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/firmware.h>
20 #include <asm/vas.h>
21 #include "vas.h"
22 
23 #define VAS_INVALID_WIN_ADDRESS	0xFFFFFFFFFFFFFFFFul
24 #define VAS_DEFAULT_DOMAIN_ID	0xFFFFFFFFFFFFFFFFul
25 /* The hypervisor allows one credit per window right now */
26 #define DEF_WIN_CREDS		1
27 
28 static struct vas_all_caps caps_all;
29 static bool copypaste_feat;
30 static struct hv_vas_cop_feat_caps hv_cop_caps;
31 
32 static struct vas_caps vascaps[VAS_MAX_FEAT_TYPE];
33 static DEFINE_MUTEX(vas_pseries_mutex);
34 static bool migration_in_progress;
35 
36 static long hcall_return_busy_check(long rc)
37 {
38 	/* Check if we are stalled for some time */
39 	if (H_IS_LONG_BUSY(rc)) {
40 		msleep(get_longbusy_msecs(rc));
41 		rc = H_BUSY;
42 	} else if (rc == H_BUSY) {
43 		cond_resched();
44 	}
45 
46 	return rc;
47 }
48 
49 /*
50  * Allocate VAS window hcall
51  */
52 static int h_allocate_vas_window(struct pseries_vas_window *win, u64 *domain,
53 				     u8 wintype, u16 credits)
54 {
55 	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
56 	long rc;
57 
58 	do {
59 		rc = plpar_hcall9(H_ALLOCATE_VAS_WINDOW, retbuf, wintype,
60 				  credits, domain[0], domain[1], domain[2],
61 				  domain[3], domain[4], domain[5]);
62 
63 		rc = hcall_return_busy_check(rc);
64 	} while (rc == H_BUSY);
65 
66 	if (rc == H_SUCCESS) {
67 		if (win->win_addr == VAS_INVALID_WIN_ADDRESS) {
68 			pr_err("H_ALLOCATE_VAS_WINDOW: COPY/PASTE is not supported\n");
69 			return -ENOTSUPP;
70 		}
71 		win->vas_win.winid = retbuf[0];
72 		win->win_addr = retbuf[1];
73 		win->complete_irq = retbuf[2];
74 		win->fault_irq = retbuf[3];
75 		return 0;
76 	}
77 
78 	pr_err("H_ALLOCATE_VAS_WINDOW error: %ld, wintype: %u, credits: %u\n",
79 		rc, wintype, credits);
80 
81 	return -EIO;
82 }
83 
84 /*
85  * Deallocate VAS window hcall.
86  */
87 static int h_deallocate_vas_window(u64 winid)
88 {
89 	long rc;
90 
91 	do {
92 		rc = plpar_hcall_norets(H_DEALLOCATE_VAS_WINDOW, winid);
93 
94 		rc = hcall_return_busy_check(rc);
95 	} while (rc == H_BUSY);
96 
97 	if (rc == H_SUCCESS)
98 		return 0;
99 
100 	pr_err("H_DEALLOCATE_VAS_WINDOW error: %ld, winid: %llu\n",
101 		rc, winid);
102 	return -EIO;
103 }
104 
105 /*
106  * Modify VAS window.
107  * After the window is opened with allocate window hcall, configure it
108  * with flags and LPAR PID before using.
109  */
110 static int h_modify_vas_window(struct pseries_vas_window *win)
111 {
112 	long rc;
113 
114 	/*
115 	 * AMR value is not supported in Linux VAS implementation.
116 	 * The hypervisor ignores it if 0 is passed.
117 	 */
118 	do {
119 		rc = plpar_hcall_norets(H_MODIFY_VAS_WINDOW,
120 					win->vas_win.winid, win->pid, 0,
121 					VAS_MOD_WIN_FLAGS, 0);
122 
123 		rc = hcall_return_busy_check(rc);
124 	} while (rc == H_BUSY);
125 
126 	if (rc == H_SUCCESS)
127 		return 0;
128 
129 	pr_err("H_MODIFY_VAS_WINDOW error: %ld, winid %u pid %u\n",
130 			rc, win->vas_win.winid, win->pid);
131 	return -EIO;
132 }
133 
134 /*
135  * This hcall is used to determine the capabilities from the hypervisor.
136  * @hcall: H_QUERY_VAS_CAPABILITIES or H_QUERY_NX_CAPABILITIES
137  * @query_type: If 0 is passed, the hypervisor returns the overall
138  *		capabilities which provides all feature(s) that are
139  *		available. Then query the hypervisor to get the
140  *		corresponding capabilities for the specific feature.
141  *		Example: H_QUERY_VAS_CAPABILITIES provides VAS GZIP QoS
142  *			and VAS GZIP Default capabilities.
143  *			H_QUERY_NX_CAPABILITIES provides NX GZIP
144  *			capabilities.
145  * @result: Return buffer to save capabilities.
146  */
147 int h_query_vas_capabilities(const u64 hcall, u8 query_type, u64 result)
148 {
149 	long rc;
150 
151 	rc = plpar_hcall_norets(hcall, query_type, result);
152 
153 	if (rc == H_SUCCESS)
154 		return 0;
155 
156 	/* H_FUNCTION means HV does not support VAS so don't print an error */
157 	if (rc != H_FUNCTION) {
158 		pr_err("%s error %ld, query_type %u, result buffer 0x%llx\n",
159 			(hcall == H_QUERY_VAS_CAPABILITIES) ?
160 				"H_QUERY_VAS_CAPABILITIES" :
161 				"H_QUERY_NX_CAPABILITIES",
162 			rc, query_type, result);
163 	}
164 
165 	return -EIO;
166 }
167 EXPORT_SYMBOL_GPL(h_query_vas_capabilities);
168 
169 /*
170  * hcall to get fault CRB from the hypervisor.
171  */
172 static int h_get_nx_fault(u32 winid, u64 buffer)
173 {
174 	long rc;
175 
176 	rc = plpar_hcall_norets(H_GET_NX_FAULT, winid, buffer);
177 
178 	if (rc == H_SUCCESS)
179 		return 0;
180 
181 	pr_err("H_GET_NX_FAULT error: %ld, winid %u, buffer 0x%llx\n",
182 		rc, winid, buffer);
183 	return -EIO;
184 
185 }
186 
187 /*
188  * Handle the fault interrupt.
189  * When the fault interrupt is received for each window, query the
190  * hypervisor to get the fault CRB on the specific fault. Then
191  * process the CRB by updating CSB or send signal if the user space
192  * CSB is invalid.
193  * Note: The hypervisor forwards an interrupt for each fault request.
194  *	So one fault CRB to process for each H_GET_NX_FAULT hcall.
195  */
196 static irqreturn_t pseries_vas_fault_thread_fn(int irq, void *data)
197 {
198 	struct pseries_vas_window *txwin = data;
199 	struct coprocessor_request_block crb;
200 	struct vas_user_win_ref *tsk_ref;
201 	int rc;
202 
203 	rc = h_get_nx_fault(txwin->vas_win.winid, (u64)virt_to_phys(&crb));
204 	if (!rc) {
205 		tsk_ref = &txwin->vas_win.task_ref;
206 		vas_dump_crb(&crb);
207 		vas_update_csb(&crb, tsk_ref);
208 	}
209 
210 	return IRQ_HANDLED;
211 }
212 
213 /*
214  * Allocate window and setup IRQ mapping.
215  */
216 static int allocate_setup_window(struct pseries_vas_window *txwin,
217 				 u64 *domain, u8 wintype)
218 {
219 	int rc;
220 
221 	rc = h_allocate_vas_window(txwin, domain, wintype, DEF_WIN_CREDS);
222 	if (rc)
223 		return rc;
224 	/*
225 	 * On PowerVM, the hypervisor setup and forwards the fault
226 	 * interrupt per window. So the IRQ setup and fault handling
227 	 * will be done for each open window separately.
228 	 */
229 	txwin->fault_virq = irq_create_mapping(NULL, txwin->fault_irq);
230 	if (!txwin->fault_virq) {
231 		pr_err("Failed irq mapping %d\n", txwin->fault_irq);
232 		rc = -EINVAL;
233 		goto out_win;
234 	}
235 
236 	txwin->name = kasprintf(GFP_KERNEL, "vas-win-%d",
237 				txwin->vas_win.winid);
238 	if (!txwin->name) {
239 		rc = -ENOMEM;
240 		goto out_irq;
241 	}
242 
243 	rc = request_threaded_irq(txwin->fault_virq, NULL,
244 				  pseries_vas_fault_thread_fn, IRQF_ONESHOT,
245 				  txwin->name, txwin);
246 	if (rc) {
247 		pr_err("VAS-Window[%d]: Request IRQ(%u) failed with %d\n",
248 		       txwin->vas_win.winid, txwin->fault_virq, rc);
249 		goto out_free;
250 	}
251 
252 	txwin->vas_win.wcreds_max = DEF_WIN_CREDS;
253 
254 	return 0;
255 out_free:
256 	kfree(txwin->name);
257 out_irq:
258 	irq_dispose_mapping(txwin->fault_virq);
259 out_win:
260 	h_deallocate_vas_window(txwin->vas_win.winid);
261 	return rc;
262 }
263 
264 static inline void free_irq_setup(struct pseries_vas_window *txwin)
265 {
266 	free_irq(txwin->fault_virq, txwin);
267 	kfree(txwin->name);
268 	irq_dispose_mapping(txwin->fault_virq);
269 }
270 
271 static struct vas_window *vas_allocate_window(int vas_id, u64 flags,
272 					      enum vas_cop_type cop_type)
273 {
274 	long domain[PLPAR_HCALL9_BUFSIZE] = {VAS_DEFAULT_DOMAIN_ID};
275 	struct vas_cop_feat_caps *cop_feat_caps;
276 	struct vas_caps *caps;
277 	struct pseries_vas_window *txwin;
278 	int rc;
279 
280 	txwin = kzalloc(sizeof(*txwin), GFP_KERNEL);
281 	if (!txwin)
282 		return ERR_PTR(-ENOMEM);
283 
284 	/*
285 	 * A VAS window can have many credits which means that many
286 	 * requests can be issued simultaneously. But the hypervisor
287 	 * restricts one credit per window.
288 	 * The hypervisor introduces 2 different types of credits:
289 	 * Default credit type (Uses normal priority FIFO):
290 	 *	A limited number of credits are assigned to partitions
291 	 *	based on processor entitlement. But these credits may be
292 	 *	over-committed on a system depends on whether the CPUs
293 	 *	are in shared or dedicated modes - that is, more requests
294 	 *	may be issued across the system than NX can service at
295 	 *	once which can result in paste command failure (RMA_busy).
296 	 *	Then the process has to resend requests or fall-back to
297 	 *	SW compression.
298 	 * Quality of Service (QoS) credit type (Uses high priority FIFO):
299 	 *	To avoid NX HW contention, the system admins can assign
300 	 *	QoS credits for each LPAR so that this partition is
301 	 *	guaranteed access to NX resources. These credits are
302 	 *	assigned to partitions via the HMC.
303 	 *	Refer PAPR for more information.
304 	 *
305 	 * Allocate window with QoS credits if user requested. Otherwise
306 	 * default credits are used.
307 	 */
308 	if (flags & VAS_TX_WIN_FLAG_QOS_CREDIT)
309 		caps = &vascaps[VAS_GZIP_QOS_FEAT_TYPE];
310 	else
311 		caps = &vascaps[VAS_GZIP_DEF_FEAT_TYPE];
312 
313 	cop_feat_caps = &caps->caps;
314 
315 	if (atomic_inc_return(&cop_feat_caps->nr_used_credits) >
316 			atomic_read(&cop_feat_caps->nr_total_credits)) {
317 		pr_err("Credits are not available to allocate window\n");
318 		rc = -EINVAL;
319 		goto out;
320 	}
321 
322 	if (vas_id == -1) {
323 		/*
324 		 * The user space is requesting to allocate a window on
325 		 * a VAS instance where the process is executing.
326 		 * On PowerVM, domain values are passed to the hypervisor
327 		 * to select VAS instance. Useful if the process is
328 		 * affinity to NUMA node.
329 		 * The hypervisor selects VAS instance if
330 		 * VAS_DEFAULT_DOMAIN_ID (-1) is passed for domain values.
331 		 * The h_allocate_vas_window hcall is defined to take a
332 		 * domain values as specified by h_home_node_associativity,
333 		 * So no unpacking needs to be done.
334 		 */
335 		rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, domain,
336 				  VPHN_FLAG_VCPU, smp_processor_id());
337 		if (rc != H_SUCCESS) {
338 			pr_err("H_HOME_NODE_ASSOCIATIVITY error: %d\n", rc);
339 			goto out;
340 		}
341 	}
342 
343 	txwin->pid = mfspr(SPRN_PID);
344 
345 	/*
346 	 * Allocate / Deallocate window hcalls and setup / free IRQs
347 	 * have to be protected with mutex.
348 	 * Open VAS window: Allocate window hcall and setup IRQ
349 	 * Close VAS window: Deallocate window hcall and free IRQ
350 	 *	The hypervisor waits until all NX requests are
351 	 *	completed before closing the window. So expects OS
352 	 *	to handle NX faults, means IRQ can be freed only
353 	 *	after the deallocate window hcall is returned.
354 	 * So once the window is closed with deallocate hcall before
355 	 * the IRQ is freed, it can be assigned to new allocate
356 	 * hcall with the same fault IRQ by the hypervisor. It can
357 	 * result in setup IRQ fail for the new window since the
358 	 * same fault IRQ is not freed by the OS before.
359 	 */
360 	mutex_lock(&vas_pseries_mutex);
361 	if (migration_in_progress)
362 		rc = -EBUSY;
363 	else
364 		rc = allocate_setup_window(txwin, (u64 *)&domain[0],
365 				   cop_feat_caps->win_type);
366 	mutex_unlock(&vas_pseries_mutex);
367 	if (rc)
368 		goto out;
369 
370 	/*
371 	 * Modify window and it is ready to use.
372 	 */
373 	rc = h_modify_vas_window(txwin);
374 	if (!rc)
375 		rc = get_vas_user_win_ref(&txwin->vas_win.task_ref);
376 	if (rc)
377 		goto out_free;
378 
379 	txwin->win_type = cop_feat_caps->win_type;
380 	mutex_lock(&vas_pseries_mutex);
381 	/*
382 	 * Possible to lose the acquired credit with DLPAR core
383 	 * removal after the window is opened. So if there are any
384 	 * closed windows (means with lost credits), do not give new
385 	 * window to user space. New windows will be opened only
386 	 * after the existing windows are reopened when credits are
387 	 * available.
388 	 */
389 	if (!caps->nr_close_wins) {
390 		list_add(&txwin->win_list, &caps->list);
391 		caps->nr_open_windows++;
392 		mutex_unlock(&vas_pseries_mutex);
393 		vas_user_win_add_mm_context(&txwin->vas_win.task_ref);
394 		return &txwin->vas_win;
395 	}
396 	mutex_unlock(&vas_pseries_mutex);
397 
398 	put_vas_user_win_ref(&txwin->vas_win.task_ref);
399 	rc = -EBUSY;
400 	pr_err("No credit is available to allocate window\n");
401 
402 out_free:
403 	/*
404 	 * Window is not operational. Free IRQ before closing
405 	 * window so that do not have to hold mutex.
406 	 */
407 	free_irq_setup(txwin);
408 	h_deallocate_vas_window(txwin->vas_win.winid);
409 out:
410 	atomic_dec(&cop_feat_caps->nr_used_credits);
411 	kfree(txwin);
412 	return ERR_PTR(rc);
413 }
414 
415 static u64 vas_paste_address(struct vas_window *vwin)
416 {
417 	struct pseries_vas_window *win;
418 
419 	win = container_of(vwin, struct pseries_vas_window, vas_win);
420 	return win->win_addr;
421 }
422 
423 static int deallocate_free_window(struct pseries_vas_window *win)
424 {
425 	int rc = 0;
426 
427 	/*
428 	 * The hypervisor waits for all requests including faults
429 	 * are processed before closing the window - Means all
430 	 * credits have to be returned. In the case of fault
431 	 * request, a credit is returned after OS issues
432 	 * H_GET_NX_FAULT hcall.
433 	 * So free IRQ after executing H_DEALLOCATE_VAS_WINDOW
434 	 * hcall.
435 	 */
436 	rc = h_deallocate_vas_window(win->vas_win.winid);
437 	if (!rc)
438 		free_irq_setup(win);
439 
440 	return rc;
441 }
442 
443 static int vas_deallocate_window(struct vas_window *vwin)
444 {
445 	struct pseries_vas_window *win;
446 	struct vas_cop_feat_caps *caps;
447 	int rc = 0;
448 
449 	if (!vwin)
450 		return -EINVAL;
451 
452 	win = container_of(vwin, struct pseries_vas_window, vas_win);
453 
454 	/* Should not happen */
455 	if (win->win_type >= VAS_MAX_FEAT_TYPE) {
456 		pr_err("Window (%u): Invalid window type %u\n",
457 				vwin->winid, win->win_type);
458 		return -EINVAL;
459 	}
460 
461 	caps = &vascaps[win->win_type].caps;
462 	mutex_lock(&vas_pseries_mutex);
463 	/*
464 	 * VAS window is already closed in the hypervisor when
465 	 * lost the credit or with migration. So just remove the entry
466 	 * from the list, remove task references and free vas_window
467 	 * struct.
468 	 */
469 	if (!(win->vas_win.status & VAS_WIN_NO_CRED_CLOSE) &&
470 		!(win->vas_win.status & VAS_WIN_MIGRATE_CLOSE)) {
471 		rc = deallocate_free_window(win);
472 		if (rc) {
473 			mutex_unlock(&vas_pseries_mutex);
474 			return rc;
475 		}
476 	} else
477 		vascaps[win->win_type].nr_close_wins--;
478 
479 	list_del(&win->win_list);
480 	atomic_dec(&caps->nr_used_credits);
481 	vascaps[win->win_type].nr_open_windows--;
482 	mutex_unlock(&vas_pseries_mutex);
483 
484 	put_vas_user_win_ref(&vwin->task_ref);
485 	mm_context_remove_vas_window(vwin->task_ref.mm);
486 
487 	kfree(win);
488 	return 0;
489 }
490 
491 static const struct vas_user_win_ops vops_pseries = {
492 	.open_win	= vas_allocate_window,	/* Open and configure window */
493 	.paste_addr	= vas_paste_address,	/* To do copy/paste */
494 	.close_win	= vas_deallocate_window, /* Close window */
495 };
496 
497 /*
498  * Supporting only nx-gzip coprocessor type now, but this API code
499  * extended to other coprocessor types later.
500  */
501 int vas_register_api_pseries(struct module *mod, enum vas_cop_type cop_type,
502 			     const char *name)
503 {
504 	int rc;
505 
506 	if (!copypaste_feat)
507 		return -ENOTSUPP;
508 
509 	rc = vas_register_coproc_api(mod, cop_type, name, &vops_pseries);
510 
511 	return rc;
512 }
513 EXPORT_SYMBOL_GPL(vas_register_api_pseries);
514 
515 void vas_unregister_api_pseries(void)
516 {
517 	vas_unregister_coproc_api();
518 }
519 EXPORT_SYMBOL_GPL(vas_unregister_api_pseries);
520 
521 /*
522  * Get the specific capabilities based on the feature type.
523  * Right now supports GZIP default and GZIP QoS capabilities.
524  */
525 static int __init get_vas_capabilities(u8 feat, enum vas_cop_feat_type type,
526 				struct hv_vas_cop_feat_caps *hv_caps)
527 {
528 	struct vas_cop_feat_caps *caps;
529 	struct vas_caps *vcaps;
530 	int rc = 0;
531 
532 	vcaps = &vascaps[type];
533 	memset(vcaps, 0, sizeof(*vcaps));
534 	INIT_LIST_HEAD(&vcaps->list);
535 
536 	vcaps->feat = feat;
537 	caps = &vcaps->caps;
538 
539 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, feat,
540 					  (u64)virt_to_phys(hv_caps));
541 	if (rc)
542 		return rc;
543 
544 	caps->user_mode = hv_caps->user_mode;
545 	if (!(caps->user_mode & VAS_COPY_PASTE_USER_MODE)) {
546 		pr_err("User space COPY/PASTE is not supported\n");
547 		return -ENOTSUPP;
548 	}
549 
550 	caps->descriptor = be64_to_cpu(hv_caps->descriptor);
551 	caps->win_type = hv_caps->win_type;
552 	if (caps->win_type >= VAS_MAX_FEAT_TYPE) {
553 		pr_err("Unsupported window type %u\n", caps->win_type);
554 		return -EINVAL;
555 	}
556 	caps->max_lpar_creds = be16_to_cpu(hv_caps->max_lpar_creds);
557 	caps->max_win_creds = be16_to_cpu(hv_caps->max_win_creds);
558 	atomic_set(&caps->nr_total_credits,
559 		   be16_to_cpu(hv_caps->target_lpar_creds));
560 	if (feat == VAS_GZIP_DEF_FEAT) {
561 		caps->def_lpar_creds = be16_to_cpu(hv_caps->def_lpar_creds);
562 
563 		if (caps->max_win_creds < DEF_WIN_CREDS) {
564 			pr_err("Window creds(%u) > max allowed window creds(%u)\n",
565 			       DEF_WIN_CREDS, caps->max_win_creds);
566 			return -EINVAL;
567 		}
568 	}
569 
570 	rc = sysfs_add_vas_caps(caps);
571 	if (rc)
572 		return rc;
573 
574 	copypaste_feat = true;
575 
576 	return 0;
577 }
578 
579 /*
580  * VAS windows can be closed due to lost credits when the core is
581  * removed. So reopen them if credits are available due to DLPAR
582  * core add and set the window active status. When NX sees the page
583  * fault on the unmapped paste address, the kernel handles the fault
584  * by setting the remapping to new paste address if the window is
585  * active.
586  */
587 static int reconfig_open_windows(struct vas_caps *vcaps, int creds,
588 				 bool migrate)
589 {
590 	long domain[PLPAR_HCALL9_BUFSIZE] = {VAS_DEFAULT_DOMAIN_ID};
591 	struct vas_cop_feat_caps *caps = &vcaps->caps;
592 	struct pseries_vas_window *win = NULL, *tmp;
593 	int rc, mv_ents = 0;
594 	int flag;
595 
596 	/*
597 	 * Nothing to do if there are no closed windows.
598 	 */
599 	if (!vcaps->nr_close_wins)
600 		return 0;
601 
602 	/*
603 	 * For the core removal, the hypervisor reduces the credits
604 	 * assigned to the LPAR and the kernel closes VAS windows
605 	 * in the hypervisor depends on reduced credits. The kernel
606 	 * uses LIFO (the last windows that are opened will be closed
607 	 * first) and expects to open in the same order when credits
608 	 * are available.
609 	 * For example, 40 windows are closed when the LPAR lost 2 cores
610 	 * (dedicated). If 1 core is added, this LPAR can have 20 more
611 	 * credits. It means the kernel can reopen 20 windows. So move
612 	 * 20 entries in the VAS windows lost and reopen next 20 windows.
613 	 * For partition migration, reopen all windows that are closed
614 	 * during resume.
615 	 */
616 	if ((vcaps->nr_close_wins > creds) && !migrate)
617 		mv_ents = vcaps->nr_close_wins - creds;
618 
619 	list_for_each_entry_safe(win, tmp, &vcaps->list, win_list) {
620 		if (!mv_ents)
621 			break;
622 
623 		mv_ents--;
624 	}
625 
626 	/*
627 	 * Open windows if they are closed only with migration or
628 	 * DLPAR (lost credit) before.
629 	 */
630 	if (migrate)
631 		flag = VAS_WIN_MIGRATE_CLOSE;
632 	else
633 		flag = VAS_WIN_NO_CRED_CLOSE;
634 
635 	list_for_each_entry_safe_from(win, tmp, &vcaps->list, win_list) {
636 		/*
637 		 * This window is closed with DLPAR and migration events.
638 		 * So reopen the window with the last event.
639 		 * The user space is not suspended with the current
640 		 * migration notifier. So the user space can issue DLPAR
641 		 * CPU hotplug while migration in progress. In this case
642 		 * this window will be opened with the last event.
643 		 */
644 		if ((win->vas_win.status & VAS_WIN_NO_CRED_CLOSE) &&
645 			(win->vas_win.status & VAS_WIN_MIGRATE_CLOSE)) {
646 			win->vas_win.status &= ~flag;
647 			continue;
648 		}
649 
650 		/*
651 		 * Nothing to do on this window if it is not closed
652 		 * with this flag
653 		 */
654 		if (!(win->vas_win.status & flag))
655 			continue;
656 
657 		rc = allocate_setup_window(win, (u64 *)&domain[0],
658 					   caps->win_type);
659 		if (rc)
660 			return rc;
661 
662 		rc = h_modify_vas_window(win);
663 		if (rc)
664 			goto out;
665 
666 		mutex_lock(&win->vas_win.task_ref.mmap_mutex);
667 		/*
668 		 * Set window status to active
669 		 */
670 		win->vas_win.status &= ~flag;
671 		mutex_unlock(&win->vas_win.task_ref.mmap_mutex);
672 		win->win_type = caps->win_type;
673 		if (!--vcaps->nr_close_wins)
674 			break;
675 	}
676 
677 	return 0;
678 out:
679 	/*
680 	 * Window modify HCALL failed. So close the window to the
681 	 * hypervisor and return.
682 	 */
683 	free_irq_setup(win);
684 	h_deallocate_vas_window(win->vas_win.winid);
685 	return rc;
686 }
687 
688 /*
689  * The hypervisor reduces the available credits if the LPAR lost core. It
690  * means the excessive windows should not be active and the user space
691  * should not be using these windows to send compression requests to NX.
692  * So the kernel closes the excessive windows and unmap the paste address
693  * such that the user space receives paste instruction failure. Then up to
694  * the user space to fall back to SW compression and manage with the
695  * existing windows.
696  */
697 static int reconfig_close_windows(struct vas_caps *vcap, int excess_creds,
698 									bool migrate)
699 {
700 	struct pseries_vas_window *win, *tmp;
701 	struct vas_user_win_ref *task_ref;
702 	struct vm_area_struct *vma;
703 	int rc = 0, flag;
704 
705 	if (migrate)
706 		flag = VAS_WIN_MIGRATE_CLOSE;
707 	else
708 		flag = VAS_WIN_NO_CRED_CLOSE;
709 
710 	list_for_each_entry_safe(win, tmp, &vcap->list, win_list) {
711 		/*
712 		 * This window is already closed due to lost credit
713 		 * or for migration before. Go for next window.
714 		 * For migration, nothing to do since this window
715 		 * closed for DLPAR and will be reopened even on
716 		 * the destination system with other DLPAR operation.
717 		 */
718 		if ((win->vas_win.status & VAS_WIN_MIGRATE_CLOSE) ||
719 			(win->vas_win.status & VAS_WIN_NO_CRED_CLOSE)) {
720 			win->vas_win.status |= flag;
721 			continue;
722 		}
723 
724 		task_ref = &win->vas_win.task_ref;
725 		mutex_lock(&task_ref->mmap_mutex);
726 		vma = task_ref->vma;
727 		/*
728 		 * Number of available credits are reduced, So select
729 		 * and close windows.
730 		 */
731 		win->vas_win.status |= flag;
732 
733 		mmap_write_lock(task_ref->mm);
734 		/*
735 		 * vma is set in the original mapping. But this mapping
736 		 * is done with mmap() after the window is opened with ioctl.
737 		 * so we may not see the original mapping if the core remove
738 		 * is done before the original mmap() and after the ioctl.
739 		 */
740 		if (vma)
741 			zap_page_range(vma, vma->vm_start,
742 					vma->vm_end - vma->vm_start);
743 
744 		mmap_write_unlock(task_ref->mm);
745 		mutex_unlock(&task_ref->mmap_mutex);
746 		/*
747 		 * Close VAS window in the hypervisor, but do not
748 		 * free vas_window struct since it may be reused
749 		 * when the credit is available later (DLPAR with
750 		 * adding cores). This struct will be used
751 		 * later when the process issued with close(FD).
752 		 */
753 		rc = deallocate_free_window(win);
754 		/*
755 		 * This failure is from the hypervisor.
756 		 * No way to stop migration for these failures.
757 		 * So ignore error and continue closing other windows.
758 		 */
759 		if (rc && !migrate)
760 			return rc;
761 
762 		vcap->nr_close_wins++;
763 
764 		/*
765 		 * For migration, do not depend on lpar_creds in case if
766 		 * mismatch with the hypervisor value (should not happen).
767 		 * So close all active windows in the list and will be
768 		 * reopened windows based on the new lpar_creds on the
769 		 * destination system during resume.
770 		 */
771 		if (!migrate && !--excess_creds)
772 			break;
773 	}
774 
775 	return 0;
776 }
777 
778 /*
779  * Get new VAS capabilities when the core add/removal configuration
780  * changes. Reconfig window configurations based on the credits
781  * availability from this new capabilities.
782  */
783 int vas_reconfig_capabilties(u8 type, int new_nr_creds)
784 {
785 	struct vas_cop_feat_caps *caps;
786 	int old_nr_creds;
787 	struct vas_caps *vcaps;
788 	int rc = 0, nr_active_wins;
789 
790 	if (type >= VAS_MAX_FEAT_TYPE) {
791 		pr_err("Invalid credit type %d\n", type);
792 		return -EINVAL;
793 	}
794 
795 	vcaps = &vascaps[type];
796 	caps = &vcaps->caps;
797 
798 	mutex_lock(&vas_pseries_mutex);
799 
800 	old_nr_creds = atomic_read(&caps->nr_total_credits);
801 
802 	atomic_set(&caps->nr_total_credits, new_nr_creds);
803 	/*
804 	 * The total number of available credits may be decreased or
805 	 * increased with DLPAR operation. Means some windows have to be
806 	 * closed / reopened. Hold the vas_pseries_mutex so that the
807 	 * user space can not open new windows.
808 	 */
809 	if (old_nr_creds <  new_nr_creds) {
810 		/*
811 		 * If the existing target credits is less than the new
812 		 * target, reopen windows if they are closed due to
813 		 * the previous DLPAR (core removal).
814 		 */
815 		rc = reconfig_open_windows(vcaps, new_nr_creds - old_nr_creds,
816 					   false);
817 	} else {
818 		/*
819 		 * # active windows is more than new LPAR available
820 		 * credits. So close the excessive windows.
821 		 * On pseries, each window will have 1 credit.
822 		 */
823 		nr_active_wins = vcaps->nr_open_windows - vcaps->nr_close_wins;
824 		if (nr_active_wins > new_nr_creds)
825 			rc = reconfig_close_windows(vcaps,
826 					nr_active_wins - new_nr_creds,
827 					false);
828 	}
829 
830 	mutex_unlock(&vas_pseries_mutex);
831 	return rc;
832 }
833 /*
834  * Total number of default credits available (target_credits)
835  * in LPAR depends on number of cores configured. It varies based on
836  * whether processors are in shared mode or dedicated mode.
837  * Get the notifier when CPU configuration is changed with DLPAR
838  * operation so that get the new target_credits (vas default capabilities)
839  * and then update the existing windows usage if needed.
840  */
841 static int pseries_vas_notifier(struct notifier_block *nb,
842 				unsigned long action, void *data)
843 {
844 	struct of_reconfig_data *rd = data;
845 	struct device_node *dn = rd->dn;
846 	const __be32 *intserv = NULL;
847 	int new_nr_creds, len, rc = 0;
848 
849 	if ((action == OF_RECONFIG_ATTACH_NODE) ||
850 		(action == OF_RECONFIG_DETACH_NODE))
851 		intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s",
852 					  &len);
853 	/*
854 	 * Processor config is not changed
855 	 */
856 	if (!intserv)
857 		return NOTIFY_OK;
858 
859 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES,
860 					vascaps[VAS_GZIP_DEF_FEAT_TYPE].feat,
861 					(u64)virt_to_phys(&hv_cop_caps));
862 	if (!rc) {
863 		new_nr_creds = be16_to_cpu(hv_cop_caps.target_lpar_creds);
864 		rc = vas_reconfig_capabilties(VAS_GZIP_DEF_FEAT_TYPE,
865 						new_nr_creds);
866 	}
867 
868 	if (rc)
869 		pr_err("Failed reconfig VAS capabilities with DLPAR\n");
870 
871 	return rc;
872 }
873 
874 static struct notifier_block pseries_vas_nb = {
875 	.notifier_call = pseries_vas_notifier,
876 };
877 
878 /*
879  * For LPM, all windows have to be closed on the source partition
880  * before migration and reopen them on the destination partition
881  * after migration. So closing windows during suspend and
882  * reopen them during resume.
883  */
884 int vas_migration_handler(int action)
885 {
886 	struct vas_cop_feat_caps *caps;
887 	int old_nr_creds, new_nr_creds = 0;
888 	struct vas_caps *vcaps;
889 	int i, rc = 0;
890 
891 	/*
892 	 * NX-GZIP is not enabled. Nothing to do for migration.
893 	 */
894 	if (!copypaste_feat)
895 		return rc;
896 
897 	mutex_lock(&vas_pseries_mutex);
898 
899 	if (action == VAS_SUSPEND)
900 		migration_in_progress = true;
901 	else
902 		migration_in_progress = false;
903 
904 	for (i = 0; i < VAS_MAX_FEAT_TYPE; i++) {
905 		vcaps = &vascaps[i];
906 		caps = &vcaps->caps;
907 		old_nr_creds = atomic_read(&caps->nr_total_credits);
908 
909 		rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES,
910 					      vcaps->feat,
911 					      (u64)virt_to_phys(&hv_cop_caps));
912 		if (!rc) {
913 			new_nr_creds = be16_to_cpu(hv_cop_caps.target_lpar_creds);
914 			/*
915 			 * Should not happen. But incase print messages, close
916 			 * all windows in the list during suspend and reopen
917 			 * windows based on new lpar_creds on the destination
918 			 * system.
919 			 */
920 			if (old_nr_creds != new_nr_creds) {
921 				pr_err("Target credits mismatch with the hypervisor\n");
922 				pr_err("state(%d): lpar creds: %d HV lpar creds: %d\n",
923 					action, old_nr_creds, new_nr_creds);
924 				pr_err("Used creds: %d, Active creds: %d\n",
925 					atomic_read(&caps->nr_used_credits),
926 					vcaps->nr_open_windows - vcaps->nr_close_wins);
927 			}
928 		} else {
929 			pr_err("state(%d): Get VAS capabilities failed with %d\n",
930 				action, rc);
931 			/*
932 			 * We can not stop migration with the current lpm
933 			 * implementation. So continue closing all windows in
934 			 * the list (during suspend) and return without
935 			 * opening windows (during resume) if VAS capabilities
936 			 * HCALL failed.
937 			 */
938 			if (action == VAS_RESUME)
939 				goto out;
940 		}
941 
942 		switch (action) {
943 		case VAS_SUSPEND:
944 			rc = reconfig_close_windows(vcaps, vcaps->nr_open_windows,
945 							true);
946 			break;
947 		case VAS_RESUME:
948 			atomic_set(&caps->nr_total_credits, new_nr_creds);
949 			rc = reconfig_open_windows(vcaps, new_nr_creds, true);
950 			break;
951 		default:
952 			/* should not happen */
953 			pr_err("Invalid migration action %d\n", action);
954 			rc = -EINVAL;
955 			goto out;
956 		}
957 
958 		/*
959 		 * Ignore errors during suspend and return for resume.
960 		 */
961 		if (rc && (action == VAS_RESUME))
962 			goto out;
963 	}
964 
965 out:
966 	mutex_unlock(&vas_pseries_mutex);
967 	return rc;
968 }
969 
970 static int __init pseries_vas_init(void)
971 {
972 	struct hv_vas_all_caps *hv_caps;
973 	int rc = 0;
974 
975 	/*
976 	 * Linux supports user space COPY/PASTE only with Radix
977 	 */
978 	if (!radix_enabled()) {
979 		pr_err("API is supported only with radix page tables\n");
980 		return -ENOTSUPP;
981 	}
982 
983 	hv_caps = kmalloc(sizeof(*hv_caps), GFP_KERNEL);
984 	if (!hv_caps)
985 		return -ENOMEM;
986 	/*
987 	 * Get VAS overall capabilities by passing 0 to feature type.
988 	 */
989 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, 0,
990 					  (u64)virt_to_phys(hv_caps));
991 	if (rc)
992 		goto out;
993 
994 	caps_all.descriptor = be64_to_cpu(hv_caps->descriptor);
995 	caps_all.feat_type = be64_to_cpu(hv_caps->feat_type);
996 
997 	sysfs_pseries_vas_init(&caps_all);
998 
999 	/*
1000 	 * QOS capabilities available
1001 	 */
1002 	if (caps_all.feat_type & VAS_GZIP_QOS_FEAT_BIT) {
1003 		rc = get_vas_capabilities(VAS_GZIP_QOS_FEAT,
1004 					  VAS_GZIP_QOS_FEAT_TYPE, &hv_cop_caps);
1005 
1006 		if (rc)
1007 			goto out;
1008 	}
1009 	/*
1010 	 * Default capabilities available
1011 	 */
1012 	if (caps_all.feat_type & VAS_GZIP_DEF_FEAT_BIT)
1013 		rc = get_vas_capabilities(VAS_GZIP_DEF_FEAT,
1014 					  VAS_GZIP_DEF_FEAT_TYPE, &hv_cop_caps);
1015 
1016 	if (!rc && copypaste_feat) {
1017 		if (firmware_has_feature(FW_FEATURE_LPAR))
1018 			of_reconfig_notifier_register(&pseries_vas_nb);
1019 
1020 		pr_info("GZIP feature is available\n");
1021 	} else {
1022 		/*
1023 		 * Should not happen, but only when get default
1024 		 * capabilities HCALL failed. So disable copy paste
1025 		 * feature.
1026 		 */
1027 		copypaste_feat = false;
1028 	}
1029 
1030 out:
1031 	kfree(hv_caps);
1032 	return rc;
1033 }
1034 machine_device_initcall(pseries, pseries_vas_init);
1035