1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright 2020-21 IBM Corp.
4  */
5 
6 #define pr_fmt(fmt) "vas: " fmt
7 
8 #include <linux/module.h>
9 #include <linux/kernel.h>
10 #include <linux/export.h>
11 #include <linux/types.h>
12 #include <linux/delay.h>
13 #include <linux/slab.h>
14 #include <linux/interrupt.h>
15 #include <linux/irqdomain.h>
16 #include <asm/machdep.h>
17 #include <asm/hvcall.h>
18 #include <asm/plpar_wrappers.h>
19 #include <asm/vas.h>
20 #include "vas.h"
21 
22 #define VAS_INVALID_WIN_ADDRESS	0xFFFFFFFFFFFFFFFFul
23 #define VAS_DEFAULT_DOMAIN_ID	0xFFFFFFFFFFFFFFFFul
24 /* The hypervisor allows one credit per window right now */
25 #define DEF_WIN_CREDS		1
26 
27 static struct vas_all_caps caps_all;
28 static bool copypaste_feat;
29 
30 static struct vas_caps vascaps[VAS_MAX_FEAT_TYPE];
31 static DEFINE_MUTEX(vas_pseries_mutex);
32 
33 static long hcall_return_busy_check(long rc)
34 {
35 	/* Check if we are stalled for some time */
36 	if (H_IS_LONG_BUSY(rc)) {
37 		msleep(get_longbusy_msecs(rc));
38 		rc = H_BUSY;
39 	} else if (rc == H_BUSY) {
40 		cond_resched();
41 	}
42 
43 	return rc;
44 }
45 
46 /*
47  * Allocate VAS window hcall
48  */
49 static int h_allocate_vas_window(struct pseries_vas_window *win, u64 *domain,
50 				     u8 wintype, u16 credits)
51 {
52 	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
53 	long rc;
54 
55 	do {
56 		rc = plpar_hcall9(H_ALLOCATE_VAS_WINDOW, retbuf, wintype,
57 				  credits, domain[0], domain[1], domain[2],
58 				  domain[3], domain[4], domain[5]);
59 
60 		rc = hcall_return_busy_check(rc);
61 	} while (rc == H_BUSY);
62 
63 	if (rc == H_SUCCESS) {
64 		if (win->win_addr == VAS_INVALID_WIN_ADDRESS) {
65 			pr_err("H_ALLOCATE_VAS_WINDOW: COPY/PASTE is not supported\n");
66 			return -ENOTSUPP;
67 		}
68 		win->vas_win.winid = retbuf[0];
69 		win->win_addr = retbuf[1];
70 		win->complete_irq = retbuf[2];
71 		win->fault_irq = retbuf[3];
72 		return 0;
73 	}
74 
75 	pr_err("H_ALLOCATE_VAS_WINDOW error: %ld, wintype: %u, credits: %u\n",
76 		rc, wintype, credits);
77 
78 	return -EIO;
79 }
80 
81 /*
82  * Deallocate VAS window hcall.
83  */
84 static int h_deallocate_vas_window(u64 winid)
85 {
86 	long rc;
87 
88 	do {
89 		rc = plpar_hcall_norets(H_DEALLOCATE_VAS_WINDOW, winid);
90 
91 		rc = hcall_return_busy_check(rc);
92 	} while (rc == H_BUSY);
93 
94 	if (rc == H_SUCCESS)
95 		return 0;
96 
97 	pr_err("H_DEALLOCATE_VAS_WINDOW error: %ld, winid: %llu\n",
98 		rc, winid);
99 	return -EIO;
100 }
101 
102 /*
103  * Modify VAS window.
104  * After the window is opened with allocate window hcall, configure it
105  * with flags and LPAR PID before using.
106  */
107 static int h_modify_vas_window(struct pseries_vas_window *win)
108 {
109 	long rc;
110 	u32 lpid = mfspr(SPRN_PID);
111 
112 	/*
113 	 * AMR value is not supported in Linux VAS implementation.
114 	 * The hypervisor ignores it if 0 is passed.
115 	 */
116 	do {
117 		rc = plpar_hcall_norets(H_MODIFY_VAS_WINDOW,
118 					win->vas_win.winid, lpid, 0,
119 					VAS_MOD_WIN_FLAGS, 0);
120 
121 		rc = hcall_return_busy_check(rc);
122 	} while (rc == H_BUSY);
123 
124 	if (rc == H_SUCCESS)
125 		return 0;
126 
127 	pr_err("H_MODIFY_VAS_WINDOW error: %ld, winid %u lpid %u\n",
128 			rc, win->vas_win.winid, lpid);
129 	return -EIO;
130 }
131 
132 /*
133  * This hcall is used to determine the capabilities from the hypervisor.
134  * @hcall: H_QUERY_VAS_CAPABILITIES or H_QUERY_NX_CAPABILITIES
135  * @query_type: If 0 is passed, the hypervisor returns the overall
136  *		capabilities which provides all feature(s) that are
137  *		available. Then query the hypervisor to get the
138  *		corresponding capabilities for the specific feature.
139  *		Example: H_QUERY_VAS_CAPABILITIES provides VAS GZIP QoS
140  *			and VAS GZIP Default capabilities.
141  *			H_QUERY_NX_CAPABILITIES provides NX GZIP
142  *			capabilities.
143  * @result: Return buffer to save capabilities.
144  */
145 int h_query_vas_capabilities(const u64 hcall, u8 query_type, u64 result)
146 {
147 	long rc;
148 
149 	rc = plpar_hcall_norets(hcall, query_type, result);
150 
151 	if (rc == H_SUCCESS)
152 		return 0;
153 
154 	pr_err("HCALL(%llx) error %ld, query_type %u, result buffer 0x%llx\n",
155 			hcall, rc, query_type, result);
156 	return -EIO;
157 }
158 EXPORT_SYMBOL_GPL(h_query_vas_capabilities);
159 
160 /*
161  * hcall to get fault CRB from the hypervisor.
162  */
163 static int h_get_nx_fault(u32 winid, u64 buffer)
164 {
165 	long rc;
166 
167 	rc = plpar_hcall_norets(H_GET_NX_FAULT, winid, buffer);
168 
169 	if (rc == H_SUCCESS)
170 		return 0;
171 
172 	pr_err("H_GET_NX_FAULT error: %ld, winid %u, buffer 0x%llx\n",
173 		rc, winid, buffer);
174 	return -EIO;
175 
176 }
177 
178 /*
179  * Handle the fault interrupt.
180  * When the fault interrupt is received for each window, query the
181  * hypervisor to get the fault CRB on the specific fault. Then
182  * process the CRB by updating CSB or send signal if the user space
183  * CSB is invalid.
184  * Note: The hypervisor forwards an interrupt for each fault request.
185  *	So one fault CRB to process for each H_GET_NX_FAULT hcall.
186  */
187 irqreturn_t pseries_vas_fault_thread_fn(int irq, void *data)
188 {
189 	struct pseries_vas_window *txwin = data;
190 	struct coprocessor_request_block crb;
191 	struct vas_user_win_ref *tsk_ref;
192 	int rc;
193 
194 	rc = h_get_nx_fault(txwin->vas_win.winid, (u64)virt_to_phys(&crb));
195 	if (!rc) {
196 		tsk_ref = &txwin->vas_win.task_ref;
197 		vas_dump_crb(&crb);
198 		vas_update_csb(&crb, tsk_ref);
199 	}
200 
201 	return IRQ_HANDLED;
202 }
203 
204 /*
205  * Allocate window and setup IRQ mapping.
206  */
207 static int allocate_setup_window(struct pseries_vas_window *txwin,
208 				 u64 *domain, u8 wintype)
209 {
210 	int rc;
211 
212 	rc = h_allocate_vas_window(txwin, domain, wintype, DEF_WIN_CREDS);
213 	if (rc)
214 		return rc;
215 	/*
216 	 * On PowerVM, the hypervisor setup and forwards the fault
217 	 * interrupt per window. So the IRQ setup and fault handling
218 	 * will be done for each open window separately.
219 	 */
220 	txwin->fault_virq = irq_create_mapping(NULL, txwin->fault_irq);
221 	if (!txwin->fault_virq) {
222 		pr_err("Failed irq mapping %d\n", txwin->fault_irq);
223 		rc = -EINVAL;
224 		goto out_win;
225 	}
226 
227 	txwin->name = kasprintf(GFP_KERNEL, "vas-win-%d",
228 				txwin->vas_win.winid);
229 	if (!txwin->name) {
230 		rc = -ENOMEM;
231 		goto out_irq;
232 	}
233 
234 	rc = request_threaded_irq(txwin->fault_virq, NULL,
235 				  pseries_vas_fault_thread_fn, IRQF_ONESHOT,
236 				  txwin->name, txwin);
237 	if (rc) {
238 		pr_err("VAS-Window[%d]: Request IRQ(%u) failed with %d\n",
239 		       txwin->vas_win.winid, txwin->fault_virq, rc);
240 		goto out_free;
241 	}
242 
243 	txwin->vas_win.wcreds_max = DEF_WIN_CREDS;
244 
245 	return 0;
246 out_free:
247 	kfree(txwin->name);
248 out_irq:
249 	irq_dispose_mapping(txwin->fault_virq);
250 out_win:
251 	h_deallocate_vas_window(txwin->vas_win.winid);
252 	return rc;
253 }
254 
255 static inline void free_irq_setup(struct pseries_vas_window *txwin)
256 {
257 	free_irq(txwin->fault_virq, txwin);
258 	kfree(txwin->name);
259 	irq_dispose_mapping(txwin->fault_virq);
260 }
261 
262 static struct vas_window *vas_allocate_window(int vas_id, u64 flags,
263 					      enum vas_cop_type cop_type)
264 {
265 	long domain[PLPAR_HCALL9_BUFSIZE] = {VAS_DEFAULT_DOMAIN_ID};
266 	struct vas_cop_feat_caps *cop_feat_caps;
267 	struct vas_caps *caps;
268 	struct pseries_vas_window *txwin;
269 	int rc;
270 
271 	txwin = kzalloc(sizeof(*txwin), GFP_KERNEL);
272 	if (!txwin)
273 		return ERR_PTR(-ENOMEM);
274 
275 	/*
276 	 * A VAS window can have many credits which means that many
277 	 * requests can be issued simultaneously. But the hypervisor
278 	 * restricts one credit per window.
279 	 * The hypervisor introduces 2 different types of credits:
280 	 * Default credit type (Uses normal priority FIFO):
281 	 *	A limited number of credits are assigned to partitions
282 	 *	based on processor entitlement. But these credits may be
283 	 *	over-committed on a system depends on whether the CPUs
284 	 *	are in shared or dedicated modes - that is, more requests
285 	 *	may be issued across the system than NX can service at
286 	 *	once which can result in paste command failure (RMA_busy).
287 	 *	Then the process has to resend requests or fall-back to
288 	 *	SW compression.
289 	 * Quality of Service (QoS) credit type (Uses high priority FIFO):
290 	 *	To avoid NX HW contention, the system admins can assign
291 	 *	QoS credits for each LPAR so that this partition is
292 	 *	guaranteed access to NX resources. These credits are
293 	 *	assigned to partitions via the HMC.
294 	 *	Refer PAPR for more information.
295 	 *
296 	 * Allocate window with QoS credits if user requested. Otherwise
297 	 * default credits are used.
298 	 */
299 	if (flags & VAS_TX_WIN_FLAG_QOS_CREDIT)
300 		caps = &vascaps[VAS_GZIP_QOS_FEAT_TYPE];
301 	else
302 		caps = &vascaps[VAS_GZIP_DEF_FEAT_TYPE];
303 
304 	cop_feat_caps = &caps->caps;
305 
306 	if (atomic_inc_return(&cop_feat_caps->used_lpar_creds) >
307 			atomic_read(&cop_feat_caps->target_lpar_creds)) {
308 		pr_err("Credits are not available to allocate window\n");
309 		rc = -EINVAL;
310 		goto out;
311 	}
312 
313 	if (vas_id == -1) {
314 		/*
315 		 * The user space is requesting to allocate a window on
316 		 * a VAS instance where the process is executing.
317 		 * On PowerVM, domain values are passed to the hypervisor
318 		 * to select VAS instance. Useful if the process is
319 		 * affinity to NUMA node.
320 		 * The hypervisor selects VAS instance if
321 		 * VAS_DEFAULT_DOMAIN_ID (-1) is passed for domain values.
322 		 * The h_allocate_vas_window hcall is defined to take a
323 		 * domain values as specified by h_home_node_associativity,
324 		 * So no unpacking needs to be done.
325 		 */
326 		rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, domain,
327 				  VPHN_FLAG_VCPU, smp_processor_id());
328 		if (rc != H_SUCCESS) {
329 			pr_err("H_HOME_NODE_ASSOCIATIVITY error: %d\n", rc);
330 			goto out;
331 		}
332 	}
333 
334 	/*
335 	 * Allocate / Deallocate window hcalls and setup / free IRQs
336 	 * have to be protected with mutex.
337 	 * Open VAS window: Allocate window hcall and setup IRQ
338 	 * Close VAS window: Deallocate window hcall and free IRQ
339 	 *	The hypervisor waits until all NX requests are
340 	 *	completed before closing the window. So expects OS
341 	 *	to handle NX faults, means IRQ can be freed only
342 	 *	after the deallocate window hcall is returned.
343 	 * So once the window is closed with deallocate hcall before
344 	 * the IRQ is freed, it can be assigned to new allocate
345 	 * hcall with the same fault IRQ by the hypervisor. It can
346 	 * result in setup IRQ fail for the new window since the
347 	 * same fault IRQ is not freed by the OS before.
348 	 */
349 	mutex_lock(&vas_pseries_mutex);
350 	rc = allocate_setup_window(txwin, (u64 *)&domain[0],
351 				   cop_feat_caps->win_type);
352 	mutex_unlock(&vas_pseries_mutex);
353 	if (rc)
354 		goto out;
355 
356 	/*
357 	 * Modify window and it is ready to use.
358 	 */
359 	rc = h_modify_vas_window(txwin);
360 	if (!rc)
361 		rc = get_vas_user_win_ref(&txwin->vas_win.task_ref);
362 	if (rc)
363 		goto out_free;
364 
365 	vas_user_win_add_mm_context(&txwin->vas_win.task_ref);
366 	txwin->win_type = cop_feat_caps->win_type;
367 	mutex_lock(&vas_pseries_mutex);
368 	list_add(&txwin->win_list, &caps->list);
369 	mutex_unlock(&vas_pseries_mutex);
370 
371 	return &txwin->vas_win;
372 
373 out_free:
374 	/*
375 	 * Window is not operational. Free IRQ before closing
376 	 * window so that do not have to hold mutex.
377 	 */
378 	free_irq_setup(txwin);
379 	h_deallocate_vas_window(txwin->vas_win.winid);
380 out:
381 	atomic_dec(&cop_feat_caps->used_lpar_creds);
382 	kfree(txwin);
383 	return ERR_PTR(rc);
384 }
385 
386 static u64 vas_paste_address(struct vas_window *vwin)
387 {
388 	struct pseries_vas_window *win;
389 
390 	win = container_of(vwin, struct pseries_vas_window, vas_win);
391 	return win->win_addr;
392 }
393 
394 static int deallocate_free_window(struct pseries_vas_window *win)
395 {
396 	int rc = 0;
397 
398 	/*
399 	 * The hypervisor waits for all requests including faults
400 	 * are processed before closing the window - Means all
401 	 * credits have to be returned. In the case of fault
402 	 * request, a credit is returned after OS issues
403 	 * H_GET_NX_FAULT hcall.
404 	 * So free IRQ after executing H_DEALLOCATE_VAS_WINDOW
405 	 * hcall.
406 	 */
407 	rc = h_deallocate_vas_window(win->vas_win.winid);
408 	if (!rc)
409 		free_irq_setup(win);
410 
411 	return rc;
412 }
413 
414 static int vas_deallocate_window(struct vas_window *vwin)
415 {
416 	struct pseries_vas_window *win;
417 	struct vas_cop_feat_caps *caps;
418 	int rc = 0;
419 
420 	if (!vwin)
421 		return -EINVAL;
422 
423 	win = container_of(vwin, struct pseries_vas_window, vas_win);
424 
425 	/* Should not happen */
426 	if (win->win_type >= VAS_MAX_FEAT_TYPE) {
427 		pr_err("Window (%u): Invalid window type %u\n",
428 				vwin->winid, win->win_type);
429 		return -EINVAL;
430 	}
431 
432 	caps = &vascaps[win->win_type].caps;
433 	mutex_lock(&vas_pseries_mutex);
434 	rc = deallocate_free_window(win);
435 	if (rc) {
436 		mutex_unlock(&vas_pseries_mutex);
437 		return rc;
438 	}
439 
440 	list_del(&win->win_list);
441 	atomic_dec(&caps->used_lpar_creds);
442 	mutex_unlock(&vas_pseries_mutex);
443 
444 	put_vas_user_win_ref(&vwin->task_ref);
445 	mm_context_remove_vas_window(vwin->task_ref.mm);
446 
447 	kfree(win);
448 	return 0;
449 }
450 
451 static const struct vas_user_win_ops vops_pseries = {
452 	.open_win	= vas_allocate_window,	/* Open and configure window */
453 	.paste_addr	= vas_paste_address,	/* To do copy/paste */
454 	.close_win	= vas_deallocate_window, /* Close window */
455 };
456 
457 /*
458  * Supporting only nx-gzip coprocessor type now, but this API code
459  * extended to other coprocessor types later.
460  */
461 int vas_register_api_pseries(struct module *mod, enum vas_cop_type cop_type,
462 			     const char *name)
463 {
464 	int rc;
465 
466 	if (!copypaste_feat)
467 		return -ENOTSUPP;
468 
469 	rc = vas_register_coproc_api(mod, cop_type, name, &vops_pseries);
470 
471 	return rc;
472 }
473 EXPORT_SYMBOL_GPL(vas_register_api_pseries);
474 
475 void vas_unregister_api_pseries(void)
476 {
477 	vas_unregister_coproc_api();
478 }
479 EXPORT_SYMBOL_GPL(vas_unregister_api_pseries);
480 
481 /*
482  * Get the specific capabilities based on the feature type.
483  * Right now supports GZIP default and GZIP QoS capabilities.
484  */
485 static int get_vas_capabilities(u8 feat, enum vas_cop_feat_type type,
486 				struct hv_vas_cop_feat_caps *hv_caps)
487 {
488 	struct vas_cop_feat_caps *caps;
489 	struct vas_caps *vcaps;
490 	int rc = 0;
491 
492 	vcaps = &vascaps[type];
493 	memset(vcaps, 0, sizeof(*vcaps));
494 	INIT_LIST_HEAD(&vcaps->list);
495 
496 	caps = &vcaps->caps;
497 
498 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, feat,
499 					  (u64)virt_to_phys(hv_caps));
500 	if (rc)
501 		return rc;
502 
503 	caps->user_mode = hv_caps->user_mode;
504 	if (!(caps->user_mode & VAS_COPY_PASTE_USER_MODE)) {
505 		pr_err("User space COPY/PASTE is not supported\n");
506 		return -ENOTSUPP;
507 	}
508 
509 	caps->descriptor = be64_to_cpu(hv_caps->descriptor);
510 	caps->win_type = hv_caps->win_type;
511 	if (caps->win_type >= VAS_MAX_FEAT_TYPE) {
512 		pr_err("Unsupported window type %u\n", caps->win_type);
513 		return -EINVAL;
514 	}
515 	caps->max_lpar_creds = be16_to_cpu(hv_caps->max_lpar_creds);
516 	caps->max_win_creds = be16_to_cpu(hv_caps->max_win_creds);
517 	atomic_set(&caps->target_lpar_creds,
518 		   be16_to_cpu(hv_caps->target_lpar_creds));
519 	if (feat == VAS_GZIP_DEF_FEAT) {
520 		caps->def_lpar_creds = be16_to_cpu(hv_caps->def_lpar_creds);
521 
522 		if (caps->max_win_creds < DEF_WIN_CREDS) {
523 			pr_err("Window creds(%u) > max allowed window creds(%u)\n",
524 			       DEF_WIN_CREDS, caps->max_win_creds);
525 			return -EINVAL;
526 		}
527 	}
528 
529 	copypaste_feat = true;
530 
531 	return 0;
532 }
533 
534 static int __init pseries_vas_init(void)
535 {
536 	struct hv_vas_cop_feat_caps *hv_cop_caps;
537 	struct hv_vas_all_caps *hv_caps;
538 	int rc;
539 
540 	/*
541 	 * Linux supports user space COPY/PASTE only with Radix
542 	 */
543 	if (!radix_enabled()) {
544 		pr_err("API is supported only with radix page tables\n");
545 		return -ENOTSUPP;
546 	}
547 
548 	hv_caps = kmalloc(sizeof(*hv_caps), GFP_KERNEL);
549 	if (!hv_caps)
550 		return -ENOMEM;
551 	/*
552 	 * Get VAS overall capabilities by passing 0 to feature type.
553 	 */
554 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, 0,
555 					  (u64)virt_to_phys(hv_caps));
556 	if (rc)
557 		goto out;
558 
559 	caps_all.descriptor = be64_to_cpu(hv_caps->descriptor);
560 	caps_all.feat_type = be64_to_cpu(hv_caps->feat_type);
561 
562 	hv_cop_caps = kmalloc(sizeof(*hv_cop_caps), GFP_KERNEL);
563 	if (!hv_cop_caps) {
564 		rc = -ENOMEM;
565 		goto out;
566 	}
567 	/*
568 	 * QOS capabilities available
569 	 */
570 	if (caps_all.feat_type & VAS_GZIP_QOS_FEAT_BIT) {
571 		rc = get_vas_capabilities(VAS_GZIP_QOS_FEAT,
572 					  VAS_GZIP_QOS_FEAT_TYPE, hv_cop_caps);
573 
574 		if (rc)
575 			goto out_cop;
576 	}
577 	/*
578 	 * Default capabilities available
579 	 */
580 	if (caps_all.feat_type & VAS_GZIP_DEF_FEAT_BIT) {
581 		rc = get_vas_capabilities(VAS_GZIP_DEF_FEAT,
582 					  VAS_GZIP_DEF_FEAT_TYPE, hv_cop_caps);
583 		if (rc)
584 			goto out_cop;
585 	}
586 
587 	pr_info("GZIP feature is available\n");
588 
589 out_cop:
590 	kfree(hv_cop_caps);
591 out:
592 	kfree(hv_caps);
593 	return rc;
594 }
595 machine_device_initcall(pseries, pseries_vas_init);
596