xref: /linux/arch/powerpc/platforms/powernv/pci-ioda.c (revision 0883c2c06fb5bcf5b9e008270827e63c09a88c1e)
1 /*
2  * Support PCI/PCIe on PowerNV platforms
3  *
4  * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11 
12 #undef DEBUG
13 
14 #include <linux/kernel.h>
15 #include <linux/pci.h>
16 #include <linux/crash_dump.h>
17 #include <linux/debugfs.h>
18 #include <linux/delay.h>
19 #include <linux/string.h>
20 #include <linux/init.h>
21 #include <linux/bootmem.h>
22 #include <linux/irq.h>
23 #include <linux/io.h>
24 #include <linux/msi.h>
25 #include <linux/memblock.h>
26 #include <linux/iommu.h>
27 #include <linux/rculist.h>
28 #include <linux/sizes.h>
29 
30 #include <asm/sections.h>
31 #include <asm/io.h>
32 #include <asm/prom.h>
33 #include <asm/pci-bridge.h>
34 #include <asm/machdep.h>
35 #include <asm/msi_bitmap.h>
36 #include <asm/ppc-pci.h>
37 #include <asm/opal.h>
38 #include <asm/iommu.h>
39 #include <asm/tce.h>
40 #include <asm/xics.h>
41 #include <asm/debug.h>
42 #include <asm/firmware.h>
43 #include <asm/pnv-pci.h>
44 #include <asm/mmzone.h>
45 
46 #include <misc/cxl-base.h>
47 
48 #include "powernv.h"
49 #include "pci.h"
50 
51 #define PNV_IODA1_M64_NUM	16	/* Number of M64 BARs	*/
52 #define PNV_IODA1_M64_SEGS	8	/* Segments per M64 BAR	*/
53 #define PNV_IODA1_DMA32_SEGSIZE	0x10000000
54 
55 #define POWERNV_IOMMU_DEFAULT_LEVELS	1
56 #define POWERNV_IOMMU_MAX_LEVELS	5
57 
58 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU" };
59 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl);
60 
61 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
62 			    const char *fmt, ...)
63 {
64 	struct va_format vaf;
65 	va_list args;
66 	char pfix[32];
67 
68 	va_start(args, fmt);
69 
70 	vaf.fmt = fmt;
71 	vaf.va = &args;
72 
73 	if (pe->flags & PNV_IODA_PE_DEV)
74 		strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
75 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
76 		sprintf(pfix, "%04x:%02x     ",
77 			pci_domain_nr(pe->pbus), pe->pbus->number);
78 #ifdef CONFIG_PCI_IOV
79 	else if (pe->flags & PNV_IODA_PE_VF)
80 		sprintf(pfix, "%04x:%02x:%2x.%d",
81 			pci_domain_nr(pe->parent_dev->bus),
82 			(pe->rid & 0xff00) >> 8,
83 			PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
84 #endif /* CONFIG_PCI_IOV*/
85 
86 	printk("%spci %s: [PE# %.3d] %pV",
87 	       level, pfix, pe->pe_number, &vaf);
88 
89 	va_end(args);
90 }
91 
92 static bool pnv_iommu_bypass_disabled __read_mostly;
93 
94 static int __init iommu_setup(char *str)
95 {
96 	if (!str)
97 		return -EINVAL;
98 
99 	while (*str) {
100 		if (!strncmp(str, "nobypass", 8)) {
101 			pnv_iommu_bypass_disabled = true;
102 			pr_info("PowerNV: IOMMU bypass window disabled.\n");
103 			break;
104 		}
105 		str += strcspn(str, ",");
106 		if (*str == ',')
107 			str++;
108 	}
109 
110 	return 0;
111 }
112 early_param("iommu", iommu_setup);
113 
114 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags)
115 {
116 	return ((flags & (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)) ==
117 		(IORESOURCE_MEM_64 | IORESOURCE_PREFETCH));
118 }
119 
120 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
121 {
122 	phb->ioda.pe_array[pe_no].phb = phb;
123 	phb->ioda.pe_array[pe_no].pe_number = pe_no;
124 
125 	return &phb->ioda.pe_array[pe_no];
126 }
127 
128 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
129 {
130 	if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
131 		pr_warn("%s: Invalid PE %d on PHB#%x\n",
132 			__func__, pe_no, phb->hose->global_number);
133 		return;
134 	}
135 
136 	if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
137 		pr_debug("%s: PE %d was reserved on PHB#%x\n",
138 			 __func__, pe_no, phb->hose->global_number);
139 
140 	pnv_ioda_init_pe(phb, pe_no);
141 }
142 
143 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
144 {
145 	unsigned long pe = phb->ioda.total_pe_num - 1;
146 
147 	for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
148 		if (!test_and_set_bit(pe, phb->ioda.pe_alloc))
149 			return pnv_ioda_init_pe(phb, pe);
150 	}
151 
152 	return NULL;
153 }
154 
155 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
156 {
157 	struct pnv_phb *phb = pe->phb;
158 
159 	WARN_ON(pe->pdev);
160 
161 	memset(pe, 0, sizeof(struct pnv_ioda_pe));
162 	clear_bit(pe->pe_number, phb->ioda.pe_alloc);
163 }
164 
165 /* The default M64 BAR is shared by all PEs */
166 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
167 {
168 	const char *desc;
169 	struct resource *r;
170 	s64 rc;
171 
172 	/* Configure the default M64 BAR */
173 	rc = opal_pci_set_phb_mem_window(phb->opal_id,
174 					 OPAL_M64_WINDOW_TYPE,
175 					 phb->ioda.m64_bar_idx,
176 					 phb->ioda.m64_base,
177 					 0, /* unused */
178 					 phb->ioda.m64_size);
179 	if (rc != OPAL_SUCCESS) {
180 		desc = "configuring";
181 		goto fail;
182 	}
183 
184 	/* Enable the default M64 BAR */
185 	rc = opal_pci_phb_mmio_enable(phb->opal_id,
186 				      OPAL_M64_WINDOW_TYPE,
187 				      phb->ioda.m64_bar_idx,
188 				      OPAL_ENABLE_M64_SPLIT);
189 	if (rc != OPAL_SUCCESS) {
190 		desc = "enabling";
191 		goto fail;
192 	}
193 
194 	/* Mark the M64 BAR assigned */
195 	set_bit(phb->ioda.m64_bar_idx, &phb->ioda.m64_bar_alloc);
196 
197 	/*
198 	 * Exclude the segments for reserved and root bus PE, which
199 	 * are first or last two PEs.
200 	 */
201 	r = &phb->hose->mem_resources[1];
202 	if (phb->ioda.reserved_pe_idx == 0)
203 		r->start += (2 * phb->ioda.m64_segsize);
204 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
205 		r->end -= (2 * phb->ioda.m64_segsize);
206 	else
207 		pr_warn("  Cannot strip M64 segment for reserved PE#%d\n",
208 			phb->ioda.reserved_pe_idx);
209 
210 	return 0;
211 
212 fail:
213 	pr_warn("  Failure %lld %s M64 BAR#%d\n",
214 		rc, desc, phb->ioda.m64_bar_idx);
215 	opal_pci_phb_mmio_enable(phb->opal_id,
216 				 OPAL_M64_WINDOW_TYPE,
217 				 phb->ioda.m64_bar_idx,
218 				 OPAL_DISABLE_M64);
219 	return -EIO;
220 }
221 
222 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
223 					 unsigned long *pe_bitmap)
224 {
225 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
226 	struct pnv_phb *phb = hose->private_data;
227 	struct resource *r;
228 	resource_size_t base, sgsz, start, end;
229 	int segno, i;
230 
231 	base = phb->ioda.m64_base;
232 	sgsz = phb->ioda.m64_segsize;
233 	for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
234 		r = &pdev->resource[i];
235 		if (!r->parent || !pnv_pci_is_mem_pref_64(r->flags))
236 			continue;
237 
238 		start = _ALIGN_DOWN(r->start - base, sgsz);
239 		end = _ALIGN_UP(r->end - base, sgsz);
240 		for (segno = start / sgsz; segno < end / sgsz; segno++) {
241 			if (pe_bitmap)
242 				set_bit(segno, pe_bitmap);
243 			else
244 				pnv_ioda_reserve_pe(phb, segno);
245 		}
246 	}
247 }
248 
249 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
250 {
251 	struct resource *r;
252 	int index;
253 
254 	/*
255 	 * There are 16 M64 BARs, each of which has 8 segments. So
256 	 * there are as many M64 segments as the maximum number of
257 	 * PEs, which is 128.
258 	 */
259 	for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
260 		unsigned long base, segsz = phb->ioda.m64_segsize;
261 		int64_t rc;
262 
263 		base = phb->ioda.m64_base +
264 		       index * PNV_IODA1_M64_SEGS * segsz;
265 		rc = opal_pci_set_phb_mem_window(phb->opal_id,
266 				OPAL_M64_WINDOW_TYPE, index, base, 0,
267 				PNV_IODA1_M64_SEGS * segsz);
268 		if (rc != OPAL_SUCCESS) {
269 			pr_warn("  Error %lld setting M64 PHB#%d-BAR#%d\n",
270 				rc, phb->hose->global_number, index);
271 			goto fail;
272 		}
273 
274 		rc = opal_pci_phb_mmio_enable(phb->opal_id,
275 				OPAL_M64_WINDOW_TYPE, index,
276 				OPAL_ENABLE_M64_SPLIT);
277 		if (rc != OPAL_SUCCESS) {
278 			pr_warn("  Error %lld enabling M64 PHB#%d-BAR#%d\n",
279 				rc, phb->hose->global_number, index);
280 			goto fail;
281 		}
282 	}
283 
284 	/*
285 	 * Exclude the segments for reserved and root bus PE, which
286 	 * are first or last two PEs.
287 	 */
288 	r = &phb->hose->mem_resources[1];
289 	if (phb->ioda.reserved_pe_idx == 0)
290 		r->start += (2 * phb->ioda.m64_segsize);
291 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
292 		r->end -= (2 * phb->ioda.m64_segsize);
293 	else
294 		WARN(1, "Wrong reserved PE#%d on PHB#%d\n",
295 		     phb->ioda.reserved_pe_idx, phb->hose->global_number);
296 
297 	return 0;
298 
299 fail:
300 	for ( ; index >= 0; index--)
301 		opal_pci_phb_mmio_enable(phb->opal_id,
302 			OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
303 
304 	return -EIO;
305 }
306 
307 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
308 				    unsigned long *pe_bitmap,
309 				    bool all)
310 {
311 	struct pci_dev *pdev;
312 
313 	list_for_each_entry(pdev, &bus->devices, bus_list) {
314 		pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
315 
316 		if (all && pdev->subordinate)
317 			pnv_ioda_reserve_m64_pe(pdev->subordinate,
318 						pe_bitmap, all);
319 	}
320 }
321 
322 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
323 {
324 	struct pci_controller *hose = pci_bus_to_host(bus);
325 	struct pnv_phb *phb = hose->private_data;
326 	struct pnv_ioda_pe *master_pe, *pe;
327 	unsigned long size, *pe_alloc;
328 	int i;
329 
330 	/* Root bus shouldn't use M64 */
331 	if (pci_is_root_bus(bus))
332 		return NULL;
333 
334 	/* Allocate bitmap */
335 	size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
336 	pe_alloc = kzalloc(size, GFP_KERNEL);
337 	if (!pe_alloc) {
338 		pr_warn("%s: Out of memory !\n",
339 			__func__);
340 		return NULL;
341 	}
342 
343 	/* Figure out reserved PE numbers by the PE */
344 	pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
345 
346 	/*
347 	 * the current bus might not own M64 window and that's all
348 	 * contributed by its child buses. For the case, we needn't
349 	 * pick M64 dependent PE#.
350 	 */
351 	if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
352 		kfree(pe_alloc);
353 		return NULL;
354 	}
355 
356 	/*
357 	 * Figure out the master PE and put all slave PEs to master
358 	 * PE's list to form compound PE.
359 	 */
360 	master_pe = NULL;
361 	i = -1;
362 	while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
363 		phb->ioda.total_pe_num) {
364 		pe = &phb->ioda.pe_array[i];
365 
366 		phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
367 		if (!master_pe) {
368 			pe->flags |= PNV_IODA_PE_MASTER;
369 			INIT_LIST_HEAD(&pe->slaves);
370 			master_pe = pe;
371 		} else {
372 			pe->flags |= PNV_IODA_PE_SLAVE;
373 			pe->master = master_pe;
374 			list_add_tail(&pe->list, &master_pe->slaves);
375 		}
376 
377 		/*
378 		 * P7IOC supports M64DT, which helps mapping M64 segment
379 		 * to one particular PE#. However, PHB3 has fixed mapping
380 		 * between M64 segment and PE#. In order to have same logic
381 		 * for P7IOC and PHB3, we enforce fixed mapping between M64
382 		 * segment and PE# on P7IOC.
383 		 */
384 		if (phb->type == PNV_PHB_IODA1) {
385 			int64_t rc;
386 
387 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
388 					pe->pe_number, OPAL_M64_WINDOW_TYPE,
389 					pe->pe_number / PNV_IODA1_M64_SEGS,
390 					pe->pe_number % PNV_IODA1_M64_SEGS);
391 			if (rc != OPAL_SUCCESS)
392 				pr_warn("%s: Error %lld mapping M64 for PHB#%d-PE#%d\n",
393 					__func__, rc, phb->hose->global_number,
394 					pe->pe_number);
395 		}
396 	}
397 
398 	kfree(pe_alloc);
399 	return master_pe;
400 }
401 
402 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
403 {
404 	struct pci_controller *hose = phb->hose;
405 	struct device_node *dn = hose->dn;
406 	struct resource *res;
407 	const u32 *r;
408 	u64 pci_addr;
409 
410 	if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
411 		pr_info("  Not support M64 window\n");
412 		return;
413 	}
414 
415 	if (!firmware_has_feature(FW_FEATURE_OPAL)) {
416 		pr_info("  Firmware too old to support M64 window\n");
417 		return;
418 	}
419 
420 	r = of_get_property(dn, "ibm,opal-m64-window", NULL);
421 	if (!r) {
422 		pr_info("  No <ibm,opal-m64-window> on %s\n",
423 			dn->full_name);
424 		return;
425 	}
426 
427 	res = &hose->mem_resources[1];
428 	res->name = dn->full_name;
429 	res->start = of_translate_address(dn, r + 2);
430 	res->end = res->start + of_read_number(r + 4, 2) - 1;
431 	res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
432 	pci_addr = of_read_number(r, 2);
433 	hose->mem_offset[1] = res->start - pci_addr;
434 
435 	phb->ioda.m64_size = resource_size(res);
436 	phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
437 	phb->ioda.m64_base = pci_addr;
438 
439 	pr_info(" MEM64 0x%016llx..0x%016llx -> 0x%016llx\n",
440 			res->start, res->end, pci_addr);
441 
442 	/* Use last M64 BAR to cover M64 window */
443 	phb->ioda.m64_bar_idx = 15;
444 	if (phb->type == PNV_PHB_IODA1)
445 		phb->init_m64 = pnv_ioda1_init_m64;
446 	else
447 		phb->init_m64 = pnv_ioda2_init_m64;
448 	phb->reserve_m64_pe = pnv_ioda_reserve_m64_pe;
449 	phb->pick_m64_pe = pnv_ioda_pick_m64_pe;
450 }
451 
452 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
453 {
454 	struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
455 	struct pnv_ioda_pe *slave;
456 	s64 rc;
457 
458 	/* Fetch master PE */
459 	if (pe->flags & PNV_IODA_PE_SLAVE) {
460 		pe = pe->master;
461 		if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
462 			return;
463 
464 		pe_no = pe->pe_number;
465 	}
466 
467 	/* Freeze master PE */
468 	rc = opal_pci_eeh_freeze_set(phb->opal_id,
469 				     pe_no,
470 				     OPAL_EEH_ACTION_SET_FREEZE_ALL);
471 	if (rc != OPAL_SUCCESS) {
472 		pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
473 			__func__, rc, phb->hose->global_number, pe_no);
474 		return;
475 	}
476 
477 	/* Freeze slave PEs */
478 	if (!(pe->flags & PNV_IODA_PE_MASTER))
479 		return;
480 
481 	list_for_each_entry(slave, &pe->slaves, list) {
482 		rc = opal_pci_eeh_freeze_set(phb->opal_id,
483 					     slave->pe_number,
484 					     OPAL_EEH_ACTION_SET_FREEZE_ALL);
485 		if (rc != OPAL_SUCCESS)
486 			pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
487 				__func__, rc, phb->hose->global_number,
488 				slave->pe_number);
489 	}
490 }
491 
492 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
493 {
494 	struct pnv_ioda_pe *pe, *slave;
495 	s64 rc;
496 
497 	/* Find master PE */
498 	pe = &phb->ioda.pe_array[pe_no];
499 	if (pe->flags & PNV_IODA_PE_SLAVE) {
500 		pe = pe->master;
501 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
502 		pe_no = pe->pe_number;
503 	}
504 
505 	/* Clear frozen state for master PE */
506 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
507 	if (rc != OPAL_SUCCESS) {
508 		pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
509 			__func__, rc, opt, phb->hose->global_number, pe_no);
510 		return -EIO;
511 	}
512 
513 	if (!(pe->flags & PNV_IODA_PE_MASTER))
514 		return 0;
515 
516 	/* Clear frozen state for slave PEs */
517 	list_for_each_entry(slave, &pe->slaves, list) {
518 		rc = opal_pci_eeh_freeze_clear(phb->opal_id,
519 					     slave->pe_number,
520 					     opt);
521 		if (rc != OPAL_SUCCESS) {
522 			pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
523 				__func__, rc, opt, phb->hose->global_number,
524 				slave->pe_number);
525 			return -EIO;
526 		}
527 	}
528 
529 	return 0;
530 }
531 
532 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
533 {
534 	struct pnv_ioda_pe *slave, *pe;
535 	u8 fstate, state;
536 	__be16 pcierr;
537 	s64 rc;
538 
539 	/* Sanity check on PE number */
540 	if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
541 		return OPAL_EEH_STOPPED_PERM_UNAVAIL;
542 
543 	/*
544 	 * Fetch the master PE and the PE instance might be
545 	 * not initialized yet.
546 	 */
547 	pe = &phb->ioda.pe_array[pe_no];
548 	if (pe->flags & PNV_IODA_PE_SLAVE) {
549 		pe = pe->master;
550 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
551 		pe_no = pe->pe_number;
552 	}
553 
554 	/* Check the master PE */
555 	rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
556 					&state, &pcierr, NULL);
557 	if (rc != OPAL_SUCCESS) {
558 		pr_warn("%s: Failure %lld getting "
559 			"PHB#%x-PE#%x state\n",
560 			__func__, rc,
561 			phb->hose->global_number, pe_no);
562 		return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
563 	}
564 
565 	/* Check the slave PE */
566 	if (!(pe->flags & PNV_IODA_PE_MASTER))
567 		return state;
568 
569 	list_for_each_entry(slave, &pe->slaves, list) {
570 		rc = opal_pci_eeh_freeze_status(phb->opal_id,
571 						slave->pe_number,
572 						&fstate,
573 						&pcierr,
574 						NULL);
575 		if (rc != OPAL_SUCCESS) {
576 			pr_warn("%s: Failure %lld getting "
577 				"PHB#%x-PE#%x state\n",
578 				__func__, rc,
579 				phb->hose->global_number, slave->pe_number);
580 			return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
581 		}
582 
583 		/*
584 		 * Override the result based on the ascending
585 		 * priority.
586 		 */
587 		if (fstate > state)
588 			state = fstate;
589 	}
590 
591 	return state;
592 }
593 
594 /* Currently those 2 are only used when MSIs are enabled, this will change
595  * but in the meantime, we need to protect them to avoid warnings
596  */
597 #ifdef CONFIG_PCI_MSI
598 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
599 {
600 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
601 	struct pnv_phb *phb = hose->private_data;
602 	struct pci_dn *pdn = pci_get_pdn(dev);
603 
604 	if (!pdn)
605 		return NULL;
606 	if (pdn->pe_number == IODA_INVALID_PE)
607 		return NULL;
608 	return &phb->ioda.pe_array[pdn->pe_number];
609 }
610 #endif /* CONFIG_PCI_MSI */
611 
612 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
613 				  struct pnv_ioda_pe *parent,
614 				  struct pnv_ioda_pe *child,
615 				  bool is_add)
616 {
617 	const char *desc = is_add ? "adding" : "removing";
618 	uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
619 			      OPAL_REMOVE_PE_FROM_DOMAIN;
620 	struct pnv_ioda_pe *slave;
621 	long rc;
622 
623 	/* Parent PE affects child PE */
624 	rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
625 				child->pe_number, op);
626 	if (rc != OPAL_SUCCESS) {
627 		pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
628 			rc, desc);
629 		return -ENXIO;
630 	}
631 
632 	if (!(child->flags & PNV_IODA_PE_MASTER))
633 		return 0;
634 
635 	/* Compound case: parent PE affects slave PEs */
636 	list_for_each_entry(slave, &child->slaves, list) {
637 		rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
638 					slave->pe_number, op);
639 		if (rc != OPAL_SUCCESS) {
640 			pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
641 				rc, desc);
642 			return -ENXIO;
643 		}
644 	}
645 
646 	return 0;
647 }
648 
649 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
650 			      struct pnv_ioda_pe *pe,
651 			      bool is_add)
652 {
653 	struct pnv_ioda_pe *slave;
654 	struct pci_dev *pdev = NULL;
655 	int ret;
656 
657 	/*
658 	 * Clear PE frozen state. If it's master PE, we need
659 	 * clear slave PE frozen state as well.
660 	 */
661 	if (is_add) {
662 		opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
663 					  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
664 		if (pe->flags & PNV_IODA_PE_MASTER) {
665 			list_for_each_entry(slave, &pe->slaves, list)
666 				opal_pci_eeh_freeze_clear(phb->opal_id,
667 							  slave->pe_number,
668 							  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
669 		}
670 	}
671 
672 	/*
673 	 * Associate PE in PELT. We need add the PE into the
674 	 * corresponding PELT-V as well. Otherwise, the error
675 	 * originated from the PE might contribute to other
676 	 * PEs.
677 	 */
678 	ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
679 	if (ret)
680 		return ret;
681 
682 	/* For compound PEs, any one affects all of them */
683 	if (pe->flags & PNV_IODA_PE_MASTER) {
684 		list_for_each_entry(slave, &pe->slaves, list) {
685 			ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
686 			if (ret)
687 				return ret;
688 		}
689 	}
690 
691 	if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
692 		pdev = pe->pbus->self;
693 	else if (pe->flags & PNV_IODA_PE_DEV)
694 		pdev = pe->pdev->bus->self;
695 #ifdef CONFIG_PCI_IOV
696 	else if (pe->flags & PNV_IODA_PE_VF)
697 		pdev = pe->parent_dev;
698 #endif /* CONFIG_PCI_IOV */
699 	while (pdev) {
700 		struct pci_dn *pdn = pci_get_pdn(pdev);
701 		struct pnv_ioda_pe *parent;
702 
703 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
704 			parent = &phb->ioda.pe_array[pdn->pe_number];
705 			ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
706 			if (ret)
707 				return ret;
708 		}
709 
710 		pdev = pdev->bus->self;
711 	}
712 
713 	return 0;
714 }
715 
716 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
717 {
718 	struct pci_dev *parent;
719 	uint8_t bcomp, dcomp, fcomp;
720 	int64_t rc;
721 	long rid_end, rid;
722 
723 	/* Currently, we just deconfigure VF PE. Bus PE will always there.*/
724 	if (pe->pbus) {
725 		int count;
726 
727 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
728 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
729 		parent = pe->pbus->self;
730 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
731 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
732 		else
733 			count = 1;
734 
735 		switch(count) {
736 		case  1: bcomp = OpalPciBusAll;         break;
737 		case  2: bcomp = OpalPciBus7Bits;       break;
738 		case  4: bcomp = OpalPciBus6Bits;       break;
739 		case  8: bcomp = OpalPciBus5Bits;       break;
740 		case 16: bcomp = OpalPciBus4Bits;       break;
741 		case 32: bcomp = OpalPciBus3Bits;       break;
742 		default:
743 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
744 			        count);
745 			/* Do an exact match only */
746 			bcomp = OpalPciBusAll;
747 		}
748 		rid_end = pe->rid + (count << 8);
749 	} else {
750 #ifdef CONFIG_PCI_IOV
751 		if (pe->flags & PNV_IODA_PE_VF)
752 			parent = pe->parent_dev;
753 		else
754 #endif
755 			parent = pe->pdev->bus->self;
756 		bcomp = OpalPciBusAll;
757 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
758 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
759 		rid_end = pe->rid + 1;
760 	}
761 
762 	/* Clear the reverse map */
763 	for (rid = pe->rid; rid < rid_end; rid++)
764 		phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
765 
766 	/* Release from all parents PELT-V */
767 	while (parent) {
768 		struct pci_dn *pdn = pci_get_pdn(parent);
769 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
770 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
771 						pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
772 			/* XXX What to do in case of error ? */
773 		}
774 		parent = parent->bus->self;
775 	}
776 
777 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
778 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
779 
780 	/* Disassociate PE in PELT */
781 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
782 				pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
783 	if (rc)
784 		pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
785 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
786 			     bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
787 	if (rc)
788 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
789 
790 	pe->pbus = NULL;
791 	pe->pdev = NULL;
792 #ifdef CONFIG_PCI_IOV
793 	pe->parent_dev = NULL;
794 #endif
795 
796 	return 0;
797 }
798 
799 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
800 {
801 	struct pci_dev *parent;
802 	uint8_t bcomp, dcomp, fcomp;
803 	long rc, rid_end, rid;
804 
805 	/* Bus validation ? */
806 	if (pe->pbus) {
807 		int count;
808 
809 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
810 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
811 		parent = pe->pbus->self;
812 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
813 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
814 		else
815 			count = 1;
816 
817 		switch(count) {
818 		case  1: bcomp = OpalPciBusAll;		break;
819 		case  2: bcomp = OpalPciBus7Bits;	break;
820 		case  4: bcomp = OpalPciBus6Bits;	break;
821 		case  8: bcomp = OpalPciBus5Bits;	break;
822 		case 16: bcomp = OpalPciBus4Bits;	break;
823 		case 32: bcomp = OpalPciBus3Bits;	break;
824 		default:
825 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
826 			        count);
827 			/* Do an exact match only */
828 			bcomp = OpalPciBusAll;
829 		}
830 		rid_end = pe->rid + (count << 8);
831 	} else {
832 #ifdef CONFIG_PCI_IOV
833 		if (pe->flags & PNV_IODA_PE_VF)
834 			parent = pe->parent_dev;
835 		else
836 #endif /* CONFIG_PCI_IOV */
837 			parent = pe->pdev->bus->self;
838 		bcomp = OpalPciBusAll;
839 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
840 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
841 		rid_end = pe->rid + 1;
842 	}
843 
844 	/*
845 	 * Associate PE in PELT. We need add the PE into the
846 	 * corresponding PELT-V as well. Otherwise, the error
847 	 * originated from the PE might contribute to other
848 	 * PEs.
849 	 */
850 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
851 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
852 	if (rc) {
853 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
854 		return -ENXIO;
855 	}
856 
857 	/*
858 	 * Configure PELTV. NPUs don't have a PELTV table so skip
859 	 * configuration on them.
860 	 */
861 	if (phb->type != PNV_PHB_NPU)
862 		pnv_ioda_set_peltv(phb, pe, true);
863 
864 	/* Setup reverse map */
865 	for (rid = pe->rid; rid < rid_end; rid++)
866 		phb->ioda.pe_rmap[rid] = pe->pe_number;
867 
868 	/* Setup one MVTs on IODA1 */
869 	if (phb->type != PNV_PHB_IODA1) {
870 		pe->mve_number = 0;
871 		goto out;
872 	}
873 
874 	pe->mve_number = pe->pe_number;
875 	rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
876 	if (rc != OPAL_SUCCESS) {
877 		pe_err(pe, "OPAL error %ld setting up MVE %d\n",
878 		       rc, pe->mve_number);
879 		pe->mve_number = -1;
880 	} else {
881 		rc = opal_pci_set_mve_enable(phb->opal_id,
882 					     pe->mve_number, OPAL_ENABLE_MVE);
883 		if (rc) {
884 			pe_err(pe, "OPAL error %ld enabling MVE %d\n",
885 			       rc, pe->mve_number);
886 			pe->mve_number = -1;
887 		}
888 	}
889 
890 out:
891 	return 0;
892 }
893 
894 #ifdef CONFIG_PCI_IOV
895 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
896 {
897 	struct pci_dn *pdn = pci_get_pdn(dev);
898 	int i;
899 	struct resource *res, res2;
900 	resource_size_t size;
901 	u16 num_vfs;
902 
903 	if (!dev->is_physfn)
904 		return -EINVAL;
905 
906 	/*
907 	 * "offset" is in VFs.  The M64 windows are sized so that when they
908 	 * are segmented, each segment is the same size as the IOV BAR.
909 	 * Each segment is in a separate PE, and the high order bits of the
910 	 * address are the PE number.  Therefore, each VF's BAR is in a
911 	 * separate PE, and changing the IOV BAR start address changes the
912 	 * range of PEs the VFs are in.
913 	 */
914 	num_vfs = pdn->num_vfs;
915 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
916 		res = &dev->resource[i + PCI_IOV_RESOURCES];
917 		if (!res->flags || !res->parent)
918 			continue;
919 
920 		/*
921 		 * The actual IOV BAR range is determined by the start address
922 		 * and the actual size for num_vfs VFs BAR.  This check is to
923 		 * make sure that after shifting, the range will not overlap
924 		 * with another device.
925 		 */
926 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
927 		res2.flags = res->flags;
928 		res2.start = res->start + (size * offset);
929 		res2.end = res2.start + (size * num_vfs) - 1;
930 
931 		if (res2.end > res->end) {
932 			dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
933 				i, &res2, res, num_vfs, offset);
934 			return -EBUSY;
935 		}
936 	}
937 
938 	/*
939 	 * After doing so, there would be a "hole" in the /proc/iomem when
940 	 * offset is a positive value. It looks like the device return some
941 	 * mmio back to the system, which actually no one could use it.
942 	 */
943 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
944 		res = &dev->resource[i + PCI_IOV_RESOURCES];
945 		if (!res->flags || !res->parent)
946 			continue;
947 
948 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
949 		res2 = *res;
950 		res->start += size * offset;
951 
952 		dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
953 			 i, &res2, res, (offset > 0) ? "En" : "Dis",
954 			 num_vfs, offset);
955 		pci_update_resource(dev, i + PCI_IOV_RESOURCES);
956 	}
957 	return 0;
958 }
959 #endif /* CONFIG_PCI_IOV */
960 
961 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
962 {
963 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
964 	struct pnv_phb *phb = hose->private_data;
965 	struct pci_dn *pdn = pci_get_pdn(dev);
966 	struct pnv_ioda_pe *pe;
967 
968 	if (!pdn) {
969 		pr_err("%s: Device tree node not associated properly\n",
970 			   pci_name(dev));
971 		return NULL;
972 	}
973 	if (pdn->pe_number != IODA_INVALID_PE)
974 		return NULL;
975 
976 	pe = pnv_ioda_alloc_pe(phb);
977 	if (!pe) {
978 		pr_warning("%s: Not enough PE# available, disabling device\n",
979 			   pci_name(dev));
980 		return NULL;
981 	}
982 
983 	/* NOTE: We get only one ref to the pci_dev for the pdn, not for the
984 	 * pointer in the PE data structure, both should be destroyed at the
985 	 * same time. However, this needs to be looked at more closely again
986 	 * once we actually start removing things (Hotplug, SR-IOV, ...)
987 	 *
988 	 * At some point we want to remove the PDN completely anyways
989 	 */
990 	pci_dev_get(dev);
991 	pdn->pcidev = dev;
992 	pdn->pe_number = pe->pe_number;
993 	pe->flags = PNV_IODA_PE_DEV;
994 	pe->pdev = dev;
995 	pe->pbus = NULL;
996 	pe->mve_number = -1;
997 	pe->rid = dev->bus->number << 8 | pdn->devfn;
998 
999 	pe_info(pe, "Associated device to PE\n");
1000 
1001 	if (pnv_ioda_configure_pe(phb, pe)) {
1002 		/* XXX What do we do here ? */
1003 		pnv_ioda_free_pe(pe);
1004 		pdn->pe_number = IODA_INVALID_PE;
1005 		pe->pdev = NULL;
1006 		pci_dev_put(dev);
1007 		return NULL;
1008 	}
1009 
1010 	/* Put PE to the list */
1011 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1012 
1013 	return pe;
1014 }
1015 
1016 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1017 {
1018 	struct pci_dev *dev;
1019 
1020 	list_for_each_entry(dev, &bus->devices, bus_list) {
1021 		struct pci_dn *pdn = pci_get_pdn(dev);
1022 
1023 		if (pdn == NULL) {
1024 			pr_warn("%s: No device node associated with device !\n",
1025 				pci_name(dev));
1026 			continue;
1027 		}
1028 
1029 		/*
1030 		 * In partial hotplug case, the PCI device might be still
1031 		 * associated with the PE and needn't attach it to the PE
1032 		 * again.
1033 		 */
1034 		if (pdn->pe_number != IODA_INVALID_PE)
1035 			continue;
1036 
1037 		pe->device_count++;
1038 		pdn->pcidev = dev;
1039 		pdn->pe_number = pe->pe_number;
1040 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1041 			pnv_ioda_setup_same_PE(dev->subordinate, pe);
1042 	}
1043 }
1044 
1045 /*
1046  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1047  * single PCI bus. Another one that contains the primary PCI bus and its
1048  * subordinate PCI devices and buses. The second type of PE is normally
1049  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1050  */
1051 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1052 {
1053 	struct pci_controller *hose = pci_bus_to_host(bus);
1054 	struct pnv_phb *phb = hose->private_data;
1055 	struct pnv_ioda_pe *pe = NULL;
1056 	unsigned int pe_num;
1057 
1058 	/*
1059 	 * In partial hotplug case, the PE instance might be still alive.
1060 	 * We should reuse it instead of allocating a new one.
1061 	 */
1062 	pe_num = phb->ioda.pe_rmap[bus->number << 8];
1063 	if (pe_num != IODA_INVALID_PE) {
1064 		pe = &phb->ioda.pe_array[pe_num];
1065 		pnv_ioda_setup_same_PE(bus, pe);
1066 		return NULL;
1067 	}
1068 
1069 	/* PE number for root bus should have been reserved */
1070 	if (pci_is_root_bus(bus) &&
1071 	    phb->ioda.root_pe_idx != IODA_INVALID_PE)
1072 		pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1073 
1074 	/* Check if PE is determined by M64 */
1075 	if (!pe && phb->pick_m64_pe)
1076 		pe = phb->pick_m64_pe(bus, all);
1077 
1078 	/* The PE number isn't pinned by M64 */
1079 	if (!pe)
1080 		pe = pnv_ioda_alloc_pe(phb);
1081 
1082 	if (!pe) {
1083 		pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1084 			__func__, pci_domain_nr(bus), bus->number);
1085 		return NULL;
1086 	}
1087 
1088 	pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1089 	pe->pbus = bus;
1090 	pe->pdev = NULL;
1091 	pe->mve_number = -1;
1092 	pe->rid = bus->busn_res.start << 8;
1093 
1094 	if (all)
1095 		pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
1096 			bus->busn_res.start, bus->busn_res.end, pe->pe_number);
1097 	else
1098 		pe_info(pe, "Secondary bus %d associated with PE#%d\n",
1099 			bus->busn_res.start, pe->pe_number);
1100 
1101 	if (pnv_ioda_configure_pe(phb, pe)) {
1102 		/* XXX What do we do here ? */
1103 		pnv_ioda_free_pe(pe);
1104 		pe->pbus = NULL;
1105 		return NULL;
1106 	}
1107 
1108 	/* Associate it with all child devices */
1109 	pnv_ioda_setup_same_PE(bus, pe);
1110 
1111 	/* Put PE to the list */
1112 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1113 
1114 	return pe;
1115 }
1116 
1117 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1118 {
1119 	int pe_num, found_pe = false, rc;
1120 	long rid;
1121 	struct pnv_ioda_pe *pe;
1122 	struct pci_dev *gpu_pdev;
1123 	struct pci_dn *npu_pdn;
1124 	struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1125 	struct pnv_phb *phb = hose->private_data;
1126 
1127 	/*
1128 	 * Due to a hardware errata PE#0 on the NPU is reserved for
1129 	 * error handling. This means we only have three PEs remaining
1130 	 * which need to be assigned to four links, implying some
1131 	 * links must share PEs.
1132 	 *
1133 	 * To achieve this we assign PEs such that NPUs linking the
1134 	 * same GPU get assigned the same PE.
1135 	 */
1136 	gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1137 	for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1138 		pe = &phb->ioda.pe_array[pe_num];
1139 		if (!pe->pdev)
1140 			continue;
1141 
1142 		if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1143 			/*
1144 			 * This device has the same peer GPU so should
1145 			 * be assigned the same PE as the existing
1146 			 * peer NPU.
1147 			 */
1148 			dev_info(&npu_pdev->dev,
1149 				"Associating to existing PE %d\n", pe_num);
1150 			pci_dev_get(npu_pdev);
1151 			npu_pdn = pci_get_pdn(npu_pdev);
1152 			rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1153 			npu_pdn->pcidev = npu_pdev;
1154 			npu_pdn->pe_number = pe_num;
1155 			phb->ioda.pe_rmap[rid] = pe->pe_number;
1156 
1157 			/* Map the PE to this link */
1158 			rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1159 					OpalPciBusAll,
1160 					OPAL_COMPARE_RID_DEVICE_NUMBER,
1161 					OPAL_COMPARE_RID_FUNCTION_NUMBER,
1162 					OPAL_MAP_PE);
1163 			WARN_ON(rc != OPAL_SUCCESS);
1164 			found_pe = true;
1165 			break;
1166 		}
1167 	}
1168 
1169 	if (!found_pe)
1170 		/*
1171 		 * Could not find an existing PE so allocate a new
1172 		 * one.
1173 		 */
1174 		return pnv_ioda_setup_dev_PE(npu_pdev);
1175 	else
1176 		return pe;
1177 }
1178 
1179 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1180 {
1181 	struct pci_dev *pdev;
1182 
1183 	list_for_each_entry(pdev, &bus->devices, bus_list)
1184 		pnv_ioda_setup_npu_PE(pdev);
1185 }
1186 
1187 static void pnv_pci_ioda_setup_PEs(void)
1188 {
1189 	struct pci_controller *hose, *tmp;
1190 	struct pnv_phb *phb;
1191 
1192 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1193 		phb = hose->private_data;
1194 		if (phb->type == PNV_PHB_NPU) {
1195 			/* PE#0 is needed for error reporting */
1196 			pnv_ioda_reserve_pe(phb, 0);
1197 			pnv_ioda_setup_npu_PEs(hose->bus);
1198 		}
1199 	}
1200 }
1201 
1202 #ifdef CONFIG_PCI_IOV
1203 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1204 {
1205 	struct pci_bus        *bus;
1206 	struct pci_controller *hose;
1207 	struct pnv_phb        *phb;
1208 	struct pci_dn         *pdn;
1209 	int                    i, j;
1210 	int                    m64_bars;
1211 
1212 	bus = pdev->bus;
1213 	hose = pci_bus_to_host(bus);
1214 	phb = hose->private_data;
1215 	pdn = pci_get_pdn(pdev);
1216 
1217 	if (pdn->m64_single_mode)
1218 		m64_bars = num_vfs;
1219 	else
1220 		m64_bars = 1;
1221 
1222 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1223 		for (j = 0; j < m64_bars; j++) {
1224 			if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1225 				continue;
1226 			opal_pci_phb_mmio_enable(phb->opal_id,
1227 				OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1228 			clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1229 			pdn->m64_map[j][i] = IODA_INVALID_M64;
1230 		}
1231 
1232 	kfree(pdn->m64_map);
1233 	return 0;
1234 }
1235 
1236 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1237 {
1238 	struct pci_bus        *bus;
1239 	struct pci_controller *hose;
1240 	struct pnv_phb        *phb;
1241 	struct pci_dn         *pdn;
1242 	unsigned int           win;
1243 	struct resource       *res;
1244 	int                    i, j;
1245 	int64_t                rc;
1246 	int                    total_vfs;
1247 	resource_size_t        size, start;
1248 	int                    pe_num;
1249 	int                    m64_bars;
1250 
1251 	bus = pdev->bus;
1252 	hose = pci_bus_to_host(bus);
1253 	phb = hose->private_data;
1254 	pdn = pci_get_pdn(pdev);
1255 	total_vfs = pci_sriov_get_totalvfs(pdev);
1256 
1257 	if (pdn->m64_single_mode)
1258 		m64_bars = num_vfs;
1259 	else
1260 		m64_bars = 1;
1261 
1262 	pdn->m64_map = kmalloc(sizeof(*pdn->m64_map) * m64_bars, GFP_KERNEL);
1263 	if (!pdn->m64_map)
1264 		return -ENOMEM;
1265 	/* Initialize the m64_map to IODA_INVALID_M64 */
1266 	for (i = 0; i < m64_bars ; i++)
1267 		for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1268 			pdn->m64_map[i][j] = IODA_INVALID_M64;
1269 
1270 
1271 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1272 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
1273 		if (!res->flags || !res->parent)
1274 			continue;
1275 
1276 		for (j = 0; j < m64_bars; j++) {
1277 			do {
1278 				win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1279 						phb->ioda.m64_bar_idx + 1, 0);
1280 
1281 				if (win >= phb->ioda.m64_bar_idx + 1)
1282 					goto m64_failed;
1283 			} while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1284 
1285 			pdn->m64_map[j][i] = win;
1286 
1287 			if (pdn->m64_single_mode) {
1288 				size = pci_iov_resource_size(pdev,
1289 							PCI_IOV_RESOURCES + i);
1290 				start = res->start + size * j;
1291 			} else {
1292 				size = resource_size(res);
1293 				start = res->start;
1294 			}
1295 
1296 			/* Map the M64 here */
1297 			if (pdn->m64_single_mode) {
1298 				pe_num = pdn->pe_num_map[j];
1299 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1300 						pe_num, OPAL_M64_WINDOW_TYPE,
1301 						pdn->m64_map[j][i], 0);
1302 			}
1303 
1304 			rc = opal_pci_set_phb_mem_window(phb->opal_id,
1305 						 OPAL_M64_WINDOW_TYPE,
1306 						 pdn->m64_map[j][i],
1307 						 start,
1308 						 0, /* unused */
1309 						 size);
1310 
1311 
1312 			if (rc != OPAL_SUCCESS) {
1313 				dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1314 					win, rc);
1315 				goto m64_failed;
1316 			}
1317 
1318 			if (pdn->m64_single_mode)
1319 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1320 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1321 			else
1322 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1323 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1324 
1325 			if (rc != OPAL_SUCCESS) {
1326 				dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1327 					win, rc);
1328 				goto m64_failed;
1329 			}
1330 		}
1331 	}
1332 	return 0;
1333 
1334 m64_failed:
1335 	pnv_pci_vf_release_m64(pdev, num_vfs);
1336 	return -EBUSY;
1337 }
1338 
1339 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1340 		int num);
1341 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
1342 
1343 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1344 {
1345 	struct iommu_table    *tbl;
1346 	int64_t               rc;
1347 
1348 	tbl = pe->table_group.tables[0];
1349 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1350 	if (rc)
1351 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1352 
1353 	pnv_pci_ioda2_set_bypass(pe, false);
1354 	if (pe->table_group.group) {
1355 		iommu_group_put(pe->table_group.group);
1356 		BUG_ON(pe->table_group.group);
1357 	}
1358 	pnv_pci_ioda2_table_free_pages(tbl);
1359 	iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
1360 }
1361 
1362 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1363 {
1364 	struct pci_bus        *bus;
1365 	struct pci_controller *hose;
1366 	struct pnv_phb        *phb;
1367 	struct pnv_ioda_pe    *pe, *pe_n;
1368 	struct pci_dn         *pdn;
1369 
1370 	bus = pdev->bus;
1371 	hose = pci_bus_to_host(bus);
1372 	phb = hose->private_data;
1373 	pdn = pci_get_pdn(pdev);
1374 
1375 	if (!pdev->is_physfn)
1376 		return;
1377 
1378 	list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1379 		if (pe->parent_dev != pdev)
1380 			continue;
1381 
1382 		pnv_pci_ioda2_release_dma_pe(pdev, pe);
1383 
1384 		/* Remove from list */
1385 		mutex_lock(&phb->ioda.pe_list_mutex);
1386 		list_del(&pe->list);
1387 		mutex_unlock(&phb->ioda.pe_list_mutex);
1388 
1389 		pnv_ioda_deconfigure_pe(phb, pe);
1390 
1391 		pnv_ioda_free_pe(pe);
1392 	}
1393 }
1394 
1395 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1396 {
1397 	struct pci_bus        *bus;
1398 	struct pci_controller *hose;
1399 	struct pnv_phb        *phb;
1400 	struct pnv_ioda_pe    *pe;
1401 	struct pci_dn         *pdn;
1402 	struct pci_sriov      *iov;
1403 	u16                    num_vfs, i;
1404 
1405 	bus = pdev->bus;
1406 	hose = pci_bus_to_host(bus);
1407 	phb = hose->private_data;
1408 	pdn = pci_get_pdn(pdev);
1409 	iov = pdev->sriov;
1410 	num_vfs = pdn->num_vfs;
1411 
1412 	/* Release VF PEs */
1413 	pnv_ioda_release_vf_PE(pdev);
1414 
1415 	if (phb->type == PNV_PHB_IODA2) {
1416 		if (!pdn->m64_single_mode)
1417 			pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1418 
1419 		/* Release M64 windows */
1420 		pnv_pci_vf_release_m64(pdev, num_vfs);
1421 
1422 		/* Release PE numbers */
1423 		if (pdn->m64_single_mode) {
1424 			for (i = 0; i < num_vfs; i++) {
1425 				if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1426 					continue;
1427 
1428 				pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1429 				pnv_ioda_free_pe(pe);
1430 			}
1431 		} else
1432 			bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1433 		/* Releasing pe_num_map */
1434 		kfree(pdn->pe_num_map);
1435 	}
1436 }
1437 
1438 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1439 				       struct pnv_ioda_pe *pe);
1440 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1441 {
1442 	struct pci_bus        *bus;
1443 	struct pci_controller *hose;
1444 	struct pnv_phb        *phb;
1445 	struct pnv_ioda_pe    *pe;
1446 	int                    pe_num;
1447 	u16                    vf_index;
1448 	struct pci_dn         *pdn;
1449 
1450 	bus = pdev->bus;
1451 	hose = pci_bus_to_host(bus);
1452 	phb = hose->private_data;
1453 	pdn = pci_get_pdn(pdev);
1454 
1455 	if (!pdev->is_physfn)
1456 		return;
1457 
1458 	/* Reserve PE for each VF */
1459 	for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1460 		if (pdn->m64_single_mode)
1461 			pe_num = pdn->pe_num_map[vf_index];
1462 		else
1463 			pe_num = *pdn->pe_num_map + vf_index;
1464 
1465 		pe = &phb->ioda.pe_array[pe_num];
1466 		pe->pe_number = pe_num;
1467 		pe->phb = phb;
1468 		pe->flags = PNV_IODA_PE_VF;
1469 		pe->pbus = NULL;
1470 		pe->parent_dev = pdev;
1471 		pe->mve_number = -1;
1472 		pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1473 			   pci_iov_virtfn_devfn(pdev, vf_index);
1474 
1475 		pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%d\n",
1476 			hose->global_number, pdev->bus->number,
1477 			PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1478 			PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1479 
1480 		if (pnv_ioda_configure_pe(phb, pe)) {
1481 			/* XXX What do we do here ? */
1482 			pnv_ioda_free_pe(pe);
1483 			pe->pdev = NULL;
1484 			continue;
1485 		}
1486 
1487 		/* Put PE to the list */
1488 		mutex_lock(&phb->ioda.pe_list_mutex);
1489 		list_add_tail(&pe->list, &phb->ioda.pe_list);
1490 		mutex_unlock(&phb->ioda.pe_list_mutex);
1491 
1492 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
1493 	}
1494 }
1495 
1496 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1497 {
1498 	struct pci_bus        *bus;
1499 	struct pci_controller *hose;
1500 	struct pnv_phb        *phb;
1501 	struct pnv_ioda_pe    *pe;
1502 	struct pci_dn         *pdn;
1503 	int                    ret;
1504 	u16                    i;
1505 
1506 	bus = pdev->bus;
1507 	hose = pci_bus_to_host(bus);
1508 	phb = hose->private_data;
1509 	pdn = pci_get_pdn(pdev);
1510 
1511 	if (phb->type == PNV_PHB_IODA2) {
1512 		if (!pdn->vfs_expanded) {
1513 			dev_info(&pdev->dev, "don't support this SRIOV device"
1514 				" with non 64bit-prefetchable IOV BAR\n");
1515 			return -ENOSPC;
1516 		}
1517 
1518 		/*
1519 		 * When M64 BARs functions in Single PE mode, the number of VFs
1520 		 * could be enabled must be less than the number of M64 BARs.
1521 		 */
1522 		if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1523 			dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1524 			return -EBUSY;
1525 		}
1526 
1527 		/* Allocating pe_num_map */
1528 		if (pdn->m64_single_mode)
1529 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map) * num_vfs,
1530 					GFP_KERNEL);
1531 		else
1532 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1533 
1534 		if (!pdn->pe_num_map)
1535 			return -ENOMEM;
1536 
1537 		if (pdn->m64_single_mode)
1538 			for (i = 0; i < num_vfs; i++)
1539 				pdn->pe_num_map[i] = IODA_INVALID_PE;
1540 
1541 		/* Calculate available PE for required VFs */
1542 		if (pdn->m64_single_mode) {
1543 			for (i = 0; i < num_vfs; i++) {
1544 				pe = pnv_ioda_alloc_pe(phb);
1545 				if (!pe) {
1546 					ret = -EBUSY;
1547 					goto m64_failed;
1548 				}
1549 
1550 				pdn->pe_num_map[i] = pe->pe_number;
1551 			}
1552 		} else {
1553 			mutex_lock(&phb->ioda.pe_alloc_mutex);
1554 			*pdn->pe_num_map = bitmap_find_next_zero_area(
1555 				phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1556 				0, num_vfs, 0);
1557 			if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1558 				mutex_unlock(&phb->ioda.pe_alloc_mutex);
1559 				dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1560 				kfree(pdn->pe_num_map);
1561 				return -EBUSY;
1562 			}
1563 			bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1564 			mutex_unlock(&phb->ioda.pe_alloc_mutex);
1565 		}
1566 		pdn->num_vfs = num_vfs;
1567 
1568 		/* Assign M64 window accordingly */
1569 		ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1570 		if (ret) {
1571 			dev_info(&pdev->dev, "Not enough M64 window resources\n");
1572 			goto m64_failed;
1573 		}
1574 
1575 		/*
1576 		 * When using one M64 BAR to map one IOV BAR, we need to shift
1577 		 * the IOV BAR according to the PE# allocated to the VFs.
1578 		 * Otherwise, the PE# for the VF will conflict with others.
1579 		 */
1580 		if (!pdn->m64_single_mode) {
1581 			ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1582 			if (ret)
1583 				goto m64_failed;
1584 		}
1585 	}
1586 
1587 	/* Setup VF PEs */
1588 	pnv_ioda_setup_vf_PE(pdev, num_vfs);
1589 
1590 	return 0;
1591 
1592 m64_failed:
1593 	if (pdn->m64_single_mode) {
1594 		for (i = 0; i < num_vfs; i++) {
1595 			if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1596 				continue;
1597 
1598 			pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1599 			pnv_ioda_free_pe(pe);
1600 		}
1601 	} else
1602 		bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1603 
1604 	/* Releasing pe_num_map */
1605 	kfree(pdn->pe_num_map);
1606 
1607 	return ret;
1608 }
1609 
1610 int pcibios_sriov_disable(struct pci_dev *pdev)
1611 {
1612 	pnv_pci_sriov_disable(pdev);
1613 
1614 	/* Release PCI data */
1615 	remove_dev_pci_data(pdev);
1616 	return 0;
1617 }
1618 
1619 int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1620 {
1621 	/* Allocate PCI data */
1622 	add_dev_pci_data(pdev);
1623 
1624 	return pnv_pci_sriov_enable(pdev, num_vfs);
1625 }
1626 #endif /* CONFIG_PCI_IOV */
1627 
1628 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1629 {
1630 	struct pci_dn *pdn = pci_get_pdn(pdev);
1631 	struct pnv_ioda_pe *pe;
1632 
1633 	/*
1634 	 * The function can be called while the PE#
1635 	 * hasn't been assigned. Do nothing for the
1636 	 * case.
1637 	 */
1638 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1639 		return;
1640 
1641 	pe = &phb->ioda.pe_array[pdn->pe_number];
1642 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1643 	set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1644 	set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1645 	/*
1646 	 * Note: iommu_add_device() will fail here as
1647 	 * for physical PE: the device is already added by now;
1648 	 * for virtual PE: sysfs entries are not ready yet and
1649 	 * tce_iommu_bus_notifier will add the device to a group later.
1650 	 */
1651 }
1652 
1653 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1654 {
1655 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1656 	struct pnv_phb *phb = hose->private_data;
1657 	struct pci_dn *pdn = pci_get_pdn(pdev);
1658 	struct pnv_ioda_pe *pe;
1659 	uint64_t top;
1660 	bool bypass = false;
1661 
1662 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1663 		return -ENODEV;;
1664 
1665 	pe = &phb->ioda.pe_array[pdn->pe_number];
1666 	if (pe->tce_bypass_enabled) {
1667 		top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1668 		bypass = (dma_mask >= top);
1669 	}
1670 
1671 	if (bypass) {
1672 		dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1673 		set_dma_ops(&pdev->dev, &dma_direct_ops);
1674 	} else {
1675 		dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1676 		set_dma_ops(&pdev->dev, &dma_iommu_ops);
1677 	}
1678 	*pdev->dev.dma_mask = dma_mask;
1679 
1680 	/* Update peer npu devices */
1681 	pnv_npu_try_dma_set_bypass(pdev, bypass);
1682 
1683 	return 0;
1684 }
1685 
1686 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev)
1687 {
1688 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1689 	struct pnv_phb *phb = hose->private_data;
1690 	struct pci_dn *pdn = pci_get_pdn(pdev);
1691 	struct pnv_ioda_pe *pe;
1692 	u64 end, mask;
1693 
1694 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1695 		return 0;
1696 
1697 	pe = &phb->ioda.pe_array[pdn->pe_number];
1698 	if (!pe->tce_bypass_enabled)
1699 		return __dma_get_required_mask(&pdev->dev);
1700 
1701 
1702 	end = pe->tce_bypass_base + memblock_end_of_DRAM();
1703 	mask = 1ULL << (fls64(end) - 1);
1704 	mask += mask - 1;
1705 
1706 	return mask;
1707 }
1708 
1709 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1710 				   struct pci_bus *bus)
1711 {
1712 	struct pci_dev *dev;
1713 
1714 	list_for_each_entry(dev, &bus->devices, bus_list) {
1715 		set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1716 		set_dma_offset(&dev->dev, pe->tce_bypass_base);
1717 		iommu_add_device(&dev->dev);
1718 
1719 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1720 			pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1721 	}
1722 }
1723 
1724 static void pnv_pci_ioda1_tce_invalidate(struct iommu_table *tbl,
1725 		unsigned long index, unsigned long npages, bool rm)
1726 {
1727 	struct iommu_table_group_link *tgl = list_first_entry_or_null(
1728 			&tbl->it_group_list, struct iommu_table_group_link,
1729 			next);
1730 	struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1731 			struct pnv_ioda_pe, table_group);
1732 	__be64 __iomem *invalidate = rm ?
1733 		(__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys :
1734 		pe->phb->ioda.tce_inval_reg;
1735 	unsigned long start, end, inc;
1736 	const unsigned shift = tbl->it_page_shift;
1737 
1738 	start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1739 	end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1740 			npages - 1);
1741 
1742 	/* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
1743 	if (tbl->it_busno) {
1744 		start <<= shift;
1745 		end <<= shift;
1746 		inc = 128ull << shift;
1747 		start |= tbl->it_busno;
1748 		end |= tbl->it_busno;
1749 	} else if (tbl->it_type & TCE_PCI_SWINV_PAIR) {
1750 		/* p7ioc-style invalidation, 2 TCEs per write */
1751 		start |= (1ull << 63);
1752 		end |= (1ull << 63);
1753 		inc = 16;
1754         } else {
1755 		/* Default (older HW) */
1756                 inc = 128;
1757 	}
1758 
1759         end |= inc - 1;	/* round up end to be different than start */
1760 
1761         mb(); /* Ensure above stores are visible */
1762         while (start <= end) {
1763 		if (rm)
1764 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1765 		else
1766 			__raw_writeq(cpu_to_be64(start), invalidate);
1767                 start += inc;
1768         }
1769 
1770 	/*
1771 	 * The iommu layer will do another mb() for us on build()
1772 	 * and we don't care on free()
1773 	 */
1774 }
1775 
1776 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1777 		long npages, unsigned long uaddr,
1778 		enum dma_data_direction direction,
1779 		struct dma_attrs *attrs)
1780 {
1781 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1782 			attrs);
1783 
1784 	if (!ret && (tbl->it_type & TCE_PCI_SWINV_CREATE))
1785 		pnv_pci_ioda1_tce_invalidate(tbl, index, npages, false);
1786 
1787 	return ret;
1788 }
1789 
1790 #ifdef CONFIG_IOMMU_API
1791 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
1792 		unsigned long *hpa, enum dma_data_direction *direction)
1793 {
1794 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1795 
1796 	if (!ret && (tbl->it_type &
1797 			(TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE)))
1798 		pnv_pci_ioda1_tce_invalidate(tbl, index, 1, false);
1799 
1800 	return ret;
1801 }
1802 #endif
1803 
1804 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1805 		long npages)
1806 {
1807 	pnv_tce_free(tbl, index, npages);
1808 
1809 	if (tbl->it_type & TCE_PCI_SWINV_FREE)
1810 		pnv_pci_ioda1_tce_invalidate(tbl, index, npages, false);
1811 }
1812 
1813 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1814 	.set = pnv_ioda1_tce_build,
1815 #ifdef CONFIG_IOMMU_API
1816 	.exchange = pnv_ioda1_tce_xchg,
1817 #endif
1818 	.clear = pnv_ioda1_tce_free,
1819 	.get = pnv_tce_get,
1820 };
1821 
1822 #define TCE_KILL_INVAL_ALL  PPC_BIT(0)
1823 #define TCE_KILL_INVAL_PE   PPC_BIT(1)
1824 #define TCE_KILL_INVAL_TCE  PPC_BIT(2)
1825 
1826 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
1827 {
1828 	const unsigned long val = TCE_KILL_INVAL_ALL;
1829 
1830 	mb(); /* Ensure previous TCE table stores are visible */
1831 	if (rm)
1832 		__raw_rm_writeq(cpu_to_be64(val),
1833 				(__be64 __iomem *)
1834 				phb->ioda.tce_inval_reg_phys);
1835 	else
1836 		__raw_writeq(cpu_to_be64(val), phb->ioda.tce_inval_reg);
1837 }
1838 
1839 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1840 {
1841 	/* 01xb - invalidate TCEs that match the specified PE# */
1842 	unsigned long val = TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1843 	struct pnv_phb *phb = pe->phb;
1844 
1845 	if (!phb->ioda.tce_inval_reg)
1846 		return;
1847 
1848 	mb(); /* Ensure above stores are visible */
1849 	__raw_writeq(cpu_to_be64(val), phb->ioda.tce_inval_reg);
1850 }
1851 
1852 static void pnv_pci_ioda2_do_tce_invalidate(unsigned pe_number, bool rm,
1853 		__be64 __iomem *invalidate, unsigned shift,
1854 		unsigned long index, unsigned long npages)
1855 {
1856 	unsigned long start, end, inc;
1857 
1858 	/* We'll invalidate DMA address in PE scope */
1859 	start = TCE_KILL_INVAL_TCE;
1860 	start |= (pe_number & 0xFF);
1861 	end = start;
1862 
1863 	/* Figure out the start, end and step */
1864 	start |= (index << shift);
1865 	end |= ((index + npages - 1) << shift);
1866 	inc = (0x1ull << shift);
1867 	mb();
1868 
1869 	while (start <= end) {
1870 		if (rm)
1871 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1872 		else
1873 			__raw_writeq(cpu_to_be64(start), invalidate);
1874 		start += inc;
1875 	}
1876 }
1877 
1878 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1879 		unsigned long index, unsigned long npages, bool rm)
1880 {
1881 	struct iommu_table_group_link *tgl;
1882 
1883 	list_for_each_entry_rcu(tgl, &tbl->it_group_list, next) {
1884 		struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1885 				struct pnv_ioda_pe, table_group);
1886 		__be64 __iomem *invalidate = rm ?
1887 			(__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys :
1888 			pe->phb->ioda.tce_inval_reg;
1889 
1890 		if (pe->phb->type == PNV_PHB_NPU) {
1891 			/*
1892 			 * The NVLink hardware does not support TCE kill
1893 			 * per TCE entry so we have to invalidate
1894 			 * the entire cache for it.
1895 			 */
1896 			pnv_pci_ioda2_tce_invalidate_entire(pe->phb, rm);
1897 			continue;
1898 		}
1899 		pnv_pci_ioda2_do_tce_invalidate(pe->pe_number, rm,
1900 			invalidate, tbl->it_page_shift,
1901 			index, npages);
1902 	}
1903 }
1904 
1905 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1906 		long npages, unsigned long uaddr,
1907 		enum dma_data_direction direction,
1908 		struct dma_attrs *attrs)
1909 {
1910 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1911 			attrs);
1912 
1913 	if (!ret && (tbl->it_type & TCE_PCI_SWINV_CREATE))
1914 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1915 
1916 	return ret;
1917 }
1918 
1919 #ifdef CONFIG_IOMMU_API
1920 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
1921 		unsigned long *hpa, enum dma_data_direction *direction)
1922 {
1923 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1924 
1925 	if (!ret && (tbl->it_type &
1926 			(TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE)))
1927 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
1928 
1929 	return ret;
1930 }
1931 #endif
1932 
1933 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1934 		long npages)
1935 {
1936 	pnv_tce_free(tbl, index, npages);
1937 
1938 	if (tbl->it_type & TCE_PCI_SWINV_FREE)
1939 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1940 }
1941 
1942 static void pnv_ioda2_table_free(struct iommu_table *tbl)
1943 {
1944 	pnv_pci_ioda2_table_free_pages(tbl);
1945 	iommu_free_table(tbl, "pnv");
1946 }
1947 
1948 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
1949 	.set = pnv_ioda2_tce_build,
1950 #ifdef CONFIG_IOMMU_API
1951 	.exchange = pnv_ioda2_tce_xchg,
1952 #endif
1953 	.clear = pnv_ioda2_tce_free,
1954 	.get = pnv_tce_get,
1955 	.free = pnv_ioda2_table_free,
1956 };
1957 
1958 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
1959 {
1960 	unsigned int *weight = (unsigned int *)data;
1961 
1962 	/* This is quite simplistic. The "base" weight of a device
1963 	 * is 10. 0 means no DMA is to be accounted for it.
1964 	 */
1965 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1966 		return 0;
1967 
1968 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
1969 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
1970 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
1971 		*weight += 3;
1972 	else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
1973 		*weight += 15;
1974 	else
1975 		*weight += 10;
1976 
1977 	return 0;
1978 }
1979 
1980 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
1981 {
1982 	unsigned int weight = 0;
1983 
1984 	/* SRIOV VF has same DMA32 weight as its PF */
1985 #ifdef CONFIG_PCI_IOV
1986 	if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
1987 		pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
1988 		return weight;
1989 	}
1990 #endif
1991 
1992 	if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
1993 		pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
1994 	} else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
1995 		struct pci_dev *pdev;
1996 
1997 		list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
1998 			pnv_pci_ioda_dev_dma_weight(pdev, &weight);
1999 	} else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2000 		pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2001 	}
2002 
2003 	return weight;
2004 }
2005 
2006 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2007 				       struct pnv_ioda_pe *pe)
2008 {
2009 
2010 	struct page *tce_mem = NULL;
2011 	struct iommu_table *tbl;
2012 	unsigned int weight, total_weight = 0;
2013 	unsigned int tce32_segsz, base, segs, avail, i;
2014 	int64_t rc;
2015 	void *addr;
2016 
2017 	/* XXX FIXME: Handle 64-bit only DMA devices */
2018 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2019 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
2020 	weight = pnv_pci_ioda_pe_dma_weight(pe);
2021 	if (!weight)
2022 		return;
2023 
2024 	pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2025 		     &total_weight);
2026 	segs = (weight * phb->ioda.dma32_count) / total_weight;
2027 	if (!segs)
2028 		segs = 1;
2029 
2030 	/*
2031 	 * Allocate contiguous DMA32 segments. We begin with the expected
2032 	 * number of segments. With one more attempt, the number of DMA32
2033 	 * segments to be allocated is decreased by one until one segment
2034 	 * is allocated successfully.
2035 	 */
2036 	do {
2037 		for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2038 			for (avail = 0, i = base; i < base + segs; i++) {
2039 				if (phb->ioda.dma32_segmap[i] ==
2040 				    IODA_INVALID_PE)
2041 					avail++;
2042 			}
2043 
2044 			if (avail == segs)
2045 				goto found;
2046 		}
2047 	} while (--segs);
2048 
2049 	if (!segs) {
2050 		pe_warn(pe, "No available DMA32 segments\n");
2051 		return;
2052 	}
2053 
2054 found:
2055 	tbl = pnv_pci_table_alloc(phb->hose->node);
2056 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2057 			pe->pe_number);
2058 	pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2059 
2060 	/* Grab a 32-bit TCE table */
2061 	pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2062 		weight, total_weight, base, segs);
2063 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2064 		base * PNV_IODA1_DMA32_SEGSIZE,
2065 		(base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2066 
2067 	/* XXX Currently, we allocate one big contiguous table for the
2068 	 * TCEs. We only really need one chunk per 256M of TCE space
2069 	 * (ie per segment) but that's an optimization for later, it
2070 	 * requires some added smarts with our get/put_tce implementation
2071 	 *
2072 	 * Each TCE page is 4KB in size and each TCE entry occupies 8
2073 	 * bytes
2074 	 */
2075 	tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2076 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2077 				   get_order(tce32_segsz * segs));
2078 	if (!tce_mem) {
2079 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2080 		goto fail;
2081 	}
2082 	addr = page_address(tce_mem);
2083 	memset(addr, 0, tce32_segsz * segs);
2084 
2085 	/* Configure HW */
2086 	for (i = 0; i < segs; i++) {
2087 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
2088 					      pe->pe_number,
2089 					      base + i, 1,
2090 					      __pa(addr) + tce32_segsz * i,
2091 					      tce32_segsz, IOMMU_PAGE_SIZE_4K);
2092 		if (rc) {
2093 			pe_err(pe, " Failed to configure 32-bit TCE table,"
2094 			       " err %ld\n", rc);
2095 			goto fail;
2096 		}
2097 	}
2098 
2099 	/* Setup DMA32 segment mapping */
2100 	for (i = base; i < base + segs; i++)
2101 		phb->ioda.dma32_segmap[i] = pe->pe_number;
2102 
2103 	/* Setup linux iommu table */
2104 	pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2105 				  base * PNV_IODA1_DMA32_SEGSIZE,
2106 				  IOMMU_PAGE_SHIFT_4K);
2107 
2108 	/* OPAL variant of P7IOC SW invalidated TCEs */
2109 	if (phb->ioda.tce_inval_reg)
2110 		tbl->it_type |= (TCE_PCI_SWINV_CREATE |
2111 				 TCE_PCI_SWINV_FREE   |
2112 				 TCE_PCI_SWINV_PAIR);
2113 
2114 	tbl->it_ops = &pnv_ioda1_iommu_ops;
2115 	pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2116 	pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2117 	iommu_init_table(tbl, phb->hose->node);
2118 
2119 	if (pe->flags & PNV_IODA_PE_DEV) {
2120 		/*
2121 		 * Setting table base here only for carrying iommu_group
2122 		 * further down to let iommu_add_device() do the job.
2123 		 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2124 		 */
2125 		set_iommu_table_base(&pe->pdev->dev, tbl);
2126 		iommu_add_device(&pe->pdev->dev);
2127 	} else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2128 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2129 
2130 	return;
2131  fail:
2132 	/* XXX Failure: Try to fallback to 64-bit only ? */
2133 	if (tce_mem)
2134 		__free_pages(tce_mem, get_order(tce32_segsz * segs));
2135 	if (tbl) {
2136 		pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2137 		iommu_free_table(tbl, "pnv");
2138 	}
2139 }
2140 
2141 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2142 		int num, struct iommu_table *tbl)
2143 {
2144 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2145 			table_group);
2146 	struct pnv_phb *phb = pe->phb;
2147 	int64_t rc;
2148 	const unsigned long size = tbl->it_indirect_levels ?
2149 			tbl->it_level_size : tbl->it_size;
2150 	const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2151 	const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2152 
2153 	pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num,
2154 			start_addr, start_addr + win_size - 1,
2155 			IOMMU_PAGE_SIZE(tbl));
2156 
2157 	/*
2158 	 * Map TCE table through TVT. The TVE index is the PE number
2159 	 * shifted by 1 bit for 32-bits DMA space.
2160 	 */
2161 	rc = opal_pci_map_pe_dma_window(phb->opal_id,
2162 			pe->pe_number,
2163 			(pe->pe_number << 1) + num,
2164 			tbl->it_indirect_levels + 1,
2165 			__pa(tbl->it_base),
2166 			size << 3,
2167 			IOMMU_PAGE_SIZE(tbl));
2168 	if (rc) {
2169 		pe_err(pe, "Failed to configure TCE table, err %ld\n", rc);
2170 		return rc;
2171 	}
2172 
2173 	pnv_pci_link_table_and_group(phb->hose->node, num,
2174 			tbl, &pe->table_group);
2175 	pnv_pci_ioda2_tce_invalidate_pe(pe);
2176 
2177 	return 0;
2178 }
2179 
2180 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2181 {
2182 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
2183 	int64_t rc;
2184 
2185 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2186 	if (enable) {
2187 		phys_addr_t top = memblock_end_of_DRAM();
2188 
2189 		top = roundup_pow_of_two(top);
2190 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2191 						     pe->pe_number,
2192 						     window_id,
2193 						     pe->tce_bypass_base,
2194 						     top);
2195 	} else {
2196 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2197 						     pe->pe_number,
2198 						     window_id,
2199 						     pe->tce_bypass_base,
2200 						     0);
2201 	}
2202 	if (rc)
2203 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2204 	else
2205 		pe->tce_bypass_enabled = enable;
2206 }
2207 
2208 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2209 		__u32 page_shift, __u64 window_size, __u32 levels,
2210 		struct iommu_table *tbl);
2211 
2212 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2213 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
2214 		struct iommu_table **ptbl)
2215 {
2216 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2217 			table_group);
2218 	int nid = pe->phb->hose->node;
2219 	__u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2220 	long ret;
2221 	struct iommu_table *tbl;
2222 
2223 	tbl = pnv_pci_table_alloc(nid);
2224 	if (!tbl)
2225 		return -ENOMEM;
2226 
2227 	ret = pnv_pci_ioda2_table_alloc_pages(nid,
2228 			bus_offset, page_shift, window_size,
2229 			levels, tbl);
2230 	if (ret) {
2231 		iommu_free_table(tbl, "pnv");
2232 		return ret;
2233 	}
2234 
2235 	tbl->it_ops = &pnv_ioda2_iommu_ops;
2236 	if (pe->phb->ioda.tce_inval_reg)
2237 		tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
2238 
2239 	*ptbl = tbl;
2240 
2241 	return 0;
2242 }
2243 
2244 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2245 {
2246 	struct iommu_table *tbl = NULL;
2247 	long rc;
2248 
2249 	/*
2250 	 * crashkernel= specifies the kdump kernel's maximum memory at
2251 	 * some offset and there is no guaranteed the result is a power
2252 	 * of 2, which will cause errors later.
2253 	 */
2254 	const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2255 
2256 	/*
2257 	 * In memory constrained environments, e.g. kdump kernel, the
2258 	 * DMA window can be larger than available memory, which will
2259 	 * cause errors later.
2260 	 */
2261 	const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2262 
2263 	rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2264 			IOMMU_PAGE_SHIFT_4K,
2265 			window_size,
2266 			POWERNV_IOMMU_DEFAULT_LEVELS, &tbl);
2267 	if (rc) {
2268 		pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2269 				rc);
2270 		return rc;
2271 	}
2272 
2273 	iommu_init_table(tbl, pe->phb->hose->node);
2274 
2275 	rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2276 	if (rc) {
2277 		pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2278 				rc);
2279 		pnv_ioda2_table_free(tbl);
2280 		return rc;
2281 	}
2282 
2283 	if (!pnv_iommu_bypass_disabled)
2284 		pnv_pci_ioda2_set_bypass(pe, true);
2285 
2286 	/* OPAL variant of PHB3 invalidated TCEs */
2287 	if (pe->phb->ioda.tce_inval_reg)
2288 		tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
2289 
2290 	/*
2291 	 * Setting table base here only for carrying iommu_group
2292 	 * further down to let iommu_add_device() do the job.
2293 	 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2294 	 */
2295 	if (pe->flags & PNV_IODA_PE_DEV)
2296 		set_iommu_table_base(&pe->pdev->dev, tbl);
2297 
2298 	return 0;
2299 }
2300 
2301 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2302 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2303 		int num)
2304 {
2305 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2306 			table_group);
2307 	struct pnv_phb *phb = pe->phb;
2308 	long ret;
2309 
2310 	pe_info(pe, "Removing DMA window #%d\n", num);
2311 
2312 	ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2313 			(pe->pe_number << 1) + num,
2314 			0/* levels */, 0/* table address */,
2315 			0/* table size */, 0/* page size */);
2316 	if (ret)
2317 		pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2318 	else
2319 		pnv_pci_ioda2_tce_invalidate_pe(pe);
2320 
2321 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2322 
2323 	return ret;
2324 }
2325 #endif
2326 
2327 #ifdef CONFIG_IOMMU_API
2328 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2329 		__u64 window_size, __u32 levels)
2330 {
2331 	unsigned long bytes = 0;
2332 	const unsigned window_shift = ilog2(window_size);
2333 	unsigned entries_shift = window_shift - page_shift;
2334 	unsigned table_shift = entries_shift + 3;
2335 	unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2336 	unsigned long direct_table_size;
2337 
2338 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2339 			(window_size > memory_hotplug_max()) ||
2340 			!is_power_of_2(window_size))
2341 		return 0;
2342 
2343 	/* Calculate a direct table size from window_size and levels */
2344 	entries_shift = (entries_shift + levels - 1) / levels;
2345 	table_shift = entries_shift + 3;
2346 	table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2347 	direct_table_size =  1UL << table_shift;
2348 
2349 	for ( ; levels; --levels) {
2350 		bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2351 
2352 		tce_table_size /= direct_table_size;
2353 		tce_table_size <<= 3;
2354 		tce_table_size = _ALIGN_UP(tce_table_size, direct_table_size);
2355 	}
2356 
2357 	return bytes;
2358 }
2359 
2360 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2361 {
2362 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2363 						table_group);
2364 	/* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2365 	struct iommu_table *tbl = pe->table_group.tables[0];
2366 
2367 	pnv_pci_ioda2_set_bypass(pe, false);
2368 	pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2369 	pnv_ioda2_table_free(tbl);
2370 }
2371 
2372 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2373 {
2374 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2375 						table_group);
2376 
2377 	pnv_pci_ioda2_setup_default_config(pe);
2378 }
2379 
2380 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2381 	.get_table_size = pnv_pci_ioda2_get_table_size,
2382 	.create_table = pnv_pci_ioda2_create_table,
2383 	.set_window = pnv_pci_ioda2_set_window,
2384 	.unset_window = pnv_pci_ioda2_unset_window,
2385 	.take_ownership = pnv_ioda2_take_ownership,
2386 	.release_ownership = pnv_ioda2_release_ownership,
2387 };
2388 
2389 static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque)
2390 {
2391 	struct pci_controller *hose;
2392 	struct pnv_phb *phb;
2393 	struct pnv_ioda_pe **ptmppe = opaque;
2394 	struct pci_dev *pdev = container_of(dev, struct pci_dev, dev);
2395 	struct pci_dn *pdn = pci_get_pdn(pdev);
2396 
2397 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2398 		return 0;
2399 
2400 	hose = pci_bus_to_host(pdev->bus);
2401 	phb = hose->private_data;
2402 	if (phb->type != PNV_PHB_NPU)
2403 		return 0;
2404 
2405 	*ptmppe = &phb->ioda.pe_array[pdn->pe_number];
2406 
2407 	return 1;
2408 }
2409 
2410 /*
2411  * This returns PE of associated NPU.
2412  * This assumes that NPU is in the same IOMMU group with GPU and there is
2413  * no other PEs.
2414  */
2415 static struct pnv_ioda_pe *gpe_table_group_to_npe(
2416 		struct iommu_table_group *table_group)
2417 {
2418 	struct pnv_ioda_pe *npe = NULL;
2419 	int ret = iommu_group_for_each_dev(table_group->group, &npe,
2420 			gpe_table_group_to_npe_cb);
2421 
2422 	BUG_ON(!ret || !npe);
2423 
2424 	return npe;
2425 }
2426 
2427 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group *table_group,
2428 		int num, struct iommu_table *tbl)
2429 {
2430 	long ret = pnv_pci_ioda2_set_window(table_group, num, tbl);
2431 
2432 	if (ret)
2433 		return ret;
2434 
2435 	ret = pnv_npu_set_window(gpe_table_group_to_npe(table_group), num, tbl);
2436 	if (ret)
2437 		pnv_pci_ioda2_unset_window(table_group, num);
2438 
2439 	return ret;
2440 }
2441 
2442 static long pnv_pci_ioda2_npu_unset_window(
2443 		struct iommu_table_group *table_group,
2444 		int num)
2445 {
2446 	long ret = pnv_pci_ioda2_unset_window(table_group, num);
2447 
2448 	if (ret)
2449 		return ret;
2450 
2451 	return pnv_npu_unset_window(gpe_table_group_to_npe(table_group), num);
2452 }
2453 
2454 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group *table_group)
2455 {
2456 	/*
2457 	 * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2458 	 * the iommu_table if 32bit DMA is enabled.
2459 	 */
2460 	pnv_npu_take_ownership(gpe_table_group_to_npe(table_group));
2461 	pnv_ioda2_take_ownership(table_group);
2462 }
2463 
2464 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = {
2465 	.get_table_size = pnv_pci_ioda2_get_table_size,
2466 	.create_table = pnv_pci_ioda2_create_table,
2467 	.set_window = pnv_pci_ioda2_npu_set_window,
2468 	.unset_window = pnv_pci_ioda2_npu_unset_window,
2469 	.take_ownership = pnv_ioda2_npu_take_ownership,
2470 	.release_ownership = pnv_ioda2_release_ownership,
2471 };
2472 
2473 static void pnv_pci_ioda_setup_iommu_api(void)
2474 {
2475 	struct pci_controller *hose, *tmp;
2476 	struct pnv_phb *phb;
2477 	struct pnv_ioda_pe *pe, *gpe;
2478 
2479 	/*
2480 	 * Now we have all PHBs discovered, time to add NPU devices to
2481 	 * the corresponding IOMMU groups.
2482 	 */
2483 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2484 		phb = hose->private_data;
2485 
2486 		if (phb->type != PNV_PHB_NPU)
2487 			continue;
2488 
2489 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2490 			gpe = pnv_pci_npu_setup_iommu(pe);
2491 			if (gpe)
2492 				gpe->table_group.ops = &pnv_pci_ioda2_npu_ops;
2493 		}
2494 	}
2495 }
2496 #else /* !CONFIG_IOMMU_API */
2497 static void pnv_pci_ioda_setup_iommu_api(void) { };
2498 #endif
2499 
2500 static void pnv_pci_ioda_setup_opal_tce_kill(struct pnv_phb *phb)
2501 {
2502 	const __be64 *swinvp;
2503 
2504 	/* OPAL variant of PHB3 invalidated TCEs */
2505 	swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
2506 	if (!swinvp)
2507 		return;
2508 
2509 	phb->ioda.tce_inval_reg_phys = be64_to_cpup(swinvp);
2510 	phb->ioda.tce_inval_reg = ioremap(phb->ioda.tce_inval_reg_phys, 8);
2511 }
2512 
2513 static __be64 *pnv_pci_ioda2_table_do_alloc_pages(int nid, unsigned shift,
2514 		unsigned levels, unsigned long limit,
2515 		unsigned long *current_offset, unsigned long *total_allocated)
2516 {
2517 	struct page *tce_mem = NULL;
2518 	__be64 *addr, *tmp;
2519 	unsigned order = max_t(unsigned, shift, PAGE_SHIFT) - PAGE_SHIFT;
2520 	unsigned long allocated = 1UL << (order + PAGE_SHIFT);
2521 	unsigned entries = 1UL << (shift - 3);
2522 	long i;
2523 
2524 	tce_mem = alloc_pages_node(nid, GFP_KERNEL, order);
2525 	if (!tce_mem) {
2526 		pr_err("Failed to allocate a TCE memory, order=%d\n", order);
2527 		return NULL;
2528 	}
2529 	addr = page_address(tce_mem);
2530 	memset(addr, 0, allocated);
2531 	*total_allocated += allocated;
2532 
2533 	--levels;
2534 	if (!levels) {
2535 		*current_offset += allocated;
2536 		return addr;
2537 	}
2538 
2539 	for (i = 0; i < entries; ++i) {
2540 		tmp = pnv_pci_ioda2_table_do_alloc_pages(nid, shift,
2541 				levels, limit, current_offset, total_allocated);
2542 		if (!tmp)
2543 			break;
2544 
2545 		addr[i] = cpu_to_be64(__pa(tmp) |
2546 				TCE_PCI_READ | TCE_PCI_WRITE);
2547 
2548 		if (*current_offset >= limit)
2549 			break;
2550 	}
2551 
2552 	return addr;
2553 }
2554 
2555 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2556 		unsigned long size, unsigned level);
2557 
2558 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2559 		__u32 page_shift, __u64 window_size, __u32 levels,
2560 		struct iommu_table *tbl)
2561 {
2562 	void *addr;
2563 	unsigned long offset = 0, level_shift, total_allocated = 0;
2564 	const unsigned window_shift = ilog2(window_size);
2565 	unsigned entries_shift = window_shift - page_shift;
2566 	unsigned table_shift = max_t(unsigned, entries_shift + 3, PAGE_SHIFT);
2567 	const unsigned long tce_table_size = 1UL << table_shift;
2568 
2569 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS))
2570 		return -EINVAL;
2571 
2572 	if ((window_size > memory_hotplug_max()) || !is_power_of_2(window_size))
2573 		return -EINVAL;
2574 
2575 	/* Adjust direct table size from window_size and levels */
2576 	entries_shift = (entries_shift + levels - 1) / levels;
2577 	level_shift = entries_shift + 3;
2578 	level_shift = max_t(unsigned, level_shift, PAGE_SHIFT);
2579 
2580 	/* Allocate TCE table */
2581 	addr = pnv_pci_ioda2_table_do_alloc_pages(nid, level_shift,
2582 			levels, tce_table_size, &offset, &total_allocated);
2583 
2584 	/* addr==NULL means that the first level allocation failed */
2585 	if (!addr)
2586 		return -ENOMEM;
2587 
2588 	/*
2589 	 * First level was allocated but some lower level failed as
2590 	 * we did not allocate as much as we wanted,
2591 	 * release partially allocated table.
2592 	 */
2593 	if (offset < tce_table_size) {
2594 		pnv_pci_ioda2_table_do_free_pages(addr,
2595 				1ULL << (level_shift - 3), levels - 1);
2596 		return -ENOMEM;
2597 	}
2598 
2599 	/* Setup linux iommu table */
2600 	pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, bus_offset,
2601 			page_shift);
2602 	tbl->it_level_size = 1ULL << (level_shift - 3);
2603 	tbl->it_indirect_levels = levels - 1;
2604 	tbl->it_allocated_size = total_allocated;
2605 
2606 	pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2607 			window_size, tce_table_size, bus_offset);
2608 
2609 	return 0;
2610 }
2611 
2612 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2613 		unsigned long size, unsigned level)
2614 {
2615 	const unsigned long addr_ul = (unsigned long) addr &
2616 			~(TCE_PCI_READ | TCE_PCI_WRITE);
2617 
2618 	if (level) {
2619 		long i;
2620 		u64 *tmp = (u64 *) addr_ul;
2621 
2622 		for (i = 0; i < size; ++i) {
2623 			unsigned long hpa = be64_to_cpu(tmp[i]);
2624 
2625 			if (!(hpa & (TCE_PCI_READ | TCE_PCI_WRITE)))
2626 				continue;
2627 
2628 			pnv_pci_ioda2_table_do_free_pages(__va(hpa), size,
2629 					level - 1);
2630 		}
2631 	}
2632 
2633 	free_pages(addr_ul, get_order(size << 3));
2634 }
2635 
2636 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl)
2637 {
2638 	const unsigned long size = tbl->it_indirect_levels ?
2639 			tbl->it_level_size : tbl->it_size;
2640 
2641 	if (!tbl->it_size)
2642 		return;
2643 
2644 	pnv_pci_ioda2_table_do_free_pages((__be64 *)tbl->it_base, size,
2645 			tbl->it_indirect_levels);
2646 }
2647 
2648 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2649 				       struct pnv_ioda_pe *pe)
2650 {
2651 	int64_t rc;
2652 
2653 	if (!pnv_pci_ioda_pe_dma_weight(pe))
2654 		return;
2655 
2656 	/* TVE #1 is selected by PCI address bit 59 */
2657 	pe->tce_bypass_base = 1ull << 59;
2658 
2659 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2660 			pe->pe_number);
2661 
2662 	/* The PE will reserve all possible 32-bits space */
2663 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2664 		phb->ioda.m32_pci_base);
2665 
2666 	/* Setup linux iommu table */
2667 	pe->table_group.tce32_start = 0;
2668 	pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2669 	pe->table_group.max_dynamic_windows_supported =
2670 			IOMMU_TABLE_GROUP_MAX_TABLES;
2671 	pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2672 	pe->table_group.pgsizes = SZ_4K | SZ_64K | SZ_16M;
2673 #ifdef CONFIG_IOMMU_API
2674 	pe->table_group.ops = &pnv_pci_ioda2_ops;
2675 #endif
2676 
2677 	rc = pnv_pci_ioda2_setup_default_config(pe);
2678 	if (rc)
2679 		return;
2680 
2681 	if (pe->flags & PNV_IODA_PE_DEV)
2682 		iommu_add_device(&pe->pdev->dev);
2683 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2684 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2685 }
2686 
2687 #ifdef CONFIG_PCI_MSI
2688 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2689 {
2690 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2691 	struct irq_chip *chip = irq_data_get_irq_chip(d);
2692 	struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2693 					   ioda.irq_chip);
2694 	int64_t rc;
2695 
2696 	rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
2697 	WARN_ON_ONCE(rc);
2698 
2699 	icp_native_eoi(d);
2700 }
2701 
2702 
2703 static void set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2704 {
2705 	struct irq_data *idata;
2706 	struct irq_chip *ichip;
2707 
2708 	if (phb->type != PNV_PHB_IODA2)
2709 		return;
2710 
2711 	if (!phb->ioda.irq_chip_init) {
2712 		/*
2713 		 * First time we setup an MSI IRQ, we need to setup the
2714 		 * corresponding IRQ chip to route correctly.
2715 		 */
2716 		idata = irq_get_irq_data(virq);
2717 		ichip = irq_data_get_irq_chip(idata);
2718 		phb->ioda.irq_chip_init = 1;
2719 		phb->ioda.irq_chip = *ichip;
2720 		phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2721 	}
2722 	irq_set_chip(virq, &phb->ioda.irq_chip);
2723 }
2724 
2725 #ifdef CONFIG_CXL_BASE
2726 
2727 struct device_node *pnv_pci_get_phb_node(struct pci_dev *dev)
2728 {
2729 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2730 
2731 	return of_node_get(hose->dn);
2732 }
2733 EXPORT_SYMBOL(pnv_pci_get_phb_node);
2734 
2735 int pnv_phb_to_cxl_mode(struct pci_dev *dev, uint64_t mode)
2736 {
2737 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2738 	struct pnv_phb *phb = hose->private_data;
2739 	struct pnv_ioda_pe *pe;
2740 	int rc;
2741 
2742 	pe = pnv_ioda_get_pe(dev);
2743 	if (!pe)
2744 		return -ENODEV;
2745 
2746 	pe_info(pe, "Switching PHB to CXL\n");
2747 
2748 	rc = opal_pci_set_phb_cxl_mode(phb->opal_id, mode, pe->pe_number);
2749 	if (rc == OPAL_UNSUPPORTED)
2750 		dev_err(&dev->dev, "Required cxl mode not supported by firmware - update skiboot\n");
2751 	else if (rc)
2752 		dev_err(&dev->dev, "opal_pci_set_phb_cxl_mode failed: %i\n", rc);
2753 
2754 	return rc;
2755 }
2756 EXPORT_SYMBOL(pnv_phb_to_cxl_mode);
2757 
2758 /* Find PHB for cxl dev and allocate MSI hwirqs?
2759  * Returns the absolute hardware IRQ number
2760  */
2761 int pnv_cxl_alloc_hwirqs(struct pci_dev *dev, int num)
2762 {
2763 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2764 	struct pnv_phb *phb = hose->private_data;
2765 	int hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, num);
2766 
2767 	if (hwirq < 0) {
2768 		dev_warn(&dev->dev, "Failed to find a free MSI\n");
2769 		return -ENOSPC;
2770 	}
2771 
2772 	return phb->msi_base + hwirq;
2773 }
2774 EXPORT_SYMBOL(pnv_cxl_alloc_hwirqs);
2775 
2776 void pnv_cxl_release_hwirqs(struct pci_dev *dev, int hwirq, int num)
2777 {
2778 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2779 	struct pnv_phb *phb = hose->private_data;
2780 
2781 	msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq - phb->msi_base, num);
2782 }
2783 EXPORT_SYMBOL(pnv_cxl_release_hwirqs);
2784 
2785 void pnv_cxl_release_hwirq_ranges(struct cxl_irq_ranges *irqs,
2786 				  struct pci_dev *dev)
2787 {
2788 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2789 	struct pnv_phb *phb = hose->private_data;
2790 	int i, hwirq;
2791 
2792 	for (i = 1; i < CXL_IRQ_RANGES; i++) {
2793 		if (!irqs->range[i])
2794 			continue;
2795 		pr_devel("cxl release irq range 0x%x: offset: 0x%lx  limit: %ld\n",
2796 			 i, irqs->offset[i],
2797 			 irqs->range[i]);
2798 		hwirq = irqs->offset[i] - phb->msi_base;
2799 		msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq,
2800 				       irqs->range[i]);
2801 	}
2802 }
2803 EXPORT_SYMBOL(pnv_cxl_release_hwirq_ranges);
2804 
2805 int pnv_cxl_alloc_hwirq_ranges(struct cxl_irq_ranges *irqs,
2806 			       struct pci_dev *dev, int num)
2807 {
2808 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2809 	struct pnv_phb *phb = hose->private_data;
2810 	int i, hwirq, try;
2811 
2812 	memset(irqs, 0, sizeof(struct cxl_irq_ranges));
2813 
2814 	/* 0 is reserved for the multiplexed PSL DSI interrupt */
2815 	for (i = 1; i < CXL_IRQ_RANGES && num; i++) {
2816 		try = num;
2817 		while (try) {
2818 			hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, try);
2819 			if (hwirq >= 0)
2820 				break;
2821 			try /= 2;
2822 		}
2823 		if (!try)
2824 			goto fail;
2825 
2826 		irqs->offset[i] = phb->msi_base + hwirq;
2827 		irqs->range[i] = try;
2828 		pr_devel("cxl alloc irq range 0x%x: offset: 0x%lx  limit: %li\n",
2829 			 i, irqs->offset[i], irqs->range[i]);
2830 		num -= try;
2831 	}
2832 	if (num)
2833 		goto fail;
2834 
2835 	return 0;
2836 fail:
2837 	pnv_cxl_release_hwirq_ranges(irqs, dev);
2838 	return -ENOSPC;
2839 }
2840 EXPORT_SYMBOL(pnv_cxl_alloc_hwirq_ranges);
2841 
2842 int pnv_cxl_get_irq_count(struct pci_dev *dev)
2843 {
2844 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2845 	struct pnv_phb *phb = hose->private_data;
2846 
2847 	return phb->msi_bmp.irq_count;
2848 }
2849 EXPORT_SYMBOL(pnv_cxl_get_irq_count);
2850 
2851 int pnv_cxl_ioda_msi_setup(struct pci_dev *dev, unsigned int hwirq,
2852 			   unsigned int virq)
2853 {
2854 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
2855 	struct pnv_phb *phb = hose->private_data;
2856 	unsigned int xive_num = hwirq - phb->msi_base;
2857 	struct pnv_ioda_pe *pe;
2858 	int rc;
2859 
2860 	if (!(pe = pnv_ioda_get_pe(dev)))
2861 		return -ENODEV;
2862 
2863 	/* Assign XIVE to PE */
2864 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2865 	if (rc) {
2866 		pe_warn(pe, "%s: OPAL error %d setting msi_base 0x%x "
2867 			"hwirq 0x%x XIVE 0x%x PE\n",
2868 			pci_name(dev), rc, phb->msi_base, hwirq, xive_num);
2869 		return -EIO;
2870 	}
2871 	set_msi_irq_chip(phb, virq);
2872 
2873 	return 0;
2874 }
2875 EXPORT_SYMBOL(pnv_cxl_ioda_msi_setup);
2876 #endif
2877 
2878 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2879 				  unsigned int hwirq, unsigned int virq,
2880 				  unsigned int is_64, struct msi_msg *msg)
2881 {
2882 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2883 	unsigned int xive_num = hwirq - phb->msi_base;
2884 	__be32 data;
2885 	int rc;
2886 
2887 	/* No PE assigned ? bail out ... no MSI for you ! */
2888 	if (pe == NULL)
2889 		return -ENXIO;
2890 
2891 	/* Check if we have an MVE */
2892 	if (pe->mve_number < 0)
2893 		return -ENXIO;
2894 
2895 	/* Force 32-bit MSI on some broken devices */
2896 	if (dev->no_64bit_msi)
2897 		is_64 = 0;
2898 
2899 	/* Assign XIVE to PE */
2900 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2901 	if (rc) {
2902 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2903 			pci_name(dev), rc, xive_num);
2904 		return -EIO;
2905 	}
2906 
2907 	if (is_64) {
2908 		__be64 addr64;
2909 
2910 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2911 				     &addr64, &data);
2912 		if (rc) {
2913 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2914 				pci_name(dev), rc);
2915 			return -EIO;
2916 		}
2917 		msg->address_hi = be64_to_cpu(addr64) >> 32;
2918 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2919 	} else {
2920 		__be32 addr32;
2921 
2922 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2923 				     &addr32, &data);
2924 		if (rc) {
2925 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2926 				pci_name(dev), rc);
2927 			return -EIO;
2928 		}
2929 		msg->address_hi = 0;
2930 		msg->address_lo = be32_to_cpu(addr32);
2931 	}
2932 	msg->data = be32_to_cpu(data);
2933 
2934 	set_msi_irq_chip(phb, virq);
2935 
2936 	pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2937 		 " address=%x_%08x data=%x PE# %d\n",
2938 		 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2939 		 msg->address_hi, msg->address_lo, data, pe->pe_number);
2940 
2941 	return 0;
2942 }
2943 
2944 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2945 {
2946 	unsigned int count;
2947 	const __be32 *prop = of_get_property(phb->hose->dn,
2948 					     "ibm,opal-msi-ranges", NULL);
2949 	if (!prop) {
2950 		/* BML Fallback */
2951 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2952 	}
2953 	if (!prop)
2954 		return;
2955 
2956 	phb->msi_base = be32_to_cpup(prop);
2957 	count = be32_to_cpup(prop + 1);
2958 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2959 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2960 		       phb->hose->global_number);
2961 		return;
2962 	}
2963 
2964 	phb->msi_setup = pnv_pci_ioda_msi_setup;
2965 	phb->msi32_support = 1;
2966 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2967 		count, phb->msi_base);
2968 }
2969 #else
2970 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
2971 #endif /* CONFIG_PCI_MSI */
2972 
2973 #ifdef CONFIG_PCI_IOV
2974 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2975 {
2976 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
2977 	struct pnv_phb *phb = hose->private_data;
2978 	const resource_size_t gate = phb->ioda.m64_segsize >> 2;
2979 	struct resource *res;
2980 	int i;
2981 	resource_size_t size, total_vf_bar_sz;
2982 	struct pci_dn *pdn;
2983 	int mul, total_vfs;
2984 
2985 	if (!pdev->is_physfn || pdev->is_added)
2986 		return;
2987 
2988 	pdn = pci_get_pdn(pdev);
2989 	pdn->vfs_expanded = 0;
2990 	pdn->m64_single_mode = false;
2991 
2992 	total_vfs = pci_sriov_get_totalvfs(pdev);
2993 	mul = phb->ioda.total_pe_num;
2994 	total_vf_bar_sz = 0;
2995 
2996 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2997 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2998 		if (!res->flags || res->parent)
2999 			continue;
3000 		if (!pnv_pci_is_mem_pref_64(res->flags)) {
3001 			dev_warn(&pdev->dev, "Don't support SR-IOV with"
3002 					" non M64 VF BAR%d: %pR. \n",
3003 				 i, res);
3004 			goto truncate_iov;
3005 		}
3006 
3007 		total_vf_bar_sz += pci_iov_resource_size(pdev,
3008 				i + PCI_IOV_RESOURCES);
3009 
3010 		/*
3011 		 * If bigger than quarter of M64 segment size, just round up
3012 		 * power of two.
3013 		 *
3014 		 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
3015 		 * with other devices, IOV BAR size is expanded to be
3016 		 * (total_pe * VF_BAR_size).  When VF_BAR_size is half of M64
3017 		 * segment size , the expanded size would equal to half of the
3018 		 * whole M64 space size, which will exhaust the M64 Space and
3019 		 * limit the system flexibility.  This is a design decision to
3020 		 * set the boundary to quarter of the M64 segment size.
3021 		 */
3022 		if (total_vf_bar_sz > gate) {
3023 			mul = roundup_pow_of_two(total_vfs);
3024 			dev_info(&pdev->dev,
3025 				"VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
3026 				total_vf_bar_sz, gate, mul);
3027 			pdn->m64_single_mode = true;
3028 			break;
3029 		}
3030 	}
3031 
3032 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3033 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
3034 		if (!res->flags || res->parent)
3035 			continue;
3036 
3037 		size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
3038 		/*
3039 		 * On PHB3, the minimum size alignment of M64 BAR in single
3040 		 * mode is 32MB.
3041 		 */
3042 		if (pdn->m64_single_mode && (size < SZ_32M))
3043 			goto truncate_iov;
3044 		dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
3045 		res->end = res->start + size * mul - 1;
3046 		dev_dbg(&pdev->dev, "                       %pR\n", res);
3047 		dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
3048 			 i, res, mul);
3049 	}
3050 	pdn->vfs_expanded = mul;
3051 
3052 	return;
3053 
3054 truncate_iov:
3055 	/* To save MMIO space, IOV BAR is truncated. */
3056 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3057 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
3058 		res->flags = 0;
3059 		res->end = res->start - 1;
3060 	}
3061 }
3062 #endif /* CONFIG_PCI_IOV */
3063 
3064 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
3065 				  struct resource *res)
3066 {
3067 	struct pnv_phb *phb = pe->phb;
3068 	struct pci_bus_region region;
3069 	int index;
3070 	int64_t rc;
3071 
3072 	if (!res || !res->flags || res->start > res->end)
3073 		return;
3074 
3075 	if (res->flags & IORESOURCE_IO) {
3076 		region.start = res->start - phb->ioda.io_pci_base;
3077 		region.end   = res->end - phb->ioda.io_pci_base;
3078 		index = region.start / phb->ioda.io_segsize;
3079 
3080 		while (index < phb->ioda.total_pe_num &&
3081 		       region.start <= region.end) {
3082 			phb->ioda.io_segmap[index] = pe->pe_number;
3083 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3084 				pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
3085 			if (rc != OPAL_SUCCESS) {
3086 				pr_err("%s: Error %lld mapping IO segment#%d to PE#%d\n",
3087 				       __func__, rc, index, pe->pe_number);
3088 				break;
3089 			}
3090 
3091 			region.start += phb->ioda.io_segsize;
3092 			index++;
3093 		}
3094 	} else if ((res->flags & IORESOURCE_MEM) &&
3095 		   !pnv_pci_is_mem_pref_64(res->flags)) {
3096 		region.start = res->start -
3097 			       phb->hose->mem_offset[0] -
3098 			       phb->ioda.m32_pci_base;
3099 		region.end   = res->end -
3100 			       phb->hose->mem_offset[0] -
3101 			       phb->ioda.m32_pci_base;
3102 		index = region.start / phb->ioda.m32_segsize;
3103 
3104 		while (index < phb->ioda.total_pe_num &&
3105 		       region.start <= region.end) {
3106 			phb->ioda.m32_segmap[index] = pe->pe_number;
3107 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3108 				pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
3109 			if (rc != OPAL_SUCCESS) {
3110 				pr_err("%s: Error %lld mapping M32 segment#%d to PE#%d",
3111 				       __func__, rc, index, pe->pe_number);
3112 				break;
3113 			}
3114 
3115 			region.start += phb->ioda.m32_segsize;
3116 			index++;
3117 		}
3118 	}
3119 }
3120 
3121 /*
3122  * This function is supposed to be called on basis of PE from top
3123  * to bottom style. So the the I/O or MMIO segment assigned to
3124  * parent PE could be overrided by its child PEs if necessary.
3125  */
3126 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
3127 {
3128 	struct pci_dev *pdev;
3129 	int i;
3130 
3131 	/*
3132 	 * NOTE: We only care PCI bus based PE for now. For PCI
3133 	 * device based PE, for example SRIOV sensitive VF should
3134 	 * be figured out later.
3135 	 */
3136 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3137 
3138 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3139 		for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3140 			pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3141 
3142 		/*
3143 		 * If the PE contains all subordinate PCI buses, the
3144 		 * windows of the child bridges should be mapped to
3145 		 * the PE as well.
3146 		 */
3147 		if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3148 			continue;
3149 		for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3150 			pnv_ioda_setup_pe_res(pe,
3151 				&pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3152 	}
3153 }
3154 
3155 static void pnv_pci_ioda_create_dbgfs(void)
3156 {
3157 #ifdef CONFIG_DEBUG_FS
3158 	struct pci_controller *hose, *tmp;
3159 	struct pnv_phb *phb;
3160 	char name[16];
3161 
3162 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3163 		phb = hose->private_data;
3164 
3165 		/* Notify initialization of PHB done */
3166 		phb->initialized = 1;
3167 
3168 		sprintf(name, "PCI%04x", hose->global_number);
3169 		phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3170 		if (!phb->dbgfs)
3171 			pr_warning("%s: Error on creating debugfs on PHB#%x\n",
3172 				__func__, hose->global_number);
3173 	}
3174 #endif /* CONFIG_DEBUG_FS */
3175 }
3176 
3177 static void pnv_pci_ioda_fixup(void)
3178 {
3179 	pnv_pci_ioda_setup_PEs();
3180 	pnv_pci_ioda_setup_iommu_api();
3181 	pnv_pci_ioda_create_dbgfs();
3182 
3183 #ifdef CONFIG_EEH
3184 	eeh_init();
3185 	eeh_addr_cache_build();
3186 #endif
3187 }
3188 
3189 /*
3190  * Returns the alignment for I/O or memory windows for P2P
3191  * bridges. That actually depends on how PEs are segmented.
3192  * For now, we return I/O or M32 segment size for PE sensitive
3193  * P2P bridges. Otherwise, the default values (4KiB for I/O,
3194  * 1MiB for memory) will be returned.
3195  *
3196  * The current PCI bus might be put into one PE, which was
3197  * create against the parent PCI bridge. For that case, we
3198  * needn't enlarge the alignment so that we can save some
3199  * resources.
3200  */
3201 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3202 						unsigned long type)
3203 {
3204 	struct pci_dev *bridge;
3205 	struct pci_controller *hose = pci_bus_to_host(bus);
3206 	struct pnv_phb *phb = hose->private_data;
3207 	int num_pci_bridges = 0;
3208 
3209 	bridge = bus->self;
3210 	while (bridge) {
3211 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3212 			num_pci_bridges++;
3213 			if (num_pci_bridges >= 2)
3214 				return 1;
3215 		}
3216 
3217 		bridge = bridge->bus->self;
3218 	}
3219 
3220 	/* We fail back to M32 if M64 isn't supported */
3221 	if (phb->ioda.m64_segsize &&
3222 	    pnv_pci_is_mem_pref_64(type))
3223 		return phb->ioda.m64_segsize;
3224 	if (type & IORESOURCE_MEM)
3225 		return phb->ioda.m32_segsize;
3226 
3227 	return phb->ioda.io_segsize;
3228 }
3229 
3230 /*
3231  * We are updating root port or the upstream port of the
3232  * bridge behind the root port with PHB's windows in order
3233  * to accommodate the changes on required resources during
3234  * PCI (slot) hotplug, which is connected to either root
3235  * port or the downstream ports of PCIe switch behind the
3236  * root port.
3237  */
3238 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
3239 					   unsigned long type)
3240 {
3241 	struct pci_controller *hose = pci_bus_to_host(bus);
3242 	struct pnv_phb *phb = hose->private_data;
3243 	struct pci_dev *bridge = bus->self;
3244 	struct resource *r, *w;
3245 	bool msi_region = false;
3246 	int i;
3247 
3248 	/* Check if we need apply fixup to the bridge's windows */
3249 	if (!pci_is_root_bus(bridge->bus) &&
3250 	    !pci_is_root_bus(bridge->bus->self->bus))
3251 		return;
3252 
3253 	/* Fixup the resources */
3254 	for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
3255 		r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
3256 		if (!r->flags || !r->parent)
3257 			continue;
3258 
3259 		w = NULL;
3260 		if (r->flags & type & IORESOURCE_IO)
3261 			w = &hose->io_resource;
3262 		else if (pnv_pci_is_mem_pref_64(r->flags) &&
3263 			 (type & IORESOURCE_PREFETCH) &&
3264 			 phb->ioda.m64_segsize)
3265 			w = &hose->mem_resources[1];
3266 		else if (r->flags & type & IORESOURCE_MEM) {
3267 			w = &hose->mem_resources[0];
3268 			msi_region = true;
3269 		}
3270 
3271 		r->start = w->start;
3272 		r->end = w->end;
3273 
3274 		/* The 64KB 32-bits MSI region shouldn't be included in
3275 		 * the 32-bits bridge window. Otherwise, we can see strange
3276 		 * issues. One of them is EEH error observed on Garrison.
3277 		 *
3278 		 * Exclude top 1MB region which is the minimal alignment of
3279 		 * 32-bits bridge window.
3280 		 */
3281 		if (msi_region) {
3282 			r->end += 0x10000;
3283 			r->end -= 0x100000;
3284 		}
3285 	}
3286 }
3287 
3288 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type)
3289 {
3290 	struct pci_controller *hose = pci_bus_to_host(bus);
3291 	struct pnv_phb *phb = hose->private_data;
3292 	struct pci_dev *bridge = bus->self;
3293 	struct pnv_ioda_pe *pe;
3294 	bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
3295 
3296 	/* Extend bridge's windows if necessary */
3297 	pnv_pci_fixup_bridge_resources(bus, type);
3298 
3299 	/* The PE for root bus should be realized before any one else */
3300 	if (!phb->ioda.root_pe_populated) {
3301 		pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false);
3302 		if (pe) {
3303 			phb->ioda.root_pe_idx = pe->pe_number;
3304 			phb->ioda.root_pe_populated = true;
3305 		}
3306 	}
3307 
3308 	/* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3309 	if (list_empty(&bus->devices))
3310 		return;
3311 
3312 	/* Reserve PEs according to used M64 resources */
3313 	if (phb->reserve_m64_pe)
3314 		phb->reserve_m64_pe(bus, NULL, all);
3315 
3316 	/*
3317 	 * Assign PE. We might run here because of partial hotplug.
3318 	 * For the case, we just pick up the existing PE and should
3319 	 * not allocate resources again.
3320 	 */
3321 	pe = pnv_ioda_setup_bus_PE(bus, all);
3322 	if (!pe)
3323 		return;
3324 
3325 	pnv_ioda_setup_pe_seg(pe);
3326 	switch (phb->type) {
3327 	case PNV_PHB_IODA1:
3328 		pnv_pci_ioda1_setup_dma_pe(phb, pe);
3329 		break;
3330 	case PNV_PHB_IODA2:
3331 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
3332 		break;
3333 	default:
3334 		pr_warn("%s: No DMA for PHB#%d (type %d)\n",
3335 			__func__, phb->hose->global_number, phb->type);
3336 	}
3337 }
3338 
3339 #ifdef CONFIG_PCI_IOV
3340 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3341 						      int resno)
3342 {
3343 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3344 	struct pnv_phb *phb = hose->private_data;
3345 	struct pci_dn *pdn = pci_get_pdn(pdev);
3346 	resource_size_t align;
3347 
3348 	/*
3349 	 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3350 	 * SR-IOV. While from hardware perspective, the range mapped by M64
3351 	 * BAR should be size aligned.
3352 	 *
3353 	 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3354 	 * powernv-specific hardware restriction is gone. But if just use the
3355 	 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3356 	 * in one segment of M64 #15, which introduces the PE conflict between
3357 	 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3358 	 * m64_segsize.
3359 	 *
3360 	 * This function returns the total IOV BAR size if M64 BAR is in
3361 	 * Shared PE mode or just VF BAR size if not.
3362 	 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3363 	 * M64 segment size if IOV BAR size is less.
3364 	 */
3365 	align = pci_iov_resource_size(pdev, resno);
3366 	if (!pdn->vfs_expanded)
3367 		return align;
3368 	if (pdn->m64_single_mode)
3369 		return max(align, (resource_size_t)phb->ioda.m64_segsize);
3370 
3371 	return pdn->vfs_expanded * align;
3372 }
3373 #endif /* CONFIG_PCI_IOV */
3374 
3375 /* Prevent enabling devices for which we couldn't properly
3376  * assign a PE
3377  */
3378 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3379 {
3380 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
3381 	struct pnv_phb *phb = hose->private_data;
3382 	struct pci_dn *pdn;
3383 
3384 	/* The function is probably called while the PEs have
3385 	 * not be created yet. For example, resource reassignment
3386 	 * during PCI probe period. We just skip the check if
3387 	 * PEs isn't ready.
3388 	 */
3389 	if (!phb->initialized)
3390 		return true;
3391 
3392 	pdn = pci_get_pdn(dev);
3393 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3394 		return false;
3395 
3396 	return true;
3397 }
3398 
3399 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
3400 				       int num)
3401 {
3402 	struct pnv_ioda_pe *pe = container_of(table_group,
3403 					      struct pnv_ioda_pe, table_group);
3404 	struct pnv_phb *phb = pe->phb;
3405 	unsigned int idx;
3406 	long rc;
3407 
3408 	pe_info(pe, "Removing DMA window #%d\n", num);
3409 	for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
3410 		if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
3411 			continue;
3412 
3413 		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
3414 						idx, 0, 0ul, 0ul, 0ul);
3415 		if (rc != OPAL_SUCCESS) {
3416 			pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
3417 				rc, idx);
3418 			return rc;
3419 		}
3420 
3421 		phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
3422 	}
3423 
3424 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
3425 	return OPAL_SUCCESS;
3426 }
3427 
3428 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
3429 {
3430 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3431 	struct iommu_table *tbl = pe->table_group.tables[0];
3432 	int64_t rc;
3433 
3434 	if (!weight)
3435 		return;
3436 
3437 	rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
3438 	if (rc != OPAL_SUCCESS)
3439 		return;
3440 
3441 	pnv_pci_ioda1_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
3442 	if (pe->table_group.group) {
3443 		iommu_group_put(pe->table_group.group);
3444 		WARN_ON(pe->table_group.group);
3445 	}
3446 
3447 	free_pages(tbl->it_base, get_order(tbl->it_size << 3));
3448 	iommu_free_table(tbl, "pnv");
3449 }
3450 
3451 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
3452 {
3453 	struct iommu_table *tbl = pe->table_group.tables[0];
3454 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3455 #ifdef CONFIG_IOMMU_API
3456 	int64_t rc;
3457 #endif
3458 
3459 	if (!weight)
3460 		return;
3461 
3462 #ifdef CONFIG_IOMMU_API
3463 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
3464 	if (rc)
3465 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
3466 #endif
3467 
3468 	pnv_pci_ioda2_set_bypass(pe, false);
3469 	if (pe->table_group.group) {
3470 		iommu_group_put(pe->table_group.group);
3471 		WARN_ON(pe->table_group.group);
3472 	}
3473 
3474 	pnv_pci_ioda2_table_free_pages(tbl);
3475 	iommu_free_table(tbl, "pnv");
3476 }
3477 
3478 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
3479 				 unsigned short win,
3480 				 unsigned int *map)
3481 {
3482 	struct pnv_phb *phb = pe->phb;
3483 	int idx;
3484 	int64_t rc;
3485 
3486 	for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
3487 		if (map[idx] != pe->pe_number)
3488 			continue;
3489 
3490 		if (win == OPAL_M64_WINDOW_TYPE)
3491 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3492 					phb->ioda.reserved_pe_idx, win,
3493 					idx / PNV_IODA1_M64_SEGS,
3494 					idx % PNV_IODA1_M64_SEGS);
3495 		else
3496 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3497 					phb->ioda.reserved_pe_idx, win, 0, idx);
3498 
3499 		if (rc != OPAL_SUCCESS)
3500 			pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
3501 				rc, win, idx);
3502 
3503 		map[idx] = IODA_INVALID_PE;
3504 	}
3505 }
3506 
3507 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
3508 {
3509 	struct pnv_phb *phb = pe->phb;
3510 
3511 	if (phb->type == PNV_PHB_IODA1) {
3512 		pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
3513 				     phb->ioda.io_segmap);
3514 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3515 				     phb->ioda.m32_segmap);
3516 		pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE,
3517 				     phb->ioda.m64_segmap);
3518 	} else if (phb->type == PNV_PHB_IODA2) {
3519 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3520 				     phb->ioda.m32_segmap);
3521 	}
3522 }
3523 
3524 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
3525 {
3526 	struct pnv_phb *phb = pe->phb;
3527 	struct pnv_ioda_pe *slave, *tmp;
3528 
3529 	/* Release slave PEs in compound PE */
3530 	if (pe->flags & PNV_IODA_PE_MASTER) {
3531 		list_for_each_entry_safe(slave, tmp, &pe->slaves, list)
3532 			pnv_ioda_release_pe(slave);
3533 	}
3534 
3535 	list_del(&pe->list);
3536 	switch (phb->type) {
3537 	case PNV_PHB_IODA1:
3538 		pnv_pci_ioda1_release_pe_dma(pe);
3539 		break;
3540 	case PNV_PHB_IODA2:
3541 		pnv_pci_ioda2_release_pe_dma(pe);
3542 		break;
3543 	default:
3544 		WARN_ON(1);
3545 	}
3546 
3547 	pnv_ioda_release_pe_seg(pe);
3548 	pnv_ioda_deconfigure_pe(pe->phb, pe);
3549 	pnv_ioda_free_pe(pe);
3550 }
3551 
3552 static void pnv_pci_release_device(struct pci_dev *pdev)
3553 {
3554 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3555 	struct pnv_phb *phb = hose->private_data;
3556 	struct pci_dn *pdn = pci_get_pdn(pdev);
3557 	struct pnv_ioda_pe *pe;
3558 
3559 	if (pdev->is_virtfn)
3560 		return;
3561 
3562 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3563 		return;
3564 
3565 	pe = &phb->ioda.pe_array[pdn->pe_number];
3566 	WARN_ON(--pe->device_count < 0);
3567 	if (pe->device_count == 0)
3568 		pnv_ioda_release_pe(pe);
3569 }
3570 
3571 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3572 {
3573 	struct pnv_phb *phb = hose->private_data;
3574 
3575 	opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3576 		       OPAL_ASSERT_RESET);
3577 }
3578 
3579 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3580 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3581 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3582 #ifdef CONFIG_PCI_MSI
3583 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3584 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3585 #endif
3586 	.enable_device_hook	= pnv_pci_enable_device_hook,
3587 	.release_device		= pnv_pci_release_device,
3588 	.window_alignment	= pnv_pci_window_alignment,
3589 	.setup_bridge		= pnv_pci_setup_bridge,
3590 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3591 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3592 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3593 	.shutdown		= pnv_pci_ioda_shutdown,
3594 };
3595 
3596 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask)
3597 {
3598 	dev_err_once(&npdev->dev,
3599 			"%s operation unsupported for NVLink devices\n",
3600 			__func__);
3601 	return -EPERM;
3602 }
3603 
3604 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3605 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3606 #ifdef CONFIG_PCI_MSI
3607 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3608 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3609 #endif
3610 	.enable_device_hook	= pnv_pci_enable_device_hook,
3611 	.window_alignment	= pnv_pci_window_alignment,
3612 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3613 	.dma_set_mask		= pnv_npu_dma_set_mask,
3614 	.shutdown		= pnv_pci_ioda_shutdown,
3615 };
3616 
3617 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3618 					 u64 hub_id, int ioda_type)
3619 {
3620 	struct pci_controller *hose;
3621 	struct pnv_phb *phb;
3622 	unsigned long size, m64map_off, m32map_off, pemap_off;
3623 	unsigned long iomap_off = 0, dma32map_off = 0;
3624 	const __be64 *prop64;
3625 	const __be32 *prop32;
3626 	int len;
3627 	unsigned int segno;
3628 	u64 phb_id;
3629 	void *aux;
3630 	long rc;
3631 
3632 	pr_info("Initializing %s PHB (%s)\n",
3633 		pnv_phb_names[ioda_type], of_node_full_name(np));
3634 
3635 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3636 	if (!prop64) {
3637 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3638 		return;
3639 	}
3640 	phb_id = be64_to_cpup(prop64);
3641 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3642 
3643 	phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
3644 
3645 	/* Allocate PCI controller */
3646 	phb->hose = hose = pcibios_alloc_controller(np);
3647 	if (!phb->hose) {
3648 		pr_err("  Can't allocate PCI controller for %s\n",
3649 		       np->full_name);
3650 		memblock_free(__pa(phb), sizeof(struct pnv_phb));
3651 		return;
3652 	}
3653 
3654 	spin_lock_init(&phb->lock);
3655 	prop32 = of_get_property(np, "bus-range", &len);
3656 	if (prop32 && len == 8) {
3657 		hose->first_busno = be32_to_cpu(prop32[0]);
3658 		hose->last_busno = be32_to_cpu(prop32[1]);
3659 	} else {
3660 		pr_warn("  Broken <bus-range> on %s\n", np->full_name);
3661 		hose->first_busno = 0;
3662 		hose->last_busno = 0xff;
3663 	}
3664 	hose->private_data = phb;
3665 	phb->hub_id = hub_id;
3666 	phb->opal_id = phb_id;
3667 	phb->type = ioda_type;
3668 	mutex_init(&phb->ioda.pe_alloc_mutex);
3669 
3670 	/* Detect specific models for error handling */
3671 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3672 		phb->model = PNV_PHB_MODEL_P7IOC;
3673 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3674 		phb->model = PNV_PHB_MODEL_PHB3;
3675 	else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3676 		phb->model = PNV_PHB_MODEL_NPU;
3677 	else
3678 		phb->model = PNV_PHB_MODEL_UNKNOWN;
3679 
3680 	/* Parse 32-bit and IO ranges (if any) */
3681 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3682 
3683 	/* Get registers */
3684 	phb->regs = of_iomap(np, 0);
3685 	if (phb->regs == NULL)
3686 		pr_err("  Failed to map registers !\n");
3687 
3688 	/* Initialize TCE kill register */
3689 	pnv_pci_ioda_setup_opal_tce_kill(phb);
3690 
3691 	/* Initialize more IODA stuff */
3692 	phb->ioda.total_pe_num = 1;
3693 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3694 	if (prop32)
3695 		phb->ioda.total_pe_num = be32_to_cpup(prop32);
3696 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3697 	if (prop32)
3698 		phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3699 
3700 	/* Invalidate RID to PE# mapping */
3701 	for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3702 		phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3703 
3704 	/* Parse 64-bit MMIO range */
3705 	pnv_ioda_parse_m64_window(phb);
3706 
3707 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3708 	/* FW Has already off top 64k of M32 space (MSI space) */
3709 	phb->ioda.m32_size += 0x10000;
3710 
3711 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3712 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3713 	phb->ioda.io_size = hose->pci_io_size;
3714 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3715 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3716 
3717 	/* Calculate how many 32-bit TCE segments we have */
3718 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3719 				PNV_IODA1_DMA32_SEGSIZE;
3720 
3721 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3722 	size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3723 			sizeof(unsigned long));
3724 	m64map_off = size;
3725 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3726 	m32map_off = size;
3727 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3728 	if (phb->type == PNV_PHB_IODA1) {
3729 		iomap_off = size;
3730 		size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3731 		dma32map_off = size;
3732 		size += phb->ioda.dma32_count *
3733 			sizeof(phb->ioda.dma32_segmap[0]);
3734 	}
3735 	pemap_off = size;
3736 	size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3737 	aux = memblock_virt_alloc(size, 0);
3738 	phb->ioda.pe_alloc = aux;
3739 	phb->ioda.m64_segmap = aux + m64map_off;
3740 	phb->ioda.m32_segmap = aux + m32map_off;
3741 	for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3742 		phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3743 		phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3744 	}
3745 	if (phb->type == PNV_PHB_IODA1) {
3746 		phb->ioda.io_segmap = aux + iomap_off;
3747 		for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3748 			phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3749 
3750 		phb->ioda.dma32_segmap = aux + dma32map_off;
3751 		for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3752 			phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3753 	}
3754 	phb->ioda.pe_array = aux + pemap_off;
3755 
3756 	/*
3757 	 * Choose PE number for root bus, which shouldn't have
3758 	 * M64 resources consumed by its child devices. To pick
3759 	 * the PE number adjacent to the reserved one if possible.
3760 	 */
3761 	pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3762 	if (phb->ioda.reserved_pe_idx == 0) {
3763 		phb->ioda.root_pe_idx = 1;
3764 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3765 	} else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3766 		phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3767 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3768 	} else {
3769 		phb->ioda.root_pe_idx = IODA_INVALID_PE;
3770 	}
3771 
3772 	INIT_LIST_HEAD(&phb->ioda.pe_list);
3773 	mutex_init(&phb->ioda.pe_list_mutex);
3774 
3775 	/* Calculate how many 32-bit TCE segments we have */
3776 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3777 				PNV_IODA1_DMA32_SEGSIZE;
3778 
3779 #if 0 /* We should really do that ... */
3780 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
3781 					 window_type,
3782 					 window_num,
3783 					 starting_real_address,
3784 					 starting_pci_address,
3785 					 segment_size);
3786 #endif
3787 
3788 	pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3789 		phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3790 		phb->ioda.m32_size, phb->ioda.m32_segsize);
3791 	if (phb->ioda.m64_size)
3792 		pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3793 			phb->ioda.m64_size, phb->ioda.m64_segsize);
3794 	if (phb->ioda.io_size)
3795 		pr_info("                  IO: 0x%x [segment=0x%x]\n",
3796 			phb->ioda.io_size, phb->ioda.io_segsize);
3797 
3798 
3799 	phb->hose->ops = &pnv_pci_ops;
3800 	phb->get_pe_state = pnv_ioda_get_pe_state;
3801 	phb->freeze_pe = pnv_ioda_freeze_pe;
3802 	phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3803 
3804 	/* Setup MSI support */
3805 	pnv_pci_init_ioda_msis(phb);
3806 
3807 	/*
3808 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3809 	 * to let the PCI core do resource assignment. It's supposed
3810 	 * that the PCI core will do correct I/O and MMIO alignment
3811 	 * for the P2P bridge bars so that each PCI bus (excluding
3812 	 * the child P2P bridges) can form individual PE.
3813 	 */
3814 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3815 
3816 	if (phb->type == PNV_PHB_NPU) {
3817 		hose->controller_ops = pnv_npu_ioda_controller_ops;
3818 	} else {
3819 		phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
3820 		hose->controller_ops = pnv_pci_ioda_controller_ops;
3821 	}
3822 
3823 #ifdef CONFIG_PCI_IOV
3824 	ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
3825 	ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3826 #endif
3827 
3828 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3829 
3830 	/* Reset IODA tables to a clean state */
3831 	rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3832 	if (rc)
3833 		pr_warning("  OPAL Error %ld performing IODA table reset !\n", rc);
3834 
3835 	/* If we're running in kdump kerenl, the previous kerenl never
3836 	 * shutdown PCI devices correctly. We already got IODA table
3837 	 * cleaned out. So we have to issue PHB reset to stop all PCI
3838 	 * transactions from previous kerenl.
3839 	 */
3840 	if (is_kdump_kernel()) {
3841 		pr_info("  Issue PHB reset ...\n");
3842 		pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3843 		pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3844 	}
3845 
3846 	/* Remove M64 resource if we can't configure it successfully */
3847 	if (!phb->init_m64 || phb->init_m64(phb))
3848 		hose->mem_resources[1].flags = 0;
3849 }
3850 
3851 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3852 {
3853 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3854 }
3855 
3856 void __init pnv_pci_init_npu_phb(struct device_node *np)
3857 {
3858 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU);
3859 }
3860 
3861 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3862 {
3863 	struct device_node *phbn;
3864 	const __be64 *prop64;
3865 	u64 hub_id;
3866 
3867 	pr_info("Probing IODA IO-Hub %s\n", np->full_name);
3868 
3869 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3870 	if (!prop64) {
3871 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3872 		return;
3873 	}
3874 	hub_id = be64_to_cpup(prop64);
3875 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3876 
3877 	/* Count child PHBs */
3878 	for_each_child_of_node(np, phbn) {
3879 		/* Look for IODA1 PHBs */
3880 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3881 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3882 	}
3883 }
3884