xref: /linux/drivers/misc/ocxl/link.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 // SPDX-License-Identifier: GPL-2.0+
2 // Copyright 2017 IBM Corp.
3 #include <linux/sched/mm.h>
4 #include <linux/mutex.h>
5 #include <linux/mm.h>
6 #include <linux/mm_types.h>
7 #include <linux/mmu_context.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/irqdomain.h>
10 #include <asm/copro.h>
11 #include <asm/pnv-ocxl.h>
12 #include <asm/xive.h>
13 #include <misc/ocxl.h>
14 #include "ocxl_internal.h"
15 #include "trace.h"
16 
17 
18 #define SPA_PASID_BITS		15
19 #define SPA_PASID_MAX		((1 << SPA_PASID_BITS) - 1)
20 #define SPA_PE_MASK		SPA_PASID_MAX
21 #define SPA_SPA_SIZE_LOG	22 /* Each SPA is 4 Mb */
22 
23 #define SPA_CFG_SF		(1ull << (63-0))
24 #define SPA_CFG_TA		(1ull << (63-1))
25 #define SPA_CFG_HV		(1ull << (63-3))
26 #define SPA_CFG_UV		(1ull << (63-4))
27 #define SPA_CFG_XLAT_hpt	(0ull << (63-6)) /* Hashed page table (HPT) mode */
28 #define SPA_CFG_XLAT_roh	(2ull << (63-6)) /* Radix on HPT mode */
29 #define SPA_CFG_XLAT_ror	(3ull << (63-6)) /* Radix on Radix mode */
30 #define SPA_CFG_PR		(1ull << (63-49))
31 #define SPA_CFG_TC		(1ull << (63-54))
32 #define SPA_CFG_DR		(1ull << (63-59))
33 
34 #define SPA_XSL_TF		(1ull << (63-3))  /* Translation fault */
35 #define SPA_XSL_S		(1ull << (63-38)) /* Store operation */
36 
37 #define SPA_PE_VALID		0x80000000
38 
39 struct ocxl_link;
40 
41 struct pe_data {
42 	struct mm_struct *mm;
43 	/* callback to trigger when a translation fault occurs */
44 	void (*xsl_err_cb)(void *data, u64 addr, u64 dsisr);
45 	/* opaque pointer to be passed to the above callback */
46 	void *xsl_err_data;
47 	struct rcu_head rcu;
48 	struct ocxl_link *link;
49 	struct mmu_notifier mmu_notifier;
50 };
51 
52 struct spa {
53 	struct ocxl_process_element *spa_mem;
54 	int spa_order;
55 	struct mutex spa_lock;
56 	struct radix_tree_root pe_tree; /* Maps PE handles to pe_data */
57 	char *irq_name;
58 	int virq;
59 	void __iomem *reg_dsisr;
60 	void __iomem *reg_dar;
61 	void __iomem *reg_tfc;
62 	void __iomem *reg_pe_handle;
63 	/*
64 	 * The following field are used by the memory fault
65 	 * interrupt handler. We can only have one interrupt at a
66 	 * time. The NPU won't raise another interrupt until the
67 	 * previous one has been ack'd by writing to the TFC register
68 	 */
69 	struct xsl_fault {
70 		struct work_struct fault_work;
71 		u64 pe;
72 		u64 dsisr;
73 		u64 dar;
74 		struct pe_data pe_data;
75 	} xsl_fault;
76 };
77 
78 /*
79  * A opencapi link can be used be by several PCI functions. We have
80  * one link per device slot.
81  *
82  * A linked list of opencapi links should suffice, as there's a
83  * limited number of opencapi slots on a system and lookup is only
84  * done when the device is probed
85  */
86 struct ocxl_link {
87 	struct list_head list;
88 	struct kref ref;
89 	int domain;
90 	int bus;
91 	int dev;
92 	void __iomem *arva;     /* ATSD register virtual address */
93 	spinlock_t atsd_lock;   /* to serialize shootdowns */
94 	atomic_t irq_available;
95 	struct spa *spa;
96 	void *platform_data;
97 };
98 static LIST_HEAD(links_list);
99 static DEFINE_MUTEX(links_list_lock);
100 
101 enum xsl_response {
102 	CONTINUE,
103 	ADDRESS_ERROR,
104 	RESTART,
105 };
106 
107 
108 static void read_irq(struct spa *spa, u64 *dsisr, u64 *dar, u64 *pe)
109 {
110 	u64 reg;
111 
112 	*dsisr = in_be64(spa->reg_dsisr);
113 	*dar = in_be64(spa->reg_dar);
114 	reg = in_be64(spa->reg_pe_handle);
115 	*pe = reg & SPA_PE_MASK;
116 }
117 
118 static void ack_irq(struct spa *spa, enum xsl_response r)
119 {
120 	u64 reg = 0;
121 
122 	/* continue is not supported */
123 	if (r == RESTART)
124 		reg = PPC_BIT(31);
125 	else if (r == ADDRESS_ERROR)
126 		reg = PPC_BIT(30);
127 	else
128 		WARN(1, "Invalid irq response %d\n", r);
129 
130 	if (reg) {
131 		trace_ocxl_fault_ack(spa->spa_mem, spa->xsl_fault.pe,
132 				spa->xsl_fault.dsisr, spa->xsl_fault.dar, reg);
133 		out_be64(spa->reg_tfc, reg);
134 	}
135 }
136 
137 static void xsl_fault_handler_bh(struct work_struct *fault_work)
138 {
139 	vm_fault_t flt = 0;
140 	unsigned long access, flags, inv_flags = 0;
141 	enum xsl_response r;
142 	struct xsl_fault *fault = container_of(fault_work, struct xsl_fault,
143 					fault_work);
144 	struct spa *spa = container_of(fault, struct spa, xsl_fault);
145 
146 	int rc;
147 
148 	/*
149 	 * We must release a reference on mm_users whenever exiting this
150 	 * function (taken in the memory fault interrupt handler)
151 	 */
152 	rc = copro_handle_mm_fault(fault->pe_data.mm, fault->dar, fault->dsisr,
153 				&flt);
154 	if (rc) {
155 		pr_debug("copro_handle_mm_fault failed: %d\n", rc);
156 		if (fault->pe_data.xsl_err_cb) {
157 			fault->pe_data.xsl_err_cb(
158 				fault->pe_data.xsl_err_data,
159 				fault->dar, fault->dsisr);
160 		}
161 		r = ADDRESS_ERROR;
162 		goto ack;
163 	}
164 
165 	if (!radix_enabled()) {
166 		/*
167 		 * update_mmu_cache() will not have loaded the hash
168 		 * since current->trap is not a 0x400 or 0x300, so
169 		 * just call hash_page_mm() here.
170 		 */
171 		access = _PAGE_PRESENT | _PAGE_READ;
172 		if (fault->dsisr & SPA_XSL_S)
173 			access |= _PAGE_WRITE;
174 
175 		if (get_region_id(fault->dar) != USER_REGION_ID)
176 			access |= _PAGE_PRIVILEGED;
177 
178 		local_irq_save(flags);
179 		hash_page_mm(fault->pe_data.mm, fault->dar, access, 0x300,
180 			inv_flags);
181 		local_irq_restore(flags);
182 	}
183 	r = RESTART;
184 ack:
185 	mmput(fault->pe_data.mm);
186 	ack_irq(spa, r);
187 }
188 
189 static irqreturn_t xsl_fault_handler(int irq, void *data)
190 {
191 	struct ocxl_link *link = data;
192 	struct spa *spa = link->spa;
193 	u64 dsisr, dar, pe_handle;
194 	struct pe_data *pe_data;
195 	struct ocxl_process_element *pe;
196 	int pid;
197 	bool schedule = false;
198 
199 	read_irq(spa, &dsisr, &dar, &pe_handle);
200 	trace_ocxl_fault(spa->spa_mem, pe_handle, dsisr, dar, -1);
201 
202 	WARN_ON(pe_handle > SPA_PE_MASK);
203 	pe = spa->spa_mem + pe_handle;
204 	pid = be32_to_cpu(pe->pid);
205 	/* We could be reading all null values here if the PE is being
206 	 * removed while an interrupt kicks in. It's not supposed to
207 	 * happen if the driver notified the AFU to terminate the
208 	 * PASID, and the AFU waited for pending operations before
209 	 * acknowledging. But even if it happens, we won't find a
210 	 * memory context below and fail silently, so it should be ok.
211 	 */
212 	if (!(dsisr & SPA_XSL_TF)) {
213 		WARN(1, "Invalid xsl interrupt fault register %#llx\n", dsisr);
214 		ack_irq(spa, ADDRESS_ERROR);
215 		return IRQ_HANDLED;
216 	}
217 
218 	rcu_read_lock();
219 	pe_data = radix_tree_lookup(&spa->pe_tree, pe_handle);
220 	if (!pe_data) {
221 		/*
222 		 * Could only happen if the driver didn't notify the
223 		 * AFU about PASID termination before removing the PE,
224 		 * or the AFU didn't wait for all memory access to
225 		 * have completed.
226 		 *
227 		 * Either way, we fail early, but we shouldn't log an
228 		 * error message, as it is a valid (if unexpected)
229 		 * scenario
230 		 */
231 		rcu_read_unlock();
232 		pr_debug("Unknown mm context for xsl interrupt\n");
233 		ack_irq(spa, ADDRESS_ERROR);
234 		return IRQ_HANDLED;
235 	}
236 
237 	if (!pe_data->mm) {
238 		/*
239 		 * translation fault from a kernel context - an OpenCAPI
240 		 * device tried to access a bad kernel address
241 		 */
242 		rcu_read_unlock();
243 		pr_warn("Unresolved OpenCAPI xsl fault in kernel context\n");
244 		ack_irq(spa, ADDRESS_ERROR);
245 		return IRQ_HANDLED;
246 	}
247 	WARN_ON(pe_data->mm->context.id != pid);
248 
249 	if (mmget_not_zero(pe_data->mm)) {
250 			spa->xsl_fault.pe = pe_handle;
251 			spa->xsl_fault.dar = dar;
252 			spa->xsl_fault.dsisr = dsisr;
253 			spa->xsl_fault.pe_data = *pe_data;
254 			schedule = true;
255 			/* mm_users count released by bottom half */
256 	}
257 	rcu_read_unlock();
258 	if (schedule)
259 		schedule_work(&spa->xsl_fault.fault_work);
260 	else
261 		ack_irq(spa, ADDRESS_ERROR);
262 	return IRQ_HANDLED;
263 }
264 
265 static void unmap_irq_registers(struct spa *spa)
266 {
267 	pnv_ocxl_unmap_xsl_regs(spa->reg_dsisr, spa->reg_dar, spa->reg_tfc,
268 				spa->reg_pe_handle);
269 }
270 
271 static int map_irq_registers(struct pci_dev *dev, struct spa *spa)
272 {
273 	return pnv_ocxl_map_xsl_regs(dev, &spa->reg_dsisr, &spa->reg_dar,
274 				&spa->reg_tfc, &spa->reg_pe_handle);
275 }
276 
277 static int setup_xsl_irq(struct pci_dev *dev, struct ocxl_link *link)
278 {
279 	struct spa *spa = link->spa;
280 	int rc;
281 	int hwirq;
282 
283 	rc = pnv_ocxl_get_xsl_irq(dev, &hwirq);
284 	if (rc)
285 		return rc;
286 
287 	rc = map_irq_registers(dev, spa);
288 	if (rc)
289 		return rc;
290 
291 	spa->irq_name = kasprintf(GFP_KERNEL, "ocxl-xsl-%x-%x-%x",
292 				link->domain, link->bus, link->dev);
293 	if (!spa->irq_name) {
294 		dev_err(&dev->dev, "Can't allocate name for xsl interrupt\n");
295 		rc = -ENOMEM;
296 		goto err_xsl;
297 	}
298 	/*
299 	 * At some point, we'll need to look into allowing a higher
300 	 * number of interrupts. Could we have an IRQ domain per link?
301 	 */
302 	spa->virq = irq_create_mapping(NULL, hwirq);
303 	if (!spa->virq) {
304 		dev_err(&dev->dev,
305 			"irq_create_mapping failed for translation interrupt\n");
306 		rc = -EINVAL;
307 		goto err_name;
308 	}
309 
310 	dev_dbg(&dev->dev, "hwirq %d mapped to virq %d\n", hwirq, spa->virq);
311 
312 	rc = request_irq(spa->virq, xsl_fault_handler, 0, spa->irq_name,
313 			link);
314 	if (rc) {
315 		dev_err(&dev->dev,
316 			"request_irq failed for translation interrupt: %d\n",
317 			rc);
318 		rc = -EINVAL;
319 		goto err_mapping;
320 	}
321 	return 0;
322 
323 err_mapping:
324 	irq_dispose_mapping(spa->virq);
325 err_name:
326 	kfree(spa->irq_name);
327 err_xsl:
328 	unmap_irq_registers(spa);
329 	return rc;
330 }
331 
332 static void release_xsl_irq(struct ocxl_link *link)
333 {
334 	struct spa *spa = link->spa;
335 
336 	if (spa->virq) {
337 		free_irq(spa->virq, link);
338 		irq_dispose_mapping(spa->virq);
339 	}
340 	kfree(spa->irq_name);
341 	unmap_irq_registers(spa);
342 }
343 
344 static int alloc_spa(struct pci_dev *dev, struct ocxl_link *link)
345 {
346 	struct spa *spa;
347 
348 	spa = kzalloc(sizeof(struct spa), GFP_KERNEL);
349 	if (!spa)
350 		return -ENOMEM;
351 
352 	mutex_init(&spa->spa_lock);
353 	INIT_RADIX_TREE(&spa->pe_tree, GFP_KERNEL);
354 	INIT_WORK(&spa->xsl_fault.fault_work, xsl_fault_handler_bh);
355 
356 	spa->spa_order = SPA_SPA_SIZE_LOG - PAGE_SHIFT;
357 	spa->spa_mem = (struct ocxl_process_element *)
358 		__get_free_pages(GFP_KERNEL | __GFP_ZERO, spa->spa_order);
359 	if (!spa->spa_mem) {
360 		dev_err(&dev->dev, "Can't allocate Shared Process Area\n");
361 		kfree(spa);
362 		return -ENOMEM;
363 	}
364 	pr_debug("Allocated SPA for %x:%x:%x at %p\n", link->domain, link->bus,
365 		link->dev, spa->spa_mem);
366 
367 	link->spa = spa;
368 	return 0;
369 }
370 
371 static void free_spa(struct ocxl_link *link)
372 {
373 	struct spa *spa = link->spa;
374 
375 	pr_debug("Freeing SPA for %x:%x:%x\n", link->domain, link->bus,
376 		link->dev);
377 
378 	if (spa && spa->spa_mem) {
379 		free_pages((unsigned long) spa->spa_mem, spa->spa_order);
380 		kfree(spa);
381 		link->spa = NULL;
382 	}
383 }
384 
385 static int alloc_link(struct pci_dev *dev, int PE_mask, struct ocxl_link **out_link)
386 {
387 	struct ocxl_link *link;
388 	int rc;
389 
390 	link = kzalloc(sizeof(struct ocxl_link), GFP_KERNEL);
391 	if (!link)
392 		return -ENOMEM;
393 
394 	kref_init(&link->ref);
395 	link->domain = pci_domain_nr(dev->bus);
396 	link->bus = dev->bus->number;
397 	link->dev = PCI_SLOT(dev->devfn);
398 	atomic_set(&link->irq_available, MAX_IRQ_PER_LINK);
399 	spin_lock_init(&link->atsd_lock);
400 
401 	rc = alloc_spa(dev, link);
402 	if (rc)
403 		goto err_free;
404 
405 	rc = setup_xsl_irq(dev, link);
406 	if (rc)
407 		goto err_spa;
408 
409 	/* platform specific hook */
410 	rc = pnv_ocxl_spa_setup(dev, link->spa->spa_mem, PE_mask,
411 				&link->platform_data);
412 	if (rc)
413 		goto err_xsl_irq;
414 
415 	/* if link->arva is not defeined, MMIO registers are not used to
416 	 * generate TLB invalidate. PowerBus snooping is enabled.
417 	 * Otherwise, PowerBus snooping is disabled. TLB Invalidates are
418 	 * initiated using MMIO registers.
419 	 */
420 	pnv_ocxl_map_lpar(dev, mfspr(SPRN_LPID), 0, &link->arva);
421 
422 	*out_link = link;
423 	return 0;
424 
425 err_xsl_irq:
426 	release_xsl_irq(link);
427 err_spa:
428 	free_spa(link);
429 err_free:
430 	kfree(link);
431 	return rc;
432 }
433 
434 static void free_link(struct ocxl_link *link)
435 {
436 	release_xsl_irq(link);
437 	free_spa(link);
438 	kfree(link);
439 }
440 
441 int ocxl_link_setup(struct pci_dev *dev, int PE_mask, void **link_handle)
442 {
443 	int rc = 0;
444 	struct ocxl_link *link;
445 
446 	mutex_lock(&links_list_lock);
447 	list_for_each_entry(link, &links_list, list) {
448 		/* The functions of a device all share the same link */
449 		if (link->domain == pci_domain_nr(dev->bus) &&
450 			link->bus == dev->bus->number &&
451 			link->dev == PCI_SLOT(dev->devfn)) {
452 			kref_get(&link->ref);
453 			*link_handle = link;
454 			goto unlock;
455 		}
456 	}
457 	rc = alloc_link(dev, PE_mask, &link);
458 	if (rc)
459 		goto unlock;
460 
461 	list_add(&link->list, &links_list);
462 	*link_handle = link;
463 unlock:
464 	mutex_unlock(&links_list_lock);
465 	return rc;
466 }
467 EXPORT_SYMBOL_GPL(ocxl_link_setup);
468 
469 static void release_xsl(struct kref *ref)
470 {
471 	struct ocxl_link *link = container_of(ref, struct ocxl_link, ref);
472 
473 	if (link->arva) {
474 		pnv_ocxl_unmap_lpar(link->arva);
475 		link->arva = NULL;
476 	}
477 
478 	list_del(&link->list);
479 	/* call platform code before releasing data */
480 	pnv_ocxl_spa_release(link->platform_data);
481 	free_link(link);
482 }
483 
484 void ocxl_link_release(struct pci_dev *dev, void *link_handle)
485 {
486 	struct ocxl_link *link = link_handle;
487 
488 	mutex_lock(&links_list_lock);
489 	kref_put(&link->ref, release_xsl);
490 	mutex_unlock(&links_list_lock);
491 }
492 EXPORT_SYMBOL_GPL(ocxl_link_release);
493 
494 static void arch_invalidate_secondary_tlbs(struct mmu_notifier *mn,
495 					struct mm_struct *mm,
496 					unsigned long start, unsigned long end)
497 {
498 	struct pe_data *pe_data = container_of(mn, struct pe_data, mmu_notifier);
499 	struct ocxl_link *link = pe_data->link;
500 	unsigned long addr, pid, page_size = PAGE_SIZE;
501 
502 	pid = mm->context.id;
503 	trace_ocxl_mmu_notifier_range(start, end, pid);
504 
505 	spin_lock(&link->atsd_lock);
506 	for (addr = start; addr < end; addr += page_size)
507 		pnv_ocxl_tlb_invalidate(link->arva, pid, addr, page_size);
508 	spin_unlock(&link->atsd_lock);
509 }
510 
511 static const struct mmu_notifier_ops ocxl_mmu_notifier_ops = {
512 	.arch_invalidate_secondary_tlbs = arch_invalidate_secondary_tlbs,
513 };
514 
515 static u64 calculate_cfg_state(bool kernel)
516 {
517 	u64 state;
518 
519 	state = SPA_CFG_DR;
520 	if (mfspr(SPRN_LPCR) & LPCR_TC)
521 		state |= SPA_CFG_TC;
522 	if (radix_enabled())
523 		state |= SPA_CFG_XLAT_ror;
524 	else
525 		state |= SPA_CFG_XLAT_hpt;
526 	state |= SPA_CFG_HV;
527 	if (kernel) {
528 		if (mfmsr() & MSR_SF)
529 			state |= SPA_CFG_SF;
530 	} else {
531 		state |= SPA_CFG_PR;
532 		if (!test_tsk_thread_flag(current, TIF_32BIT))
533 			state |= SPA_CFG_SF;
534 	}
535 	return state;
536 }
537 
538 int ocxl_link_add_pe(void *link_handle, int pasid, u32 pidr, u32 tidr,
539 		u64 amr, u16 bdf, struct mm_struct *mm,
540 		void (*xsl_err_cb)(void *data, u64 addr, u64 dsisr),
541 		void *xsl_err_data)
542 {
543 	struct ocxl_link *link = link_handle;
544 	struct spa *spa = link->spa;
545 	struct ocxl_process_element *pe;
546 	int pe_handle, rc = 0;
547 	struct pe_data *pe_data;
548 
549 	BUILD_BUG_ON(sizeof(struct ocxl_process_element) != 128);
550 	if (pasid > SPA_PASID_MAX)
551 		return -EINVAL;
552 
553 	mutex_lock(&spa->spa_lock);
554 	pe_handle = pasid & SPA_PE_MASK;
555 	pe = spa->spa_mem + pe_handle;
556 
557 	if (pe->software_state) {
558 		rc = -EBUSY;
559 		goto unlock;
560 	}
561 
562 	pe_data = kmalloc(sizeof(*pe_data), GFP_KERNEL);
563 	if (!pe_data) {
564 		rc = -ENOMEM;
565 		goto unlock;
566 	}
567 
568 	pe_data->mm = mm;
569 	pe_data->xsl_err_cb = xsl_err_cb;
570 	pe_data->xsl_err_data = xsl_err_data;
571 	pe_data->link = link;
572 	pe_data->mmu_notifier.ops = &ocxl_mmu_notifier_ops;
573 
574 	memset(pe, 0, sizeof(struct ocxl_process_element));
575 	pe->config_state = cpu_to_be64(calculate_cfg_state(pidr == 0));
576 	pe->pasid = cpu_to_be32(pasid << (31 - 19));
577 	pe->bdf = cpu_to_be16(bdf);
578 	pe->lpid = cpu_to_be32(mfspr(SPRN_LPID));
579 	pe->pid = cpu_to_be32(pidr);
580 	pe->tid = cpu_to_be32(tidr);
581 	pe->amr = cpu_to_be64(amr);
582 	pe->software_state = cpu_to_be32(SPA_PE_VALID);
583 
584 	/*
585 	 * For user contexts, register a copro so that TLBIs are seen
586 	 * by the nest MMU. If we have a kernel context, TLBIs are
587 	 * already global.
588 	 */
589 	if (mm) {
590 		mm_context_add_copro(mm);
591 		if (link->arva) {
592 			/* Use MMIO registers for the TLB Invalidate
593 			 * operations.
594 			 */
595 			trace_ocxl_init_mmu_notifier(pasid, mm->context.id);
596 			mmu_notifier_register(&pe_data->mmu_notifier, mm);
597 		}
598 	}
599 
600 	/*
601 	 * Barrier is to make sure PE is visible in the SPA before it
602 	 * is used by the device. It also helps with the global TLBI
603 	 * invalidation
604 	 */
605 	mb();
606 	radix_tree_insert(&spa->pe_tree, pe_handle, pe_data);
607 
608 	/*
609 	 * The mm must stay valid for as long as the device uses it. We
610 	 * lower the count when the context is removed from the SPA.
611 	 *
612 	 * We grab mm_count (and not mm_users), as we don't want to
613 	 * end up in a circular dependency if a process mmaps its
614 	 * mmio, therefore incrementing the file ref count when
615 	 * calling mmap(), and forgets to unmap before exiting. In
616 	 * that scenario, when the kernel handles the death of the
617 	 * process, the file is not cleaned because unmap was not
618 	 * called, and the mm wouldn't be freed because we would still
619 	 * have a reference on mm_users. Incrementing mm_count solves
620 	 * the problem.
621 	 */
622 	if (mm)
623 		mmgrab(mm);
624 	trace_ocxl_context_add(current->pid, spa->spa_mem, pasid, pidr, tidr);
625 unlock:
626 	mutex_unlock(&spa->spa_lock);
627 	return rc;
628 }
629 EXPORT_SYMBOL_GPL(ocxl_link_add_pe);
630 
631 int ocxl_link_update_pe(void *link_handle, int pasid, __u16 tid)
632 {
633 	struct ocxl_link *link = link_handle;
634 	struct spa *spa = link->spa;
635 	struct ocxl_process_element *pe;
636 	int pe_handle, rc;
637 
638 	if (pasid > SPA_PASID_MAX)
639 		return -EINVAL;
640 
641 	pe_handle = pasid & SPA_PE_MASK;
642 	pe = spa->spa_mem + pe_handle;
643 
644 	mutex_lock(&spa->spa_lock);
645 
646 	pe->tid = cpu_to_be32(tid);
647 
648 	/*
649 	 * The barrier makes sure the PE is updated
650 	 * before we clear the NPU context cache below, so that the
651 	 * old PE cannot be reloaded erroneously.
652 	 */
653 	mb();
654 
655 	/*
656 	 * hook to platform code
657 	 * On powerpc, the entry needs to be cleared from the context
658 	 * cache of the NPU.
659 	 */
660 	rc = pnv_ocxl_spa_remove_pe_from_cache(link->platform_data, pe_handle);
661 	WARN_ON(rc);
662 
663 	mutex_unlock(&spa->spa_lock);
664 	return rc;
665 }
666 
667 int ocxl_link_remove_pe(void *link_handle, int pasid)
668 {
669 	struct ocxl_link *link = link_handle;
670 	struct spa *spa = link->spa;
671 	struct ocxl_process_element *pe;
672 	struct pe_data *pe_data;
673 	int pe_handle, rc;
674 
675 	if (pasid > SPA_PASID_MAX)
676 		return -EINVAL;
677 
678 	/*
679 	 * About synchronization with our memory fault handler:
680 	 *
681 	 * Before removing the PE, the driver is supposed to have
682 	 * notified the AFU, which should have cleaned up and make
683 	 * sure the PASID is no longer in use, including pending
684 	 * interrupts. However, there's no way to be sure...
685 	 *
686 	 * We clear the PE and remove the context from our radix
687 	 * tree. From that point on, any new interrupt for that
688 	 * context will fail silently, which is ok. As mentioned
689 	 * above, that's not expected, but it could happen if the
690 	 * driver or AFU didn't do the right thing.
691 	 *
692 	 * There could still be a bottom half running, but we don't
693 	 * need to wait/flush, as it is managing a reference count on
694 	 * the mm it reads from the radix tree.
695 	 */
696 	pe_handle = pasid & SPA_PE_MASK;
697 	pe = spa->spa_mem + pe_handle;
698 
699 	mutex_lock(&spa->spa_lock);
700 
701 	if (!(be32_to_cpu(pe->software_state) & SPA_PE_VALID)) {
702 		rc = -EINVAL;
703 		goto unlock;
704 	}
705 
706 	trace_ocxl_context_remove(current->pid, spa->spa_mem, pasid,
707 				be32_to_cpu(pe->pid), be32_to_cpu(pe->tid));
708 
709 	memset(pe, 0, sizeof(struct ocxl_process_element));
710 	/*
711 	 * The barrier makes sure the PE is removed from the SPA
712 	 * before we clear the NPU context cache below, so that the
713 	 * old PE cannot be reloaded erroneously.
714 	 */
715 	mb();
716 
717 	/*
718 	 * hook to platform code
719 	 * On powerpc, the entry needs to be cleared from the context
720 	 * cache of the NPU.
721 	 */
722 	rc = pnv_ocxl_spa_remove_pe_from_cache(link->platform_data, pe_handle);
723 	WARN_ON(rc);
724 
725 	pe_data = radix_tree_delete(&spa->pe_tree, pe_handle);
726 	if (!pe_data) {
727 		WARN(1, "Couldn't find pe data when removing PE\n");
728 	} else {
729 		if (pe_data->mm) {
730 			if (link->arva) {
731 				trace_ocxl_release_mmu_notifier(pasid,
732 								pe_data->mm->context.id);
733 				mmu_notifier_unregister(&pe_data->mmu_notifier,
734 							pe_data->mm);
735 				spin_lock(&link->atsd_lock);
736 				pnv_ocxl_tlb_invalidate(link->arva,
737 							pe_data->mm->context.id,
738 							0ull,
739 							PAGE_SIZE);
740 				spin_unlock(&link->atsd_lock);
741 			}
742 			mm_context_remove_copro(pe_data->mm);
743 			mmdrop(pe_data->mm);
744 		}
745 		kfree_rcu(pe_data, rcu);
746 	}
747 unlock:
748 	mutex_unlock(&spa->spa_lock);
749 	return rc;
750 }
751 EXPORT_SYMBOL_GPL(ocxl_link_remove_pe);
752 
753 int ocxl_link_irq_alloc(void *link_handle, int *hw_irq)
754 {
755 	struct ocxl_link *link = link_handle;
756 	int irq;
757 
758 	if (atomic_dec_if_positive(&link->irq_available) < 0)
759 		return -ENOSPC;
760 
761 	irq = xive_native_alloc_irq();
762 	if (!irq) {
763 		atomic_inc(&link->irq_available);
764 		return -ENXIO;
765 	}
766 
767 	*hw_irq = irq;
768 	return 0;
769 }
770 EXPORT_SYMBOL_GPL(ocxl_link_irq_alloc);
771 
772 void ocxl_link_free_irq(void *link_handle, int hw_irq)
773 {
774 	struct ocxl_link *link = link_handle;
775 
776 	xive_native_free_irq(hw_irq);
777 	atomic_inc(&link->irq_available);
778 }
779 EXPORT_SYMBOL_GPL(ocxl_link_free_irq);
780