1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Support for Partition Mobility/Migration
4  *
5  * Copyright (C) 2010 Nathan Fontenot
6  * Copyright (C) 2010 IBM Corporation
7  */
8 
9 
10 #define pr_fmt(fmt) "mobility: " fmt
11 
12 #include <linux/cpu.h>
13 #include <linux/kernel.h>
14 #include <linux/kobject.h>
15 #include <linux/nmi.h>
16 #include <linux/sched.h>
17 #include <linux/smp.h>
18 #include <linux/stat.h>
19 #include <linux/stop_machine.h>
20 #include <linux/completion.h>
21 #include <linux/device.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/stringify.h>
25 
26 #include <asm/machdep.h>
27 #include <asm/rtas.h>
28 #include "pseries.h"
29 #include "vas.h"	/* vas_migration_handler() */
30 #include "../../kernel/cacheinfo.h"
31 
32 static struct kobject *mobility_kobj;
33 
34 struct update_props_workarea {
35 	__be32 phandle;
36 	__be32 state;
37 	__be64 reserved;
38 	__be32 nprops;
39 } __packed;
40 
41 #define NODE_ACTION_MASK	0xff000000
42 #define NODE_COUNT_MASK		0x00ffffff
43 
44 #define DELETE_DT_NODE	0x01000000
45 #define UPDATE_DT_NODE	0x02000000
46 #define ADD_DT_NODE	0x03000000
47 
48 #define MIGRATION_SCOPE	(1)
49 #define PRRN_SCOPE -2
50 
51 static int mobility_rtas_call(int token, char *buf, s32 scope)
52 {
53 	int rc;
54 
55 	spin_lock(&rtas_data_buf_lock);
56 
57 	memcpy(rtas_data_buf, buf, RTAS_DATA_BUF_SIZE);
58 	rc = rtas_call(token, 2, 1, NULL, rtas_data_buf, scope);
59 	memcpy(buf, rtas_data_buf, RTAS_DATA_BUF_SIZE);
60 
61 	spin_unlock(&rtas_data_buf_lock);
62 	return rc;
63 }
64 
65 static int delete_dt_node(struct device_node *dn)
66 {
67 	struct device_node *pdn;
68 	bool is_platfac;
69 
70 	pdn = of_get_parent(dn);
71 	is_platfac = of_node_is_type(dn, "ibm,platform-facilities") ||
72 		     of_node_is_type(pdn, "ibm,platform-facilities");
73 	of_node_put(pdn);
74 
75 	/*
76 	 * The drivers that bind to nodes in the platform-facilities
77 	 * hierarchy don't support node removal, and the removal directive
78 	 * from firmware is always followed by an add of an equivalent
79 	 * node. The capability (e.g. RNG, encryption, compression)
80 	 * represented by the node is never interrupted by the migration.
81 	 * So ignore changes to this part of the tree.
82 	 */
83 	if (is_platfac) {
84 		pr_notice("ignoring remove operation for %pOFfp\n", dn);
85 		return 0;
86 	}
87 
88 	pr_debug("removing node %pOFfp\n", dn);
89 	dlpar_detach_node(dn);
90 	return 0;
91 }
92 
93 static int update_dt_property(struct device_node *dn, struct property **prop,
94 			      const char *name, u32 vd, char *value)
95 {
96 	struct property *new_prop = *prop;
97 	int more = 0;
98 
99 	/* A negative 'vd' value indicates that only part of the new property
100 	 * value is contained in the buffer and we need to call
101 	 * ibm,update-properties again to get the rest of the value.
102 	 *
103 	 * A negative value is also the two's compliment of the actual value.
104 	 */
105 	if (vd & 0x80000000) {
106 		vd = ~vd + 1;
107 		more = 1;
108 	}
109 
110 	if (new_prop) {
111 		/* partial property fixup */
112 		char *new_data = kzalloc(new_prop->length + vd, GFP_KERNEL);
113 		if (!new_data)
114 			return -ENOMEM;
115 
116 		memcpy(new_data, new_prop->value, new_prop->length);
117 		memcpy(new_data + new_prop->length, value, vd);
118 
119 		kfree(new_prop->value);
120 		new_prop->value = new_data;
121 		new_prop->length += vd;
122 	} else {
123 		new_prop = kzalloc(sizeof(*new_prop), GFP_KERNEL);
124 		if (!new_prop)
125 			return -ENOMEM;
126 
127 		new_prop->name = kstrdup(name, GFP_KERNEL);
128 		if (!new_prop->name) {
129 			kfree(new_prop);
130 			return -ENOMEM;
131 		}
132 
133 		new_prop->length = vd;
134 		new_prop->value = kzalloc(new_prop->length, GFP_KERNEL);
135 		if (!new_prop->value) {
136 			kfree(new_prop->name);
137 			kfree(new_prop);
138 			return -ENOMEM;
139 		}
140 
141 		memcpy(new_prop->value, value, vd);
142 		*prop = new_prop;
143 	}
144 
145 	if (!more) {
146 		pr_debug("updating node %pOF property %s\n", dn, name);
147 		of_update_property(dn, new_prop);
148 		*prop = NULL;
149 	}
150 
151 	return 0;
152 }
153 
154 static int update_dt_node(struct device_node *dn, s32 scope)
155 {
156 	struct update_props_workarea *upwa;
157 	struct property *prop = NULL;
158 	int i, rc, rtas_rc;
159 	char *prop_data;
160 	char *rtas_buf;
161 	int update_properties_token;
162 	u32 nprops;
163 	u32 vd;
164 
165 	update_properties_token = rtas_token("ibm,update-properties");
166 	if (update_properties_token == RTAS_UNKNOWN_SERVICE)
167 		return -EINVAL;
168 
169 	rtas_buf = kzalloc(RTAS_DATA_BUF_SIZE, GFP_KERNEL);
170 	if (!rtas_buf)
171 		return -ENOMEM;
172 
173 	upwa = (struct update_props_workarea *)&rtas_buf[0];
174 	upwa->phandle = cpu_to_be32(dn->phandle);
175 
176 	do {
177 		rtas_rc = mobility_rtas_call(update_properties_token, rtas_buf,
178 					scope);
179 		if (rtas_rc < 0)
180 			break;
181 
182 		prop_data = rtas_buf + sizeof(*upwa);
183 		nprops = be32_to_cpu(upwa->nprops);
184 
185 		/* On the first call to ibm,update-properties for a node the
186 		 * the first property value descriptor contains an empty
187 		 * property name, the property value length encoded as u32,
188 		 * and the property value is the node path being updated.
189 		 */
190 		if (*prop_data == 0) {
191 			prop_data++;
192 			vd = be32_to_cpu(*(__be32 *)prop_data);
193 			prop_data += vd + sizeof(vd);
194 			nprops--;
195 		}
196 
197 		for (i = 0; i < nprops; i++) {
198 			char *prop_name;
199 
200 			prop_name = prop_data;
201 			prop_data += strlen(prop_name) + 1;
202 			vd = be32_to_cpu(*(__be32 *)prop_data);
203 			prop_data += sizeof(vd);
204 
205 			switch (vd) {
206 			case 0x00000000:
207 				/* name only property, nothing to do */
208 				break;
209 
210 			case 0x80000000:
211 				of_remove_property(dn, of_find_property(dn,
212 							prop_name, NULL));
213 				prop = NULL;
214 				break;
215 
216 			default:
217 				rc = update_dt_property(dn, &prop, prop_name,
218 							vd, prop_data);
219 				if (rc) {
220 					pr_err("updating %s property failed: %d\n",
221 					       prop_name, rc);
222 				}
223 
224 				prop_data += vd;
225 				break;
226 			}
227 
228 			cond_resched();
229 		}
230 
231 		cond_resched();
232 	} while (rtas_rc == 1);
233 
234 	kfree(rtas_buf);
235 	return 0;
236 }
237 
238 static int add_dt_node(struct device_node *parent_dn, __be32 drc_index)
239 {
240 	struct device_node *dn;
241 	int rc;
242 
243 	dn = dlpar_configure_connector(drc_index, parent_dn);
244 	if (!dn)
245 		return -ENOENT;
246 
247 	/*
248 	 * Since delete_dt_node() ignores this node type, this is the
249 	 * necessary counterpart. We also know that a platform-facilities
250 	 * node returned from dlpar_configure_connector() has children
251 	 * attached, and dlpar_attach_node() only adds the parent, leaking
252 	 * the children. So ignore these on the add side for now.
253 	 */
254 	if (of_node_is_type(dn, "ibm,platform-facilities")) {
255 		pr_notice("ignoring add operation for %pOF\n", dn);
256 		dlpar_free_cc_nodes(dn);
257 		return 0;
258 	}
259 
260 	rc = dlpar_attach_node(dn, parent_dn);
261 	if (rc)
262 		dlpar_free_cc_nodes(dn);
263 
264 	pr_debug("added node %pOFfp\n", dn);
265 
266 	return rc;
267 }
268 
269 static int pseries_devicetree_update(s32 scope)
270 {
271 	char *rtas_buf;
272 	__be32 *data;
273 	int update_nodes_token;
274 	int rc;
275 
276 	update_nodes_token = rtas_token("ibm,update-nodes");
277 	if (update_nodes_token == RTAS_UNKNOWN_SERVICE)
278 		return 0;
279 
280 	rtas_buf = kzalloc(RTAS_DATA_BUF_SIZE, GFP_KERNEL);
281 	if (!rtas_buf)
282 		return -ENOMEM;
283 
284 	do {
285 		rc = mobility_rtas_call(update_nodes_token, rtas_buf, scope);
286 		if (rc && rc != 1)
287 			break;
288 
289 		data = (__be32 *)rtas_buf + 4;
290 		while (be32_to_cpu(*data) & NODE_ACTION_MASK) {
291 			int i;
292 			u32 action = be32_to_cpu(*data) & NODE_ACTION_MASK;
293 			u32 node_count = be32_to_cpu(*data) & NODE_COUNT_MASK;
294 
295 			data++;
296 
297 			for (i = 0; i < node_count; i++) {
298 				struct device_node *np;
299 				__be32 phandle = *data++;
300 				__be32 drc_index;
301 
302 				np = of_find_node_by_phandle(be32_to_cpu(phandle));
303 				if (!np) {
304 					pr_warn("Failed lookup: phandle 0x%x for action 0x%x\n",
305 						be32_to_cpu(phandle), action);
306 					continue;
307 				}
308 
309 				switch (action) {
310 				case DELETE_DT_NODE:
311 					delete_dt_node(np);
312 					break;
313 				case UPDATE_DT_NODE:
314 					update_dt_node(np, scope);
315 					break;
316 				case ADD_DT_NODE:
317 					drc_index = *data++;
318 					add_dt_node(np, drc_index);
319 					break;
320 				}
321 
322 				of_node_put(np);
323 				cond_resched();
324 			}
325 		}
326 
327 		cond_resched();
328 	} while (rc == 1);
329 
330 	kfree(rtas_buf);
331 	return rc;
332 }
333 
334 void post_mobility_fixup(void)
335 {
336 	int rc;
337 
338 	rtas_activate_firmware();
339 
340 	/*
341 	 * We don't want CPUs to go online/offline while the device
342 	 * tree is being updated.
343 	 */
344 	cpus_read_lock();
345 
346 	/*
347 	 * It's common for the destination firmware to replace cache
348 	 * nodes.  Release all of the cacheinfo hierarchy's references
349 	 * before updating the device tree.
350 	 */
351 	cacheinfo_teardown();
352 
353 	rc = pseries_devicetree_update(MIGRATION_SCOPE);
354 	if (rc)
355 		pr_err("device tree update failed: %d\n", rc);
356 
357 	cacheinfo_rebuild();
358 
359 	cpus_read_unlock();
360 
361 	/* Possibly switch to a new L1 flush type */
362 	pseries_setup_security_mitigations();
363 
364 	/* Reinitialise system information for hv-24x7 */
365 	read_24x7_sys_info();
366 
367 	return;
368 }
369 
370 static int poll_vasi_state(u64 handle, unsigned long *res)
371 {
372 	unsigned long retbuf[PLPAR_HCALL_BUFSIZE];
373 	long hvrc;
374 	int ret;
375 
376 	hvrc = plpar_hcall(H_VASI_STATE, retbuf, handle);
377 	switch (hvrc) {
378 	case H_SUCCESS:
379 		ret = 0;
380 		*res = retbuf[0];
381 		break;
382 	case H_PARAMETER:
383 		ret = -EINVAL;
384 		break;
385 	case H_FUNCTION:
386 		ret = -EOPNOTSUPP;
387 		break;
388 	case H_HARDWARE:
389 	default:
390 		pr_err("unexpected H_VASI_STATE result %ld\n", hvrc);
391 		ret = -EIO;
392 		break;
393 	}
394 	return ret;
395 }
396 
397 static int wait_for_vasi_session_suspending(u64 handle)
398 {
399 	unsigned long state;
400 	int ret;
401 
402 	/*
403 	 * Wait for transition from H_VASI_ENABLED to
404 	 * H_VASI_SUSPENDING. Treat anything else as an error.
405 	 */
406 	while (true) {
407 		ret = poll_vasi_state(handle, &state);
408 
409 		if (ret != 0 || state == H_VASI_SUSPENDING) {
410 			break;
411 		} else if (state == H_VASI_ENABLED) {
412 			ssleep(1);
413 		} else {
414 			pr_err("unexpected H_VASI_STATE result %lu\n", state);
415 			ret = -EIO;
416 			break;
417 		}
418 	}
419 
420 	/*
421 	 * Proceed even if H_VASI_STATE is unavailable. If H_JOIN or
422 	 * ibm,suspend-me are also unimplemented, we'll recover then.
423 	 */
424 	if (ret == -EOPNOTSUPP)
425 		ret = 0;
426 
427 	return ret;
428 }
429 
430 static void prod_single(unsigned int target_cpu)
431 {
432 	long hvrc;
433 	int hwid;
434 
435 	hwid = get_hard_smp_processor_id(target_cpu);
436 	hvrc = plpar_hcall_norets(H_PROD, hwid);
437 	if (hvrc == H_SUCCESS)
438 		return;
439 	pr_err_ratelimited("H_PROD of CPU %u (hwid %d) error: %ld\n",
440 			   target_cpu, hwid, hvrc);
441 }
442 
443 static void prod_others(void)
444 {
445 	unsigned int cpu;
446 
447 	for_each_online_cpu(cpu) {
448 		if (cpu != smp_processor_id())
449 			prod_single(cpu);
450 	}
451 }
452 
453 static u16 clamp_slb_size(void)
454 {
455 #ifdef CONFIG_PPC_64S_HASH_MMU
456 	u16 prev = mmu_slb_size;
457 
458 	slb_set_size(SLB_MIN_SIZE);
459 
460 	return prev;
461 #else
462 	return 0;
463 #endif
464 }
465 
466 static int do_suspend(void)
467 {
468 	u16 saved_slb_size;
469 	int status;
470 	int ret;
471 
472 	pr_info("calling ibm,suspend-me on CPU %i\n", smp_processor_id());
473 
474 	/*
475 	 * The destination processor model may have fewer SLB entries
476 	 * than the source. We reduce mmu_slb_size to a safe minimum
477 	 * before suspending in order to minimize the possibility of
478 	 * programming non-existent entries on the destination. If
479 	 * suspend fails, we restore it before returning. On success
480 	 * the OF reconfig path will update it from the new device
481 	 * tree after resuming on the destination.
482 	 */
483 	saved_slb_size = clamp_slb_size();
484 
485 	ret = rtas_ibm_suspend_me(&status);
486 	if (ret != 0) {
487 		pr_err("ibm,suspend-me error: %d\n", status);
488 		slb_set_size(saved_slb_size);
489 	}
490 
491 	return ret;
492 }
493 
494 /**
495  * struct pseries_suspend_info - State shared between CPUs for join/suspend.
496  * @counter: Threads are to increment this upon resuming from suspend
497  *           or if an error is received from H_JOIN. The thread which performs
498  *           the first increment (i.e. sets it to 1) is responsible for
499  *           waking the other threads.
500  * @done: False if join/suspend is in progress. True if the operation is
501  *        complete (successful or not).
502  */
503 struct pseries_suspend_info {
504 	atomic_t counter;
505 	bool done;
506 };
507 
508 static int do_join(void *arg)
509 {
510 	struct pseries_suspend_info *info = arg;
511 	atomic_t *counter = &info->counter;
512 	long hvrc;
513 	int ret;
514 
515 retry:
516 	/* Must ensure MSR.EE off for H_JOIN. */
517 	hard_irq_disable();
518 	hvrc = plpar_hcall_norets(H_JOIN);
519 
520 	switch (hvrc) {
521 	case H_CONTINUE:
522 		/*
523 		 * All other CPUs are offline or in H_JOIN. This CPU
524 		 * attempts the suspend.
525 		 */
526 		ret = do_suspend();
527 		break;
528 	case H_SUCCESS:
529 		/*
530 		 * The suspend is complete and this cpu has received a
531 		 * prod, or we've received a stray prod from unrelated
532 		 * code (e.g. paravirt spinlocks) and we need to join
533 		 * again.
534 		 *
535 		 * This barrier orders the return from H_JOIN above vs
536 		 * the load of info->done. It pairs with the barrier
537 		 * in the wakeup/prod path below.
538 		 */
539 		smp_mb();
540 		if (READ_ONCE(info->done) == false) {
541 			pr_info_ratelimited("premature return from H_JOIN on CPU %i, retrying",
542 					    smp_processor_id());
543 			goto retry;
544 		}
545 		ret = 0;
546 		break;
547 	case H_BAD_MODE:
548 	case H_HARDWARE:
549 	default:
550 		ret = -EIO;
551 		pr_err_ratelimited("H_JOIN error %ld on CPU %i\n",
552 				   hvrc, smp_processor_id());
553 		break;
554 	}
555 
556 	if (atomic_inc_return(counter) == 1) {
557 		pr_info("CPU %u waking all threads\n", smp_processor_id());
558 		WRITE_ONCE(info->done, true);
559 		/*
560 		 * This barrier orders the store to info->done vs subsequent
561 		 * H_PRODs to wake the other CPUs. It pairs with the barrier
562 		 * in the H_SUCCESS case above.
563 		 */
564 		smp_mb();
565 		prod_others();
566 	}
567 	/*
568 	 * Execution may have been suspended for several seconds, so
569 	 * reset the watchdog.
570 	 */
571 	touch_nmi_watchdog();
572 	return ret;
573 }
574 
575 /*
576  * Abort reason code byte 0. We use only the 'Migrating partition' value.
577  */
578 enum vasi_aborting_entity {
579 	ORCHESTRATOR        = 1,
580 	VSP_SOURCE          = 2,
581 	PARTITION_FIRMWARE  = 3,
582 	PLATFORM_FIRMWARE   = 4,
583 	VSP_TARGET          = 5,
584 	MIGRATING_PARTITION = 6,
585 };
586 
587 static void pseries_cancel_migration(u64 handle, int err)
588 {
589 	u32 reason_code;
590 	u32 detail;
591 	u8 entity;
592 	long hvrc;
593 
594 	entity = MIGRATING_PARTITION;
595 	detail = abs(err) & 0xffffff;
596 	reason_code = (entity << 24) | detail;
597 
598 	hvrc = plpar_hcall_norets(H_VASI_SIGNAL, handle,
599 				  H_VASI_SIGNAL_CANCEL, reason_code);
600 	if (hvrc)
601 		pr_err("H_VASI_SIGNAL error: %ld\n", hvrc);
602 }
603 
604 static int pseries_suspend(u64 handle)
605 {
606 	const unsigned int max_attempts = 5;
607 	unsigned int retry_interval_ms = 1;
608 	unsigned int attempt = 1;
609 	int ret;
610 
611 	while (true) {
612 		struct pseries_suspend_info info;
613 		unsigned long vasi_state;
614 		int vasi_err;
615 
616 		info = (struct pseries_suspend_info) {
617 			.counter = ATOMIC_INIT(0),
618 			.done = false,
619 		};
620 
621 		ret = stop_machine(do_join, &info, cpu_online_mask);
622 		if (ret == 0)
623 			break;
624 		/*
625 		 * Encountered an error. If the VASI stream is still
626 		 * in Suspending state, it's likely a transient
627 		 * condition related to some device in the partition
628 		 * and we can retry in the hope that the cause has
629 		 * cleared after some delay.
630 		 *
631 		 * A better design would allow drivers etc to prepare
632 		 * for the suspend and avoid conditions which prevent
633 		 * the suspend from succeeding. For now, we have this
634 		 * mitigation.
635 		 */
636 		pr_notice("Partition suspend attempt %u of %u error: %d\n",
637 			  attempt, max_attempts, ret);
638 
639 		if (attempt == max_attempts)
640 			break;
641 
642 		vasi_err = poll_vasi_state(handle, &vasi_state);
643 		if (vasi_err == 0) {
644 			if (vasi_state != H_VASI_SUSPENDING) {
645 				pr_notice("VASI state %lu after failed suspend\n",
646 					  vasi_state);
647 				break;
648 			}
649 		} else if (vasi_err != -EOPNOTSUPP) {
650 			pr_err("VASI state poll error: %d", vasi_err);
651 			break;
652 		}
653 
654 		pr_notice("Will retry partition suspend after %u ms\n",
655 			  retry_interval_ms);
656 
657 		msleep(retry_interval_ms);
658 		retry_interval_ms *= 10;
659 		attempt++;
660 	}
661 
662 	return ret;
663 }
664 
665 static int pseries_migrate_partition(u64 handle)
666 {
667 	int ret;
668 
669 	ret = wait_for_vasi_session_suspending(handle);
670 	if (ret)
671 		return ret;
672 
673 	vas_migration_handler(VAS_SUSPEND);
674 
675 	ret = pseries_suspend(handle);
676 	if (ret == 0)
677 		post_mobility_fixup();
678 	else
679 		pseries_cancel_migration(handle, ret);
680 
681 	vas_migration_handler(VAS_RESUME);
682 
683 	return ret;
684 }
685 
686 int rtas_syscall_dispatch_ibm_suspend_me(u64 handle)
687 {
688 	return pseries_migrate_partition(handle);
689 }
690 
691 static ssize_t migration_store(struct class *class,
692 			       struct class_attribute *attr, const char *buf,
693 			       size_t count)
694 {
695 	u64 streamid;
696 	int rc;
697 
698 	rc = kstrtou64(buf, 0, &streamid);
699 	if (rc)
700 		return rc;
701 
702 	rc = pseries_migrate_partition(streamid);
703 	if (rc)
704 		return rc;
705 
706 	return count;
707 }
708 
709 /*
710  * Used by drmgr to determine the kernel behavior of the migration interface.
711  *
712  * Version 1: Performs all PAPR requirements for migration including
713  *	firmware activation and device tree update.
714  */
715 #define MIGRATION_API_VERSION	1
716 
717 static CLASS_ATTR_WO(migration);
718 static CLASS_ATTR_STRING(api_version, 0444, __stringify(MIGRATION_API_VERSION));
719 
720 static int __init mobility_sysfs_init(void)
721 {
722 	int rc;
723 
724 	mobility_kobj = kobject_create_and_add("mobility", kernel_kobj);
725 	if (!mobility_kobj)
726 		return -ENOMEM;
727 
728 	rc = sysfs_create_file(mobility_kobj, &class_attr_migration.attr);
729 	if (rc)
730 		pr_err("unable to create migration sysfs file (%d)\n", rc);
731 
732 	rc = sysfs_create_file(mobility_kobj, &class_attr_api_version.attr.attr);
733 	if (rc)
734 		pr_err("unable to create api_version sysfs file (%d)\n", rc);
735 
736 	return 0;
737 }
738 machine_device_initcall(pseries, mobility_sysfs_init);
739