1 #include "amd64_edac.h"
3 static struct edac_pci_ctl_info *amd64_ctl_pci;
5 static int report_gart_errors;
6 module_param(report_gart_errors, int, 0644);
9 * Set by command line parameter. If BIOS has enabled the ECC, this override is
10 * cleared to prevent re-enabling the hardware by this driver.
12 static int ecc_enable_override;
13 module_param(ecc_enable_override, int, 0644);
15 /* Lookup table for all possible MC control instances */
17 static struct mem_ctl_info *mci_lookup[MAX_NUMNODES];
18 static struct amd64_pvt *pvt_lookup[MAX_NUMNODES];
21 * Memory scrubber control interface. For K8, memory scrubbing is handled by
22 * hardware and can involve L2 cache, dcache as well as the main memory. With
23 * F10, this is extended to L3 cache scrubbing on CPU models sporting that
26 * This causes the "units" for the scrubbing speed to vary from 64 byte blocks
27 * (dram) over to cache lines. This is nasty, so we will use bandwidth in
28 * bytes/sec for the setting.
30 * Currently, we only do dram scrubbing. If the scrubbing is done in software on
31 * other archs, we might not have access to the caches directly.
35 * scan the scrub rate mapping table for a close or matching bandwidth value to
36 * issue. If requested is too big, then use last maximum value found.
38 static int amd64_search_set_scrub_rate(struct pci_dev *ctl, u32 new_bw,
45 * map the configured rate (new_bw) to a value specific to the AMD64
46 * memory controller and apply to register. Search for the first
47 * bandwidth entry that is greater or equal than the setting requested
48 * and program that. If at last entry, turn off DRAM scrubbing.
50 for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
52 * skip scrub rates which aren't recommended
53 * (see F10 BKDG, F3x58)
55 if (scrubrates[i].scrubval < min_scrubrate)
58 if (scrubrates[i].bandwidth <= new_bw)
62 * if no suitable bandwidth found, turn off DRAM scrubbing
63 * entirely by falling back to the last element in the
68 scrubval = scrubrates[i].scrubval;
70 edac_printk(KERN_DEBUG, EDAC_MC,
71 "Setting scrub rate bandwidth: %u\n",
72 scrubrates[i].bandwidth);
74 edac_printk(KERN_DEBUG, EDAC_MC, "Turning scrubbing off.\n");
76 pci_write_bits32(ctl, K8_SCRCTRL, scrubval, 0x001F);
81 static int amd64_set_scrub_rate(struct mem_ctl_info *mci, u32 *bandwidth)
83 struct amd64_pvt *pvt = mci->pvt_info;
84 u32 min_scrubrate = 0x0;
86 switch (boot_cpu_data.x86) {
88 min_scrubrate = K8_MIN_SCRUB_RATE_BITS;
91 min_scrubrate = F10_MIN_SCRUB_RATE_BITS;
94 min_scrubrate = F11_MIN_SCRUB_RATE_BITS;
98 amd64_printk(KERN_ERR, "Unsupported family!\n");
101 return amd64_search_set_scrub_rate(pvt->misc_f3_ctl, *bandwidth,
105 static int amd64_get_scrub_rate(struct mem_ctl_info *mci, u32 *bw)
107 struct amd64_pvt *pvt = mci->pvt_info;
109 int status = -1, i, ret = 0;
111 ret = pci_read_config_dword(pvt->misc_f3_ctl, K8_SCRCTRL, &scrubval);
113 debugf0("Reading K8_SCRCTRL failed\n");
115 scrubval = scrubval & 0x001F;
117 edac_printk(KERN_DEBUG, EDAC_MC,
118 "pci-read, sdram scrub control value: %d \n", scrubval);
120 for (i = 0; ARRAY_SIZE(scrubrates); i++) {
121 if (scrubrates[i].scrubval == scrubval) {
122 *bw = scrubrates[i].bandwidth;
131 /* Map from a CSROW entry to the mask entry that operates on it */
132 static inline u32 amd64_map_to_dcs_mask(struct amd64_pvt *pvt, int csrow)
134 return csrow >> (pvt->num_dcsm >> 3);
137 /* return the 'base' address the i'th CS entry of the 'dct' DRAM controller */
138 static u32 amd64_get_dct_base(struct amd64_pvt *pvt, int dct, int csrow)
141 return pvt->dcsb0[csrow];
143 return pvt->dcsb1[csrow];
147 * Return the 'mask' address the i'th CS entry. This function is needed because
148 * there number of DCSM registers on Rev E and prior vs Rev F and later is
151 static u32 amd64_get_dct_mask(struct amd64_pvt *pvt, int dct, int csrow)
154 return pvt->dcsm0[amd64_map_to_dcs_mask(pvt, csrow)];
156 return pvt->dcsm1[amd64_map_to_dcs_mask(pvt, csrow)];
161 * In *base and *limit, pass back the full 40-bit base and limit physical
162 * addresses for the node given by node_id. This information is obtained from
163 * DRAM Base (section 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers. The
164 * base and limit addresses are of type SysAddr, as defined at the start of
165 * section 3.4.4 (p. 70). They are the lowest and highest physical addresses
166 * in the address range they represent.
168 static void amd64_get_base_and_limit(struct amd64_pvt *pvt, int node_id,
169 u64 *base, u64 *limit)
171 *base = pvt->dram_base[node_id];
172 *limit = pvt->dram_limit[node_id];
176 * Return 1 if the SysAddr given by sys_addr matches the base/limit associated
179 static int amd64_base_limit_match(struct amd64_pvt *pvt,
180 u64 sys_addr, int node_id)
182 u64 base, limit, addr;
184 amd64_get_base_and_limit(pvt, node_id, &base, &limit);
186 /* The K8 treats this as a 40-bit value. However, bits 63-40 will be
187 * all ones if the most significant implemented address bit is 1.
188 * Here we discard bits 63-40. See section 3.4.2 of AMD publication
189 * 24592: AMD x86-64 Architecture Programmer's Manual Volume 1
190 * Application Programming.
192 addr = sys_addr & 0x000000ffffffffffull;
194 return (addr >= base) && (addr <= limit);
198 * Attempt to map a SysAddr to a node. On success, return a pointer to the
199 * mem_ctl_info structure for the node that the SysAddr maps to.
201 * On failure, return NULL.
203 static struct mem_ctl_info *find_mc_by_sys_addr(struct mem_ctl_info *mci,
206 struct amd64_pvt *pvt;
211 * Here we use the DRAM Base (section 3.4.4.1) and DRAM Limit (section
212 * 3.4.4.2) registers to map the SysAddr to a node ID.
217 * The value of this field should be the same for all DRAM Base
218 * registers. Therefore we arbitrarily choose to read it from the
219 * register for node 0.
221 intlv_en = pvt->dram_IntlvEn[0];
224 for (node_id = 0; ; ) {
225 if (amd64_base_limit_match(pvt, sys_addr, node_id))
228 if (++node_id >= DRAM_REG_COUNT)
234 if (unlikely((intlv_en != (0x01 << 8)) &&
235 (intlv_en != (0x03 << 8)) &&
236 (intlv_en != (0x07 << 8)))) {
237 amd64_printk(KERN_WARNING, "junk value of 0x%x extracted from "
238 "IntlvEn field of DRAM Base Register for node 0: "
239 "This probably indicates a BIOS bug.\n", intlv_en);
243 bits = (((u32) sys_addr) >> 12) & intlv_en;
245 for (node_id = 0; ; ) {
246 if ((pvt->dram_limit[node_id] & intlv_en) == bits)
247 break; /* intlv_sel field matches */
249 if (++node_id >= DRAM_REG_COUNT)
253 /* sanity test for sys_addr */
254 if (unlikely(!amd64_base_limit_match(pvt, sys_addr, node_id))) {
255 amd64_printk(KERN_WARNING,
256 "%s(): sys_addr 0x%lx falls outside base/limit "
257 "address range for node %d with node interleaving "
258 "enabled.\n", __func__, (unsigned long)sys_addr,
264 return edac_mc_find(node_id);
267 debugf2("sys_addr 0x%lx doesn't match any node\n",
268 (unsigned long)sys_addr);
274 * Extract the DRAM CS base address from selected csrow register.
276 static u64 base_from_dct_base(struct amd64_pvt *pvt, int csrow)
278 return ((u64) (amd64_get_dct_base(pvt, 0, csrow) & pvt->dcsb_base)) <<
283 * Extract the mask from the dcsb0[csrow] entry in a CPU revision-specific way.
285 static u64 mask_from_dct_mask(struct amd64_pvt *pvt, int csrow)
287 u64 dcsm_bits, other_bits;
290 /* Extract bits from DRAM CS Mask. */
291 dcsm_bits = amd64_get_dct_mask(pvt, 0, csrow) & pvt->dcsm_mask;
293 other_bits = pvt->dcsm_mask;
294 other_bits = ~(other_bits << pvt->dcs_shift);
297 * The extracted bits from DCSM belong in the spaces represented by
298 * the cleared bits in other_bits.
300 mask = (dcsm_bits << pvt->dcs_shift) | other_bits;
306 * @input_addr is an InputAddr associated with the node given by mci. Return the
307 * csrow that input_addr maps to, or -1 on failure (no csrow claims input_addr).
309 static int input_addr_to_csrow(struct mem_ctl_info *mci, u64 input_addr)
311 struct amd64_pvt *pvt;
318 * Here we use the DRAM CS Base and DRAM CS Mask registers. For each CS
319 * base/mask register pair, test the condition shown near the start of
320 * section 3.5.4 (p. 84, BKDG #26094, K8, revA-E).
322 for (csrow = 0; csrow < CHIPSELECT_COUNT; csrow++) {
324 /* This DRAM chip select is disabled on this node */
325 if ((pvt->dcsb0[csrow] & K8_DCSB_CS_ENABLE) == 0)
328 base = base_from_dct_base(pvt, csrow);
329 mask = ~mask_from_dct_mask(pvt, csrow);
331 if ((input_addr & mask) == (base & mask)) {
332 debugf2("InputAddr 0x%lx matches csrow %d (node %d)\n",
333 (unsigned long)input_addr, csrow,
340 debugf2("no matching csrow for InputAddr 0x%lx (MC node %d)\n",
341 (unsigned long)input_addr, pvt->mc_node_id);
347 * Return the base value defined by the DRAM Base register for the node
348 * represented by mci. This function returns the full 40-bit value despite the
349 * fact that the register only stores bits 39-24 of the value. See section
350 * 3.4.4.1 (BKDG #26094, K8, revA-E)
352 static inline u64 get_dram_base(struct mem_ctl_info *mci)
354 struct amd64_pvt *pvt = mci->pvt_info;
356 return pvt->dram_base[pvt->mc_node_id];
360 * Obtain info from the DRAM Hole Address Register (section 3.4.8, pub #26094)
361 * for the node represented by mci. Info is passed back in *hole_base,
362 * *hole_offset, and *hole_size. Function returns 0 if info is valid or 1 if
363 * info is invalid. Info may be invalid for either of the following reasons:
365 * - The revision of the node is not E or greater. In this case, the DRAM Hole
366 * Address Register does not exist.
368 * - The DramHoleValid bit is cleared in the DRAM Hole Address Register,
369 * indicating that its contents are not valid.
371 * The values passed back in *hole_base, *hole_offset, and *hole_size are
372 * complete 32-bit values despite the fact that the bitfields in the DHAR
373 * only represent bits 31-24 of the base and offset values.
375 int amd64_get_dram_hole_info(struct mem_ctl_info *mci, u64 *hole_base,
376 u64 *hole_offset, u64 *hole_size)
378 struct amd64_pvt *pvt = mci->pvt_info;
381 /* only revE and later have the DRAM Hole Address Register */
382 if (boot_cpu_data.x86 == 0xf && pvt->ext_model < OPTERON_CPU_REV_E) {
383 debugf1(" revision %d for node %d does not support DHAR\n",
384 pvt->ext_model, pvt->mc_node_id);
388 /* only valid for Fam10h */
389 if (boot_cpu_data.x86 == 0x10 &&
390 (pvt->dhar & F10_DRAM_MEM_HOIST_VALID) == 0) {
391 debugf1(" Dram Memory Hoisting is DISABLED on this system\n");
395 if ((pvt->dhar & DHAR_VALID) == 0) {
396 debugf1(" Dram Memory Hoisting is DISABLED on this node %d\n",
401 /* This node has Memory Hoisting */
403 /* +------------------+--------------------+--------------------+-----
404 * | memory | DRAM hole | relocated |
405 * | [0, (x - 1)] | [x, 0xffffffff] | addresses from |
407 * | | | [0x100000000, |
408 * | | | (0x100000000+ |
409 * | | | (0xffffffff-x))] |
410 * +------------------+--------------------+--------------------+-----
412 * Above is a diagram of physical memory showing the DRAM hole and the
413 * relocated addresses from the DRAM hole. As shown, the DRAM hole
414 * starts at address x (the base address) and extends through address
415 * 0xffffffff. The DRAM Hole Address Register (DHAR) relocates the
416 * addresses in the hole so that they start at 0x100000000.
419 base = dhar_base(pvt->dhar);
422 *hole_size = (0x1ull << 32) - base;
424 if (boot_cpu_data.x86 > 0xf)
425 *hole_offset = f10_dhar_offset(pvt->dhar);
427 *hole_offset = k8_dhar_offset(pvt->dhar);
429 debugf1(" DHAR info for node %d base 0x%lx offset 0x%lx size 0x%lx\n",
430 pvt->mc_node_id, (unsigned long)*hole_base,
431 (unsigned long)*hole_offset, (unsigned long)*hole_size);
435 EXPORT_SYMBOL_GPL(amd64_get_dram_hole_info);
438 * Return the DramAddr that the SysAddr given by @sys_addr maps to. It is
439 * assumed that sys_addr maps to the node given by mci.
441 * The first part of section 3.4.4 (p. 70) shows how the DRAM Base (section
442 * 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers are used to translate a
443 * SysAddr to a DramAddr. If the DRAM Hole Address Register (DHAR) is enabled,
444 * then it is also involved in translating a SysAddr to a DramAddr. Sections
445 * 3.4.8 and 3.5.8.2 describe the DHAR and how it is used for memory hoisting.
446 * These parts of the documentation are unclear. I interpret them as follows:
448 * When node n receives a SysAddr, it processes the SysAddr as follows:
450 * 1. It extracts the DRAMBase and DRAMLimit values from the DRAM Base and DRAM
451 * Limit registers for node n. If the SysAddr is not within the range
452 * specified by the base and limit values, then node n ignores the Sysaddr
453 * (since it does not map to node n). Otherwise continue to step 2 below.
455 * 2. If the DramHoleValid bit of the DHAR for node n is clear, the DHAR is
456 * disabled so skip to step 3 below. Otherwise see if the SysAddr is within
457 * the range of relocated addresses (starting at 0x100000000) from the DRAM
458 * hole. If not, skip to step 3 below. Else get the value of the
459 * DramHoleOffset field from the DHAR. To obtain the DramAddr, subtract the
460 * offset defined by this value from the SysAddr.
462 * 3. Obtain the base address for node n from the DRAMBase field of the DRAM
463 * Base register for node n. To obtain the DramAddr, subtract the base
464 * address from the SysAddr, as shown near the start of section 3.4.4 (p.70).
466 static u64 sys_addr_to_dram_addr(struct mem_ctl_info *mci, u64 sys_addr)
468 u64 dram_base, hole_base, hole_offset, hole_size, dram_addr;
471 dram_base = get_dram_base(mci);
473 ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
476 if ((sys_addr >= (1ull << 32)) &&
477 (sys_addr < ((1ull << 32) + hole_size))) {
478 /* use DHAR to translate SysAddr to DramAddr */
479 dram_addr = sys_addr - hole_offset;
481 debugf2("using DHAR to translate SysAddr 0x%lx to "
483 (unsigned long)sys_addr,
484 (unsigned long)dram_addr);
491 * Translate the SysAddr to a DramAddr as shown near the start of
492 * section 3.4.4 (p. 70). Although sys_addr is a 64-bit value, the k8
493 * only deals with 40-bit values. Therefore we discard bits 63-40 of
494 * sys_addr below. If bit 39 of sys_addr is 1 then the bits we
495 * discard are all 1s. Otherwise the bits we discard are all 0s. See
496 * section 3.4.2 of AMD publication 24592: AMD x86-64 Architecture
497 * Programmer's Manual Volume 1 Application Programming.
499 dram_addr = (sys_addr & 0xffffffffffull) - dram_base;
501 debugf2("using DRAM Base register to translate SysAddr 0x%lx to "
502 "DramAddr 0x%lx\n", (unsigned long)sys_addr,
503 (unsigned long)dram_addr);
508 * @intlv_en is the value of the IntlvEn field from a DRAM Base register
509 * (section 3.4.4.1). Return the number of bits from a SysAddr that are used
510 * for node interleaving.
512 static int num_node_interleave_bits(unsigned intlv_en)
514 static const int intlv_shift_table[] = { 0, 1, 0, 2, 0, 0, 0, 3 };
517 BUG_ON(intlv_en > 7);
518 n = intlv_shift_table[intlv_en];
522 /* Translate the DramAddr given by @dram_addr to an InputAddr. */
523 static u64 dram_addr_to_input_addr(struct mem_ctl_info *mci, u64 dram_addr)
525 struct amd64_pvt *pvt;
532 * See the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
533 * concerning translating a DramAddr to an InputAddr.
535 intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
536 input_addr = ((dram_addr >> intlv_shift) & 0xffffff000ull) +
539 debugf2(" Intlv Shift=%d DramAddr=0x%lx maps to InputAddr=0x%lx\n",
540 intlv_shift, (unsigned long)dram_addr,
541 (unsigned long)input_addr);
547 * Translate the SysAddr represented by @sys_addr to an InputAddr. It is
548 * assumed that @sys_addr maps to the node given by mci.
550 static u64 sys_addr_to_input_addr(struct mem_ctl_info *mci, u64 sys_addr)
555 dram_addr_to_input_addr(mci, sys_addr_to_dram_addr(mci, sys_addr));
557 debugf2("SysAdddr 0x%lx translates to InputAddr 0x%lx\n",
558 (unsigned long)sys_addr, (unsigned long)input_addr);
565 * @input_addr is an InputAddr associated with the node represented by mci.
566 * Translate @input_addr to a DramAddr and return the result.
568 static u64 input_addr_to_dram_addr(struct mem_ctl_info *mci, u64 input_addr)
570 struct amd64_pvt *pvt;
571 int node_id, intlv_shift;
576 * Near the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
577 * shows how to translate a DramAddr to an InputAddr. Here we reverse
578 * this procedure. When translating from a DramAddr to an InputAddr, the
579 * bits used for node interleaving are discarded. Here we recover these
580 * bits from the IntlvSel field of the DRAM Limit register (section
581 * 3.4.4.2) for the node that input_addr is associated with.
584 node_id = pvt->mc_node_id;
585 BUG_ON((node_id < 0) || (node_id > 7));
587 intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
589 if (intlv_shift == 0) {
590 debugf1(" InputAddr 0x%lx translates to DramAddr of "
591 "same value\n", (unsigned long)input_addr);
596 bits = ((input_addr & 0xffffff000ull) << intlv_shift) +
597 (input_addr & 0xfff);
599 intlv_sel = pvt->dram_IntlvSel[node_id] & ((1 << intlv_shift) - 1);
600 dram_addr = bits + (intlv_sel << 12);
602 debugf1("InputAddr 0x%lx translates to DramAddr 0x%lx "
603 "(%d node interleave bits)\n", (unsigned long)input_addr,
604 (unsigned long)dram_addr, intlv_shift);
610 * @dram_addr is a DramAddr that maps to the node represented by mci. Convert
611 * @dram_addr to a SysAddr.
613 static u64 dram_addr_to_sys_addr(struct mem_ctl_info *mci, u64 dram_addr)
615 struct amd64_pvt *pvt = mci->pvt_info;
616 u64 hole_base, hole_offset, hole_size, base, limit, sys_addr;
619 ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
622 if ((dram_addr >= hole_base) &&
623 (dram_addr < (hole_base + hole_size))) {
624 sys_addr = dram_addr + hole_offset;
626 debugf1("using DHAR to translate DramAddr 0x%lx to "
627 "SysAddr 0x%lx\n", (unsigned long)dram_addr,
628 (unsigned long)sys_addr);
634 amd64_get_base_and_limit(pvt, pvt->mc_node_id, &base, &limit);
635 sys_addr = dram_addr + base;
638 * The sys_addr we have computed up to this point is a 40-bit value
639 * because the k8 deals with 40-bit values. However, the value we are
640 * supposed to return is a full 64-bit physical address. The AMD
641 * x86-64 architecture specifies that the most significant implemented
642 * address bit through bit 63 of a physical address must be either all
643 * 0s or all 1s. Therefore we sign-extend the 40-bit sys_addr to a
644 * 64-bit value below. See section 3.4.2 of AMD publication 24592:
645 * AMD x86-64 Architecture Programmer's Manual Volume 1 Application
648 sys_addr |= ~((sys_addr & (1ull << 39)) - 1);
650 debugf1(" Node %d, DramAddr 0x%lx to SysAddr 0x%lx\n",
651 pvt->mc_node_id, (unsigned long)dram_addr,
652 (unsigned long)sys_addr);
658 * @input_addr is an InputAddr associated with the node given by mci. Translate
659 * @input_addr to a SysAddr.
661 static inline u64 input_addr_to_sys_addr(struct mem_ctl_info *mci,
664 return dram_addr_to_sys_addr(mci,
665 input_addr_to_dram_addr(mci, input_addr));
669 * Find the minimum and maximum InputAddr values that map to the given @csrow.
670 * Pass back these values in *input_addr_min and *input_addr_max.
672 static void find_csrow_limits(struct mem_ctl_info *mci, int csrow,
673 u64 *input_addr_min, u64 *input_addr_max)
675 struct amd64_pvt *pvt;
679 BUG_ON((csrow < 0) || (csrow >= CHIPSELECT_COUNT));
681 base = base_from_dct_base(pvt, csrow);
682 mask = mask_from_dct_mask(pvt, csrow);
684 *input_addr_min = base & ~mask;
685 *input_addr_max = base | mask | pvt->dcs_mask_notused;
689 * Extract error address from MCA NB Address Low (section 3.6.4.5) and MCA NB
690 * Address High (section 3.6.4.6) register values and return the result. Address
691 * is located in the info structure (nbeah and nbeal), the encoding is device
694 static u64 extract_error_address(struct mem_ctl_info *mci,
695 struct amd64_error_info_regs *info)
697 struct amd64_pvt *pvt = mci->pvt_info;
699 return pvt->ops->get_error_address(mci, info);
703 /* Map the Error address to a PAGE and PAGE OFFSET. */
704 static inline void error_address_to_page_and_offset(u64 error_address,
705 u32 *page, u32 *offset)
707 *page = (u32) (error_address >> PAGE_SHIFT);
708 *offset = ((u32) error_address) & ~PAGE_MASK;
712 * @sys_addr is an error address (a SysAddr) extracted from the MCA NB Address
713 * Low (section 3.6.4.5) and MCA NB Address High (section 3.6.4.6) registers
714 * of a node that detected an ECC memory error. mci represents the node that
715 * the error address maps to (possibly different from the node that detected
716 * the error). Return the number of the csrow that sys_addr maps to, or -1 on
719 static int sys_addr_to_csrow(struct mem_ctl_info *mci, u64 sys_addr)
723 csrow = input_addr_to_csrow(mci, sys_addr_to_input_addr(mci, sys_addr));
726 amd64_mc_printk(mci, KERN_ERR,
727 "Failed to translate InputAddr to csrow for "
728 "address 0x%lx\n", (unsigned long)sys_addr);
732 static int get_channel_from_ecc_syndrome(unsigned short syndrome);
734 static void amd64_cpu_display_info(struct amd64_pvt *pvt)
736 if (boot_cpu_data.x86 == 0x11)
737 edac_printk(KERN_DEBUG, EDAC_MC, "F11h CPU detected\n");
738 else if (boot_cpu_data.x86 == 0x10)
739 edac_printk(KERN_DEBUG, EDAC_MC, "F10h CPU detected\n");
740 else if (boot_cpu_data.x86 == 0xf)
741 edac_printk(KERN_DEBUG, EDAC_MC, "%s detected\n",
742 (pvt->ext_model >= OPTERON_CPU_REV_F) ?
743 "Rev F or later" : "Rev E or earlier");
745 /* we'll hardly ever ever get here */
746 edac_printk(KERN_ERR, EDAC_MC, "Unknown cpu!\n");
750 * Determine if the DIMMs have ECC enabled. ECC is enabled ONLY if all the DIMMs
753 static enum edac_type amd64_determine_edac_cap(struct amd64_pvt *pvt)
756 enum dev_type edac_cap = EDAC_NONE;
758 bit = (boot_cpu_data.x86 > 0xf || pvt->ext_model >= OPTERON_CPU_REV_F)
762 if (pvt->dclr0 >> BIT(bit))
763 edac_cap = EDAC_FLAG_SECDED;
769 static void f10_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt,
772 /* Display and decode various NB registers for debug purposes. */
773 static void amd64_dump_misc_regs(struct amd64_pvt *pvt)
777 debugf1(" nbcap:0x%8.08x DctDualCap=%s DualNode=%s 8-Node=%s\n",
779 (pvt->nbcap & K8_NBCAP_DCT_DUAL) ? "True" : "False",
780 (pvt->nbcap & K8_NBCAP_DUAL_NODE) ? "True" : "False",
781 (pvt->nbcap & K8_NBCAP_8_NODE) ? "True" : "False");
782 debugf1(" ECC Capable=%s ChipKill Capable=%s\n",
783 (pvt->nbcap & K8_NBCAP_SECDED) ? "True" : "False",
784 (pvt->nbcap & K8_NBCAP_CHIPKILL) ? "True" : "False");
785 debugf1(" DramCfg0-low=0x%08x DIMM-ECC=%s Parity=%s Width=%s\n",
787 (pvt->dclr0 & BIT(19)) ? "Enabled" : "Disabled",
788 (pvt->dclr0 & BIT(8)) ? "Enabled" : "Disabled",
789 (pvt->dclr0 & BIT(11)) ? "128b" : "64b");
790 debugf1(" DIMM x4 Present: L0=%s L1=%s L2=%s L3=%s DIMM Type=%s\n",
791 (pvt->dclr0 & BIT(12)) ? "Y" : "N",
792 (pvt->dclr0 & BIT(13)) ? "Y" : "N",
793 (pvt->dclr0 & BIT(14)) ? "Y" : "N",
794 (pvt->dclr0 & BIT(15)) ? "Y" : "N",
795 (pvt->dclr0 & BIT(16)) ? "UN-Buffered" : "Buffered");
798 debugf1(" online-spare: 0x%8.08x\n", pvt->online_spare);
800 if (boot_cpu_data.x86 == 0xf) {
801 debugf1(" dhar: 0x%8.08x Base=0x%08x Offset=0x%08x\n",
802 pvt->dhar, dhar_base(pvt->dhar),
803 k8_dhar_offset(pvt->dhar));
804 debugf1(" DramHoleValid=%s\n",
805 (pvt->dhar & DHAR_VALID) ? "True" : "False");
807 debugf1(" dbam-dkt: 0x%8.08x\n", pvt->dbam0);
809 /* everything below this point is Fam10h and above */
813 debugf1(" dhar: 0x%8.08x Base=0x%08x Offset=0x%08x\n",
814 pvt->dhar, dhar_base(pvt->dhar),
815 f10_dhar_offset(pvt->dhar));
816 debugf1(" DramMemHoistValid=%s DramHoleValid=%s\n",
817 (pvt->dhar & F10_DRAM_MEM_HOIST_VALID) ?
819 (pvt->dhar & DHAR_VALID) ?
823 /* Only if NOT ganged does dcl1 have valid info */
824 if (!dct_ganging_enabled(pvt)) {
825 debugf1(" DramCfg1-low=0x%08x DIMM-ECC=%s Parity=%s "
826 "Width=%s\n", pvt->dclr1,
827 (pvt->dclr1 & BIT(19)) ? "Enabled" : "Disabled",
828 (pvt->dclr1 & BIT(8)) ? "Enabled" : "Disabled",
829 (pvt->dclr1 & BIT(11)) ? "128b" : "64b");
830 debugf1(" DIMM x4 Present: L0=%s L1=%s L2=%s L3=%s "
832 (pvt->dclr1 & BIT(12)) ? "Y" : "N",
833 (pvt->dclr1 & BIT(13)) ? "Y" : "N",
834 (pvt->dclr1 & BIT(14)) ? "Y" : "N",
835 (pvt->dclr1 & BIT(15)) ? "Y" : "N",
836 (pvt->dclr1 & BIT(16)) ? "UN-Buffered" : "Buffered");
840 * Determine if ganged and then dump memory sizes for first controller,
841 * and if NOT ganged dump info for 2nd controller.
843 ganged = dct_ganging_enabled(pvt);
845 f10_debug_display_dimm_sizes(0, pvt, ganged);
848 f10_debug_display_dimm_sizes(1, pvt, ganged);
851 /* Read in both of DBAM registers */
852 static void amd64_read_dbam_reg(struct amd64_pvt *pvt)
858 err = pci_read_config_dword(pvt->dram_f2_ctl, reg, &pvt->dbam0);
862 if (boot_cpu_data.x86 >= 0x10) {
864 err = pci_read_config_dword(pvt->dram_f2_ctl, reg, &pvt->dbam1);
871 debugf0("Error reading F2x%03x.\n", reg);
875 * NOTE: CPU Revision Dependent code: Rev E and Rev F
877 * Set the DCSB and DCSM mask values depending on the CPU revision value. Also
878 * set the shift factor for the DCSB and DCSM values.
880 * ->dcs_mask_notused, RevE:
882 * To find the max InputAddr for the csrow, start with the base address and set
883 * all bits that are "don't care" bits in the test at the start of section
886 * The "don't care" bits are all set bits in the mask and all bits in the gaps
887 * between bit ranges [35:25] and [19:13]. The value REV_E_DCS_NOTUSED_BITS
888 * represents bits [24:20] and [12:0], which are all bits in the above-mentioned
891 * ->dcs_mask_notused, RevF and later:
893 * To find the max InputAddr for the csrow, start with the base address and set
894 * all bits that are "don't care" bits in the test at the start of NPT section
897 * The "don't care" bits are all set bits in the mask and all bits in the gaps
898 * between bit ranges [36:27] and [21:13].
900 * The value REV_F_F1Xh_DCS_NOTUSED_BITS represents bits [26:22] and [12:0],
901 * which are all bits in the above-mentioned gaps.
903 static void amd64_set_dct_base_and_mask(struct amd64_pvt *pvt)
905 if (pvt->ext_model >= OPTERON_CPU_REV_F) {
906 pvt->dcsb_base = REV_F_F1Xh_DCSB_BASE_BITS;
907 pvt->dcsm_mask = REV_F_F1Xh_DCSM_MASK_BITS;
908 pvt->dcs_mask_notused = REV_F_F1Xh_DCS_NOTUSED_BITS;
909 pvt->dcs_shift = REV_F_F1Xh_DCS_SHIFT;
911 switch (boot_cpu_data.x86) {
913 pvt->num_dcsm = REV_F_DCSM_COUNT;
917 pvt->num_dcsm = F10_DCSM_COUNT;
921 pvt->num_dcsm = F11_DCSM_COUNT;
925 amd64_printk(KERN_ERR, "Unsupported family!\n");
929 pvt->dcsb_base = REV_E_DCSB_BASE_BITS;
930 pvt->dcsm_mask = REV_E_DCSM_MASK_BITS;
931 pvt->dcs_mask_notused = REV_E_DCS_NOTUSED_BITS;
932 pvt->dcs_shift = REV_E_DCS_SHIFT;
933 pvt->num_dcsm = REV_E_DCSM_COUNT;
938 * Function 2 Offset F10_DCSB0; read in the DCS Base and DCS Mask hw registers
940 static void amd64_read_dct_base_mask(struct amd64_pvt *pvt)
942 int cs, reg, err = 0;
944 amd64_set_dct_base_and_mask(pvt);
946 for (cs = 0; cs < CHIPSELECT_COUNT; cs++) {
947 reg = K8_DCSB0 + (cs * 4);
948 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
951 debugf0("Reading K8_DCSB0[%d] failed\n", cs);
953 debugf0(" DCSB0[%d]=0x%08x reg: F2x%x\n",
954 cs, pvt->dcsb0[cs], reg);
956 /* If DCT are NOT ganged, then read in DCT1's base */
957 if (boot_cpu_data.x86 >= 0x10 && !dct_ganging_enabled(pvt)) {
958 reg = F10_DCSB1 + (cs * 4);
959 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
962 debugf0("Reading F10_DCSB1[%d] failed\n", cs);
964 debugf0(" DCSB1[%d]=0x%08x reg: F2x%x\n",
965 cs, pvt->dcsb1[cs], reg);
971 for (cs = 0; cs < pvt->num_dcsm; cs++) {
972 reg = K8_DCSB0 + (cs * 4);
973 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
976 debugf0("Reading K8_DCSM0 failed\n");
978 debugf0(" DCSM0[%d]=0x%08x reg: F2x%x\n",
979 cs, pvt->dcsm0[cs], reg);
981 /* If DCT are NOT ganged, then read in DCT1's mask */
982 if (boot_cpu_data.x86 >= 0x10 && !dct_ganging_enabled(pvt)) {
983 reg = F10_DCSM1 + (cs * 4);
984 err = pci_read_config_dword(pvt->dram_f2_ctl, reg,
987 debugf0("Reading F10_DCSM1[%d] failed\n", cs);
989 debugf0(" DCSM1[%d]=0x%08x reg: F2x%x\n",
990 cs, pvt->dcsm1[cs], reg);
996 static enum mem_type amd64_determine_memory_type(struct amd64_pvt *pvt)
1000 if (boot_cpu_data.x86 >= 0x10 || pvt->ext_model >= OPTERON_CPU_REV_F) {
1001 /* Rev F and later */
1002 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR2 : MEM_RDDR2;
1004 /* Rev E and earlier */
1005 type = (pvt->dclr0 & BIT(18)) ? MEM_DDR : MEM_RDDR;
1008 debugf1(" Memory type is: %s\n",
1009 (type == MEM_DDR2) ? "MEM_DDR2" :
1010 (type == MEM_RDDR2) ? "MEM_RDDR2" :
1011 (type == MEM_DDR) ? "MEM_DDR" : "MEM_RDDR");
1017 * Read the DRAM Configuration Low register. It differs between CG, D & E revs
1018 * and the later RevF memory controllers (DDR vs DDR2)
1021 * number of memory channels in operation
1023 * contents of the DCL0_LOW register
1025 static int k8_early_channel_count(struct amd64_pvt *pvt)
1029 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_0, &pvt->dclr0);
1033 if ((boot_cpu_data.x86_model >> 4) >= OPTERON_CPU_REV_F) {
1034 /* RevF (NPT) and later */
1035 flag = pvt->dclr0 & F10_WIDTH_128;
1037 /* RevE and earlier */
1038 flag = pvt->dclr0 & REVE_WIDTH_128;
1044 return (flag) ? 2 : 1;
1047 /* extract the ERROR ADDRESS for the K8 CPUs */
1048 static u64 k8_get_error_address(struct mem_ctl_info *mci,
1049 struct amd64_error_info_regs *info)
1051 return (((u64) (info->nbeah & 0xff)) << 32) +
1052 (info->nbeal & ~0x03);
1056 * Read the Base and Limit registers for K8 based Memory controllers; extract
1057 * fields from the 'raw' reg into separate data fields
1059 * Isolates: BASE, LIMIT, IntlvEn, IntlvSel, RW_EN
1061 static void k8_read_dram_base_limit(struct amd64_pvt *pvt, int dram)
1064 u32 off = dram << 3; /* 8 bytes between DRAM entries */
1067 err = pci_read_config_dword(pvt->addr_f1_ctl,
1068 K8_DRAM_BASE_LOW + off, &low);
1070 debugf0("Reading K8_DRAM_BASE_LOW failed\n");
1072 /* Extract parts into separate data entries */
1073 pvt->dram_base[dram] = ((u64) low & 0xFFFF0000) << 8;
1074 pvt->dram_IntlvEn[dram] = (low >> 8) & 0x7;
1075 pvt->dram_rw_en[dram] = (low & 0x3);
1077 err = pci_read_config_dword(pvt->addr_f1_ctl,
1078 K8_DRAM_LIMIT_LOW + off, &low);
1080 debugf0("Reading K8_DRAM_LIMIT_LOW failed\n");
1083 * Extract parts into separate data entries. Limit is the HIGHEST memory
1084 * location of the region, so lower 24 bits need to be all ones
1086 pvt->dram_limit[dram] = (((u64) low & 0xFFFF0000) << 8) | 0x00FFFFFF;
1087 pvt->dram_IntlvSel[dram] = (low >> 8) & 0x7;
1088 pvt->dram_DstNode[dram] = (low & 0x7);
1091 static void k8_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
1092 struct amd64_error_info_regs *info,
1095 struct mem_ctl_info *src_mci;
1096 unsigned short syndrome;
1100 /* Extract the syndrome parts and form a 16-bit syndrome */
1101 syndrome = EXTRACT_HIGH_SYNDROME(info->nbsl) << 8;
1102 syndrome |= EXTRACT_LOW_SYNDROME(info->nbsh);
1104 /* CHIPKILL enabled */
1105 if (info->nbcfg & K8_NBCFG_CHIPKILL) {
1106 channel = get_channel_from_ecc_syndrome(syndrome);
1109 * Syndrome didn't map, so we don't know which of the
1110 * 2 DIMMs is in error. So we need to ID 'both' of them
1113 amd64_mc_printk(mci, KERN_WARNING,
1114 "unknown syndrome 0x%x - possible error "
1115 "reporting race\n", syndrome);
1116 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1121 * non-chipkill ecc mode
1123 * The k8 documentation is unclear about how to determine the
1124 * channel number when using non-chipkill memory. This method
1125 * was obtained from email communication with someone at AMD.
1126 * (Wish the email was placed in this comment - norsk)
1128 channel = ((SystemAddress & BIT(3)) != 0);
1132 * Find out which node the error address belongs to. This may be
1133 * different from the node that detected the error.
1135 src_mci = find_mc_by_sys_addr(mci, SystemAddress);
1137 amd64_mc_printk(mci, KERN_ERR,
1138 "failed to map error address 0x%lx to a node\n",
1139 (unsigned long)SystemAddress);
1140 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1144 /* Now map the SystemAddress to a CSROW */
1145 csrow = sys_addr_to_csrow(src_mci, SystemAddress);
1147 edac_mc_handle_ce_no_info(src_mci, EDAC_MOD_STR);
1149 error_address_to_page_and_offset(SystemAddress, &page, &offset);
1151 edac_mc_handle_ce(src_mci, page, offset, syndrome, csrow,
1152 channel, EDAC_MOD_STR);
1157 * determrine the number of PAGES in for this DIMM's size based on its DRAM
1160 * First step is to calc the number of bits to shift a value of 1 left to
1161 * indicate show many pages. Start with the DBAM value as the starting bits,
1162 * then proceed to adjust those shift bits, based on CPU rev and the table.
1163 * See BKDG on the DBAM
1165 static int k8_dbam_map_to_pages(struct amd64_pvt *pvt, int dram_map)
1169 if (pvt->ext_model >= OPTERON_CPU_REV_F) {
1170 nr_pages = 1 << (revf_quad_ddr2_shift[dram_map] - PAGE_SHIFT);
1173 * RevE and less section; this line is tricky. It collapses the
1174 * table used by RevD and later to one that matches revisions CG
1177 dram_map -= (pvt->ext_model >= OPTERON_CPU_REV_D) ?
1178 (dram_map > 8 ? 4 : (dram_map > 5 ?
1179 3 : (dram_map > 2 ? 1 : 0))) : 0;
1181 /* 25 shift is 32MiB minimum DIMM size in RevE and prior */
1182 nr_pages = 1 << (dram_map + 25 - PAGE_SHIFT);
1189 * Get the number of DCT channels in use.
1192 * number of Memory Channels in operation
1194 * contents of the DCL0_LOW register
1196 static int f10_early_channel_count(struct amd64_pvt *pvt)
1198 int err = 0, channels = 0;
1201 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_0, &pvt->dclr0);
1205 err = pci_read_config_dword(pvt->dram_f2_ctl, F10_DCLR_1, &pvt->dclr1);
1209 /* If we are in 128 bit mode, then we are using 2 channels */
1210 if (pvt->dclr0 & F10_WIDTH_128) {
1211 debugf0("Data WIDTH is 128 bits - 2 channels\n");
1217 * Need to check if in UN-ganged mode: In such, there are 2 channels,
1218 * but they are NOT in 128 bit mode and thus the above 'dcl0' status bit
1221 * Need to check DCT0[0] and DCT1[0] to see if only one of them has
1222 * their CSEnable bit on. If so, then SINGLE DIMM case.
1224 debugf0("Data WIDTH is NOT 128 bits - need more decoding\n");
1227 * Check DRAM Bank Address Mapping values for each DIMM to see if there
1228 * is more than just one DIMM present in unganged mode. Need to check
1229 * both controllers since DIMMs can be placed in either one.
1232 err = pci_read_config_dword(pvt->dram_f2_ctl, DBAM0, &dbam);
1236 if (DBAM_DIMM(0, dbam) > 0)
1238 if (DBAM_DIMM(1, dbam) > 0)
1240 if (DBAM_DIMM(2, dbam) > 0)
1242 if (DBAM_DIMM(3, dbam) > 0)
1245 /* If more than 2 DIMMs are present, then we have 2 channels */
1248 else if (channels == 0) {
1249 /* No DIMMs on DCT0, so look at DCT1 */
1250 err = pci_read_config_dword(pvt->dram_f2_ctl, DBAM1, &dbam);
1254 if (DBAM_DIMM(0, dbam) > 0)
1256 if (DBAM_DIMM(1, dbam) > 0)
1258 if (DBAM_DIMM(2, dbam) > 0)
1260 if (DBAM_DIMM(3, dbam) > 0)
1267 /* If we found ALL 0 values, then assume just ONE DIMM-ONE Channel */
1271 debugf0("DIMM count= %d\n", channels);
1280 static int f10_dbam_map_to_pages(struct amd64_pvt *pvt, int dram_map)
1282 return 1 << (revf_quad_ddr2_shift[dram_map] - PAGE_SHIFT);
1285 /* Enable extended configuration access via 0xCF8 feature */
1286 static void amd64_setup(struct amd64_pvt *pvt)
1290 pci_read_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, ®);
1292 pvt->flags.cf8_extcfg = !!(reg & F10_NB_CFG_LOW_ENABLE_EXT_CFG);
1293 reg |= F10_NB_CFG_LOW_ENABLE_EXT_CFG;
1294 pci_write_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, reg);
1297 /* Restore the extended configuration access via 0xCF8 feature */
1298 static void amd64_teardown(struct amd64_pvt *pvt)
1302 pci_read_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, ®);
1304 reg &= ~F10_NB_CFG_LOW_ENABLE_EXT_CFG;
1305 if (pvt->flags.cf8_extcfg)
1306 reg |= F10_NB_CFG_LOW_ENABLE_EXT_CFG;
1307 pci_write_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, reg);
1310 static u64 f10_get_error_address(struct mem_ctl_info *mci,
1311 struct amd64_error_info_regs *info)
1313 return (((u64) (info->nbeah & 0xffff)) << 32) +
1314 (info->nbeal & ~0x01);
1318 * Read the Base and Limit registers for F10 based Memory controllers. Extract
1319 * fields from the 'raw' reg into separate data fields.
1321 * Isolates: BASE, LIMIT, IntlvEn, IntlvSel, RW_EN.
1323 static void f10_read_dram_base_limit(struct amd64_pvt *pvt, int dram)
1325 u32 high_offset, low_offset, high_base, low_base, high_limit, low_limit;
1327 low_offset = K8_DRAM_BASE_LOW + (dram << 3);
1328 high_offset = F10_DRAM_BASE_HIGH + (dram << 3);
1330 /* read the 'raw' DRAM BASE Address register */
1331 pci_read_config_dword(pvt->addr_f1_ctl, low_offset, &low_base);
1333 /* Read from the ECS data register */
1334 pci_read_config_dword(pvt->addr_f1_ctl, high_offset, &high_base);
1336 /* Extract parts into separate data entries */
1337 pvt->dram_rw_en[dram] = (low_base & 0x3);
1339 if (pvt->dram_rw_en[dram] == 0)
1342 pvt->dram_IntlvEn[dram] = (low_base >> 8) & 0x7;
1344 pvt->dram_base[dram] = (((((u64) high_base & 0x000000FF) << 32) |
1345 ((u64) low_base & 0xFFFF0000))) << 8;
1347 low_offset = K8_DRAM_LIMIT_LOW + (dram << 3);
1348 high_offset = F10_DRAM_LIMIT_HIGH + (dram << 3);
1350 /* read the 'raw' LIMIT registers */
1351 pci_read_config_dword(pvt->addr_f1_ctl, low_offset, &low_limit);
1353 /* Read from the ECS data register for the HIGH portion */
1354 pci_read_config_dword(pvt->addr_f1_ctl, high_offset, &high_limit);
1356 debugf0(" HW Regs: BASE=0x%08x-%08x LIMIT= 0x%08x-%08x\n",
1357 high_base, low_base, high_limit, low_limit);
1359 pvt->dram_DstNode[dram] = (low_limit & 0x7);
1360 pvt->dram_IntlvSel[dram] = (low_limit >> 8) & 0x7;
1363 * Extract address values and form a LIMIT address. Limit is the HIGHEST
1364 * memory location of the region, so low 24 bits need to be all ones.
1366 low_limit |= 0x0000FFFF;
1367 pvt->dram_limit[dram] =
1368 ((((u64) high_limit << 32) + (u64) low_limit) << 8) | (0xFF);