1 #include "amd64_edac.h"
2 #include <asm/amd_nb.h>
4 static struct edac_pci_ctl_info *amd64_ctl_pci;
6 static int report_gart_errors;
7 module_param(report_gart_errors, int, 0644);
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.
13 static int ecc_enable_override;
14 module_param(ecc_enable_override, int, 0644);
16 static struct msr __percpu *msrs;
19 * count successfully initialized driver instances for setup_pci_device()
21 static atomic_t drv_instances = ATOMIC_INIT(0);
23 /* Per-node driver instances */
24 static struct mem_ctl_info **mcis;
25 static struct ecc_settings **ecc_stngs;
28 * Address to DRAM bank mapping: see F2x80 for K8 and F2x[1,0]80 for Fam10 and
31 static int ddr2_dbam_revCG[] = {
41 static int ddr2_dbam_revD[] = {
53 static int ddr2_dbam[] = { [0] = 128,
62 static int ddr3_dbam[] = { [0] = -1,
73 * Valid scrub rates for the K8 hardware memory scrubber. We map the scrubbing
74 * bandwidth to a valid bit pattern. The 'set' operation finds the 'matching-
77 *FIXME: Produce a better mapping/linearisation.
82 u32 scrubval; /* bit pattern for scrub rate */
83 u32 bandwidth; /* bandwidth consumed (bytes/sec) */
85 { 0x01, 1600000000UL},
107 { 0x00, 0UL}, /* scrubbing off */
110 static int __amd64_read_pci_cfg_dword(struct pci_dev *pdev, int offset,
111 u32 *val, const char *func)
115 err = pci_read_config_dword(pdev, offset, val);
117 amd64_warn("%s: error reading F%dx%03x.\n",
118 func, PCI_FUNC(pdev->devfn), offset);
123 int __amd64_write_pci_cfg_dword(struct pci_dev *pdev, int offset,
124 u32 val, const char *func)
128 err = pci_write_config_dword(pdev, offset, val);
130 amd64_warn("%s: error writing to F%dx%03x.\n",
131 func, PCI_FUNC(pdev->devfn), offset);
138 * Depending on the family, F2 DCT reads need special handling:
140 * K8: has a single DCT only
142 * F10h: each DCT has its own set of regs
146 * F15h: we select which DCT we access using F1x10C[DctCfgSel]
149 static int k8_read_dct_pci_cfg(struct amd64_pvt *pvt, int addr, u32 *val,
155 return __amd64_read_pci_cfg_dword(pvt->F2, addr, val, func);
158 static int f10_read_dct_pci_cfg(struct amd64_pvt *pvt, int addr, u32 *val,
161 return __amd64_read_pci_cfg_dword(pvt->F2, addr, val, func);
164 static int f15_read_dct_pci_cfg(struct amd64_pvt *pvt, int addr, u32 *val,
170 if (addr >= 0x140 && addr <= 0x1a0) {
175 amd64_read_pci_cfg(pvt->F1, DCT_CFG_SEL, ®);
178 amd64_write_pci_cfg(pvt->F1, DCT_CFG_SEL, reg);
180 return __amd64_read_pci_cfg_dword(pvt->F2, addr, val, func);
184 * Memory scrubber control interface. For K8, memory scrubbing is handled by
185 * hardware and can involve L2 cache, dcache as well as the main memory. With
186 * F10, this is extended to L3 cache scrubbing on CPU models sporting that
189 * This causes the "units" for the scrubbing speed to vary from 64 byte blocks
190 * (dram) over to cache lines. This is nasty, so we will use bandwidth in
191 * bytes/sec for the setting.
193 * Currently, we only do dram scrubbing. If the scrubbing is done in software on
194 * other archs, we might not have access to the caches directly.
198 * scan the scrub rate mapping table for a close or matching bandwidth value to
199 * issue. If requested is too big, then use last maximum value found.
201 static int __amd64_set_scrub_rate(struct pci_dev *ctl, u32 new_bw, u32 min_rate)
207 * map the configured rate (new_bw) to a value specific to the AMD64
208 * memory controller and apply to register. Search for the first
209 * bandwidth entry that is greater or equal than the setting requested
210 * and program that. If at last entry, turn off DRAM scrubbing.
212 for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
214 * skip scrub rates which aren't recommended
215 * (see F10 BKDG, F3x58)
217 if (scrubrates[i].scrubval < min_rate)
220 if (scrubrates[i].bandwidth <= new_bw)
224 * if no suitable bandwidth found, turn off DRAM scrubbing
225 * entirely by falling back to the last element in the
230 scrubval = scrubrates[i].scrubval;
232 pci_write_bits32(ctl, K8_SCRCTRL, scrubval, 0x001F);
235 return scrubrates[i].bandwidth;
240 static int amd64_set_scrub_rate(struct mem_ctl_info *mci, u32 bw)
242 struct amd64_pvt *pvt = mci->pvt_info;
244 return __amd64_set_scrub_rate(pvt->F3, bw, pvt->min_scrubrate);
247 static int amd64_get_scrub_rate(struct mem_ctl_info *mci)
249 struct amd64_pvt *pvt = mci->pvt_info;
251 int i, retval = -EINVAL;
253 amd64_read_pci_cfg(pvt->F3, K8_SCRCTRL, &scrubval);
255 scrubval = scrubval & 0x001F;
257 amd64_debug("pci-read, sdram scrub control value: %d\n", scrubval);
259 for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
260 if (scrubrates[i].scrubval == scrubval) {
261 retval = scrubrates[i].bandwidth;
269 * returns true if the SysAddr given by sys_addr matches the
270 * DRAM base/limit associated with node_id
272 static bool amd64_base_limit_match(struct amd64_pvt *pvt, u64 sys_addr, int nid)
276 /* The K8 treats this as a 40-bit value. However, bits 63-40 will be
277 * all ones if the most significant implemented address bit is 1.
278 * Here we discard bits 63-40. See section 3.4.2 of AMD publication
279 * 24592: AMD x86-64 Architecture Programmer's Manual Volume 1
280 * Application Programming.
282 addr = sys_addr & 0x000000ffffffffffull;
284 return ((addr >= get_dram_base(pvt, nid)) &&
285 (addr <= get_dram_limit(pvt, nid)));
289 * Attempt to map a SysAddr to a node. On success, return a pointer to the
290 * mem_ctl_info structure for the node that the SysAddr maps to.
292 * On failure, return NULL.
294 static struct mem_ctl_info *find_mc_by_sys_addr(struct mem_ctl_info *mci,
297 struct amd64_pvt *pvt;
302 * Here we use the DRAM Base (section 3.4.4.1) and DRAM Limit (section
303 * 3.4.4.2) registers to map the SysAddr to a node ID.
308 * The value of this field should be the same for all DRAM Base
309 * registers. Therefore we arbitrarily choose to read it from the
310 * register for node 0.
312 intlv_en = dram_intlv_en(pvt, 0);
315 for (node_id = 0; node_id < DRAM_RANGES; node_id++) {
316 if (amd64_base_limit_match(pvt, sys_addr, node_id))
322 if (unlikely((intlv_en != 0x01) &&
323 (intlv_en != 0x03) &&
324 (intlv_en != 0x07))) {
325 amd64_warn("DRAM Base[IntlvEn] junk value: 0x%x, BIOS bug?\n", intlv_en);
329 bits = (((u32) sys_addr) >> 12) & intlv_en;
331 for (node_id = 0; ; ) {
332 if ((dram_intlv_sel(pvt, node_id) & intlv_en) == bits)
333 break; /* intlv_sel field matches */
335 if (++node_id >= DRAM_RANGES)
339 /* sanity test for sys_addr */
340 if (unlikely(!amd64_base_limit_match(pvt, sys_addr, node_id))) {
341 amd64_warn("%s: sys_addr 0x%llx falls outside base/limit address"
342 "range for node %d with node interleaving enabled.\n",
343 __func__, sys_addr, node_id);
348 return edac_mc_find(node_id);
351 debugf2("sys_addr 0x%lx doesn't match any node\n",
352 (unsigned long)sys_addr);
358 * compute the CS base address of the @csrow on the DRAM controller @dct.
359 * For details see F2x[5C:40] in the processor's BKDG
361 static void get_cs_base_and_mask(struct amd64_pvt *pvt, int csrow, u8 dct,
362 u64 *base, u64 *mask)
364 u64 csbase, csmask, base_bits, mask_bits;
367 if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_F) {
368 csbase = pvt->csels[dct].csbases[csrow];
369 csmask = pvt->csels[dct].csmasks[csrow];
370 base_bits = GENMASK(21, 31) | GENMASK(9, 15);
371 mask_bits = GENMASK(21, 29) | GENMASK(9, 15);
374 csbase = pvt->csels[dct].csbases[csrow];
375 csmask = pvt->csels[dct].csmasks[csrow >> 1];
378 if (boot_cpu_data.x86 == 0x15)
379 base_bits = mask_bits = GENMASK(19,30) | GENMASK(5,13);
381 base_bits = mask_bits = GENMASK(19,28) | GENMASK(5,13);
384 *base = (csbase & base_bits) << addr_shift;
387 /* poke holes for the csmask */
388 *mask &= ~(mask_bits << addr_shift);
390 *mask |= (csmask & mask_bits) << addr_shift;
393 #define for_each_chip_select(i, dct, pvt) \
394 for (i = 0; i < pvt->csels[dct].b_cnt; i++)
396 #define for_each_chip_select_mask(i, dct, pvt) \
397 for (i = 0; i < pvt->csels[dct].m_cnt; i++)
400 * @input_addr is an InputAddr associated with the node given by mci. Return the
401 * csrow that input_addr maps to, or -1 on failure (no csrow claims input_addr).
403 static int input_addr_to_csrow(struct mem_ctl_info *mci, u64 input_addr)
405 struct amd64_pvt *pvt;
411 for_each_chip_select(csrow, 0, pvt) {
412 if (!csrow_enabled(csrow, 0, pvt))
415 get_cs_base_and_mask(pvt, csrow, 0, &base, &mask);
419 if ((input_addr & mask) == (base & mask)) {
420 debugf2("InputAddr 0x%lx matches csrow %d (node %d)\n",
421 (unsigned long)input_addr, csrow,
427 debugf2("no matching csrow for InputAddr 0x%lx (MC node %d)\n",
428 (unsigned long)input_addr, pvt->mc_node_id);
434 * Obtain info from the DRAM Hole Address Register (section 3.4.8, pub #26094)
435 * for the node represented by mci. Info is passed back in *hole_base,
436 * *hole_offset, and *hole_size. Function returns 0 if info is valid or 1 if
437 * info is invalid. Info may be invalid for either of the following reasons:
439 * - The revision of the node is not E or greater. In this case, the DRAM Hole
440 * Address Register does not exist.
442 * - The DramHoleValid bit is cleared in the DRAM Hole Address Register,
443 * indicating that its contents are not valid.
445 * The values passed back in *hole_base, *hole_offset, and *hole_size are
446 * complete 32-bit values despite the fact that the bitfields in the DHAR
447 * only represent bits 31-24 of the base and offset values.
449 int amd64_get_dram_hole_info(struct mem_ctl_info *mci, u64 *hole_base,
450 u64 *hole_offset, u64 *hole_size)
452 struct amd64_pvt *pvt = mci->pvt_info;
455 /* only revE and later have the DRAM Hole Address Register */
456 if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_E) {
457 debugf1(" revision %d for node %d does not support DHAR\n",
458 pvt->ext_model, pvt->mc_node_id);
462 /* valid for Fam10h and above */
463 if (boot_cpu_data.x86 >= 0x10 && !dhar_mem_hoist_valid(pvt)) {
464 debugf1(" Dram Memory Hoisting is DISABLED on this system\n");
468 if (!dhar_valid(pvt)) {
469 debugf1(" Dram Memory Hoisting is DISABLED on this node %d\n",
474 /* This node has Memory Hoisting */
476 /* +------------------+--------------------+--------------------+-----
477 * | memory | DRAM hole | relocated |
478 * | [0, (x - 1)] | [x, 0xffffffff] | addresses from |
480 * | | | [0x100000000, |
481 * | | | (0x100000000+ |
482 * | | | (0xffffffff-x))] |
483 * +------------------+--------------------+--------------------+-----
485 * Above is a diagram of physical memory showing the DRAM hole and the
486 * relocated addresses from the DRAM hole. As shown, the DRAM hole
487 * starts at address x (the base address) and extends through address
488 * 0xffffffff. The DRAM Hole Address Register (DHAR) relocates the
489 * addresses in the hole so that they start at 0x100000000.
492 base = dhar_base(pvt);
495 *hole_size = (0x1ull << 32) - base;
497 if (boot_cpu_data.x86 > 0xf)
498 *hole_offset = f10_dhar_offset(pvt);
500 *hole_offset = k8_dhar_offset(pvt);
502 debugf1(" DHAR info for node %d base 0x%lx offset 0x%lx size 0x%lx\n",
503 pvt->mc_node_id, (unsigned long)*hole_base,
504 (unsigned long)*hole_offset, (unsigned long)*hole_size);
508 EXPORT_SYMBOL_GPL(amd64_get_dram_hole_info);
511 * Return the DramAddr that the SysAddr given by @sys_addr maps to. It is
512 * assumed that sys_addr maps to the node given by mci.
514 * The first part of section 3.4.4 (p. 70) shows how the DRAM Base (section
515 * 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers are used to translate a
516 * SysAddr to a DramAddr. If the DRAM Hole Address Register (DHAR) is enabled,
517 * then it is also involved in translating a SysAddr to a DramAddr. Sections
518 * 3.4.8 and 3.5.8.2 describe the DHAR and how it is used for memory hoisting.
519 * These parts of the documentation are unclear. I interpret them as follows:
521 * When node n receives a SysAddr, it processes the SysAddr as follows:
523 * 1. It extracts the DRAMBase and DRAMLimit values from the DRAM Base and DRAM
524 * Limit registers for node n. If the SysAddr is not within the range
525 * specified by the base and limit values, then node n ignores the Sysaddr
526 * (since it does not map to node n). Otherwise continue to step 2 below.
528 * 2. If the DramHoleValid bit of the DHAR for node n is clear, the DHAR is
529 * disabled so skip to step 3 below. Otherwise see if the SysAddr is within
530 * the range of relocated addresses (starting at 0x100000000) from the DRAM
531 * hole. If not, skip to step 3 below. Else get the value of the
532 * DramHoleOffset field from the DHAR. To obtain the DramAddr, subtract the
533 * offset defined by this value from the SysAddr.
535 * 3. Obtain the base address for node n from the DRAMBase field of the DRAM
536 * Base register for node n. To obtain the DramAddr, subtract the base
537 * address from the SysAddr, as shown near the start of section 3.4.4 (p.70).
539 static u64 sys_addr_to_dram_addr(struct mem_ctl_info *mci, u64 sys_addr)
541 struct amd64_pvt *pvt = mci->pvt_info;
542 u64 dram_base, hole_base, hole_offset, hole_size, dram_addr;
545 dram_base = get_dram_base(pvt, pvt->mc_node_id);
547 ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
550 if ((sys_addr >= (1ull << 32)) &&
551 (sys_addr < ((1ull << 32) + hole_size))) {
552 /* use DHAR to translate SysAddr to DramAddr */
553 dram_addr = sys_addr - hole_offset;
555 debugf2("using DHAR to translate SysAddr 0x%lx to "
557 (unsigned long)sys_addr,
558 (unsigned long)dram_addr);
565 * Translate the SysAddr to a DramAddr as shown near the start of
566 * section 3.4.4 (p. 70). Although sys_addr is a 64-bit value, the k8
567 * only deals with 40-bit values. Therefore we discard bits 63-40 of
568 * sys_addr below. If bit 39 of sys_addr is 1 then the bits we
569 * discard are all 1s. Otherwise the bits we discard are all 0s. See
570 * section 3.4.2 of AMD publication 24592: AMD x86-64 Architecture
571 * Programmer's Manual Volume 1 Application Programming.
573 dram_addr = (sys_addr & GENMASK(0, 39)) - dram_base;
575 debugf2("using DRAM Base register to translate SysAddr 0x%lx to "
576 "DramAddr 0x%lx\n", (unsigned long)sys_addr,
577 (unsigned long)dram_addr);
582 * @intlv_en is the value of the IntlvEn field from a DRAM Base register
583 * (section 3.4.4.1). Return the number of bits from a SysAddr that are used
584 * for node interleaving.
586 static int num_node_interleave_bits(unsigned intlv_en)
588 static const int intlv_shift_table[] = { 0, 1, 0, 2, 0, 0, 0, 3 };
591 BUG_ON(intlv_en > 7);
592 n = intlv_shift_table[intlv_en];
596 /* Translate the DramAddr given by @dram_addr to an InputAddr. */
597 static u64 dram_addr_to_input_addr(struct mem_ctl_info *mci, u64 dram_addr)
599 struct amd64_pvt *pvt;
606 * See the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
607 * concerning translating a DramAddr to an InputAddr.
609 intlv_shift = num_node_interleave_bits(dram_intlv_en(pvt, 0));
610 input_addr = ((dram_addr >> intlv_shift) & GENMASK(12, 35)) +
613 debugf2(" Intlv Shift=%d DramAddr=0x%lx maps to InputAddr=0x%lx\n",
614 intlv_shift, (unsigned long)dram_addr,
615 (unsigned long)input_addr);
621 * Translate the SysAddr represented by @sys_addr to an InputAddr. It is
622 * assumed that @sys_addr maps to the node given by mci.
624 static u64 sys_addr_to_input_addr(struct mem_ctl_info *mci, u64 sys_addr)
629 dram_addr_to_input_addr(mci, sys_addr_to_dram_addr(mci, sys_addr));
631 debugf2("SysAdddr 0x%lx translates to InputAddr 0x%lx\n",
632 (unsigned long)sys_addr, (unsigned long)input_addr);
639 * @input_addr is an InputAddr associated with the node represented by mci.
640 * Translate @input_addr to a DramAddr and return the result.
642 static u64 input_addr_to_dram_addr(struct mem_ctl_info *mci, u64 input_addr)
644 struct amd64_pvt *pvt;
645 int node_id, intlv_shift;
650 * Near the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
651 * shows how to translate a DramAddr to an InputAddr. Here we reverse
652 * this procedure. When translating from a DramAddr to an InputAddr, the
653 * bits used for node interleaving are discarded. Here we recover these
654 * bits from the IntlvSel field of the DRAM Limit register (section
655 * 3.4.4.2) for the node that input_addr is associated with.
658 node_id = pvt->mc_node_id;
659 BUG_ON((node_id < 0) || (node_id > 7));
661 intlv_shift = num_node_interleave_bits(dram_intlv_en(pvt, 0));
663 if (intlv_shift == 0) {
664 debugf1(" InputAddr 0x%lx translates to DramAddr of "
665 "same value\n", (unsigned long)input_addr);
670 bits = ((input_addr & GENMASK(12, 35)) << intlv_shift) +
671 (input_addr & 0xfff);
673 intlv_sel = dram_intlv_sel(pvt, node_id) & ((1 << intlv_shift) - 1);
674 dram_addr = bits + (intlv_sel << 12);
676 debugf1("InputAddr 0x%lx translates to DramAddr 0x%lx "
677 "(%d node interleave bits)\n", (unsigned long)input_addr,
678 (unsigned long)dram_addr, intlv_shift);
684 * @dram_addr is a DramAddr that maps to the node represented by mci. Convert
685 * @dram_addr to a SysAddr.
687 static u64 dram_addr_to_sys_addr(struct mem_ctl_info *mci, u64 dram_addr)
689 struct amd64_pvt *pvt = mci->pvt_info;
690 u64 hole_base, hole_offset, hole_size, base, sys_addr;
693 ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
696 if ((dram_addr >= hole_base) &&
697 (dram_addr < (hole_base + hole_size))) {
698 sys_addr = dram_addr + hole_offset;
700 debugf1("using DHAR to translate DramAddr 0x%lx to "
701 "SysAddr 0x%lx\n", (unsigned long)dram_addr,
702 (unsigned long)sys_addr);
708 base = get_dram_base(pvt, pvt->mc_node_id);
709 sys_addr = dram_addr + base;
712 * The sys_addr we have computed up to this point is a 40-bit value
713 * because the k8 deals with 40-bit values. However, the value we are
714 * supposed to return is a full 64-bit physical address. The AMD
715 * x86-64 architecture specifies that the most significant implemented
716 * address bit through bit 63 of a physical address must be either all
717 * 0s or all 1s. Therefore we sign-extend the 40-bit sys_addr to a
718 * 64-bit value below. See section 3.4.2 of AMD publication 24592:
719 * AMD x86-64 Architecture Programmer's Manual Volume 1 Application
722 sys_addr |= ~((sys_addr & (1ull << 39)) - 1);
724 debugf1(" Node %d, DramAddr 0x%lx to SysAddr 0x%lx\n",
725 pvt->mc_node_id, (unsigned long)dram_addr,
726 (unsigned long)sys_addr);
732 * @input_addr is an InputAddr associated with the node given by mci. Translate
733 * @input_addr to a SysAddr.
735 static inline u64 input_addr_to_sys_addr(struct mem_ctl_info *mci,
738 return dram_addr_to_sys_addr(mci,
739 input_addr_to_dram_addr(mci, input_addr));
743 * Find the minimum and maximum InputAddr values that map to the given @csrow.
744 * Pass back these values in *input_addr_min and *input_addr_max.
746 static void find_csrow_limits(struct mem_ctl_info *mci, int csrow,
747 u64 *input_addr_min, u64 *input_addr_max)
749 struct amd64_pvt *pvt;
753 BUG_ON((csrow < 0) || (csrow >= pvt->csels[0].b_cnt));
755 get_cs_base_and_mask(pvt, csrow, 0, &base, &mask);
757 *input_addr_min = base & ~mask;
758 *input_addr_max = base | mask;
761 /* Map the Error address to a PAGE and PAGE OFFSET. */
762 static inline void error_address_to_page_and_offset(u64 error_address,
763 u32 *page, u32 *offset)
765 *page = (u32) (error_address >> PAGE_SHIFT);
766 *offset = ((u32) error_address) & ~PAGE_MASK;
770 * @sys_addr is an error address (a SysAddr) extracted from the MCA NB Address
771 * Low (section 3.6.4.5) and MCA NB Address High (section 3.6.4.6) registers
772 * of a node that detected an ECC memory error. mci represents the node that
773 * the error address maps to (possibly different from the node that detected
774 * the error). Return the number of the csrow that sys_addr maps to, or -1 on
777 static int sys_addr_to_csrow(struct mem_ctl_info *mci, u64 sys_addr)
781 csrow = input_addr_to_csrow(mci, sys_addr_to_input_addr(mci, sys_addr));
784 amd64_mc_err(mci, "Failed to translate InputAddr to csrow for "
785 "address 0x%lx\n", (unsigned long)sys_addr);
789 static int get_channel_from_ecc_syndrome(struct mem_ctl_info *, u16);
791 static u16 extract_syndrome(struct err_regs *err)
793 return ((err->nbsh >> 15) & 0xff) | ((err->nbsl >> 16) & 0xff00);
797 * Determine if the DIMMs have ECC enabled. ECC is enabled ONLY if all the DIMMs
800 static enum edac_type amd64_determine_edac_cap(struct amd64_pvt *pvt)
803 enum dev_type edac_cap = EDAC_FLAG_NONE;
805 bit = (boot_cpu_data.x86 > 0xf || pvt->ext_model >= K8_REV_F)
809 if (pvt->dclr0 & BIT(bit))
810 edac_cap = EDAC_FLAG_SECDED;
816 static void amd64_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt);
818 static void amd64_dump_dramcfg_low(u32 dclr, int chan)
820 debugf1("F2x%d90 (DRAM Cfg Low): 0x%08x\n", chan, dclr);
822 debugf1(" DIMM type: %sbuffered; all DIMMs support ECC: %s\n",
823 (dclr & BIT(16)) ? "un" : "",
824 (dclr & BIT(19)) ? "yes" : "no");
826 debugf1(" PAR/ERR parity: %s\n",
827 (dclr & BIT(8)) ? "enabled" : "disabled");
829 debugf1(" DCT 128bit mode width: %s\n",
830 (dclr & BIT(11)) ? "128b" : "64b");
832 debugf1(" x4 logical DIMMs present: L0: %s L1: %s L2: %s L3: %s\n",
833 (dclr & BIT(12)) ? "yes" : "no",
834 (dclr & BIT(13)) ? "yes" : "no",
835 (dclr & BIT(14)) ? "yes" : "no",
836 (dclr & BIT(15)) ? "yes" : "no");
839 /* Display and decode various NB registers for debug purposes. */
840 static void dump_misc_regs(struct amd64_pvt *pvt)
842 debugf1("F3xE8 (NB Cap): 0x%08x\n", pvt->nbcap);
844 debugf1(" NB two channel DRAM capable: %s\n",
845 (pvt->nbcap & K8_NBCAP_DCT_DUAL) ? "yes" : "no");
847 debugf1(" ECC capable: %s, ChipKill ECC capable: %s\n",
848 (pvt->nbcap & K8_NBCAP_SECDED) ? "yes" : "no",
849 (pvt->nbcap & K8_NBCAP_CHIPKILL) ? "yes" : "no");
851 amd64_dump_dramcfg_low(pvt->dclr0, 0);
853 debugf1("F3xB0 (Online Spare): 0x%08x\n", pvt->online_spare);
855 debugf1("F1xF0 (DRAM Hole Address): 0x%08x, base: 0x%08x, "
857 pvt->dhar, dhar_base(pvt),
858 (boot_cpu_data.x86 == 0xf) ? k8_dhar_offset(pvt)
859 : f10_dhar_offset(pvt));
861 debugf1(" DramHoleValid: %s\n", dhar_valid(pvt) ? "yes" : "no");
863 amd64_debug_display_dimm_sizes(0, pvt);
865 /* everything below this point is Fam10h and above */
866 if (boot_cpu_data.x86 == 0xf)
869 amd64_debug_display_dimm_sizes(1, pvt);
871 amd64_info("using %s syndromes.\n", ((pvt->syn_type == 8) ? "x8" : "x4"));
873 /* Only if NOT ganged does dclr1 have valid info */
874 if (!dct_ganging_enabled(pvt))
875 amd64_dump_dramcfg_low(pvt->dclr1, 1);
879 * see BKDG, F2x[1,0][5C:40], F2[1,0][6C:60]
881 static void prep_chip_selects(struct amd64_pvt *pvt)
883 if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_F) {
884 pvt->csels[0].b_cnt = pvt->csels[1].b_cnt = 8;
885 pvt->csels[0].m_cnt = pvt->csels[1].m_cnt = 8;
887 pvt->csels[0].b_cnt = pvt->csels[1].b_cnt = 8;
888 pvt->csels[0].m_cnt = pvt->csels[1].m_cnt = 4;
893 * Function 2 Offset F10_DCSB0; read in the DCS Base and DCS Mask registers
895 static void read_dct_base_mask(struct amd64_pvt *pvt)
899 prep_chip_selects(pvt);
901 for_each_chip_select(cs, 0, pvt) {
902 u32 reg0 = DCSB0 + (cs * 4);
903 u32 reg1 = DCSB1 + (cs * 4);
904 u32 *base0 = &pvt->csels[0].csbases[cs];
905 u32 *base1 = &pvt->csels[1].csbases[cs];
907 if (!amd64_read_dct_pci_cfg(pvt, reg0, base0))
908 debugf0(" DCSB0[%d]=0x%08x reg: F2x%x\n",
911 if (boot_cpu_data.x86 == 0xf || dct_ganging_enabled(pvt))
914 if (!amd64_read_dct_pci_cfg(pvt, reg1, base1))
915 debugf0(" DCSB1[%d]=0x%08x reg: F2x%x\n",
919 for_each_chip_select_mask(cs, 0, pvt) {
920 u32 reg0 = DCSM0 + (cs * 4);
921 u32 reg1 = DCSM1 + (cs * 4);
922 u32 *mask0 = &pvt->csels[0].csmasks[cs];
923 u32 *mask1 = &pvt->csels[1].csmasks[cs];
925 if (!amd64_read_dct_pci_cfg(pvt, reg0, mask0))
926 debugf0(" DCSM0[%d]=0x%08x reg: F2x%x\n",
929 if (boot_cpu_data.x86 == 0xf || dct_ganging_enabled(pvt))
932 if (!amd64_read_dct_pci_cfg(pvt, reg1, mask1))
933 debugf0(" DCSM1[%d]=0x%08x reg: F2x%x\n",
938 static enum mem_type amd64_determine_memory_type(struct amd64_pvt *pvt, int cs)
942 if (boot_cpu_data.x86 >= 0x10 || pvt->ext_model >= K8_REV_F) {
943 if (pvt->dchr0 & DDR3_MODE)
944 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR3 : MEM_RDDR3;
946 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR2 : MEM_RDDR2;
948 type = (pvt->dclr0 & BIT(18)) ? MEM_DDR : MEM_RDDR;
951 amd64_info("CS%d: %s\n", cs, edac_mem_types[type]);
957 * Read the DRAM Configuration Low register. It differs between CG, D & E revs
958 * and the later RevF memory controllers (DDR vs DDR2)
961 * number of memory channels in operation
963 * contents of the DCL0_LOW register
965 static int k8_early_channel_count(struct amd64_pvt *pvt)
969 err = amd64_read_dct_pci_cfg(pvt, F10_DCLR_0, &pvt->dclr0);
973 if (pvt->ext_model >= K8_REV_F)
974 /* RevF (NPT) and later */
975 flag = pvt->dclr0 & F10_WIDTH_128;
977 /* RevE and earlier */
978 flag = pvt->dclr0 & REVE_WIDTH_128;
983 return (flag) ? 2 : 1;
986 /* extract the ERROR ADDRESS for the K8 CPUs */
987 static u64 k8_get_error_address(struct mem_ctl_info *mci,
988 struct err_regs *info)
990 return (((u64) (info->nbeah & 0xff)) << 32) +
991 (info->nbeal & ~0x03);
994 static void read_dram_base_limit_regs(struct amd64_pvt *pvt, unsigned range)
996 u32 off = range << 3;
998 amd64_read_pci_cfg(pvt->F1, DRAM_BASE_LO + off, &pvt->ranges[range].base.lo);
999 amd64_read_pci_cfg(pvt->F1, DRAM_LIMIT_LO + off, &pvt->ranges[range].lim.lo);
1001 if (boot_cpu_data.x86 == 0xf)
1004 if (!dram_rw(pvt, range))
1007 amd64_read_pci_cfg(pvt->F1, DRAM_BASE_HI + off, &pvt->ranges[range].base.hi);
1008 amd64_read_pci_cfg(pvt->F1, DRAM_LIMIT_HI + off, &pvt->ranges[range].lim.hi);
1011 static void k8_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
1012 struct err_regs *err_info, u64 sys_addr)
1014 struct mem_ctl_info *src_mci;
1019 syndrome = extract_syndrome(err_info);
1021 /* CHIPKILL enabled */
1022 if (err_info->nbcfg & K8_NBCFG_CHIPKILL) {
1023 channel = get_channel_from_ecc_syndrome(mci, syndrome);
1026 * Syndrome didn't map, so we don't know which of the
1027 * 2 DIMMs is in error. So we need to ID 'both' of them
1030 amd64_mc_warn(mci, "unknown syndrome 0x%04x - possible "
1031 "error reporting race\n", syndrome);
1032 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1037 * non-chipkill ecc mode
1039 * The k8 documentation is unclear about how to determine the
1040 * channel number when using non-chipkill memory. This method
1041 * was obtained from email communication with someone at AMD.
1042 * (Wish the email was placed in this comment - norsk)
1044 channel = ((sys_addr & BIT(3)) != 0);
1048 * Find out which node the error address belongs to. This may be
1049 * different from the node that detected the error.
1051 src_mci = find_mc_by_sys_addr(mci, sys_addr);
1053 amd64_mc_err(mci, "failed to map error addr 0x%lx to a node\n",
1054 (unsigned long)sys_addr);
1055 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1059 /* Now map the sys_addr to a CSROW */
1060 csrow = sys_addr_to_csrow(src_mci, sys_addr);
1062 edac_mc_handle_ce_no_info(src_mci, EDAC_MOD_STR);
1064 error_address_to_page_and_offset(sys_addr, &page, &offset);
1066 edac_mc_handle_ce(src_mci, page, offset, syndrome, csrow,
1067 channel, EDAC_MOD_STR);
1071 static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, int cs_mode)
1075 if (pvt->ext_model >= K8_REV_F)
1076 dbam_map = ddr2_dbam;
1077 else if (pvt->ext_model >= K8_REV_D)
1078 dbam_map = ddr2_dbam_revD;
1080 dbam_map = ddr2_dbam_revCG;
1082 return dbam_map[cs_mode];
1086 * Get the number of DCT channels in use.
1089 * number of Memory Channels in operation
1091 * contents of the DCL0_LOW register
1093 static int f10_early_channel_count(struct amd64_pvt *pvt)
1095 int i, j, channels = 0;
1097 /* If we are in 128 bit mode, then we are using 2 channels */
1098 if (pvt->dclr0 & F10_WIDTH_128) {
1104 * Need to check if in unganged mode: In such, there are 2 channels,
1105 * but they are not in 128 bit mode and thus the above 'dclr0' status
1108 * Need to check DCT0[0] and DCT1[0] to see if only one of them has
1109 * their CSEnable bit on. If so, then SINGLE DIMM case.
1111 debugf0("Data width is not 128 bits - need more decoding\n");
1114 * Check DRAM Bank Address Mapping values for each DIMM to see if there
1115 * is more than just one DIMM present in unganged mode. Need to check
1116 * both controllers since DIMMs can be placed in either one.
1118 for (i = 0; i < 2; i++) {
1119 u32 dbam = (i ? pvt->dbam1 : pvt->dbam0);
1121 for (j = 0; j < 4; j++) {
1122 if (DBAM_DIMM(j, dbam) > 0) {
1132 amd64_info("MCT channel count: %d\n", channels);
1137 static int f10_dbam_to_chip_select(struct amd64_pvt *pvt, int cs_mode)
1141 if (pvt->dchr0 & DDR3_MODE || pvt->dchr1 & DDR3_MODE)
1142 dbam_map = ddr3_dbam;
1144 dbam_map = ddr2_dbam;
1146 return dbam_map[cs_mode];
1149 static u64 f10_get_error_address(struct mem_ctl_info *mci,
1150 struct err_regs *info)
1152 return (((u64) (info->nbeah & 0xffff)) << 32) +
1153 (info->nbeal & ~0x01);
1156 static void f10_read_dram_ctl_register(struct amd64_pvt *pvt)
1159 if (!amd64_read_dct_pci_cfg(pvt, F10_DCTL_SEL_LOW, &pvt->dct_sel_low)) {
1160 debugf0("F2x110 (DCTL Sel. Low): 0x%08x, High range addrs at: 0x%x\n",
1161 pvt->dct_sel_low, dct_sel_baseaddr(pvt));
1163 debugf0(" DCT mode: %s, All DCTs on: %s\n",
1164 (dct_ganging_enabled(pvt) ? "ganged" : "unganged"),
1165 (dct_dram_enabled(pvt) ? "yes" : "no"));
1167 if (!dct_ganging_enabled(pvt))
1168 debugf0(" Address range split per DCT: %s\n",
1169 (dct_high_range_enabled(pvt) ? "yes" : "no"));
1171 debugf0(" DCT data interleave for ECC: %s, "
1172 "DRAM cleared since last warm reset: %s\n",
1173 (dct_data_intlv_enabled(pvt) ? "enabled" : "disabled"),
1174 (dct_memory_cleared(pvt) ? "yes" : "no"));
1176 debugf0(" DCT channel interleave: %s, "
1177 "DCT interleave bits selector: 0x%x\n",
1178 (dct_interleave_enabled(pvt) ? "enabled" : "disabled"),
1179 dct_sel_interleave_addr(pvt));
1182 amd64_read_dct_pci_cfg(pvt, F10_DCTL_SEL_HIGH, &pvt->dct_sel_hi);
1186 * Determine channel (DCT) based on the interleaving mode: F10h BKDG, 2.8.9 Memory
1187 * Interleaving Modes.
1189 static u8 f10_determine_channel(struct amd64_pvt *pvt, u64 sys_addr,
1190 bool hi_range_sel, u8 intlv_en)
1192 u32 dct_sel_high = (pvt->dct_sel_low >> 1) & 1;
1194 if (dct_ganging_enabled(pvt))
1198 return dct_sel_high;
1201 * see F2x110[DctSelIntLvAddr] - channel interleave mode
1203 if (dct_interleave_enabled(pvt)) {
1204 u8 intlv_addr = dct_sel_interleave_addr(pvt);
1206 /* return DCT select function: 0=DCT0, 1=DCT1 */
1208 return sys_addr >> 6 & 1;
1210 if (intlv_addr & 0x2) {
1211 u8 shift = intlv_addr & 0x1 ? 9 : 6;
1212 u32 temp = hweight_long((u32) ((sys_addr >> 16) & 0x1F)) % 2;
1214 return ((sys_addr >> shift) & 1) ^ temp;
1217 return (sys_addr >> (12 + hweight8(intlv_en))) & 1;
1220 if (dct_high_range_enabled(pvt))
1221 return ~dct_sel_high & 1;
1226 /* Convert the sys_addr to the normalized DCT address */
1227 static u64 f10_get_norm_dct_addr(struct amd64_pvt *pvt, int range,
1228 u64 sys_addr, bool hi_rng,
1229 u32 dct_sel_base_addr)
1232 u64 dram_base = get_dram_base(pvt, range);
1233 u64 hole_off = f10_dhar_offset(pvt);
1234 u32 hole_valid = dhar_valid(pvt);
1235 u64 dct_sel_base_off = (pvt->dct_sel_hi & 0xFFFFFC00) << 16;
1240 * base address of high range is below 4Gb
1241 * (bits [47:27] at [31:11])
1242 * DRAM address space on this DCT is hoisted above 4Gb &&
1245 * remove hole offset from sys_addr
1247 * remove high range offset from sys_addr
1249 if ((!(dct_sel_base_addr >> 16) ||
1250 dct_sel_base_addr < dhar_base(pvt)) &&
1252 (sys_addr >= BIT_64(32)))
1253 chan_off = hole_off;
1255 chan_off = dct_sel_base_off;
1259 * we have a valid hole &&
1264 * remove dram base to normalize to DCT address
1266 if (hole_valid && (sys_addr >= BIT_64(32)))
1267 chan_off = hole_off;
1269 chan_off = dram_base;
1272 return (sys_addr & GENMASK(6,47)) - (chan_off & GENMASK(23,47));
1275 /* Hack for the time being - Can we get this from BIOS?? */
1276 #define CH0SPARE_RANK 0
1277 #define CH1SPARE_RANK 1
1280 * checks if the csrow passed in is marked as SPARED, if so returns the new
1283 static int f10_process_possible_spare(struct amd64_pvt *pvt, u8 dct, int csrow)
1288 /* Depending on channel, isolate respective SPARING info */
1290 swap_done = F10_ONLINE_SPARE_SWAPDONE1(pvt->online_spare);
1291 bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS1(pvt->online_spare);
1292 if (swap_done && (csrow == bad_dram_cs))
1293 csrow = CH1SPARE_RANK;
1295 swap_done = F10_ONLINE_SPARE_SWAPDONE0(pvt->online_spare);
1296 bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS0(pvt->online_spare);
1297 if (swap_done && (csrow == bad_dram_cs))
1298 csrow = CH0SPARE_RANK;
1304 * Iterate over the DRAM DCT "base" and "mask" registers looking for a
1305 * SystemAddr match on the specified 'ChannelSelect' and 'NodeID'
1308 * -EINVAL: NOT FOUND
1309 * 0..csrow = Chip-Select Row
1311 static int f10_lookup_addr_in_dct(u64 in_addr, u32 nid, u8 dct)
1313 struct mem_ctl_info *mci;
1314 struct amd64_pvt *pvt;
1315 u64 cs_base, cs_mask;
1316 int cs_found = -EINVAL;
1323 pvt = mci->pvt_info;
1325 debugf1("input addr: 0x%llx, DCT: %d\n", in_addr, dct);
1327 for_each_chip_select(csrow, dct, pvt) {
1328 if (!csrow_enabled(csrow, dct, pvt))
1331 get_cs_base_and_mask(pvt, csrow, dct, &cs_base, &cs_mask);
1333 debugf1(" CSROW=%d CSBase=0x%llx CSMask=0x%llx\n",
1334 csrow, cs_base, cs_mask);
1338 debugf1(" (InputAddr & ~CSMask)=0x%llx "
1339 "(CSBase & ~CSMask)=0x%llx\n",
1340 (in_addr & cs_mask), (cs_base & cs_mask));
1342 if ((in_addr & cs_mask) == (cs_base & cs_mask)) {
1343 cs_found = f10_process_possible_spare(pvt, dct, csrow);
1345 debugf1(" MATCH csrow=%d\n", cs_found);
1352 /* For a given @dram_range, check if @sys_addr falls within it. */
1353 static int f10_match_to_this_node(struct amd64_pvt *pvt, int range,
1354 u64 sys_addr, int *nid, int *chan_sel)
1356 int cs_found = -EINVAL;
1358 u32 tmp, dct_sel_base;
1360 bool high_range = false;
1362 u8 node_id = dram_dst_node(pvt, range);
1363 u8 intlv_en = dram_intlv_en(pvt, range);
1364 u32 intlv_sel = dram_intlv_sel(pvt, range);
1366 debugf1("(range %d) SystemAddr= 0x%llx Limit=0x%llx\n",
1367 range, sys_addr, get_dram_limit(pvt, range));
1370 (intlv_sel != ((sys_addr >> 12) & intlv_en)))
1373 dct_sel_base = dct_sel_baseaddr(pvt);
1376 * check whether addresses >= DctSelBaseAddr[47:27] are to be used to
1377 * select between DCT0 and DCT1.
1379 if (dct_high_range_enabled(pvt) &&
1380 !dct_ganging_enabled(pvt) &&
1381 ((sys_addr >> 27) >= (dct_sel_base >> 11)))
1384 channel = f10_determine_channel(pvt, sys_addr, high_range, intlv_en);
1386 chan_addr = f10_get_norm_dct_addr(pvt, range, sys_addr,
1387 high_range, dct_sel_base);
1389 /* remove Node ID (in case of node interleaving) */
1390 tmp = chan_addr & 0xFC0;
1392 chan_addr = ((chan_addr >> hweight8(intlv_en)) & GENMASK(12, 47)) | tmp;
1394 /* remove channel interleave and hash */
1395 if (dct_interleave_enabled(pvt) &&
1396 !dct_high_range_enabled(pvt) &&
1397 !dct_ganging_enabled(pvt)) {
1398 if (dct_sel_interleave_addr(pvt) != 1)
1399 chan_addr = (chan_addr >> 1) & GENMASK(6, 63);
1401 tmp = chan_addr & 0xFC0;
1402 chan_addr = ((chan_addr & GENMASK(14, 63)) >> 1) | tmp;
1406 debugf1(" (ChannelAddrLong=0x%llx)\n", chan_addr);
1408 cs_found = f10_lookup_addr_in_dct(chan_addr, node_id, channel);
1410 if (cs_found >= 0) {
1412 *chan_sel = channel;
1417 static int f10_translate_sysaddr_to_cs(struct amd64_pvt *pvt, u64 sys_addr,
1418 int *node, int *chan_sel)
1420 int range, cs_found = -EINVAL;
1422 for (range = 0; range < DRAM_RANGES; range++) {
1424 if (!dram_rw(pvt, range))
1427 if ((get_dram_base(pvt, range) <= sys_addr) &&
1428 (get_dram_limit(pvt, range) >= sys_addr)) {
1430 cs_found = f10_match_to_this_node(pvt, range,
1441 * For reference see "2.8.5 Routing DRAM Requests" in F10 BKDG. This code maps
1442 * a @sys_addr to NodeID, DCT (channel) and chip select (CSROW).
1444 * The @sys_addr is usually an error address received from the hardware
1447 static void f10_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
1448 struct err_regs *err_info,
1451 struct amd64_pvt *pvt = mci->pvt_info;
1453 int nid, csrow, chan = 0;
1456 csrow = f10_translate_sysaddr_to_cs(pvt, sys_addr, &nid, &chan);
1459 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1463 error_address_to_page_and_offset(sys_addr, &page, &offset);
1465 syndrome = extract_syndrome(err_info);
1468 * We need the syndromes for channel detection only when we're
1469 * ganged. Otherwise @chan should already contain the channel at
1472 if (dct_ganging_enabled(pvt) && (pvt->nbcfg & K8_NBCFG_CHIPKILL))
1473 chan = get_channel_from_ecc_syndrome(mci, syndrome);
1476 edac_mc_handle_ce(mci, page, offset, syndrome, csrow, chan,
1480 * Channel unknown, report all channels on this CSROW as failed.
1482 for (chan = 0; chan < mci->csrows[csrow].nr_channels; chan++)
1483 edac_mc_handle_ce(mci, page, offset, syndrome,
1484 csrow, chan, EDAC_MOD_STR);
1488 * debug routine to display the memory sizes of all logical DIMMs and its
1491 static void amd64_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt)
1493 int dimm, size0, size1, factor = 0;
1494 u32 *dcsb = ctrl ? pvt->csels[1].csbases : pvt->csels[0].csbases;
1495 u32 dbam = ctrl ? pvt->dbam1 : pvt->dbam0;
1497 if (boot_cpu_data.x86 == 0xf) {
1498 if (pvt->dclr0 & F10_WIDTH_128)
1501 /* K8 families < revF not supported yet */
1502 if (pvt->ext_model < K8_REV_F)
1508 dbam = (ctrl && !dct_ganging_enabled(pvt)) ? pvt->dbam1 : pvt->dbam0;
1509 dcsb = (ctrl && !dct_ganging_enabled(pvt)) ? pvt->csels[1].csbases
1510 : pvt->csels[0].csbases;
1512 debugf1("F2x%d80 (DRAM Bank Address Mapping): 0x%08x\n", ctrl, dbam);
1514 edac_printk(KERN_DEBUG, EDAC_MC, "DCT%d chip selects:\n", ctrl);
1516 /* Dump memory sizes for DIMM and its CSROWs */
1517 for (dimm = 0; dimm < 4; dimm++) {
1520 if (dcsb[dimm*2] & DCSB_CS_ENABLE)
1521 size0 = pvt->ops->dbam_to_cs(pvt, DBAM_DIMM(dimm, dbam));
1524 if (dcsb[dimm*2 + 1] & DCSB_CS_ENABLE)
1525 size1 = pvt->ops->dbam_to_cs(pvt, DBAM_DIMM(dimm, dbam));
1527 amd64_info(EDAC_MC ": %d: %5dMB %d: %5dMB\n",
1528 dimm * 2, size0 << factor,
1529 dimm * 2 + 1, size1 << factor);
1533 static struct amd64_family_type amd64_family_types[] = {
1536 .f1_id = PCI_DEVICE_ID_AMD_K8_NB_ADDRMAP,
1537 .f3_id = PCI_DEVICE_ID_AMD_K8_NB_MISC,
1539 .early_channel_count = k8_early_channel_count,
1540 .get_error_address = k8_get_error_address,
1541 .map_sysaddr_to_csrow = k8_map_sysaddr_to_csrow,
1542 .dbam_to_cs = k8_dbam_to_chip_select,
1543 .read_dct_pci_cfg = k8_read_dct_pci_cfg,
1548 .f1_id = PCI_DEVICE_ID_AMD_10H_NB_MAP,
1549 .f3_id = PCI_DEVICE_ID_AMD_10H_NB_MISC,
1551 .early_channel_count = f10_early_channel_count,
1552 .get_error_address = f10_get_error_address,
1553 .read_dram_ctl_register = f10_read_dram_ctl_register,
1554 .map_sysaddr_to_csrow = f10_map_sysaddr_to_csrow,
1555 .dbam_to_cs = f10_dbam_to_chip_select,
1556 .read_dct_pci_cfg = f10_read_dct_pci_cfg,
1562 .read_dct_pci_cfg = f15_read_dct_pci_cfg,
1567 static struct pci_dev *pci_get_related_function(unsigned int vendor,
1568 unsigned int device,
1569 struct pci_dev *related)
1571 struct pci_dev *dev = NULL;
1573 dev = pci_get_device(vendor, device, dev);
1575 if ((dev->bus->number == related->bus->number) &&
1576 (PCI_SLOT(dev->devfn) == PCI_SLOT(related->devfn)))
1578 dev = pci_get_device(vendor, device, dev);
1585 * These are tables of eigenvectors (one per line) which can be used for the
1586 * construction of the syndrome tables. The modified syndrome search algorithm
1587 * uses those to find the symbol in error and thus the DIMM.
1589 * Algorithm courtesy of Ross LaFetra from AMD.
1591 static u16 x4_vectors[] = {
1592 0x2f57, 0x1afe, 0x66cc, 0xdd88,
1593 0x11eb, 0x3396, 0x7f4c, 0xeac8,
1594 0x0001, 0x0002, 0x0004, 0x0008,
1595 0x1013, 0x3032, 0x4044, 0x8088,
1596 0x106b, 0x30d6, 0x70fc, 0xe0a8,
1597 0x4857, 0xc4fe, 0x13cc, 0x3288,
1598 0x1ac5, 0x2f4a, 0x5394, 0xa1e8,
1599 0x1f39, 0x251e, 0xbd6c, 0x6bd8,
1600 0x15c1, 0x2a42, 0x89ac, 0x4758,
1601 0x2b03, 0x1602, 0x4f0c, 0xca08,
1602 0x1f07, 0x3a0e, 0x6b04, 0xbd08,
1603 0x8ba7, 0x465e, 0x244c, 0x1cc8,
1604 0x2b87, 0x164e, 0x642c, 0xdc18,
1605 0x40b9, 0x80de, 0x1094, 0x20e8,
1606 0x27db, 0x1eb6, 0x9dac, 0x7b58,
1607 0x11c1, 0x2242, 0x84ac, 0x4c58,
1608 0x1be5, 0x2d7a, 0x5e34, 0xa718,
1609 0x4b39, 0x8d1e, 0x14b4, 0x28d8,
1610 0x4c97, 0xc87e, 0x11fc, 0x33a8,
1611 0x8e97, 0x497e, 0x2ffc, 0x1aa8,
1612 0x16b3, 0x3d62, 0x4f34, 0x8518,
1613 0x1e2f, 0x391a, 0x5cac, 0xf858,
1614 0x1d9f, 0x3b7a, 0x572c, 0xfe18,
1615 0x15f5, 0x2a5a, 0x5264, 0xa3b8,
1616 0x1dbb, 0x3b66, 0x715c, 0xe3f8,
1617 0x4397, 0xc27e, 0x17fc, 0x3ea8,
1618 0x1617, 0x3d3e, 0x6464, 0xb8b8,
1619 0x23ff, 0x12aa, 0xab6c, 0x56d8,
1620 0x2dfb, 0x1ba6, 0x913c, 0x7328,
1621 0x185d, 0x2ca6, 0x7914, 0x9e28,
1622 0x171b, 0x3e36, 0x7d7c, 0xebe8,
1623 0x4199, 0x82ee, 0x19f4, 0x2e58,
1624 0x4807, 0xc40e, 0x130c, 0x3208,
1625 0x1905, 0x2e0a, 0x5804, 0xac08,
1626 0x213f, 0x132a, 0xadfc, 0x5ba8,
1627 0x19a9, 0x2efe, 0xb5cc, 0x6f88,
1630 static u16 x8_vectors[] = {
1631 0x0145, 0x028a, 0x2374, 0x43c8, 0xa1f0, 0x0520, 0x0a40, 0x1480,
1632 0x0211, 0x0422, 0x0844, 0x1088, 0x01b0, 0x44e0, 0x23c0, 0xed80,
1633 0x1011, 0x0116, 0x022c, 0x0458, 0x08b0, 0x8c60, 0x2740, 0x4e80,
1634 0x0411, 0x0822, 0x1044, 0x0158, 0x02b0, 0x2360, 0x46c0, 0xab80,
1635 0x0811, 0x1022, 0x012c, 0x0258, 0x04b0, 0x4660, 0x8cc0, 0x2780,
1636 0x2071, 0x40e2, 0xa0c4, 0x0108, 0x0210, 0x0420, 0x0840, 0x1080,
1637 0x4071, 0x80e2, 0x0104, 0x0208, 0x0410, 0x0820, 0x1040, 0x2080,
1638 0x8071, 0x0102, 0x0204, 0x0408, 0x0810, 0x1020, 0x2040, 0x4080,
1639 0x019d, 0x03d6, 0x136c, 0x2198, 0x50b0, 0xb2e0, 0x0740, 0x0e80,
1640 0x0189, 0x03ea, 0x072c, 0x0e58, 0x1cb0, 0x56e0, 0x37c0, 0xf580,
1641 0x01fd, 0x0376, 0x06ec, 0x0bb8, 0x1110, 0x2220, 0x4440, 0x8880,
1642 0x0163, 0x02c6, 0x1104, 0x0758, 0x0eb0, 0x2be0, 0x6140, 0xc280,
1643 0x02fd, 0x01c6, 0x0b5c, 0x1108, 0x07b0, 0x25a0, 0x8840, 0x6180,
1644 0x0801, 0x012e, 0x025c, 0x04b8, 0x1370, 0x26e0, 0x57c0, 0xb580,
1645 0x0401, 0x0802, 0x015c, 0x02b8, 0x22b0, 0x13e0, 0x7140, 0xe280,
1646 0x0201, 0x0402, 0x0804, 0x01b8, 0x11b0, 0x31a0, 0x8040, 0x7180,
1647 0x0101, 0x0202, 0x0404, 0x0808, 0x1010, 0x2020, 0x4040, 0x8080,
1648 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
1649 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x8000,
1652 static int decode_syndrome(u16 syndrome, u16 *vectors, int num_vecs,
1655 unsigned int i, err_sym;
1657 for (err_sym = 0; err_sym < num_vecs / v_dim; err_sym++) {
1659 int v_idx = err_sym * v_dim;
1660 int v_end = (err_sym + 1) * v_dim;
1662 /* walk over all 16 bits of the syndrome */
1663 for (i = 1; i < (1U << 16); i <<= 1) {
1665 /* if bit is set in that eigenvector... */
1666 if (v_idx < v_end && vectors[v_idx] & i) {
1667 u16 ev_comp = vectors[v_idx++];
1669 /* ... and bit set in the modified syndrome, */
1679 /* can't get to zero, move to next symbol */
1684 debugf0("syndrome(%x) not found\n", syndrome);
1688 static int map_err_sym_to_channel(int err_sym, int sym_size)
1701 return err_sym >> 4;
1707 /* imaginary bits not in a DIMM */
1709 WARN(1, KERN_ERR "Invalid error symbol: 0x%x\n",
1721 return err_sym >> 3;
1727 static int get_channel_from_ecc_syndrome(struct mem_ctl_info *mci, u16 syndrome)
1729 struct amd64_pvt *pvt = mci->pvt_info;
1732 if (pvt->syn_type == 8)
1733 err_sym = decode_syndrome(syndrome, x8_vectors,
1734 ARRAY_SIZE(x8_vectors),
1736 else if (pvt->syn_type == 4)
1737 err_sym = decode_syndrome(syndrome, x4_vectors,
1738 ARRAY_SIZE(x4_vectors),
1741 amd64_warn("Illegal syndrome type: %u\n", pvt->syn_type);
1745 return map_err_sym_to_channel(err_sym, pvt->syn_type);
1749 * Handle any Correctable Errors (CEs) that have occurred. Check for valid ERROR
1750 * ADDRESS and process.
1752 static void amd64_handle_ce(struct mem_ctl_info *mci,
1753 struct err_regs *info)
1755 struct amd64_pvt *pvt = mci->pvt_info;
1758 /* Ensure that the Error Address is VALID */
1759 if (!(info->nbsh & K8_NBSH_VALID_ERROR_ADDR)) {
1760 amd64_mc_err(mci, "HW has no ERROR_ADDRESS available\n");
1761 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1765 sys_addr = pvt->ops->get_error_address(mci, info);
1767 amd64_mc_err(mci, "CE ERROR_ADDRESS= 0x%llx\n", sys_addr);
1769 pvt->ops->map_sysaddr_to_csrow(mci, info, sys_addr);
1772 /* Handle any Un-correctable Errors (UEs) */
1773 static void amd64_handle_ue(struct mem_ctl_info *mci,
1774 struct err_regs *info)
1776 struct amd64_pvt *pvt = mci->pvt_info;
1777 struct mem_ctl_info *log_mci, *src_mci = NULL;
1784 if (!(info->nbsh & K8_NBSH_VALID_ERROR_ADDR)) {
1785 amd64_mc_err(mci, "HW has no ERROR_ADDRESS available\n");
1786 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
1790 sys_addr = pvt->ops->get_error_address(mci, info);
1793 * Find out which node the error address belongs to. This may be
1794 * different from the node that detected the error.
1796 src_mci = find_mc_by_sys_addr(mci, sys_addr);
1798 amd64_mc_err(mci, "ERROR ADDRESS (0x%lx) NOT mapped to a MC\n",
1799 (unsigned long)sys_addr);
1800 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
1806 csrow = sys_addr_to_csrow(log_mci, sys_addr);
1808 amd64_mc_err(mci, "ERROR_ADDRESS (0x%lx) NOT mapped to CS\n",
1809 (unsigned long)sys_addr);
1810 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
1812 error_address_to_page_and_offset(sys_addr, &page, &offset);
1813 edac_mc_handle_ue(log_mci, page, offset, csrow, EDAC_MOD_STR);
1817 static inline void __amd64_decode_bus_error(struct mem_ctl_info *mci,
1818 struct err_regs *info)
1820 u16 ec = EC(info->nbsl);
1821 u8 xec = XEC(info->nbsl, 0x1f);
1822 int ecc_type = (info->nbsh >> 13) & 0x3;
1824 /* Bail early out if this was an 'observed' error */
1825 if (PP(ec) == K8_NBSL_PP_OBS)
1828 /* Do only ECC errors */
1829 if (xec && xec != F10_NBSL_EXT_ERR_ECC)
1833 amd64_handle_ce(mci, info);
1834 else if (ecc_type == 1)
1835 amd64_handle_ue(mci, info);
1838 void amd64_decode_bus_error(int node_id, struct mce *m, u32 nbcfg)
1840 struct mem_ctl_info *mci = mcis[node_id];
1841 struct err_regs regs;
1843 regs.nbsl = (u32) m->status;
1844 regs.nbsh = (u32)(m->status >> 32);
1845 regs.nbeal = (u32) m->addr;
1846 regs.nbeah = (u32)(m->addr >> 32);
1849 __amd64_decode_bus_error(mci, ®s);
1852 * Check the UE bit of the NB status high register, if set generate some
1853 * logs. If NOT a GART error, then process the event as a NO-INFO event.
1854 * If it was a GART error, skip that process.
1856 * FIXME: this should go somewhere else, if at all.
1858 if (regs.nbsh & K8_NBSH_UC_ERR && !report_gart_errors)
1859 edac_mc_handle_ue_no_info(mci, "UE bit is set");
1864 * Use pvt->F2 which contains the F2 CPU PCI device to get the related
1865 * F1 (AddrMap) and F3 (Misc) devices. Return negative value on error.
1867 static int reserve_mc_sibling_devs(struct amd64_pvt *pvt, u16 f1_id, u16 f3_id)
1869 /* Reserve the ADDRESS MAP Device */
1870 pvt->F1 = pci_get_related_function(pvt->F2->vendor, f1_id, pvt->F2);
1872 amd64_err("error address map device not found: "
1873 "vendor %x device 0x%x (broken BIOS?)\n",
1874 PCI_VENDOR_ID_AMD, f1_id);
1878 /* Reserve the MISC Device */
1879 pvt->F3 = pci_get_related_function(pvt->F2->vendor, f3_id, pvt->F2);
1881 pci_dev_put(pvt->F1);
1884 amd64_err("error F3 device not found: "
1885 "vendor %x device 0x%x (broken BIOS?)\n",
1886 PCI_VENDOR_ID_AMD, f3_id);
1890 debugf1("F1: %s\n", pci_name(pvt->F1));
1891 debugf1("F2: %s\n", pci_name(pvt->F2));
1892 debugf1("F3: %s\n", pci_name(pvt->F3));
1897 static void free_mc_sibling_devs(struct amd64_pvt *pvt)
1899 pci_dev_put(pvt->F1);
1900 pci_dev_put(pvt->F3);
1904 * Retrieve the hardware registers of the memory controller (this includes the
1905 * 'Address Map' and 'Misc' device regs)
1907 static void read_mc_regs(struct amd64_pvt *pvt)
1914 * Retrieve TOP_MEM and TOP_MEM2; no masking off of reserved bits since
1915 * those are Read-As-Zero
1917 rdmsrl(MSR_K8_TOP_MEM1, pvt->top_mem);
1918 debugf0(" TOP_MEM: 0x%016llx\n", pvt->top_mem);
1920 /* check first whether TOP_MEM2 is enabled */
1921 rdmsrl(MSR_K8_SYSCFG, msr_val);
1922 if (msr_val & (1U << 21)) {
1923 rdmsrl(MSR_K8_TOP_MEM2, pvt->top_mem2);
1924 debugf0(" TOP_MEM2: 0x%016llx\n", pvt->top_mem2);
1926 debugf0(" TOP_MEM2 disabled.\n");
1928 amd64_read_pci_cfg(pvt->F3, K8_NBCAP, &pvt->nbcap);
1930 if (pvt->ops->read_dram_ctl_register)
1931 pvt->ops->read_dram_ctl_register(pvt);
1933 for (range = 0; range < DRAM_RANGES; range++) {
1936 /* read settings for this DRAM range */
1937 read_dram_base_limit_regs(pvt, range);
1939 rw = dram_rw(pvt, range);
1943 debugf1(" DRAM range[%d], base: 0x%016llx; limit: 0x%016llx\n",
1945 get_dram_base(pvt, range),
1946 get_dram_limit(pvt, range));
1948 debugf1(" IntlvEn=%s; Range access: %s%s IntlvSel=%d DstNode=%d\n",
1949 dram_intlv_en(pvt, range) ? "Enabled" : "Disabled",
1950 (rw & 0x1) ? "R" : "-",
1951 (rw & 0x2) ? "W" : "-",
1952 dram_intlv_sel(pvt, range),
1953 dram_dst_node(pvt, range));
1956 read_dct_base_mask(pvt);
1958 amd64_read_pci_cfg(pvt->F1, DHAR, &pvt->dhar);
1959 amd64_read_dct_pci_cfg(pvt, DBAM0, &pvt->dbam0);
1961 amd64_read_pci_cfg(pvt->F3, F10_ONLINE_SPARE, &pvt->online_spare);
1963 amd64_read_dct_pci_cfg(pvt, F10_DCLR_0, &pvt->dclr0);
1964 amd64_read_dct_pci_cfg(pvt, F10_DCHR_0, &pvt->dchr0);
1966 if (!dct_ganging_enabled(pvt)) {
1967 amd64_read_dct_pci_cfg(pvt, F10_DCLR_1, &pvt->dclr1);
1968 amd64_read_dct_pci_cfg(pvt, F10_DCHR_1, &pvt->dchr1);
1971 if (boot_cpu_data.x86 >= 0x10) {
1972 amd64_read_pci_cfg(pvt->F3, EXT_NB_MCA_CFG, &tmp);
1973 amd64_read_dct_pci_cfg(pvt, DBAM1, &pvt->dbam1);
1976 if (boot_cpu_data.x86 == 0x10 &&
1977 boot_cpu_data.x86_model > 7 &&
1978 /* F3x180[EccSymbolSize]=1 => x8 symbols */
1984 dump_misc_regs(pvt);
1988 * NOTE: CPU Revision Dependent code
1991 * @csrow_nr ChipSelect Row Number (0..NUM_CHIPSELECTS-1)
1992 * k8 private pointer to -->
1993 * DRAM Bank Address mapping register
1995 * DCL register where dual_channel_active is
1997 * The DBAM register consists of 4 sets of 4 bits each definitions:
2000 * 0-3 CSROWs 0 and 1
2001 * 4-7 CSROWs 2 and 3
2002 * 8-11 CSROWs 4 and 5
2003 * 12-15 CSROWs 6 and 7
2005 * Values range from: 0 to 15
2006 * The meaning of the values depends on CPU revision and dual-channel state,
2007 * see relevant BKDG more info.
2009 * The memory controller provides for total of only 8 CSROWs in its current
2010 * architecture. Each "pair" of CSROWs normally represents just one DIMM in
2011 * single channel or two (2) DIMMs in dual channel mode.
2013 * The following code logic collapses the various tables for CSROW based on CPU
2017 * The number of PAGE_SIZE pages on the specified CSROW number it
2021 static u32 amd64_csrow_nr_pages(int csrow_nr, struct amd64_pvt *pvt)
2023 u32 cs_mode, nr_pages;
2026 * The math on this doesn't look right on the surface because x/2*4 can
2027 * be simplified to x*2 but this expression makes use of the fact that
2028 * it is integral math where 1/2=0. This intermediate value becomes the
2029 * number of bits to shift the DBAM register to extract the proper CSROW
2032 cs_mode = (pvt->dbam0 >> ((csrow_nr / 2) * 4)) & 0xF;
2034 nr_pages = pvt->ops->dbam_to_cs(pvt, cs_mode) << (20 - PAGE_SHIFT);
2037 * If dual channel then double the memory size of single channel.
2038 * Channel count is 1 or 2
2040 nr_pages <<= (pvt->channel_count - 1);
2042 debugf0(" (csrow=%d) DBAM map index= %d\n", csrow_nr, cs_mode);
2043 debugf0(" nr_pages= %u channel-count = %d\n",
2044 nr_pages, pvt->channel_count);
2050 * Initialize the array of csrow attribute instances, based on the values
2051 * from pci config hardware registers.
2053 static int init_csrows(struct mem_ctl_info *mci)
2055 struct csrow_info *csrow;
2056 struct amd64_pvt *pvt = mci->pvt_info;
2057 u64 input_addr_min, input_addr_max, sys_addr, base, mask;
2061 amd64_read_pci_cfg(pvt->F3, K8_NBCFG, &val);
2064 pvt->ctl_error_info.nbcfg = val;
2066 debugf0("node %d, NBCFG=0x%08x[ChipKillEccCap: %d|DramEccEn: %d]\n",
2067 pvt->mc_node_id, val,
2068 !!(val & K8_NBCFG_CHIPKILL), !!(val & K8_NBCFG_ECC_ENABLE));
2070 for_each_chip_select(i, 0, pvt) {
2071 csrow = &mci->csrows[i];
2073 if (!csrow_enabled(i, 0, pvt)) {
2074 debugf1("----CSROW %d EMPTY for node %d\n", i,
2079 debugf1("----CSROW %d VALID for MC node %d\n",
2080 i, pvt->mc_node_id);
2083 csrow->nr_pages = amd64_csrow_nr_pages(i, pvt);
2084 find_csrow_limits(mci, i, &input_addr_min, &input_addr_max);
2085 sys_addr = input_addr_to_sys_addr(mci, input_addr_min);
2086 csrow->first_page = (u32) (sys_addr >> PAGE_SHIFT);
2087 sys_addr = input_addr_to_sys_addr(mci, input_addr_max);
2088 csrow->last_page = (u32) (sys_addr >> PAGE_SHIFT);
2090 get_cs_base_and_mask(pvt, i, 0, &base, &mask);
2091 csrow->page_mask = ~mask;
2092 /* 8 bytes of resolution */
2094 csrow->mtype = amd64_determine_memory_type(pvt, i);
2096 debugf1(" for MC node %d csrow %d:\n", pvt->mc_node_id, i);
2097 debugf1(" input_addr_min: 0x%lx input_addr_max: 0x%lx\n",
2098 (unsigned long)input_addr_min,
2099 (unsigned long)input_addr_max);
2100 debugf1(" sys_addr: 0x%lx page_mask: 0x%lx\n",
2101 (unsigned long)sys_addr, csrow->page_mask);
2102 debugf1(" nr_pages: %u first_page: 0x%lx "
2103 "last_page: 0x%lx\n",
2104 (unsigned)csrow->nr_pages,
2105 csrow->first_page, csrow->last_page);
2108 * determine whether CHIPKILL or JUST ECC or NO ECC is operating
2110 if (pvt->nbcfg & K8_NBCFG_ECC_ENABLE)
2112 (pvt->nbcfg & K8_NBCFG_CHIPKILL) ?
2113 EDAC_S4ECD4ED : EDAC_SECDED;
2115 csrow->edac_mode = EDAC_NONE;
2121 /* get all cores on this DCT */
2122 static void get_cpus_on_this_dct_cpumask(struct cpumask *mask, int nid)
2126 for_each_online_cpu(cpu)
2127 if (amd_get_nb_id(cpu) == nid)
2128 cpumask_set_cpu(cpu, mask);
2131 /* check MCG_CTL on all the cpus on this node */
2132 static bool amd64_nb_mce_bank_enabled_on_node(int nid)
2138 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2139 amd64_warn("%s: Error allocating mask\n", __func__);
2143 get_cpus_on_this_dct_cpumask(mask, nid);
2145 rdmsr_on_cpus(mask, MSR_IA32_MCG_CTL, msrs);
2147 for_each_cpu(cpu, mask) {
2148 struct msr *reg = per_cpu_ptr(msrs, cpu);
2149 nbe = reg->l & K8_MSR_MCGCTL_NBE;
2151 debugf0("core: %u, MCG_CTL: 0x%llx, NB MSR is %s\n",
2153 (nbe ? "enabled" : "disabled"));
2161 free_cpumask_var(mask);
2165 static int toggle_ecc_err_reporting(struct ecc_settings *s, u8 nid, bool on)
2167 cpumask_var_t cmask;
2170 if (!zalloc_cpumask_var(&cmask, GFP_KERNEL)) {
2171 amd64_warn("%s: error allocating mask\n", __func__);
2175 get_cpus_on_this_dct_cpumask(cmask, nid);
2177 rdmsr_on_cpus(cmask, MSR_IA32_MCG_CTL, msrs);
2179 for_each_cpu(cpu, cmask) {
2181 struct msr *reg = per_cpu_ptr(msrs, cpu);
2184 if (reg->l & K8_MSR_MCGCTL_NBE)
2185 s->flags.nb_mce_enable = 1;
2187 reg->l |= K8_MSR_MCGCTL_NBE;
2190 * Turn off NB MCE reporting only when it was off before
2192 if (!s->flags.nb_mce_enable)
2193 reg->l &= ~K8_MSR_MCGCTL_NBE;
2196 wrmsr_on_cpus(cmask, MSR_IA32_MCG_CTL, msrs);
2198 free_cpumask_var(cmask);
2203 static bool enable_ecc_error_reporting(struct ecc_settings *s, u8 nid,
2207 u32 value, mask = K8_NBCTL_CECCEn | K8_NBCTL_UECCEn;
2209 if (toggle_ecc_err_reporting(s, nid, ON)) {
2210 amd64_warn("Error enabling ECC reporting over MCGCTL!\n");
2214 amd64_read_pci_cfg(F3, K8_NBCTL, &value);
2216 /* turn on UECCEn and CECCEn bits */
2217 s->old_nbctl = value & mask;
2218 s->nbctl_valid = true;
2221 amd64_write_pci_cfg(F3, K8_NBCTL, value);
2223 amd64_read_pci_cfg(F3, K8_NBCFG, &value);
2225 debugf0("1: node %d, NBCFG=0x%08x[ChipKillEccCap: %d|DramEccEn: %d]\n",
2227 !!(value & K8_NBCFG_CHIPKILL), !!(value & K8_NBCFG_ECC_ENABLE));
2229 if (!(value & K8_NBCFG_ECC_ENABLE)) {
2230 amd64_warn("DRAM ECC disabled on this node, enabling...\n");
2232 s->flags.nb_ecc_prev = 0;
2234 /* Attempt to turn on DRAM ECC Enable */
2235 value |= K8_NBCFG_ECC_ENABLE;
2236 amd64_write_pci_cfg(F3, K8_NBCFG, value);
2238 amd64_read_pci_cfg(F3, K8_NBCFG, &value);
2240 if (!(value & K8_NBCFG_ECC_ENABLE)) {
2241 amd64_warn("Hardware rejected DRAM ECC enable,"
2242 "check memory DIMM configuration.\n");
2245 amd64_info("Hardware accepted DRAM ECC Enable\n");
2248 s->flags.nb_ecc_prev = 1;
2251 debugf0("2: node %d, NBCFG=0x%08x[ChipKillEccCap: %d|DramEccEn: %d]\n",
2253 !!(value & K8_NBCFG_CHIPKILL), !!(value & K8_NBCFG_ECC_ENABLE));
2258 static void restore_ecc_error_reporting(struct ecc_settings *s, u8 nid,
2261 u32 value, mask = K8_NBCTL_CECCEn | K8_NBCTL_UECCEn;
2263 if (!s->nbctl_valid)
2266 amd64_read_pci_cfg(F3, K8_NBCTL, &value);
2268 value |= s->old_nbctl;
2270 amd64_write_pci_cfg(F3, K8_NBCTL, value);
2272 /* restore previous BIOS DRAM ECC "off" setting we force-enabled */
2273 if (!s->flags.nb_ecc_prev) {
2274 amd64_read_pci_cfg(F3, K8_NBCFG, &value);
2275 value &= ~K8_NBCFG_ECC_ENABLE;
2276 amd64_write_pci_cfg(F3, K8_NBCFG, value);
2279 /* restore the NB Enable MCGCTL bit */
2280 if (toggle_ecc_err_reporting(s, nid, OFF))
2281 amd64_warn("Error restoring NB MCGCTL settings!\n");
2285 * EDAC requires that the BIOS have ECC enabled before
2286 * taking over the processing of ECC errors. A command line
2287 * option allows to force-enable hardware ECC later in
2288 * enable_ecc_error_reporting().
2290 static const char *ecc_msg =
2291 "ECC disabled in the BIOS or no ECC capability, module will not load.\n"
2292 " Either enable ECC checking or force module loading by setting "
2293 "'ecc_enable_override'.\n"
2294 " (Note that use of the override may cause unknown side effects.)\n";
2296 static bool ecc_enabled(struct pci_dev *F3, u8 nid)
2300 bool nb_mce_en = false;
2302 amd64_read_pci_cfg(F3, K8_NBCFG, &value);
2304 ecc_en = !!(value & K8_NBCFG_ECC_ENABLE);
2305 amd64_info("DRAM ECC %s.\n", (ecc_en ? "enabled" : "disabled"));
2307 nb_mce_en = amd64_nb_mce_bank_enabled_on_node(nid);
2309 amd64_notice("NB MCE bank disabled, set MSR "
2310 "0x%08x[4] on node %d to enable.\n",
2311 MSR_IA32_MCG_CTL, nid);
2313 if (!ecc_en || !nb_mce_en) {
2314 amd64_notice("%s", ecc_msg);
2320 struct mcidev_sysfs_attribute sysfs_attrs[ARRAY_SIZE(amd64_dbg_attrs) +
2321 ARRAY_SIZE(amd64_inj_attrs) +
2324 struct mcidev_sysfs_attribute terminator = { .attr = { .name = NULL } };
2326 static void set_mc_sysfs_attrs(struct mem_ctl_info *mci)
2328 unsigned int i = 0, j = 0;
2330 for (; i < ARRAY_SIZE(amd64_dbg_attrs); i++)
2331 sysfs_attrs[i] = amd64_dbg_attrs[i];
2333 if (boot_cpu_data.x86 >= 0x10)
2334 for (j = 0; j < ARRAY_SIZE(amd64_inj_attrs); j++, i++)
2335 sysfs_attrs[i] = amd64_inj_attrs[j];
2337 sysfs_attrs[i] = terminator;
2339 mci->mc_driver_sysfs_attributes = sysfs_attrs;
2342 static void setup_mci_misc_attrs(struct mem_ctl_info *mci)
2344 struct amd64_pvt *pvt = mci->pvt_info;
2346 mci->mtype_cap = MEM_FLAG_DDR2 | MEM_FLAG_RDDR2;
2347 mci->edac_ctl_cap = EDAC_FLAG_NONE;
2349 if (pvt->nbcap & K8_NBCAP_SECDED)
2350 mci->edac_ctl_cap |= EDAC_FLAG_SECDED;
2352 if (pvt->nbcap & K8_NBCAP_CHIPKILL)
2353 mci->edac_ctl_cap |= EDAC_FLAG_S4ECD4ED;
2355 mci->edac_cap = amd64_determine_edac_cap(pvt);
2356 mci->mod_name = EDAC_MOD_STR;
2357 mci->mod_ver = EDAC_AMD64_VERSION;
2358 mci->ctl_name = pvt->ctl_name;
2359 mci->dev_name = pci_name(pvt->F2);
2360 mci->ctl_page_to_phys = NULL;
2362 /* memory scrubber interface */
2363 mci->set_sdram_scrub_rate = amd64_set_scrub_rate;
2364 mci->get_sdram_scrub_rate = amd64_get_scrub_rate;
2368 * returns a pointer to the family descriptor on success, NULL otherwise.
2370 static struct amd64_family_type *amd64_per_family_init(struct amd64_pvt *pvt)
2372 u8 fam = boot_cpu_data.x86;
2373 struct amd64_family_type *fam_type = NULL;
2377 fam_type = &amd64_family_types[K8_CPUS];
2378 pvt->ops = &amd64_family_types[K8_CPUS].ops;
2379 pvt->ctl_name = fam_type->ctl_name;
2380 pvt->min_scrubrate = K8_MIN_SCRUB_RATE_BITS;
2383 fam_type = &amd64_family_types[F10_CPUS];
2384 pvt->ops = &amd64_family_types[F10_CPUS].ops;
2385 pvt->ctl_name = fam_type->ctl_name;
2386 pvt->min_scrubrate = F10_MIN_SCRUB_RATE_BITS;
2390 amd64_err("Unsupported family!\n");
2394 pvt->ext_model = boot_cpu_data.x86_model >> 4;
2396 amd64_info("%s %sdetected (node %d).\n", pvt->ctl_name,
2398 (pvt->ext_model >= K8_REV_F ? "revF or later "
2399 : "revE or earlier ")
2400 : ""), pvt->mc_node_id);
2404 static int amd64_init_one_instance(struct pci_dev *F2)
2406 struct amd64_pvt *pvt = NULL;
2407 struct amd64_family_type *fam_type = NULL;
2408 struct mem_ctl_info *mci = NULL;
2410 u8 nid = get_node_id(F2);
2413 pvt = kzalloc(sizeof(struct amd64_pvt), GFP_KERNEL);
2417 pvt->mc_node_id = nid;
2421 fam_type = amd64_per_family_init(pvt);
2426 err = reserve_mc_sibling_devs(pvt, fam_type->f1_id, fam_type->f3_id);
2433 * We need to determine how many memory channels there are. Then use
2434 * that information for calculating the size of the dynamic instance
2435 * tables in the 'mci' structure.
2438 pvt->channel_count = pvt->ops->early_channel_count(pvt);
2439 if (pvt->channel_count < 0)
2443 mci = edac_mc_alloc(0, pvt->csels[0].b_cnt, pvt->channel_count, nid);
2447 mci->pvt_info = pvt;
2448 mci->dev = &pvt->F2->dev;
2450 setup_mci_misc_attrs(mci);
2452 if (init_csrows(mci))
2453 mci->edac_cap = EDAC_FLAG_NONE;
2455 set_mc_sysfs_attrs(mci);
2458 if (edac_mc_add_mc(mci)) {
2459 debugf1("failed edac_mc_add_mc()\n");
2463 /* register stuff with EDAC MCE */
2464 if (report_gart_errors)
2465 amd_report_gart_errors(true);
2467 amd_register_ecc_decoder(amd64_decode_bus_error);
2471 atomic_inc(&drv_instances);
2479 free_mc_sibling_devs(pvt);
2488 static int __devinit amd64_probe_one_instance(struct pci_dev *pdev,
2489 const struct pci_device_id *mc_type)
2491 u8 nid = get_node_id(pdev);
2492 struct pci_dev *F3 = node_to_amd_nb(nid)->misc;
2493 struct ecc_settings *s;
2496 ret = pci_enable_device(pdev);
2498 debugf0("ret=%d\n", ret);
2503 s = kzalloc(sizeof(struct ecc_settings), GFP_KERNEL);
2509 if (!ecc_enabled(F3, nid)) {
2512 if (!ecc_enable_override)
2515 amd64_warn("Forcing ECC on!\n");
2517 if (!enable_ecc_error_reporting(s, nid, F3))
2521 ret = amd64_init_one_instance(pdev);
2523 amd64_err("Error probing instance: %d\n", nid);
2524 restore_ecc_error_reporting(s, nid, F3);
2531 ecc_stngs[nid] = NULL;
2537 static void __devexit amd64_remove_one_instance(struct pci_dev *pdev)
2539 struct mem_ctl_info *mci;
2540 struct amd64_pvt *pvt;
2541 u8 nid = get_node_id(pdev);
2542 struct pci_dev *F3 = node_to_amd_nb(nid)->misc;
2543 struct ecc_settings *s = ecc_stngs[nid];
2545 /* Remove from EDAC CORE tracking list */
2546 mci = edac_mc_del_mc(&pdev->dev);
2550 pvt = mci->pvt_info;
2552 restore_ecc_error_reporting(s, nid, F3);
2554 free_mc_sibling_devs(pvt);
2556 /* unregister from EDAC MCE */
2557 amd_report_gart_errors(false);
2558 amd_unregister_ecc_decoder(amd64_decode_bus_error);
2560 kfree(ecc_stngs[nid]);
2561 ecc_stngs[nid] = NULL;
2563 /* Free the EDAC CORE resources */
2564 mci->pvt_info = NULL;
2572 * This table is part of the interface for loading drivers for PCI devices. The
2573 * PCI core identifies what devices are on a system during boot, and then
2574 * inquiry this table to see if this driver is for a given device found.
2576 static const struct pci_device_id amd64_pci_table[] __devinitdata = {
2578 .vendor = PCI_VENDOR_ID_AMD,
2579 .device = PCI_DEVICE_ID_AMD_K8_NB_MEMCTL,
2580 .subvendor = PCI_ANY_ID,
2581 .subdevice = PCI_ANY_ID,
2586 .vendor = PCI_VENDOR_ID_AMD,
2587 .device = PCI_DEVICE_ID_AMD_10H_NB_DRAM,
2588 .subvendor = PCI_ANY_ID,
2589 .subdevice = PCI_ANY_ID,
2595 MODULE_DEVICE_TABLE(pci, amd64_pci_table);
2597 static struct pci_driver amd64_pci_driver = {
2598 .name = EDAC_MOD_STR,
2599 .probe = amd64_probe_one_instance,
2600 .remove = __devexit_p(amd64_remove_one_instance),
2601 .id_table = amd64_pci_table,
2604 static void setup_pci_device(void)
2606 struct mem_ctl_info *mci;
2607 struct amd64_pvt *pvt;
2615 pvt = mci->pvt_info;
2617 edac_pci_create_generic_ctl(&pvt->F2->dev, EDAC_MOD_STR);
2619 if (!amd64_ctl_pci) {
2620 pr_warning("%s(): Unable to create PCI control\n",
2623 pr_warning("%s(): PCI error report via EDAC not set\n",
2629 static int __init amd64_edac_init(void)
2633 edac_printk(KERN_INFO, EDAC_MOD_STR, EDAC_AMD64_VERSION "\n");
2637 if (amd_cache_northbridges() < 0)
2641 mcis = kzalloc(amd_nb_num() * sizeof(mcis[0]), GFP_KERNEL);
2642 ecc_stngs = kzalloc(amd_nb_num() * sizeof(ecc_stngs[0]), GFP_KERNEL);
2643 if (!(mcis && ecc_stngs))
2646 msrs = msrs_alloc();
2650 err = pci_register_driver(&amd64_pci_driver);
2655 if (!atomic_read(&drv_instances))
2656 goto err_no_instances;
2662 pci_unregister_driver(&amd64_pci_driver);
2679 static void __exit amd64_edac_exit(void)
2682 edac_pci_release_generic_ctl(amd64_ctl_pci);
2684 pci_unregister_driver(&amd64_pci_driver);
2696 module_init(amd64_edac_init);
2697 module_exit(amd64_edac_exit);
2699 MODULE_LICENSE("GPL");
2700 MODULE_AUTHOR("SoftwareBitMaker: Doug Thompson, "
2701 "Dave Peterson, Thayne Harbaugh");
2702 MODULE_DESCRIPTION("MC support for AMD64 memory controllers - "
2703 EDAC_AMD64_VERSION);
2705 module_param(edac_op_state, int, 0444);
2706 MODULE_PARM_DESC(edac_op_state, "EDAC Error Reporting state: 0=Poll,1=NMI");