1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Copyright 2014 IBM Corp.
4 */
5
6 #include <linux/spinlock.h>
7 #include <linux/sched.h>
8 #include <linux/sched/clock.h>
9 #include <linux/slab.h>
10 #include <linux/mutex.h>
11 #include <linux/mm.h>
12 #include <linux/uaccess.h>
13 #include <linux/delay.h>
14 #include <linux/irqdomain.h>
15 #include <asm/synch.h>
16 #include <asm/switch_to.h>
17 #include <misc/cxl-base.h>
18
19 #include "cxl.h"
20 #include "trace.h"
21
afu_control(struct cxl_afu * afu,u64 command,u64 clear,u64 result,u64 mask,bool enabled)22 static int afu_control(struct cxl_afu *afu, u64 command, u64 clear,
23 u64 result, u64 mask, bool enabled)
24 {
25 u64 AFU_Cntl;
26 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
27 int rc = 0;
28
29 spin_lock(&afu->afu_cntl_lock);
30 pr_devel("AFU command starting: %llx\n", command);
31
32 trace_cxl_afu_ctrl(afu, command);
33
34 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
35 cxl_p2n_write(afu, CXL_AFU_Cntl_An, (AFU_Cntl & ~clear) | command);
36
37 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
38 while ((AFU_Cntl & mask) != result) {
39 if (time_after_eq(jiffies, timeout)) {
40 dev_warn(&afu->dev, "WARNING: AFU control timed out!\n");
41 rc = -EBUSY;
42 goto out;
43 }
44
45 if (!cxl_ops->link_ok(afu->adapter, afu)) {
46 afu->enabled = enabled;
47 rc = -EIO;
48 goto out;
49 }
50
51 pr_devel_ratelimited("AFU control... (0x%016llx)\n",
52 AFU_Cntl | command);
53 cpu_relax();
54 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
55 }
56
57 if (AFU_Cntl & CXL_AFU_Cntl_An_RA) {
58 /*
59 * Workaround for a bug in the XSL used in the Mellanox CX4
60 * that fails to clear the RA bit after an AFU reset,
61 * preventing subsequent AFU resets from working.
62 */
63 cxl_p2n_write(afu, CXL_AFU_Cntl_An, AFU_Cntl & ~CXL_AFU_Cntl_An_RA);
64 }
65
66 pr_devel("AFU command complete: %llx\n", command);
67 afu->enabled = enabled;
68 out:
69 trace_cxl_afu_ctrl_done(afu, command, rc);
70 spin_unlock(&afu->afu_cntl_lock);
71
72 return rc;
73 }
74
afu_enable(struct cxl_afu * afu)75 static int afu_enable(struct cxl_afu *afu)
76 {
77 pr_devel("AFU enable request\n");
78
79 return afu_control(afu, CXL_AFU_Cntl_An_E, 0,
80 CXL_AFU_Cntl_An_ES_Enabled,
81 CXL_AFU_Cntl_An_ES_MASK, true);
82 }
83
cxl_afu_disable(struct cxl_afu * afu)84 int cxl_afu_disable(struct cxl_afu *afu)
85 {
86 pr_devel("AFU disable request\n");
87
88 return afu_control(afu, 0, CXL_AFU_Cntl_An_E,
89 CXL_AFU_Cntl_An_ES_Disabled,
90 CXL_AFU_Cntl_An_ES_MASK, false);
91 }
92
93 /* This will disable as well as reset */
native_afu_reset(struct cxl_afu * afu)94 static int native_afu_reset(struct cxl_afu *afu)
95 {
96 int rc;
97 u64 serr;
98
99 pr_devel("AFU reset request\n");
100
101 rc = afu_control(afu, CXL_AFU_Cntl_An_RA, 0,
102 CXL_AFU_Cntl_An_RS_Complete | CXL_AFU_Cntl_An_ES_Disabled,
103 CXL_AFU_Cntl_An_RS_MASK | CXL_AFU_Cntl_An_ES_MASK,
104 false);
105
106 /*
107 * Re-enable any masked interrupts when the AFU is not
108 * activated to avoid side effects after attaching a process
109 * in dedicated mode.
110 */
111 if (afu->current_mode == 0) {
112 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
113 serr &= ~CXL_PSL_SERR_An_IRQ_MASKS;
114 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
115 }
116
117 return rc;
118 }
119
native_afu_check_and_enable(struct cxl_afu * afu)120 static int native_afu_check_and_enable(struct cxl_afu *afu)
121 {
122 if (!cxl_ops->link_ok(afu->adapter, afu)) {
123 WARN(1, "Refusing to enable afu while link down!\n");
124 return -EIO;
125 }
126 if (afu->enabled)
127 return 0;
128 return afu_enable(afu);
129 }
130
cxl_psl_purge(struct cxl_afu * afu)131 int cxl_psl_purge(struct cxl_afu *afu)
132 {
133 u64 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An);
134 u64 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
135 u64 dsisr, dar;
136 u64 start, end;
137 u64 trans_fault = 0x0ULL;
138 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
139 int rc = 0;
140
141 trace_cxl_psl_ctrl(afu, CXL_PSL_SCNTL_An_Pc);
142
143 pr_devel("PSL purge request\n");
144
145 if (cxl_is_power8())
146 trans_fault = CXL_PSL_DSISR_TRANS;
147 if (cxl_is_power9())
148 trans_fault = CXL_PSL9_DSISR_An_TF;
149
150 if (!cxl_ops->link_ok(afu->adapter, afu)) {
151 dev_warn(&afu->dev, "PSL Purge called with link down, ignoring\n");
152 rc = -EIO;
153 goto out;
154 }
155
156 if ((AFU_Cntl & CXL_AFU_Cntl_An_ES_MASK) != CXL_AFU_Cntl_An_ES_Disabled) {
157 WARN(1, "psl_purge request while AFU not disabled!\n");
158 cxl_afu_disable(afu);
159 }
160
161 cxl_p1n_write(afu, CXL_PSL_SCNTL_An,
162 PSL_CNTL | CXL_PSL_SCNTL_An_Pc);
163 start = local_clock();
164 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An);
165 while ((PSL_CNTL & CXL_PSL_SCNTL_An_Ps_MASK)
166 == CXL_PSL_SCNTL_An_Ps_Pending) {
167 if (time_after_eq(jiffies, timeout)) {
168 dev_warn(&afu->dev, "WARNING: PSL Purge timed out!\n");
169 rc = -EBUSY;
170 goto out;
171 }
172 if (!cxl_ops->link_ok(afu->adapter, afu)) {
173 rc = -EIO;
174 goto out;
175 }
176
177 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
178 pr_devel_ratelimited("PSL purging... PSL_CNTL: 0x%016llx PSL_DSISR: 0x%016llx\n",
179 PSL_CNTL, dsisr);
180
181 if (dsisr & trans_fault) {
182 dar = cxl_p2n_read(afu, CXL_PSL_DAR_An);
183 dev_notice(&afu->dev, "PSL purge terminating pending translation, DSISR: 0x%016llx, DAR: 0x%016llx\n",
184 dsisr, dar);
185 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
186 } else if (dsisr) {
187 dev_notice(&afu->dev, "PSL purge acknowledging pending non-translation fault, DSISR: 0x%016llx\n",
188 dsisr);
189 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
190 } else {
191 cpu_relax();
192 }
193 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An);
194 }
195 end = local_clock();
196 pr_devel("PSL purged in %lld ns\n", end - start);
197
198 cxl_p1n_write(afu, CXL_PSL_SCNTL_An,
199 PSL_CNTL & ~CXL_PSL_SCNTL_An_Pc);
200 out:
201 trace_cxl_psl_ctrl_done(afu, CXL_PSL_SCNTL_An_Pc, rc);
202 return rc;
203 }
204
spa_max_procs(int spa_size)205 static int spa_max_procs(int spa_size)
206 {
207 /*
208 * From the CAIA:
209 * end_of_SPA_area = SPA_Base + ((n+4) * 128) + (( ((n*8) + 127) >> 7) * 128) + 255
210 * Most of that junk is really just an overly-complicated way of saying
211 * the last 256 bytes are __aligned(128), so it's really:
212 * end_of_SPA_area = end_of_PSL_queue_area + __aligned(128) 255
213 * and
214 * end_of_PSL_queue_area = SPA_Base + ((n+4) * 128) + (n*8) - 1
215 * so
216 * sizeof(SPA) = ((n+4) * 128) + (n*8) + __aligned(128) 256
217 * Ignore the alignment (which is safe in this case as long as we are
218 * careful with our rounding) and solve for n:
219 */
220 return ((spa_size / 8) - 96) / 17;
221 }
222
cxl_alloc_spa(struct cxl_afu * afu,int mode)223 static int cxl_alloc_spa(struct cxl_afu *afu, int mode)
224 {
225 unsigned spa_size;
226
227 /* Work out how many pages to allocate */
228 afu->native->spa_order = -1;
229 do {
230 afu->native->spa_order++;
231 spa_size = (1 << afu->native->spa_order) * PAGE_SIZE;
232
233 if (spa_size > 0x100000) {
234 dev_warn(&afu->dev, "num_of_processes too large for the SPA, limiting to %i (0x%x)\n",
235 afu->native->spa_max_procs, afu->native->spa_size);
236 if (mode != CXL_MODE_DEDICATED)
237 afu->num_procs = afu->native->spa_max_procs;
238 break;
239 }
240
241 afu->native->spa_size = spa_size;
242 afu->native->spa_max_procs = spa_max_procs(afu->native->spa_size);
243 } while (afu->native->spa_max_procs < afu->num_procs);
244
245 if (!(afu->native->spa = (struct cxl_process_element *)
246 __get_free_pages(GFP_KERNEL | __GFP_ZERO, afu->native->spa_order))) {
247 pr_err("cxl_alloc_spa: Unable to allocate scheduled process area\n");
248 return -ENOMEM;
249 }
250 pr_devel("spa pages: %i afu->spa_max_procs: %i afu->num_procs: %i\n",
251 1<<afu->native->spa_order, afu->native->spa_max_procs, afu->num_procs);
252
253 return 0;
254 }
255
attach_spa(struct cxl_afu * afu)256 static void attach_spa(struct cxl_afu *afu)
257 {
258 u64 spap;
259
260 afu->native->sw_command_status = (__be64 *)((char *)afu->native->spa +
261 ((afu->native->spa_max_procs + 3) * 128));
262
263 spap = virt_to_phys(afu->native->spa) & CXL_PSL_SPAP_Addr;
264 spap |= ((afu->native->spa_size >> (12 - CXL_PSL_SPAP_Size_Shift)) - 1) & CXL_PSL_SPAP_Size;
265 spap |= CXL_PSL_SPAP_V;
266 pr_devel("cxl: SPA allocated at 0x%p. Max processes: %i, sw_command_status: 0x%p CXL_PSL_SPAP_An=0x%016llx\n",
267 afu->native->spa, afu->native->spa_max_procs,
268 afu->native->sw_command_status, spap);
269 cxl_p1n_write(afu, CXL_PSL_SPAP_An, spap);
270 }
271
detach_spa(struct cxl_afu * afu)272 static inline void detach_spa(struct cxl_afu *afu)
273 {
274 cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0);
275 }
276
cxl_release_spa(struct cxl_afu * afu)277 void cxl_release_spa(struct cxl_afu *afu)
278 {
279 if (afu->native->spa) {
280 free_pages((unsigned long) afu->native->spa,
281 afu->native->spa_order);
282 afu->native->spa = NULL;
283 }
284 }
285
286 /*
287 * Invalidation of all ERAT entries is no longer required by CAIA2. Use
288 * only for debug.
289 */
cxl_invalidate_all_psl9(struct cxl * adapter)290 int cxl_invalidate_all_psl9(struct cxl *adapter)
291 {
292 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
293 u64 ierat;
294
295 pr_devel("CXL adapter - invalidation of all ERAT entries\n");
296
297 /* Invalidates all ERAT entries for Radix or HPT */
298 ierat = CXL_XSL9_IERAT_IALL;
299 if (radix_enabled())
300 ierat |= CXL_XSL9_IERAT_INVR;
301 cxl_p1_write(adapter, CXL_XSL9_IERAT, ierat);
302
303 while (cxl_p1_read(adapter, CXL_XSL9_IERAT) & CXL_XSL9_IERAT_IINPROG) {
304 if (time_after_eq(jiffies, timeout)) {
305 dev_warn(&adapter->dev,
306 "WARNING: CXL adapter invalidation of all ERAT entries timed out!\n");
307 return -EBUSY;
308 }
309 if (!cxl_ops->link_ok(adapter, NULL))
310 return -EIO;
311 cpu_relax();
312 }
313 return 0;
314 }
315
cxl_invalidate_all_psl8(struct cxl * adapter)316 int cxl_invalidate_all_psl8(struct cxl *adapter)
317 {
318 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
319
320 pr_devel("CXL adapter wide TLBIA & SLBIA\n");
321
322 cxl_p1_write(adapter, CXL_PSL_AFUSEL, CXL_PSL_AFUSEL_A);
323
324 cxl_p1_write(adapter, CXL_PSL_TLBIA, CXL_TLB_SLB_IQ_ALL);
325 while (cxl_p1_read(adapter, CXL_PSL_TLBIA) & CXL_TLB_SLB_P) {
326 if (time_after_eq(jiffies, timeout)) {
327 dev_warn(&adapter->dev, "WARNING: CXL adapter wide TLBIA timed out!\n");
328 return -EBUSY;
329 }
330 if (!cxl_ops->link_ok(adapter, NULL))
331 return -EIO;
332 cpu_relax();
333 }
334
335 cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_ALL);
336 while (cxl_p1_read(adapter, CXL_PSL_SLBIA) & CXL_TLB_SLB_P) {
337 if (time_after_eq(jiffies, timeout)) {
338 dev_warn(&adapter->dev, "WARNING: CXL adapter wide SLBIA timed out!\n");
339 return -EBUSY;
340 }
341 if (!cxl_ops->link_ok(adapter, NULL))
342 return -EIO;
343 cpu_relax();
344 }
345 return 0;
346 }
347
cxl_data_cache_flush(struct cxl * adapter)348 int cxl_data_cache_flush(struct cxl *adapter)
349 {
350 u64 reg;
351 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
352
353 /*
354 * Do a datacache flush only if datacache is available.
355 * In case of PSL9D datacache absent hence flush operation.
356 * would timeout.
357 */
358 if (adapter->native->no_data_cache) {
359 pr_devel("No PSL data cache. Ignoring cache flush req.\n");
360 return 0;
361 }
362
363 pr_devel("Flushing data cache\n");
364 reg = cxl_p1_read(adapter, CXL_PSL_Control);
365 reg |= CXL_PSL_Control_Fr;
366 cxl_p1_write(adapter, CXL_PSL_Control, reg);
367
368 reg = cxl_p1_read(adapter, CXL_PSL_Control);
369 while ((reg & CXL_PSL_Control_Fs_MASK) != CXL_PSL_Control_Fs_Complete) {
370 if (time_after_eq(jiffies, timeout)) {
371 dev_warn(&adapter->dev, "WARNING: cache flush timed out!\n");
372 return -EBUSY;
373 }
374
375 if (!cxl_ops->link_ok(adapter, NULL)) {
376 dev_warn(&adapter->dev, "WARNING: link down when flushing cache\n");
377 return -EIO;
378 }
379 cpu_relax();
380 reg = cxl_p1_read(adapter, CXL_PSL_Control);
381 }
382
383 reg &= ~CXL_PSL_Control_Fr;
384 cxl_p1_write(adapter, CXL_PSL_Control, reg);
385 return 0;
386 }
387
cxl_write_sstp(struct cxl_afu * afu,u64 sstp0,u64 sstp1)388 static int cxl_write_sstp(struct cxl_afu *afu, u64 sstp0, u64 sstp1)
389 {
390 int rc;
391
392 /* 1. Disable SSTP by writing 0 to SSTP1[V] */
393 cxl_p2n_write(afu, CXL_SSTP1_An, 0);
394
395 /* 2. Invalidate all SLB entries */
396 if ((rc = cxl_afu_slbia(afu)))
397 return rc;
398
399 /* 3. Set SSTP0_An */
400 cxl_p2n_write(afu, CXL_SSTP0_An, sstp0);
401
402 /* 4. Set SSTP1_An */
403 cxl_p2n_write(afu, CXL_SSTP1_An, sstp1);
404
405 return 0;
406 }
407
408 /* Using per slice version may improve performance here. (ie. SLBIA_An) */
slb_invalid(struct cxl_context * ctx)409 static void slb_invalid(struct cxl_context *ctx)
410 {
411 struct cxl *adapter = ctx->afu->adapter;
412 u64 slbia;
413
414 WARN_ON(!mutex_is_locked(&ctx->afu->native->spa_mutex));
415
416 cxl_p1_write(adapter, CXL_PSL_LBISEL,
417 ((u64)be32_to_cpu(ctx->elem->common.pid) << 32) |
418 be32_to_cpu(ctx->elem->lpid));
419 cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_LPIDPID);
420
421 while (1) {
422 if (!cxl_ops->link_ok(adapter, NULL))
423 break;
424 slbia = cxl_p1_read(adapter, CXL_PSL_SLBIA);
425 if (!(slbia & CXL_TLB_SLB_P))
426 break;
427 cpu_relax();
428 }
429 }
430
do_process_element_cmd(struct cxl_context * ctx,u64 cmd,u64 pe_state)431 static int do_process_element_cmd(struct cxl_context *ctx,
432 u64 cmd, u64 pe_state)
433 {
434 u64 state;
435 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT);
436 int rc = 0;
437
438 trace_cxl_llcmd(ctx, cmd);
439
440 WARN_ON(!ctx->afu->enabled);
441
442 ctx->elem->software_state = cpu_to_be32(pe_state);
443 smp_wmb();
444 *(ctx->afu->native->sw_command_status) = cpu_to_be64(cmd | 0 | ctx->pe);
445 smp_mb();
446 cxl_p1n_write(ctx->afu, CXL_PSL_LLCMD_An, cmd | ctx->pe);
447 while (1) {
448 if (time_after_eq(jiffies, timeout)) {
449 dev_warn(&ctx->afu->dev, "WARNING: Process Element Command timed out!\n");
450 rc = -EBUSY;
451 goto out;
452 }
453 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
454 dev_warn(&ctx->afu->dev, "WARNING: Device link down, aborting Process Element Command!\n");
455 rc = -EIO;
456 goto out;
457 }
458 state = be64_to_cpup(ctx->afu->native->sw_command_status);
459 if (state == ~0ULL) {
460 pr_err("cxl: Error adding process element to AFU\n");
461 rc = -1;
462 goto out;
463 }
464 if ((state & (CXL_SPA_SW_CMD_MASK | CXL_SPA_SW_STATE_MASK | CXL_SPA_SW_LINK_MASK)) ==
465 (cmd | (cmd >> 16) | ctx->pe))
466 break;
467 /*
468 * The command won't finish in the PSL if there are
469 * outstanding DSIs. Hence we need to yield here in
470 * case there are outstanding DSIs that we need to
471 * service. Tuning possiblity: we could wait for a
472 * while before sched
473 */
474 schedule();
475
476 }
477 out:
478 trace_cxl_llcmd_done(ctx, cmd, rc);
479 return rc;
480 }
481
add_process_element(struct cxl_context * ctx)482 static int add_process_element(struct cxl_context *ctx)
483 {
484 int rc = 0;
485
486 mutex_lock(&ctx->afu->native->spa_mutex);
487 pr_devel("%s Adding pe: %i started\n", __func__, ctx->pe);
488 if (!(rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_ADD, CXL_PE_SOFTWARE_STATE_V)))
489 ctx->pe_inserted = true;
490 pr_devel("%s Adding pe: %i finished\n", __func__, ctx->pe);
491 mutex_unlock(&ctx->afu->native->spa_mutex);
492 return rc;
493 }
494
terminate_process_element(struct cxl_context * ctx)495 static int terminate_process_element(struct cxl_context *ctx)
496 {
497 int rc = 0;
498
499 /* fast path terminate if it's already invalid */
500 if (!(ctx->elem->software_state & cpu_to_be32(CXL_PE_SOFTWARE_STATE_V)))
501 return rc;
502
503 mutex_lock(&ctx->afu->native->spa_mutex);
504 pr_devel("%s Terminate pe: %i started\n", __func__, ctx->pe);
505 /* We could be asked to terminate when the hw is down. That
506 * should always succeed: it's not running if the hw has gone
507 * away and is being reset.
508 */
509 if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
510 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_TERMINATE,
511 CXL_PE_SOFTWARE_STATE_V | CXL_PE_SOFTWARE_STATE_T);
512 ctx->elem->software_state = 0; /* Remove Valid bit */
513 pr_devel("%s Terminate pe: %i finished\n", __func__, ctx->pe);
514 mutex_unlock(&ctx->afu->native->spa_mutex);
515 return rc;
516 }
517
remove_process_element(struct cxl_context * ctx)518 static int remove_process_element(struct cxl_context *ctx)
519 {
520 int rc = 0;
521
522 mutex_lock(&ctx->afu->native->spa_mutex);
523 pr_devel("%s Remove pe: %i started\n", __func__, ctx->pe);
524
525 /* We could be asked to remove when the hw is down. Again, if
526 * the hw is down, the PE is gone, so we succeed.
527 */
528 if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
529 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_REMOVE, 0);
530
531 if (!rc)
532 ctx->pe_inserted = false;
533 if (cxl_is_power8())
534 slb_invalid(ctx);
535 pr_devel("%s Remove pe: %i finished\n", __func__, ctx->pe);
536 mutex_unlock(&ctx->afu->native->spa_mutex);
537
538 return rc;
539 }
540
cxl_assign_psn_space(struct cxl_context * ctx)541 void cxl_assign_psn_space(struct cxl_context *ctx)
542 {
543 if (!ctx->afu->pp_size || ctx->master) {
544 ctx->psn_phys = ctx->afu->psn_phys;
545 ctx->psn_size = ctx->afu->adapter->ps_size;
546 } else {
547 ctx->psn_phys = ctx->afu->psn_phys +
548 (ctx->afu->native->pp_offset + ctx->afu->pp_size * ctx->pe);
549 ctx->psn_size = ctx->afu->pp_size;
550 }
551 }
552
activate_afu_directed(struct cxl_afu * afu)553 static int activate_afu_directed(struct cxl_afu *afu)
554 {
555 int rc;
556
557 dev_info(&afu->dev, "Activating AFU directed mode\n");
558
559 afu->num_procs = afu->max_procs_virtualised;
560 if (afu->native->spa == NULL) {
561 if (cxl_alloc_spa(afu, CXL_MODE_DIRECTED))
562 return -ENOMEM;
563 }
564 attach_spa(afu);
565
566 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_AFU);
567 if (cxl_is_power8())
568 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL);
569 cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L);
570
571 afu->current_mode = CXL_MODE_DIRECTED;
572
573 if ((rc = cxl_chardev_m_afu_add(afu)))
574 return rc;
575
576 if ((rc = cxl_sysfs_afu_m_add(afu)))
577 goto err;
578
579 if ((rc = cxl_chardev_s_afu_add(afu)))
580 goto err1;
581
582 return 0;
583 err1:
584 cxl_sysfs_afu_m_remove(afu);
585 err:
586 cxl_chardev_afu_remove(afu);
587 return rc;
588 }
589
590 #ifdef CONFIG_CPU_LITTLE_ENDIAN
591 #define set_endian(sr) ((sr) |= CXL_PSL_SR_An_LE)
592 #else
593 #define set_endian(sr) ((sr) &= ~(CXL_PSL_SR_An_LE))
594 #endif
595
cxl_calculate_sr(bool master,bool kernel,bool real_mode,bool p9)596 u64 cxl_calculate_sr(bool master, bool kernel, bool real_mode, bool p9)
597 {
598 u64 sr = 0;
599
600 set_endian(sr);
601 if (master)
602 sr |= CXL_PSL_SR_An_MP;
603 if (mfspr(SPRN_LPCR) & LPCR_TC)
604 sr |= CXL_PSL_SR_An_TC;
605
606 if (kernel) {
607 if (!real_mode)
608 sr |= CXL_PSL_SR_An_R;
609 sr |= (mfmsr() & MSR_SF) | CXL_PSL_SR_An_HV;
610 } else {
611 sr |= CXL_PSL_SR_An_PR | CXL_PSL_SR_An_R;
612 if (radix_enabled())
613 sr |= CXL_PSL_SR_An_HV;
614 else
615 sr &= ~(CXL_PSL_SR_An_HV);
616 if (!test_tsk_thread_flag(current, TIF_32BIT))
617 sr |= CXL_PSL_SR_An_SF;
618 }
619 if (p9) {
620 if (radix_enabled())
621 sr |= CXL_PSL_SR_An_XLAT_ror;
622 else
623 sr |= CXL_PSL_SR_An_XLAT_hpt;
624 }
625 return sr;
626 }
627
calculate_sr(struct cxl_context * ctx)628 static u64 calculate_sr(struct cxl_context *ctx)
629 {
630 return cxl_calculate_sr(ctx->master, ctx->kernel, false,
631 cxl_is_power9());
632 }
633
update_ivtes_directed(struct cxl_context * ctx)634 static void update_ivtes_directed(struct cxl_context *ctx)
635 {
636 bool need_update = (ctx->status == STARTED);
637 int r;
638
639 if (need_update) {
640 WARN_ON(terminate_process_element(ctx));
641 WARN_ON(remove_process_element(ctx));
642 }
643
644 for (r = 0; r < CXL_IRQ_RANGES; r++) {
645 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]);
646 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]);
647 }
648
649 /*
650 * Theoretically we could use the update llcmd, instead of a
651 * terminate/remove/add (or if an atomic update was required we could
652 * do a suspend/update/resume), however it seems there might be issues
653 * with the update llcmd on some cards (including those using an XSL on
654 * an ASIC) so for now it's safest to go with the commands that are
655 * known to work. In the future if we come across a situation where the
656 * card may be performing transactions using the same PE while we are
657 * doing this update we might need to revisit this.
658 */
659 if (need_update)
660 WARN_ON(add_process_element(ctx));
661 }
662
process_element_entry_psl9(struct cxl_context * ctx,u64 wed,u64 amr)663 static int process_element_entry_psl9(struct cxl_context *ctx, u64 wed, u64 amr)
664 {
665 u32 pid;
666 int rc;
667
668 cxl_assign_psn_space(ctx);
669
670 ctx->elem->ctxtime = 0; /* disable */
671 ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID));
672 ctx->elem->haurp = 0; /* disable */
673
674 if (ctx->kernel)
675 pid = 0;
676 else {
677 if (ctx->mm == NULL) {
678 pr_devel("%s: unable to get mm for pe=%d pid=%i\n",
679 __func__, ctx->pe, pid_nr(ctx->pid));
680 return -EINVAL;
681 }
682 pid = ctx->mm->context.id;
683 }
684
685 /* Assign a unique TIDR (thread id) for the current thread */
686 if (!(ctx->tidr) && (ctx->assign_tidr)) {
687 rc = set_thread_tidr(current);
688 if (rc)
689 return -ENODEV;
690 ctx->tidr = current->thread.tidr;
691 pr_devel("%s: current tidr: %d\n", __func__, ctx->tidr);
692 }
693
694 ctx->elem->common.tid = cpu_to_be32(ctx->tidr);
695 ctx->elem->common.pid = cpu_to_be32(pid);
696
697 ctx->elem->sr = cpu_to_be64(calculate_sr(ctx));
698
699 ctx->elem->common.csrp = 0; /* disable */
700
701 cxl_prefault(ctx, wed);
702
703 /*
704 * Ensure we have the multiplexed PSL interrupt set up to take faults
705 * for kernel contexts that may not have allocated any AFU IRQs at all:
706 */
707 if (ctx->irqs.range[0] == 0) {
708 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq;
709 ctx->irqs.range[0] = 1;
710 }
711
712 ctx->elem->common.amr = cpu_to_be64(amr);
713 ctx->elem->common.wed = cpu_to_be64(wed);
714
715 return 0;
716 }
717
cxl_attach_afu_directed_psl9(struct cxl_context * ctx,u64 wed,u64 amr)718 int cxl_attach_afu_directed_psl9(struct cxl_context *ctx, u64 wed, u64 amr)
719 {
720 int result;
721
722 /* fill the process element entry */
723 result = process_element_entry_psl9(ctx, wed, amr);
724 if (result)
725 return result;
726
727 update_ivtes_directed(ctx);
728
729 /* first guy needs to enable */
730 result = cxl_ops->afu_check_and_enable(ctx->afu);
731 if (result)
732 return result;
733
734 return add_process_element(ctx);
735 }
736
cxl_attach_afu_directed_psl8(struct cxl_context * ctx,u64 wed,u64 amr)737 int cxl_attach_afu_directed_psl8(struct cxl_context *ctx, u64 wed, u64 amr)
738 {
739 u32 pid;
740 int result;
741
742 cxl_assign_psn_space(ctx);
743
744 ctx->elem->ctxtime = 0; /* disable */
745 ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID));
746 ctx->elem->haurp = 0; /* disable */
747 ctx->elem->u.sdr = cpu_to_be64(mfspr(SPRN_SDR1));
748
749 pid = current->pid;
750 if (ctx->kernel)
751 pid = 0;
752 ctx->elem->common.tid = 0;
753 ctx->elem->common.pid = cpu_to_be32(pid);
754
755 ctx->elem->sr = cpu_to_be64(calculate_sr(ctx));
756
757 ctx->elem->common.csrp = 0; /* disable */
758 ctx->elem->common.u.psl8.aurp0 = 0; /* disable */
759 ctx->elem->common.u.psl8.aurp1 = 0; /* disable */
760
761 cxl_prefault(ctx, wed);
762
763 ctx->elem->common.u.psl8.sstp0 = cpu_to_be64(ctx->sstp0);
764 ctx->elem->common.u.psl8.sstp1 = cpu_to_be64(ctx->sstp1);
765
766 /*
767 * Ensure we have the multiplexed PSL interrupt set up to take faults
768 * for kernel contexts that may not have allocated any AFU IRQs at all:
769 */
770 if (ctx->irqs.range[0] == 0) {
771 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq;
772 ctx->irqs.range[0] = 1;
773 }
774
775 update_ivtes_directed(ctx);
776
777 ctx->elem->common.amr = cpu_to_be64(amr);
778 ctx->elem->common.wed = cpu_to_be64(wed);
779
780 /* first guy needs to enable */
781 if ((result = cxl_ops->afu_check_and_enable(ctx->afu)))
782 return result;
783
784 return add_process_element(ctx);
785 }
786
deactivate_afu_directed(struct cxl_afu * afu)787 static int deactivate_afu_directed(struct cxl_afu *afu)
788 {
789 dev_info(&afu->dev, "Deactivating AFU directed mode\n");
790
791 afu->current_mode = 0;
792 afu->num_procs = 0;
793
794 cxl_sysfs_afu_m_remove(afu);
795 cxl_chardev_afu_remove(afu);
796
797 /*
798 * The CAIA section 2.2.1 indicates that the procedure for starting and
799 * stopping an AFU in AFU directed mode is AFU specific, which is not
800 * ideal since this code is generic and with one exception has no
801 * knowledge of the AFU. This is in contrast to the procedure for
802 * disabling a dedicated process AFU, which is documented to just
803 * require a reset. The architecture does indicate that both an AFU
804 * reset and an AFU disable should result in the AFU being disabled and
805 * we do both followed by a PSL purge for safety.
806 *
807 * Notably we used to have some issues with the disable sequence on PSL
808 * cards, which is why we ended up using this heavy weight procedure in
809 * the first place, however a bug was discovered that had rendered the
810 * disable operation ineffective, so it is conceivable that was the
811 * sole explanation for those difficulties. Careful regression testing
812 * is recommended if anyone attempts to remove or reorder these
813 * operations.
814 *
815 * The XSL on the Mellanox CX4 behaves a little differently from the
816 * PSL based cards and will time out an AFU reset if the AFU is still
817 * enabled. That card is special in that we do have a means to identify
818 * it from this code, so in that case we skip the reset and just use a
819 * disable/purge to avoid the timeout and corresponding noise in the
820 * kernel log.
821 */
822 if (afu->adapter->native->sl_ops->needs_reset_before_disable)
823 cxl_ops->afu_reset(afu);
824 cxl_afu_disable(afu);
825 cxl_psl_purge(afu);
826
827 return 0;
828 }
829
cxl_activate_dedicated_process_psl9(struct cxl_afu * afu)830 int cxl_activate_dedicated_process_psl9(struct cxl_afu *afu)
831 {
832 dev_info(&afu->dev, "Activating dedicated process mode\n");
833
834 /*
835 * If XSL is set to dedicated mode (Set in PSL_SCNTL reg), the
836 * XSL and AFU are programmed to work with a single context.
837 * The context information should be configured in the SPA area
838 * index 0 (so PSL_SPAP must be configured before enabling the
839 * AFU).
840 */
841 afu->num_procs = 1;
842 if (afu->native->spa == NULL) {
843 if (cxl_alloc_spa(afu, CXL_MODE_DEDICATED))
844 return -ENOMEM;
845 }
846 attach_spa(afu);
847
848 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process);
849 cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L);
850
851 afu->current_mode = CXL_MODE_DEDICATED;
852
853 return cxl_chardev_d_afu_add(afu);
854 }
855
cxl_activate_dedicated_process_psl8(struct cxl_afu * afu)856 int cxl_activate_dedicated_process_psl8(struct cxl_afu *afu)
857 {
858 dev_info(&afu->dev, "Activating dedicated process mode\n");
859
860 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process);
861
862 cxl_p1n_write(afu, CXL_PSL_CtxTime_An, 0); /* disable */
863 cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0); /* disable */
864 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL);
865 cxl_p1n_write(afu, CXL_PSL_LPID_An, mfspr(SPRN_LPID));
866 cxl_p1n_write(afu, CXL_HAURP_An, 0); /* disable */
867 cxl_p1n_write(afu, CXL_PSL_SDR_An, mfspr(SPRN_SDR1));
868
869 cxl_p2n_write(afu, CXL_CSRP_An, 0); /* disable */
870 cxl_p2n_write(afu, CXL_AURP0_An, 0); /* disable */
871 cxl_p2n_write(afu, CXL_AURP1_An, 0); /* disable */
872
873 afu->current_mode = CXL_MODE_DEDICATED;
874 afu->num_procs = 1;
875
876 return cxl_chardev_d_afu_add(afu);
877 }
878
cxl_update_dedicated_ivtes_psl9(struct cxl_context * ctx)879 void cxl_update_dedicated_ivtes_psl9(struct cxl_context *ctx)
880 {
881 int r;
882
883 for (r = 0; r < CXL_IRQ_RANGES; r++) {
884 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]);
885 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]);
886 }
887 }
888
cxl_update_dedicated_ivtes_psl8(struct cxl_context * ctx)889 void cxl_update_dedicated_ivtes_psl8(struct cxl_context *ctx)
890 {
891 struct cxl_afu *afu = ctx->afu;
892
893 cxl_p1n_write(afu, CXL_PSL_IVTE_Offset_An,
894 (((u64)ctx->irqs.offset[0] & 0xffff) << 48) |
895 (((u64)ctx->irqs.offset[1] & 0xffff) << 32) |
896 (((u64)ctx->irqs.offset[2] & 0xffff) << 16) |
897 ((u64)ctx->irqs.offset[3] & 0xffff));
898 cxl_p1n_write(afu, CXL_PSL_IVTE_Limit_An, (u64)
899 (((u64)ctx->irqs.range[0] & 0xffff) << 48) |
900 (((u64)ctx->irqs.range[1] & 0xffff) << 32) |
901 (((u64)ctx->irqs.range[2] & 0xffff) << 16) |
902 ((u64)ctx->irqs.range[3] & 0xffff));
903 }
904
cxl_attach_dedicated_process_psl9(struct cxl_context * ctx,u64 wed,u64 amr)905 int cxl_attach_dedicated_process_psl9(struct cxl_context *ctx, u64 wed, u64 amr)
906 {
907 struct cxl_afu *afu = ctx->afu;
908 int result;
909
910 /* fill the process element entry */
911 result = process_element_entry_psl9(ctx, wed, amr);
912 if (result)
913 return result;
914
915 if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes)
916 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx);
917
918 ctx->elem->software_state = cpu_to_be32(CXL_PE_SOFTWARE_STATE_V);
919 /*
920 * Ideally we should do a wmb() here to make sure the changes to the
921 * PE are visible to the card before we call afu_enable.
922 * On ppc64 though all mmios are preceded by a 'sync' instruction hence
923 * we dont dont need one here.
924 */
925
926 result = cxl_ops->afu_reset(afu);
927 if (result)
928 return result;
929
930 return afu_enable(afu);
931 }
932
cxl_attach_dedicated_process_psl8(struct cxl_context * ctx,u64 wed,u64 amr)933 int cxl_attach_dedicated_process_psl8(struct cxl_context *ctx, u64 wed, u64 amr)
934 {
935 struct cxl_afu *afu = ctx->afu;
936 u64 pid;
937 int rc;
938
939 pid = (u64)current->pid << 32;
940 if (ctx->kernel)
941 pid = 0;
942 cxl_p2n_write(afu, CXL_PSL_PID_TID_An, pid);
943
944 cxl_p1n_write(afu, CXL_PSL_SR_An, calculate_sr(ctx));
945
946 if ((rc = cxl_write_sstp(afu, ctx->sstp0, ctx->sstp1)))
947 return rc;
948
949 cxl_prefault(ctx, wed);
950
951 if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes)
952 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx);
953
954 cxl_p2n_write(afu, CXL_PSL_AMR_An, amr);
955
956 /* master only context for dedicated */
957 cxl_assign_psn_space(ctx);
958
959 if ((rc = cxl_ops->afu_reset(afu)))
960 return rc;
961
962 cxl_p2n_write(afu, CXL_PSL_WED_An, wed);
963
964 return afu_enable(afu);
965 }
966
deactivate_dedicated_process(struct cxl_afu * afu)967 static int deactivate_dedicated_process(struct cxl_afu *afu)
968 {
969 dev_info(&afu->dev, "Deactivating dedicated process mode\n");
970
971 afu->current_mode = 0;
972 afu->num_procs = 0;
973
974 cxl_chardev_afu_remove(afu);
975
976 return 0;
977 }
978
native_afu_deactivate_mode(struct cxl_afu * afu,int mode)979 static int native_afu_deactivate_mode(struct cxl_afu *afu, int mode)
980 {
981 if (mode == CXL_MODE_DIRECTED)
982 return deactivate_afu_directed(afu);
983 if (mode == CXL_MODE_DEDICATED)
984 return deactivate_dedicated_process(afu);
985 return 0;
986 }
987
native_afu_activate_mode(struct cxl_afu * afu,int mode)988 static int native_afu_activate_mode(struct cxl_afu *afu, int mode)
989 {
990 if (!mode)
991 return 0;
992 if (!(mode & afu->modes_supported))
993 return -EINVAL;
994
995 if (!cxl_ops->link_ok(afu->adapter, afu)) {
996 WARN(1, "Device link is down, refusing to activate!\n");
997 return -EIO;
998 }
999
1000 if (mode == CXL_MODE_DIRECTED)
1001 return activate_afu_directed(afu);
1002 if ((mode == CXL_MODE_DEDICATED) &&
1003 (afu->adapter->native->sl_ops->activate_dedicated_process))
1004 return afu->adapter->native->sl_ops->activate_dedicated_process(afu);
1005
1006 return -EINVAL;
1007 }
1008
native_attach_process(struct cxl_context * ctx,bool kernel,u64 wed,u64 amr)1009 static int native_attach_process(struct cxl_context *ctx, bool kernel,
1010 u64 wed, u64 amr)
1011 {
1012 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
1013 WARN(1, "Device link is down, refusing to attach process!\n");
1014 return -EIO;
1015 }
1016
1017 ctx->kernel = kernel;
1018 if ((ctx->afu->current_mode == CXL_MODE_DIRECTED) &&
1019 (ctx->afu->adapter->native->sl_ops->attach_afu_directed))
1020 return ctx->afu->adapter->native->sl_ops->attach_afu_directed(ctx, wed, amr);
1021
1022 if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) &&
1023 (ctx->afu->adapter->native->sl_ops->attach_dedicated_process))
1024 return ctx->afu->adapter->native->sl_ops->attach_dedicated_process(ctx, wed, amr);
1025
1026 return -EINVAL;
1027 }
1028
detach_process_native_dedicated(struct cxl_context * ctx)1029 static inline int detach_process_native_dedicated(struct cxl_context *ctx)
1030 {
1031 /*
1032 * The CAIA section 2.1.1 indicates that we need to do an AFU reset to
1033 * stop the AFU in dedicated mode (we therefore do not make that
1034 * optional like we do in the afu directed path). It does not indicate
1035 * that we need to do an explicit disable (which should occur
1036 * implicitly as part of the reset) or purge, but we do these as well
1037 * to be on the safe side.
1038 *
1039 * Notably we used to have some issues with the disable sequence
1040 * (before the sequence was spelled out in the architecture) which is
1041 * why we were so heavy weight in the first place, however a bug was
1042 * discovered that had rendered the disable operation ineffective, so
1043 * it is conceivable that was the sole explanation for those
1044 * difficulties. Point is, we should be careful and do some regression
1045 * testing if we ever attempt to remove any part of this procedure.
1046 */
1047 cxl_ops->afu_reset(ctx->afu);
1048 cxl_afu_disable(ctx->afu);
1049 cxl_psl_purge(ctx->afu);
1050 return 0;
1051 }
1052
native_update_ivtes(struct cxl_context * ctx)1053 static void native_update_ivtes(struct cxl_context *ctx)
1054 {
1055 if (ctx->afu->current_mode == CXL_MODE_DIRECTED)
1056 return update_ivtes_directed(ctx);
1057 if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) &&
1058 (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes))
1059 return ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx);
1060 WARN(1, "native_update_ivtes: Bad mode\n");
1061 }
1062
detach_process_native_afu_directed(struct cxl_context * ctx)1063 static inline int detach_process_native_afu_directed(struct cxl_context *ctx)
1064 {
1065 if (!ctx->pe_inserted)
1066 return 0;
1067 if (terminate_process_element(ctx))
1068 return -1;
1069 if (remove_process_element(ctx))
1070 return -1;
1071
1072 return 0;
1073 }
1074
native_detach_process(struct cxl_context * ctx)1075 static int native_detach_process(struct cxl_context *ctx)
1076 {
1077 trace_cxl_detach(ctx);
1078
1079 if (ctx->afu->current_mode == CXL_MODE_DEDICATED)
1080 return detach_process_native_dedicated(ctx);
1081
1082 return detach_process_native_afu_directed(ctx);
1083 }
1084
native_get_irq_info(struct cxl_afu * afu,struct cxl_irq_info * info)1085 static int native_get_irq_info(struct cxl_afu *afu, struct cxl_irq_info *info)
1086 {
1087 /* If the adapter has gone away, we can't get any meaningful
1088 * information.
1089 */
1090 if (!cxl_ops->link_ok(afu->adapter, afu))
1091 return -EIO;
1092
1093 info->dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
1094 info->dar = cxl_p2n_read(afu, CXL_PSL_DAR_An);
1095 if (cxl_is_power8())
1096 info->dsr = cxl_p2n_read(afu, CXL_PSL_DSR_An);
1097 info->afu_err = cxl_p2n_read(afu, CXL_AFU_ERR_An);
1098 info->errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
1099 info->proc_handle = 0;
1100
1101 return 0;
1102 }
1103
cxl_native_irq_dump_regs_psl9(struct cxl_context * ctx)1104 void cxl_native_irq_dump_regs_psl9(struct cxl_context *ctx)
1105 {
1106 u64 fir1, serr;
1107
1108 fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL9_FIR1);
1109
1110 dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1);
1111 if (ctx->afu->adapter->native->sl_ops->register_serr_irq) {
1112 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An);
1113 cxl_afu_decode_psl_serr(ctx->afu, serr);
1114 }
1115 }
1116
cxl_native_irq_dump_regs_psl8(struct cxl_context * ctx)1117 void cxl_native_irq_dump_regs_psl8(struct cxl_context *ctx)
1118 {
1119 u64 fir1, fir2, fir_slice, serr, afu_debug;
1120
1121 fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR1);
1122 fir2 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR2);
1123 fir_slice = cxl_p1n_read(ctx->afu, CXL_PSL_FIR_SLICE_An);
1124 afu_debug = cxl_p1n_read(ctx->afu, CXL_AFU_DEBUG_An);
1125
1126 dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1);
1127 dev_crit(&ctx->afu->dev, "PSL_FIR2: 0x%016llx\n", fir2);
1128 if (ctx->afu->adapter->native->sl_ops->register_serr_irq) {
1129 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An);
1130 cxl_afu_decode_psl_serr(ctx->afu, serr);
1131 }
1132 dev_crit(&ctx->afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice);
1133 dev_crit(&ctx->afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug);
1134 }
1135
native_handle_psl_slice_error(struct cxl_context * ctx,u64 dsisr,u64 errstat)1136 static irqreturn_t native_handle_psl_slice_error(struct cxl_context *ctx,
1137 u64 dsisr, u64 errstat)
1138 {
1139
1140 dev_crit(&ctx->afu->dev, "PSL ERROR STATUS: 0x%016llx\n", errstat);
1141
1142 if (ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers)
1143 ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers(ctx);
1144
1145 if (ctx->afu->adapter->native->sl_ops->debugfs_stop_trace) {
1146 dev_crit(&ctx->afu->dev, "STOPPING CXL TRACE\n");
1147 ctx->afu->adapter->native->sl_ops->debugfs_stop_trace(ctx->afu->adapter);
1148 }
1149
1150 return cxl_ops->ack_irq(ctx, 0, errstat);
1151 }
1152
cxl_is_translation_fault(struct cxl_afu * afu,u64 dsisr)1153 static bool cxl_is_translation_fault(struct cxl_afu *afu, u64 dsisr)
1154 {
1155 if ((cxl_is_power8()) && (dsisr & CXL_PSL_DSISR_TRANS))
1156 return true;
1157
1158 if ((cxl_is_power9()) && (dsisr & CXL_PSL9_DSISR_An_TF))
1159 return true;
1160
1161 return false;
1162 }
1163
cxl_fail_irq_psl(struct cxl_afu * afu,struct cxl_irq_info * irq_info)1164 irqreturn_t cxl_fail_irq_psl(struct cxl_afu *afu, struct cxl_irq_info *irq_info)
1165 {
1166 if (cxl_is_translation_fault(afu, irq_info->dsisr))
1167 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
1168 else
1169 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
1170
1171 return IRQ_HANDLED;
1172 }
1173
native_irq_multiplexed(int irq,void * data)1174 static irqreturn_t native_irq_multiplexed(int irq, void *data)
1175 {
1176 struct cxl_afu *afu = data;
1177 struct cxl_context *ctx;
1178 struct cxl_irq_info irq_info;
1179 u64 phreg = cxl_p2n_read(afu, CXL_PSL_PEHandle_An);
1180 int ph, ret = IRQ_HANDLED, res;
1181
1182 /* check if eeh kicked in while the interrupt was in flight */
1183 if (unlikely(phreg == ~0ULL)) {
1184 dev_warn(&afu->dev,
1185 "Ignoring slice interrupt(%d) due to fenced card",
1186 irq);
1187 return IRQ_HANDLED;
1188 }
1189 /* Mask the pe-handle from register value */
1190 ph = phreg & 0xffff;
1191 if ((res = native_get_irq_info(afu, &irq_info))) {
1192 WARN(1, "Unable to get CXL IRQ Info: %i\n", res);
1193 if (afu->adapter->native->sl_ops->fail_irq)
1194 return afu->adapter->native->sl_ops->fail_irq(afu, &irq_info);
1195 return ret;
1196 }
1197
1198 rcu_read_lock();
1199 ctx = idr_find(&afu->contexts_idr, ph);
1200 if (ctx) {
1201 if (afu->adapter->native->sl_ops->handle_interrupt)
1202 ret = afu->adapter->native->sl_ops->handle_interrupt(irq, ctx, &irq_info);
1203 rcu_read_unlock();
1204 return ret;
1205 }
1206 rcu_read_unlock();
1207
1208 WARN(1, "Unable to demultiplex CXL PSL IRQ for PE %i DSISR %016llx DAR"
1209 " %016llx\n(Possible AFU HW issue - was a term/remove acked"
1210 " with outstanding transactions?)\n", ph, irq_info.dsisr,
1211 irq_info.dar);
1212 if (afu->adapter->native->sl_ops->fail_irq)
1213 ret = afu->adapter->native->sl_ops->fail_irq(afu, &irq_info);
1214 return ret;
1215 }
1216
native_irq_wait(struct cxl_context * ctx)1217 static void native_irq_wait(struct cxl_context *ctx)
1218 {
1219 u64 dsisr;
1220 int timeout = 1000;
1221 int ph;
1222
1223 /*
1224 * Wait until no further interrupts are presented by the PSL
1225 * for this context.
1226 */
1227 while (timeout--) {
1228 ph = cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) & 0xffff;
1229 if (ph != ctx->pe)
1230 return;
1231 dsisr = cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An);
1232 if (cxl_is_power8() &&
1233 ((dsisr & CXL_PSL_DSISR_PENDING) == 0))
1234 return;
1235 if (cxl_is_power9() &&
1236 ((dsisr & CXL_PSL9_DSISR_PENDING) == 0))
1237 return;
1238 /*
1239 * We are waiting for the workqueue to process our
1240 * irq, so need to let that run here.
1241 */
1242 msleep(1);
1243 }
1244
1245 dev_warn(&ctx->afu->dev, "WARNING: waiting on DSI for PE %i"
1246 " DSISR %016llx!\n", ph, dsisr);
1247 return;
1248 }
1249
native_slice_irq_err(int irq,void * data)1250 static irqreturn_t native_slice_irq_err(int irq, void *data)
1251 {
1252 struct cxl_afu *afu = data;
1253 u64 errstat, serr, afu_error, dsisr;
1254 u64 fir_slice, afu_debug, irq_mask;
1255
1256 /*
1257 * slice err interrupt is only used with full PSL (no XSL)
1258 */
1259 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
1260 errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
1261 afu_error = cxl_p2n_read(afu, CXL_AFU_ERR_An);
1262 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
1263 cxl_afu_decode_psl_serr(afu, serr);
1264
1265 if (cxl_is_power8()) {
1266 fir_slice = cxl_p1n_read(afu, CXL_PSL_FIR_SLICE_An);
1267 afu_debug = cxl_p1n_read(afu, CXL_AFU_DEBUG_An);
1268 dev_crit(&afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice);
1269 dev_crit(&afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug);
1270 }
1271 dev_crit(&afu->dev, "CXL_PSL_ErrStat_An: 0x%016llx\n", errstat);
1272 dev_crit(&afu->dev, "AFU_ERR_An: 0x%.16llx\n", afu_error);
1273 dev_crit(&afu->dev, "PSL_DSISR_An: 0x%.16llx\n", dsisr);
1274
1275 /* mask off the IRQ so it won't retrigger until the AFU is reset */
1276 irq_mask = (serr & CXL_PSL_SERR_An_IRQS) >> 32;
1277 serr |= irq_mask;
1278 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
1279 dev_info(&afu->dev, "Further such interrupts will be masked until the AFU is reset\n");
1280
1281 return IRQ_HANDLED;
1282 }
1283
cxl_native_err_irq_dump_regs_psl9(struct cxl * adapter)1284 void cxl_native_err_irq_dump_regs_psl9(struct cxl *adapter)
1285 {
1286 u64 fir1;
1287
1288 fir1 = cxl_p1_read(adapter, CXL_PSL9_FIR1);
1289 dev_crit(&adapter->dev, "PSL_FIR: 0x%016llx\n", fir1);
1290 }
1291
cxl_native_err_irq_dump_regs_psl8(struct cxl * adapter)1292 void cxl_native_err_irq_dump_regs_psl8(struct cxl *adapter)
1293 {
1294 u64 fir1, fir2;
1295
1296 fir1 = cxl_p1_read(adapter, CXL_PSL_FIR1);
1297 fir2 = cxl_p1_read(adapter, CXL_PSL_FIR2);
1298 dev_crit(&adapter->dev,
1299 "PSL_FIR1: 0x%016llx\nPSL_FIR2: 0x%016llx\n",
1300 fir1, fir2);
1301 }
1302
native_irq_err(int irq,void * data)1303 static irqreturn_t native_irq_err(int irq, void *data)
1304 {
1305 struct cxl *adapter = data;
1306 u64 err_ivte;
1307
1308 WARN(1, "CXL ERROR interrupt %i\n", irq);
1309
1310 err_ivte = cxl_p1_read(adapter, CXL_PSL_ErrIVTE);
1311 dev_crit(&adapter->dev, "PSL_ErrIVTE: 0x%016llx\n", err_ivte);
1312
1313 if (adapter->native->sl_ops->debugfs_stop_trace) {
1314 dev_crit(&adapter->dev, "STOPPING CXL TRACE\n");
1315 adapter->native->sl_ops->debugfs_stop_trace(adapter);
1316 }
1317
1318 if (adapter->native->sl_ops->err_irq_dump_registers)
1319 adapter->native->sl_ops->err_irq_dump_registers(adapter);
1320
1321 return IRQ_HANDLED;
1322 }
1323
cxl_native_register_psl_err_irq(struct cxl * adapter)1324 int cxl_native_register_psl_err_irq(struct cxl *adapter)
1325 {
1326 int rc;
1327
1328 adapter->irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err",
1329 dev_name(&adapter->dev));
1330 if (!adapter->irq_name)
1331 return -ENOMEM;
1332
1333 if ((rc = cxl_register_one_irq(adapter, native_irq_err, adapter,
1334 &adapter->native->err_hwirq,
1335 &adapter->native->err_virq,
1336 adapter->irq_name))) {
1337 kfree(adapter->irq_name);
1338 adapter->irq_name = NULL;
1339 return rc;
1340 }
1341
1342 cxl_p1_write(adapter, CXL_PSL_ErrIVTE, adapter->native->err_hwirq & 0xffff);
1343
1344 return 0;
1345 }
1346
cxl_native_release_psl_err_irq(struct cxl * adapter)1347 void cxl_native_release_psl_err_irq(struct cxl *adapter)
1348 {
1349 if (adapter->native->err_virq == 0 ||
1350 adapter->native->err_virq !=
1351 irq_find_mapping(NULL, adapter->native->err_hwirq))
1352 return;
1353
1354 cxl_p1_write(adapter, CXL_PSL_ErrIVTE, 0x0000000000000000);
1355 cxl_unmap_irq(adapter->native->err_virq, adapter);
1356 cxl_ops->release_one_irq(adapter, adapter->native->err_hwirq);
1357 kfree(adapter->irq_name);
1358 adapter->native->err_virq = 0;
1359 }
1360
cxl_native_register_serr_irq(struct cxl_afu * afu)1361 int cxl_native_register_serr_irq(struct cxl_afu *afu)
1362 {
1363 u64 serr;
1364 int rc;
1365
1366 afu->err_irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err",
1367 dev_name(&afu->dev));
1368 if (!afu->err_irq_name)
1369 return -ENOMEM;
1370
1371 if ((rc = cxl_register_one_irq(afu->adapter, native_slice_irq_err, afu,
1372 &afu->serr_hwirq,
1373 &afu->serr_virq, afu->err_irq_name))) {
1374 kfree(afu->err_irq_name);
1375 afu->err_irq_name = NULL;
1376 return rc;
1377 }
1378
1379 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
1380 if (cxl_is_power8())
1381 serr = (serr & 0x00ffffffffff0000ULL) | (afu->serr_hwirq & 0xffff);
1382 if (cxl_is_power9()) {
1383 /*
1384 * By default, all errors are masked. So don't set all masks.
1385 * Slice errors will be transfered.
1386 */
1387 serr = (serr & ~0xff0000007fffffffULL) | (afu->serr_hwirq & 0xffff);
1388 }
1389 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
1390
1391 return 0;
1392 }
1393
cxl_native_release_serr_irq(struct cxl_afu * afu)1394 void cxl_native_release_serr_irq(struct cxl_afu *afu)
1395 {
1396 if (afu->serr_virq == 0 ||
1397 afu->serr_virq != irq_find_mapping(NULL, afu->serr_hwirq))
1398 return;
1399
1400 cxl_p1n_write(afu, CXL_PSL_SERR_An, 0x0000000000000000);
1401 cxl_unmap_irq(afu->serr_virq, afu);
1402 cxl_ops->release_one_irq(afu->adapter, afu->serr_hwirq);
1403 kfree(afu->err_irq_name);
1404 afu->serr_virq = 0;
1405 }
1406
cxl_native_register_psl_irq(struct cxl_afu * afu)1407 int cxl_native_register_psl_irq(struct cxl_afu *afu)
1408 {
1409 int rc;
1410
1411 afu->psl_irq_name = kasprintf(GFP_KERNEL, "cxl-%s",
1412 dev_name(&afu->dev));
1413 if (!afu->psl_irq_name)
1414 return -ENOMEM;
1415
1416 if ((rc = cxl_register_one_irq(afu->adapter, native_irq_multiplexed,
1417 afu, &afu->native->psl_hwirq, &afu->native->psl_virq,
1418 afu->psl_irq_name))) {
1419 kfree(afu->psl_irq_name);
1420 afu->psl_irq_name = NULL;
1421 }
1422 return rc;
1423 }
1424
cxl_native_release_psl_irq(struct cxl_afu * afu)1425 void cxl_native_release_psl_irq(struct cxl_afu *afu)
1426 {
1427 if (afu->native->psl_virq == 0 ||
1428 afu->native->psl_virq !=
1429 irq_find_mapping(NULL, afu->native->psl_hwirq))
1430 return;
1431
1432 cxl_unmap_irq(afu->native->psl_virq, afu);
1433 cxl_ops->release_one_irq(afu->adapter, afu->native->psl_hwirq);
1434 kfree(afu->psl_irq_name);
1435 afu->native->psl_virq = 0;
1436 }
1437
recover_psl_err(struct cxl_afu * afu,u64 errstat)1438 static void recover_psl_err(struct cxl_afu *afu, u64 errstat)
1439 {
1440 u64 dsisr;
1441
1442 pr_devel("RECOVERING FROM PSL ERROR... (0x%016llx)\n", errstat);
1443
1444 /* Clear PSL_DSISR[PE] */
1445 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
1446 cxl_p2n_write(afu, CXL_PSL_DSISR_An, dsisr & ~CXL_PSL_DSISR_An_PE);
1447
1448 /* Write 1s to clear error status bits */
1449 cxl_p2n_write(afu, CXL_PSL_ErrStat_An, errstat);
1450 }
1451
native_ack_irq(struct cxl_context * ctx,u64 tfc,u64 psl_reset_mask)1452 static int native_ack_irq(struct cxl_context *ctx, u64 tfc, u64 psl_reset_mask)
1453 {
1454 trace_cxl_psl_irq_ack(ctx, tfc);
1455 if (tfc)
1456 cxl_p2n_write(ctx->afu, CXL_PSL_TFC_An, tfc);
1457 if (psl_reset_mask)
1458 recover_psl_err(ctx->afu, psl_reset_mask);
1459
1460 return 0;
1461 }
1462
cxl_check_error(struct cxl_afu * afu)1463 int cxl_check_error(struct cxl_afu *afu)
1464 {
1465 return (cxl_p1n_read(afu, CXL_PSL_SCNTL_An) == ~0ULL);
1466 }
1467
native_support_attributes(const char * attr_name,enum cxl_attrs type)1468 static bool native_support_attributes(const char *attr_name,
1469 enum cxl_attrs type)
1470 {
1471 return true;
1472 }
1473
native_afu_cr_read64(struct cxl_afu * afu,int cr,u64 off,u64 * out)1474 static int native_afu_cr_read64(struct cxl_afu *afu, int cr, u64 off, u64 *out)
1475 {
1476 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu)))
1477 return -EIO;
1478 if (unlikely(off >= afu->crs_len))
1479 return -ERANGE;
1480 *out = in_le64(afu->native->afu_desc_mmio + afu->crs_offset +
1481 (cr * afu->crs_len) + off);
1482 return 0;
1483 }
1484
native_afu_cr_read32(struct cxl_afu * afu,int cr,u64 off,u32 * out)1485 static int native_afu_cr_read32(struct cxl_afu *afu, int cr, u64 off, u32 *out)
1486 {
1487 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu)))
1488 return -EIO;
1489 if (unlikely(off >= afu->crs_len))
1490 return -ERANGE;
1491 *out = in_le32(afu->native->afu_desc_mmio + afu->crs_offset +
1492 (cr * afu->crs_len) + off);
1493 return 0;
1494 }
1495
native_afu_cr_read16(struct cxl_afu * afu,int cr,u64 off,u16 * out)1496 static int native_afu_cr_read16(struct cxl_afu *afu, int cr, u64 off, u16 *out)
1497 {
1498 u64 aligned_off = off & ~0x3L;
1499 u32 val;
1500 int rc;
1501
1502 rc = native_afu_cr_read32(afu, cr, aligned_off, &val);
1503 if (!rc)
1504 *out = (val >> ((off & 0x3) * 8)) & 0xffff;
1505 return rc;
1506 }
1507
native_afu_cr_read8(struct cxl_afu * afu,int cr,u64 off,u8 * out)1508 static int native_afu_cr_read8(struct cxl_afu *afu, int cr, u64 off, u8 *out)
1509 {
1510 u64 aligned_off = off & ~0x3L;
1511 u32 val;
1512 int rc;
1513
1514 rc = native_afu_cr_read32(afu, cr, aligned_off, &val);
1515 if (!rc)
1516 *out = (val >> ((off & 0x3) * 8)) & 0xff;
1517 return rc;
1518 }
1519
native_afu_cr_write32(struct cxl_afu * afu,int cr,u64 off,u32 in)1520 static int native_afu_cr_write32(struct cxl_afu *afu, int cr, u64 off, u32 in)
1521 {
1522 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu)))
1523 return -EIO;
1524 if (unlikely(off >= afu->crs_len))
1525 return -ERANGE;
1526 out_le32(afu->native->afu_desc_mmio + afu->crs_offset +
1527 (cr * afu->crs_len) + off, in);
1528 return 0;
1529 }
1530
native_afu_cr_write16(struct cxl_afu * afu,int cr,u64 off,u16 in)1531 static int native_afu_cr_write16(struct cxl_afu *afu, int cr, u64 off, u16 in)
1532 {
1533 u64 aligned_off = off & ~0x3L;
1534 u32 val32, mask, shift;
1535 int rc;
1536
1537 rc = native_afu_cr_read32(afu, cr, aligned_off, &val32);
1538 if (rc)
1539 return rc;
1540 shift = (off & 0x3) * 8;
1541 WARN_ON(shift == 24);
1542 mask = 0xffff << shift;
1543 val32 = (val32 & ~mask) | (in << shift);
1544
1545 rc = native_afu_cr_write32(afu, cr, aligned_off, val32);
1546 return rc;
1547 }
1548
native_afu_cr_write8(struct cxl_afu * afu,int cr,u64 off,u8 in)1549 static int native_afu_cr_write8(struct cxl_afu *afu, int cr, u64 off, u8 in)
1550 {
1551 u64 aligned_off = off & ~0x3L;
1552 u32 val32, mask, shift;
1553 int rc;
1554
1555 rc = native_afu_cr_read32(afu, cr, aligned_off, &val32);
1556 if (rc)
1557 return rc;
1558 shift = (off & 0x3) * 8;
1559 mask = 0xff << shift;
1560 val32 = (val32 & ~mask) | (in << shift);
1561
1562 rc = native_afu_cr_write32(afu, cr, aligned_off, val32);
1563 return rc;
1564 }
1565
1566 const struct cxl_backend_ops cxl_native_ops = {
1567 .module = THIS_MODULE,
1568 .adapter_reset = cxl_pci_reset,
1569 .alloc_one_irq = cxl_pci_alloc_one_irq,
1570 .release_one_irq = cxl_pci_release_one_irq,
1571 .alloc_irq_ranges = cxl_pci_alloc_irq_ranges,
1572 .release_irq_ranges = cxl_pci_release_irq_ranges,
1573 .setup_irq = cxl_pci_setup_irq,
1574 .handle_psl_slice_error = native_handle_psl_slice_error,
1575 .psl_interrupt = NULL,
1576 .ack_irq = native_ack_irq,
1577 .irq_wait = native_irq_wait,
1578 .attach_process = native_attach_process,
1579 .detach_process = native_detach_process,
1580 .update_ivtes = native_update_ivtes,
1581 .support_attributes = native_support_attributes,
1582 .link_ok = cxl_adapter_link_ok,
1583 .release_afu = cxl_pci_release_afu,
1584 .afu_read_err_buffer = cxl_pci_afu_read_err_buffer,
1585 .afu_check_and_enable = native_afu_check_and_enable,
1586 .afu_activate_mode = native_afu_activate_mode,
1587 .afu_deactivate_mode = native_afu_deactivate_mode,
1588 .afu_reset = native_afu_reset,
1589 .afu_cr_read8 = native_afu_cr_read8,
1590 .afu_cr_read16 = native_afu_cr_read16,
1591 .afu_cr_read32 = native_afu_cr_read32,
1592 .afu_cr_read64 = native_afu_cr_read64,
1593 .afu_cr_write8 = native_afu_cr_write8,
1594 .afu_cr_write16 = native_afu_cr_write16,
1595 .afu_cr_write32 = native_afu_cr_write32,
1596 .read_adapter_vpd = cxl_pci_read_adapter_vpd,
1597 };
1598