• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2014 IBM Corp.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation; either version
7  * 2 of the License, or (at your option) any later version.
8  */
9 
10 #include <linux/interrupt.h>
11 #include <linux/workqueue.h>
12 #include <linux/sched.h>
13 #include <linux/wait.h>
14 #include <linux/slab.h>
15 #include <linux/pid.h>
16 #include <asm/cputable.h>
17 #include <misc/cxl-base.h>
18 
19 #include "cxl.h"
20 #include "trace.h"
21 
22 /* XXX: This is implementation specific */
handle_psl_slice_error(struct cxl_context * ctx,u64 dsisr,u64 errstat)23 static irqreturn_t handle_psl_slice_error(struct cxl_context *ctx, u64 dsisr, u64 errstat)
24 {
25 	u64 fir1, fir2, fir_slice, serr, afu_debug;
26 
27 	fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR1);
28 	fir2 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR2);
29 	fir_slice = cxl_p1n_read(ctx->afu, CXL_PSL_FIR_SLICE_An);
30 	serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An);
31 	afu_debug = cxl_p1n_read(ctx->afu, CXL_AFU_DEBUG_An);
32 
33 	dev_crit(&ctx->afu->dev, "PSL ERROR STATUS: 0x%016llx\n", errstat);
34 	dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1);
35 	dev_crit(&ctx->afu->dev, "PSL_FIR2: 0x%016llx\n", fir2);
36 	dev_crit(&ctx->afu->dev, "PSL_SERR_An: 0x%016llx\n", serr);
37 	dev_crit(&ctx->afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice);
38 	dev_crit(&ctx->afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug);
39 
40 	dev_crit(&ctx->afu->dev, "STOPPING CXL TRACE\n");
41 	cxl_stop_trace(ctx->afu->adapter);
42 
43 	return cxl_ack_irq(ctx, 0, errstat);
44 }
45 
cxl_slice_irq_err(int irq,void * data)46 irqreturn_t cxl_slice_irq_err(int irq, void *data)
47 {
48 	struct cxl_afu *afu = data;
49 	u64 fir_slice, errstat, serr, afu_debug;
50 
51 	WARN(irq, "CXL SLICE ERROR interrupt %i\n", irq);
52 
53 	serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
54 	fir_slice = cxl_p1n_read(afu, CXL_PSL_FIR_SLICE_An);
55 	errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
56 	afu_debug = cxl_p1n_read(afu, CXL_AFU_DEBUG_An);
57 	dev_crit(&afu->dev, "PSL_SERR_An: 0x%016llx\n", serr);
58 	dev_crit(&afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice);
59 	dev_crit(&afu->dev, "CXL_PSL_ErrStat_An: 0x%016llx\n", errstat);
60 	dev_crit(&afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug);
61 
62 	cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
63 
64 	return IRQ_HANDLED;
65 }
66 
cxl_irq_err(int irq,void * data)67 static irqreturn_t cxl_irq_err(int irq, void *data)
68 {
69 	struct cxl *adapter = data;
70 	u64 fir1, fir2, err_ivte;
71 
72 	WARN(1, "CXL ERROR interrupt %i\n", irq);
73 
74 	err_ivte = cxl_p1_read(adapter, CXL_PSL_ErrIVTE);
75 	dev_crit(&adapter->dev, "PSL_ErrIVTE: 0x%016llx\n", err_ivte);
76 
77 	dev_crit(&adapter->dev, "STOPPING CXL TRACE\n");
78 	cxl_stop_trace(adapter);
79 
80 	fir1 = cxl_p1_read(adapter, CXL_PSL_FIR1);
81 	fir2 = cxl_p1_read(adapter, CXL_PSL_FIR2);
82 
83 	dev_crit(&adapter->dev, "PSL_FIR1: 0x%016llx\nPSL_FIR2: 0x%016llx\n", fir1, fir2);
84 
85 	return IRQ_HANDLED;
86 }
87 
schedule_cxl_fault(struct cxl_context * ctx,u64 dsisr,u64 dar)88 static irqreturn_t schedule_cxl_fault(struct cxl_context *ctx, u64 dsisr, u64 dar)
89 {
90 	ctx->dsisr = dsisr;
91 	ctx->dar = dar;
92 	schedule_work(&ctx->fault_work);
93 	return IRQ_HANDLED;
94 }
95 
cxl_irq(int irq,void * data,struct cxl_irq_info * irq_info)96 static irqreturn_t cxl_irq(int irq, void *data, struct cxl_irq_info *irq_info)
97 {
98 	struct cxl_context *ctx = data;
99 	u64 dsisr, dar;
100 
101 	dsisr = irq_info->dsisr;
102 	dar = irq_info->dar;
103 
104 	trace_cxl_psl_irq(ctx, irq, dsisr, dar);
105 
106 	pr_devel("CXL interrupt %i for afu pe: %i DSISR: %#llx DAR: %#llx\n", irq, ctx->pe, dsisr, dar);
107 
108 	if (dsisr & CXL_PSL_DSISR_An_DS) {
109 		/*
110 		 * We don't inherently need to sleep to handle this, but we do
111 		 * need to get a ref to the task's mm, which we can't do from
112 		 * irq context without the potential for a deadlock since it
113 		 * takes the task_lock. An alternate option would be to keep a
114 		 * reference to the task's mm the entire time it has cxl open,
115 		 * but to do that we need to solve the issue where we hold a
116 		 * ref to the mm, but the mm can hold a ref to the fd after an
117 		 * mmap preventing anything from being cleaned up.
118 		 */
119 		pr_devel("Scheduling segment miss handling for later pe: %i\n", ctx->pe);
120 		return schedule_cxl_fault(ctx, dsisr, dar);
121 	}
122 
123 	if (dsisr & CXL_PSL_DSISR_An_M)
124 		pr_devel("CXL interrupt: PTE not found\n");
125 	if (dsisr & CXL_PSL_DSISR_An_P)
126 		pr_devel("CXL interrupt: Storage protection violation\n");
127 	if (dsisr & CXL_PSL_DSISR_An_A)
128 		pr_devel("CXL interrupt: AFU lock access to write through or cache inhibited storage\n");
129 	if (dsisr & CXL_PSL_DSISR_An_S)
130 		pr_devel("CXL interrupt: Access was afu_wr or afu_zero\n");
131 	if (dsisr & CXL_PSL_DSISR_An_K)
132 		pr_devel("CXL interrupt: Access not permitted by virtual page class key protection\n");
133 
134 	if (dsisr & CXL_PSL_DSISR_An_DM) {
135 		/*
136 		 * In some cases we might be able to handle the fault
137 		 * immediately if hash_page would succeed, but we still need
138 		 * the task's mm, which as above we can't get without a lock
139 		 */
140 		pr_devel("Scheduling page fault handling for later pe: %i\n", ctx->pe);
141 		return schedule_cxl_fault(ctx, dsisr, dar);
142 	}
143 	if (dsisr & CXL_PSL_DSISR_An_ST)
144 		WARN(1, "CXL interrupt: Segment Table PTE not found\n");
145 	if (dsisr & CXL_PSL_DSISR_An_UR)
146 		pr_devel("CXL interrupt: AURP PTE not found\n");
147 	if (dsisr & CXL_PSL_DSISR_An_PE)
148 		return handle_psl_slice_error(ctx, dsisr, irq_info->errstat);
149 	if (dsisr & CXL_PSL_DSISR_An_AE) {
150 		pr_devel("CXL interrupt: AFU Error 0x%016llx\n", irq_info->afu_err);
151 
152 		if (ctx->pending_afu_err) {
153 			/*
154 			 * This shouldn't happen - the PSL treats these errors
155 			 * as fatal and will have reset the AFU, so there's not
156 			 * much point buffering multiple AFU errors.
157 			 * OTOH if we DO ever see a storm of these come in it's
158 			 * probably best that we log them somewhere:
159 			 */
160 			dev_err_ratelimited(&ctx->afu->dev, "CXL AFU Error "
161 					    "undelivered to pe %i: 0x%016llx\n",
162 					    ctx->pe, irq_info->afu_err);
163 		} else {
164 			spin_lock(&ctx->lock);
165 			ctx->afu_err = irq_info->afu_err;
166 			ctx->pending_afu_err = 1;
167 			spin_unlock(&ctx->lock);
168 
169 			wake_up_all(&ctx->wq);
170 		}
171 
172 		cxl_ack_irq(ctx, CXL_PSL_TFC_An_A, 0);
173 		return IRQ_HANDLED;
174 	}
175 	if (dsisr & CXL_PSL_DSISR_An_OC)
176 		pr_devel("CXL interrupt: OS Context Warning\n");
177 
178 	WARN(1, "Unhandled CXL PSL IRQ\n");
179 	return IRQ_HANDLED;
180 }
181 
fail_psl_irq(struct cxl_afu * afu,struct cxl_irq_info * irq_info)182 static irqreturn_t fail_psl_irq(struct cxl_afu *afu, struct cxl_irq_info *irq_info)
183 {
184 	if (irq_info->dsisr & CXL_PSL_DSISR_TRANS)
185 		cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
186 	else
187 		cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
188 
189 	return IRQ_HANDLED;
190 }
191 
cxl_irq_multiplexed(int irq,void * data)192 static irqreturn_t cxl_irq_multiplexed(int irq, void *data)
193 {
194 	struct cxl_afu *afu = data;
195 	struct cxl_context *ctx;
196 	struct cxl_irq_info irq_info;
197 	int ph = cxl_p2n_read(afu, CXL_PSL_PEHandle_An) & 0xffff;
198 	int ret;
199 
200 	if ((ret = cxl_get_irq(afu, &irq_info))) {
201 		WARN(1, "Unable to get CXL IRQ Info: %i\n", ret);
202 		return fail_psl_irq(afu, &irq_info);
203 	}
204 
205 	rcu_read_lock();
206 	ctx = idr_find(&afu->contexts_idr, ph);
207 	if (ctx) {
208 		ret = cxl_irq(irq, ctx, &irq_info);
209 		rcu_read_unlock();
210 		return ret;
211 	}
212 	rcu_read_unlock();
213 
214 	WARN(1, "Unable to demultiplex CXL PSL IRQ for PE %i DSISR %016llx DAR"
215 		" %016llx\n(Possible AFU HW issue - was a term/remove acked"
216 		" with outstanding transactions?)\n", ph, irq_info.dsisr,
217 		irq_info.dar);
218 	return fail_psl_irq(afu, &irq_info);
219 }
220 
cxl_irq_afu(int irq,void * data)221 static irqreturn_t cxl_irq_afu(int irq, void *data)
222 {
223 	struct cxl_context *ctx = data;
224 	irq_hw_number_t hwirq = irqd_to_hwirq(irq_get_irq_data(irq));
225 	int irq_off, afu_irq = 1;
226 	__u16 range;
227 	int r;
228 
229 	for (r = 1; r < CXL_IRQ_RANGES; r++) {
230 		irq_off = hwirq - ctx->irqs.offset[r];
231 		range = ctx->irqs.range[r];
232 		if (irq_off >= 0 && irq_off < range) {
233 			afu_irq += irq_off;
234 			break;
235 		}
236 		afu_irq += range;
237 	}
238 	if (unlikely(r >= CXL_IRQ_RANGES)) {
239 		WARN(1, "Recieved AFU IRQ out of range for pe %i (virq %i hwirq %lx)\n",
240 		     ctx->pe, irq, hwirq);
241 		return IRQ_HANDLED;
242 	}
243 
244 	trace_cxl_afu_irq(ctx, afu_irq, irq, hwirq);
245 	pr_devel("Received AFU interrupt %i for pe: %i (virq %i hwirq %lx)\n",
246 	       afu_irq, ctx->pe, irq, hwirq);
247 
248 	if (unlikely(!ctx->irq_bitmap)) {
249 		WARN(1, "Recieved AFU IRQ for context with no IRQ bitmap\n");
250 		return IRQ_HANDLED;
251 	}
252 	spin_lock(&ctx->lock);
253 	set_bit(afu_irq - 1, ctx->irq_bitmap);
254 	ctx->pending_irq = true;
255 	spin_unlock(&ctx->lock);
256 
257 	wake_up_all(&ctx->wq);
258 
259 	return IRQ_HANDLED;
260 }
261 
cxl_map_irq(struct cxl * adapter,irq_hw_number_t hwirq,irq_handler_t handler,void * cookie,const char * name)262 unsigned int cxl_map_irq(struct cxl *adapter, irq_hw_number_t hwirq,
263 			 irq_handler_t handler, void *cookie, const char *name)
264 {
265 	unsigned int virq;
266 	int result;
267 
268 	/* IRQ Domain? */
269 	virq = irq_create_mapping(NULL, hwirq);
270 	if (!virq) {
271 		dev_warn(&adapter->dev, "cxl_map_irq: irq_create_mapping failed\n");
272 		return 0;
273 	}
274 
275 	cxl_setup_irq(adapter, hwirq, virq);
276 
277 	pr_devel("hwirq %#lx mapped to virq %u\n", hwirq, virq);
278 
279 	result = request_irq(virq, handler, 0, name, cookie);
280 	if (result) {
281 		dev_warn(&adapter->dev, "cxl_map_irq: request_irq failed: %i\n", result);
282 		return 0;
283 	}
284 
285 	return virq;
286 }
287 
cxl_unmap_irq(unsigned int virq,void * cookie)288 void cxl_unmap_irq(unsigned int virq, void *cookie)
289 {
290 	free_irq(virq, cookie);
291 }
292 
cxl_register_one_irq(struct cxl * adapter,irq_handler_t handler,void * cookie,irq_hw_number_t * dest_hwirq,unsigned int * dest_virq,const char * name)293 static int cxl_register_one_irq(struct cxl *adapter,
294 				irq_handler_t handler,
295 				void *cookie,
296 				irq_hw_number_t *dest_hwirq,
297 				unsigned int *dest_virq,
298 				const char *name)
299 {
300 	int hwirq, virq;
301 
302 	if ((hwirq = cxl_alloc_one_irq(adapter)) < 0)
303 		return hwirq;
304 
305 	if (!(virq = cxl_map_irq(adapter, hwirq, handler, cookie, name)))
306 		goto err;
307 
308 	*dest_hwirq = hwirq;
309 	*dest_virq = virq;
310 
311 	return 0;
312 
313 err:
314 	cxl_release_one_irq(adapter, hwirq);
315 	return -ENOMEM;
316 }
317 
cxl_register_psl_err_irq(struct cxl * adapter)318 int cxl_register_psl_err_irq(struct cxl *adapter)
319 {
320 	int rc;
321 
322 	adapter->irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err",
323 				      dev_name(&adapter->dev));
324 	if (!adapter->irq_name)
325 		return -ENOMEM;
326 
327 	if ((rc = cxl_register_one_irq(adapter, cxl_irq_err, adapter,
328 				       &adapter->err_hwirq,
329 				       &adapter->err_virq,
330 				       adapter->irq_name))) {
331 		kfree(adapter->irq_name);
332 		adapter->irq_name = NULL;
333 		return rc;
334 	}
335 
336 	cxl_p1_write(adapter, CXL_PSL_ErrIVTE, adapter->err_hwirq & 0xffff);
337 
338 	return 0;
339 }
340 
cxl_release_psl_err_irq(struct cxl * adapter)341 void cxl_release_psl_err_irq(struct cxl *adapter)
342 {
343 	if (adapter->err_virq != irq_find_mapping(NULL, adapter->err_hwirq))
344 		return;
345 
346 	cxl_p1_write(adapter, CXL_PSL_ErrIVTE, 0x0000000000000000);
347 	cxl_unmap_irq(adapter->err_virq, adapter);
348 	cxl_release_one_irq(adapter, adapter->err_hwirq);
349 	kfree(adapter->irq_name);
350 }
351 
cxl_register_serr_irq(struct cxl_afu * afu)352 int cxl_register_serr_irq(struct cxl_afu *afu)
353 {
354 	u64 serr;
355 	int rc;
356 
357 	afu->err_irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err",
358 				      dev_name(&afu->dev));
359 	if (!afu->err_irq_name)
360 		return -ENOMEM;
361 
362 	if ((rc = cxl_register_one_irq(afu->adapter, cxl_slice_irq_err, afu,
363 				       &afu->serr_hwirq,
364 				       &afu->serr_virq, afu->err_irq_name))) {
365 		kfree(afu->err_irq_name);
366 		afu->err_irq_name = NULL;
367 		return rc;
368 	}
369 
370 	serr = cxl_p1n_read(afu, CXL_PSL_SERR_An);
371 	serr = (serr & 0x00ffffffffff0000ULL) | (afu->serr_hwirq & 0xffff);
372 	cxl_p1n_write(afu, CXL_PSL_SERR_An, serr);
373 
374 	return 0;
375 }
376 
cxl_release_serr_irq(struct cxl_afu * afu)377 void cxl_release_serr_irq(struct cxl_afu *afu)
378 {
379 	if (afu->serr_virq != irq_find_mapping(NULL, afu->serr_hwirq))
380 		return;
381 
382 	cxl_p1n_write(afu, CXL_PSL_SERR_An, 0x0000000000000000);
383 	cxl_unmap_irq(afu->serr_virq, afu);
384 	cxl_release_one_irq(afu->adapter, afu->serr_hwirq);
385 	kfree(afu->err_irq_name);
386 }
387 
cxl_register_psl_irq(struct cxl_afu * afu)388 int cxl_register_psl_irq(struct cxl_afu *afu)
389 {
390 	int rc;
391 
392 	afu->psl_irq_name = kasprintf(GFP_KERNEL, "cxl-%s",
393 				      dev_name(&afu->dev));
394 	if (!afu->psl_irq_name)
395 		return -ENOMEM;
396 
397 	if ((rc = cxl_register_one_irq(afu->adapter, cxl_irq_multiplexed, afu,
398 				    &afu->psl_hwirq, &afu->psl_virq,
399 				    afu->psl_irq_name))) {
400 		kfree(afu->psl_irq_name);
401 		afu->psl_irq_name = NULL;
402 	}
403 	return rc;
404 }
405 
cxl_release_psl_irq(struct cxl_afu * afu)406 void cxl_release_psl_irq(struct cxl_afu *afu)
407 {
408 	if (afu->psl_virq != irq_find_mapping(NULL, afu->psl_hwirq))
409 		return;
410 
411 	cxl_unmap_irq(afu->psl_virq, afu);
412 	cxl_release_one_irq(afu->adapter, afu->psl_hwirq);
413 	kfree(afu->psl_irq_name);
414 }
415 
afu_irq_name_free(struct cxl_context * ctx)416 void afu_irq_name_free(struct cxl_context *ctx)
417 {
418 	struct cxl_irq_name *irq_name, *tmp;
419 
420 	list_for_each_entry_safe(irq_name, tmp, &ctx->irq_names, list) {
421 		kfree(irq_name->name);
422 		list_del(&irq_name->list);
423 		kfree(irq_name);
424 	}
425 }
426 
afu_allocate_irqs(struct cxl_context * ctx,u32 count)427 int afu_allocate_irqs(struct cxl_context *ctx, u32 count)
428 {
429 	int rc, r, i, j = 1;
430 	struct cxl_irq_name *irq_name;
431 
432 	/* Initialize the list head to hold irq names */
433 	INIT_LIST_HEAD(&ctx->irq_names);
434 
435 	if ((rc = cxl_alloc_irq_ranges(&ctx->irqs, ctx->afu->adapter, count)))
436 		return rc;
437 
438 	/* Multiplexed PSL Interrupt */
439 	ctx->irqs.offset[0] = ctx->afu->psl_hwirq;
440 	ctx->irqs.range[0] = 1;
441 
442 	ctx->irq_count = count;
443 	ctx->irq_bitmap = kcalloc(BITS_TO_LONGS(count),
444 				  sizeof(*ctx->irq_bitmap), GFP_KERNEL);
445 	if (!ctx->irq_bitmap)
446 		goto out;
447 
448 	/*
449 	 * Allocate names first.  If any fail, bail out before allocating
450 	 * actual hardware IRQs.
451 	 */
452 	for (r = 1; r < CXL_IRQ_RANGES; r++) {
453 		for (i = 0; i < ctx->irqs.range[r]; i++) {
454 			irq_name = kmalloc(sizeof(struct cxl_irq_name),
455 					   GFP_KERNEL);
456 			if (!irq_name)
457 				goto out;
458 			irq_name->name = kasprintf(GFP_KERNEL, "cxl-%s-pe%i-%i",
459 						   dev_name(&ctx->afu->dev),
460 						   ctx->pe, j);
461 			if (!irq_name->name) {
462 				kfree(irq_name);
463 				goto out;
464 			}
465 			/* Add to tail so next look get the correct order */
466 			list_add_tail(&irq_name->list, &ctx->irq_names);
467 			j++;
468 		}
469 	}
470 	return 0;
471 
472 out:
473 	cxl_release_irq_ranges(&ctx->irqs, ctx->afu->adapter);
474 	afu_irq_name_free(ctx);
475 	return -ENOMEM;
476 }
477 
afu_register_hwirqs(struct cxl_context * ctx)478 static void afu_register_hwirqs(struct cxl_context *ctx)
479 {
480 	irq_hw_number_t hwirq;
481 	struct cxl_irq_name *irq_name;
482 	int r,i;
483 
484 	/* We've allocated all memory now, so let's do the irq allocations */
485 	irq_name = list_first_entry(&ctx->irq_names, struct cxl_irq_name, list);
486 	for (r = 1; r < CXL_IRQ_RANGES; r++) {
487 		hwirq = ctx->irqs.offset[r];
488 		for (i = 0; i < ctx->irqs.range[r]; hwirq++, i++) {
489 			cxl_map_irq(ctx->afu->adapter, hwirq,
490 				    cxl_irq_afu, ctx, irq_name->name);
491 			irq_name = list_next_entry(irq_name, list);
492 		}
493 	}
494 }
495 
afu_register_irqs(struct cxl_context * ctx,u32 count)496 int afu_register_irqs(struct cxl_context *ctx, u32 count)
497 {
498 	int rc;
499 
500 	rc = afu_allocate_irqs(ctx, count);
501 	if (rc)
502 		return rc;
503 
504 	afu_register_hwirqs(ctx);
505 	return 0;
506  }
507 
afu_release_irqs(struct cxl_context * ctx,void * cookie)508 void afu_release_irqs(struct cxl_context *ctx, void *cookie)
509 {
510 	irq_hw_number_t hwirq;
511 	unsigned int virq;
512 	int r, i;
513 
514 	for (r = 1; r < CXL_IRQ_RANGES; r++) {
515 		hwirq = ctx->irqs.offset[r];
516 		for (i = 0; i < ctx->irqs.range[r]; hwirq++, i++) {
517 			virq = irq_find_mapping(NULL, hwirq);
518 			if (virq)
519 				cxl_unmap_irq(virq, cookie);
520 		}
521 	}
522 
523 	afu_irq_name_free(ctx);
524 	cxl_release_irq_ranges(&ctx->irqs, ctx->afu->adapter);
525 
526 	ctx->irq_count = 0;
527 }
528