From d0cdf9003140e9b40d2488aaee2838babe7e212c Mon Sep 17 00:00:00 2001 From: Jim Snow Date: Thu, 3 Dec 2015 10:48:54 +0100 Subject: [PATCH] EDAC, sb_edac: Add Knights Landing (Xeon Phi gen 2) support Knights Landing is the next generation architecture for HPC market. KNL introduces concept of a tile and CHA - Cache/Home Agent for memory accesses. Some things are fixed in KNL: () There's single DIMM slot per channel () There's 2 memory controllers with 3 channels each, however, from EDAC standpoint, it is presented as single memory controller with 6 channels. In order to represent 2 MCs w/ 3 CH, it would require major redesign of EDAC core driver. Basically, two functionalities are added/extended: () during driver initialization KNL topology is being recognized, i.e. which channels are populated with what DIMM sizes (knl_get_dimm_capacity function) () handle MCE errors - channel swizzling Reviewed-by: Tony Luck Signed-off-by: Jim Snow Cc: Mauro Carvalho Chehab Cc: linux-edac Cc: lukasz.anaczkowski@intel.com Link: http://lkml.kernel.org/r/1449136134-23706-5-git-send-email-hubert.chrzaniuk@intel.com [ Rebase to 4.4-rc3. ] Signed-off-by: Hubert Chrzaniuk Signed-off-by: Borislav Petkov --- drivers/edac/sb_edac.c | 966 +++++++++++++++++++++++++++++++++++++++-- 1 file changed, 921 insertions(+), 45 deletions(-) diff --git a/drivers/edac/sb_edac.c b/drivers/edac/sb_edac.c index c8fbde2bd20a..b3d924da5985 100644 --- a/drivers/edac/sb_edac.c +++ b/drivers/edac/sb_edac.c @@ -65,6 +65,14 @@ static const u32 ibridge_dram_rule[] = { 0xd8, 0xe0, 0xe8, 0xf0, 0xf8, }; +static const u32 knl_dram_rule[] = { + 0x60, 0x68, 0x70, 0x78, 0x80, /* 0-4 */ + 0x88, 0x90, 0x98, 0xa0, 0xa8, /* 5-9 */ + 0xb0, 0xb8, 0xc0, 0xc8, 0xd0, /* 10-14 */ + 0xd8, 0xe0, 0xe8, 0xf0, 0xf8, /* 15-19 */ + 0x100, 0x108, 0x110, 0x118, /* 20-23 */ +}; + #define DRAM_RULE_ENABLE(reg) GET_BITFIELD(reg, 0, 0) #define A7MODE(reg) GET_BITFIELD(reg, 26, 26) @@ -94,6 +102,14 @@ static const u32 ibridge_interleave_list[] = { 0xdc, 0xe4, 0xec, 0xf4, 0xfc, }; +static const u32 knl_interleave_list[] = { + 0x64, 0x6c, 0x74, 0x7c, 0x84, /* 0-4 */ + 0x8c, 0x94, 0x9c, 0xa4, 0xac, /* 5-9 */ + 0xb4, 0xbc, 0xc4, 0xcc, 0xd4, /* 10-14 */ + 0xdc, 0xe4, 0xec, 0xf4, 0xfc, /* 15-19 */ + 0x104, 0x10c, 0x114, 0x11c, /* 20-23 */ +}; + struct interleave_pkg { unsigned char start; unsigned char end; @@ -131,10 +147,13 @@ static inline int sad_pkg(const struct interleave_pkg *table, u32 reg, /* Devices 12 Function 7 */ #define TOLM 0x80 -#define TOHM 0x84 +#define TOHM 0x84 #define HASWELL_TOLM 0xd0 #define HASWELL_TOHM_0 0xd4 #define HASWELL_TOHM_1 0xd8 +#define KNL_TOLM 0xd0 +#define KNL_TOHM_0 0xd4 +#define KNL_TOHM_1 0xd8 #define GET_TOLM(reg) ((GET_BITFIELD(reg, 0, 3) << 28) | 0x3ffffff) #define GET_TOHM(reg) ((GET_BITFIELD(reg, 0, 20) << 25) | 0x3ffffff) @@ -145,6 +164,8 @@ static inline int sad_pkg(const struct interleave_pkg *table, u32 reg, #define SOURCE_ID(reg) GET_BITFIELD(reg, 9, 11) +#define SOURCE_ID_KNL(reg) GET_BITFIELD(reg, 12, 14) + #define SAD_CONTROL 0xf4 /* Device 14 function 0 */ @@ -167,6 +188,7 @@ static const u32 tad_dram_rule[] = { /* Device 15, function 0 */ #define MCMTR 0x7c +#define KNL_MCMTR 0x624 #define IS_ECC_ENABLED(mcmtr) GET_BITFIELD(mcmtr, 2, 2) #define IS_LOCKSTEP_ENABLED(mcmtr) GET_BITFIELD(mcmtr, 1, 1) @@ -183,6 +205,8 @@ static const int mtr_regs[] = { 0x80, 0x84, 0x88, }; +static const int knl_mtr_reg = 0xb60; + #define RANK_DISABLE(mtr) GET_BITFIELD(mtr, 16, 19) #define IS_DIMM_PRESENT(mtr) GET_BITFIELD(mtr, 14, 14) #define RANK_CNT_BITS(mtr) GET_BITFIELD(mtr, 12, 13) @@ -253,6 +277,9 @@ static const u32 correrrthrsld[] = { #define NUM_CHANNELS 8 /* 2MC per socket, four chan per MC */ #define MAX_DIMMS 3 /* Max DIMMS per channel */ +#define KNL_MAX_CHAS 38 /* KNL max num. of Cache Home Agents */ +#define KNL_MAX_CHANNELS 6 /* KNL max num. of PCI channels */ +#define KNL_MAX_EDCS 8 /* Embedded DRAM controllers */ #define CHANNEL_UNSPECIFIED 0xf /* Intel IA32 SDM 15-14 */ enum type { @@ -260,6 +287,7 @@ enum type { IVY_BRIDGE, HASWELL, BROADWELL, + KNIGHTS_LANDING, }; struct sbridge_pvt; @@ -309,6 +337,16 @@ struct sbridge_dev { struct mem_ctl_info *mci; }; +struct knl_pvt { + struct pci_dev *pci_cha[KNL_MAX_CHAS]; + struct pci_dev *pci_channel[KNL_MAX_CHANNELS]; + struct pci_dev *pci_mc0; + struct pci_dev *pci_mc1; + struct pci_dev *pci_mc0_misc; + struct pci_dev *pci_mc1_misc; + struct pci_dev *pci_mc_info; /* tolm, tohm */ +}; + struct sbridge_pvt { struct pci_dev *pci_ta, *pci_ddrio, *pci_ras; struct pci_dev *pci_sad0, *pci_sad1; @@ -337,6 +375,7 @@ struct sbridge_pvt { /* Memory description */ u64 tolm, tohm; + struct knl_pvt knl; }; #define PCI_DESCR(device_id, opt) \ @@ -510,6 +549,50 @@ static const struct pci_id_table pci_dev_descr_haswell_table[] = { {0,} /* 0 terminated list. */ }; +/* Knight's Landing Support */ +/* + * KNL's memory channels are swizzled between memory controllers. + * MC0 is mapped to CH3,5,6 and MC1 is mapped to CH0,1,2 + */ +#define knl_channel_remap(channel) ((channel + 3) % 6) + +/* Memory controller, TAD tables, error injection - 2-8-0, 2-9-0 (2 of these) */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_MC 0x7840 +/* DRAM channel stuff; bank addrs, dimmmtr, etc.. 2-8-2 - 2-9-4 (6 of these) */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_CHANNEL 0x7843 +/* kdrwdbu TAD limits/offsets, MCMTR - 2-10-1, 2-11-1 (2 of these) */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_TA 0x7844 +/* CHA broadcast registers, dram rules - 1-29-0 (1 of these) */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_SAD0 0x782a +/* SAD target - 1-29-1 (1 of these) */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_SAD1 0x782b +/* Caching / Home Agent */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_CHA 0x782c +/* Device with TOLM and TOHM, 0-5-0 (1 of these) */ +#define PCI_DEVICE_ID_INTEL_KNL_IMC_TOLHM 0x7810 + +/* + * KNL differs from SB, IB, and Haswell in that it has multiple + * instances of the same device with the same device ID, so we handle that + * by creating as many copies in the table as we expect to find. + * (Like device ID must be grouped together.) + */ + +static const struct pci_id_descr pci_dev_descr_knl[] = { + [0] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_SAD0, 0) }, + [1] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_SAD1, 0) }, + [2 ... 3] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_MC, 0)}, + [4 ... 41] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_CHA, 0) }, + [42 ... 47] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_CHANNEL, 0) }, + [48] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_TA, 0) }, + [49] = { PCI_DESCR(PCI_DEVICE_ID_INTEL_KNL_IMC_TOLHM, 0) }, +}; + +static const struct pci_id_table pci_dev_descr_knl_table[] = { + PCI_ID_TABLE_ENTRY(pci_dev_descr_knl), + {0,} +}; + /* * Broadwell support * @@ -586,6 +669,7 @@ static const struct pci_device_id sbridge_pci_tbl[] = { {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IBRIDGE_IMC_HA0_TA)}, {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_HASWELL_IMC_HA0)}, {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_BROADWELL_IMC_HA0)}, + {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_KNL_IMC_SAD0)}, {0,} /* 0 terminated list. */ }; @@ -599,7 +683,7 @@ static inline int numrank(enum type type, u32 mtr) int ranks = (1 << RANK_CNT_BITS(mtr)); int max = 4; - if (type == HASWELL || type == BROADWELL) + if (type == HASWELL || type == BROADWELL || type == KNIGHTS_LANDING) max = 8; if (ranks > max) { @@ -748,6 +832,47 @@ static u32 dram_attr(u32 reg) return GET_BITFIELD(reg, 2, 3); } +static u64 knl_sad_limit(u32 reg) +{ + return (GET_BITFIELD(reg, 7, 26) << 26) | 0x3ffffff; +} + +static u32 knl_interleave_mode(u32 reg) +{ + return GET_BITFIELD(reg, 1, 2); +} + +static char *knl_show_interleave_mode(u32 reg) +{ + char *s; + + switch (knl_interleave_mode(reg)) { + case 0: + s = "use address bits [8:6]"; + break; + case 1: + s = "use address bits [10:8]"; + break; + case 2: + s = "use address bits [14:12]"; + break; + case 3: + s = "use address bits [32:30]"; + break; + default: + WARN_ON(1); + break; + } + + return s; +} + +static u32 dram_attr_knl(u32 reg) +{ + return GET_BITFIELD(reg, 3, 4); +} + + static enum mem_type get_memory_type(struct sbridge_pvt *pvt) { u32 reg; @@ -842,6 +967,12 @@ static enum dev_type broadwell_get_width(struct sbridge_pvt *pvt, u32 mtr) return __ibridge_get_width(GET_BITFIELD(mtr, 8, 9)); } +static enum mem_type knl_get_memory_type(struct sbridge_pvt *pvt) +{ + /* DDR4 RDIMMS and LRDIMMS are supported */ + return MEM_RDDR4; +} + static u8 get_node_id(struct sbridge_pvt *pvt) { u32 reg; @@ -857,6 +988,15 @@ static u8 haswell_get_node_id(struct sbridge_pvt *pvt) return GET_BITFIELD(reg, 0, 3); } +static u8 knl_get_node_id(struct sbridge_pvt *pvt) +{ + u32 reg; + + pci_read_config_dword(pvt->pci_sad1, SAD_CONTROL, ®); + return GET_BITFIELD(reg, 0, 2); +} + + static u64 haswell_get_tolm(struct sbridge_pvt *pvt) { u32 reg; @@ -878,6 +1018,26 @@ static u64 haswell_get_tohm(struct sbridge_pvt *pvt) return rc | 0x1ffffff; } +static u64 knl_get_tolm(struct sbridge_pvt *pvt) +{ + u32 reg; + + pci_read_config_dword(pvt->knl.pci_mc_info, KNL_TOLM, ®); + return (GET_BITFIELD(reg, 26, 31) << 26) | 0x3ffffff; +} + +static u64 knl_get_tohm(struct sbridge_pvt *pvt) +{ + u64 rc; + u32 reg_lo, reg_hi; + + pci_read_config_dword(pvt->knl.pci_mc_info, KNL_TOHM_0, ®_lo); + pci_read_config_dword(pvt->knl.pci_mc_info, KNL_TOHM_1, ®_hi); + rc = ((u64)reg_hi << 32) | reg_lo; + return rc | 0x3ffffff; +} + + static u64 haswell_rir_limit(u32 reg) { return (((u64)GET_BITFIELD(reg, 1, 11) + 1) << 29) - 1; @@ -935,11 +1095,22 @@ static int check_if_ecc_is_active(const u8 bus, enum type type) case BROADWELL: id = PCI_DEVICE_ID_INTEL_BROADWELL_IMC_HA0_TA; break; + case KNIGHTS_LANDING: + /* + * KNL doesn't group things by bus the same way + * SB/IB/Haswell does. + */ + id = PCI_DEVICE_ID_INTEL_KNL_IMC_TA; + break; default: return -ENODEV; } - pdev = get_pdev_same_bus(bus, id); + if (type != KNIGHTS_LANDING) + pdev = get_pdev_same_bus(bus, id); + else + pdev = pci_get_device(PCI_VENDOR_ID_INTEL, id, 0); + if (!pdev) { sbridge_printk(KERN_ERR, "Couldn't find PCI device " "%04x:%04x! on bus %02d\n", @@ -947,7 +1118,8 @@ static int check_if_ecc_is_active(const u8 bus, enum type type) return -ENODEV; } - pci_read_config_dword(pdev, MCMTR, &mcmtr); + pci_read_config_dword(pdev, + type == KNIGHTS_LANDING ? KNL_MCMTR : MCMTR, &mcmtr); if (!IS_ECC_ENABLED(mcmtr)) { sbridge_printk(KERN_ERR, "ECC is disabled. Aborting\n"); return -ENODEV; @@ -955,6 +1127,476 @@ static int check_if_ecc_is_active(const u8 bus, enum type type) return 0; } +/* Low bits of TAD limit, and some metadata. */ +static const u32 knl_tad_dram_limit_lo[] = { + 0x400, 0x500, 0x600, 0x700, + 0x800, 0x900, 0xa00, 0xb00, +}; + +/* Low bits of TAD offset. */ +static const u32 knl_tad_dram_offset_lo[] = { + 0x404, 0x504, 0x604, 0x704, + 0x804, 0x904, 0xa04, 0xb04, +}; + +/* High 16 bits of TAD limit and offset. */ +static const u32 knl_tad_dram_hi[] = { + 0x408, 0x508, 0x608, 0x708, + 0x808, 0x908, 0xa08, 0xb08, +}; + +/* Number of ways a tad entry is interleaved. */ +static const u32 knl_tad_ways[] = { + 8, 6, 4, 3, 2, 1, +}; + +/* + * Retrieve the n'th Target Address Decode table entry + * from the memory controller's TAD table. + * + * @pvt: driver private data + * @entry: which entry you want to retrieve + * @mc: which memory controller (0 or 1) + * @offset: output tad range offset + * @limit: output address of first byte above tad range + * @ways: output number of interleave ways + * + * The offset value has curious semantics. It's a sort of running total + * of the sizes of all the memory regions that aren't mapped in this + * tad table. + */ +static int knl_get_tad(const struct sbridge_pvt *pvt, + const int entry, + const int mc, + u64 *offset, + u64 *limit, + int *ways) +{ + u32 reg_limit_lo, reg_offset_lo, reg_hi; + struct pci_dev *pci_mc; + int way_id; + + switch (mc) { + case 0: + pci_mc = pvt->knl.pci_mc0; + break; + case 1: + pci_mc = pvt->knl.pci_mc1; + break; + default: + WARN_ON(1); + return -EINVAL; + } + + pci_read_config_dword(pci_mc, + knl_tad_dram_limit_lo[entry], ®_limit_lo); + pci_read_config_dword(pci_mc, + knl_tad_dram_offset_lo[entry], ®_offset_lo); + pci_read_config_dword(pci_mc, + knl_tad_dram_hi[entry], ®_hi); + + /* Is this TAD entry enabled? */ + if (!GET_BITFIELD(reg_limit_lo, 0, 0)) + return -ENODEV; + + way_id = GET_BITFIELD(reg_limit_lo, 3, 5); + + if (way_id < ARRAY_SIZE(knl_tad_ways)) { + *ways = knl_tad_ways[way_id]; + } else { + *ways = 0; + sbridge_printk(KERN_ERR, + "Unexpected value %d in mc_tad_limit_lo wayness field\n", + way_id); + return -ENODEV; + } + + /* + * The least significant 6 bits of base and limit are truncated. + * For limit, we fill the missing bits with 1s. + */ + *offset = ((u64) GET_BITFIELD(reg_offset_lo, 6, 31) << 6) | + ((u64) GET_BITFIELD(reg_hi, 0, 15) << 32); + *limit = ((u64) GET_BITFIELD(reg_limit_lo, 6, 31) << 6) | 63 | + ((u64) GET_BITFIELD(reg_hi, 16, 31) << 32); + + return 0; +} + +/* Determine which memory controller is responsible for a given channel. */ +static int knl_channel_mc(int channel) +{ + WARN_ON(channel < 0 || channel >= 6); + + return channel < 3 ? 1 : 0; +} + +/* + * Get the Nth entry from EDC_ROUTE_TABLE register. + * (This is the per-tile mapping of logical interleave targets to + * physical EDC modules.) + * + * entry 0: 0:2 + * 1: 3:5 + * 2: 6:8 + * 3: 9:11 + * 4: 12:14 + * 5: 15:17 + * 6: 18:20 + * 7: 21:23 + * reserved: 24:31 + */ +static u32 knl_get_edc_route(int entry, u32 reg) +{ + WARN_ON(entry >= KNL_MAX_EDCS); + return GET_BITFIELD(reg, entry*3, (entry*3)+2); +} + +/* + * Get the Nth entry from MC_ROUTE_TABLE register. + * (This is the per-tile mapping of logical interleave targets to + * physical DRAM channels modules.) + * + * entry 0: mc 0:2 channel 18:19 + * 1: mc 3:5 channel 20:21 + * 2: mc 6:8 channel 22:23 + * 3: mc 9:11 channel 24:25 + * 4: mc 12:14 channel 26:27 + * 5: mc 15:17 channel 28:29 + * reserved: 30:31 + * + * Though we have 3 bits to identify the MC, we should only see + * the values 0 or 1. + */ + +static u32 knl_get_mc_route(int entry, u32 reg) +{ + int mc, chan; + + WARN_ON(entry >= KNL_MAX_CHANNELS); + + mc = GET_BITFIELD(reg, entry*3, (entry*3)+2); + chan = GET_BITFIELD(reg, (entry*2) + 18, (entry*2) + 18 + 1); + + return knl_channel_remap(mc*3 + chan); +} + +/* + * Render the EDC_ROUTE register in human-readable form. + * Output string s should be at least KNL_MAX_EDCS*2 bytes. + */ +static void knl_show_edc_route(u32 reg, char *s) +{ + int i; + + for (i = 0; i < KNL_MAX_EDCS; i++) { + s[i*2] = knl_get_edc_route(i, reg) + '0'; + s[i*2+1] = '-'; + } + + s[KNL_MAX_EDCS*2 - 1] = '\0'; +} + +/* + * Render the MC_ROUTE register in human-readable form. + * Output string s should be at least KNL_MAX_CHANNELS*2 bytes. + */ +static void knl_show_mc_route(u32 reg, char *s) +{ + int i; + + for (i = 0; i < KNL_MAX_CHANNELS; i++) { + s[i*2] = knl_get_mc_route(i, reg) + '0'; + s[i*2+1] = '-'; + } + + s[KNL_MAX_CHANNELS*2 - 1] = '\0'; +} + +#define KNL_EDC_ROUTE 0xb8 +#define KNL_MC_ROUTE 0xb4 + +/* Is this dram rule backed by regular DRAM in flat mode? */ +#define KNL_EDRAM(reg) GET_BITFIELD(reg, 29, 29) + +/* Is this dram rule cached? */ +#define KNL_CACHEABLE(reg) GET_BITFIELD(reg, 28, 28) + +/* Is this rule backed by edc ? */ +#define KNL_EDRAM_ONLY(reg) GET_BITFIELD(reg, 29, 29) + +/* Is this rule backed by DRAM, cacheable in EDRAM? */ +#define KNL_CACHEABLE(reg) GET_BITFIELD(reg, 28, 28) + +/* Is this rule mod3? */ +#define KNL_MOD3(reg) GET_BITFIELD(reg, 27, 27) + +/* + * Figure out how big our RAM modules are. + * + * The DIMMMTR register in KNL doesn't tell us the size of the DIMMs, so we + * have to figure this out from the SAD rules, interleave lists, route tables, + * and TAD rules. + * + * SAD rules can have holes in them (e.g. the 3G-4G hole), so we have to + * inspect the TAD rules to figure out how large the SAD regions really are. + * + * When we know the real size of a SAD region and how many ways it's + * interleaved, we know the individual contribution of each channel to + * TAD is size/ways. + * + * Finally, we have to check whether each channel participates in each SAD + * region. + * + * Fortunately, KNL only supports one DIMM per channel, so once we know how + * much memory the channel uses, we know the DIMM is at least that large. + * (The BIOS might possibly choose not to map all available memory, in which + * case we will underreport the size of the DIMM.) + * + * In theory, we could try to determine the EDC sizes as well, but that would + * only work in flat mode, not in cache mode. + * + * @mc_sizes: Output sizes of channels (must have space for KNL_MAX_CHANNELS + * elements) + */ +static int knl_get_dimm_capacity(struct sbridge_pvt *pvt, u64 *mc_sizes) +{ + u64 sad_base, sad_size, sad_limit = 0; + u64 tad_base, tad_size, tad_limit, tad_deadspace, tad_livespace; + int sad_rule = 0; + int tad_rule = 0; + int intrlv_ways, tad_ways; + u32 first_pkg, pkg; + int i; + u64 sad_actual_size[2]; /* sad size accounting for holes, per mc */ + u32 dram_rule, interleave_reg; + u32 mc_route_reg[KNL_MAX_CHAS]; + u32 edc_route_reg[KNL_MAX_CHAS]; + int edram_only; + char edc_route_string[KNL_MAX_EDCS*2]; + char mc_route_string[KNL_MAX_CHANNELS*2]; + int cur_reg_start; + int mc; + int channel; + int way; + int participants[KNL_MAX_CHANNELS]; + int participant_count = 0; + + for (i = 0; i < KNL_MAX_CHANNELS; i++) + mc_sizes[i] = 0; + + /* Read the EDC route table in each CHA. */ + cur_reg_start = 0; + for (i = 0; i < KNL_MAX_CHAS; i++) { + pci_read_config_dword(pvt->knl.pci_cha[i], + KNL_EDC_ROUTE, &edc_route_reg[i]); + + if (i > 0 && edc_route_reg[i] != edc_route_reg[i-1]) { + knl_show_edc_route(edc_route_reg[i-1], + edc_route_string); + if (cur_reg_start == i-1) + edac_dbg(0, "edc route table for CHA %d: %s\n", + cur_reg_start, edc_route_string); + else + edac_dbg(0, "edc route table for CHA %d-%d: %s\n", + cur_reg_start, i-1, edc_route_string); + cur_reg_start = i; + } + } + knl_show_edc_route(edc_route_reg[i-1], edc_route_string); + if (cur_reg_start == i-1) + edac_dbg(0, "edc route table for CHA %d: %s\n", + cur_reg_start, edc_route_string); + else + edac_dbg(0, "edc route table for CHA %d-%d: %s\n", + cur_reg_start, i-1, edc_route_string); + + /* Read the MC route table in each CHA. */ + cur_reg_start = 0; + for (i = 0; i < KNL_MAX_CHAS; i++) { + pci_read_config_dword(pvt->knl.pci_cha[i], + KNL_MC_ROUTE, &mc_route_reg[i]); + + if (i > 0 && mc_route_reg[i] != mc_route_reg[i-1]) { + knl_show_mc_route(mc_route_reg[i-1], mc_route_string); + if (cur_reg_start == i-1) + edac_dbg(0, "mc route table for CHA %d: %s\n", + cur_reg_start, mc_route_string); + else + edac_dbg(0, "mc route table for CHA %d-%d: %s\n", + cur_reg_start, i-1, mc_route_string); + cur_reg_start = i; + } + } + knl_show_mc_route(mc_route_reg[i-1], mc_route_string); + if (cur_reg_start == i-1) + edac_dbg(0, "mc route table for CHA %d: %s\n", + cur_reg_start, mc_route_string); + else + edac_dbg(0, "mc route table for CHA %d-%d: %s\n", + cur_reg_start, i-1, mc_route_string); + + /* Process DRAM rules */ + for (sad_rule = 0; sad_rule < pvt->info.max_sad; sad_rule++) { + /* previous limit becomes the new base */ + sad_base = sad_limit; + + pci_read_config_dword(pvt->pci_sad0, + pvt->info.dram_rule[sad_rule], &dram_rule); + + if (!DRAM_RULE_ENABLE(dram_rule)) + break; + + edram_only = KNL_EDRAM_ONLY(dram_rule); + + sad_limit = pvt->info.sad_limit(dram_rule)+1; + sad_size = sad_limit - sad_base; + + pci_read_config_dword(pvt->pci_sad0, + pvt->info.interleave_list[sad_rule], &interleave_reg); + + /* + * Find out how many ways this dram rule is interleaved. + * We stop when we see the first channel again. + */ + first_pkg = sad_pkg(pvt->info.interleave_pkg, + interleave_reg, 0); + for (intrlv_ways = 1; intrlv_ways < 8; intrlv_ways++) { + pkg = sad_pkg(pvt->info.interleave_pkg, + interleave_reg, intrlv_ways); + + if ((pkg & 0x8) == 0) { + /* + * 0 bit means memory is non-local, + * which KNL doesn't support + */ + edac_dbg(0, "Unexpected interleave target %d\n", + pkg); + return -1; + } + + if (pkg == first_pkg) + break; + } + if (KNL_MOD3(dram_rule)) + intrlv_ways *= 3; + + edac_dbg(3, "dram rule %d (base 0x%llx, limit 0x%llx), %d way interleave%s\n", + sad_rule, + sad_base, + sad_limit, + intrlv_ways, + edram_only ? ", EDRAM" : ""); + + /* + * Find out how big the SAD region really is by iterating + * over TAD tables (SAD regions may contain holes). + * Each memory controller might have a different TAD table, so + * we have to look at both. + * + * Livespace is the memory that's mapped in this TAD table, + * deadspace is the holes (this could be the MMIO hole, or it + * could be memory that's mapped by the other TAD table but + * not this one). + */ + for (mc = 0; mc < 2; mc++) { + sad_actual_size[mc] = 0; + tad_livespace = 0; + for (tad_rule = 0; + tad_rule < ARRAY_SIZE( + knl_tad_dram_limit_lo); + tad_rule++) { + if (knl_get_tad(pvt, + tad_rule, + mc, + &tad_deadspace, + &tad_limit, + &tad_ways)) + break; + + tad_size = (tad_limit+1) - + (tad_livespace + tad_deadspace); + tad_livespace += tad_size; + tad_base = (tad_limit+1) - tad_size; + + if (tad_base < sad_base) { + if (tad_limit > sad_base) + edac_dbg(0, "TAD region overlaps lower SAD boundary -- TAD tables may be configured incorrectly.\n"); + } else if (tad_base < sad_limit) { + if (tad_limit+1 > sad_limit) { + edac_dbg(0, "TAD region overlaps upper SAD boundary -- TAD tables may be configured incorrectly.\n"); + } else { + /* TAD region is completely inside SAD region */ + edac_dbg(3, "TAD region %d 0x%llx - 0x%llx (%lld bytes) table%d\n", + tad_rule, tad_base, + tad_limit, tad_size, + mc); + sad_actual_size[mc] += tad_size; + } + } + tad_base = tad_limit+1; + } + } + + for (mc = 0; mc < 2; mc++) { + edac_dbg(3, " total TAD DRAM footprint in table%d : 0x%llx (%lld bytes)\n", + mc, sad_actual_size[mc], sad_actual_size[mc]); + } + + /* Ignore EDRAM rule */ + if (edram_only) + continue; + + /* Figure out which channels participate in interleave. */ + for (channel = 0; channel < KNL_MAX_CHANNELS; channel++) + participants[channel] = 0; + + /* For each channel, does at least one CHA have + * this channel mapped to the given target? + */ + for (channel = 0; channel < KNL_MAX_CHANNELS; channel++) { + for (way = 0; way < intrlv_ways; way++) { + int target; + int cha; + + if (KNL_MOD3(dram_rule)) + target = way; + else + target = 0x7 & sad_pkg( + pvt->info.interleave_pkg, interleave_reg, way); + + for (cha = 0; cha < KNL_MAX_CHAS; cha++) { + if (knl_get_mc_route(target, + mc_route_reg[cha]) == channel + && participants[channel]) { + participant_count++; + participants[channel] = 1; + break; + } + } + } + } + + if (participant_count != intrlv_ways) + edac_dbg(0, "participant_count (%d) != interleave_ways (%d): DIMM size may be incorrect\n", + participant_count, intrlv_ways); + + for (channel = 0; channel < KNL_MAX_CHANNELS; channel++) { + mc = knl_channel_mc(channel); + if (participants[channel]) { + edac_dbg(4, "mc channel %d contributes %lld bytes via sad entry %d\n", + channel, + sad_actual_size[mc]/intrlv_ways, + sad_rule); + mc_sizes[channel] += + sad_actual_size[mc]/intrlv_ways; + } + } + } + + return 0; +} + static int get_dimm_config(struct mem_ctl_info *mci) { struct sbridge_pvt *pvt = mci->pvt_info; @@ -964,13 +1606,20 @@ static int get_dimm_config(struct mem_ctl_info *mci) u32 reg; enum edac_type mode; enum mem_type mtype; + int channels = pvt->info.type == KNIGHTS_LANDING ? + KNL_MAX_CHANNELS : NUM_CHANNELS; + u64 knl_mc_sizes[KNL_MAX_CHANNELS]; - if (pvt->info.type == HASWELL || pvt->info.type == BROADWELL) + if (pvt->info.type == HASWELL || pvt->info.type == BROADWELL || + pvt->info.type == KNIGHTS_LANDING) pci_read_config_dword(pvt->pci_sad1, SAD_TARGET, ®); else pci_read_config_dword(pvt->pci_br0, SAD_TARGET, ®); - pvt->sbridge_dev->source_id = SOURCE_ID(reg); + if (pvt->info.type == KNIGHTS_LANDING) + pvt->sbridge_dev->source_id = SOURCE_ID_KNL(reg); + else + pvt->sbridge_dev->source_id = SOURCE_ID(reg); pvt->sbridge_dev->node_id = pvt->info.get_node_id(pvt); edac_dbg(0, "mc#%d: Node ID: %d, source ID: %d\n", @@ -978,31 +1627,42 @@ static int get_dimm_config(struct mem_ctl_info *mci) pvt->sbridge_dev->node_id, pvt->sbridge_dev->source_id); - pci_read_config_dword(pvt->pci_ras, RASENABLES, ®); - if (IS_MIRROR_ENABLED(reg)) { - edac_dbg(0, "Memory mirror is enabled\n"); - pvt->is_mirrored = true; - } else { - edac_dbg(0, "Memory mirror is disabled\n"); + /* KNL doesn't support mirroring or lockstep, + * and is always closed page + */ + if (pvt->info.type == KNIGHTS_LANDING) { + mode = EDAC_S4ECD4ED; pvt->is_mirrored = false; - } - pci_read_config_dword(pvt->pci_ta, MCMTR, &pvt->info.mcmtr); - if (IS_LOCKSTEP_ENABLED(pvt->info.mcmtr)) { - edac_dbg(0, "Lockstep is enabled\n"); - mode = EDAC_S8ECD8ED; - pvt->is_lockstep = true; + if (knl_get_dimm_capacity(pvt, knl_mc_sizes) != 0) + return -1; } else { - edac_dbg(0, "Lockstep is disabled\n"); - mode = EDAC_S4ECD4ED; - pvt->is_lockstep = false; - } - if (IS_CLOSE_PG(pvt->info.mcmtr)) { - edac_dbg(0, "address map is on closed page mode\n"); - pvt->is_close_pg = true; - } else { - edac_dbg(0, "address map is on open page mode\n"); - pvt->is_close_pg = false; + pci_read_config_dword(pvt->pci_ras, RASENABLES, ®); + if (IS_MIRROR_ENABLED(reg)) { + edac_dbg(0, "Memory mirror is enabled\n"); + pvt->is_mirrored = true; + } else { + edac_dbg(0, "Memory mirror is disabled\n"); + pvt->is_mirrored = false; + } + + pci_read_config_dword(pvt->pci_ta, MCMTR, &pvt->info.mcmtr); + if (IS_LOCKSTEP_ENABLED(pvt->info.mcmtr)) { + edac_dbg(0, "Lockstep is enabled\n"); + mode = EDAC_S8ECD8ED; + pvt->is_lockstep = true; + } else { + edac_dbg(0, "Lockstep is disabled\n"); + mode = EDAC_S4ECD4ED; + pvt->is_lockstep = false; + } + if (IS_CLOSE_PG(pvt->info.mcmtr)) { + edac_dbg(0, "address map is on closed page mode\n"); + pvt->is_close_pg = true; + } else { + edac_dbg(0, "address map is on open page mode\n"); + pvt->is_close_pg = false; + } } mtype = pvt->info.get_memory_type(pvt); @@ -1018,23 +1678,46 @@ static int get_dimm_config(struct mem_ctl_info *mci) else banks = 8; - for (i = 0; i < NUM_CHANNELS; i++) { + for (i = 0; i < channels; i++) { u32 mtr; - if (!pvt->pci_tad[i]) - continue; - for (j = 0; j < ARRAY_SIZE(mtr_regs); j++) { + int max_dimms_per_channel; + + if (pvt->info.type == KNIGHTS_LANDING) { + max_dimms_per_channel = 1; + if (!pvt->knl.pci_channel[i]) + continue; + } else { + max_dimms_per_channel = ARRAY_SIZE(mtr_regs); + if (!pvt->pci_tad[i]) + continue; + } + + for (j = 0; j < max_dimms_per_channel; j++) { dimm = EDAC_DIMM_PTR(mci->layers, mci->dimms, mci->n_layers, i, j, 0); - pci_read_config_dword(pvt->pci_tad[i], - mtr_regs[j], &mtr); + if (pvt->info.type == KNIGHTS_LANDING) { + pci_read_config_dword(pvt->knl.pci_channel[i], + knl_mtr_reg, &mtr); + } else { + pci_read_config_dword(pvt->pci_tad[i], + mtr_regs[j], &mtr); + } edac_dbg(4, "Channel #%d MTR%d = %x\n", i, j, mtr); if (IS_DIMM_PRESENT(mtr)) { pvt->channel[i].dimms++; ranks = numrank(pvt->info.type, mtr); - rows = numrow(mtr); - cols = numcol(mtr); + + if (pvt->info.type == KNIGHTS_LANDING) { + /* For DDR4, this is fixed. */ + cols = 1 << 10; + rows = knl_mc_sizes[i] / + ((u64) cols * ranks * banks * 8); + } else { + rows = numrow(mtr); + cols = numcol(mtr); + } size = ((u64)rows * cols * banks * ranks) >> (20 - 3); npages = MiB_TO_PAGES(size); @@ -1131,6 +1814,9 @@ static void get_memory_layout(const struct mem_ctl_info *mci) } } + if (pvt->info.type == KNIGHTS_LANDING) + return; + /* * Step 3) Get TAD range */ @@ -1727,6 +2413,8 @@ static int sbridge_get_all_devices_full(u8 *num_mc, #define sbridge_get_all_devices(num_mc, table) \ sbridge_get_all_devices_full(num_mc, table, 0, 0) +#define sbridge_get_all_devices_knl(num_mc, table) \ + sbridge_get_all_devices_full(num_mc, table, 1, 1) static int sbridge_mci_bind_devs(struct mem_ctl_info *mci, struct sbridge_dev *sbridge_dev) @@ -2083,6 +2771,131 @@ enodev: return -ENODEV; } +static int knl_mci_bind_devs(struct mem_ctl_info *mci, + struct sbridge_dev *sbridge_dev) +{ + struct sbridge_pvt *pvt = mci->pvt_info; + struct pci_dev *pdev; + int dev, func; + + int i; + int devidx; + + for (i = 0; i < sbridge_dev->n_devs; i++) { + pdev = sbridge_dev->pdev[i]; + if (!pdev) + continue; + + /* Extract PCI device and function. */ + dev = (pdev->devfn >> 3) & 0x1f; + func = pdev->devfn & 0x7; + + switch (pdev->device) { + case PCI_DEVICE_ID_INTEL_KNL_IMC_MC: + if (dev == 8) + pvt->knl.pci_mc0 = pdev; + else if (dev == 9) + pvt->knl.pci_mc1 = pdev; + else { + sbridge_printk(KERN_ERR, + "Memory controller in unexpected place! (dev %d, fn %d)\n", + dev, func); + continue; + } + break; + + case PCI_DEVICE_ID_INTEL_KNL_IMC_SAD0: + pvt->pci_sad0 = pdev; + break; + + case PCI_DEVICE_ID_INTEL_KNL_IMC_SAD1: + pvt->pci_sad1 = pdev; + break; + + case PCI_DEVICE_ID_INTEL_KNL_IMC_CHA: + /* There are one of these per tile, and range from + * 1.14.0 to 1.18.5. + */ + devidx = ((dev-14)*8)+func; + + if (devidx < 0 || devidx >= KNL_MAX_CHAS) { + sbridge_printk(KERN_ERR, + "Caching and Home Agent in unexpected place! (dev %d, fn %d)\n", + dev, func); + continue; + } + + WARN_ON(pvt->knl.pci_cha[devidx] != NULL); + + pvt->knl.pci_cha[devidx] = pdev; + break; + + case PCI_DEVICE_ID_INTEL_KNL_IMC_CHANNEL: + devidx = -1; + + /* + * MC0 channels 0-2 are device 9 function 2-4, + * MC1 channels 3-5 are device 8 function 2-4. + */ + + if (dev == 9) + devidx = func-2; + else if (dev == 8) + devidx = 3 + (func-2); + + if (devidx < 0 || devidx >= KNL_MAX_CHANNELS) { + sbridge_printk(KERN_ERR, + "DRAM Channel Registers in unexpected place! (dev %d, fn %d)\n", + dev, func); + continue; + } + + WARN_ON(pvt->knl.pci_channel[devidx] != NULL); + pvt->knl.pci_channel[devidx] = pdev; + break; + + case PCI_DEVICE_ID_INTEL_KNL_IMC_TOLHM: + pvt->knl.pci_mc_info = pdev; + break; + + case PCI_DEVICE_ID_INTEL_KNL_IMC_TA: + pvt->pci_ta = pdev; + break; + + default: + sbridge_printk(KERN_ERR, "Unexpected device %d\n", + pdev->device); + break; + } + } + + if (!pvt->knl.pci_mc0 || !pvt->knl.pci_mc1 || + !pvt->pci_sad0 || !pvt->pci_sad1 || + !pvt->pci_ta) { + goto enodev; + } + + for (i = 0; i < KNL_MAX_CHANNELS; i++) { + if (!pvt->knl.pci_channel[i]) { + sbridge_printk(KERN_ERR, "Missing channel %d\n", i); + goto enodev; + } + } + + for (i = 0; i < KNL_MAX_CHAS; i++) { + if (!pvt->knl.pci_cha[i]) { + sbridge_printk(KERN_ERR, "Missing CHA %d\n", i); + goto enodev; + } + } + + return 0; + +enodev: + sbridge_printk(KERN_ERR, "Some needed devices are missing\n"); + return -ENODEV; +} + /**************************************************************************** Error check routines ****************************************************************************/ @@ -2172,8 +2985,36 @@ static void sbridge_mce_output_error(struct mem_ctl_info *mci, if (!GET_BITFIELD(m->status, 58, 58)) return; - rc = get_memory_error_data(mci, m->addr, &socket, &ha, - &channel_mask, &rank, &area_type, msg); + if (pvt->info.type == KNIGHTS_LANDING) { + if (channel == 14) { + edac_dbg(0, "%s%s err_code:%04x:%04x EDRAM bank %d\n", + overflow ? " OVERFLOW" : "", + (uncorrected_error && recoverable) + ? " recoverable" : "", + mscod, errcode, + m->bank); + } else { + char A = *("A"); + + channel = knl_channel_remap(channel); + channel_mask = 1 << channel; + snprintf(msg, sizeof(msg), + "%s%s err_code:%04x:%04x channel:%d (DIMM_%c)", + overflow ? " OVERFLOW" : "", + (uncorrected_error && recoverable) + ? " recoverable" : " ", + mscod, errcode, channel, A + channel); + edac_mc_handle_error(tp_event, mci, core_err_cnt, + m->addr >> PAGE_SHIFT, m->addr & ~PAGE_MASK, 0, + channel, 0, -1, + optype, msg); + } + return; + } else { + rc = get_memory_error_data(mci, m->addr, &socket, &ha, + &channel_mask, &rank, &area_type, msg); + } + if (rc < 0) goto err_parsing; new_mci = get_mci_for_node_id(socket); @@ -2404,10 +3245,11 @@ static int sbridge_register_mci(struct sbridge_dev *sbridge_dev, enum type type) /* allocate a new MC control structure */ layers[0].type = EDAC_MC_LAYER_CHANNEL; - layers[0].size = NUM_CHANNELS; + layers[0].size = type == KNIGHTS_LANDING ? + KNL_MAX_CHANNELS : NUM_CHANNELS; layers[0].is_virt_csrow = false; layers[1].type = EDAC_MC_LAYER_SLOT; - layers[1].size = MAX_DIMMS; + layers[1].size = type == KNIGHTS_LANDING ? 1 : MAX_DIMMS; layers[1].is_virt_csrow = true; mci = edac_mc_alloc(sbridge_dev->mc, ARRAY_SIZE(layers), layers, sizeof(*pvt)); @@ -2425,7 +3267,8 @@ static int sbridge_register_mci(struct sbridge_dev *sbridge_dev, enum type type) pvt->sbridge_dev = sbridge_dev; sbridge_dev->mci = mci; - mci->mtype_cap = MEM_FLAG_DDR3; + mci->mtype_cap = type == KNIGHTS_LANDING ? + MEM_FLAG_DDR4 : MEM_FLAG_DDR3; mci->edac_ctl_cap = EDAC_FLAG_NONE; mci->edac_cap = EDAC_FLAG_NONE; mci->mod_name = "sbridge_edac.c"; @@ -2534,6 +3377,30 @@ static int sbridge_register_mci(struct sbridge_dev *sbridge_dev, enum type type) if (unlikely(rc < 0)) goto fail0; break; + case KNIGHTS_LANDING: + /* pvt->info.rankcfgr == ??? */ + pvt->info.get_tolm = knl_get_tolm; + pvt->info.get_tohm = knl_get_tohm; + pvt->info.dram_rule = knl_dram_rule; + pvt->info.get_memory_type = knl_get_memory_type; + pvt->info.get_node_id = knl_get_node_id; + pvt->info.rir_limit = NULL; + pvt->info.sad_limit = knl_sad_limit; + pvt->info.interleave_mode = knl_interleave_mode; + pvt->info.show_interleave_mode = knl_show_interleave_mode; + pvt->info.dram_attr = dram_attr_knl; + pvt->info.max_sad = ARRAY_SIZE(knl_dram_rule); + pvt->info.interleave_list = knl_interleave_list; + pvt->info.max_interleave = ARRAY_SIZE(knl_interleave_list); + pvt->info.interleave_pkg = ibridge_interleave_pkg; + pvt->info.get_width = ibridge_get_width; + mci->ctl_name = kasprintf(GFP_KERNEL, + "Knights Landing Socket#%d", mci->mc_idx); + + rc = knl_mci_bind_devs(mci, sbridge_dev); + if (unlikely(rc < 0)) + goto fail0; + break; } /* Get dimm basic config and the memory layout */ @@ -2588,20 +3455,29 @@ static int sbridge_probe(struct pci_dev *pdev, const struct pci_device_id *id) switch (pdev->device) { case PCI_DEVICE_ID_INTEL_IBRIDGE_IMC_HA0_TA: - rc = sbridge_get_all_devices(&num_mc, pci_dev_descr_ibridge_table); + rc = sbridge_get_all_devices(&num_mc, + pci_dev_descr_ibridge_table); type = IVY_BRIDGE; break; case PCI_DEVICE_ID_INTEL_SBRIDGE_IMC_HA0: - rc = sbridge_get_all_devices(&num_mc, pci_dev_descr_sbridge_table); + rc = sbridge_get_all_devices(&num_mc, + pci_dev_descr_sbridge_table); type = SANDY_BRIDGE; break; case PCI_DEVICE_ID_INTEL_HASWELL_IMC_HA0: - rc = sbridge_get_all_devices(&num_mc, pci_dev_descr_haswell_table); + rc = sbridge_get_all_devices(&num_mc, + pci_dev_descr_haswell_table); type = HASWELL; break; case PCI_DEVICE_ID_INTEL_BROADWELL_IMC_HA0: - rc = sbridge_get_all_devices(&num_mc, pci_dev_descr_broadwell_table); + rc = sbridge_get_all_devices(&num_mc, + pci_dev_descr_broadwell_table); type = BROADWELL; + break; + case PCI_DEVICE_ID_INTEL_KNL_IMC_SAD0: + rc = sbridge_get_all_devices_knl(&num_mc, + pci_dev_descr_knl_table); + type = KNIGHTS_LANDING; break; } if (unlikely(rc < 0)) { -- 2.20.1