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, 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, 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);
792 * Determine if the DIMMs have ECC enabled. ECC is enabled ONLY if all the DIMMs
795 static enum edac_type amd64_determine_edac_cap(struct amd64_pvt *pvt)
798 enum dev_type edac_cap = EDAC_FLAG_NONE;
800 bit = (boot_cpu_data.x86 > 0xf || pvt->ext_model >= K8_REV_F)
804 if (pvt->dclr0 & BIT(bit))
805 edac_cap = EDAC_FLAG_SECDED;
811 static void amd64_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt);
813 static void amd64_dump_dramcfg_low(u32 dclr, int chan)
815 debugf1("F2x%d90 (DRAM Cfg Low): 0x%08x\n", chan, dclr);
817 debugf1(" DIMM type: %sbuffered; all DIMMs support ECC: %s\n",
818 (dclr & BIT(16)) ? "un" : "",
819 (dclr & BIT(19)) ? "yes" : "no");
821 debugf1(" PAR/ERR parity: %s\n",
822 (dclr & BIT(8)) ? "enabled" : "disabled");
824 if (boot_cpu_data.x86 == 0x10)
825 debugf1(" DCT 128bit mode width: %s\n",
826 (dclr & BIT(11)) ? "128b" : "64b");
828 debugf1(" x4 logical DIMMs present: L0: %s L1: %s L2: %s L3: %s\n",
829 (dclr & BIT(12)) ? "yes" : "no",
830 (dclr & BIT(13)) ? "yes" : "no",
831 (dclr & BIT(14)) ? "yes" : "no",
832 (dclr & BIT(15)) ? "yes" : "no");
835 /* Display and decode various NB registers for debug purposes. */
836 static void dump_misc_regs(struct amd64_pvt *pvt)
838 debugf1("F3xE8 (NB Cap): 0x%08x\n", pvt->nbcap);
840 debugf1(" NB two channel DRAM capable: %s\n",
841 (pvt->nbcap & NBCAP_DCT_DUAL) ? "yes" : "no");
843 debugf1(" ECC capable: %s, ChipKill ECC capable: %s\n",
844 (pvt->nbcap & NBCAP_SECDED) ? "yes" : "no",
845 (pvt->nbcap & NBCAP_CHIPKILL) ? "yes" : "no");
847 amd64_dump_dramcfg_low(pvt->dclr0, 0);
849 debugf1("F3xB0 (Online Spare): 0x%08x\n", pvt->online_spare);
851 debugf1("F1xF0 (DRAM Hole Address): 0x%08x, base: 0x%08x, "
853 pvt->dhar, dhar_base(pvt),
854 (boot_cpu_data.x86 == 0xf) ? k8_dhar_offset(pvt)
855 : f10_dhar_offset(pvt));
857 debugf1(" DramHoleValid: %s\n", dhar_valid(pvt) ? "yes" : "no");
859 amd64_debug_display_dimm_sizes(0, pvt);
861 /* everything below this point is Fam10h and above */
862 if (boot_cpu_data.x86 == 0xf)
865 amd64_debug_display_dimm_sizes(1, pvt);
867 amd64_info("using %s syndromes.\n", ((pvt->syn_type == 8) ? "x8" : "x4"));
869 /* Only if NOT ganged does dclr1 have valid info */
870 if (!dct_ganging_enabled(pvt))
871 amd64_dump_dramcfg_low(pvt->dclr1, 1);
875 * see BKDG, F2x[1,0][5C:40], F2[1,0][6C:60]
877 static void prep_chip_selects(struct amd64_pvt *pvt)
879 if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_F) {
880 pvt->csels[0].b_cnt = pvt->csels[1].b_cnt = 8;
881 pvt->csels[0].m_cnt = pvt->csels[1].m_cnt = 8;
883 pvt->csels[0].b_cnt = pvt->csels[1].b_cnt = 8;
884 pvt->csels[0].m_cnt = pvt->csels[1].m_cnt = 4;
889 * Function 2 Offset F10_DCSB0; read in the DCS Base and DCS Mask registers
891 static void read_dct_base_mask(struct amd64_pvt *pvt)
895 prep_chip_selects(pvt);
897 for_each_chip_select(cs, 0, pvt) {
898 u32 reg0 = DCSB0 + (cs * 4);
899 u32 reg1 = DCSB1 + (cs * 4);
900 u32 *base0 = &pvt->csels[0].csbases[cs];
901 u32 *base1 = &pvt->csels[1].csbases[cs];
903 if (!amd64_read_dct_pci_cfg(pvt, reg0, base0))
904 debugf0(" DCSB0[%d]=0x%08x reg: F2x%x\n",
907 if (boot_cpu_data.x86 == 0xf || dct_ganging_enabled(pvt))
910 if (!amd64_read_dct_pci_cfg(pvt, reg1, base1))
911 debugf0(" DCSB1[%d]=0x%08x reg: F2x%x\n",
915 for_each_chip_select_mask(cs, 0, pvt) {
916 u32 reg0 = DCSM0 + (cs * 4);
917 u32 reg1 = DCSM1 + (cs * 4);
918 u32 *mask0 = &pvt->csels[0].csmasks[cs];
919 u32 *mask1 = &pvt->csels[1].csmasks[cs];
921 if (!amd64_read_dct_pci_cfg(pvt, reg0, mask0))
922 debugf0(" DCSM0[%d]=0x%08x reg: F2x%x\n",
925 if (boot_cpu_data.x86 == 0xf || dct_ganging_enabled(pvt))
928 if (!amd64_read_dct_pci_cfg(pvt, reg1, mask1))
929 debugf0(" DCSM1[%d]=0x%08x reg: F2x%x\n",
934 static enum mem_type amd64_determine_memory_type(struct amd64_pvt *pvt, int cs)
938 /* F15h supports only DDR3 */
939 if (boot_cpu_data.x86 >= 0x15)
940 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR3 : MEM_RDDR3;
941 else if (boot_cpu_data.x86 == 0x10 || pvt->ext_model >= K8_REV_F) {
942 if (pvt->dchr0 & DDR3_MODE)
943 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR3 : MEM_RDDR3;
945 type = (pvt->dclr0 & BIT(16)) ? MEM_DDR2 : MEM_RDDR2;
947 type = (pvt->dclr0 & BIT(18)) ? MEM_DDR : MEM_RDDR;
950 amd64_info("CS%d: %s\n", cs, edac_mem_types[type]);
955 /* Get the number of DCT channels the memory controller is using. */
956 static int k8_early_channel_count(struct amd64_pvt *pvt)
960 if (pvt->ext_model >= K8_REV_F)
961 /* RevF (NPT) and later */
962 flag = pvt->dclr0 & F10_WIDTH_128;
964 /* RevE and earlier */
965 flag = pvt->dclr0 & REVE_WIDTH_128;
970 return (flag) ? 2 : 1;
973 /* On F10h and later ErrAddr is MC4_ADDR[47:1] */
974 static u64 get_error_address(struct mce *m)
979 if (boot_cpu_data.x86 == 0xf) {
984 return m->addr & GENMASK(start_bit, end_bit);
987 static void read_dram_base_limit_regs(struct amd64_pvt *pvt, unsigned range)
989 u32 off = range << 3;
991 amd64_read_pci_cfg(pvt->F1, DRAM_BASE_LO + off, &pvt->ranges[range].base.lo);
992 amd64_read_pci_cfg(pvt->F1, DRAM_LIMIT_LO + off, &pvt->ranges[range].lim.lo);
994 if (boot_cpu_data.x86 == 0xf)
997 if (!dram_rw(pvt, range))
1000 amd64_read_pci_cfg(pvt->F1, DRAM_BASE_HI + off, &pvt->ranges[range].base.hi);
1001 amd64_read_pci_cfg(pvt->F1, DRAM_LIMIT_HI + off, &pvt->ranges[range].lim.hi);
1004 static void k8_map_sysaddr_to_csrow(struct mem_ctl_info *mci, u64 sys_addr,
1007 struct mem_ctl_info *src_mci;
1008 struct amd64_pvt *pvt = mci->pvt_info;
1012 /* CHIPKILL enabled */
1013 if (pvt->nbcfg & NBCFG_CHIPKILL) {
1014 channel = get_channel_from_ecc_syndrome(mci, syndrome);
1017 * Syndrome didn't map, so we don't know which of the
1018 * 2 DIMMs is in error. So we need to ID 'both' of them
1021 amd64_mc_warn(mci, "unknown syndrome 0x%04x - possible "
1022 "error reporting race\n", syndrome);
1023 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1028 * non-chipkill ecc mode
1030 * The k8 documentation is unclear about how to determine the
1031 * channel number when using non-chipkill memory. This method
1032 * was obtained from email communication with someone at AMD.
1033 * (Wish the email was placed in this comment - norsk)
1035 channel = ((sys_addr & BIT(3)) != 0);
1039 * Find out which node the error address belongs to. This may be
1040 * different from the node that detected the error.
1042 src_mci = find_mc_by_sys_addr(mci, sys_addr);
1044 amd64_mc_err(mci, "failed to map error addr 0x%lx to a node\n",
1045 (unsigned long)sys_addr);
1046 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1050 /* Now map the sys_addr to a CSROW */
1051 csrow = sys_addr_to_csrow(src_mci, sys_addr);
1053 edac_mc_handle_ce_no_info(src_mci, EDAC_MOD_STR);
1055 error_address_to_page_and_offset(sys_addr, &page, &offset);
1057 edac_mc_handle_ce(src_mci, page, offset, syndrome, csrow,
1058 channel, EDAC_MOD_STR);
1062 static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, int cs_mode)
1066 if (pvt->ext_model >= K8_REV_F)
1067 dbam_map = ddr2_dbam;
1068 else if (pvt->ext_model >= K8_REV_D)
1069 dbam_map = ddr2_dbam_revD;
1071 dbam_map = ddr2_dbam_revCG;
1073 return dbam_map[cs_mode];
1077 * Get the number of DCT channels in use.
1080 * number of Memory Channels in operation
1082 * contents of the DCL0_LOW register
1084 static int f1x_early_channel_count(struct amd64_pvt *pvt)
1086 int i, j, channels = 0;
1088 /* On F10h, if we are in 128 bit mode, then we are using 2 channels */
1089 if (boot_cpu_data.x86 == 0x10 && (pvt->dclr0 & F10_WIDTH_128))
1093 * Need to check if in unganged mode: In such, there are 2 channels,
1094 * but they are not in 128 bit mode and thus the above 'dclr0' status
1097 * Need to check DCT0[0] and DCT1[0] to see if only one of them has
1098 * their CSEnable bit on. If so, then SINGLE DIMM case.
1100 debugf0("Data width is not 128 bits - need more decoding\n");
1103 * Check DRAM Bank Address Mapping values for each DIMM to see if there
1104 * is more than just one DIMM present in unganged mode. Need to check
1105 * both controllers since DIMMs can be placed in either one.
1107 for (i = 0; i < 2; i++) {
1108 u32 dbam = (i ? pvt->dbam1 : pvt->dbam0);
1110 for (j = 0; j < 4; j++) {
1111 if (DBAM_DIMM(j, dbam) > 0) {
1121 amd64_info("MCT channel count: %d\n", channels);
1126 static int f10_dbam_to_chip_select(struct amd64_pvt *pvt, int cs_mode)
1130 if (pvt->dchr0 & DDR3_MODE || pvt->dchr1 & DDR3_MODE)
1131 dbam_map = ddr3_dbam;
1133 dbam_map = ddr2_dbam;
1135 return dbam_map[cs_mode];
1138 static void f10_read_dram_ctl_register(struct amd64_pvt *pvt)
1141 if (!amd64_read_dct_pci_cfg(pvt, DCT_SEL_LO, &pvt->dct_sel_lo)) {
1142 debugf0("F2x110 (DCTSelLow): 0x%08x, High range addrs at: 0x%x\n",
1143 pvt->dct_sel_lo, dct_sel_baseaddr(pvt));
1145 debugf0(" mode: %s, All DCTs on: %s\n",
1146 (dct_ganging_enabled(pvt) ? "ganged" : "unganged"),
1147 (dct_dram_enabled(pvt) ? "yes" : "no"));
1149 if (!dct_ganging_enabled(pvt))
1150 debugf0(" Address range split per DCT: %s\n",
1151 (dct_high_range_enabled(pvt) ? "yes" : "no"));
1153 debugf0(" data interleave for ECC: %s, "
1154 "DRAM cleared since last warm reset: %s\n",
1155 (dct_data_intlv_enabled(pvt) ? "enabled" : "disabled"),
1156 (dct_memory_cleared(pvt) ? "yes" : "no"));
1158 debugf0(" channel interleave: %s, "
1159 "interleave bits selector: 0x%x\n",
1160 (dct_interleave_enabled(pvt) ? "enabled" : "disabled"),
1161 dct_sel_interleave_addr(pvt));
1164 amd64_read_dct_pci_cfg(pvt, DCT_SEL_HI, &pvt->dct_sel_hi);
1168 * Determine channel (DCT) based on the interleaving mode: F10h BKDG, 2.8.9 Memory
1169 * Interleaving Modes.
1171 static u8 f10_determine_channel(struct amd64_pvt *pvt, u64 sys_addr,
1172 bool hi_range_sel, u8 intlv_en)
1174 u32 dct_sel_high = (pvt->dct_sel_lo >> 1) & 1;
1176 if (dct_ganging_enabled(pvt))
1180 return dct_sel_high;
1183 * see F2x110[DctSelIntLvAddr] - channel interleave mode
1185 if (dct_interleave_enabled(pvt)) {
1186 u8 intlv_addr = dct_sel_interleave_addr(pvt);
1188 /* return DCT select function: 0=DCT0, 1=DCT1 */
1190 return sys_addr >> 6 & 1;
1192 if (intlv_addr & 0x2) {
1193 u8 shift = intlv_addr & 0x1 ? 9 : 6;
1194 u32 temp = hweight_long((u32) ((sys_addr >> 16) & 0x1F)) % 2;
1196 return ((sys_addr >> shift) & 1) ^ temp;
1199 return (sys_addr >> (12 + hweight8(intlv_en))) & 1;
1202 if (dct_high_range_enabled(pvt))
1203 return ~dct_sel_high & 1;
1208 /* Convert the sys_addr to the normalized DCT address */
1209 static u64 f10_get_norm_dct_addr(struct amd64_pvt *pvt, int range,
1210 u64 sys_addr, bool hi_rng,
1211 u32 dct_sel_base_addr)
1214 u64 dram_base = get_dram_base(pvt, range);
1215 u64 hole_off = f10_dhar_offset(pvt);
1216 u32 hole_valid = dhar_valid(pvt);
1217 u64 dct_sel_base_off = (pvt->dct_sel_hi & 0xFFFFFC00) << 16;
1222 * base address of high range is below 4Gb
1223 * (bits [47:27] at [31:11])
1224 * DRAM address space on this DCT is hoisted above 4Gb &&
1227 * remove hole offset from sys_addr
1229 * remove high range offset from sys_addr
1231 if ((!(dct_sel_base_addr >> 16) ||
1232 dct_sel_base_addr < dhar_base(pvt)) &&
1234 (sys_addr >= BIT_64(32)))
1235 chan_off = hole_off;
1237 chan_off = dct_sel_base_off;
1241 * we have a valid hole &&
1246 * remove dram base to normalize to DCT address
1248 if (hole_valid && (sys_addr >= BIT_64(32)))
1249 chan_off = hole_off;
1251 chan_off = dram_base;
1254 return (sys_addr & GENMASK(6,47)) - (chan_off & GENMASK(23,47));
1257 /* Hack for the time being - Can we get this from BIOS?? */
1258 #define CH0SPARE_RANK 0
1259 #define CH1SPARE_RANK 1
1262 * checks if the csrow passed in is marked as SPARED, if so returns the new
1265 static int f10_process_possible_spare(struct amd64_pvt *pvt, u8 dct, int csrow)
1270 /* Depending on channel, isolate respective SPARING info */
1272 swap_done = F10_ONLINE_SPARE_SWAPDONE1(pvt->online_spare);
1273 bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS1(pvt->online_spare);
1274 if (swap_done && (csrow == bad_dram_cs))
1275 csrow = CH1SPARE_RANK;
1277 swap_done = F10_ONLINE_SPARE_SWAPDONE0(pvt->online_spare);
1278 bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS0(pvt->online_spare);
1279 if (swap_done && (csrow == bad_dram_cs))
1280 csrow = CH0SPARE_RANK;
1286 * Iterate over the DRAM DCT "base" and "mask" registers looking for a
1287 * SystemAddr match on the specified 'ChannelSelect' and 'NodeID'
1290 * -EINVAL: NOT FOUND
1291 * 0..csrow = Chip-Select Row
1293 static int f10_lookup_addr_in_dct(u64 in_addr, u32 nid, u8 dct)
1295 struct mem_ctl_info *mci;
1296 struct amd64_pvt *pvt;
1297 u64 cs_base, cs_mask;
1298 int cs_found = -EINVAL;
1305 pvt = mci->pvt_info;
1307 debugf1("input addr: 0x%llx, DCT: %d\n", in_addr, dct);
1309 for_each_chip_select(csrow, dct, pvt) {
1310 if (!csrow_enabled(csrow, dct, pvt))
1313 get_cs_base_and_mask(pvt, csrow, dct, &cs_base, &cs_mask);
1315 debugf1(" CSROW=%d CSBase=0x%llx CSMask=0x%llx\n",
1316 csrow, cs_base, cs_mask);
1320 debugf1(" (InputAddr & ~CSMask)=0x%llx "
1321 "(CSBase & ~CSMask)=0x%llx\n",
1322 (in_addr & cs_mask), (cs_base & cs_mask));
1324 if ((in_addr & cs_mask) == (cs_base & cs_mask)) {
1325 cs_found = f10_process_possible_spare(pvt, dct, csrow);
1327 debugf1(" MATCH csrow=%d\n", cs_found);
1335 * See F2x10C. Non-interleaved graphics framebuffer memory under the 16G is
1336 * swapped with a region located at the bottom of memory so that the GPU can use
1337 * the interleaved region and thus two channels.
1339 static u64 f10_swap_interleaved_region(struct amd64_pvt *pvt, u64 sys_addr)
1341 u32 swap_reg, swap_base, swap_limit, rgn_size, tmp_addr;
1343 if (boot_cpu_data.x86 == 0x10) {
1344 /* only revC3 and revE have that feature */
1345 if (boot_cpu_data.x86_model < 4 ||
1346 (boot_cpu_data.x86_model < 0xa &&
1347 boot_cpu_data.x86_mask < 3))
1351 amd64_read_dct_pci_cfg(pvt, SWAP_INTLV_REG, &swap_reg);
1353 if (!(swap_reg & 0x1))
1356 swap_base = (swap_reg >> 3) & 0x7f;
1357 swap_limit = (swap_reg >> 11) & 0x7f;
1358 rgn_size = (swap_reg >> 20) & 0x7f;
1359 tmp_addr = sys_addr >> 27;
1361 if (!(sys_addr >> 34) &&
1362 (((tmp_addr >= swap_base) &&
1363 (tmp_addr <= swap_limit)) ||
1364 (tmp_addr < rgn_size)))
1365 return sys_addr ^ (u64)swap_base << 27;
1370 /* For a given @dram_range, check if @sys_addr falls within it. */
1371 static int f10_match_to_this_node(struct amd64_pvt *pvt, int range,
1372 u64 sys_addr, int *nid, int *chan_sel)
1374 int cs_found = -EINVAL;
1376 u32 tmp, dct_sel_base;
1378 bool high_range = false;
1380 u8 node_id = dram_dst_node(pvt, range);
1381 u8 intlv_en = dram_intlv_en(pvt, range);
1382 u32 intlv_sel = dram_intlv_sel(pvt, range);
1384 debugf1("(range %d) SystemAddr= 0x%llx Limit=0x%llx\n",
1385 range, sys_addr, get_dram_limit(pvt, range));
1388 (intlv_sel != ((sys_addr >> 12) & intlv_en)))
1391 sys_addr = f10_swap_interleaved_region(pvt, sys_addr);
1393 dct_sel_base = dct_sel_baseaddr(pvt);
1396 * check whether addresses >= DctSelBaseAddr[47:27] are to be used to
1397 * select between DCT0 and DCT1.
1399 if (dct_high_range_enabled(pvt) &&
1400 !dct_ganging_enabled(pvt) &&
1401 ((sys_addr >> 27) >= (dct_sel_base >> 11)))
1404 channel = f10_determine_channel(pvt, sys_addr, high_range, intlv_en);
1406 chan_addr = f10_get_norm_dct_addr(pvt, range, sys_addr,
1407 high_range, dct_sel_base);
1409 /* Remove node interleaving, see F1x120 */
1411 chan_addr = ((chan_addr >> (12 + hweight8(intlv_en))) << 12) |
1412 (chan_addr & 0xfff);
1414 /* remove channel interleave and hash */
1415 if (dct_interleave_enabled(pvt) &&
1416 !dct_high_range_enabled(pvt) &&
1417 !dct_ganging_enabled(pvt)) {
1418 if (dct_sel_interleave_addr(pvt) != 1)
1419 chan_addr = (chan_addr >> 1) & GENMASK(6, 63);
1421 tmp = chan_addr & 0xFC0;
1422 chan_addr = ((chan_addr & GENMASK(14, 63)) >> 1) | tmp;
1426 debugf1(" (ChannelAddrLong=0x%llx)\n", chan_addr);
1428 cs_found = f10_lookup_addr_in_dct(chan_addr, node_id, channel);
1430 if (cs_found >= 0) {
1432 *chan_sel = channel;
1437 static int f10_translate_sysaddr_to_cs(struct amd64_pvt *pvt, u64 sys_addr,
1438 int *node, int *chan_sel)
1440 int range, cs_found = -EINVAL;
1442 for (range = 0; range < DRAM_RANGES; range++) {
1444 if (!dram_rw(pvt, range))
1447 if ((get_dram_base(pvt, range) <= sys_addr) &&
1448 (get_dram_limit(pvt, range) >= sys_addr)) {
1450 cs_found = f10_match_to_this_node(pvt, range,
1461 * For reference see "2.8.5 Routing DRAM Requests" in F10 BKDG. This code maps
1462 * a @sys_addr to NodeID, DCT (channel) and chip select (CSROW).
1464 * The @sys_addr is usually an error address received from the hardware
1467 static void f10_map_sysaddr_to_csrow(struct mem_ctl_info *mci, u64 sys_addr,
1470 struct amd64_pvt *pvt = mci->pvt_info;
1472 int nid, csrow, chan = 0;
1474 csrow = f10_translate_sysaddr_to_cs(pvt, sys_addr, &nid, &chan);
1477 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1481 error_address_to_page_and_offset(sys_addr, &page, &offset);
1484 * We need the syndromes for channel detection only when we're
1485 * ganged. Otherwise @chan should already contain the channel at
1488 if (dct_ganging_enabled(pvt))
1489 chan = get_channel_from_ecc_syndrome(mci, syndrome);
1492 edac_mc_handle_ce(mci, page, offset, syndrome, csrow, chan,
1496 * Channel unknown, report all channels on this CSROW as failed.
1498 for (chan = 0; chan < mci->csrows[csrow].nr_channels; chan++)
1499 edac_mc_handle_ce(mci, page, offset, syndrome,
1500 csrow, chan, EDAC_MOD_STR);
1504 * debug routine to display the memory sizes of all logical DIMMs and its
1507 static void amd64_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt)
1509 int dimm, size0, size1, factor = 0;
1510 u32 *dcsb = ctrl ? pvt->csels[1].csbases : pvt->csels[0].csbases;
1511 u32 dbam = ctrl ? pvt->dbam1 : pvt->dbam0;
1513 if (boot_cpu_data.x86 == 0xf) {
1514 if (pvt->dclr0 & F10_WIDTH_128)
1517 /* K8 families < revF not supported yet */
1518 if (pvt->ext_model < K8_REV_F)
1524 dbam = (ctrl && !dct_ganging_enabled(pvt)) ? pvt->dbam1 : pvt->dbam0;
1525 dcsb = (ctrl && !dct_ganging_enabled(pvt)) ? pvt->csels[1].csbases
1526 : pvt->csels[0].csbases;
1528 debugf1("F2x%d80 (DRAM Bank Address Mapping): 0x%08x\n", ctrl, dbam);
1530 edac_printk(KERN_DEBUG, EDAC_MC, "DCT%d chip selects:\n", ctrl);
1532 /* Dump memory sizes for DIMM and its CSROWs */
1533 for (dimm = 0; dimm < 4; dimm++) {
1536 if (dcsb[dimm*2] & DCSB_CS_ENABLE)
1537 size0 = pvt->ops->dbam_to_cs(pvt, DBAM_DIMM(dimm, dbam));
1540 if (dcsb[dimm*2 + 1] & DCSB_CS_ENABLE)
1541 size1 = pvt->ops->dbam_to_cs(pvt, DBAM_DIMM(dimm, dbam));
1543 amd64_info(EDAC_MC ": %d: %5dMB %d: %5dMB\n",
1544 dimm * 2, size0 << factor,
1545 dimm * 2 + 1, size1 << factor);
1549 static struct amd64_family_type amd64_family_types[] = {
1552 .f1_id = PCI_DEVICE_ID_AMD_K8_NB_ADDRMAP,
1553 .f3_id = PCI_DEVICE_ID_AMD_K8_NB_MISC,
1555 .early_channel_count = k8_early_channel_count,
1556 .map_sysaddr_to_csrow = k8_map_sysaddr_to_csrow,
1557 .dbam_to_cs = k8_dbam_to_chip_select,
1558 .read_dct_pci_cfg = k8_read_dct_pci_cfg,
1563 .f1_id = PCI_DEVICE_ID_AMD_10H_NB_MAP,
1564 .f3_id = PCI_DEVICE_ID_AMD_10H_NB_MISC,
1566 .early_channel_count = f1x_early_channel_count,
1567 .read_dram_ctl_register = f10_read_dram_ctl_register,
1568 .map_sysaddr_to_csrow = f10_map_sysaddr_to_csrow,
1569 .dbam_to_cs = f10_dbam_to_chip_select,
1570 .read_dct_pci_cfg = f10_read_dct_pci_cfg,
1576 .early_channel_count = f1x_early_channel_count,
1577 .read_dct_pci_cfg = f15_read_dct_pci_cfg,
1582 static struct pci_dev *pci_get_related_function(unsigned int vendor,
1583 unsigned int device,
1584 struct pci_dev *related)
1586 struct pci_dev *dev = NULL;
1588 dev = pci_get_device(vendor, device, dev);
1590 if ((dev->bus->number == related->bus->number) &&
1591 (PCI_SLOT(dev->devfn) == PCI_SLOT(related->devfn)))
1593 dev = pci_get_device(vendor, device, dev);
1600 * These are tables of eigenvectors (one per line) which can be used for the
1601 * construction of the syndrome tables. The modified syndrome search algorithm
1602 * uses those to find the symbol in error and thus the DIMM.
1604 * Algorithm courtesy of Ross LaFetra from AMD.
1606 static u16 x4_vectors[] = {
1607 0x2f57, 0x1afe, 0x66cc, 0xdd88,
1608 0x11eb, 0x3396, 0x7f4c, 0xeac8,
1609 0x0001, 0x0002, 0x0004, 0x0008,
1610 0x1013, 0x3032, 0x4044, 0x8088,
1611 0x106b, 0x30d6, 0x70fc, 0xe0a8,
1612 0x4857, 0xc4fe, 0x13cc, 0x3288,
1613 0x1ac5, 0x2f4a, 0x5394, 0xa1e8,
1614 0x1f39, 0x251e, 0xbd6c, 0x6bd8,
1615 0x15c1, 0x2a42, 0x89ac, 0x4758,
1616 0x2b03, 0x1602, 0x4f0c, 0xca08,
1617 0x1f07, 0x3a0e, 0x6b04, 0xbd08,
1618 0x8ba7, 0x465e, 0x244c, 0x1cc8,
1619 0x2b87, 0x164e, 0x642c, 0xdc18,
1620 0x40b9, 0x80de, 0x1094, 0x20e8,
1621 0x27db, 0x1eb6, 0x9dac, 0x7b58,
1622 0x11c1, 0x2242, 0x84ac, 0x4c58,
1623 0x1be5, 0x2d7a, 0x5e34, 0xa718,
1624 0x4b39, 0x8d1e, 0x14b4, 0x28d8,
1625 0x4c97, 0xc87e, 0x11fc, 0x33a8,
1626 0x8e97, 0x497e, 0x2ffc, 0x1aa8,
1627 0x16b3, 0x3d62, 0x4f34, 0x8518,
1628 0x1e2f, 0x391a, 0x5cac, 0xf858,
1629 0x1d9f, 0x3b7a, 0x572c, 0xfe18,
1630 0x15f5, 0x2a5a, 0x5264, 0xa3b8,
1631 0x1dbb, 0x3b66, 0x715c, 0xe3f8,
1632 0x4397, 0xc27e, 0x17fc, 0x3ea8,
1633 0x1617, 0x3d3e, 0x6464, 0xb8b8,
1634 0x23ff, 0x12aa, 0xab6c, 0x56d8,
1635 0x2dfb, 0x1ba6, 0x913c, 0x7328,
1636 0x185d, 0x2ca6, 0x7914, 0x9e28,
1637 0x171b, 0x3e36, 0x7d7c, 0xebe8,
1638 0x4199, 0x82ee, 0x19f4, 0x2e58,
1639 0x4807, 0xc40e, 0x130c, 0x3208,
1640 0x1905, 0x2e0a, 0x5804, 0xac08,
1641 0x213f, 0x132a, 0xadfc, 0x5ba8,
1642 0x19a9, 0x2efe, 0xb5cc, 0x6f88,
1645 static u16 x8_vectors[] = {
1646 0x0145, 0x028a, 0x2374, 0x43c8, 0xa1f0, 0x0520, 0x0a40, 0x1480,
1647 0x0211, 0x0422, 0x0844, 0x1088, 0x01b0, 0x44e0, 0x23c0, 0xed80,
1648 0x1011, 0x0116, 0x022c, 0x0458, 0x08b0, 0x8c60, 0x2740, 0x4e80,
1649 0x0411, 0x0822, 0x1044, 0x0158, 0x02b0, 0x2360, 0x46c0, 0xab80,
1650 0x0811, 0x1022, 0x012c, 0x0258, 0x04b0, 0x4660, 0x8cc0, 0x2780,
1651 0x2071, 0x40e2, 0xa0c4, 0x0108, 0x0210, 0x0420, 0x0840, 0x1080,
1652 0x4071, 0x80e2, 0x0104, 0x0208, 0x0410, 0x0820, 0x1040, 0x2080,
1653 0x8071, 0x0102, 0x0204, 0x0408, 0x0810, 0x1020, 0x2040, 0x4080,
1654 0x019d, 0x03d6, 0x136c, 0x2198, 0x50b0, 0xb2e0, 0x0740, 0x0e80,
1655 0x0189, 0x03ea, 0x072c, 0x0e58, 0x1cb0, 0x56e0, 0x37c0, 0xf580,
1656 0x01fd, 0x0376, 0x06ec, 0x0bb8, 0x1110, 0x2220, 0x4440, 0x8880,
1657 0x0163, 0x02c6, 0x1104, 0x0758, 0x0eb0, 0x2be0, 0x6140, 0xc280,
1658 0x02fd, 0x01c6, 0x0b5c, 0x1108, 0x07b0, 0x25a0, 0x8840, 0x6180,
1659 0x0801, 0x012e, 0x025c, 0x04b8, 0x1370, 0x26e0, 0x57c0, 0xb580,
1660 0x0401, 0x0802, 0x015c, 0x02b8, 0x22b0, 0x13e0, 0x7140, 0xe280,
1661 0x0201, 0x0402, 0x0804, 0x01b8, 0x11b0, 0x31a0, 0x8040, 0x7180,
1662 0x0101, 0x0202, 0x0404, 0x0808, 0x1010, 0x2020, 0x4040, 0x8080,
1663 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
1664 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x8000,
1667 static int decode_syndrome(u16 syndrome, u16 *vectors, int num_vecs,
1670 unsigned int i, err_sym;
1672 for (err_sym = 0; err_sym < num_vecs / v_dim; err_sym++) {
1674 int v_idx = err_sym * v_dim;
1675 int v_end = (err_sym + 1) * v_dim;
1677 /* walk over all 16 bits of the syndrome */
1678 for (i = 1; i < (1U << 16); i <<= 1) {
1680 /* if bit is set in that eigenvector... */
1681 if (v_idx < v_end && vectors[v_idx] & i) {
1682 u16 ev_comp = vectors[v_idx++];
1684 /* ... and bit set in the modified syndrome, */
1694 /* can't get to zero, move to next symbol */
1699 debugf0("syndrome(%x) not found\n", syndrome);
1703 static int map_err_sym_to_channel(int err_sym, int sym_size)
1716 return err_sym >> 4;
1722 /* imaginary bits not in a DIMM */
1724 WARN(1, KERN_ERR "Invalid error symbol: 0x%x\n",
1736 return err_sym >> 3;
1742 static int get_channel_from_ecc_syndrome(struct mem_ctl_info *mci, u16 syndrome)
1744 struct amd64_pvt *pvt = mci->pvt_info;
1747 if (pvt->syn_type == 8)
1748 err_sym = decode_syndrome(syndrome, x8_vectors,
1749 ARRAY_SIZE(x8_vectors),
1751 else if (pvt->syn_type == 4)
1752 err_sym = decode_syndrome(syndrome, x4_vectors,
1753 ARRAY_SIZE(x4_vectors),
1756 amd64_warn("Illegal syndrome type: %u\n", pvt->syn_type);
1760 return map_err_sym_to_channel(err_sym, pvt->syn_type);
1764 * Handle any Correctable Errors (CEs) that have occurred. Check for valid ERROR
1765 * ADDRESS and process.
1767 static void amd64_handle_ce(struct mem_ctl_info *mci, struct mce *m)
1769 struct amd64_pvt *pvt = mci->pvt_info;
1773 /* Ensure that the Error Address is VALID */
1774 if (!(m->status & MCI_STATUS_ADDRV)) {
1775 amd64_mc_err(mci, "HW has no ERROR_ADDRESS available\n");
1776 edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
1780 sys_addr = get_error_address(m);
1781 syndrome = extract_syndrome(m->status);
1783 amd64_mc_err(mci, "CE ERROR_ADDRESS= 0x%llx\n", sys_addr);
1785 pvt->ops->map_sysaddr_to_csrow(mci, sys_addr, syndrome);
1788 /* Handle any Un-correctable Errors (UEs) */
1789 static void amd64_handle_ue(struct mem_ctl_info *mci, struct mce *m)
1791 struct mem_ctl_info *log_mci, *src_mci = NULL;
1798 if (!(m->status & MCI_STATUS_ADDRV)) {
1799 amd64_mc_err(mci, "HW has no ERROR_ADDRESS available\n");
1800 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
1804 sys_addr = get_error_address(m);
1807 * Find out which node the error address belongs to. This may be
1808 * different from the node that detected the error.
1810 src_mci = find_mc_by_sys_addr(mci, sys_addr);
1812 amd64_mc_err(mci, "ERROR ADDRESS (0x%lx) NOT mapped to a MC\n",
1813 (unsigned long)sys_addr);
1814 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
1820 csrow = sys_addr_to_csrow(log_mci, sys_addr);
1822 amd64_mc_err(mci, "ERROR_ADDRESS (0x%lx) NOT mapped to CS\n",
1823 (unsigned long)sys_addr);
1824 edac_mc_handle_ue_no_info(log_mci, EDAC_MOD_STR);
1826 error_address_to_page_and_offset(sys_addr, &page, &offset);
1827 edac_mc_handle_ue(log_mci, page, offset, csrow, EDAC_MOD_STR);
1831 static inline void __amd64_decode_bus_error(struct mem_ctl_info *mci,
1834 u16 ec = EC(m->status);
1835 u8 xec = XEC(m->status, 0x1f);
1836 u8 ecc_type = (m->status >> 45) & 0x3;
1838 /* Bail early out if this was an 'observed' error */
1839 if (PP(ec) == NBSL_PP_OBS)
1842 /* Do only ECC errors */
1843 if (xec && xec != F10_NBSL_EXT_ERR_ECC)
1847 amd64_handle_ce(mci, m);
1848 else if (ecc_type == 1)
1849 amd64_handle_ue(mci, m);
1852 void amd64_decode_bus_error(int node_id, struct mce *m, u32 nbcfg)
1854 struct mem_ctl_info *mci = mcis[node_id];
1856 __amd64_decode_bus_error(mci, m);
1860 * Use pvt->F2 which contains the F2 CPU PCI device to get the related
1861 * F1 (AddrMap) and F3 (Misc) devices. Return negative value on error.
1863 static int reserve_mc_sibling_devs(struct amd64_pvt *pvt, u16 f1_id, u16 f3_id)
1865 /* Reserve the ADDRESS MAP Device */
1866 pvt->F1 = pci_get_related_function(pvt->F2->vendor, f1_id, pvt->F2);
1868 amd64_err("error address map device not found: "
1869 "vendor %x device 0x%x (broken BIOS?)\n",
1870 PCI_VENDOR_ID_AMD, f1_id);
1874 /* Reserve the MISC Device */
1875 pvt->F3 = pci_get_related_function(pvt->F2->vendor, f3_id, pvt->F2);
1877 pci_dev_put(pvt->F1);
1880 amd64_err("error F3 device not found: "
1881 "vendor %x device 0x%x (broken BIOS?)\n",
1882 PCI_VENDOR_ID_AMD, f3_id);
1886 debugf1("F1: %s\n", pci_name(pvt->F1));
1887 debugf1("F2: %s\n", pci_name(pvt->F2));
1888 debugf1("F3: %s\n", pci_name(pvt->F3));
1893 static void free_mc_sibling_devs(struct amd64_pvt *pvt)
1895 pci_dev_put(pvt->F1);
1896 pci_dev_put(pvt->F3);
1900 * Retrieve the hardware registers of the memory controller (this includes the
1901 * 'Address Map' and 'Misc' device regs)
1903 static void read_mc_regs(struct amd64_pvt *pvt)
1910 * Retrieve TOP_MEM and TOP_MEM2; no masking off of reserved bits since
1911 * those are Read-As-Zero
1913 rdmsrl(MSR_K8_TOP_MEM1, pvt->top_mem);
1914 debugf0(" TOP_MEM: 0x%016llx\n", pvt->top_mem);
1916 /* check first whether TOP_MEM2 is enabled */
1917 rdmsrl(MSR_K8_SYSCFG, msr_val);
1918 if (msr_val & (1U << 21)) {
1919 rdmsrl(MSR_K8_TOP_MEM2, pvt->top_mem2);
1920 debugf0(" TOP_MEM2: 0x%016llx\n", pvt->top_mem2);
1922 debugf0(" TOP_MEM2 disabled.\n");
1924 amd64_read_pci_cfg(pvt->F3, NBCAP, &pvt->nbcap);
1926 if (pvt->ops->read_dram_ctl_register)
1927 pvt->ops->read_dram_ctl_register(pvt);
1929 for (range = 0; range < DRAM_RANGES; range++) {
1932 /* read settings for this DRAM range */
1933 read_dram_base_limit_regs(pvt, range);
1935 rw = dram_rw(pvt, range);
1939 debugf1(" DRAM range[%d], base: 0x%016llx; limit: 0x%016llx\n",
1941 get_dram_base(pvt, range),
1942 get_dram_limit(pvt, range));
1944 debugf1(" IntlvEn=%s; Range access: %s%s IntlvSel=%d DstNode=%d\n",
1945 dram_intlv_en(pvt, range) ? "Enabled" : "Disabled",
1946 (rw & 0x1) ? "R" : "-",
1947 (rw & 0x2) ? "W" : "-",
1948 dram_intlv_sel(pvt, range),
1949 dram_dst_node(pvt, range));
1952 read_dct_base_mask(pvt);
1954 amd64_read_pci_cfg(pvt->F1, DHAR, &pvt->dhar);
1955 amd64_read_dct_pci_cfg(pvt, DBAM0, &pvt->dbam0);
1957 amd64_read_pci_cfg(pvt->F3, F10_ONLINE_SPARE, &pvt->online_spare);
1959 amd64_read_dct_pci_cfg(pvt, DCLR0, &pvt->dclr0);
1960 amd64_read_dct_pci_cfg(pvt, DCHR0, &pvt->dchr0);
1962 if (!dct_ganging_enabled(pvt)) {
1963 amd64_read_dct_pci_cfg(pvt, DCLR1, &pvt->dclr1);
1964 amd64_read_dct_pci_cfg(pvt, DCHR1, &pvt->dchr1);
1967 if (boot_cpu_data.x86 >= 0x10) {
1968 amd64_read_pci_cfg(pvt->F3, EXT_NB_MCA_CFG, &tmp);
1969 amd64_read_dct_pci_cfg(pvt, DBAM1, &pvt->dbam1);
1972 if (boot_cpu_data.x86 == 0x10 &&
1973 boot_cpu_data.x86_model > 7 &&
1974 /* F3x180[EccSymbolSize]=1 => x8 symbols */
1980 dump_misc_regs(pvt);
1984 * NOTE: CPU Revision Dependent code
1987 * @csrow_nr ChipSelect Row Number (0..NUM_CHIPSELECTS-1)
1988 * k8 private pointer to -->
1989 * DRAM Bank Address mapping register
1991 * DCL register where dual_channel_active is
1993 * The DBAM register consists of 4 sets of 4 bits each definitions:
1996 * 0-3 CSROWs 0 and 1
1997 * 4-7 CSROWs 2 and 3
1998 * 8-11 CSROWs 4 and 5
1999 * 12-15 CSROWs 6 and 7
2001 * Values range from: 0 to 15
2002 * The meaning of the values depends on CPU revision and dual-channel state,
2003 * see relevant BKDG more info.
2005 * The memory controller provides for total of only 8 CSROWs in its current
2006 * architecture. Each "pair" of CSROWs normally represents just one DIMM in
2007 * single channel or two (2) DIMMs in dual channel mode.
2009 * The following code logic collapses the various tables for CSROW based on CPU
2013 * The number of PAGE_SIZE pages on the specified CSROW number it
2017 static u32 amd64_csrow_nr_pages(int csrow_nr, struct amd64_pvt *pvt)
2019 u32 cs_mode, nr_pages;
2022 * The math on this doesn't look right on the surface because x/2*4 can
2023 * be simplified to x*2 but this expression makes use of the fact that
2024 * it is integral math where 1/2=0. This intermediate value becomes the
2025 * number of bits to shift the DBAM register to extract the proper CSROW
2028 cs_mode = (pvt->dbam0 >> ((csrow_nr / 2) * 4)) & 0xF;
2030 nr_pages = pvt->ops->dbam_to_cs(pvt, cs_mode) << (20 - PAGE_SHIFT);
2033 * If dual channel then double the memory size of single channel.
2034 * Channel count is 1 or 2
2036 nr_pages <<= (pvt->channel_count - 1);
2038 debugf0(" (csrow=%d) DBAM map index= %d\n", csrow_nr, cs_mode);
2039 debugf0(" nr_pages= %u channel-count = %d\n",
2040 nr_pages, pvt->channel_count);
2046 * Initialize the array of csrow attribute instances, based on the values
2047 * from pci config hardware registers.
2049 static int init_csrows(struct mem_ctl_info *mci)
2051 struct csrow_info *csrow;
2052 struct amd64_pvt *pvt = mci->pvt_info;
2053 u64 input_addr_min, input_addr_max, sys_addr, base, mask;
2057 amd64_read_pci_cfg(pvt->F3, NBCFG, &val);
2061 debugf0("node %d, NBCFG=0x%08x[ChipKillEccCap: %d|DramEccEn: %d]\n",
2062 pvt->mc_node_id, val,
2063 !!(val & NBCFG_CHIPKILL), !!(val & NBCFG_ECC_ENABLE));
2065 for_each_chip_select(i, 0, pvt) {
2066 csrow = &mci->csrows[i];
2068 if (!csrow_enabled(i, 0, pvt)) {
2069 debugf1("----CSROW %d EMPTY for node %d\n", i,
2074 debugf1("----CSROW %d VALID for MC node %d\n",
2075 i, pvt->mc_node_id);
2078 csrow->nr_pages = amd64_csrow_nr_pages(i, pvt);
2079 find_csrow_limits(mci, i, &input_addr_min, &input_addr_max);
2080 sys_addr = input_addr_to_sys_addr(mci, input_addr_min);
2081 csrow->first_page = (u32) (sys_addr >> PAGE_SHIFT);
2082 sys_addr = input_addr_to_sys_addr(mci, input_addr_max);
2083 csrow->last_page = (u32) (sys_addr >> PAGE_SHIFT);
2085 get_cs_base_and_mask(pvt, i, 0, &base, &mask);
2086 csrow->page_mask = ~mask;
2087 /* 8 bytes of resolution */
2089 csrow->mtype = amd64_determine_memory_type(pvt, i);
2091 debugf1(" for MC node %d csrow %d:\n", pvt->mc_node_id, i);
2092 debugf1(" input_addr_min: 0x%lx input_addr_max: 0x%lx\n",
2093 (unsigned long)input_addr_min,
2094 (unsigned long)input_addr_max);
2095 debugf1(" sys_addr: 0x%lx page_mask: 0x%lx\n",
2096 (unsigned long)sys_addr, csrow->page_mask);
2097 debugf1(" nr_pages: %u first_page: 0x%lx "
2098 "last_page: 0x%lx\n",
2099 (unsigned)csrow->nr_pages,
2100 csrow->first_page, csrow->last_page);
2103 * determine whether CHIPKILL or JUST ECC or NO ECC is operating
2105 if (pvt->nbcfg & NBCFG_ECC_ENABLE)
2107 (pvt->nbcfg & NBCFG_CHIPKILL) ?
2108 EDAC_S4ECD4ED : EDAC_SECDED;
2110 csrow->edac_mode = EDAC_NONE;
2116 /* get all cores on this DCT */
2117 static void get_cpus_on_this_dct_cpumask(struct cpumask *mask, int nid)
2121 for_each_online_cpu(cpu)
2122 if (amd_get_nb_id(cpu) == nid)
2123 cpumask_set_cpu(cpu, mask);
2126 /* check MCG_CTL on all the cpus on this node */
2127 static bool amd64_nb_mce_bank_enabled_on_node(int nid)
2133 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2134 amd64_warn("%s: Error allocating mask\n", __func__);
2138 get_cpus_on_this_dct_cpumask(mask, nid);
2140 rdmsr_on_cpus(mask, MSR_IA32_MCG_CTL, msrs);
2142 for_each_cpu(cpu, mask) {
2143 struct msr *reg = per_cpu_ptr(msrs, cpu);
2144 nbe = reg->l & MSR_MCGCTL_NBE;
2146 debugf0("core: %u, MCG_CTL: 0x%llx, NB MSR is %s\n",
2148 (nbe ? "enabled" : "disabled"));
2156 free_cpumask_var(mask);
2160 static int toggle_ecc_err_reporting(struct ecc_settings *s, u8 nid, bool on)
2162 cpumask_var_t cmask;
2165 if (!zalloc_cpumask_var(&cmask, GFP_KERNEL)) {
2166 amd64_warn("%s: error allocating mask\n", __func__);
2170 get_cpus_on_this_dct_cpumask(cmask, nid);
2172 rdmsr_on_cpus(cmask, MSR_IA32_MCG_CTL, msrs);
2174 for_each_cpu(cpu, cmask) {
2176 struct msr *reg = per_cpu_ptr(msrs, cpu);
2179 if (reg->l & MSR_MCGCTL_NBE)
2180 s->flags.nb_mce_enable = 1;
2182 reg->l |= MSR_MCGCTL_NBE;
2185 * Turn off NB MCE reporting only when it was off before
2187 if (!s->flags.nb_mce_enable)
2188 reg->l &= ~MSR_MCGCTL_NBE;
2191 wrmsr_on_cpus(cmask, MSR_IA32_MCG_CTL, msrs);
2193 free_cpumask_var(cmask);
2198 static bool enable_ecc_error_reporting(struct ecc_settings *s, u8 nid,
2202 u32 value, mask = 0x3; /* UECC/CECC enable */
2204 if (toggle_ecc_err_reporting(s, nid, ON)) {
2205 amd64_warn("Error enabling ECC reporting over MCGCTL!\n");
2209 amd64_read_pci_cfg(F3, NBCTL, &value);
2211 s->old_nbctl = value & mask;
2212 s->nbctl_valid = true;
2215 amd64_write_pci_cfg(F3, NBCTL, value);
2217 amd64_read_pci_cfg(F3, NBCFG, &value);
2219 debugf0("1: node %d, NBCFG=0x%08x[DramEccEn: %d]\n",
2220 nid, value, !!(value & NBCFG_ECC_ENABLE));
2222 if (!(value & NBCFG_ECC_ENABLE)) {
2223 amd64_warn("DRAM ECC disabled on this node, enabling...\n");
2225 s->flags.nb_ecc_prev = 0;
2227 /* Attempt to turn on DRAM ECC Enable */
2228 value |= NBCFG_ECC_ENABLE;
2229 amd64_write_pci_cfg(F3, NBCFG, value);
2231 amd64_read_pci_cfg(F3, NBCFG, &value);
2233 if (!(value & NBCFG_ECC_ENABLE)) {
2234 amd64_warn("Hardware rejected DRAM ECC enable,"
2235 "check memory DIMM configuration.\n");
2238 amd64_info("Hardware accepted DRAM ECC Enable\n");
2241 s->flags.nb_ecc_prev = 1;
2244 debugf0("2: node %d, NBCFG=0x%08x[DramEccEn: %d]\n",
2245 nid, value, !!(value & NBCFG_ECC_ENABLE));
2250 static void restore_ecc_error_reporting(struct ecc_settings *s, u8 nid,
2253 u32 value, mask = 0x3; /* UECC/CECC enable */
2256 if (!s->nbctl_valid)
2259 amd64_read_pci_cfg(F3, NBCTL, &value);
2261 value |= s->old_nbctl;
2263 amd64_write_pci_cfg(F3, NBCTL, value);
2265 /* restore previous BIOS DRAM ECC "off" setting we force-enabled */
2266 if (!s->flags.nb_ecc_prev) {
2267 amd64_read_pci_cfg(F3, NBCFG, &value);
2268 value &= ~NBCFG_ECC_ENABLE;
2269 amd64_write_pci_cfg(F3, NBCFG, value);
2272 /* restore the NB Enable MCGCTL bit */
2273 if (toggle_ecc_err_reporting(s, nid, OFF))
2274 amd64_warn("Error restoring NB MCGCTL settings!\n");
2278 * EDAC requires that the BIOS have ECC enabled before
2279 * taking over the processing of ECC errors. A command line
2280 * option allows to force-enable hardware ECC later in
2281 * enable_ecc_error_reporting().
2283 static const char *ecc_msg =
2284 "ECC disabled in the BIOS or no ECC capability, module will not load.\n"
2285 " Either enable ECC checking or force module loading by setting "
2286 "'ecc_enable_override'.\n"
2287 " (Note that use of the override may cause unknown side effects.)\n";
2289 static bool ecc_enabled(struct pci_dev *F3, u8 nid)
2293 bool nb_mce_en = false;
2295 amd64_read_pci_cfg(F3, NBCFG, &value);
2297 ecc_en = !!(value & NBCFG_ECC_ENABLE);
2298 amd64_info("DRAM ECC %s.\n", (ecc_en ? "enabled" : "disabled"));
2300 nb_mce_en = amd64_nb_mce_bank_enabled_on_node(nid);
2302 amd64_notice("NB MCE bank disabled, set MSR "
2303 "0x%08x[4] on node %d to enable.\n",
2304 MSR_IA32_MCG_CTL, nid);
2306 if (!ecc_en || !nb_mce_en) {
2307 amd64_notice("%s", ecc_msg);
2313 struct mcidev_sysfs_attribute sysfs_attrs[ARRAY_SIZE(amd64_dbg_attrs) +
2314 ARRAY_SIZE(amd64_inj_attrs) +
2317 struct mcidev_sysfs_attribute terminator = { .attr = { .name = NULL } };
2319 static void set_mc_sysfs_attrs(struct mem_ctl_info *mci)
2321 unsigned int i = 0, j = 0;
2323 for (; i < ARRAY_SIZE(amd64_dbg_attrs); i++)
2324 sysfs_attrs[i] = amd64_dbg_attrs[i];
2326 if (boot_cpu_data.x86 >= 0x10)
2327 for (j = 0; j < ARRAY_SIZE(amd64_inj_attrs); j++, i++)
2328 sysfs_attrs[i] = amd64_inj_attrs[j];
2330 sysfs_attrs[i] = terminator;
2332 mci->mc_driver_sysfs_attributes = sysfs_attrs;
2335 static void setup_mci_misc_attrs(struct mem_ctl_info *mci)
2337 struct amd64_pvt *pvt = mci->pvt_info;
2339 mci->mtype_cap = MEM_FLAG_DDR2 | MEM_FLAG_RDDR2;
2340 mci->edac_ctl_cap = EDAC_FLAG_NONE;
2342 if (pvt->nbcap & NBCAP_SECDED)
2343 mci->edac_ctl_cap |= EDAC_FLAG_SECDED;
2345 if (pvt->nbcap & NBCAP_CHIPKILL)
2346 mci->edac_ctl_cap |= EDAC_FLAG_S4ECD4ED;
2348 mci->edac_cap = amd64_determine_edac_cap(pvt);
2349 mci->mod_name = EDAC_MOD_STR;
2350 mci->mod_ver = EDAC_AMD64_VERSION;
2351 mci->ctl_name = pvt->ctl_name;
2352 mci->dev_name = pci_name(pvt->F2);
2353 mci->ctl_page_to_phys = NULL;
2355 /* memory scrubber interface */
2356 mci->set_sdram_scrub_rate = amd64_set_scrub_rate;
2357 mci->get_sdram_scrub_rate = amd64_get_scrub_rate;
2361 * returns a pointer to the family descriptor on success, NULL otherwise.
2363 static struct amd64_family_type *amd64_per_family_init(struct amd64_pvt *pvt)
2365 u8 fam = boot_cpu_data.x86;
2366 struct amd64_family_type *fam_type = NULL;
2370 fam_type = &amd64_family_types[K8_CPUS];
2371 pvt->ops = &amd64_family_types[K8_CPUS].ops;
2372 pvt->ctl_name = fam_type->ctl_name;
2373 pvt->min_scrubrate = K8_MIN_SCRUB_RATE_BITS;
2376 fam_type = &amd64_family_types[F10_CPUS];
2377 pvt->ops = &amd64_family_types[F10_CPUS].ops;
2378 pvt->ctl_name = fam_type->ctl_name;
2379 pvt->min_scrubrate = F10_MIN_SCRUB_RATE_BITS;
2383 amd64_err("Unsupported family!\n");
2387 pvt->ext_model = boot_cpu_data.x86_model >> 4;
2389 amd64_info("%s %sdetected (node %d).\n", pvt->ctl_name,
2391 (pvt->ext_model >= K8_REV_F ? "revF or later "
2392 : "revE or earlier ")
2393 : ""), pvt->mc_node_id);
2397 static int amd64_init_one_instance(struct pci_dev *F2)
2399 struct amd64_pvt *pvt = NULL;
2400 struct amd64_family_type *fam_type = NULL;
2401 struct mem_ctl_info *mci = NULL;
2403 u8 nid = get_node_id(F2);
2406 pvt = kzalloc(sizeof(struct amd64_pvt), GFP_KERNEL);
2410 pvt->mc_node_id = nid;
2414 fam_type = amd64_per_family_init(pvt);
2419 err = reserve_mc_sibling_devs(pvt, fam_type->f1_id, fam_type->f3_id);
2426 * We need to determine how many memory channels there are. Then use
2427 * that information for calculating the size of the dynamic instance
2428 * tables in the 'mci' structure.
2431 pvt->channel_count = pvt->ops->early_channel_count(pvt);
2432 if (pvt->channel_count < 0)
2436 mci = edac_mc_alloc(0, pvt->csels[0].b_cnt, pvt->channel_count, nid);
2440 mci->pvt_info = pvt;
2441 mci->dev = &pvt->F2->dev;
2443 setup_mci_misc_attrs(mci);
2445 if (init_csrows(mci))
2446 mci->edac_cap = EDAC_FLAG_NONE;
2448 set_mc_sysfs_attrs(mci);
2451 if (edac_mc_add_mc(mci)) {
2452 debugf1("failed edac_mc_add_mc()\n");
2456 /* register stuff with EDAC MCE */
2457 if (report_gart_errors)
2458 amd_report_gart_errors(true);
2460 amd_register_ecc_decoder(amd64_decode_bus_error);
2464 atomic_inc(&drv_instances);
2472 free_mc_sibling_devs(pvt);
2481 static int __devinit amd64_probe_one_instance(struct pci_dev *pdev,
2482 const struct pci_device_id *mc_type)
2484 u8 nid = get_node_id(pdev);
2485 struct pci_dev *F3 = node_to_amd_nb(nid)->misc;
2486 struct ecc_settings *s;
2489 ret = pci_enable_device(pdev);
2491 debugf0("ret=%d\n", ret);
2496 s = kzalloc(sizeof(struct ecc_settings), GFP_KERNEL);
2502 if (!ecc_enabled(F3, nid)) {
2505 if (!ecc_enable_override)
2508 amd64_warn("Forcing ECC on!\n");
2510 if (!enable_ecc_error_reporting(s, nid, F3))
2514 ret = amd64_init_one_instance(pdev);
2516 amd64_err("Error probing instance: %d\n", nid);
2517 restore_ecc_error_reporting(s, nid, F3);
2524 ecc_stngs[nid] = NULL;
2530 static void __devexit amd64_remove_one_instance(struct pci_dev *pdev)
2532 struct mem_ctl_info *mci;
2533 struct amd64_pvt *pvt;
2534 u8 nid = get_node_id(pdev);
2535 struct pci_dev *F3 = node_to_amd_nb(nid)->misc;
2536 struct ecc_settings *s = ecc_stngs[nid];
2538 /* Remove from EDAC CORE tracking list */
2539 mci = edac_mc_del_mc(&pdev->dev);
2543 pvt = mci->pvt_info;
2545 restore_ecc_error_reporting(s, nid, F3);
2547 free_mc_sibling_devs(pvt);
2549 /* unregister from EDAC MCE */
2550 amd_report_gart_errors(false);
2551 amd_unregister_ecc_decoder(amd64_decode_bus_error);
2553 kfree(ecc_stngs[nid]);
2554 ecc_stngs[nid] = NULL;
2556 /* Free the EDAC CORE resources */
2557 mci->pvt_info = NULL;
2565 * This table is part of the interface for loading drivers for PCI devices. The
2566 * PCI core identifies what devices are on a system during boot, and then
2567 * inquiry this table to see if this driver is for a given device found.
2569 static const struct pci_device_id amd64_pci_table[] __devinitdata = {
2571 .vendor = PCI_VENDOR_ID_AMD,
2572 .device = PCI_DEVICE_ID_AMD_K8_NB_MEMCTL,
2573 .subvendor = PCI_ANY_ID,
2574 .subdevice = PCI_ANY_ID,
2579 .vendor = PCI_VENDOR_ID_AMD,
2580 .device = PCI_DEVICE_ID_AMD_10H_NB_DRAM,
2581 .subvendor = PCI_ANY_ID,
2582 .subdevice = PCI_ANY_ID,
2588 MODULE_DEVICE_TABLE(pci, amd64_pci_table);
2590 static struct pci_driver amd64_pci_driver = {
2591 .name = EDAC_MOD_STR,
2592 .probe = amd64_probe_one_instance,
2593 .remove = __devexit_p(amd64_remove_one_instance),
2594 .id_table = amd64_pci_table,
2597 static void setup_pci_device(void)
2599 struct mem_ctl_info *mci;
2600 struct amd64_pvt *pvt;
2608 pvt = mci->pvt_info;
2610 edac_pci_create_generic_ctl(&pvt->F2->dev, EDAC_MOD_STR);
2612 if (!amd64_ctl_pci) {
2613 pr_warning("%s(): Unable to create PCI control\n",
2616 pr_warning("%s(): PCI error report via EDAC not set\n",
2622 static int __init amd64_edac_init(void)
2626 edac_printk(KERN_INFO, EDAC_MOD_STR, EDAC_AMD64_VERSION "\n");
2630 if (amd_cache_northbridges() < 0)
2634 mcis = kzalloc(amd_nb_num() * sizeof(mcis[0]), GFP_KERNEL);
2635 ecc_stngs = kzalloc(amd_nb_num() * sizeof(ecc_stngs[0]), GFP_KERNEL);
2636 if (!(mcis && ecc_stngs))
2639 msrs = msrs_alloc();
2643 err = pci_register_driver(&amd64_pci_driver);
2648 if (!atomic_read(&drv_instances))
2649 goto err_no_instances;
2655 pci_unregister_driver(&amd64_pci_driver);
2672 static void __exit amd64_edac_exit(void)
2675 edac_pci_release_generic_ctl(amd64_ctl_pci);
2677 pci_unregister_driver(&amd64_pci_driver);
2689 module_init(amd64_edac_init);
2690 module_exit(amd64_edac_exit);
2692 MODULE_LICENSE("GPL");
2693 MODULE_AUTHOR("SoftwareBitMaker: Doug Thompson, "
2694 "Dave Peterson, Thayne Harbaugh");
2695 MODULE_DESCRIPTION("MC support for AMD64 memory controllers - "
2696 EDAC_AMD64_VERSION);
2698 module_param(edac_op_state, int, 0444);
2699 MODULE_PARM_DESC(edac_op_state, "EDAC Error Reporting state: 0=Poll,1=NMI");