1 /*
2 * Copyright 2023 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23 #include "umc_v12_0.h"
24 #include "amdgpu_ras.h"
25 #include "amdgpu_umc.h"
26 #include "amdgpu.h"
27 #include "umc/umc_12_0_0_offset.h"
28 #include "umc/umc_12_0_0_sh_mask.h"
29 #include "mp/mp_13_0_6_sh_mask.h"
30
31 #define MAX_ECC_NUM_PER_RETIREMENT 32
32 #define DELAYED_TIME_FOR_GPU_RESET 1000 //ms
33
get_umc_v12_0_reg_offset(struct amdgpu_device * adev,uint32_t node_inst,uint32_t umc_inst,uint32_t ch_inst)34 static inline uint64_t get_umc_v12_0_reg_offset(struct amdgpu_device *adev,
35 uint32_t node_inst,
36 uint32_t umc_inst,
37 uint32_t ch_inst)
38 {
39 uint32_t index = umc_inst * adev->umc.channel_inst_num + ch_inst;
40 uint64_t cross_node_offset = (node_inst == 0) ? 0 : UMC_V12_0_CROSS_NODE_OFFSET;
41
42 umc_inst = index / 4;
43 ch_inst = index % 4;
44
45 return adev->umc.channel_offs * ch_inst + UMC_V12_0_INST_DIST * umc_inst +
46 UMC_V12_0_NODE_DIST * node_inst + cross_node_offset;
47 }
48
umc_v12_0_reset_error_count_per_channel(struct amdgpu_device * adev,uint32_t node_inst,uint32_t umc_inst,uint32_t ch_inst,void * data)49 static int umc_v12_0_reset_error_count_per_channel(struct amdgpu_device *adev,
50 uint32_t node_inst, uint32_t umc_inst,
51 uint32_t ch_inst, void *data)
52 {
53 uint64_t odecc_err_cnt_addr;
54 uint64_t umc_reg_offset =
55 get_umc_v12_0_reg_offset(adev, node_inst, umc_inst, ch_inst);
56
57 odecc_err_cnt_addr =
58 SOC15_REG_OFFSET(UMC, 0, regUMCCH0_OdEccErrCnt);
59
60 /* clear error count */
61 WREG32_PCIE_EXT((odecc_err_cnt_addr + umc_reg_offset) * 4,
62 UMC_V12_0_CE_CNT_INIT);
63
64 return 0;
65 }
66
umc_v12_0_reset_error_count(struct amdgpu_device * adev)67 static void umc_v12_0_reset_error_count(struct amdgpu_device *adev)
68 {
69 amdgpu_umc_loop_channels(adev,
70 umc_v12_0_reset_error_count_per_channel, NULL);
71 }
72
umc_v12_0_is_deferred_error(struct amdgpu_device * adev,uint64_t mc_umc_status)73 bool umc_v12_0_is_deferred_error(struct amdgpu_device *adev, uint64_t mc_umc_status)
74 {
75 dev_dbg(adev->dev,
76 "MCA_UMC_STATUS(0x%llx): Val:%llu, Poison:%llu, Deferred:%llu, PCC:%llu, UC:%llu, TCC:%llu\n",
77 mc_umc_status,
78 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val),
79 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Poison),
80 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Deferred),
81 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, PCC),
82 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, UC),
83 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, TCC)
84 );
85
86 return (amdgpu_ras_is_poison_mode_supported(adev) &&
87 (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val) == 1) &&
88 ((REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Deferred) == 1) ||
89 (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Poison) == 1)));
90 }
91
umc_v12_0_is_uncorrectable_error(struct amdgpu_device * adev,uint64_t mc_umc_status)92 bool umc_v12_0_is_uncorrectable_error(struct amdgpu_device *adev, uint64_t mc_umc_status)
93 {
94 if (umc_v12_0_is_deferred_error(adev, mc_umc_status))
95 return false;
96
97 return ((REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val) == 1) &&
98 (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, PCC) == 1 ||
99 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, UC) == 1 ||
100 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, TCC) == 1));
101 }
102
umc_v12_0_is_correctable_error(struct amdgpu_device * adev,uint64_t mc_umc_status)103 bool umc_v12_0_is_correctable_error(struct amdgpu_device *adev, uint64_t mc_umc_status)
104 {
105 if (umc_v12_0_is_deferred_error(adev, mc_umc_status))
106 return false;
107
108 return (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val) == 1 &&
109 (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, CECC) == 1 ||
110 (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, UECC) == 1 &&
111 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, UC) == 0) ||
112 /* Identify data parity error in replay mode */
113 ((REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, ErrorCodeExt) == 0x5 ||
114 REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, ErrorCodeExt) == 0xb) &&
115 !(umc_v12_0_is_uncorrectable_error(adev, mc_umc_status)))));
116 }
117
umc_v12_0_query_error_count_per_type(struct amdgpu_device * adev,uint64_t umc_reg_offset,unsigned long * error_count,check_error_type_func error_type_func)118 static void umc_v12_0_query_error_count_per_type(struct amdgpu_device *adev,
119 uint64_t umc_reg_offset,
120 unsigned long *error_count,
121 check_error_type_func error_type_func)
122 {
123 uint64_t mc_umc_status;
124 uint64_t mc_umc_status_addr;
125
126 mc_umc_status_addr =
127 SOC15_REG_OFFSET(UMC, 0, regMCA_UMC_UMC0_MCUMC_STATUST0);
128
129 /* Check MCUMC_STATUS */
130 mc_umc_status =
131 RREG64_PCIE_EXT((mc_umc_status_addr + umc_reg_offset) * 4);
132
133 if (error_type_func(adev, mc_umc_status))
134 *error_count += 1;
135 }
136
umc_v12_0_query_error_count(struct amdgpu_device * adev,uint32_t node_inst,uint32_t umc_inst,uint32_t ch_inst,void * data)137 static int umc_v12_0_query_error_count(struct amdgpu_device *adev,
138 uint32_t node_inst, uint32_t umc_inst,
139 uint32_t ch_inst, void *data)
140 {
141 struct ras_err_data *err_data = (struct ras_err_data *)data;
142 unsigned long ue_count = 0, ce_count = 0, de_count = 0;
143
144 /* NOTE: node_inst is converted by adev->umc.active_mask and the range is [0-3],
145 * which can be used as die ID directly */
146 struct amdgpu_smuio_mcm_config_info mcm_info = {
147 .socket_id = adev->smuio.funcs->get_socket_id(adev),
148 .die_id = node_inst,
149 };
150
151 uint64_t umc_reg_offset =
152 get_umc_v12_0_reg_offset(adev, node_inst, umc_inst, ch_inst);
153
154 umc_v12_0_query_error_count_per_type(adev, umc_reg_offset,
155 &ce_count, umc_v12_0_is_correctable_error);
156 umc_v12_0_query_error_count_per_type(adev, umc_reg_offset,
157 &ue_count, umc_v12_0_is_uncorrectable_error);
158 umc_v12_0_query_error_count_per_type(adev, umc_reg_offset,
159 &de_count, umc_v12_0_is_deferred_error);
160
161 amdgpu_ras_error_statistic_ue_count(err_data, &mcm_info, ue_count);
162 amdgpu_ras_error_statistic_ce_count(err_data, &mcm_info, ce_count);
163 amdgpu_ras_error_statistic_de_count(err_data, &mcm_info, de_count);
164
165 return 0;
166 }
167
umc_v12_0_query_ras_error_count(struct amdgpu_device * adev,void * ras_error_status)168 static void umc_v12_0_query_ras_error_count(struct amdgpu_device *adev,
169 void *ras_error_status)
170 {
171 amdgpu_umc_loop_channels(adev,
172 umc_v12_0_query_error_count, ras_error_status);
173
174 umc_v12_0_reset_error_count(adev);
175 }
176
umc_v12_0_get_retire_flip_bits(struct amdgpu_device * adev)177 static void umc_v12_0_get_retire_flip_bits(struct amdgpu_device *adev)
178 {
179 enum amdgpu_memory_partition nps = AMDGPU_NPS1_PARTITION_MODE;
180 uint32_t vram_type = adev->gmc.vram_type;
181 struct amdgpu_umc_flip_bits *flip_bits = &(adev->umc.flip_bits);
182
183 if (adev->gmc.gmc_funcs->query_mem_partition_mode)
184 nps = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
185
186 /* default setting */
187 flip_bits->flip_bits_in_pa[0] = UMC_V12_0_PA_C2_BIT;
188 flip_bits->flip_bits_in_pa[1] = UMC_V12_0_PA_C3_BIT;
189 flip_bits->flip_bits_in_pa[2] = UMC_V12_0_PA_C4_BIT;
190 flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R13_BIT;
191 flip_bits->flip_row_bit = 13;
192 flip_bits->bit_num = 4;
193 flip_bits->r13_in_pa = UMC_V12_0_PA_R13_BIT;
194
195 if (nps == AMDGPU_NPS2_PARTITION_MODE) {
196 flip_bits->flip_bits_in_pa[0] = UMC_V12_0_PA_CH5_BIT;
197 flip_bits->flip_bits_in_pa[1] = UMC_V12_0_PA_C2_BIT;
198 flip_bits->flip_bits_in_pa[2] = UMC_V12_0_PA_B1_BIT;
199 flip_bits->r13_in_pa = UMC_V12_0_PA_R12_BIT;
200 } else if (nps == AMDGPU_NPS4_PARTITION_MODE) {
201 flip_bits->flip_bits_in_pa[0] = UMC_V12_0_PA_CH4_BIT;
202 flip_bits->flip_bits_in_pa[1] = UMC_V12_0_PA_CH5_BIT;
203 flip_bits->flip_bits_in_pa[2] = UMC_V12_0_PA_B0_BIT;
204 flip_bits->r13_in_pa = UMC_V12_0_PA_R11_BIT;
205 }
206
207 switch (vram_type) {
208 case AMDGPU_VRAM_TYPE_HBM:
209 /* other nps modes are taken as nps1 */
210 if (nps == AMDGPU_NPS2_PARTITION_MODE)
211 flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R12_BIT;
212 else if (nps == AMDGPU_NPS4_PARTITION_MODE)
213 flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R11_BIT;
214
215 break;
216 case AMDGPU_VRAM_TYPE_HBM3E:
217 flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R12_BIT;
218 flip_bits->flip_row_bit = 12;
219
220 if (nps == AMDGPU_NPS2_PARTITION_MODE)
221 flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R11_BIT;
222 else if (nps == AMDGPU_NPS4_PARTITION_MODE)
223 flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R10_BIT;
224
225 break;
226 default:
227 dev_warn(adev->dev,
228 "Unknown HBM type, set RAS retire flip bits to the value in NPS1 mode.\n");
229 break;
230 }
231
232 adev->umc.retire_unit = 0x1 << flip_bits->bit_num;
233 }
234
umc_v12_0_convert_error_address(struct amdgpu_device * adev,struct ras_err_data * err_data,struct ta_ras_query_address_input * addr_in,struct ta_ras_query_address_output * addr_out,bool dump_addr)235 static int umc_v12_0_convert_error_address(struct amdgpu_device *adev,
236 struct ras_err_data *err_data,
237 struct ta_ras_query_address_input *addr_in,
238 struct ta_ras_query_address_output *addr_out,
239 bool dump_addr)
240 {
241 uint32_t col, col_lower, row, row_lower, row_high, bank;
242 uint32_t channel_index = 0, umc_inst = 0;
243 uint32_t i, bit_num, retire_unit, *flip_bits;
244 uint64_t soc_pa, column, err_addr;
245 struct ta_ras_query_address_output addr_out_tmp;
246 struct ta_ras_query_address_output *paddr_out;
247 int ret = 0;
248
249 if (!addr_out)
250 paddr_out = &addr_out_tmp;
251 else
252 paddr_out = addr_out;
253
254 err_addr = bank = 0;
255 if (addr_in) {
256 err_addr = addr_in->ma.err_addr;
257 addr_in->addr_type = TA_RAS_MCA_TO_PA;
258 ret = psp_ras_query_address(&adev->psp, addr_in, paddr_out);
259 if (ret) {
260 dev_warn(adev->dev, "Failed to query RAS physical address for 0x%llx",
261 err_addr);
262
263 goto out;
264 }
265
266 bank = paddr_out->pa.bank;
267 /* no need to care about umc inst if addr_in is NULL */
268 umc_inst = addr_in->ma.umc_inst;
269 }
270
271 flip_bits = adev->umc.flip_bits.flip_bits_in_pa;
272 bit_num = adev->umc.flip_bits.bit_num;
273 retire_unit = adev->umc.retire_unit;
274
275 soc_pa = paddr_out->pa.pa;
276 channel_index = paddr_out->pa.channel_idx;
277 /* clear loop bits in soc physical address */
278 for (i = 0; i < bit_num; i++)
279 soc_pa &= ~BIT_ULL(flip_bits[i]);
280
281 paddr_out->pa.pa = soc_pa;
282 /* get column bit 0 and 1 in mca address */
283 col_lower = (err_addr >> 1) & 0x3ULL;
284 /* extra row bit will be handled later */
285 row_lower = (err_addr >> UMC_V12_0_MA_R0_BIT) & 0x1fffULL;
286 row_lower &= ~BIT_ULL(adev->umc.flip_bits.flip_row_bit);
287
288 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(9, 5, 0)) {
289 row_high = (soc_pa >> adev->umc.flip_bits.r13_in_pa) & 0x3ULL;
290 /* it's 2.25GB in each channel, from MCA address to PA
291 * [R14 R13] is converted if the two bits value are 0x3,
292 * get them from PA instead of MCA address.
293 */
294 row_lower |= (row_high << 13);
295 }
296
297 if (!err_data && !dump_addr)
298 goto out;
299
300 /* loop for all possibilities of retired bits */
301 for (column = 0; column < retire_unit; column++) {
302 soc_pa = paddr_out->pa.pa;
303 for (i = 0; i < bit_num; i++)
304 soc_pa |= (((column >> i) & 0x1ULL) << flip_bits[i]);
305
306 col = ((column & 0x7) << 2) | col_lower;
307 /* handle extra row bit */
308 if (bit_num == RETIRE_FLIP_BITS_NUM)
309 row = ((column >> 3) << adev->umc.flip_bits.flip_row_bit) |
310 row_lower;
311
312 if (dump_addr)
313 dev_info(adev->dev,
314 "Error Address(PA):0x%-10llx Row:0x%-4x Col:0x%-2x Bank:0x%x Channel:0x%x\n",
315 soc_pa, row, col, bank, channel_index);
316
317 if (err_data)
318 amdgpu_umc_fill_error_record(err_data, err_addr,
319 soc_pa, channel_index, umc_inst);
320 }
321
322 out:
323 return ret;
324 }
325
umc_v12_0_query_error_address(struct amdgpu_device * adev,uint32_t node_inst,uint32_t umc_inst,uint32_t ch_inst,void * data)326 static int umc_v12_0_query_error_address(struct amdgpu_device *adev,
327 uint32_t node_inst, uint32_t umc_inst,
328 uint32_t ch_inst, void *data)
329 {
330 struct ras_err_data *err_data = (struct ras_err_data *)data;
331 struct ta_ras_query_address_input addr_in;
332 uint64_t mc_umc_status_addr;
333 uint64_t mc_umc_status, err_addr;
334 uint64_t mc_umc_addrt0;
335 uint64_t umc_reg_offset =
336 get_umc_v12_0_reg_offset(adev, node_inst, umc_inst, ch_inst);
337
338 mc_umc_status_addr =
339 SOC15_REG_OFFSET(UMC, 0, regMCA_UMC_UMC0_MCUMC_STATUST0);
340
341 mc_umc_status = RREG64_PCIE_EXT((mc_umc_status_addr + umc_reg_offset) * 4);
342
343 if (mc_umc_status == 0)
344 return 0;
345
346 if (!err_data->err_addr) {
347 /* clear umc status */
348 WREG64_PCIE_EXT((mc_umc_status_addr + umc_reg_offset) * 4, 0x0ULL);
349
350 return 0;
351 }
352
353 /* calculate error address if ue error is detected */
354 if (umc_v12_0_is_uncorrectable_error(adev, mc_umc_status) ||
355 umc_v12_0_is_deferred_error(adev, mc_umc_status)) {
356 mc_umc_addrt0 =
357 SOC15_REG_OFFSET(UMC, 0, regMCA_UMC_UMC0_MCUMC_ADDRT0);
358
359 err_addr = RREG64_PCIE_EXT((mc_umc_addrt0 + umc_reg_offset) * 4);
360
361 err_addr = REG_GET_FIELD(err_addr, MCA_UMC_UMC0_MCUMC_ADDRT0, ErrorAddr);
362
363 if (!adev->aid_mask &&
364 adev->smuio.funcs &&
365 adev->smuio.funcs->get_socket_id)
366 addr_in.ma.socket_id = adev->smuio.funcs->get_socket_id(adev);
367 else
368 addr_in.ma.socket_id = 0;
369
370 addr_in.ma.err_addr = err_addr;
371 addr_in.ma.ch_inst = ch_inst;
372 addr_in.ma.umc_inst = umc_inst;
373 addr_in.ma.node_inst = node_inst;
374
375 umc_v12_0_convert_error_address(adev, err_data, &addr_in, NULL, true);
376 }
377
378 /* clear umc status */
379 WREG64_PCIE_EXT((mc_umc_status_addr + umc_reg_offset) * 4, 0x0ULL);
380
381 return 0;
382 }
383
umc_v12_0_query_ras_error_address(struct amdgpu_device * adev,void * ras_error_status)384 static void umc_v12_0_query_ras_error_address(struct amdgpu_device *adev,
385 void *ras_error_status)
386 {
387 amdgpu_umc_loop_channels(adev,
388 umc_v12_0_query_error_address, ras_error_status);
389 }
390
umc_v12_0_err_cnt_init_per_channel(struct amdgpu_device * adev,uint32_t node_inst,uint32_t umc_inst,uint32_t ch_inst,void * data)391 static int umc_v12_0_err_cnt_init_per_channel(struct amdgpu_device *adev,
392 uint32_t node_inst, uint32_t umc_inst,
393 uint32_t ch_inst, void *data)
394 {
395 uint32_t odecc_cnt_sel;
396 uint64_t odecc_cnt_sel_addr, odecc_err_cnt_addr;
397 uint64_t umc_reg_offset =
398 get_umc_v12_0_reg_offset(adev, node_inst, umc_inst, ch_inst);
399
400 odecc_cnt_sel_addr =
401 SOC15_REG_OFFSET(UMC, 0, regUMCCH0_OdEccCntSel);
402 odecc_err_cnt_addr =
403 SOC15_REG_OFFSET(UMC, 0, regUMCCH0_OdEccErrCnt);
404
405 odecc_cnt_sel = RREG32_PCIE_EXT((odecc_cnt_sel_addr + umc_reg_offset) * 4);
406
407 /* set ce error interrupt type to APIC based interrupt */
408 odecc_cnt_sel = REG_SET_FIELD(odecc_cnt_sel, UMCCH0_OdEccCntSel,
409 OdEccErrInt, 0x1);
410 WREG32_PCIE_EXT((odecc_cnt_sel_addr + umc_reg_offset) * 4, odecc_cnt_sel);
411
412 /* set error count to initial value */
413 WREG32_PCIE_EXT((odecc_err_cnt_addr + umc_reg_offset) * 4, UMC_V12_0_CE_CNT_INIT);
414
415 return 0;
416 }
417
umc_v12_0_check_ecc_err_status(struct amdgpu_device * adev,enum amdgpu_mca_error_type type,void * ras_error_status)418 static bool umc_v12_0_check_ecc_err_status(struct amdgpu_device *adev,
419 enum amdgpu_mca_error_type type, void *ras_error_status)
420 {
421 uint64_t mc_umc_status = *(uint64_t *)ras_error_status;
422
423 switch (type) {
424 case AMDGPU_MCA_ERROR_TYPE_UE:
425 return umc_v12_0_is_uncorrectable_error(adev, mc_umc_status);
426 case AMDGPU_MCA_ERROR_TYPE_CE:
427 return umc_v12_0_is_correctable_error(adev, mc_umc_status);
428 case AMDGPU_MCA_ERROR_TYPE_DE:
429 return umc_v12_0_is_deferred_error(adev, mc_umc_status);
430 default:
431 return false;
432 }
433
434 return false;
435 }
436
umc_v12_0_err_cnt_init(struct amdgpu_device * adev)437 static void umc_v12_0_err_cnt_init(struct amdgpu_device *adev)
438 {
439 amdgpu_umc_loop_channels(adev,
440 umc_v12_0_err_cnt_init_per_channel, NULL);
441 }
442
umc_v12_0_query_ras_poison_mode(struct amdgpu_device * adev)443 static bool umc_v12_0_query_ras_poison_mode(struct amdgpu_device *adev)
444 {
445 /*
446 * Force return true, because regUMCCH0_EccCtrl
447 * is not accessible from host side
448 */
449 return true;
450 }
451
452 const struct amdgpu_ras_block_hw_ops umc_v12_0_ras_hw_ops = {
453 .query_ras_error_count = umc_v12_0_query_ras_error_count,
454 .query_ras_error_address = umc_v12_0_query_ras_error_address,
455 };
456
umc_v12_0_aca_bank_parser(struct aca_handle * handle,struct aca_bank * bank,enum aca_smu_type type,void * data)457 static int umc_v12_0_aca_bank_parser(struct aca_handle *handle, struct aca_bank *bank,
458 enum aca_smu_type type, void *data)
459 {
460 struct amdgpu_device *adev = handle->adev;
461 struct aca_bank_info info;
462 enum aca_error_type err_type;
463 u64 status, count;
464 u32 ext_error_code;
465 int ret;
466
467 status = bank->regs[ACA_REG_IDX_STATUS];
468 if (umc_v12_0_is_deferred_error(adev, status))
469 err_type = ACA_ERROR_TYPE_DEFERRED;
470 else if (umc_v12_0_is_uncorrectable_error(adev, status))
471 err_type = ACA_ERROR_TYPE_UE;
472 else if (umc_v12_0_is_correctable_error(adev, status))
473 err_type = ACA_ERROR_TYPE_CE;
474 else
475 return 0;
476 bank->aca_err_type = err_type;
477
478 ret = aca_bank_info_decode(bank, &info);
479 if (ret)
480 return ret;
481
482 amdgpu_umc_update_ecc_status(adev,
483 bank->regs[ACA_REG_IDX_STATUS],
484 bank->regs[ACA_REG_IDX_IPID],
485 bank->regs[ACA_REG_IDX_ADDR]);
486
487 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status);
488 if (umc_v12_0_is_deferred_error(adev, status))
489 count = ext_error_code == 0 ?
490 adev->umc.err_addr_cnt / adev->umc.retire_unit : 1ULL;
491 else
492 count = ext_error_code == 0 ?
493 ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]) : 1ULL;
494
495 return aca_error_cache_log_bank_error(handle, &info, err_type, count);
496 }
497
498 static const struct aca_bank_ops umc_v12_0_aca_bank_ops = {
499 .aca_bank_parser = umc_v12_0_aca_bank_parser,
500 };
501
502 const struct aca_info umc_v12_0_aca_info = {
503 .hwip = ACA_HWIP_TYPE_UMC,
504 .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK | ACA_ERROR_DEFERRED_MASK,
505 .bank_ops = &umc_v12_0_aca_bank_ops,
506 };
507
umc_v12_0_ras_late_init(struct amdgpu_device * adev,struct ras_common_if * ras_block)508 static int umc_v12_0_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
509 {
510 int ret;
511
512 ret = amdgpu_umc_ras_late_init(adev, ras_block);
513 if (ret)
514 return ret;
515
516 ret = amdgpu_ras_bind_aca(adev, AMDGPU_RAS_BLOCK__UMC,
517 &umc_v12_0_aca_info, NULL);
518 if (ret)
519 return ret;
520
521 return 0;
522 }
523
umc_v12_0_update_ecc_status(struct amdgpu_device * adev,uint64_t status,uint64_t ipid,uint64_t addr)524 static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev,
525 uint64_t status, uint64_t ipid, uint64_t addr)
526 {
527 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
528 uint16_t hwid, mcatype;
529 uint64_t page_pfn[UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL];
530 uint64_t err_addr, pa_addr = 0;
531 struct ras_ecc_err *ecc_err;
532 struct ta_ras_query_address_output addr_out;
533 uint32_t shift_bit = adev->umc.flip_bits.flip_bits_in_pa[2];
534 int count, ret, i;
535
536 hwid = REG_GET_FIELD(ipid, MCMP1_IPIDT0, HardwareID);
537 mcatype = REG_GET_FIELD(ipid, MCMP1_IPIDT0, McaType);
538
539 if ((hwid != MCA_UMC_HWID_V12_0) || (mcatype != MCA_UMC_MCATYPE_V12_0))
540 return 0;
541
542 if (!status)
543 return 0;
544
545 if (!umc_v12_0_is_deferred_error(adev, status))
546 return 0;
547
548 err_addr = REG_GET_FIELD(addr,
549 MCA_UMC_UMC0_MCUMC_ADDRT0, ErrorAddr);
550
551 dev_dbg(adev->dev,
552 "UMC:IPID:0x%llx, socket:%llu, aid:%llu, inst:%llu, ch:%llu, err_addr:0x%llx\n",
553 ipid,
554 MCA_IPID_2_SOCKET_ID(ipid),
555 MCA_IPID_2_DIE_ID(ipid),
556 MCA_IPID_2_UMC_INST(ipid),
557 MCA_IPID_2_UMC_CH(ipid),
558 err_addr);
559
560 ret = amdgpu_umc_mca_to_addr(adev,
561 err_addr, MCA_IPID_2_UMC_CH(ipid),
562 MCA_IPID_2_UMC_INST(ipid), MCA_IPID_2_DIE_ID(ipid),
563 MCA_IPID_2_SOCKET_ID(ipid), &addr_out, true);
564 if (ret)
565 return ret;
566
567 ecc_err = kzalloc(sizeof(*ecc_err), GFP_KERNEL);
568 if (!ecc_err)
569 return -ENOMEM;
570
571 pa_addr = addr_out.pa.pa;
572 ecc_err->status = status;
573 ecc_err->ipid = ipid;
574 ecc_err->addr = addr;
575 ecc_err->pa_pfn = pa_addr >> AMDGPU_GPU_PAGE_SHIFT;
576 ecc_err->channel_idx = addr_out.pa.channel_idx;
577
578 /* If converted pa_pfn is 0, use pa C4 pfn. */
579 if (!ecc_err->pa_pfn)
580 ecc_err->pa_pfn = BIT_ULL(shift_bit) >> AMDGPU_GPU_PAGE_SHIFT;
581
582 ret = amdgpu_umc_logs_ecc_err(adev, &con->umc_ecc_log.de_page_tree, ecc_err);
583 if (ret) {
584 if (ret == -EEXIST)
585 con->umc_ecc_log.de_queried_count++;
586 else
587 dev_err(adev->dev, "Fail to log ecc error! ret:%d\n", ret);
588
589 kfree(ecc_err);
590 return ret;
591 }
592
593 con->umc_ecc_log.de_queried_count++;
594
595 memset(page_pfn, 0, sizeof(page_pfn));
596 count = amdgpu_umc_lookup_bad_pages_in_a_row(adev,
597 pa_addr,
598 page_pfn, ARRAY_SIZE(page_pfn));
599 if (count <= 0) {
600 dev_warn(adev->dev, "Fail to convert error address! count:%d\n", count);
601 return 0;
602 }
603
604 /* Reserve memory */
605 for (i = 0; i < count; i++)
606 amdgpu_ras_reserve_page(adev, page_pfn[i]);
607
608 /* The problem case is as follows:
609 * 1. GPU A triggers a gpu ras reset, and GPU A drives
610 * GPU B to also perform a gpu ras reset.
611 * 2. After gpu B ras reset started, gpu B queried a DE
612 * data. Since the DE data was queried in the ras reset
613 * thread instead of the page retirement thread, bad
614 * page retirement work would not be triggered. Then
615 * even if all gpu resets are completed, the bad pages
616 * will be cached in RAM until GPU B's bad page retirement
617 * work is triggered again and then saved to eeprom.
618 * Trigger delayed work to save the bad pages to eeprom in time
619 * after gpu ras reset is completed.
620 */
621 if (amdgpu_ras_in_recovery(adev))
622 schedule_delayed_work(&con->page_retirement_dwork,
623 msecs_to_jiffies(DELAYED_TIME_FOR_GPU_RESET));
624
625 return 0;
626 }
627
umc_v12_0_fill_error_record(struct amdgpu_device * adev,struct ras_ecc_err * ecc_err,void * ras_error_status)628 static int umc_v12_0_fill_error_record(struct amdgpu_device *adev,
629 struct ras_ecc_err *ecc_err, void *ras_error_status)
630 {
631 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
632 uint64_t page_pfn[UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL];
633 int ret, i, count;
634
635 if (!err_data || !ecc_err)
636 return -EINVAL;
637
638 memset(page_pfn, 0, sizeof(page_pfn));
639 count = amdgpu_umc_lookup_bad_pages_in_a_row(adev,
640 ecc_err->pa_pfn << AMDGPU_GPU_PAGE_SHIFT,
641 page_pfn, ARRAY_SIZE(page_pfn));
642
643 for (i = 0; i < count; i++) {
644 ret = amdgpu_umc_fill_error_record(err_data,
645 ecc_err->addr,
646 page_pfn[i] << AMDGPU_GPU_PAGE_SHIFT,
647 ecc_err->channel_idx,
648 MCA_IPID_2_UMC_INST(ecc_err->ipid));
649 if (ret)
650 break;
651 }
652
653 err_data->de_count++;
654
655 return ret;
656 }
657
umc_v12_0_query_ras_ecc_err_addr(struct amdgpu_device * adev,void * ras_error_status)658 static void umc_v12_0_query_ras_ecc_err_addr(struct amdgpu_device *adev,
659 void *ras_error_status)
660 {
661 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
662 struct ras_ecc_err *entries[MAX_ECC_NUM_PER_RETIREMENT];
663 struct radix_tree_root *ecc_tree;
664 int new_detected, ret, i;
665
666 ecc_tree = &con->umc_ecc_log.de_page_tree;
667
668 mutex_lock(&con->umc_ecc_log.lock);
669 new_detected = radix_tree_gang_lookup_tag(ecc_tree, (void **)entries,
670 0, ARRAY_SIZE(entries), UMC_ECC_NEW_DETECTED_TAG);
671 for (i = 0; i < new_detected; i++) {
672 if (!entries[i])
673 continue;
674
675 ret = umc_v12_0_fill_error_record(adev, entries[i], ras_error_status);
676 if (ret) {
677 dev_err(adev->dev, "Fail to fill umc error record, ret:%d\n", ret);
678 break;
679 }
680 radix_tree_tag_clear(ecc_tree,
681 entries[i]->pa_pfn, UMC_ECC_NEW_DETECTED_TAG);
682 }
683 mutex_unlock(&con->umc_ecc_log.lock);
684 }
685
umc_v12_0_get_die_id(struct amdgpu_device * adev,uint64_t mca_addr,uint64_t retired_page)686 static uint32_t umc_v12_0_get_die_id(struct amdgpu_device *adev,
687 uint64_t mca_addr, uint64_t retired_page)
688 {
689 uint32_t die = 0;
690
691 /* we only calculate die id for nps1 mode right now */
692 die += ((((retired_page >> 12) & 0x1ULL)^
693 ((retired_page >> 20) & 0x1ULL) ^
694 ((retired_page >> 27) & 0x1ULL) ^
695 ((retired_page >> 34) & 0x1ULL) ^
696 ((retired_page >> 41) & 0x1ULL)) << 0);
697
698 /* the original PA_C4 and PA_R13 may be cleared in retired_page, so
699 * get them from mca_addr.
700 */
701 die += ((((retired_page >> 13) & 0x1ULL) ^
702 ((mca_addr >> 5) & 0x1ULL) ^
703 ((retired_page >> 28) & 0x1ULL) ^
704 ((mca_addr >> 23) & 0x1ULL) ^
705 ((retired_page >> 42) & 0x1ULL)) << 1);
706 die &= 3;
707
708 return die;
709 }
710
711 struct amdgpu_umc_ras umc_v12_0_ras = {
712 .ras_block = {
713 .hw_ops = &umc_v12_0_ras_hw_ops,
714 .ras_late_init = umc_v12_0_ras_late_init,
715 },
716 .err_cnt_init = umc_v12_0_err_cnt_init,
717 .query_ras_poison_mode = umc_v12_0_query_ras_poison_mode,
718 .ecc_info_query_ras_error_address = umc_v12_0_query_ras_ecc_err_addr,
719 .check_ecc_err_status = umc_v12_0_check_ecc_err_status,
720 .update_ecc_status = umc_v12_0_update_ecc_status,
721 .convert_ras_err_addr = umc_v12_0_convert_error_address,
722 .get_die_id_from_pa = umc_v12_0_get_die_id,
723 .get_retire_flip_bits = umc_v12_0_get_retire_flip_bits,
724 };
725
726