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 	while (atomic_read(&txwin->pending_faults)) {
204 		rc = h_get_nx_fault(txwin->vas_win.winid, (u64)virt_to_phys(&crb));
205 		if (!rc) {
206 			tsk_ref = &txwin->vas_win.task_ref;
207 			vas_dump_crb(&crb);
208 			vas_update_csb(&crb, tsk_ref);
209 		}
210 		atomic_dec(&txwin->pending_faults);
211 	}
212 
213 	return IRQ_HANDLED;
214 }
215 
216 /*
217  * irq_default_primary_handler() can be used only with IRQF_ONESHOT
218  * which disables IRQ before executing the thread handler and enables
219  * it after. But this disabling interrupt sets the VAS IRQ OFF
220  * state in the hypervisor. If the NX generates fault interrupt
221  * during this window, the hypervisor will not deliver this
222  * interrupt to the LPAR. So use VAS specific IRQ handler instead
223  * of calling the default primary handler.
224  */
225 static irqreturn_t pseries_vas_irq_handler(int irq, void *data)
226 {
227 	struct pseries_vas_window *txwin = data;
228 
229 	/*
230 	 * The thread hanlder will process this interrupt if it is
231 	 * already running.
232 	 */
233 	atomic_inc(&txwin->pending_faults);
234 
235 	return IRQ_WAKE_THREAD;
236 }
237 
238 /*
239  * Allocate window and setup IRQ mapping.
240  */
241 static int allocate_setup_window(struct pseries_vas_window *txwin,
242 				 u64 *domain, u8 wintype)
243 {
244 	int rc;
245 
246 	rc = h_allocate_vas_window(txwin, domain, wintype, DEF_WIN_CREDS);
247 	if (rc)
248 		return rc;
249 	/*
250 	 * On PowerVM, the hypervisor setup and forwards the fault
251 	 * interrupt per window. So the IRQ setup and fault handling
252 	 * will be done for each open window separately.
253 	 */
254 	txwin->fault_virq = irq_create_mapping(NULL, txwin->fault_irq);
255 	if (!txwin->fault_virq) {
256 		pr_err("Failed irq mapping %d\n", txwin->fault_irq);
257 		rc = -EINVAL;
258 		goto out_win;
259 	}
260 
261 	txwin->name = kasprintf(GFP_KERNEL, "vas-win-%d",
262 				txwin->vas_win.winid);
263 	if (!txwin->name) {
264 		rc = -ENOMEM;
265 		goto out_irq;
266 	}
267 
268 	rc = request_threaded_irq(txwin->fault_virq,
269 				  pseries_vas_irq_handler,
270 				  pseries_vas_fault_thread_fn, 0,
271 				  txwin->name, txwin);
272 	if (rc) {
273 		pr_err("VAS-Window[%d]: Request IRQ(%u) failed with %d\n",
274 		       txwin->vas_win.winid, txwin->fault_virq, rc);
275 		goto out_free;
276 	}
277 
278 	txwin->vas_win.wcreds_max = DEF_WIN_CREDS;
279 
280 	return 0;
281 out_free:
282 	kfree(txwin->name);
283 out_irq:
284 	irq_dispose_mapping(txwin->fault_virq);
285 out_win:
286 	h_deallocate_vas_window(txwin->vas_win.winid);
287 	return rc;
288 }
289 
290 static inline void free_irq_setup(struct pseries_vas_window *txwin)
291 {
292 	free_irq(txwin->fault_virq, txwin);
293 	kfree(txwin->name);
294 	irq_dispose_mapping(txwin->fault_virq);
295 }
296 
297 static struct vas_window *vas_allocate_window(int vas_id, u64 flags,
298 					      enum vas_cop_type cop_type)
299 {
300 	long domain[PLPAR_HCALL9_BUFSIZE] = {VAS_DEFAULT_DOMAIN_ID};
301 	struct vas_cop_feat_caps *cop_feat_caps;
302 	struct vas_caps *caps;
303 	struct pseries_vas_window *txwin;
304 	int rc;
305 
306 	txwin = kzalloc(sizeof(*txwin), GFP_KERNEL);
307 	if (!txwin)
308 		return ERR_PTR(-ENOMEM);
309 
310 	/*
311 	 * A VAS window can have many credits which means that many
312 	 * requests can be issued simultaneously. But the hypervisor
313 	 * restricts one credit per window.
314 	 * The hypervisor introduces 2 different types of credits:
315 	 * Default credit type (Uses normal priority FIFO):
316 	 *	A limited number of credits are assigned to partitions
317 	 *	based on processor entitlement. But these credits may be
318 	 *	over-committed on a system depends on whether the CPUs
319 	 *	are in shared or dedicated modes - that is, more requests
320 	 *	may be issued across the system than NX can service at
321 	 *	once which can result in paste command failure (RMA_busy).
322 	 *	Then the process has to resend requests or fall-back to
323 	 *	SW compression.
324 	 * Quality of Service (QoS) credit type (Uses high priority FIFO):
325 	 *	To avoid NX HW contention, the system admins can assign
326 	 *	QoS credits for each LPAR so that this partition is
327 	 *	guaranteed access to NX resources. These credits are
328 	 *	assigned to partitions via the HMC.
329 	 *	Refer PAPR for more information.
330 	 *
331 	 * Allocate window with QoS credits if user requested. Otherwise
332 	 * default credits are used.
333 	 */
334 	if (flags & VAS_TX_WIN_FLAG_QOS_CREDIT)
335 		caps = &vascaps[VAS_GZIP_QOS_FEAT_TYPE];
336 	else
337 		caps = &vascaps[VAS_GZIP_DEF_FEAT_TYPE];
338 
339 	cop_feat_caps = &caps->caps;
340 
341 	if (atomic_inc_return(&cop_feat_caps->nr_used_credits) >
342 			atomic_read(&cop_feat_caps->nr_total_credits)) {
343 		pr_err("Credits are not available to allocate window\n");
344 		rc = -EINVAL;
345 		goto out;
346 	}
347 
348 	if (vas_id == -1) {
349 		/*
350 		 * The user space is requesting to allocate a window on
351 		 * a VAS instance where the process is executing.
352 		 * On PowerVM, domain values are passed to the hypervisor
353 		 * to select VAS instance. Useful if the process is
354 		 * affinity to NUMA node.
355 		 * The hypervisor selects VAS instance if
356 		 * VAS_DEFAULT_DOMAIN_ID (-1) is passed for domain values.
357 		 * The h_allocate_vas_window hcall is defined to take a
358 		 * domain values as specified by h_home_node_associativity,
359 		 * So no unpacking needs to be done.
360 		 */
361 		rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, domain,
362 				  VPHN_FLAG_VCPU, hard_smp_processor_id());
363 		if (rc != H_SUCCESS) {
364 			pr_err("H_HOME_NODE_ASSOCIATIVITY error: %d\n", rc);
365 			goto out;
366 		}
367 	}
368 
369 	txwin->pid = mfspr(SPRN_PID);
370 
371 	/*
372 	 * Allocate / Deallocate window hcalls and setup / free IRQs
373 	 * have to be protected with mutex.
374 	 * Open VAS window: Allocate window hcall and setup IRQ
375 	 * Close VAS window: Deallocate window hcall and free IRQ
376 	 *	The hypervisor waits until all NX requests are
377 	 *	completed before closing the window. So expects OS
378 	 *	to handle NX faults, means IRQ can be freed only
379 	 *	after the deallocate window hcall is returned.
380 	 * So once the window is closed with deallocate hcall before
381 	 * the IRQ is freed, it can be assigned to new allocate
382 	 * hcall with the same fault IRQ by the hypervisor. It can
383 	 * result in setup IRQ fail for the new window since the
384 	 * same fault IRQ is not freed by the OS before.
385 	 */
386 	mutex_lock(&vas_pseries_mutex);
387 	if (migration_in_progress)
388 		rc = -EBUSY;
389 	else
390 		rc = allocate_setup_window(txwin, (u64 *)&domain[0],
391 				   cop_feat_caps->win_type);
392 	mutex_unlock(&vas_pseries_mutex);
393 	if (rc)
394 		goto out;
395 
396 	/*
397 	 * Modify window and it is ready to use.
398 	 */
399 	rc = h_modify_vas_window(txwin);
400 	if (!rc)
401 		rc = get_vas_user_win_ref(&txwin->vas_win.task_ref);
402 	if (rc)
403 		goto out_free;
404 
405 	txwin->win_type = cop_feat_caps->win_type;
406 	mutex_lock(&vas_pseries_mutex);
407 	/*
408 	 * Possible to lose the acquired credit with DLPAR core
409 	 * removal after the window is opened. So if there are any
410 	 * closed windows (means with lost credits), do not give new
411 	 * window to user space. New windows will be opened only
412 	 * after the existing windows are reopened when credits are
413 	 * available.
414 	 */
415 	if (!caps->nr_close_wins) {
416 		list_add(&txwin->win_list, &caps->list);
417 		caps->nr_open_windows++;
418 		mutex_unlock(&vas_pseries_mutex);
419 		vas_user_win_add_mm_context(&txwin->vas_win.task_ref);
420 		return &txwin->vas_win;
421 	}
422 	mutex_unlock(&vas_pseries_mutex);
423 
424 	put_vas_user_win_ref(&txwin->vas_win.task_ref);
425 	rc = -EBUSY;
426 	pr_err("No credit is available to allocate window\n");
427 
428 out_free:
429 	/*
430 	 * Window is not operational. Free IRQ before closing
431 	 * window so that do not have to hold mutex.
432 	 */
433 	free_irq_setup(txwin);
434 	h_deallocate_vas_window(txwin->vas_win.winid);
435 out:
436 	atomic_dec(&cop_feat_caps->nr_used_credits);
437 	kfree(txwin);
438 	return ERR_PTR(rc);
439 }
440 
441 static u64 vas_paste_address(struct vas_window *vwin)
442 {
443 	struct pseries_vas_window *win;
444 
445 	win = container_of(vwin, struct pseries_vas_window, vas_win);
446 	return win->win_addr;
447 }
448 
449 static int deallocate_free_window(struct pseries_vas_window *win)
450 {
451 	int rc = 0;
452 
453 	/*
454 	 * The hypervisor waits for all requests including faults
455 	 * are processed before closing the window - Means all
456 	 * credits have to be returned. In the case of fault
457 	 * request, a credit is returned after OS issues
458 	 * H_GET_NX_FAULT hcall.
459 	 * So free IRQ after executing H_DEALLOCATE_VAS_WINDOW
460 	 * hcall.
461 	 */
462 	rc = h_deallocate_vas_window(win->vas_win.winid);
463 	if (!rc)
464 		free_irq_setup(win);
465 
466 	return rc;
467 }
468 
469 static int vas_deallocate_window(struct vas_window *vwin)
470 {
471 	struct pseries_vas_window *win;
472 	struct vas_cop_feat_caps *caps;
473 	int rc = 0;
474 
475 	if (!vwin)
476 		return -EINVAL;
477 
478 	win = container_of(vwin, struct pseries_vas_window, vas_win);
479 
480 	/* Should not happen */
481 	if (win->win_type >= VAS_MAX_FEAT_TYPE) {
482 		pr_err("Window (%u): Invalid window type %u\n",
483 				vwin->winid, win->win_type);
484 		return -EINVAL;
485 	}
486 
487 	caps = &vascaps[win->win_type].caps;
488 	mutex_lock(&vas_pseries_mutex);
489 	/*
490 	 * VAS window is already closed in the hypervisor when
491 	 * lost the credit or with migration. So just remove the entry
492 	 * from the list, remove task references and free vas_window
493 	 * struct.
494 	 */
495 	if (!(win->vas_win.status & VAS_WIN_NO_CRED_CLOSE) &&
496 		!(win->vas_win.status & VAS_WIN_MIGRATE_CLOSE)) {
497 		rc = deallocate_free_window(win);
498 		if (rc) {
499 			mutex_unlock(&vas_pseries_mutex);
500 			return rc;
501 		}
502 	} else
503 		vascaps[win->win_type].nr_close_wins--;
504 
505 	list_del(&win->win_list);
506 	atomic_dec(&caps->nr_used_credits);
507 	vascaps[win->win_type].nr_open_windows--;
508 	mutex_unlock(&vas_pseries_mutex);
509 
510 	mm_context_remove_vas_window(vwin->task_ref.mm);
511 	put_vas_user_win_ref(&vwin->task_ref);
512 
513 	kfree(win);
514 	return 0;
515 }
516 
517 static const struct vas_user_win_ops vops_pseries = {
518 	.open_win	= vas_allocate_window,	/* Open and configure window */
519 	.paste_addr	= vas_paste_address,	/* To do copy/paste */
520 	.close_win	= vas_deallocate_window, /* Close window */
521 };
522 
523 /*
524  * Supporting only nx-gzip coprocessor type now, but this API code
525  * extended to other coprocessor types later.
526  */
527 int vas_register_api_pseries(struct module *mod, enum vas_cop_type cop_type,
528 			     const char *name)
529 {
530 	if (!copypaste_feat)
531 		return -ENOTSUPP;
532 
533 	return vas_register_coproc_api(mod, cop_type, name, &vops_pseries);
534 }
535 EXPORT_SYMBOL_GPL(vas_register_api_pseries);
536 
537 void vas_unregister_api_pseries(void)
538 {
539 	vas_unregister_coproc_api();
540 }
541 EXPORT_SYMBOL_GPL(vas_unregister_api_pseries);
542 
543 /*
544  * Get the specific capabilities based on the feature type.
545  * Right now supports GZIP default and GZIP QoS capabilities.
546  */
547 static int __init get_vas_capabilities(u8 feat, enum vas_cop_feat_type type,
548 				struct hv_vas_cop_feat_caps *hv_caps)
549 {
550 	struct vas_cop_feat_caps *caps;
551 	struct vas_caps *vcaps;
552 	int rc = 0;
553 
554 	vcaps = &vascaps[type];
555 	memset(vcaps, 0, sizeof(*vcaps));
556 	INIT_LIST_HEAD(&vcaps->list);
557 
558 	vcaps->feat = feat;
559 	caps = &vcaps->caps;
560 
561 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, feat,
562 					  (u64)virt_to_phys(hv_caps));
563 	if (rc)
564 		return rc;
565 
566 	caps->user_mode = hv_caps->user_mode;
567 	if (!(caps->user_mode & VAS_COPY_PASTE_USER_MODE)) {
568 		pr_err("User space COPY/PASTE is not supported\n");
569 		return -ENOTSUPP;
570 	}
571 
572 	caps->descriptor = be64_to_cpu(hv_caps->descriptor);
573 	caps->win_type = hv_caps->win_type;
574 	if (caps->win_type >= VAS_MAX_FEAT_TYPE) {
575 		pr_err("Unsupported window type %u\n", caps->win_type);
576 		return -EINVAL;
577 	}
578 	caps->max_lpar_creds = be16_to_cpu(hv_caps->max_lpar_creds);
579 	caps->max_win_creds = be16_to_cpu(hv_caps->max_win_creds);
580 	atomic_set(&caps->nr_total_credits,
581 		   be16_to_cpu(hv_caps->target_lpar_creds));
582 	if (feat == VAS_GZIP_DEF_FEAT) {
583 		caps->def_lpar_creds = be16_to_cpu(hv_caps->def_lpar_creds);
584 
585 		if (caps->max_win_creds < DEF_WIN_CREDS) {
586 			pr_err("Window creds(%u) > max allowed window creds(%u)\n",
587 			       DEF_WIN_CREDS, caps->max_win_creds);
588 			return -EINVAL;
589 		}
590 	}
591 
592 	rc = sysfs_add_vas_caps(caps);
593 	if (rc)
594 		return rc;
595 
596 	copypaste_feat = true;
597 
598 	return 0;
599 }
600 
601 /*
602  * VAS windows can be closed due to lost credits when the core is
603  * removed. So reopen them if credits are available due to DLPAR
604  * core add and set the window active status. When NX sees the page
605  * fault on the unmapped paste address, the kernel handles the fault
606  * by setting the remapping to new paste address if the window is
607  * active.
608  */
609 static int reconfig_open_windows(struct vas_caps *vcaps, int creds,
610 				 bool migrate)
611 {
612 	long domain[PLPAR_HCALL9_BUFSIZE] = {VAS_DEFAULT_DOMAIN_ID};
613 	struct vas_cop_feat_caps *caps = &vcaps->caps;
614 	struct pseries_vas_window *win = NULL, *tmp;
615 	int rc, mv_ents = 0;
616 	int flag;
617 
618 	/*
619 	 * Nothing to do if there are no closed windows.
620 	 */
621 	if (!vcaps->nr_close_wins)
622 		return 0;
623 
624 	/*
625 	 * For the core removal, the hypervisor reduces the credits
626 	 * assigned to the LPAR and the kernel closes VAS windows
627 	 * in the hypervisor depends on reduced credits. The kernel
628 	 * uses LIFO (the last windows that are opened will be closed
629 	 * first) and expects to open in the same order when credits
630 	 * are available.
631 	 * For example, 40 windows are closed when the LPAR lost 2 cores
632 	 * (dedicated). If 1 core is added, this LPAR can have 20 more
633 	 * credits. It means the kernel can reopen 20 windows. So move
634 	 * 20 entries in the VAS windows lost and reopen next 20 windows.
635 	 * For partition migration, reopen all windows that are closed
636 	 * during resume.
637 	 */
638 	if ((vcaps->nr_close_wins > creds) && !migrate)
639 		mv_ents = vcaps->nr_close_wins - creds;
640 
641 	list_for_each_entry_safe(win, tmp, &vcaps->list, win_list) {
642 		if (!mv_ents)
643 			break;
644 
645 		mv_ents--;
646 	}
647 
648 	/*
649 	 * Open windows if they are closed only with migration or
650 	 * DLPAR (lost credit) before.
651 	 */
652 	if (migrate)
653 		flag = VAS_WIN_MIGRATE_CLOSE;
654 	else
655 		flag = VAS_WIN_NO_CRED_CLOSE;
656 
657 	list_for_each_entry_safe_from(win, tmp, &vcaps->list, win_list) {
658 		/*
659 		 * This window is closed with DLPAR and migration events.
660 		 * So reopen the window with the last event.
661 		 * The user space is not suspended with the current
662 		 * migration notifier. So the user space can issue DLPAR
663 		 * CPU hotplug while migration in progress. In this case
664 		 * this window will be opened with the last event.
665 		 */
666 		if ((win->vas_win.status & VAS_WIN_NO_CRED_CLOSE) &&
667 			(win->vas_win.status & VAS_WIN_MIGRATE_CLOSE)) {
668 			win->vas_win.status &= ~flag;
669 			continue;
670 		}
671 
672 		/*
673 		 * Nothing to do on this window if it is not closed
674 		 * with this flag
675 		 */
676 		if (!(win->vas_win.status & flag))
677 			continue;
678 
679 		rc = allocate_setup_window(win, (u64 *)&domain[0],
680 					   caps->win_type);
681 		if (rc)
682 			return rc;
683 
684 		rc = h_modify_vas_window(win);
685 		if (rc)
686 			goto out;
687 
688 		mutex_lock(&win->vas_win.task_ref.mmap_mutex);
689 		/*
690 		 * Set window status to active
691 		 */
692 		win->vas_win.status &= ~flag;
693 		mutex_unlock(&win->vas_win.task_ref.mmap_mutex);
694 		win->win_type = caps->win_type;
695 		if (!--vcaps->nr_close_wins)
696 			break;
697 	}
698 
699 	return 0;
700 out:
701 	/*
702 	 * Window modify HCALL failed. So close the window to the
703 	 * hypervisor and return.
704 	 */
705 	free_irq_setup(win);
706 	h_deallocate_vas_window(win->vas_win.winid);
707 	return rc;
708 }
709 
710 /*
711  * The hypervisor reduces the available credits if the LPAR lost core. It
712  * means the excessive windows should not be active and the user space
713  * should not be using these windows to send compression requests to NX.
714  * So the kernel closes the excessive windows and unmap the paste address
715  * such that the user space receives paste instruction failure. Then up to
716  * the user space to fall back to SW compression and manage with the
717  * existing windows.
718  */
719 static int reconfig_close_windows(struct vas_caps *vcap, int excess_creds,
720 									bool migrate)
721 {
722 	struct pseries_vas_window *win, *tmp;
723 	struct vas_user_win_ref *task_ref;
724 	struct vm_area_struct *vma;
725 	int rc = 0, flag;
726 
727 	if (migrate)
728 		flag = VAS_WIN_MIGRATE_CLOSE;
729 	else
730 		flag = VAS_WIN_NO_CRED_CLOSE;
731 
732 	list_for_each_entry_safe(win, tmp, &vcap->list, win_list) {
733 		/*
734 		 * This window is already closed due to lost credit
735 		 * or for migration before. Go for next window.
736 		 * For migration, nothing to do since this window
737 		 * closed for DLPAR and will be reopened even on
738 		 * the destination system with other DLPAR operation.
739 		 */
740 		if ((win->vas_win.status & VAS_WIN_MIGRATE_CLOSE) ||
741 			(win->vas_win.status & VAS_WIN_NO_CRED_CLOSE)) {
742 			win->vas_win.status |= flag;
743 			continue;
744 		}
745 
746 		task_ref = &win->vas_win.task_ref;
747 		mutex_lock(&task_ref->mmap_mutex);
748 		vma = task_ref->vma;
749 		/*
750 		 * Number of available credits are reduced, So select
751 		 * and close windows.
752 		 */
753 		win->vas_win.status |= flag;
754 
755 		mmap_write_lock(task_ref->mm);
756 		/*
757 		 * vma is set in the original mapping. But this mapping
758 		 * is done with mmap() after the window is opened with ioctl.
759 		 * so we may not see the original mapping if the core remove
760 		 * is done before the original mmap() and after the ioctl.
761 		 */
762 		if (vma)
763 			zap_vma_pages(vma);
764 
765 		mmap_write_unlock(task_ref->mm);
766 		mutex_unlock(&task_ref->mmap_mutex);
767 		/*
768 		 * Close VAS window in the hypervisor, but do not
769 		 * free vas_window struct since it may be reused
770 		 * when the credit is available later (DLPAR with
771 		 * adding cores). This struct will be used
772 		 * later when the process issued with close(FD).
773 		 */
774 		rc = deallocate_free_window(win);
775 		/*
776 		 * This failure is from the hypervisor.
777 		 * No way to stop migration for these failures.
778 		 * So ignore error and continue closing other windows.
779 		 */
780 		if (rc && !migrate)
781 			return rc;
782 
783 		vcap->nr_close_wins++;
784 
785 		/*
786 		 * For migration, do not depend on lpar_creds in case if
787 		 * mismatch with the hypervisor value (should not happen).
788 		 * So close all active windows in the list and will be
789 		 * reopened windows based on the new lpar_creds on the
790 		 * destination system during resume.
791 		 */
792 		if (!migrate && !--excess_creds)
793 			break;
794 	}
795 
796 	return 0;
797 }
798 
799 /*
800  * Get new VAS capabilities when the core add/removal configuration
801  * changes. Reconfig window configurations based on the credits
802  * availability from this new capabilities.
803  */
804 int vas_reconfig_capabilties(u8 type, int new_nr_creds)
805 {
806 	struct vas_cop_feat_caps *caps;
807 	int old_nr_creds;
808 	struct vas_caps *vcaps;
809 	int rc = 0, nr_active_wins;
810 
811 	if (type >= VAS_MAX_FEAT_TYPE) {
812 		pr_err("Invalid credit type %d\n", type);
813 		return -EINVAL;
814 	}
815 
816 	vcaps = &vascaps[type];
817 	caps = &vcaps->caps;
818 
819 	mutex_lock(&vas_pseries_mutex);
820 
821 	old_nr_creds = atomic_read(&caps->nr_total_credits);
822 
823 	atomic_set(&caps->nr_total_credits, new_nr_creds);
824 	/*
825 	 * The total number of available credits may be decreased or
826 	 * increased with DLPAR operation. Means some windows have to be
827 	 * closed / reopened. Hold the vas_pseries_mutex so that the
828 	 * user space can not open new windows.
829 	 */
830 	if (old_nr_creds <  new_nr_creds) {
831 		/*
832 		 * If the existing target credits is less than the new
833 		 * target, reopen windows if they are closed due to
834 		 * the previous DLPAR (core removal).
835 		 */
836 		rc = reconfig_open_windows(vcaps, new_nr_creds - old_nr_creds,
837 					   false);
838 	} else {
839 		/*
840 		 * # active windows is more than new LPAR available
841 		 * credits. So close the excessive windows.
842 		 * On pseries, each window will have 1 credit.
843 		 */
844 		nr_active_wins = vcaps->nr_open_windows - vcaps->nr_close_wins;
845 		if (nr_active_wins > new_nr_creds)
846 			rc = reconfig_close_windows(vcaps,
847 					nr_active_wins - new_nr_creds,
848 					false);
849 	}
850 
851 	mutex_unlock(&vas_pseries_mutex);
852 	return rc;
853 }
854 
855 int pseries_vas_dlpar_cpu(void)
856 {
857 	int new_nr_creds, rc;
858 
859 	/*
860 	 * NX-GZIP is not enabled. Nothing to do for DLPAR event
861 	 */
862 	if (!copypaste_feat)
863 		return 0;
864 
865 
866 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES,
867 				      vascaps[VAS_GZIP_DEF_FEAT_TYPE].feat,
868 				      (u64)virt_to_phys(&hv_cop_caps));
869 	if (!rc) {
870 		new_nr_creds = be16_to_cpu(hv_cop_caps.target_lpar_creds);
871 		rc = vas_reconfig_capabilties(VAS_GZIP_DEF_FEAT_TYPE, new_nr_creds);
872 	}
873 
874 	if (rc)
875 		pr_err("Failed reconfig VAS capabilities with DLPAR\n");
876 
877 	return rc;
878 }
879 
880 /*
881  * Total number of default credits available (target_credits)
882  * in LPAR depends on number of cores configured. It varies based on
883  * whether processors are in shared mode or dedicated mode.
884  * Get the notifier when CPU configuration is changed with DLPAR
885  * operation so that get the new target_credits (vas default capabilities)
886  * and then update the existing windows usage if needed.
887  */
888 static int pseries_vas_notifier(struct notifier_block *nb,
889 				unsigned long action, void *data)
890 {
891 	struct of_reconfig_data *rd = data;
892 	struct device_node *dn = rd->dn;
893 	const __be32 *intserv = NULL;
894 	int len;
895 
896 	/*
897 	 * For shared CPU partition, the hypervisor assigns total credits
898 	 * based on entitled core capacity. So updating VAS windows will
899 	 * be called from lparcfg_write().
900 	 */
901 	if (is_shared_processor())
902 		return NOTIFY_OK;
903 
904 	if ((action == OF_RECONFIG_ATTACH_NODE) ||
905 		(action == OF_RECONFIG_DETACH_NODE))
906 		intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s",
907 					  &len);
908 	/*
909 	 * Processor config is not changed
910 	 */
911 	if (!intserv)
912 		return NOTIFY_OK;
913 
914 	return pseries_vas_dlpar_cpu();
915 }
916 
917 static struct notifier_block pseries_vas_nb = {
918 	.notifier_call = pseries_vas_notifier,
919 };
920 
921 /*
922  * For LPM, all windows have to be closed on the source partition
923  * before migration and reopen them on the destination partition
924  * after migration. So closing windows during suspend and
925  * reopen them during resume.
926  */
927 int vas_migration_handler(int action)
928 {
929 	struct vas_cop_feat_caps *caps;
930 	int old_nr_creds, new_nr_creds = 0;
931 	struct vas_caps *vcaps;
932 	int i, rc = 0;
933 
934 	/*
935 	 * NX-GZIP is not enabled. Nothing to do for migration.
936 	 */
937 	if (!copypaste_feat)
938 		return rc;
939 
940 	mutex_lock(&vas_pseries_mutex);
941 
942 	if (action == VAS_SUSPEND)
943 		migration_in_progress = true;
944 	else
945 		migration_in_progress = false;
946 
947 	for (i = 0; i < VAS_MAX_FEAT_TYPE; i++) {
948 		vcaps = &vascaps[i];
949 		caps = &vcaps->caps;
950 		old_nr_creds = atomic_read(&caps->nr_total_credits);
951 
952 		rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES,
953 					      vcaps->feat,
954 					      (u64)virt_to_phys(&hv_cop_caps));
955 		if (!rc) {
956 			new_nr_creds = be16_to_cpu(hv_cop_caps.target_lpar_creds);
957 			/*
958 			 * Should not happen. But incase print messages, close
959 			 * all windows in the list during suspend and reopen
960 			 * windows based on new lpar_creds on the destination
961 			 * system.
962 			 */
963 			if (old_nr_creds != new_nr_creds) {
964 				pr_err("Target credits mismatch with the hypervisor\n");
965 				pr_err("state(%d): lpar creds: %d HV lpar creds: %d\n",
966 					action, old_nr_creds, new_nr_creds);
967 				pr_err("Used creds: %d, Active creds: %d\n",
968 					atomic_read(&caps->nr_used_credits),
969 					vcaps->nr_open_windows - vcaps->nr_close_wins);
970 			}
971 		} else {
972 			pr_err("state(%d): Get VAS capabilities failed with %d\n",
973 				action, rc);
974 			/*
975 			 * We can not stop migration with the current lpm
976 			 * implementation. So continue closing all windows in
977 			 * the list (during suspend) and return without
978 			 * opening windows (during resume) if VAS capabilities
979 			 * HCALL failed.
980 			 */
981 			if (action == VAS_RESUME)
982 				goto out;
983 		}
984 
985 		switch (action) {
986 		case VAS_SUSPEND:
987 			rc = reconfig_close_windows(vcaps, vcaps->nr_open_windows,
988 							true);
989 			break;
990 		case VAS_RESUME:
991 			atomic_set(&caps->nr_total_credits, new_nr_creds);
992 			rc = reconfig_open_windows(vcaps, new_nr_creds, true);
993 			break;
994 		default:
995 			/* should not happen */
996 			pr_err("Invalid migration action %d\n", action);
997 			rc = -EINVAL;
998 			goto out;
999 		}
1000 
1001 		/*
1002 		 * Ignore errors during suspend and return for resume.
1003 		 */
1004 		if (rc && (action == VAS_RESUME))
1005 			goto out;
1006 	}
1007 
1008 out:
1009 	mutex_unlock(&vas_pseries_mutex);
1010 	return rc;
1011 }
1012 
1013 static int __init pseries_vas_init(void)
1014 {
1015 	struct hv_vas_all_caps *hv_caps;
1016 	int rc = 0;
1017 
1018 	/*
1019 	 * Linux supports user space COPY/PASTE only with Radix
1020 	 */
1021 	if (!radix_enabled()) {
1022 		copypaste_feat = false;
1023 		pr_err("API is supported only with radix page tables\n");
1024 		return -ENOTSUPP;
1025 	}
1026 
1027 	hv_caps = kmalloc(sizeof(*hv_caps), GFP_KERNEL);
1028 	if (!hv_caps)
1029 		return -ENOMEM;
1030 	/*
1031 	 * Get VAS overall capabilities by passing 0 to feature type.
1032 	 */
1033 	rc = h_query_vas_capabilities(H_QUERY_VAS_CAPABILITIES, 0,
1034 					  (u64)virt_to_phys(hv_caps));
1035 	if (rc)
1036 		goto out;
1037 
1038 	caps_all.descriptor = be64_to_cpu(hv_caps->descriptor);
1039 	caps_all.feat_type = be64_to_cpu(hv_caps->feat_type);
1040 
1041 	sysfs_pseries_vas_init(&caps_all);
1042 
1043 	/*
1044 	 * QOS capabilities available
1045 	 */
1046 	if (caps_all.feat_type & VAS_GZIP_QOS_FEAT_BIT) {
1047 		rc = get_vas_capabilities(VAS_GZIP_QOS_FEAT,
1048 					  VAS_GZIP_QOS_FEAT_TYPE, &hv_cop_caps);
1049 
1050 		if (rc)
1051 			goto out;
1052 	}
1053 	/*
1054 	 * Default capabilities available
1055 	 */
1056 	if (caps_all.feat_type & VAS_GZIP_DEF_FEAT_BIT)
1057 		rc = get_vas_capabilities(VAS_GZIP_DEF_FEAT,
1058 					  VAS_GZIP_DEF_FEAT_TYPE, &hv_cop_caps);
1059 
1060 	if (!rc && copypaste_feat) {
1061 		if (firmware_has_feature(FW_FEATURE_LPAR))
1062 			of_reconfig_notifier_register(&pseries_vas_nb);
1063 
1064 		pr_info("GZIP feature is available\n");
1065 	} else {
1066 		/*
1067 		 * Should not happen, but only when get default
1068 		 * capabilities HCALL failed. So disable copy paste
1069 		 * feature.
1070 		 */
1071 		copypaste_feat = false;
1072 	}
1073 
1074 out:
1075 	kfree(hv_caps);
1076 	return rc;
1077 }
1078 machine_device_initcall(pseries, pseries_vas_init);
1079