k8_read_dram_base_limit,
1848 .map_sysaddr_to_csrow = k8_map_sysaddr_to_csrow,
1849 .dbam_map_to_pages = k8_dbam_map_to_pages,
1850 }
1851 },
1852 [F10_CPUS] = {
1853 .ctl_name = "Family 10h",
1854 .addr_f1_ctl = PCI_DEVICE_ID_AMD_10H_NB_MAP,
1855 .misc_f3_ctl = PCI_DEVICE_ID_AMD_10H_NB_MISC,
1856 .ops = {
1857 .probe_valid_hardware = f10_probe_valid_hardware,
1858 .early_channel_count = f10_early_channel_count,
1859 .get_error_address = f10_get_error_address,
1860 .read_dram_base_limit = f10_read_dram_base_limit,
1861 .read_dram_ctl_register = f10_read_dram_ctl_register,
1862 .map_sysaddr_to_csrow = f10_map_sysaddr_to_csrow,
1863 .dbam_map_to_pages = f10_dbam_map_to_pages,
1864 }
1865 },
1866 [F11_CPUS] = {
1867 .ctl_name = "Family 11h",
1868 .addr_f1_ctl = PCI_DEVICE_ID_AMD_11H_NB_MAP,
1869 .misc_f3_ctl = PCI_DEVICE_ID_AMD_11H_NB_MISC,
1870 .ops = {
1871 .probe_valid_hardware = f10_probe_valid_hardware,
1872 .early_channel_count = f10_early_channel_count,
1873 .get_error_address = f10_get_error_address,
1874 .read_dram_base_limit = f10_read_dram_base_limit,
1875 .read_dram_ctl_register = f10_read_dram_ctl_register,
1876 .map_sysaddr_to_csrow = f10_map_sysaddr_to_csrow,
1877 .dbam_map_to_pages = f10_dbam_map_to_pages,
1878 }
1879 },
1880};
1881
1882static struct pci_dev *pci_get_related_function(unsigned int vendor,
1883 unsigned int device,
1884 struct pci_dev *related)
1885{
1886 struct pci_dev *dev = NULL;
1887
1888 dev = pci_get_device(vendor, device, dev);
1889 while (dev) {
1890 if ((dev->bus->number == related->bus->number) &&
1891 (PCI_SLOT(dev->devfn) == PCI_SLOT(related->devfn)))
1892 break;
1893 dev = pci_get_device(vendor, device, dev);
1894 }
1895
1896 return dev;
1897}
1898
Doug Thompsonb1289d62009-04-27 16:37:05 +02001899/*
1900 * syndrome mapping table for ECC ChipKill devices
1901 *
1902 * The comment in each row is the token (nibble) number that is in error.
1903 * The least significant nibble of the syndrome is the mask for the bits
1904 * that are in error (need to be toggled) for the particular nibble.
1905 *
1906 * Each row contains 16 entries.
1907 * The first entry (0th) is the channel number for that row of syndromes.
1908 * The remaining 15 entries are the syndromes for the respective Error
1909 * bit mask index.
1910 *
1911 * 1st index entry is 0x0001 mask, indicating that the rightmost bit is the
1912 * bit in error.
1913 * The 2nd index entry is 0x0010 that the second bit is damaged.
1914 * The 3rd index entry is 0x0011 indicating that the rightmost 2 bits
1915 * are damaged.
1916 * Thus so on until index 15, 0x1111, whose entry has the syndrome
1917 * indicating that all 4 bits are damaged.
1918 *
1919 * A search is performed on this table looking for a given syndrome.
1920 *
1921 * See the AMD documentation for ECC syndromes. This ECC table is valid
1922 * across all the versions of the AMD64 processors.
1923 *
1924 * A fast lookup is to use the LAST four bits of the 16-bit syndrome as a
1925 * COLUMN index, then search all ROWS of that column, looking for a match
1926 * with the input syndrome. The ROW value will be the token number.
1927 *
1928 * The 0'th entry on that row, can be returned as the CHANNEL (0 or 1) of this
1929 * error.
1930 */
1931#define NUMBER_ECC_ROWS 36
1932static const unsigned short ecc_chipkill_syndromes[NUMBER_ECC_ROWS][16] = {
1933 /* Channel 0 syndromes */
1934 {/*0*/ 0, 0xe821, 0x7c32, 0x9413, 0xbb44, 0x5365, 0xc776, 0x2f57,
1935 0xdd88, 0x35a9, 0xa1ba, 0x499b, 0x66cc, 0x8eed, 0x1afe, 0xf2df },
1936 {/*1*/ 0, 0x5d31, 0xa612, 0xfb23, 0x9584, 0xc8b5, 0x3396, 0x6ea7,
1937 0xeac8, 0xb7f9, 0x4cda, 0x11eb, 0x7f4c, 0x227d, 0xd95e, 0x846f },
1938 {/*2*/ 0, 0x0001, 0x0002, 0x0003, 0x0004, 0x0005, 0x0006, 0x0007,
1939 0x0008, 0x0009, 0x000a, 0x000b, 0x000c, 0x000d, 0x000e, 0x000f },
1940 {/*3*/ 0, 0x2021, 0x3032, 0x1013, 0x4044, 0x6065, 0x7076, 0x5057,
1941 0x8088, 0xa0a9, 0xb0ba, 0x909b, 0xc0cc, 0xe0ed, 0xf0fe, 0xd0df },
1942 {/*4*/ 0, 0x5041, 0xa082, 0xf0c3, 0x9054, 0xc015, 0x30d6, 0x6097,
1943 0xe0a8, 0xb0e9, 0x402a, 0x106b, 0x70fc, 0x20bd, 0xd07e, 0x803f },
1944 {/*5*/ 0, 0xbe21, 0xd732, 0x6913, 0x2144, 0x9f65, 0xf676, 0x4857,
1945 0x3288, 0x8ca9, 0xe5ba, 0x5b9b, 0x13cc, 0xaded, 0xc4fe, 0x7adf },
1946 {/*6*/ 0, 0x4951, 0x8ea2, 0xc7f3, 0x5394, 0x1ac5, 0xdd36, 0x9467,
1947 0xa1e8, 0xe8b9, 0x2f4a, 0x661b, 0xf27c, 0xbb2d, 0x7cde, 0x358f },
1948 {/*7*/ 0, 0x74e1, 0x9872, 0xec93, 0xd6b4, 0xa255, 0x4ec6, 0x3a27,
1949 0x6bd8, 0x1f39, 0xf3aa, 0x874b, 0xbd6c, 0xc98d, 0x251e, 0x51ff },
1950 {/*8*/ 0, 0x15c1, 0x2a42, 0x3f83, 0xcef4, 0xdb35, 0xe4b6, 0xf177,
1951 0x4758, 0x5299, 0x6d1a, 0x78db, 0x89ac, 0x9c6d, 0xa3ee, 0xb62f },
1952 {/*9*/ 0, 0x3d01, 0x1602, 0x2b03, 0x8504, 0xb805, 0x9306, 0xae07,
1953 0xca08, 0xf709, 0xdc0a, 0xe10b, 0x4f0c, 0x720d, 0x590e, 0x640f },
1954 {/*a*/ 0, 0x9801, 0xec02, 0x7403, 0x6b04, 0xf305, 0x8706, 0x1f07,
1955 0xbd08, 0x2509, 0x510a, 0xc90b, 0xd60c, 0x4e0d, 0x3a0e, 0xa20f },
1956 {/*b*/ 0, 0xd131, 0x6212, 0xb323, 0x3884, 0xe9b5, 0x5a96, 0x8ba7,
1957 0x1cc8, 0xcdf9, 0x7eda, 0xafeb, 0x244c, 0xf57d, 0x465e, 0x976f },
1958 {/*c*/ 0, 0xe1d1, 0x7262, 0x93b3, 0xb834, 0x59e5, 0xca56, 0x2b87,
1959 0xdc18, 0x3dc9, 0xae7a, 0x4fab, 0x542c, 0x85fd, 0x164e, 0xf79f },
1960 {/*d*/ 0, 0x6051, 0xb0a2, 0xd0f3, 0x1094, 0x70c5, 0xa036, 0xc067,
1961 0x20e8, 0x40b9, 0x904a, 0x601b, 0x307c, 0x502d, 0x80de, 0xe08f },
1962 {/*e*/ 0, 0xa4c1, 0xf842, 0x5c83, 0xe6f4, 0x4235, 0x1eb6, 0xba77,
1963 0x7b58, 0xdf99, 0x831a, 0x27db, 0x9dac, 0x396d, 0x65ee, 0xc12f },
1964 {/*f*/ 0, 0x11c1, 0x2242, 0x3383, 0xc8f4, 0xd935, 0xeab6, 0xfb77,
1965 0x4c58, 0x5d99, 0x6e1a, 0x7fdb, 0x84ac, 0x956d, 0xa6ee, 0xb72f },
Doug Thompson4d376072009-04-27 16:25:05 +02001966
Doug Thompsonb1289d62009-04-27 16:37:05 +02001967 /* Channel 1 syndromes */
1968 {/*10*/ 1, 0x45d1, 0x8a62, 0xcfb3, 0x5e34, 0x1be5, 0xd456, 0x9187,
1969 0xa718, 0xe2c9, 0x2d7a, 0x68ab, 0xf92c, 0xbcfd, 0x734e, 0x369f },
1970 {/*11*/ 1, 0x63e1, 0xb172, 0xd293, 0x14b4, 0x7755, 0xa5c6, 0xc627,
1971 0x28d8, 0x4b39, 0x99aa, 0xfa4b, 0x3c6c, 0x5f8d, 0x8d1e, 0xeeff },
1972 {/*12*/ 1, 0xb741, 0xd982, 0x6ec3, 0x2254, 0x9515, 0xfbd6, 0x4c97,
1973 0x33a8, 0x84e9, 0xea2a, 0x5d6b, 0x11fc, 0xa6bd, 0xc87e, 0x7f3f },
1974 {/*13*/ 1, 0xdd41, 0x6682, 0xbbc3, 0x3554, 0xe815, 0x53d6, 0xce97,
1975 0x1aa8, 0xc7e9, 0x7c2a, 0xa1fb, 0x2ffc, 0xf2bd, 0x497e, 0x943f },
1976 {/*14*/ 1, 0x2bd1, 0x3d62, 0x16b3, 0x4f34, 0x64e5, 0x7256, 0x5987,
1977 0x8518, 0xaec9, 0xb87a, 0x93ab, 0xca2c, 0xe1fd, 0xf74e, 0xdc9f },
1978 {/*15*/ 1, 0x83c1, 0xc142, 0x4283, 0xa4f4, 0x2735, 0x65b6, 0xe677,
1979 0xf858, 0x7b99, 0x391a, 0xbadb, 0x5cac, 0xdf6d, 0x9dee, 0x1e2f },
1980 {/*16*/ 1, 0x8fd1, 0xc562, 0x4ab3, 0xa934, 0x26e5, 0x6c56, 0xe387,
1981 0xfe18, 0x71c9, 0x3b7a, 0xb4ab, 0x572c, 0xd8fd, 0x924e, 0x1d9f },
1982 {/*17*/ 1, 0x4791, 0x89e2, 0xce73, 0x5264, 0x15f5, 0xdb86, 0x9c17,
1983 0xa3b8, 0xe429, 0x2a5a, 0x6dcb, 0xf1dc, 0xb64d, 0x783e, 0x3faf },
1984 {/*18*/ 1, 0x5781, 0xa9c2, 0xfe43, 0x92a4, 0xc525, 0x3b66, 0x6ce7,
1985 0xe3f8, 0xb479, 0x4a3a, 0x1dbb, 0x715c, 0x26dd, 0xd89e, 0x8f1f },
1986 {/*19*/ 1, 0xbf41, 0xd582, 0x6ac3, 0x2954, 0x9615, 0xfcd6, 0x4397,
1987 0x3ea8, 0x81e9, 0xeb2a, 0x546b, 0x17fc, 0xa8bd, 0xc27e, 0x7d3f },
1988 {/*1a*/ 1, 0x9891, 0xe1e2, 0x7273, 0x6464, 0xf7f5, 0x8586, 0x1617,
1989 0xb8b8, 0x2b29, 0x595a, 0xcacb, 0xdcdc, 0x4f4d, 0x3d3e, 0xaeaf },
1990 {/*1b*/ 1, 0xcce1, 0x4472, 0x8893, 0xfdb4, 0x3f55, 0xb9c6, 0x7527,
1991 0x56d8, 0x9a39, 0x12aa, 0xde4b, 0xab6c, 0x678d, 0xef1e, 0x23ff },
1992 {/*1c*/ 1, 0xa761, 0xf9b2, 0x5ed3, 0xe214, 0x4575, 0x1ba6, 0xbcc7,
1993 0x7328, 0xd449, 0x8a9a, 0x2dfb, 0x913c, 0x365d, 0x688e, 0xcfef },
1994 {/*1d*/ 1, 0xff61, 0x55b2, 0xaad3, 0x7914, 0x8675, 0x2ca6, 0xd3c7,
1995 0x9e28, 0x6149, 0xcb9a, 0x34fb, 0xe73c, 0x185d, 0xb28e, 0x4def },
1996 {/*1e*/ 1, 0x5451, 0xa8a2, 0xfcf3, 0x9694, 0xc2c5, 0x3e36, 0x6a67,
1997 0xebe8, 0xbfb9, 0x434a, 0x171b, 0x7d7c, 0x292d, 0xd5de, 0x818f },
1998 {/*1f*/ 1, 0x6fc1, 0xb542, 0xda83, 0x19f4, 0x7635, 0xacb6, 0xc377,
1999 0x2e58, 0x4199, 0x9b1a, 0xf4db, 0x37ac, 0x586d, 0x82ee, 0xed2f },
2000
2001 /* ECC bits are also in the set of tokens and they too can go bad
2002 * first 2 cover channel 0, while the second 2 cover channel 1
2003 */
2004 {/*20*/ 0, 0xbe01, 0xd702, 0x6903, 0x2104, 0x9f05, 0xf606, 0x4807,
2005 0x3208, 0x8c09, 0xe50a, 0x5b0b, 0x130c, 0xad0d, 0xc40e, 0x7a0f },
2006 {/*21*/ 0, 0x4101, 0x8202, 0xc303, 0x5804, 0x1905, 0xda06, 0x9b07,
2007 0xac08, 0xed09, 0x2e0a, 0x6f0b, 0x640c, 0xb50d, 0x760e, 0x370f },
2008 {/*22*/ 1, 0xc441, 0x4882, 0x8cc3, 0xf654, 0x3215, 0xbed6, 0x7a97,
2009 0x5ba8, 0x9fe9, 0x132a, 0xd76b, 0xadfc, 0x69bd, 0xe57e, 0x213f },
2010 {/*23*/ 1, 0x7621, 0x9b32, 0xed13, 0xda44, 0xac65, 0x4176, 0x3757,
2011 0x6f88, 0x19a9, 0xf4ba, 0x829b, 0xb5cc, 0xc3ed, 0x2efe, 0x58df }
2012};
2013
2014/*
2015 * Given the syndrome argument, scan each of the channel tables for a syndrome
2016 * match. Depending on which table it is found, return the channel number.
2017 */
2018static int get_channel_from_ecc_syndrome(unsigned short syndrome)
2019{
2020 int row;
2021 int column;
2022
2023 /* Determine column to scan */
2024 column = syndrome & 0xF;
2025
2026 /* Scan all rows, looking for syndrome, or end of table */
2027 for (row = 0; row < NUMBER_ECC_ROWS; row++) {
2028 if (ecc_chipkill_syndromes[row][column] == syndrome)
2029 return ecc_chipkill_syndromes[row][0];
2030 }
2031
2032 debugf0("syndrome(%x) not found\n", syndrome);
2033 return -1;
2034}
Doug Thompsond27bf6f2009-05-06 17:55:27 +02002035
2036/*
2037 * Check for valid error in the NB Status High register. If so, proceed to read
2038 * NB Status Low, NB Address Low and NB Address High registers and store data
2039 * into error structure.
2040 *
2041 * Returns:
2042 * - 1: if hardware regs contains valid error info
2043 * - 0: if no valid error is indicated
2044 */
2045static int amd64_get_error_info_regs(struct mem_ctl_info *mci,
2046 struct amd64_error_info_regs *regs)
2047{
2048 struct amd64_pvt *pvt;
2049 struct pci_dev *misc_f3_ctl;
2050 int err = 0;
2051
2052 pvt = mci->pvt_info;
2053 misc_f3_ctl = pvt->misc_f3_ctl;
2054
2055 err = pci_read_config_dword(misc_f3_ctl, K8_NBSH, &regs->nbsh);
2056 if (err)
2057 goto err_reg;
2058
2059 if (!(regs->nbsh & K8_NBSH_VALID_BIT))
2060 return 0;
2061
2062 /* valid error, read remaining error information registers */
2063 err = pci_read_config_dword(misc_f3_ctl, K8_NBSL, &regs->nbsl);
2064 if (err)
2065 goto err_reg;
2066
2067 err = pci_read_config_dword(misc_f3_ctl, K8_NBEAL, &regs->nbeal);
2068 if (err)
2069 goto err_reg;
2070
2071 err = pci_read_config_dword(misc_f3_ctl, K8_NBEAH, &regs->nbeah);
2072 if (err)
2073 goto err_reg;
2074
2075 err = pci_read_config_dword(misc_f3_ctl, K8_NBCFG, &regs->nbcfg);
2076 if (err)
2077 goto err_reg;
2078
2079 return 1;
2080
2081err_reg:
2082 debugf0("Reading error info register failed\n");
2083 return 0;
2084}
2085
2086/*
2087 * This function is called to retrieve the error data from hardware and store it
2088 * in the info structure.
2089 *
2090 * Returns:
2091 * - 1: if a valid error is found
2092 * - 0: if no error is found
2093 */
2094static int amd64_get_error_info(struct mem_ctl_info *mci,
2095 struct amd64_error_info_regs *info)
2096{
2097 struct amd64_pvt *pvt;
2098 struct amd64_error_info_regs regs;
2099
2100 pvt = mci->pvt_info;
2101
2102 if (!amd64_get_error_info_regs(mci, info))
2103 return 0;
2104
2105 /*
2106 * Here's the problem with the K8's EDAC reporting: There are four
2107 * registers which report pieces of error information. They are shared
2108 * between CEs and UEs. Furthermore, contrary to what is stated in the
2109 * BKDG, the overflow bit is never used! Every error always updates the
2110 * reporting registers.
2111 *
2112 * Can you see the race condition? All four error reporting registers
2113 * must be read before a new error updates them! There is no way to read
2114 * all four registers atomically. The best than can be done is to detect
2115 * that a race has occured and then report the error without any kind of
2116 * precision.
2117 *
2118 * What is still positive is that errors are still reported and thus
2119 * problems can still be detected - just not localized because the
2120 * syndrome and address are spread out across registers.
2121 *
2122 * Grrrrr!!!!! Here's hoping that AMD fixes this in some future K8 rev.
2123 * UEs and CEs should have separate register sets with proper overflow
2124 * bits that are used! At very least the problem can be fixed by
2125 * honoring the ErrValid bit in 'nbsh' and not updating registers - just
2126 * set the overflow bit - unless the current error is CE and the new
2127 * error is UE which would be the only situation for overwriting the
2128 * current values.
2129 */
2130
2131 regs = *info;
2132
2133 /* Use info from the second read - most current */
2134 if (unlikely(!amd64_get_error_info_regs(mci, info)))
2135 return 0;
2136
2137 /* clear the error bits in hardware */
2138 pci_write_bits32(pvt->misc_f3_ctl, K8_NBSH, 0, K8_NBSH_VALID_BIT);
2139
2140 /* Check for the possible race condition */
2141 if ((regs.nbsh != info->nbsh) ||
2142 (regs.nbsl != info->nbsl) ||
2143 (regs.nbeah != info->nbeah) ||
2144 (regs.nbeal != info->nbeal)) {
2145 amd64_mc_printk(mci, KERN_WARNING,
2146 "hardware STATUS read access race condition "
2147 "detected!\n");
2148 return 0;
2149 }
2150 return 1;
2151}
2152
2153static inline void amd64_decode_gart_tlb_error(struct mem_ctl_info *mci,
2154 struct amd64_error_info_regs *info)
2155{
2156 u32 err_code;
2157 u32 ec_tt; /* error code transaction type (2b) */
2158 u32 ec_ll; /* error code cache level (2b) */
2159
2160 err_code = EXTRACT_ERROR_CODE(info->nbsl);
2161 ec_ll = EXTRACT_LL_CODE(err_code);
2162 ec_tt = EXTRACT_TT_CODE(err_code);
2163
2164 amd64_mc_printk(mci, KERN_ERR,
2165 "GART TLB event: transaction type(%s), "
2166 "cache level(%s)\n", tt_msgs[ec_tt], ll_msgs[ec_ll]);
2167}
2168
2169static inline void amd64_decode_mem_cache_error(struct mem_ctl_info *mci,
2170 struct amd64_error_info_regs *info)
2171{
2172 u32 err_code;
2173 u32 ec_rrrr; /* error code memory transaction (4b) */
2174 u32 ec_tt; /* error code transaction type (2b) */
2175 u32 ec_ll; /* error code cache level (2b) */
2176
2177 err_code = EXTRACT_ERROR_CODE(info->nbsl);
2178 ec_ll = EXTRACT_LL_CODE(err_code);
2179 ec_tt = EXTRACT_TT_CODE(err_code);
2180 ec_rrrr = EXTRACT_RRRR_CODE(err_code);
2181
2182 amd64_mc_printk(mci, KERN_ERR,
2183 "cache hierarchy error: memory transaction type(%s), "
2184 "transaction type(%s), cache level(%s)\n",
2185 rrrr_msgs[ec_rrrr], tt_msgs[ec_tt], ll_msgs[ec_ll]);
2186}
2187
2188
2189/*
2190 * Handle any Correctable Errors (CEs) that have occurred. Check for valid ERROR
2191 * ADDRESS and process.
2192 */
2193static void amd64_handle_ce(struct mem_ctl_info *mci,
2194 struct amd64_error_info_regs *info)
2195{
2196 struct amd64_pvt *pvt = mci->pvt_info;
2197 u64 SystemAddress;
2198
2199 /* Ensure that the Error Address is VALID */
2200 if ((info->nbsh & K8_NBSH_VALID_ERROR_ADDR) == 0) {
2201 amd64_mc_printk(mci, KERN_ERR,
2202 "HW has no ERROR_ADDRESS available\n");
2203 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
2204 return;
2205 }
2206
2207 SystemAddress = extract_error_address(mci, info);
2208
2209 amd64_mc_printk(mci, KERN_ERR,
2210 "CE ERROR_ADDRESS= 0x%llx\n", SystemAddress);
2211
2212 pvt->ops->map_sysaddr_to_csrow(mci, info, SystemAddress);
2213}
2214
2215/* Handle any Un-correctable Errors (UEs) */
2216static void amd64_handle_ue(struct mem_ctl_info *mci,
2217 struct amd64_error_info_regs *info)
2218{
2219 int csrow;
2220 u64 SystemAddress;
2221 u32 page, offset;
2222 struct mem_ctl_info *log_mci, *src_mci = NULL;
2223
2224 log_mci = mci;
2225
2226 if ((info->nbsh & K8_NBSH_VALID_ERROR_ADDR) == 0) {
2227 amd64_mc_printk(mci, KERN_CRIT,
2228 "HW has no ERROR_ADDRESS available\n");
2229 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
2230 return;
2231 }
2232
2233 SystemAddress = extract_error_address(mci, info);
2234
2235 /*
2236 * Find out which node the error address belongs to. This may be
2237 * different from the node that detected the error.
2238 */
2239 src_mci = find_mc_by_sys_addr(mci, SystemAddress);
2240 if (!src_mci) {
2241 amd64_mc_printk(mci, KERN_CRIT,
2242 "ERROR ADDRESS (0x%lx) value NOT mapped to a MC\n",
2243 (unsigned long)SystemAddress);
2244 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
2245 return;
2246 }
2247
2248 log_mci = src_mci;
2249
2250 csrow = sys_addr_to_csrow(log_mci, SystemAddress);
2251 if (csrow < 0) {
2252 amd64_mc_printk(mci, KERN_CRIT,
2253 "ERROR_ADDRESS (0x%lx) value NOT mapped to 'csrow'\n",
2254 (unsigned long)SystemAddress);
2255 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
2256 } else {
2257 error_address_to_page_and_offset(SystemAddress, &page, &offset);
2258 edac_mc_handle_ue(log_mci, page, offset, csrow, EDAC_MOD_STR);
2259 }
2260}
2261
2262static void amd64_decode_bus_error(struct mem_ctl_info *mci,
2263 struct amd64_error_info_regs *info)
2264{
2265 u32 err_code, ext_ec;
2266 u32 ec_pp; /* error code participating processor (2p) */
2267 u32 ec_to; /* error code timed out (1b) */
2268 u32 ec_rrrr; /* error code memory transaction (4b) */
2269 u32 ec_ii; /* error code memory or I/O (2b) */
2270 u32 ec_ll; /* error code cache level (2b) */
2271
2272 ext_ec = EXTRACT_EXT_ERROR_CODE(info->nbsl);
2273 err_code = EXTRACT_ERROR_CODE(info->nbsl);
2274
2275 ec_ll = EXTRACT_LL_CODE(err_code);
2276 ec_ii = EXTRACT_II_CODE(err_code);
2277 ec_rrrr = EXTRACT_RRRR_CODE(err_code);
2278 ec_to = EXTRACT_TO_CODE(err_code);
2279 ec_pp = EXTRACT_PP_CODE(err_code);
2280
2281 amd64_mc_printk(mci, KERN_ERR,
2282 "BUS ERROR:\n"
2283 " time-out(%s) mem or i/o(%s)\n"
2284 " participating processor(%s)\n"
2285 " memory transaction type(%s)\n"
2286 " cache level(%s) Error Found by: %s\n",
2287 to_msgs[ec_to],
2288 ii_msgs[ec_ii],
2289 pp_msgs[ec_pp],
2290 rrrr_msgs[ec_rrrr],
2291 ll_msgs[ec_ll],
2292 (info->nbsh & K8_NBSH_ERR_SCRUBER) ?
2293 "Scrubber" : "Normal Operation");
2294
2295 /* If this was an 'observed' error, early out */
2296 if (ec_pp == K8_NBSL_PP_OBS)
2297 return; /* We aren't the node involved */
2298
2299 /* Parse out the extended error code for ECC events */
2300 switch (ext_ec) {
2301 /* F10 changed to one Extended ECC error code */
2302 case F10_NBSL_EXT_ERR_RES: /* Reserved field */
2303 case F10_NBSL_EXT_ERR_ECC: /* F10 ECC ext err code */
2304 break;
2305
2306 default:
2307 amd64_mc_printk(mci, KERN_ERR, "NOT ECC: no special error "
2308 "handling for this error\n");
2309 return;
2310 }
2311
2312 if (info->nbsh & K8_NBSH_CECC)
2313 amd64_handle_ce(mci, info);
2314 else if (info->nbsh & K8_NBSH_UECC)
2315 amd64_handle_ue(mci, info);
2316
2317 /*
2318 * If main error is CE then overflow must be CE. If main error is UE
2319 * then overflow is unknown. We'll call the overflow a CE - if
2320 * panic_on_ue is set then we're already panic'ed and won't arrive
2321 * here. Else, then apparently someone doesn't think that UE's are
2322 * catastrophic.
2323 */
2324 if (info->nbsh & K8_NBSH_OVERFLOW)
2325 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR
2326 "Error Overflow set");
2327}
2328
2329int amd64_process_error_info(struct mem_ctl_info *mci,
2330 struct amd64_error_info_regs *info,
2331 int handle_errors)
2332{
2333 struct amd64_pvt *pvt;
2334 struct amd64_error_info_regs *regs;
2335 u32 err_code, ext_ec;
2336 int gart_tlb_error = 0;
2337
2338 pvt = mci->pvt_info;
2339
2340 /* If caller doesn't want us to process the error, return */
2341 if (!handle_errors)
2342 return 1;
2343
2344 regs = info;
2345
2346 debugf1("NorthBridge ERROR: mci(0x%p)\n", mci);
2347 debugf1(" MC node(%d) Error-Address(0x%.8x-%.8x)\n",
2348 pvt->mc_node_id, regs->nbeah, regs->nbeal);
2349 debugf1(" nbsh(0x%.8x) nbsl(0x%.8x)\n",
2350 regs->nbsh, regs->nbsl);
2351 debugf1(" Valid Error=%s Overflow=%s\n",
2352 (regs->nbsh & K8_NBSH_VALID_BIT) ? "True" : "False",
2353 (regs->nbsh & K8_NBSH_OVERFLOW) ? "True" : "False");
2354 debugf1(" Err Uncorrected=%s MCA Error Reporting=%s\n",
2355 (regs->nbsh & K8_NBSH_UNCORRECTED_ERR) ?
2356 "True" : "False",
2357 (regs->nbsh & K8_NBSH_ERR_ENABLE) ?
2358 "True" : "False");
2359 debugf1(" MiscErr Valid=%s ErrAddr Valid=%s PCC=%s\n",
2360 (regs->nbsh & K8_NBSH_MISC_ERR_VALID) ?
2361 "True" : "False",
2362 (regs->nbsh & K8_NBSH_VALID_ERROR_ADDR) ?
2363 "True" : "False",
2364 (regs->nbsh & K8_NBSH_PCC) ?
2365 "True" : "False");
2366 debugf1(" CECC=%s UECC=%s Found by Scruber=%s\n",
2367 (regs->nbsh & K8_NBSH_CECC) ?
2368 "True" : "False",
2369 (regs->nbsh & K8_NBSH_UECC) ?
2370 "True" : "False",
2371 (regs->nbsh & K8_NBSH_ERR_SCRUBER) ?
2372 "True" : "False");
2373 debugf1(" CORE0=%s CORE1=%s CORE2=%s CORE3=%s\n",
2374 (regs->nbsh & K8_NBSH_CORE0) ? "True" : "False",
2375 (regs->nbsh & K8_NBSH_CORE1) ? "True" : "False",
2376 (regs->nbsh & K8_NBSH_CORE2) ? "True" : "False",
2377 (regs->nbsh & K8_NBSH_CORE3) ? "True" : "False");
2378
2379
2380 err_code = EXTRACT_ERROR_CODE(regs->nbsl);
2381
2382 /* Determine which error type:
2383 * 1) GART errors - non-fatal, developmental events
2384 * 2) MEMORY errors
2385 * 3) BUS errors
2386 * 4) Unknown error
2387 */
2388 if (TEST_TLB_ERROR(err_code)) {
2389 /*
2390 * GART errors are intended to help graphics driver developers
2391 * to detect bad GART PTEs. It is recommended by AMD to disable
2392 * GART table walk error reporting by default[1] (currently
2393 * being disabled in mce_cpu_quirks()) and according to the
2394 * comment in mce_cpu_quirks(), such GART errors can be
2395 * incorrectly triggered. We may see these errors anyway and
2396 * unless requested by the user, they won't be reported.
2397 *
2398 * [1] section 13.10.1 on BIOS and Kernel Developers Guide for
2399 * AMD NPT family 0Fh processors
2400 */
2401 if (report_gart_errors == 0)
2402 return 1;
2403
2404 /*
2405 * Only if GART error reporting is requested should we generate
2406 * any logs.
2407 */
2408 gart_tlb_error = 1;
2409
2410 debugf1("GART TLB error\n");
2411 amd64_decode_gart_tlb_error(mci, info);
2412 } else if (TEST_MEM_ERROR(err_code)) {
2413 debugf1("Memory/Cache error\n");
2414 amd64_decode_mem_cache_error(mci, info);
2415 } else if (TEST_BUS_ERROR(err_code)) {
2416 debugf1("Bus (Link/DRAM) error\n");
2417 amd64_decode_bus_error(mci, info);
2418 } else {
2419 /* shouldn't reach here! */
2420 amd64_mc_printk(mci, KERN_WARNING,
2421 "%s(): unknown MCE error 0x%x\n", __func__,
2422 err_code);
2423 }
2424
2425 ext_ec = EXTRACT_EXT_ERROR_CODE(regs->nbsl);
2426 amd64_mc_printk(mci, KERN_ERR,
2427 "ExtErr=(0x%x) %s\n", ext_ec, ext_msgs[ext_ec]);
2428
2429 if (((ext_ec >= F10_NBSL_EXT_ERR_CRC &&
2430 ext_ec <= F10_NBSL_EXT_ERR_TGT) ||
2431 (ext_ec == F10_NBSL_EXT_ERR_RMW)) &&
2432 EXTRACT_LDT_LINK(info->nbsh)) {
2433
2434 amd64_mc_printk(mci, KERN_ERR,
2435 "Error on hypertransport link: %s\n",
2436 htlink_msgs[
2437 EXTRACT_LDT_LINK(info->nbsh)]);
2438 }
2439
2440 /*
2441 * Check the UE bit of the NB status high register, if set generate some
2442 * logs. If NOT a GART error, then process the event as a NO-INFO event.
2443 * If it was a GART error, skip that process.
2444 */
2445 if (regs->nbsh & K8_NBSH_UNCORRECTED_ERR) {
2446 amd64_mc_printk(mci, KERN_CRIT, "uncorrected error\n");
2447 if (!gart_tlb_error)
2448 edac_mc_handle_ue_no_info(mci, "UE bit is set\n");
2449 }
2450
2451 if (regs->nbsh & K8_NBSH_PCC)
2452 amd64_mc_printk(mci, KERN_CRIT,
2453 "PCC (processor context corrupt) set\n");
2454
2455 return 1;
2456}
2457EXPORT_SYMBOL_GPL(amd64_process_error_info);
2458
Doug Thompson0ec449e2009-04-27 19:41:25 +02002459/*
2460 * The main polling 'check' function, called FROM the edac core to perform the
2461 * error checking and if an error is encountered, error processing.
2462 */
2463static void amd64_check(struct mem_ctl_info *mci)
2464{
2465 struct amd64_error_info_regs info;
2466
2467 if (amd64_get_error_info(mci, &info))
2468 amd64_process_error_info(mci, &info, 1);
2469}
2470
2471/*
2472 * Input:
2473 * 1) struct amd64_pvt which contains pvt->dram_f2_ctl pointer
2474 * 2) AMD Family index value
2475 *
2476 * Ouput:
2477 * Upon return of 0, the following filled in:
2478 *
2479 * struct pvt->addr_f1_ctl
2480 * struct pvt->misc_f3_ctl
2481 *
2482 * Filled in with related device funcitions of 'dram_f2_ctl'
2483 * These devices are "reserved" via the pci_get_device()
2484 *
2485 * Upon return of 1 (error status):
2486 *
2487 * Nothing reserved
2488 */
2489static int amd64_reserve_mc_sibling_devices(struct amd64_pvt *pvt, int mc_idx)
2490{
2491 const struct amd64_family_type *amd64_dev = &amd64_family_types[mc_idx];
2492
2493 /* Reserve the ADDRESS MAP Device */
2494 pvt->addr_f1_ctl = pci_get_related_function(pvt->dram_f2_ctl->vendor,
2495 amd64_dev->addr_f1_ctl,
2496 pvt->dram_f2_ctl);
2497
2498 if (!pvt->addr_f1_ctl) {
2499 amd64_printk(KERN_ERR, "error address map device not found: "
2500 "vendor %x device 0x%x (broken BIOS?)\n",
2501 PCI_VENDOR_ID_AMD, amd64_dev->addr_f1_ctl);
2502 return 1;
2503 }
2504
2505 /* Reserve the MISC Device */
2506 pvt->misc_f3_ctl = pci_get_related_function(pvt->dram_f2_ctl->vendor,
2507 amd64_dev->misc_f3_ctl,
2508 pvt->dram_f2_ctl);
2509
2510 if (!pvt->misc_f3_ctl) {
2511 pci_dev_put(pvt->addr_f1_ctl);
2512 pvt->addr_f1_ctl = NULL;
2513
2514 amd64_printk(KERN_ERR, "error miscellaneous device not found: "
2515 "vendor %x device 0x%x (broken BIOS?)\n",
2516 PCI_VENDOR_ID_AMD, amd64_dev->misc_f3_ctl);
2517 return 1;
2518 }
2519
2520 debugf1(" Addr Map device PCI Bus ID:\t%s\n",
2521 pci_name(pvt->addr_f1_ctl));
2522 debugf1(" DRAM MEM-CTL PCI Bus ID:\t%s\n",
2523 pci_name(pvt->dram_f2_ctl));
2524 debugf1(" Misc device PCI Bus ID:\t%s\n",
2525 pci_name(pvt->misc_f3_ctl));
2526
2527 return 0;
2528}
2529
2530static void amd64_free_mc_sibling_devices(struct amd64_pvt *pvt)
2531{
2532 pci_dev_put(pvt->addr_f1_ctl);
2533 pci_dev_put(pvt->misc_f3_ctl);
2534}
2535
2536/*
2537 * Retrieve the hardware registers of the memory controller (this includes the
2538 * 'Address Map' and 'Misc' device regs)
2539 */
2540static void amd64_read_mc_registers(struct amd64_pvt *pvt)
2541{
2542 u64 msr_val;
2543 int dram, err = 0;
2544
2545 /*
2546 * Retrieve TOP_MEM and TOP_MEM2; no masking off of reserved bits since
2547 * those are Read-As-Zero
2548 */
2549 rdmsrl(MSR_K8_TOP_MEM1, msr_val);
2550 pvt->top_mem = msr_val >> 23;
2551 debugf0(" TOP_MEM=0x%08llx\n", pvt->top_mem);
2552
2553 /* check first whether TOP_MEM2 is enabled */
2554 rdmsrl(MSR_K8_SYSCFG, msr_val);
2555 if (msr_val & (1U << 21)) {
2556 rdmsrl(MSR_K8_TOP_MEM2, msr_val);
2557 pvt->top_mem2 = msr_val >> 23;
2558 debugf0(" TOP_MEM2=0x%08llx\n", pvt->top_mem2);
2559 } else
2560 debugf0(" TOP_MEM2 disabled.\n");
2561
2562 amd64_cpu_display_info(pvt);
2563
2564 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCAP, &pvt->nbcap);
2565 if (err)
2566 goto err_reg;
2567
2568 if (pvt->ops->read_dram_ctl_register)
2569 pvt->ops->read_dram_ctl_register(pvt);
2570
2571 for (dram = 0; dram < DRAM_REG_COUNT; dram++) {
2572 /*
2573 * Call CPU specific READ function to get the DRAM Base and
2574 * Limit values from the DCT.
2575 */
2576 pvt->ops->read_dram_base_limit(pvt, dram);
2577
2578 /*
2579 * Only print out debug info on rows with both R and W Enabled.
2580 * Normal processing, compiler should optimize this whole 'if'
2581 * debug output block away.
2582 */
2583 if (pvt->dram_rw_en[dram] != 0) {
2584 debugf1(" DRAM_BASE[%d]: 0x%8.08x-%8.08x "
2585 "DRAM_LIMIT: 0x%8.08x-%8.08x\n",
2586 dram,
2587 (u32)(pvt->dram_base[dram] >> 32),
2588 (u32)(pvt->dram_base[dram] & 0xFFFFFFFF),
2589 (u32)(pvt->dram_limit[dram] >> 32),
2590 (u32)(pvt->dram_limit[dram] & 0xFFFFFFFF));
2591 debugf1(" IntlvEn=%s %s %s "
2592 "IntlvSel=%d DstNode=%d\n",
2593 pvt->dram_IntlvEn[dram] ?
2594 "Enabled" : "Disabled",
2595 (pvt->dram_rw_en[dram] & 0x2) ? "W" : "!W",
2596 (pvt->dram_rw_en[dram] & 0x1) ? "R" : "!R",
2597 pvt->dram_IntlvSel[dram],
2598 pvt->dram_DstNode[dram]);
2599 }
2600 }
2601
2602 amd64_read_dct_base_mask(pvt);
2603
2604 err = pci_read_config_dword(pvt->addr_f1_ctl, K8_DHAR, &pvt->dhar);
2605 if (err)
2606 goto err_reg;
2607
2608 amd64_read_dbam_reg(pvt);
2609
2610 err = pci_read_config_dword(pvt->misc_f3_ctl,
2611 F10_ONLINE_SPARE, &pvt->online_spare);
2612 if (err)
2613 goto err_reg;
2614
2615 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_0, &pvt->dclr0);
2616 if (err)
2617 goto err_reg;
2618
2619 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCHR_0, &pvt->dchr0);
2620 if (err)
2621 goto err_reg;
2622
2623 if (!dct_ganging_enabled(pvt)) {
2624 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_1,
2625 &pvt->dclr1);
2626 if (err)
2627 goto err_reg;
2628
2629 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCHR_1,
2630 &pvt->dchr1);
2631 if (err)
2632 goto err_reg;
2633 }
2634
2635 amd64_dump_misc_regs(pvt);
2636
2637err_reg:
2638 debugf0("Reading an MC register failed\n");
2639
2640}
2641
2642/*
2643 * NOTE: CPU Revision Dependent code
2644 *
2645 * Input:
2646 * @csrow_nr ChipSelect Row Number (0..CHIPSELECT_COUNT-1)
2647 * k8 private pointer to -->
2648 * DRAM Bank Address mapping register
2649 * node_id
2650 * DCL register where dual_channel_active is
2651 *
2652 * The DBAM register consists of 4 sets of 4 bits each definitions:
2653 *
2654 * Bits: CSROWs
2655 * 0-3 CSROWs 0 and 1
2656 * 4-7 CSROWs 2 and 3
2657 * 8-11 CSROWs 4 and 5
2658 * 12-15 CSROWs 6 and 7
2659 *
2660 * Values range from: 0 to 15
2661 * The meaning of the values depends on CPU revision and dual-channel state,
2662 * see relevant BKDG more info.
2663 *
2664 * The memory controller provides for total of only 8 CSROWs in its current
2665 * architecture. Each "pair" of CSROWs normally represents just one DIMM in
2666 * single channel or two (2) DIMMs in dual channel mode.
2667 *
2668 * The following code logic collapses the various tables for CSROW based on CPU
2669 * revision.
2670 *
2671 * Returns:
2672 * The number of PAGE_SIZE pages on the specified CSROW number it
2673 * encompasses
2674 *
2675 */
2676static u32 amd64_csrow_nr_pages(int csrow_nr, struct amd64_pvt *pvt)
2677{
2678 u32 dram_map, nr_pages;
2679
2680 /*
2681 * The math on this doesn't look right on the surface because x/2*4 can
2682 * be simplified to x*2 but this expression makes use of the fact that
2683 * it is integral math where 1/2=0. This intermediate value becomes the
2684 * number of bits to shift the DBAM register to extract the proper CSROW
2685 * field.
2686 */
2687 dram_map = (pvt->dbam0 >> ((csrow_nr / 2) * 4)) & 0xF;
2688
2689 nr_pages = pvt->ops->dbam_map_to_pages(pvt, dram_map);
2690
2691 /*
2692 * If dual channel then double the memory size of single channel.
2693 * Channel count is 1 or 2
2694 */
2695 nr_pages <<= (pvt->channel_count - 1);
2696
2697 debugf0(" (csrow=%d) DBAM map index= %d\n", csrow_nr, dram_map);
2698 debugf0(" nr_pages= %u channel-count = %d\n",
2699 nr_pages, pvt->channel_count);
2700
2701 return nr_pages;
2702}
2703
2704/*
2705 * Initialize the array of csrow attribute instances, based on the values
2706 * from pci config hardware registers.
2707 */
2708static int amd64_init_csrows(struct mem_ctl_info *mci)
2709{
2710 struct csrow_info *csrow;
2711 struct amd64_pvt *pvt;
2712 u64 input_addr_min, input_addr_max, sys_addr;
2713 int i, err = 0, empty = 1;
2714
2715 pvt = mci->pvt_info;
2716
2717 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCFG, &pvt->nbcfg);
2718 if (err)
2719 debugf0("Reading K8_NBCFG failed\n");
2720
2721 debugf0("NBCFG= 0x%x CHIPKILL= %s DRAM ECC= %s\n", pvt->nbcfg,
2722 (pvt->nbcfg & K8_NBCFG_CHIPKILL) ? "Enabled" : "Disabled",
2723 (pvt->nbcfg & K8_NBCFG_ECC_ENABLE) ? "Enabled" : "Disabled"
2724 );
2725
2726 for (i = 0; i < CHIPSELECT_COUNT; i++) {
2727 csrow = &mci->csrows[i];
2728
2729 if ((pvt->dcsb0[i] & K8_DCSB_CS_ENABLE) == 0) {
2730 debugf1("----CSROW %d EMPTY for node %d\n", i,
2731 pvt->mc_node_id);
2732 continue;
2733 }
2734
2735 debugf1("----CSROW %d VALID for MC node %d\n",
2736 i, pvt->mc_node_id);
2737
2738 empty = 0;
2739 csrow->nr_pages = amd64_csrow_nr_pages(i, pvt);
2740 find_csrow_limits(mci, i, &input_addr_min, &input_addr_max);
2741 sys_addr = input_addr_to_sys_addr(mci, input_addr_min);
2742 csrow->first_page = (u32) (sys_addr >> PAGE_SHIFT);
2743 sys_addr = input_addr_to_sys_addr(mci, input_addr_max);
2744 csrow->last_page = (u32) (sys_addr >> PAGE_SHIFT);
2745 csrow->page_mask = ~mask_from_dct_mask(pvt, i);
2746 /* 8 bytes of resolution */
2747
2748 csrow->mtype = amd64_determine_memory_type(pvt);
2749
2750 debugf1(" for MC node %d csrow %d:\n", pvt->mc_node_id, i);
2751 debugf1(" input_addr_min: 0x%lx input_addr_max: 0x%lx\n",
2752 (unsigned long)input_addr_min,
2753 (unsigned long)input_addr_max);
2754 debugf1(" sys_addr: 0x%lx page_mask: 0x%lx\n",
2755 (unsigned long)sys_addr, csrow->page_mask);
2756 debugf1(" nr_pages: %u first_page: 0x%lx "
2757 "last_page: 0x%lx\n",
2758 (unsigned)csrow->nr_pages,
2759 csrow->first_page, csrow->last_page);
2760
2761 /*
2762 * determine whether CHIPKILL or JUST ECC or NO ECC is operating
2763 */
2764 if (pvt->nbcfg & K8_NBCFG_ECC_ENABLE)
2765 csrow->edac_mode =
2766 (pvt->nbcfg & K8_NBCFG_CHIPKILL) ?
2767 EDAC_S4ECD4ED : EDAC_SECDED;
2768 else
2769 csrow->edac_mode = EDAC_NONE;
2770 }
2771
2772 return empty;
2773}
Doug Thompsond27bf6f2009-05-06 17:55:27 +02002774
Doug Thompsonf9431992009-04-27 19:46:08 +02002775/*
2776 * Only if 'ecc_enable_override' is set AND BIOS had ECC disabled, do "we"
2777 * enable it.
2778 */
2779static void amd64_enable_ecc_error_reporting(struct mem_ctl_info *mci)
2780{
2781 struct amd64_pvt *pvt = mci->pvt_info;
2782 const cpumask_t *cpumask = cpumask_of_node(pvt->mc_node_id);
2783 int cpu, idx = 0, err = 0;
2784 struct msr msrs[cpumask_weight(cpumask)];
2785 u32 value;
2786 u32 mask = K8_NBCTL_CECCEn | K8_NBCTL_UECCEn;
2787
2788 if (!ecc_enable_override)
2789 return;
2790
2791 memset(msrs, 0, sizeof(msrs));
2792
2793 amd64_printk(KERN_WARNING,
2794 "'ecc_enable_override' parameter is active, "
2795 "Enabling AMD ECC hardware now: CAUTION\n");
2796
2797 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCTL, &value);
2798 if (err)
2799 debugf0("Reading K8_NBCTL failed\n");
2800
2801 /* turn on UECCn and CECCEn bits */
2802 pvt->old_nbctl = value & mask;
2803 pvt->nbctl_mcgctl_saved = 1;
2804
2805 value |= mask;
2806 pci_write_config_dword(pvt->misc_f3_ctl, K8_NBCTL, value);
2807
2808 rdmsr_on_cpus(cpumask, K8_MSR_MCGCTL, msrs);
2809
2810 for_each_cpu(cpu, cpumask) {
2811 if (msrs[idx].l & K8_MSR_MCGCTL_NBE)
2812 set_bit(idx, &pvt->old_mcgctl);
2813
2814 msrs[idx].l |= K8_MSR_MCGCTL_NBE;
2815 idx++;
2816 }
2817 wrmsr_on_cpus(cpumask, K8_MSR_MCGCTL, msrs);
2818
2819 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCFG, &value);
2820 if (err)
2821 debugf0("Reading K8_NBCFG failed\n");
2822
2823 debugf0("NBCFG(1)= 0x%x CHIPKILL= %s ECC_ENABLE= %s\n", value,
2824 (value & K8_NBCFG_CHIPKILL) ? "Enabled" : "Disabled",
2825 (value & K8_NBCFG_ECC_ENABLE) ? "Enabled" : "Disabled");
2826
2827 if (!(value & K8_NBCFG_ECC_ENABLE)) {
2828 amd64_printk(KERN_WARNING,
2829 "This node reports that DRAM ECC is "
2830 "currently Disabled; ENABLING now\n");
2831
2832 /* Attempt to turn on DRAM ECC Enable */
2833 value |= K8_NBCFG_ECC_ENABLE;
2834 pci_write_config_dword(pvt->misc_f3_ctl, K8_NBCFG, value);
2835
2836 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCFG, &value);
2837 if (err)
2838 debugf0("Reading K8_NBCFG failed\n");
2839
2840 if (!(value & K8_NBCFG_ECC_ENABLE)) {
2841 amd64_printk(KERN_WARNING,
2842 "Hardware rejects Enabling DRAM ECC checking\n"
2843 "Check memory DIMM configuration\n");
2844 } else {
2845 amd64_printk(KERN_DEBUG,
2846 "Hardware accepted DRAM ECC Enable\n");
2847 }
2848 }
2849 debugf0("NBCFG(2)= 0x%x CHIPKILL= %s ECC_ENABLE= %s\n", value,
2850 (value & K8_NBCFG_CHIPKILL) ? "Enabled" : "Disabled",
2851 (value & K8_NBCFG_ECC_ENABLE) ? "Enabled" : "Disabled");
2852
2853 pvt->ctl_error_info.nbcfg = value;
2854}
2855
2856static void amd64_restore_ecc_error_reporting(struct amd64_pvt *pvt)
2857{
2858 const cpumask_t *cpumask = cpumask_of_node(pvt->mc_node_id);
2859 int cpu, idx = 0, err = 0;
2860 struct msr msrs[cpumask_weight(cpumask)];
2861 u32 value;
2862 u32 mask = K8_NBCTL_CECCEn | K8_NBCTL_UECCEn;
2863
2864 if (!pvt->nbctl_mcgctl_saved)
2865 return;
2866
2867 memset(msrs, 0, sizeof(msrs));
2868
2869 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCTL, &value);
2870 if (err)
2871 debugf0("Reading K8_NBCTL failed\n");
2872 value &= ~mask;
2873 value |= pvt->old_nbctl;
2874
2875 /* restore the NB Enable MCGCTL bit */
2876 pci_write_config_dword(pvt->misc_f3_ctl, K8_NBCTL, value);
2877
2878 rdmsr_on_cpus(cpumask, K8_MSR_MCGCTL, msrs);
2879
2880 for_each_cpu(cpu, cpumask) {
2881 msrs[idx].l &= ~K8_MSR_MCGCTL_NBE;
2882 msrs[idx].l |=
2883 test_bit(idx, &pvt->old_mcgctl) << K8_MSR_MCGCTL_NBE;
2884 idx++;
2885 }
2886
2887 wrmsr_on_cpus(cpumask, K8_MSR_MCGCTL, msrs);
2888}
2889
2890static void check_mcg_ctl(void *ret)
2891{
2892 u64 msr_val = 0;
2893 u8 nbe;
2894
2895 rdmsrl(MSR_IA32_MCG_CTL, msr_val);
2896 nbe = msr_val & K8_MSR_MCGCTL_NBE;
2897
2898 debugf0("core: %u, MCG_CTL: 0x%llx, NB MSR is %s\n",
2899 raw_smp_processor_id(), msr_val,
2900 (nbe ? "enabled" : "disabled"));
2901
2902 if (!nbe)
2903 *(int *)ret = 0;
2904}
2905
2906/* check MCG_CTL on all the cpus on this node */
2907static int amd64_mcg_ctl_enabled_on_cpus(const cpumask_t *mask)
2908{
2909 int ret = 1;
2910 preempt_disable();
2911 smp_call_function_many(mask, check_mcg_ctl, &ret, 1);
2912 preempt_enable();
2913
2914 return ret;
2915}
2916
2917/*
2918 * EDAC requires that the BIOS have ECC enabled before taking over the
2919 * processing of ECC errors. This is because the BIOS can properly initialize
2920 * the memory system completely. A command line option allows to force-enable
2921 * hardware ECC later in amd64_enable_ecc_error_reporting().
2922 */
2923static int amd64_check_ecc_enabled(struct amd64_pvt *pvt)
2924{
2925 u32 value;
2926 int err = 0, ret = 0;
2927 u8 ecc_enabled = 0;
2928
2929 err = pci_read_config_dword(pvt->misc_f3_ctl, K8_NBCFG, &value);
2930 if (err)
2931 debugf0("Reading K8_NBCTL failed\n");
2932
2933 ecc_enabled = !!(value & K8_NBCFG_ECC_ENABLE);
2934
2935 ret = amd64_mcg_ctl_enabled_on_cpus(cpumask_of_node(pvt->mc_node_id));
2936
2937 debugf0("K8_NBCFG=0x%x, DRAM ECC is %s\n", value,
2938 (value & K8_NBCFG_ECC_ENABLE ? "enabled" : "disabled"));
2939
2940 if (!ecc_enabled || !ret) {
2941 if (!ecc_enabled) {
2942 amd64_printk(KERN_WARNING, "This node reports that "
2943 "Memory ECC is currently "
2944 "disabled.\n");
2945
2946 amd64_printk(KERN_WARNING, "bit 0x%lx in register "
2947 "F3x%x of the MISC_CONTROL device (%s) "
2948 "should be enabled\n", K8_NBCFG_ECC_ENABLE,
2949 K8_NBCFG, pci_name(pvt->misc_f3_ctl));
2950 }
2951 if (!ret) {
2952 amd64_printk(KERN_WARNING, "bit 0x%016lx in MSR 0x%08x "
2953 "of node %d should be enabled\n",
2954 K8_MSR_MCGCTL_NBE, MSR_IA32_MCG_CTL,
2955 pvt->mc_node_id);
2956 }
2957 if (!ecc_enable_override) {
2958 amd64_printk(KERN_WARNING, "WARNING: ECC is NOT "
2959 "currently enabled by the BIOS. Module "
2960 "will NOT be loaded.\n"
2961 " Either Enable ECC in the BIOS, "
2962 "or use the 'ecc_enable_override' "
2963 "parameter.\n"
2964 " Might be a BIOS bug, if BIOS says "
2965 "ECC is enabled\n"
2966 " Use of the override can cause "
2967 "unknown side effects.\n");
2968 ret = -ENODEV;
Borislav Petkov30c875c2009-06-22 19:42:24 +02002969 } else
2970 /*
2971 * enable further driver loading if ECC enable is
2972 * overridden.
2973 */
2974 ret = 0;
Doug Thompsonf9431992009-04-27 19:46:08 +02002975 } else {
2976 amd64_printk(KERN_INFO,
2977 "ECC is enabled by BIOS, Proceeding "
2978 "with EDAC module initialization\n");
2979
2980 /* CLEAR the override, since BIOS controlled it */
2981 ecc_enable_override = 0;
2982 }
2983
2984 return ret;
2985}
2986
Doug Thompson7d6034d2009-04-27 20:01:01 +02002987struct mcidev_sysfs_attribute sysfs_attrs[ARRAY_SIZE(amd64_dbg_attrs) +
2988 ARRAY_SIZE(amd64_inj_attrs) +
2989 1];
2990
2991struct mcidev_sysfs_attribute terminator = { .attr = { .name = NULL } };
2992
2993static void amd64_set_mc_sysfs_attributes(struct mem_ctl_info *mci)
2994{
2995 unsigned int i = 0, j = 0;
2996
2997 for (; i < ARRAY_SIZE(amd64_dbg_attrs); i++)
2998 sysfs_attrs[i] = amd64_dbg_attrs[i];
2999
3000 for (j = 0; j < ARRAY_SIZE(amd64_inj_attrs); j++, i++)
3001 sysfs_attrs[i] = amd64_inj_attrs[j];
3002
3003 sysfs_attrs[i] = terminator;
3004
3005 mci->mc_driver_sysfs_attributes = sysfs_attrs;
3006}
3007
3008static void amd64_setup_mci_misc_attributes(struct mem_ctl_info *mci)
3009{
3010 struct amd64_pvt *pvt = mci->pvt_info;
3011
3012 mci->mtype_cap = MEM_FLAG_DDR2 | MEM_FLAG_RDDR2;
3013 mci->edac_ctl_cap = EDAC_FLAG_NONE;
Doug Thompson7d6034d2009-04-27 20:01:01 +02003014
3015 if (pvt->nbcap & K8_NBCAP_SECDED)
3016 mci->edac_ctl_cap |= EDAC_FLAG_SECDED;
3017
3018 if (pvt->nbcap & K8_NBCAP_CHIPKILL)
3019 mci->edac_ctl_cap |= EDAC_FLAG_S4ECD4ED;
3020
3021 mci->edac_cap = amd64_determine_edac_cap(pvt);
3022 mci->mod_name = EDAC_MOD_STR;
3023 mci->mod_ver = EDAC_AMD64_VERSION;
3024 mci->ctl_name = get_amd_family_name(pvt->mc_type_index);
3025 mci->dev_name = pci_name(pvt->dram_f2_ctl);
3026 mci->ctl_page_to_phys = NULL;
3027
3028 /* IMPORTANT: Set the polling 'check' function in this module */
3029 mci->edac_check = amd64_check;
3030
3031 /* memory scrubber interface */
3032 mci->set_sdram_scrub_rate = amd64_set_scrub_rate;
3033 mci->get_sdram_scrub_rate = amd64_get_scrub_rate;
3034}
3035
3036/*
3037 * Init stuff for this DRAM Controller device.
3038 *
3039 * Due to a hardware feature on Fam10h CPUs, the Enable Extended Configuration
3040 * Space feature MUST be enabled on ALL Processors prior to actually reading
3041 * from the ECS registers. Since the loading of the module can occur on any
3042 * 'core', and cores don't 'see' all the other processors ECS data when the
3043 * others are NOT enabled. Our solution is to first enable ECS access in this
3044 * routine on all processors, gather some data in a amd64_pvt structure and
3045 * later come back in a finish-setup function to perform that final
3046 * initialization. See also amd64_init_2nd_stage() for that.
3047 */
3048static int amd64_probe_one_instance(struct pci_dev *dram_f2_ctl,
3049 int mc_type_index)
3050{
3051 struct amd64_pvt *pvt = NULL;
3052 int err = 0, ret;
3053
3054 ret = -ENOMEM;
3055 pvt = kzalloc(sizeof(struct amd64_pvt), GFP_KERNEL);
3056 if (!pvt)
3057 goto err_exit;
3058
Borislav Petkov37da0452009-06-10 17:36:57 +02003059 pvt->mc_node_id = get_node_id(dram_f2_ctl);
Doug Thompson7d6034d2009-04-27 20:01:01 +02003060
3061 pvt->dram_f2_ctl = dram_f2_ctl;
3062 pvt->ext_model = boot_cpu_data.x86_model >> 4;
3063 pvt->mc_type_index = mc_type_index;
3064 pvt->ops = family_ops(mc_type_index);
3065 pvt->old_mcgctl = 0;
3066
3067 /*
3068 * We have the dram_f2_ctl device as an argument, now go reserve its
3069 * sibling devices from the PCI system.
3070 */
3071 ret = -ENODEV;
3072 err = amd64_reserve_mc_sibling_devices(pvt, mc_type_index);
3073 if (err)
3074 goto err_free;
3075
3076 ret = -EINVAL;
3077 err = amd64_check_ecc_enabled(pvt);
3078 if (err)
3079 goto err_put;
3080
3081 /*
3082 * Key operation here: setup of HW prior to performing ops on it. Some
3083 * setup is required to access ECS data. After this is performed, the
3084 * 'teardown' function must be called upon error and normal exit paths.
3085 */
3086 if (boot_cpu_data.x86 >= 0x10)
3087 amd64_setup(pvt);
3088
3089 /*
3090 * Save the pointer to the private data for use in 2nd initialization
3091 * stage
3092 */
3093 pvt_lookup[pvt->mc_node_id] = pvt;
3094
3095 return 0;
3096
3097err_put:
3098 amd64_free_mc_sibling_devices(pvt);
3099
3100err_free:
3101 kfree(pvt);
3102
3103err_exit:
3104 return ret;
3105}
3106
3107/*
3108 * This is the finishing stage of the init code. Needs to be performed after all
3109 * MCs' hardware have been prepped for accessing extended config space.
3110 */
3111static int amd64_init_2nd_stage(struct amd64_pvt *pvt)
3112{
3113 int node_id = pvt->mc_node_id;
3114 struct mem_ctl_info *mci;
3115 int ret, err = 0;
3116
3117 amd64_read_mc_registers(pvt);
3118
3119 ret = -ENODEV;
3120 if (pvt->ops->probe_valid_hardware) {
3121 err = pvt->ops->probe_valid_hardware(pvt);
3122 if (err)
3123 goto err_exit;
3124 }
3125
3126 /*
3127 * We need to determine how many memory channels there are. Then use
3128 * that information for calculating the size of the dynamic instance
3129 * tables in the 'mci' structure
3130 */
3131 pvt->channel_count = pvt->ops->early_channel_count(pvt);
3132 if (pvt->channel_count < 0)
3133 goto err_exit;
3134
3135 ret = -ENOMEM;
3136 mci = edac_mc_alloc(0, CHIPSELECT_COUNT, pvt->channel_count, node_id);
3137 if (!mci)
3138 goto err_exit;
3139
3140 mci->pvt_info = pvt;
3141
3142 mci->dev = &pvt->dram_f2_ctl->dev;
3143 amd64_setup_mci_misc_attributes(mci);
3144
3145 if (amd64_init_csrows(mci))
3146 mci->edac_cap = EDAC_FLAG_NONE;
3147
3148 amd64_enable_ecc_error_reporting(mci);
3149 amd64_set_mc_sysfs_attributes(mci);
3150
3151 ret = -ENODEV;
3152 if (edac_mc_add_mc(mci)) {
3153 debugf1("failed edac_mc_add_mc()\n");
3154 goto err_add_mc;
3155 }
3156
3157 mci_lookup[node_id] = mci;
3158 pvt_lookup[node_id] = NULL;
3159 return 0;
3160
3161err_add_mc:
3162 edac_mc_free(mci);
3163
3164err_exit:
3165 debugf0("failure to init 2nd stage: ret=%d\n", ret);
3166
3167 amd64_restore_ecc_error_reporting(pvt);
3168
3169 if (boot_cpu_data.x86 > 0xf)
3170 amd64_teardown(pvt);
3171
3172 amd64_free_mc_sibling_devices(pvt);
3173
3174 kfree(pvt_lookup[pvt->mc_node_id]);
3175 pvt_lookup[node_id] = NULL;
3176
3177 return ret;
3178}
3179
3180
3181static int __devinit amd64_init_one_instance(struct pci_dev *pdev,
3182 const struct pci_device_id *mc_type)
3183{
3184 int ret = 0;
3185
Borislav Petkov37da0452009-06-10 17:36:57 +02003186 debugf0("(MC node=%d,mc_type='%s')\n", get_node_id(pdev),
Doug Thompson7d6034d2009-04-27 20:01:01 +02003187 get_amd_family_name(mc_type->driver_data));
3188
3189 ret = pci_enable_device(pdev);
3190 if (ret < 0)
3191 ret = -EIO;
3192 else
3193 ret = amd64_probe_one_instance(pdev, mc_type->driver_data);
3194
3195 if (ret < 0)
3196 debugf0("ret=%d\n", ret);
3197
3198 return ret;
3199}
3200
3201static void __devexit amd64_remove_one_instance(struct pci_dev *pdev)
3202{
3203 struct mem_ctl_info *mci;
3204 struct amd64_pvt *pvt;
3205
3206 /* Remove from EDAC CORE tracking list */
3207 mci = edac_mc_del_mc(&pdev->dev);
3208 if (!mci)
3209 return;
3210
3211 pvt = mci->pvt_info;
3212
3213 amd64_restore_ecc_error_reporting(pvt);
3214
3215 if (boot_cpu_data.x86 > 0xf)
3216 amd64_teardown(pvt);
3217
3218 amd64_free_mc_sibling_devices(pvt);
3219
3220 kfree(pvt);
3221 mci->pvt_info = NULL;
3222
3223 mci_lookup[pvt->mc_node_id] = NULL;
3224
3225 /* Free the EDAC CORE resources */
3226 edac_mc_free(mci);
3227}
3228
3229/*
3230 * This table is part of the interface for loading drivers for PCI devices. The
3231 * PCI core identifies what devices are on a system during boot, and then
3232 * inquiry this table to see if this driver is for a given device found.
3233 */
3234static const struct pci_device_id amd64_pci_table[] __devinitdata = {
3235 {
3236 .vendor = PCI_VENDOR_ID_AMD,
3237 .device = PCI_DEVICE_ID_AMD_K8_NB_MEMCTL,
3238 .subvendor = PCI_ANY_ID,
3239 .subdevice = PCI_ANY_ID,
3240 .class = 0,
3241 .class_mask = 0,
3242 .driver_data = K8_CPUS
3243 },
3244 {
3245 .vendor = PCI_VENDOR_ID_AMD,
3246 .device = PCI_DEVICE_ID_AMD_10H_NB_DRAM,
3247 .subvendor = PCI_ANY_ID,
3248 .subdevice = PCI_ANY_ID,
3249 .class = 0,
3250 .class_mask = 0,
3251 .driver_data = F10_CPUS
3252 },
3253 {
3254 .vendor = PCI_VENDOR_ID_AMD,
3255 .device = PCI_DEVICE_ID_AMD_11H_NB_DRAM,
3256 .subvendor = PCI_ANY_ID,
3257 .subdevice = PCI_ANY_ID,
3258 .class = 0,
3259 .class_mask = 0,
3260 .driver_data = F11_CPUS
3261 },
3262 {0, }
3263};
3264MODULE_DEVICE_TABLE(pci, amd64_pci_table);
3265
3266static struct pci_driver amd64_pci_driver = {
3267 .name = EDAC_MOD_STR,
3268 .probe = amd64_init_one_instance,
3269 .remove = __devexit_p(amd64_remove_one_instance),
3270 .id_table = amd64_pci_table,
3271};
3272
3273static void amd64_setup_pci_device(void)
3274{
3275 struct mem_ctl_info *mci;
3276 struct amd64_pvt *pvt;
3277
3278 if (amd64_ctl_pci)
3279 return;
3280
3281 mci = mci_lookup[0];
3282 if (mci) {
3283
3284 pvt = mci->pvt_info;
3285 amd64_ctl_pci =
3286 edac_pci_create_generic_ctl(&pvt->dram_f2_ctl->dev,
3287 EDAC_MOD_STR);
3288
3289 if (!amd64_ctl_pci) {
3290 pr_warning("%s(): Unable to create PCI control\n",
3291 __func__);
3292
3293 pr_warning("%s(): PCI error report via EDAC not set\n",
3294 __func__);
3295 }
3296 }
3297}
3298
3299static int __init amd64_edac_init(void)
3300{
3301 int nb, err = -ENODEV;
3302
3303 edac_printk(KERN_INFO, EDAC_MOD_STR, EDAC_AMD64_VERSION "\n");
3304
3305 opstate_init();
3306
3307 if (cache_k8_northbridges() < 0)
3308 goto err_exit;
3309
3310 err = pci_register_driver(&amd64_pci_driver);
3311 if (err)
3312 return err;
3313
3314 /*
3315 * At this point, the array 'pvt_lookup[]' contains pointers to alloc'd
3316 * amd64_pvt structs. These will be used in the 2nd stage init function
3317 * to finish initialization of the MC instances.
3318 */
3319 for (nb = 0; nb < num_k8_northbridges; nb++) {
3320 if (!pvt_lookup[nb])
3321 continue;
3322
3323 err = amd64_init_2nd_stage(pvt_lookup[nb]);
3324 if (err)
Borislav Petkov37da0452009-06-10 17:36:57 +02003325 goto err_2nd_stage;
Doug Thompson7d6034d2009-04-27 20:01:01 +02003326 }
3327
3328 amd64_setup_pci_device();
3329
3330 return 0;
3331
Borislav Petkov37da0452009-06-10 17:36:57 +02003332err_2nd_stage:
3333 debugf0("2nd stage failed\n");
3334
Doug Thompson7d6034d2009-04-27 20:01:01 +02003335err_exit:
Doug Thompson7d6034d2009-04-27 20:01:01 +02003336 pci_unregister_driver(&amd64_pci_driver);
3337
3338 return err;
3339}
3340
3341static void __exit amd64_edac_exit(void)
3342{
3343 if (amd64_ctl_pci)
3344 edac_pci_release_generic_ctl(amd64_ctl_pci);
3345
3346 pci_unregister_driver(&amd64_pci_driver);
3347}
3348
3349module_init(amd64_edac_init);
3350module_exit(amd64_edac_exit);
3351
3352MODULE_LICENSE("GPL");
3353MODULE_AUTHOR("SoftwareBitMaker: Doug Thompson, "
3354 "Dave Peterson, Thayne Harbaugh");
3355MODULE_DESCRIPTION("MC support for AMD64 memory controllers - "
3356 EDAC_AMD64_VERSION);
3357
3358module_param(edac_op_state, int, 0444);
3359MODULE_PARM_DESC(edac_op_state, "EDAC Error Reporting state: 0=Poll,1=NMI");