xref: /linux/arch/powerpc/platforms/powernv/pci-ioda.c (revision 9a379e77033f02c4a071891afdf0f0a01eff8ccb)
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/delay.h>
18 #include <linux/string.h>
19 #include <linux/init.h>
20 #include <linux/bootmem.h>
21 #include <linux/irq.h>
22 #include <linux/io.h>
23 #include <linux/msi.h>
24 #include <linux/memblock.h>
25 #include <linux/iommu.h>
26 #include <linux/rculist.h>
27 #include <linux/sizes.h>
28 
29 #include <asm/sections.h>
30 #include <asm/io.h>
31 #include <asm/prom.h>
32 #include <asm/pci-bridge.h>
33 #include <asm/machdep.h>
34 #include <asm/msi_bitmap.h>
35 #include <asm/ppc-pci.h>
36 #include <asm/opal.h>
37 #include <asm/iommu.h>
38 #include <asm/tce.h>
39 #include <asm/xics.h>
40 #include <asm/debugfs.h>
41 #include <asm/firmware.h>
42 #include <asm/pnv-pci.h>
43 #include <asm/mmzone.h>
44 
45 #include <misc/cxl-base.h>
46 
47 #include "powernv.h"
48 #include "pci.h"
49 
50 #define PNV_IODA1_M64_NUM	16	/* Number of M64 BARs	*/
51 #define PNV_IODA1_M64_SEGS	8	/* Segments per M64 BAR	*/
52 #define PNV_IODA1_DMA32_SEGSIZE	0x10000000
53 
54 #define POWERNV_IOMMU_DEFAULT_LEVELS	1
55 #define POWERNV_IOMMU_MAX_LEVELS	5
56 
57 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_NVLINK",
58 					      "NPU_OCAPI" };
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# %.2x] %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 static bool pci_reset_phbs __read_mostly;
94 
95 static int __init iommu_setup(char *str)
96 {
97 	if (!str)
98 		return -EINVAL;
99 
100 	while (*str) {
101 		if (!strncmp(str, "nobypass", 8)) {
102 			pnv_iommu_bypass_disabled = true;
103 			pr_info("PowerNV: IOMMU bypass window disabled.\n");
104 			break;
105 		}
106 		str += strcspn(str, ",");
107 		if (*str == ',')
108 			str++;
109 	}
110 
111 	return 0;
112 }
113 early_param("iommu", iommu_setup);
114 
115 static int __init pci_reset_phbs_setup(char *str)
116 {
117 	pci_reset_phbs = true;
118 	return 0;
119 }
120 
121 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup);
122 
123 static inline bool pnv_pci_is_m64(struct pnv_phb *phb, struct resource *r)
124 {
125 	/*
126 	 * WARNING: We cannot rely on the resource flags. The Linux PCI
127 	 * allocation code sometimes decides to put a 64-bit prefetchable
128 	 * BAR in the 32-bit window, so we have to compare the addresses.
129 	 *
130 	 * For simplicity we only test resource start.
131 	 */
132 	return (r->start >= phb->ioda.m64_base &&
133 		r->start < (phb->ioda.m64_base + phb->ioda.m64_size));
134 }
135 
136 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags)
137 {
138 	unsigned long flags = (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
139 
140 	return (resource_flags & flags) == flags;
141 }
142 
143 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
144 {
145 	s64 rc;
146 
147 	phb->ioda.pe_array[pe_no].phb = phb;
148 	phb->ioda.pe_array[pe_no].pe_number = pe_no;
149 
150 	/*
151 	 * Clear the PE frozen state as it might be put into frozen state
152 	 * in the last PCI remove path. It's not harmful to do so when the
153 	 * PE is already in unfrozen state.
154 	 */
155 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no,
156 				       OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
157 	if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED)
158 		pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
159 			__func__, rc, phb->hose->global_number, pe_no);
160 
161 	return &phb->ioda.pe_array[pe_no];
162 }
163 
164 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
165 {
166 	if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
167 		pr_warn("%s: Invalid PE %x on PHB#%x\n",
168 			__func__, pe_no, phb->hose->global_number);
169 		return;
170 	}
171 
172 	if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
173 		pr_debug("%s: PE %x was reserved on PHB#%x\n",
174 			 __func__, pe_no, phb->hose->global_number);
175 
176 	pnv_ioda_init_pe(phb, pe_no);
177 }
178 
179 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
180 {
181 	long pe;
182 
183 	for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
184 		if (!test_and_set_bit(pe, phb->ioda.pe_alloc))
185 			return pnv_ioda_init_pe(phb, pe);
186 	}
187 
188 	return NULL;
189 }
190 
191 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
192 {
193 	struct pnv_phb *phb = pe->phb;
194 	unsigned int pe_num = pe->pe_number;
195 
196 	WARN_ON(pe->pdev);
197 
198 	memset(pe, 0, sizeof(struct pnv_ioda_pe));
199 	clear_bit(pe_num, phb->ioda.pe_alloc);
200 }
201 
202 /* The default M64 BAR is shared by all PEs */
203 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
204 {
205 	const char *desc;
206 	struct resource *r;
207 	s64 rc;
208 
209 	/* Configure the default M64 BAR */
210 	rc = opal_pci_set_phb_mem_window(phb->opal_id,
211 					 OPAL_M64_WINDOW_TYPE,
212 					 phb->ioda.m64_bar_idx,
213 					 phb->ioda.m64_base,
214 					 0, /* unused */
215 					 phb->ioda.m64_size);
216 	if (rc != OPAL_SUCCESS) {
217 		desc = "configuring";
218 		goto fail;
219 	}
220 
221 	/* Enable the default M64 BAR */
222 	rc = opal_pci_phb_mmio_enable(phb->opal_id,
223 				      OPAL_M64_WINDOW_TYPE,
224 				      phb->ioda.m64_bar_idx,
225 				      OPAL_ENABLE_M64_SPLIT);
226 	if (rc != OPAL_SUCCESS) {
227 		desc = "enabling";
228 		goto fail;
229 	}
230 
231 	/*
232 	 * Exclude the segments for reserved and root bus PE, which
233 	 * are first or last two PEs.
234 	 */
235 	r = &phb->hose->mem_resources[1];
236 	if (phb->ioda.reserved_pe_idx == 0)
237 		r->start += (2 * phb->ioda.m64_segsize);
238 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
239 		r->end -= (2 * phb->ioda.m64_segsize);
240 	else
241 		pr_warn("  Cannot strip M64 segment for reserved PE#%x\n",
242 			phb->ioda.reserved_pe_idx);
243 
244 	return 0;
245 
246 fail:
247 	pr_warn("  Failure %lld %s M64 BAR#%d\n",
248 		rc, desc, phb->ioda.m64_bar_idx);
249 	opal_pci_phb_mmio_enable(phb->opal_id,
250 				 OPAL_M64_WINDOW_TYPE,
251 				 phb->ioda.m64_bar_idx,
252 				 OPAL_DISABLE_M64);
253 	return -EIO;
254 }
255 
256 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
257 					 unsigned long *pe_bitmap)
258 {
259 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
260 	struct pnv_phb *phb = hose->private_data;
261 	struct resource *r;
262 	resource_size_t base, sgsz, start, end;
263 	int segno, i;
264 
265 	base = phb->ioda.m64_base;
266 	sgsz = phb->ioda.m64_segsize;
267 	for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
268 		r = &pdev->resource[i];
269 		if (!r->parent || !pnv_pci_is_m64(phb, r))
270 			continue;
271 
272 		start = _ALIGN_DOWN(r->start - base, sgsz);
273 		end = _ALIGN_UP(r->end - base, sgsz);
274 		for (segno = start / sgsz; segno < end / sgsz; segno++) {
275 			if (pe_bitmap)
276 				set_bit(segno, pe_bitmap);
277 			else
278 				pnv_ioda_reserve_pe(phb, segno);
279 		}
280 	}
281 }
282 
283 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
284 {
285 	struct resource *r;
286 	int index;
287 
288 	/*
289 	 * There are 16 M64 BARs, each of which has 8 segments. So
290 	 * there are as many M64 segments as the maximum number of
291 	 * PEs, which is 128.
292 	 */
293 	for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
294 		unsigned long base, segsz = phb->ioda.m64_segsize;
295 		int64_t rc;
296 
297 		base = phb->ioda.m64_base +
298 		       index * PNV_IODA1_M64_SEGS * segsz;
299 		rc = opal_pci_set_phb_mem_window(phb->opal_id,
300 				OPAL_M64_WINDOW_TYPE, index, base, 0,
301 				PNV_IODA1_M64_SEGS * segsz);
302 		if (rc != OPAL_SUCCESS) {
303 			pr_warn("  Error %lld setting M64 PHB#%x-BAR#%d\n",
304 				rc, phb->hose->global_number, index);
305 			goto fail;
306 		}
307 
308 		rc = opal_pci_phb_mmio_enable(phb->opal_id,
309 				OPAL_M64_WINDOW_TYPE, index,
310 				OPAL_ENABLE_M64_SPLIT);
311 		if (rc != OPAL_SUCCESS) {
312 			pr_warn("  Error %lld enabling M64 PHB#%x-BAR#%d\n",
313 				rc, phb->hose->global_number, index);
314 			goto fail;
315 		}
316 	}
317 
318 	/*
319 	 * Exclude the segments for reserved and root bus PE, which
320 	 * are first or last two PEs.
321 	 */
322 	r = &phb->hose->mem_resources[1];
323 	if (phb->ioda.reserved_pe_idx == 0)
324 		r->start += (2 * phb->ioda.m64_segsize);
325 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
326 		r->end -= (2 * phb->ioda.m64_segsize);
327 	else
328 		WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
329 		     phb->ioda.reserved_pe_idx, phb->hose->global_number);
330 
331 	return 0;
332 
333 fail:
334 	for ( ; index >= 0; index--)
335 		opal_pci_phb_mmio_enable(phb->opal_id,
336 			OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
337 
338 	return -EIO;
339 }
340 
341 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
342 				    unsigned long *pe_bitmap,
343 				    bool all)
344 {
345 	struct pci_dev *pdev;
346 
347 	list_for_each_entry(pdev, &bus->devices, bus_list) {
348 		pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
349 
350 		if (all && pdev->subordinate)
351 			pnv_ioda_reserve_m64_pe(pdev->subordinate,
352 						pe_bitmap, all);
353 	}
354 }
355 
356 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
357 {
358 	struct pci_controller *hose = pci_bus_to_host(bus);
359 	struct pnv_phb *phb = hose->private_data;
360 	struct pnv_ioda_pe *master_pe, *pe;
361 	unsigned long size, *pe_alloc;
362 	int i;
363 
364 	/* Root bus shouldn't use M64 */
365 	if (pci_is_root_bus(bus))
366 		return NULL;
367 
368 	/* Allocate bitmap */
369 	size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
370 	pe_alloc = kzalloc(size, GFP_KERNEL);
371 	if (!pe_alloc) {
372 		pr_warn("%s: Out of memory !\n",
373 			__func__);
374 		return NULL;
375 	}
376 
377 	/* Figure out reserved PE numbers by the PE */
378 	pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
379 
380 	/*
381 	 * the current bus might not own M64 window and that's all
382 	 * contributed by its child buses. For the case, we needn't
383 	 * pick M64 dependent PE#.
384 	 */
385 	if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
386 		kfree(pe_alloc);
387 		return NULL;
388 	}
389 
390 	/*
391 	 * Figure out the master PE and put all slave PEs to master
392 	 * PE's list to form compound PE.
393 	 */
394 	master_pe = NULL;
395 	i = -1;
396 	while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
397 		phb->ioda.total_pe_num) {
398 		pe = &phb->ioda.pe_array[i];
399 
400 		phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
401 		if (!master_pe) {
402 			pe->flags |= PNV_IODA_PE_MASTER;
403 			INIT_LIST_HEAD(&pe->slaves);
404 			master_pe = pe;
405 		} else {
406 			pe->flags |= PNV_IODA_PE_SLAVE;
407 			pe->master = master_pe;
408 			list_add_tail(&pe->list, &master_pe->slaves);
409 		}
410 
411 		/*
412 		 * P7IOC supports M64DT, which helps mapping M64 segment
413 		 * to one particular PE#. However, PHB3 has fixed mapping
414 		 * between M64 segment and PE#. In order to have same logic
415 		 * for P7IOC and PHB3, we enforce fixed mapping between M64
416 		 * segment and PE# on P7IOC.
417 		 */
418 		if (phb->type == PNV_PHB_IODA1) {
419 			int64_t rc;
420 
421 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
422 					pe->pe_number, OPAL_M64_WINDOW_TYPE,
423 					pe->pe_number / PNV_IODA1_M64_SEGS,
424 					pe->pe_number % PNV_IODA1_M64_SEGS);
425 			if (rc != OPAL_SUCCESS)
426 				pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
427 					__func__, rc, phb->hose->global_number,
428 					pe->pe_number);
429 		}
430 	}
431 
432 	kfree(pe_alloc);
433 	return master_pe;
434 }
435 
436 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
437 {
438 	struct pci_controller *hose = phb->hose;
439 	struct device_node *dn = hose->dn;
440 	struct resource *res;
441 	u32 m64_range[2], i;
442 	const __be32 *r;
443 	u64 pci_addr;
444 
445 	if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
446 		pr_info("  Not support M64 window\n");
447 		return;
448 	}
449 
450 	if (!firmware_has_feature(FW_FEATURE_OPAL)) {
451 		pr_info("  Firmware too old to support M64 window\n");
452 		return;
453 	}
454 
455 	r = of_get_property(dn, "ibm,opal-m64-window", NULL);
456 	if (!r) {
457 		pr_info("  No <ibm,opal-m64-window> on %pOF\n",
458 			dn);
459 		return;
460 	}
461 
462 	/*
463 	 * Find the available M64 BAR range and pickup the last one for
464 	 * covering the whole 64-bits space. We support only one range.
465 	 */
466 	if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
467 				       m64_range, 2)) {
468 		/* In absence of the property, assume 0..15 */
469 		m64_range[0] = 0;
470 		m64_range[1] = 16;
471 	}
472 	/* We only support 64 bits in our allocator */
473 	if (m64_range[1] > 63) {
474 		pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
475 			__func__, m64_range[1], phb->hose->global_number);
476 		m64_range[1] = 63;
477 	}
478 	/* Empty range, no m64 */
479 	if (m64_range[1] <= m64_range[0]) {
480 		pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
481 			__func__, phb->hose->global_number);
482 		return;
483 	}
484 
485 	/* Configure M64 informations */
486 	res = &hose->mem_resources[1];
487 	res->name = dn->full_name;
488 	res->start = of_translate_address(dn, r + 2);
489 	res->end = res->start + of_read_number(r + 4, 2) - 1;
490 	res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
491 	pci_addr = of_read_number(r, 2);
492 	hose->mem_offset[1] = res->start - pci_addr;
493 
494 	phb->ioda.m64_size = resource_size(res);
495 	phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
496 	phb->ioda.m64_base = pci_addr;
497 
498 	/* This lines up nicely with the display from processing OF ranges */
499 	pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
500 		res->start, res->end, pci_addr, m64_range[0],
501 		m64_range[0] + m64_range[1] - 1);
502 
503 	/* Mark all M64 used up by default */
504 	phb->ioda.m64_bar_alloc = (unsigned long)-1;
505 
506 	/* Use last M64 BAR to cover M64 window */
507 	m64_range[1]--;
508 	phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
509 
510 	pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
511 
512 	/* Mark remaining ones free */
513 	for (i = m64_range[0]; i < m64_range[1]; i++)
514 		clear_bit(i, &phb->ioda.m64_bar_alloc);
515 
516 	/*
517 	 * Setup init functions for M64 based on IODA version, IODA3 uses
518 	 * the IODA2 code.
519 	 */
520 	if (phb->type == PNV_PHB_IODA1)
521 		phb->init_m64 = pnv_ioda1_init_m64;
522 	else
523 		phb->init_m64 = pnv_ioda2_init_m64;
524 	phb->reserve_m64_pe = pnv_ioda_reserve_m64_pe;
525 	phb->pick_m64_pe = pnv_ioda_pick_m64_pe;
526 }
527 
528 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
529 {
530 	struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
531 	struct pnv_ioda_pe *slave;
532 	s64 rc;
533 
534 	/* Fetch master PE */
535 	if (pe->flags & PNV_IODA_PE_SLAVE) {
536 		pe = pe->master;
537 		if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
538 			return;
539 
540 		pe_no = pe->pe_number;
541 	}
542 
543 	/* Freeze master PE */
544 	rc = opal_pci_eeh_freeze_set(phb->opal_id,
545 				     pe_no,
546 				     OPAL_EEH_ACTION_SET_FREEZE_ALL);
547 	if (rc != OPAL_SUCCESS) {
548 		pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
549 			__func__, rc, phb->hose->global_number, pe_no);
550 		return;
551 	}
552 
553 	/* Freeze slave PEs */
554 	if (!(pe->flags & PNV_IODA_PE_MASTER))
555 		return;
556 
557 	list_for_each_entry(slave, &pe->slaves, list) {
558 		rc = opal_pci_eeh_freeze_set(phb->opal_id,
559 					     slave->pe_number,
560 					     OPAL_EEH_ACTION_SET_FREEZE_ALL);
561 		if (rc != OPAL_SUCCESS)
562 			pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
563 				__func__, rc, phb->hose->global_number,
564 				slave->pe_number);
565 	}
566 }
567 
568 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
569 {
570 	struct pnv_ioda_pe *pe, *slave;
571 	s64 rc;
572 
573 	/* Find master PE */
574 	pe = &phb->ioda.pe_array[pe_no];
575 	if (pe->flags & PNV_IODA_PE_SLAVE) {
576 		pe = pe->master;
577 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
578 		pe_no = pe->pe_number;
579 	}
580 
581 	/* Clear frozen state for master PE */
582 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
583 	if (rc != OPAL_SUCCESS) {
584 		pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
585 			__func__, rc, opt, phb->hose->global_number, pe_no);
586 		return -EIO;
587 	}
588 
589 	if (!(pe->flags & PNV_IODA_PE_MASTER))
590 		return 0;
591 
592 	/* Clear frozen state for slave PEs */
593 	list_for_each_entry(slave, &pe->slaves, list) {
594 		rc = opal_pci_eeh_freeze_clear(phb->opal_id,
595 					     slave->pe_number,
596 					     opt);
597 		if (rc != OPAL_SUCCESS) {
598 			pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
599 				__func__, rc, opt, phb->hose->global_number,
600 				slave->pe_number);
601 			return -EIO;
602 		}
603 	}
604 
605 	return 0;
606 }
607 
608 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
609 {
610 	struct pnv_ioda_pe *slave, *pe;
611 	u8 fstate, state;
612 	__be16 pcierr;
613 	s64 rc;
614 
615 	/* Sanity check on PE number */
616 	if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
617 		return OPAL_EEH_STOPPED_PERM_UNAVAIL;
618 
619 	/*
620 	 * Fetch the master PE and the PE instance might be
621 	 * not initialized yet.
622 	 */
623 	pe = &phb->ioda.pe_array[pe_no];
624 	if (pe->flags & PNV_IODA_PE_SLAVE) {
625 		pe = pe->master;
626 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
627 		pe_no = pe->pe_number;
628 	}
629 
630 	/* Check the master PE */
631 	rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
632 					&state, &pcierr, NULL);
633 	if (rc != OPAL_SUCCESS) {
634 		pr_warn("%s: Failure %lld getting "
635 			"PHB#%x-PE#%x state\n",
636 			__func__, rc,
637 			phb->hose->global_number, pe_no);
638 		return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
639 	}
640 
641 	/* Check the slave PE */
642 	if (!(pe->flags & PNV_IODA_PE_MASTER))
643 		return state;
644 
645 	list_for_each_entry(slave, &pe->slaves, list) {
646 		rc = opal_pci_eeh_freeze_status(phb->opal_id,
647 						slave->pe_number,
648 						&fstate,
649 						&pcierr,
650 						NULL);
651 		if (rc != OPAL_SUCCESS) {
652 			pr_warn("%s: Failure %lld getting "
653 				"PHB#%x-PE#%x state\n",
654 				__func__, rc,
655 				phb->hose->global_number, slave->pe_number);
656 			return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
657 		}
658 
659 		/*
660 		 * Override the result based on the ascending
661 		 * priority.
662 		 */
663 		if (fstate > state)
664 			state = fstate;
665 	}
666 
667 	return state;
668 }
669 
670 /* Currently those 2 are only used when MSIs are enabled, this will change
671  * but in the meantime, we need to protect them to avoid warnings
672  */
673 #ifdef CONFIG_PCI_MSI
674 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
675 {
676 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
677 	struct pnv_phb *phb = hose->private_data;
678 	struct pci_dn *pdn = pci_get_pdn(dev);
679 
680 	if (!pdn)
681 		return NULL;
682 	if (pdn->pe_number == IODA_INVALID_PE)
683 		return NULL;
684 	return &phb->ioda.pe_array[pdn->pe_number];
685 }
686 #endif /* CONFIG_PCI_MSI */
687 
688 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
689 				  struct pnv_ioda_pe *parent,
690 				  struct pnv_ioda_pe *child,
691 				  bool is_add)
692 {
693 	const char *desc = is_add ? "adding" : "removing";
694 	uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
695 			      OPAL_REMOVE_PE_FROM_DOMAIN;
696 	struct pnv_ioda_pe *slave;
697 	long rc;
698 
699 	/* Parent PE affects child PE */
700 	rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
701 				child->pe_number, op);
702 	if (rc != OPAL_SUCCESS) {
703 		pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
704 			rc, desc);
705 		return -ENXIO;
706 	}
707 
708 	if (!(child->flags & PNV_IODA_PE_MASTER))
709 		return 0;
710 
711 	/* Compound case: parent PE affects slave PEs */
712 	list_for_each_entry(slave, &child->slaves, list) {
713 		rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
714 					slave->pe_number, op);
715 		if (rc != OPAL_SUCCESS) {
716 			pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
717 				rc, desc);
718 			return -ENXIO;
719 		}
720 	}
721 
722 	return 0;
723 }
724 
725 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
726 			      struct pnv_ioda_pe *pe,
727 			      bool is_add)
728 {
729 	struct pnv_ioda_pe *slave;
730 	struct pci_dev *pdev = NULL;
731 	int ret;
732 
733 	/*
734 	 * Clear PE frozen state. If it's master PE, we need
735 	 * clear slave PE frozen state as well.
736 	 */
737 	if (is_add) {
738 		opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
739 					  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
740 		if (pe->flags & PNV_IODA_PE_MASTER) {
741 			list_for_each_entry(slave, &pe->slaves, list)
742 				opal_pci_eeh_freeze_clear(phb->opal_id,
743 							  slave->pe_number,
744 							  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
745 		}
746 	}
747 
748 	/*
749 	 * Associate PE in PELT. We need add the PE into the
750 	 * corresponding PELT-V as well. Otherwise, the error
751 	 * originated from the PE might contribute to other
752 	 * PEs.
753 	 */
754 	ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
755 	if (ret)
756 		return ret;
757 
758 	/* For compound PEs, any one affects all of them */
759 	if (pe->flags & PNV_IODA_PE_MASTER) {
760 		list_for_each_entry(slave, &pe->slaves, list) {
761 			ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
762 			if (ret)
763 				return ret;
764 		}
765 	}
766 
767 	if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
768 		pdev = pe->pbus->self;
769 	else if (pe->flags & PNV_IODA_PE_DEV)
770 		pdev = pe->pdev->bus->self;
771 #ifdef CONFIG_PCI_IOV
772 	else if (pe->flags & PNV_IODA_PE_VF)
773 		pdev = pe->parent_dev;
774 #endif /* CONFIG_PCI_IOV */
775 	while (pdev) {
776 		struct pci_dn *pdn = pci_get_pdn(pdev);
777 		struct pnv_ioda_pe *parent;
778 
779 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
780 			parent = &phb->ioda.pe_array[pdn->pe_number];
781 			ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
782 			if (ret)
783 				return ret;
784 		}
785 
786 		pdev = pdev->bus->self;
787 	}
788 
789 	return 0;
790 }
791 
792 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
793 {
794 	struct pci_dev *parent;
795 	uint8_t bcomp, dcomp, fcomp;
796 	int64_t rc;
797 	long rid_end, rid;
798 
799 	/* Currently, we just deconfigure VF PE. Bus PE will always there.*/
800 	if (pe->pbus) {
801 		int count;
802 
803 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
804 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
805 		parent = pe->pbus->self;
806 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
807 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
808 		else
809 			count = 1;
810 
811 		switch(count) {
812 		case  1: bcomp = OpalPciBusAll;         break;
813 		case  2: bcomp = OpalPciBus7Bits;       break;
814 		case  4: bcomp = OpalPciBus6Bits;       break;
815 		case  8: bcomp = OpalPciBus5Bits;       break;
816 		case 16: bcomp = OpalPciBus4Bits;       break;
817 		case 32: bcomp = OpalPciBus3Bits;       break;
818 		default:
819 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
820 			        count);
821 			/* Do an exact match only */
822 			bcomp = OpalPciBusAll;
823 		}
824 		rid_end = pe->rid + (count << 8);
825 	} else {
826 #ifdef CONFIG_PCI_IOV
827 		if (pe->flags & PNV_IODA_PE_VF)
828 			parent = pe->parent_dev;
829 		else
830 #endif
831 			parent = pe->pdev->bus->self;
832 		bcomp = OpalPciBusAll;
833 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
834 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
835 		rid_end = pe->rid + 1;
836 	}
837 
838 	/* Clear the reverse map */
839 	for (rid = pe->rid; rid < rid_end; rid++)
840 		phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
841 
842 	/* Release from all parents PELT-V */
843 	while (parent) {
844 		struct pci_dn *pdn = pci_get_pdn(parent);
845 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
846 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
847 						pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
848 			/* XXX What to do in case of error ? */
849 		}
850 		parent = parent->bus->self;
851 	}
852 
853 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
854 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
855 
856 	/* Disassociate PE in PELT */
857 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
858 				pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
859 	if (rc)
860 		pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
861 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
862 			     bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
863 	if (rc)
864 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
865 
866 	pe->pbus = NULL;
867 	pe->pdev = NULL;
868 #ifdef CONFIG_PCI_IOV
869 	pe->parent_dev = NULL;
870 #endif
871 
872 	return 0;
873 }
874 
875 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
876 {
877 	struct pci_dev *parent;
878 	uint8_t bcomp, dcomp, fcomp;
879 	long rc, rid_end, rid;
880 
881 	/* Bus validation ? */
882 	if (pe->pbus) {
883 		int count;
884 
885 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
886 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
887 		parent = pe->pbus->self;
888 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
889 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
890 		else
891 			count = 1;
892 
893 		switch(count) {
894 		case  1: bcomp = OpalPciBusAll;		break;
895 		case  2: bcomp = OpalPciBus7Bits;	break;
896 		case  4: bcomp = OpalPciBus6Bits;	break;
897 		case  8: bcomp = OpalPciBus5Bits;	break;
898 		case 16: bcomp = OpalPciBus4Bits;	break;
899 		case 32: bcomp = OpalPciBus3Bits;	break;
900 		default:
901 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
902 			        count);
903 			/* Do an exact match only */
904 			bcomp = OpalPciBusAll;
905 		}
906 		rid_end = pe->rid + (count << 8);
907 	} else {
908 #ifdef CONFIG_PCI_IOV
909 		if (pe->flags & PNV_IODA_PE_VF)
910 			parent = pe->parent_dev;
911 		else
912 #endif /* CONFIG_PCI_IOV */
913 			parent = pe->pdev->bus->self;
914 		bcomp = OpalPciBusAll;
915 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
916 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
917 		rid_end = pe->rid + 1;
918 	}
919 
920 	/*
921 	 * Associate PE in PELT. We need add the PE into the
922 	 * corresponding PELT-V as well. Otherwise, the error
923 	 * originated from the PE might contribute to other
924 	 * PEs.
925 	 */
926 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
927 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
928 	if (rc) {
929 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
930 		return -ENXIO;
931 	}
932 
933 	/*
934 	 * Configure PELTV. NPUs don't have a PELTV table so skip
935 	 * configuration on them.
936 	 */
937 	if (phb->type != PNV_PHB_NPU_NVLINK && phb->type != PNV_PHB_NPU_OCAPI)
938 		pnv_ioda_set_peltv(phb, pe, true);
939 
940 	/* Setup reverse map */
941 	for (rid = pe->rid; rid < rid_end; rid++)
942 		phb->ioda.pe_rmap[rid] = pe->pe_number;
943 
944 	/* Setup one MVTs on IODA1 */
945 	if (phb->type != PNV_PHB_IODA1) {
946 		pe->mve_number = 0;
947 		goto out;
948 	}
949 
950 	pe->mve_number = pe->pe_number;
951 	rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
952 	if (rc != OPAL_SUCCESS) {
953 		pe_err(pe, "OPAL error %ld setting up MVE %x\n",
954 		       rc, pe->mve_number);
955 		pe->mve_number = -1;
956 	} else {
957 		rc = opal_pci_set_mve_enable(phb->opal_id,
958 					     pe->mve_number, OPAL_ENABLE_MVE);
959 		if (rc) {
960 			pe_err(pe, "OPAL error %ld enabling MVE %x\n",
961 			       rc, pe->mve_number);
962 			pe->mve_number = -1;
963 		}
964 	}
965 
966 out:
967 	return 0;
968 }
969 
970 #ifdef CONFIG_PCI_IOV
971 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
972 {
973 	struct pci_dn *pdn = pci_get_pdn(dev);
974 	int i;
975 	struct resource *res, res2;
976 	resource_size_t size;
977 	u16 num_vfs;
978 
979 	if (!dev->is_physfn)
980 		return -EINVAL;
981 
982 	/*
983 	 * "offset" is in VFs.  The M64 windows are sized so that when they
984 	 * are segmented, each segment is the same size as the IOV BAR.
985 	 * Each segment is in a separate PE, and the high order bits of the
986 	 * address are the PE number.  Therefore, each VF's BAR is in a
987 	 * separate PE, and changing the IOV BAR start address changes the
988 	 * range of PEs the VFs are in.
989 	 */
990 	num_vfs = pdn->num_vfs;
991 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
992 		res = &dev->resource[i + PCI_IOV_RESOURCES];
993 		if (!res->flags || !res->parent)
994 			continue;
995 
996 		/*
997 		 * The actual IOV BAR range is determined by the start address
998 		 * and the actual size for num_vfs VFs BAR.  This check is to
999 		 * make sure that after shifting, the range will not overlap
1000 		 * with another device.
1001 		 */
1002 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
1003 		res2.flags = res->flags;
1004 		res2.start = res->start + (size * offset);
1005 		res2.end = res2.start + (size * num_vfs) - 1;
1006 
1007 		if (res2.end > res->end) {
1008 			dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
1009 				i, &res2, res, num_vfs, offset);
1010 			return -EBUSY;
1011 		}
1012 	}
1013 
1014 	/*
1015 	 * Since M64 BAR shares segments among all possible 256 PEs,
1016 	 * we have to shift the beginning of PF IOV BAR to make it start from
1017 	 * the segment which belongs to the PE number assigned to the first VF.
1018 	 * This creates a "hole" in the /proc/iomem which could be used for
1019 	 * allocating other resources so we reserve this area below and
1020 	 * release when IOV is released.
1021 	 */
1022 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1023 		res = &dev->resource[i + PCI_IOV_RESOURCES];
1024 		if (!res->flags || !res->parent)
1025 			continue;
1026 
1027 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
1028 		res2 = *res;
1029 		res->start += size * offset;
1030 
1031 		dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
1032 			 i, &res2, res, (offset > 0) ? "En" : "Dis",
1033 			 num_vfs, offset);
1034 
1035 		if (offset < 0) {
1036 			devm_release_resource(&dev->dev, &pdn->holes[i]);
1037 			memset(&pdn->holes[i], 0, sizeof(pdn->holes[i]));
1038 		}
1039 
1040 		pci_update_resource(dev, i + PCI_IOV_RESOURCES);
1041 
1042 		if (offset > 0) {
1043 			pdn->holes[i].start = res2.start;
1044 			pdn->holes[i].end = res2.start + size * offset - 1;
1045 			pdn->holes[i].flags = IORESOURCE_BUS;
1046 			pdn->holes[i].name = "pnv_iov_reserved";
1047 			devm_request_resource(&dev->dev, res->parent,
1048 					&pdn->holes[i]);
1049 		}
1050 	}
1051 	return 0;
1052 }
1053 #endif /* CONFIG_PCI_IOV */
1054 
1055 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
1056 {
1057 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
1058 	struct pnv_phb *phb = hose->private_data;
1059 	struct pci_dn *pdn = pci_get_pdn(dev);
1060 	struct pnv_ioda_pe *pe;
1061 
1062 	if (!pdn) {
1063 		pr_err("%s: Device tree node not associated properly\n",
1064 			   pci_name(dev));
1065 		return NULL;
1066 	}
1067 	if (pdn->pe_number != IODA_INVALID_PE)
1068 		return NULL;
1069 
1070 	pe = pnv_ioda_alloc_pe(phb);
1071 	if (!pe) {
1072 		pr_warn("%s: Not enough PE# available, disabling device\n",
1073 			pci_name(dev));
1074 		return NULL;
1075 	}
1076 
1077 	/* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1078 	 * pointer in the PE data structure, both should be destroyed at the
1079 	 * same time. However, this needs to be looked at more closely again
1080 	 * once we actually start removing things (Hotplug, SR-IOV, ...)
1081 	 *
1082 	 * At some point we want to remove the PDN completely anyways
1083 	 */
1084 	pci_dev_get(dev);
1085 	pdn->pe_number = pe->pe_number;
1086 	pe->flags = PNV_IODA_PE_DEV;
1087 	pe->pdev = dev;
1088 	pe->pbus = NULL;
1089 	pe->mve_number = -1;
1090 	pe->rid = dev->bus->number << 8 | pdn->devfn;
1091 
1092 	pe_info(pe, "Associated device to PE\n");
1093 
1094 	if (pnv_ioda_configure_pe(phb, pe)) {
1095 		/* XXX What do we do here ? */
1096 		pnv_ioda_free_pe(pe);
1097 		pdn->pe_number = IODA_INVALID_PE;
1098 		pe->pdev = NULL;
1099 		pci_dev_put(dev);
1100 		return NULL;
1101 	}
1102 
1103 	/* Put PE to the list */
1104 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1105 
1106 	return pe;
1107 }
1108 
1109 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1110 {
1111 	struct pci_dev *dev;
1112 
1113 	list_for_each_entry(dev, &bus->devices, bus_list) {
1114 		struct pci_dn *pdn = pci_get_pdn(dev);
1115 
1116 		if (pdn == NULL) {
1117 			pr_warn("%s: No device node associated with device !\n",
1118 				pci_name(dev));
1119 			continue;
1120 		}
1121 
1122 		/*
1123 		 * In partial hotplug case, the PCI device might be still
1124 		 * associated with the PE and needn't attach it to the PE
1125 		 * again.
1126 		 */
1127 		if (pdn->pe_number != IODA_INVALID_PE)
1128 			continue;
1129 
1130 		pe->device_count++;
1131 		pdn->pe_number = pe->pe_number;
1132 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1133 			pnv_ioda_setup_same_PE(dev->subordinate, pe);
1134 	}
1135 }
1136 
1137 /*
1138  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1139  * single PCI bus. Another one that contains the primary PCI bus and its
1140  * subordinate PCI devices and buses. The second type of PE is normally
1141  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1142  */
1143 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1144 {
1145 	struct pci_controller *hose = pci_bus_to_host(bus);
1146 	struct pnv_phb *phb = hose->private_data;
1147 	struct pnv_ioda_pe *pe = NULL;
1148 	unsigned int pe_num;
1149 
1150 	/*
1151 	 * In partial hotplug case, the PE instance might be still alive.
1152 	 * We should reuse it instead of allocating a new one.
1153 	 */
1154 	pe_num = phb->ioda.pe_rmap[bus->number << 8];
1155 	if (pe_num != IODA_INVALID_PE) {
1156 		pe = &phb->ioda.pe_array[pe_num];
1157 		pnv_ioda_setup_same_PE(bus, pe);
1158 		return NULL;
1159 	}
1160 
1161 	/* PE number for root bus should have been reserved */
1162 	if (pci_is_root_bus(bus) &&
1163 	    phb->ioda.root_pe_idx != IODA_INVALID_PE)
1164 		pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1165 
1166 	/* Check if PE is determined by M64 */
1167 	if (!pe && phb->pick_m64_pe)
1168 		pe = phb->pick_m64_pe(bus, all);
1169 
1170 	/* The PE number isn't pinned by M64 */
1171 	if (!pe)
1172 		pe = pnv_ioda_alloc_pe(phb);
1173 
1174 	if (!pe) {
1175 		pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1176 			__func__, pci_domain_nr(bus), bus->number);
1177 		return NULL;
1178 	}
1179 
1180 	pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1181 	pe->pbus = bus;
1182 	pe->pdev = NULL;
1183 	pe->mve_number = -1;
1184 	pe->rid = bus->busn_res.start << 8;
1185 
1186 	if (all)
1187 		pe_info(pe, "Secondary bus %d..%d associated with PE#%x\n",
1188 			bus->busn_res.start, bus->busn_res.end, pe->pe_number);
1189 	else
1190 		pe_info(pe, "Secondary bus %d associated with PE#%x\n",
1191 			bus->busn_res.start, pe->pe_number);
1192 
1193 	if (pnv_ioda_configure_pe(phb, pe)) {
1194 		/* XXX What do we do here ? */
1195 		pnv_ioda_free_pe(pe);
1196 		pe->pbus = NULL;
1197 		return NULL;
1198 	}
1199 
1200 	/* Associate it with all child devices */
1201 	pnv_ioda_setup_same_PE(bus, pe);
1202 
1203 	/* Put PE to the list */
1204 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1205 
1206 	return pe;
1207 }
1208 
1209 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1210 {
1211 	int pe_num, found_pe = false, rc;
1212 	long rid;
1213 	struct pnv_ioda_pe *pe;
1214 	struct pci_dev *gpu_pdev;
1215 	struct pci_dn *npu_pdn;
1216 	struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1217 	struct pnv_phb *phb = hose->private_data;
1218 
1219 	/*
1220 	 * Due to a hardware errata PE#0 on the NPU is reserved for
1221 	 * error handling. This means we only have three PEs remaining
1222 	 * which need to be assigned to four links, implying some
1223 	 * links must share PEs.
1224 	 *
1225 	 * To achieve this we assign PEs such that NPUs linking the
1226 	 * same GPU get assigned the same PE.
1227 	 */
1228 	gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1229 	for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1230 		pe = &phb->ioda.pe_array[pe_num];
1231 		if (!pe->pdev)
1232 			continue;
1233 
1234 		if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1235 			/*
1236 			 * This device has the same peer GPU so should
1237 			 * be assigned the same PE as the existing
1238 			 * peer NPU.
1239 			 */
1240 			dev_info(&npu_pdev->dev,
1241 				"Associating to existing PE %x\n", pe_num);
1242 			pci_dev_get(npu_pdev);
1243 			npu_pdn = pci_get_pdn(npu_pdev);
1244 			rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1245 			npu_pdn->pe_number = pe_num;
1246 			phb->ioda.pe_rmap[rid] = pe->pe_number;
1247 
1248 			/* Map the PE to this link */
1249 			rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1250 					OpalPciBusAll,
1251 					OPAL_COMPARE_RID_DEVICE_NUMBER,
1252 					OPAL_COMPARE_RID_FUNCTION_NUMBER,
1253 					OPAL_MAP_PE);
1254 			WARN_ON(rc != OPAL_SUCCESS);
1255 			found_pe = true;
1256 			break;
1257 		}
1258 	}
1259 
1260 	if (!found_pe)
1261 		/*
1262 		 * Could not find an existing PE so allocate a new
1263 		 * one.
1264 		 */
1265 		return pnv_ioda_setup_dev_PE(npu_pdev);
1266 	else
1267 		return pe;
1268 }
1269 
1270 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1271 {
1272 	struct pci_dev *pdev;
1273 
1274 	list_for_each_entry(pdev, &bus->devices, bus_list)
1275 		pnv_ioda_setup_npu_PE(pdev);
1276 }
1277 
1278 static void pnv_pci_ioda_setup_PEs(void)
1279 {
1280 	struct pci_controller *hose, *tmp;
1281 	struct pnv_phb *phb;
1282 	struct pci_bus *bus;
1283 	struct pci_dev *pdev;
1284 
1285 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1286 		phb = hose->private_data;
1287 		if (phb->type == PNV_PHB_NPU_NVLINK) {
1288 			/* PE#0 is needed for error reporting */
1289 			pnv_ioda_reserve_pe(phb, 0);
1290 			pnv_ioda_setup_npu_PEs(hose->bus);
1291 			if (phb->model == PNV_PHB_MODEL_NPU2)
1292 				pnv_npu2_init(phb);
1293 		}
1294 		if (phb->type == PNV_PHB_NPU_OCAPI) {
1295 			bus = hose->bus;
1296 			list_for_each_entry(pdev, &bus->devices, bus_list)
1297 				pnv_ioda_setup_dev_PE(pdev);
1298 		}
1299 	}
1300 }
1301 
1302 #ifdef CONFIG_PCI_IOV
1303 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1304 {
1305 	struct pci_bus        *bus;
1306 	struct pci_controller *hose;
1307 	struct pnv_phb        *phb;
1308 	struct pci_dn         *pdn;
1309 	int                    i, j;
1310 	int                    m64_bars;
1311 
1312 	bus = pdev->bus;
1313 	hose = pci_bus_to_host(bus);
1314 	phb = hose->private_data;
1315 	pdn = pci_get_pdn(pdev);
1316 
1317 	if (pdn->m64_single_mode)
1318 		m64_bars = num_vfs;
1319 	else
1320 		m64_bars = 1;
1321 
1322 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1323 		for (j = 0; j < m64_bars; j++) {
1324 			if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1325 				continue;
1326 			opal_pci_phb_mmio_enable(phb->opal_id,
1327 				OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1328 			clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1329 			pdn->m64_map[j][i] = IODA_INVALID_M64;
1330 		}
1331 
1332 	kfree(pdn->m64_map);
1333 	return 0;
1334 }
1335 
1336 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1337 {
1338 	struct pci_bus        *bus;
1339 	struct pci_controller *hose;
1340 	struct pnv_phb        *phb;
1341 	struct pci_dn         *pdn;
1342 	unsigned int           win;
1343 	struct resource       *res;
1344 	int                    i, j;
1345 	int64_t                rc;
1346 	int                    total_vfs;
1347 	resource_size_t        size, start;
1348 	int                    pe_num;
1349 	int                    m64_bars;
1350 
1351 	bus = pdev->bus;
1352 	hose = pci_bus_to_host(bus);
1353 	phb = hose->private_data;
1354 	pdn = pci_get_pdn(pdev);
1355 	total_vfs = pci_sriov_get_totalvfs(pdev);
1356 
1357 	if (pdn->m64_single_mode)
1358 		m64_bars = num_vfs;
1359 	else
1360 		m64_bars = 1;
1361 
1362 	pdn->m64_map = kmalloc_array(m64_bars,
1363 				     sizeof(*pdn->m64_map),
1364 				     GFP_KERNEL);
1365 	if (!pdn->m64_map)
1366 		return -ENOMEM;
1367 	/* Initialize the m64_map to IODA_INVALID_M64 */
1368 	for (i = 0; i < m64_bars ; i++)
1369 		for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1370 			pdn->m64_map[i][j] = IODA_INVALID_M64;
1371 
1372 
1373 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1374 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
1375 		if (!res->flags || !res->parent)
1376 			continue;
1377 
1378 		for (j = 0; j < m64_bars; j++) {
1379 			do {
1380 				win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1381 						phb->ioda.m64_bar_idx + 1, 0);
1382 
1383 				if (win >= phb->ioda.m64_bar_idx + 1)
1384 					goto m64_failed;
1385 			} while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1386 
1387 			pdn->m64_map[j][i] = win;
1388 
1389 			if (pdn->m64_single_mode) {
1390 				size = pci_iov_resource_size(pdev,
1391 							PCI_IOV_RESOURCES + i);
1392 				start = res->start + size * j;
1393 			} else {
1394 				size = resource_size(res);
1395 				start = res->start;
1396 			}
1397 
1398 			/* Map the M64 here */
1399 			if (pdn->m64_single_mode) {
1400 				pe_num = pdn->pe_num_map[j];
1401 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1402 						pe_num, OPAL_M64_WINDOW_TYPE,
1403 						pdn->m64_map[j][i], 0);
1404 			}
1405 
1406 			rc = opal_pci_set_phb_mem_window(phb->opal_id,
1407 						 OPAL_M64_WINDOW_TYPE,
1408 						 pdn->m64_map[j][i],
1409 						 start,
1410 						 0, /* unused */
1411 						 size);
1412 
1413 
1414 			if (rc != OPAL_SUCCESS) {
1415 				dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1416 					win, rc);
1417 				goto m64_failed;
1418 			}
1419 
1420 			if (pdn->m64_single_mode)
1421 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1422 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1423 			else
1424 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1425 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1426 
1427 			if (rc != OPAL_SUCCESS) {
1428 				dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1429 					win, rc);
1430 				goto m64_failed;
1431 			}
1432 		}
1433 	}
1434 	return 0;
1435 
1436 m64_failed:
1437 	pnv_pci_vf_release_m64(pdev, num_vfs);
1438 	return -EBUSY;
1439 }
1440 
1441 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1442 		int num);
1443 
1444 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1445 {
1446 	struct iommu_table    *tbl;
1447 	int64_t               rc;
1448 
1449 	tbl = pe->table_group.tables[0];
1450 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1451 	if (rc)
1452 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1453 
1454 	pnv_pci_ioda2_set_bypass(pe, false);
1455 	if (pe->table_group.group) {
1456 		iommu_group_put(pe->table_group.group);
1457 		BUG_ON(pe->table_group.group);
1458 	}
1459 	iommu_tce_table_put(tbl);
1460 }
1461 
1462 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1463 {
1464 	struct pci_bus        *bus;
1465 	struct pci_controller *hose;
1466 	struct pnv_phb        *phb;
1467 	struct pnv_ioda_pe    *pe, *pe_n;
1468 	struct pci_dn         *pdn;
1469 
1470 	bus = pdev->bus;
1471 	hose = pci_bus_to_host(bus);
1472 	phb = hose->private_data;
1473 	pdn = pci_get_pdn(pdev);
1474 
1475 	if (!pdev->is_physfn)
1476 		return;
1477 
1478 	list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1479 		if (pe->parent_dev != pdev)
1480 			continue;
1481 
1482 		pnv_pci_ioda2_release_dma_pe(pdev, pe);
1483 
1484 		/* Remove from list */
1485 		mutex_lock(&phb->ioda.pe_list_mutex);
1486 		list_del(&pe->list);
1487 		mutex_unlock(&phb->ioda.pe_list_mutex);
1488 
1489 		pnv_ioda_deconfigure_pe(phb, pe);
1490 
1491 		pnv_ioda_free_pe(pe);
1492 	}
1493 }
1494 
1495 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1496 {
1497 	struct pci_bus        *bus;
1498 	struct pci_controller *hose;
1499 	struct pnv_phb        *phb;
1500 	struct pnv_ioda_pe    *pe;
1501 	struct pci_dn         *pdn;
1502 	u16                    num_vfs, i;
1503 
1504 	bus = pdev->bus;
1505 	hose = pci_bus_to_host(bus);
1506 	phb = hose->private_data;
1507 	pdn = pci_get_pdn(pdev);
1508 	num_vfs = pdn->num_vfs;
1509 
1510 	/* Release VF PEs */
1511 	pnv_ioda_release_vf_PE(pdev);
1512 
1513 	if (phb->type == PNV_PHB_IODA2) {
1514 		if (!pdn->m64_single_mode)
1515 			pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1516 
1517 		/* Release M64 windows */
1518 		pnv_pci_vf_release_m64(pdev, num_vfs);
1519 
1520 		/* Release PE numbers */
1521 		if (pdn->m64_single_mode) {
1522 			for (i = 0; i < num_vfs; i++) {
1523 				if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1524 					continue;
1525 
1526 				pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1527 				pnv_ioda_free_pe(pe);
1528 			}
1529 		} else
1530 			bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1531 		/* Releasing pe_num_map */
1532 		kfree(pdn->pe_num_map);
1533 	}
1534 }
1535 
1536 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1537 				       struct pnv_ioda_pe *pe);
1538 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1539 {
1540 	struct pci_bus        *bus;
1541 	struct pci_controller *hose;
1542 	struct pnv_phb        *phb;
1543 	struct pnv_ioda_pe    *pe;
1544 	int                    pe_num;
1545 	u16                    vf_index;
1546 	struct pci_dn         *pdn;
1547 
1548 	bus = pdev->bus;
1549 	hose = pci_bus_to_host(bus);
1550 	phb = hose->private_data;
1551 	pdn = pci_get_pdn(pdev);
1552 
1553 	if (!pdev->is_physfn)
1554 		return;
1555 
1556 	/* Reserve PE for each VF */
1557 	for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1558 		if (pdn->m64_single_mode)
1559 			pe_num = pdn->pe_num_map[vf_index];
1560 		else
1561 			pe_num = *pdn->pe_num_map + vf_index;
1562 
1563 		pe = &phb->ioda.pe_array[pe_num];
1564 		pe->pe_number = pe_num;
1565 		pe->phb = phb;
1566 		pe->flags = PNV_IODA_PE_VF;
1567 		pe->pbus = NULL;
1568 		pe->parent_dev = pdev;
1569 		pe->mve_number = -1;
1570 		pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1571 			   pci_iov_virtfn_devfn(pdev, vf_index);
1572 
1573 		pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%x\n",
1574 			hose->global_number, pdev->bus->number,
1575 			PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1576 			PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1577 
1578 		if (pnv_ioda_configure_pe(phb, pe)) {
1579 			/* XXX What do we do here ? */
1580 			pnv_ioda_free_pe(pe);
1581 			pe->pdev = NULL;
1582 			continue;
1583 		}
1584 
1585 		/* Put PE to the list */
1586 		mutex_lock(&phb->ioda.pe_list_mutex);
1587 		list_add_tail(&pe->list, &phb->ioda.pe_list);
1588 		mutex_unlock(&phb->ioda.pe_list_mutex);
1589 
1590 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
1591 	}
1592 }
1593 
1594 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1595 {
1596 	struct pci_bus        *bus;
1597 	struct pci_controller *hose;
1598 	struct pnv_phb        *phb;
1599 	struct pnv_ioda_pe    *pe;
1600 	struct pci_dn         *pdn;
1601 	int                    ret;
1602 	u16                    i;
1603 
1604 	bus = pdev->bus;
1605 	hose = pci_bus_to_host(bus);
1606 	phb = hose->private_data;
1607 	pdn = pci_get_pdn(pdev);
1608 
1609 	if (phb->type == PNV_PHB_IODA2) {
1610 		if (!pdn->vfs_expanded) {
1611 			dev_info(&pdev->dev, "don't support this SRIOV device"
1612 				" with non 64bit-prefetchable IOV BAR\n");
1613 			return -ENOSPC;
1614 		}
1615 
1616 		/*
1617 		 * When M64 BARs functions in Single PE mode, the number of VFs
1618 		 * could be enabled must be less than the number of M64 BARs.
1619 		 */
1620 		if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1621 			dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1622 			return -EBUSY;
1623 		}
1624 
1625 		/* Allocating pe_num_map */
1626 		if (pdn->m64_single_mode)
1627 			pdn->pe_num_map = kmalloc_array(num_vfs,
1628 							sizeof(*pdn->pe_num_map),
1629 							GFP_KERNEL);
1630 		else
1631 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1632 
1633 		if (!pdn->pe_num_map)
1634 			return -ENOMEM;
1635 
1636 		if (pdn->m64_single_mode)
1637 			for (i = 0; i < num_vfs; i++)
1638 				pdn->pe_num_map[i] = IODA_INVALID_PE;
1639 
1640 		/* Calculate available PE for required VFs */
1641 		if (pdn->m64_single_mode) {
1642 			for (i = 0; i < num_vfs; i++) {
1643 				pe = pnv_ioda_alloc_pe(phb);
1644 				if (!pe) {
1645 					ret = -EBUSY;
1646 					goto m64_failed;
1647 				}
1648 
1649 				pdn->pe_num_map[i] = pe->pe_number;
1650 			}
1651 		} else {
1652 			mutex_lock(&phb->ioda.pe_alloc_mutex);
1653 			*pdn->pe_num_map = bitmap_find_next_zero_area(
1654 				phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1655 				0, num_vfs, 0);
1656 			if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1657 				mutex_unlock(&phb->ioda.pe_alloc_mutex);
1658 				dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1659 				kfree(pdn->pe_num_map);
1660 				return -EBUSY;
1661 			}
1662 			bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1663 			mutex_unlock(&phb->ioda.pe_alloc_mutex);
1664 		}
1665 		pdn->num_vfs = num_vfs;
1666 
1667 		/* Assign M64 window accordingly */
1668 		ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1669 		if (ret) {
1670 			dev_info(&pdev->dev, "Not enough M64 window resources\n");
1671 			goto m64_failed;
1672 		}
1673 
1674 		/*
1675 		 * When using one M64 BAR to map one IOV BAR, we need to shift
1676 		 * the IOV BAR according to the PE# allocated to the VFs.
1677 		 * Otherwise, the PE# for the VF will conflict with others.
1678 		 */
1679 		if (!pdn->m64_single_mode) {
1680 			ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1681 			if (ret)
1682 				goto m64_failed;
1683 		}
1684 	}
1685 
1686 	/* Setup VF PEs */
1687 	pnv_ioda_setup_vf_PE(pdev, num_vfs);
1688 
1689 	return 0;
1690 
1691 m64_failed:
1692 	if (pdn->m64_single_mode) {
1693 		for (i = 0; i < num_vfs; i++) {
1694 			if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1695 				continue;
1696 
1697 			pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1698 			pnv_ioda_free_pe(pe);
1699 		}
1700 	} else
1701 		bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1702 
1703 	/* Releasing pe_num_map */
1704 	kfree(pdn->pe_num_map);
1705 
1706 	return ret;
1707 }
1708 
1709 int pnv_pcibios_sriov_disable(struct pci_dev *pdev)
1710 {
1711 	pnv_pci_sriov_disable(pdev);
1712 
1713 	/* Release PCI data */
1714 	remove_dev_pci_data(pdev);
1715 	return 0;
1716 }
1717 
1718 int pnv_pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1719 {
1720 	/* Allocate PCI data */
1721 	add_dev_pci_data(pdev);
1722 
1723 	return pnv_pci_sriov_enable(pdev, num_vfs);
1724 }
1725 #endif /* CONFIG_PCI_IOV */
1726 
1727 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1728 {
1729 	struct pci_dn *pdn = pci_get_pdn(pdev);
1730 	struct pnv_ioda_pe *pe;
1731 
1732 	/*
1733 	 * The function can be called while the PE#
1734 	 * hasn't been assigned. Do nothing for the
1735 	 * case.
1736 	 */
1737 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1738 		return;
1739 
1740 	pe = &phb->ioda.pe_array[pdn->pe_number];
1741 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1742 	set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1743 	set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1744 	/*
1745 	 * Note: iommu_add_device() will fail here as
1746 	 * for physical PE: the device is already added by now;
1747 	 * for virtual PE: sysfs entries are not ready yet and
1748 	 * tce_iommu_bus_notifier will add the device to a group later.
1749 	 */
1750 }
1751 
1752 static bool pnv_pci_ioda_pe_single_vendor(struct pnv_ioda_pe *pe)
1753 {
1754 	unsigned short vendor = 0;
1755 	struct pci_dev *pdev;
1756 
1757 	if (pe->device_count == 1)
1758 		return true;
1759 
1760 	/* pe->pdev should be set if it's a single device, pe->pbus if not */
1761 	if (!pe->pbus)
1762 		return true;
1763 
1764 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
1765 		if (!vendor) {
1766 			vendor = pdev->vendor;
1767 			continue;
1768 		}
1769 
1770 		if (pdev->vendor != vendor)
1771 			return false;
1772 	}
1773 
1774 	return true;
1775 }
1776 
1777 /*
1778  * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1779  *
1780  * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1781  * Devices can only access more than that if bit 59 of the PCI address is set
1782  * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1783  * Many PCI devices are not capable of addressing that many bits, and as a
1784  * result are limited to the 4GB of virtual memory made available to 32-bit
1785  * devices in TVE#0.
1786  *
1787  * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1788  * devices by configuring the virtual memory past the first 4GB inaccessible
1789  * by 64-bit DMAs.  This should only be used by devices that want more than
1790  * 4GB, and only on PEs that have no 32-bit devices.
1791  *
1792  * Currently this will only work on PHB3 (POWER8).
1793  */
1794 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe)
1795 {
1796 	u64 window_size, table_size, tce_count, addr;
1797 	struct page *table_pages;
1798 	u64 tce_order = 28; /* 256MB TCEs */
1799 	__be64 *tces;
1800 	s64 rc;
1801 
1802 	/*
1803 	 * Window size needs to be a power of two, but needs to account for
1804 	 * shifting memory by the 4GB offset required to skip 32bit space.
1805 	 */
1806 	window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1807 	tce_count = window_size >> tce_order;
1808 	table_size = tce_count << 3;
1809 
1810 	if (table_size < PAGE_SIZE)
1811 		table_size = PAGE_SIZE;
1812 
1813 	table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL,
1814 				       get_order(table_size));
1815 	if (!table_pages)
1816 		goto err;
1817 
1818 	tces = page_address(table_pages);
1819 	if (!tces)
1820 		goto err;
1821 
1822 	memset(tces, 0, table_size);
1823 
1824 	for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) {
1825 		tces[(addr + (1ULL << 32)) >> tce_order] =
1826 			cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE);
1827 	}
1828 
1829 	rc = opal_pci_map_pe_dma_window(pe->phb->opal_id,
1830 					pe->pe_number,
1831 					/* reconfigure window 0 */
1832 					(pe->pe_number << 1) + 0,
1833 					1,
1834 					__pa(tces),
1835 					table_size,
1836 					1 << tce_order);
1837 	if (rc == OPAL_SUCCESS) {
1838 		pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1839 		return 0;
1840 	}
1841 err:
1842 	pe_err(pe, "Error configuring 64-bit DMA bypass\n");
1843 	return -EIO;
1844 }
1845 
1846 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1847 {
1848 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1849 	struct pnv_phb *phb = hose->private_data;
1850 	struct pci_dn *pdn = pci_get_pdn(pdev);
1851 	struct pnv_ioda_pe *pe;
1852 	uint64_t top;
1853 	bool bypass = false;
1854 	s64 rc;
1855 
1856 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1857 		return -ENODEV;;
1858 
1859 	pe = &phb->ioda.pe_array[pdn->pe_number];
1860 	if (pe->tce_bypass_enabled) {
1861 		top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1862 		bypass = (dma_mask >= top);
1863 	}
1864 
1865 	if (bypass) {
1866 		dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1867 		set_dma_ops(&pdev->dev, &dma_nommu_ops);
1868 	} else {
1869 		/*
1870 		 * If the device can't set the TCE bypass bit but still wants
1871 		 * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1872 		 * bypass the 32-bit region and be usable for 64-bit DMAs.
1873 		 * The device needs to be able to address all of this space.
1874 		 */
1875 		if (dma_mask >> 32 &&
1876 		    dma_mask > (memory_hotplug_max() + (1ULL << 32)) &&
1877 		    pnv_pci_ioda_pe_single_vendor(pe) &&
1878 		    phb->model == PNV_PHB_MODEL_PHB3) {
1879 			/* Configure the bypass mode */
1880 			rc = pnv_pci_ioda_dma_64bit_bypass(pe);
1881 			if (rc)
1882 				return rc;
1883 			/* 4GB offset bypasses 32-bit space */
1884 			set_dma_offset(&pdev->dev, (1ULL << 32));
1885 			set_dma_ops(&pdev->dev, &dma_nommu_ops);
1886 		} else if (dma_mask >> 32 && dma_mask != DMA_BIT_MASK(64)) {
1887 			/*
1888 			 * Fail the request if a DMA mask between 32 and 64 bits
1889 			 * was requested but couldn't be fulfilled. Ideally we
1890 			 * would do this for 64-bits but historically we have
1891 			 * always fallen back to 32-bits.
1892 			 */
1893 			return -ENOMEM;
1894 		} else {
1895 			dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1896 			set_dma_ops(&pdev->dev, &dma_iommu_ops);
1897 		}
1898 	}
1899 	*pdev->dev.dma_mask = dma_mask;
1900 
1901 	/* Update peer npu devices */
1902 	pnv_npu_try_dma_set_bypass(pdev, bypass);
1903 
1904 	return 0;
1905 }
1906 
1907 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev)
1908 {
1909 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1910 	struct pnv_phb *phb = hose->private_data;
1911 	struct pci_dn *pdn = pci_get_pdn(pdev);
1912 	struct pnv_ioda_pe *pe;
1913 	u64 end, mask;
1914 
1915 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1916 		return 0;
1917 
1918 	pe = &phb->ioda.pe_array[pdn->pe_number];
1919 	if (!pe->tce_bypass_enabled)
1920 		return __dma_get_required_mask(&pdev->dev);
1921 
1922 
1923 	end = pe->tce_bypass_base + memblock_end_of_DRAM();
1924 	mask = 1ULL << (fls64(end) - 1);
1925 	mask += mask - 1;
1926 
1927 	return mask;
1928 }
1929 
1930 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1931 				   struct pci_bus *bus,
1932 				   bool add_to_group)
1933 {
1934 	struct pci_dev *dev;
1935 
1936 	list_for_each_entry(dev, &bus->devices, bus_list) {
1937 		set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1938 		set_dma_offset(&dev->dev, pe->tce_bypass_base);
1939 		if (add_to_group)
1940 			iommu_add_device(&dev->dev);
1941 
1942 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1943 			pnv_ioda_setup_bus_dma(pe, dev->subordinate,
1944 					add_to_group);
1945 	}
1946 }
1947 
1948 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1949 						     bool real_mode)
1950 {
1951 	return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1952 		(phb->regs + 0x210);
1953 }
1954 
1955 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1956 		unsigned long index, unsigned long npages, bool rm)
1957 {
1958 	struct iommu_table_group_link *tgl = list_first_entry_or_null(
1959 			&tbl->it_group_list, struct iommu_table_group_link,
1960 			next);
1961 	struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1962 			struct pnv_ioda_pe, table_group);
1963 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1964 	unsigned long start, end, inc;
1965 
1966 	start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1967 	end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1968 			npages - 1);
1969 
1970 	/* p7ioc-style invalidation, 2 TCEs per write */
1971 	start |= (1ull << 63);
1972 	end |= (1ull << 63);
1973 	inc = 16;
1974         end |= inc - 1;	/* round up end to be different than start */
1975 
1976         mb(); /* Ensure above stores are visible */
1977         while (start <= end) {
1978 		if (rm)
1979 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1980 		else
1981 			__raw_writeq(cpu_to_be64(start), invalidate);
1982                 start += inc;
1983         }
1984 
1985 	/*
1986 	 * The iommu layer will do another mb() for us on build()
1987 	 * and we don't care on free()
1988 	 */
1989 }
1990 
1991 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1992 		long npages, unsigned long uaddr,
1993 		enum dma_data_direction direction,
1994 		unsigned long attrs)
1995 {
1996 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1997 			attrs);
1998 
1999 	if (!ret)
2000 		pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
2001 
2002 	return ret;
2003 }
2004 
2005 #ifdef CONFIG_IOMMU_API
2006 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
2007 		unsigned long *hpa, enum dma_data_direction *direction)
2008 {
2009 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
2010 
2011 	if (!ret)
2012 		pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, false);
2013 
2014 	return ret;
2015 }
2016 
2017 static int pnv_ioda1_tce_xchg_rm(struct iommu_table *tbl, long index,
2018 		unsigned long *hpa, enum dma_data_direction *direction)
2019 {
2020 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
2021 
2022 	if (!ret)
2023 		pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, true);
2024 
2025 	return ret;
2026 }
2027 #endif
2028 
2029 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
2030 		long npages)
2031 {
2032 	pnv_tce_free(tbl, index, npages);
2033 
2034 	pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
2035 }
2036 
2037 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
2038 	.set = pnv_ioda1_tce_build,
2039 #ifdef CONFIG_IOMMU_API
2040 	.exchange = pnv_ioda1_tce_xchg,
2041 	.exchange_rm = pnv_ioda1_tce_xchg_rm,
2042 #endif
2043 	.clear = pnv_ioda1_tce_free,
2044 	.get = pnv_tce_get,
2045 };
2046 
2047 #define PHB3_TCE_KILL_INVAL_ALL		PPC_BIT(0)
2048 #define PHB3_TCE_KILL_INVAL_PE		PPC_BIT(1)
2049 #define PHB3_TCE_KILL_INVAL_ONE		PPC_BIT(2)
2050 
2051 static void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
2052 {
2053 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(phb, rm);
2054 	const unsigned long val = PHB3_TCE_KILL_INVAL_ALL;
2055 
2056 	mb(); /* Ensure previous TCE table stores are visible */
2057 	if (rm)
2058 		__raw_rm_writeq(cpu_to_be64(val), invalidate);
2059 	else
2060 		__raw_writeq(cpu_to_be64(val), invalidate);
2061 }
2062 
2063 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
2064 {
2065 	/* 01xb - invalidate TCEs that match the specified PE# */
2066 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
2067 	unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
2068 
2069 	mb(); /* Ensure above stores are visible */
2070 	__raw_writeq(cpu_to_be64(val), invalidate);
2071 }
2072 
2073 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
2074 					unsigned shift, unsigned long index,
2075 					unsigned long npages)
2076 {
2077 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
2078 	unsigned long start, end, inc;
2079 
2080 	/* We'll invalidate DMA address in PE scope */
2081 	start = PHB3_TCE_KILL_INVAL_ONE;
2082 	start |= (pe->pe_number & 0xFF);
2083 	end = start;
2084 
2085 	/* Figure out the start, end and step */
2086 	start |= (index << shift);
2087 	end |= ((index + npages - 1) << shift);
2088 	inc = (0x1ull << shift);
2089 	mb();
2090 
2091 	while (start <= end) {
2092 		if (rm)
2093 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
2094 		else
2095 			__raw_writeq(cpu_to_be64(start), invalidate);
2096 		start += inc;
2097 	}
2098 }
2099 
2100 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
2101 {
2102 	struct pnv_phb *phb = pe->phb;
2103 
2104 	if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
2105 		pnv_pci_phb3_tce_invalidate_pe(pe);
2106 	else
2107 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
2108 				  pe->pe_number, 0, 0, 0);
2109 }
2110 
2111 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
2112 		unsigned long index, unsigned long npages, bool rm)
2113 {
2114 	struct iommu_table_group_link *tgl;
2115 
2116 	list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) {
2117 		struct pnv_ioda_pe *pe = container_of(tgl->table_group,
2118 				struct pnv_ioda_pe, table_group);
2119 		struct pnv_phb *phb = pe->phb;
2120 		unsigned int shift = tbl->it_page_shift;
2121 
2122 		/*
2123 		 * NVLink1 can use the TCE kill register directly as
2124 		 * it's the same as PHB3. NVLink2 is different and
2125 		 * should go via the OPAL call.
2126 		 */
2127 		if (phb->model == PNV_PHB_MODEL_NPU) {
2128 			/*
2129 			 * The NVLink hardware does not support TCE kill
2130 			 * per TCE entry so we have to invalidate
2131 			 * the entire cache for it.
2132 			 */
2133 			pnv_pci_phb3_tce_invalidate_entire(phb, rm);
2134 			continue;
2135 		}
2136 		if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
2137 			pnv_pci_phb3_tce_invalidate(pe, rm, shift,
2138 						    index, npages);
2139 		else
2140 			opal_pci_tce_kill(phb->opal_id,
2141 					  OPAL_PCI_TCE_KILL_PAGES,
2142 					  pe->pe_number, 1u << shift,
2143 					  index << shift, npages);
2144 	}
2145 }
2146 
2147 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
2148 {
2149 	if (phb->model == PNV_PHB_MODEL_NPU || phb->model == PNV_PHB_MODEL_PHB3)
2150 		pnv_pci_phb3_tce_invalidate_entire(phb, rm);
2151 	else
2152 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL, 0, 0, 0, 0);
2153 }
2154 
2155 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
2156 		long npages, unsigned long uaddr,
2157 		enum dma_data_direction direction,
2158 		unsigned long attrs)
2159 {
2160 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
2161 			attrs);
2162 
2163 	if (!ret)
2164 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
2165 
2166 	return ret;
2167 }
2168 
2169 #ifdef CONFIG_IOMMU_API
2170 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
2171 		unsigned long *hpa, enum dma_data_direction *direction)
2172 {
2173 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
2174 
2175 	if (!ret)
2176 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
2177 
2178 	return ret;
2179 }
2180 
2181 static int pnv_ioda2_tce_xchg_rm(struct iommu_table *tbl, long index,
2182 		unsigned long *hpa, enum dma_data_direction *direction)
2183 {
2184 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
2185 
2186 	if (!ret)
2187 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, true);
2188 
2189 	return ret;
2190 }
2191 #endif
2192 
2193 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
2194 		long npages)
2195 {
2196 	pnv_tce_free(tbl, index, npages);
2197 
2198 	pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
2199 }
2200 
2201 static void pnv_ioda2_table_free(struct iommu_table *tbl)
2202 {
2203 	pnv_pci_ioda2_table_free_pages(tbl);
2204 }
2205 
2206 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
2207 	.set = pnv_ioda2_tce_build,
2208 #ifdef CONFIG_IOMMU_API
2209 	.exchange = pnv_ioda2_tce_xchg,
2210 	.exchange_rm = pnv_ioda2_tce_xchg_rm,
2211 #endif
2212 	.clear = pnv_ioda2_tce_free,
2213 	.get = pnv_tce_get,
2214 	.free = pnv_ioda2_table_free,
2215 };
2216 
2217 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
2218 {
2219 	unsigned int *weight = (unsigned int *)data;
2220 
2221 	/* This is quite simplistic. The "base" weight of a device
2222 	 * is 10. 0 means no DMA is to be accounted for it.
2223 	 */
2224 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2225 		return 0;
2226 
2227 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
2228 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
2229 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
2230 		*weight += 3;
2231 	else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
2232 		*weight += 15;
2233 	else
2234 		*weight += 10;
2235 
2236 	return 0;
2237 }
2238 
2239 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
2240 {
2241 	unsigned int weight = 0;
2242 
2243 	/* SRIOV VF has same DMA32 weight as its PF */
2244 #ifdef CONFIG_PCI_IOV
2245 	if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
2246 		pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
2247 		return weight;
2248 	}
2249 #endif
2250 
2251 	if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
2252 		pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
2253 	} else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
2254 		struct pci_dev *pdev;
2255 
2256 		list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
2257 			pnv_pci_ioda_dev_dma_weight(pdev, &weight);
2258 	} else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2259 		pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2260 	}
2261 
2262 	return weight;
2263 }
2264 
2265 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2266 				       struct pnv_ioda_pe *pe)
2267 {
2268 
2269 	struct page *tce_mem = NULL;
2270 	struct iommu_table *tbl;
2271 	unsigned int weight, total_weight = 0;
2272 	unsigned int tce32_segsz, base, segs, avail, i;
2273 	int64_t rc;
2274 	void *addr;
2275 
2276 	/* XXX FIXME: Handle 64-bit only DMA devices */
2277 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2278 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
2279 	weight = pnv_pci_ioda_pe_dma_weight(pe);
2280 	if (!weight)
2281 		return;
2282 
2283 	pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2284 		     &total_weight);
2285 	segs = (weight * phb->ioda.dma32_count) / total_weight;
2286 	if (!segs)
2287 		segs = 1;
2288 
2289 	/*
2290 	 * Allocate contiguous DMA32 segments. We begin with the expected
2291 	 * number of segments. With one more attempt, the number of DMA32
2292 	 * segments to be allocated is decreased by one until one segment
2293 	 * is allocated successfully.
2294 	 */
2295 	do {
2296 		for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2297 			for (avail = 0, i = base; i < base + segs; i++) {
2298 				if (phb->ioda.dma32_segmap[i] ==
2299 				    IODA_INVALID_PE)
2300 					avail++;
2301 			}
2302 
2303 			if (avail == segs)
2304 				goto found;
2305 		}
2306 	} while (--segs);
2307 
2308 	if (!segs) {
2309 		pe_warn(pe, "No available DMA32 segments\n");
2310 		return;
2311 	}
2312 
2313 found:
2314 	tbl = pnv_pci_table_alloc(phb->hose->node);
2315 	if (WARN_ON(!tbl))
2316 		return;
2317 
2318 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2319 			pe->pe_number);
2320 	pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2321 
2322 	/* Grab a 32-bit TCE table */
2323 	pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2324 		weight, total_weight, base, segs);
2325 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2326 		base * PNV_IODA1_DMA32_SEGSIZE,
2327 		(base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2328 
2329 	/* XXX Currently, we allocate one big contiguous table for the
2330 	 * TCEs. We only really need one chunk per 256M of TCE space
2331 	 * (ie per segment) but that's an optimization for later, it
2332 	 * requires some added smarts with our get/put_tce implementation
2333 	 *
2334 	 * Each TCE page is 4KB in size and each TCE entry occupies 8
2335 	 * bytes
2336 	 */
2337 	tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2338 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2339 				   get_order(tce32_segsz * segs));
2340 	if (!tce_mem) {
2341 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2342 		goto fail;
2343 	}
2344 	addr = page_address(tce_mem);
2345 	memset(addr, 0, tce32_segsz * segs);
2346 
2347 	/* Configure HW */
2348 	for (i = 0; i < segs; i++) {
2349 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
2350 					      pe->pe_number,
2351 					      base + i, 1,
2352 					      __pa(addr) + tce32_segsz * i,
2353 					      tce32_segsz, IOMMU_PAGE_SIZE_4K);
2354 		if (rc) {
2355 			pe_err(pe, " Failed to configure 32-bit TCE table,"
2356 			       " err %ld\n", rc);
2357 			goto fail;
2358 		}
2359 	}
2360 
2361 	/* Setup DMA32 segment mapping */
2362 	for (i = base; i < base + segs; i++)
2363 		phb->ioda.dma32_segmap[i] = pe->pe_number;
2364 
2365 	/* Setup linux iommu table */
2366 	pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2367 				  base * PNV_IODA1_DMA32_SEGSIZE,
2368 				  IOMMU_PAGE_SHIFT_4K);
2369 
2370 	tbl->it_ops = &pnv_ioda1_iommu_ops;
2371 	pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2372 	pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2373 	iommu_init_table(tbl, phb->hose->node);
2374 
2375 	if (pe->flags & PNV_IODA_PE_DEV) {
2376 		/*
2377 		 * Setting table base here only for carrying iommu_group
2378 		 * further down to let iommu_add_device() do the job.
2379 		 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2380 		 */
2381 		set_iommu_table_base(&pe->pdev->dev, tbl);
2382 		iommu_add_device(&pe->pdev->dev);
2383 	} else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2384 		pnv_ioda_setup_bus_dma(pe, pe->pbus, true);
2385 
2386 	return;
2387  fail:
2388 	/* XXX Failure: Try to fallback to 64-bit only ? */
2389 	if (tce_mem)
2390 		__free_pages(tce_mem, get_order(tce32_segsz * segs));
2391 	if (tbl) {
2392 		pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2393 		iommu_tce_table_put(tbl);
2394 	}
2395 }
2396 
2397 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2398 		int num, struct iommu_table *tbl)
2399 {
2400 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2401 			table_group);
2402 	struct pnv_phb *phb = pe->phb;
2403 	int64_t rc;
2404 	const unsigned long size = tbl->it_indirect_levels ?
2405 			tbl->it_level_size : tbl->it_size;
2406 	const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2407 	const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2408 
2409 	pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num,
2410 			start_addr, start_addr + win_size - 1,
2411 			IOMMU_PAGE_SIZE(tbl));
2412 
2413 	/*
2414 	 * Map TCE table through TVT. The TVE index is the PE number
2415 	 * shifted by 1 bit for 32-bits DMA space.
2416 	 */
2417 	rc = opal_pci_map_pe_dma_window(phb->opal_id,
2418 			pe->pe_number,
2419 			(pe->pe_number << 1) + num,
2420 			tbl->it_indirect_levels + 1,
2421 			__pa(tbl->it_base),
2422 			size << 3,
2423 			IOMMU_PAGE_SIZE(tbl));
2424 	if (rc) {
2425 		pe_err(pe, "Failed to configure TCE table, err %ld\n", rc);
2426 		return rc;
2427 	}
2428 
2429 	pnv_pci_link_table_and_group(phb->hose->node, num,
2430 			tbl, &pe->table_group);
2431 	pnv_pci_ioda2_tce_invalidate_pe(pe);
2432 
2433 	return 0;
2434 }
2435 
2436 void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2437 {
2438 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
2439 	int64_t rc;
2440 
2441 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2442 	if (enable) {
2443 		phys_addr_t top = memblock_end_of_DRAM();
2444 
2445 		top = roundup_pow_of_two(top);
2446 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2447 						     pe->pe_number,
2448 						     window_id,
2449 						     pe->tce_bypass_base,
2450 						     top);
2451 	} else {
2452 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2453 						     pe->pe_number,
2454 						     window_id,
2455 						     pe->tce_bypass_base,
2456 						     0);
2457 	}
2458 	if (rc)
2459 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2460 	else
2461 		pe->tce_bypass_enabled = enable;
2462 }
2463 
2464 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2465 		__u32 page_shift, __u64 window_size, __u32 levels,
2466 		struct iommu_table *tbl);
2467 
2468 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2469 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
2470 		struct iommu_table **ptbl)
2471 {
2472 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2473 			table_group);
2474 	int nid = pe->phb->hose->node;
2475 	__u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2476 	long ret;
2477 	struct iommu_table *tbl;
2478 
2479 	tbl = pnv_pci_table_alloc(nid);
2480 	if (!tbl)
2481 		return -ENOMEM;
2482 
2483 	tbl->it_ops = &pnv_ioda2_iommu_ops;
2484 
2485 	ret = pnv_pci_ioda2_table_alloc_pages(nid,
2486 			bus_offset, page_shift, window_size,
2487 			levels, tbl);
2488 	if (ret) {
2489 		iommu_tce_table_put(tbl);
2490 		return ret;
2491 	}
2492 
2493 	*ptbl = tbl;
2494 
2495 	return 0;
2496 }
2497 
2498 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2499 {
2500 	struct iommu_table *tbl = NULL;
2501 	long rc;
2502 
2503 	/*
2504 	 * crashkernel= specifies the kdump kernel's maximum memory at
2505 	 * some offset and there is no guaranteed the result is a power
2506 	 * of 2, which will cause errors later.
2507 	 */
2508 	const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2509 
2510 	/*
2511 	 * In memory constrained environments, e.g. kdump kernel, the
2512 	 * DMA window can be larger than available memory, which will
2513 	 * cause errors later.
2514 	 */
2515 	const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2516 
2517 	rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2518 			IOMMU_PAGE_SHIFT_4K,
2519 			window_size,
2520 			POWERNV_IOMMU_DEFAULT_LEVELS, &tbl);
2521 	if (rc) {
2522 		pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2523 				rc);
2524 		return rc;
2525 	}
2526 
2527 	iommu_init_table(tbl, pe->phb->hose->node);
2528 
2529 	rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2530 	if (rc) {
2531 		pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2532 				rc);
2533 		iommu_tce_table_put(tbl);
2534 		return rc;
2535 	}
2536 
2537 	if (!pnv_iommu_bypass_disabled)
2538 		pnv_pci_ioda2_set_bypass(pe, true);
2539 
2540 	/*
2541 	 * Setting table base here only for carrying iommu_group
2542 	 * further down to let iommu_add_device() do the job.
2543 	 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2544 	 */
2545 	if (pe->flags & PNV_IODA_PE_DEV)
2546 		set_iommu_table_base(&pe->pdev->dev, tbl);
2547 
2548 	return 0;
2549 }
2550 
2551 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2552 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2553 		int num)
2554 {
2555 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2556 			table_group);
2557 	struct pnv_phb *phb = pe->phb;
2558 	long ret;
2559 
2560 	pe_info(pe, "Removing DMA window #%d\n", num);
2561 
2562 	ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2563 			(pe->pe_number << 1) + num,
2564 			0/* levels */, 0/* table address */,
2565 			0/* table size */, 0/* page size */);
2566 	if (ret)
2567 		pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2568 	else
2569 		pnv_pci_ioda2_tce_invalidate_pe(pe);
2570 
2571 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2572 
2573 	return ret;
2574 }
2575 #endif
2576 
2577 #ifdef CONFIG_IOMMU_API
2578 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2579 		__u64 window_size, __u32 levels)
2580 {
2581 	unsigned long bytes = 0;
2582 	const unsigned window_shift = ilog2(window_size);
2583 	unsigned entries_shift = window_shift - page_shift;
2584 	unsigned table_shift = entries_shift + 3;
2585 	unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2586 	unsigned long direct_table_size;
2587 
2588 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2589 			!is_power_of_2(window_size))
2590 		return 0;
2591 
2592 	/* Calculate a direct table size from window_size and levels */
2593 	entries_shift = (entries_shift + levels - 1) / levels;
2594 	table_shift = entries_shift + 3;
2595 	table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2596 	direct_table_size =  1UL << table_shift;
2597 
2598 	for ( ; levels; --levels) {
2599 		bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2600 
2601 		tce_table_size /= direct_table_size;
2602 		tce_table_size <<= 3;
2603 		tce_table_size = max_t(unsigned long,
2604 				tce_table_size, direct_table_size);
2605 	}
2606 
2607 	return bytes;
2608 }
2609 
2610 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2611 {
2612 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2613 						table_group);
2614 	/* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2615 	struct iommu_table *tbl = pe->table_group.tables[0];
2616 
2617 	pnv_pci_ioda2_set_bypass(pe, false);
2618 	pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2619 	if (pe->pbus)
2620 		pnv_ioda_setup_bus_dma(pe, pe->pbus, false);
2621 	iommu_tce_table_put(tbl);
2622 }
2623 
2624 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2625 {
2626 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2627 						table_group);
2628 
2629 	pnv_pci_ioda2_setup_default_config(pe);
2630 	if (pe->pbus)
2631 		pnv_ioda_setup_bus_dma(pe, pe->pbus, false);
2632 }
2633 
2634 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2635 	.get_table_size = pnv_pci_ioda2_get_table_size,
2636 	.create_table = pnv_pci_ioda2_create_table,
2637 	.set_window = pnv_pci_ioda2_set_window,
2638 	.unset_window = pnv_pci_ioda2_unset_window,
2639 	.take_ownership = pnv_ioda2_take_ownership,
2640 	.release_ownership = pnv_ioda2_release_ownership,
2641 };
2642 
2643 static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque)
2644 {
2645 	struct pci_controller *hose;
2646 	struct pnv_phb *phb;
2647 	struct pnv_ioda_pe **ptmppe = opaque;
2648 	struct pci_dev *pdev = container_of(dev, struct pci_dev, dev);
2649 	struct pci_dn *pdn = pci_get_pdn(pdev);
2650 
2651 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2652 		return 0;
2653 
2654 	hose = pci_bus_to_host(pdev->bus);
2655 	phb = hose->private_data;
2656 	if (phb->type != PNV_PHB_NPU_NVLINK)
2657 		return 0;
2658 
2659 	*ptmppe = &phb->ioda.pe_array[pdn->pe_number];
2660 
2661 	return 1;
2662 }
2663 
2664 /*
2665  * This returns PE of associated NPU.
2666  * This assumes that NPU is in the same IOMMU group with GPU and there is
2667  * no other PEs.
2668  */
2669 static struct pnv_ioda_pe *gpe_table_group_to_npe(
2670 		struct iommu_table_group *table_group)
2671 {
2672 	struct pnv_ioda_pe *npe = NULL;
2673 	int ret = iommu_group_for_each_dev(table_group->group, &npe,
2674 			gpe_table_group_to_npe_cb);
2675 
2676 	BUG_ON(!ret || !npe);
2677 
2678 	return npe;
2679 }
2680 
2681 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group *table_group,
2682 		int num, struct iommu_table *tbl)
2683 {
2684 	long ret = pnv_pci_ioda2_set_window(table_group, num, tbl);
2685 
2686 	if (ret)
2687 		return ret;
2688 
2689 	ret = pnv_npu_set_window(gpe_table_group_to_npe(table_group), num, tbl);
2690 	if (ret)
2691 		pnv_pci_ioda2_unset_window(table_group, num);
2692 
2693 	return ret;
2694 }
2695 
2696 static long pnv_pci_ioda2_npu_unset_window(
2697 		struct iommu_table_group *table_group,
2698 		int num)
2699 {
2700 	long ret = pnv_pci_ioda2_unset_window(table_group, num);
2701 
2702 	if (ret)
2703 		return ret;
2704 
2705 	return pnv_npu_unset_window(gpe_table_group_to_npe(table_group), num);
2706 }
2707 
2708 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group *table_group)
2709 {
2710 	/*
2711 	 * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2712 	 * the iommu_table if 32bit DMA is enabled.
2713 	 */
2714 	pnv_npu_take_ownership(gpe_table_group_to_npe(table_group));
2715 	pnv_ioda2_take_ownership(table_group);
2716 }
2717 
2718 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = {
2719 	.get_table_size = pnv_pci_ioda2_get_table_size,
2720 	.create_table = pnv_pci_ioda2_create_table,
2721 	.set_window = pnv_pci_ioda2_npu_set_window,
2722 	.unset_window = pnv_pci_ioda2_npu_unset_window,
2723 	.take_ownership = pnv_ioda2_npu_take_ownership,
2724 	.release_ownership = pnv_ioda2_release_ownership,
2725 };
2726 
2727 static void pnv_pci_ioda_setup_iommu_api(void)
2728 {
2729 	struct pci_controller *hose, *tmp;
2730 	struct pnv_phb *phb;
2731 	struct pnv_ioda_pe *pe, *gpe;
2732 
2733 	/*
2734 	 * Now we have all PHBs discovered, time to add NPU devices to
2735 	 * the corresponding IOMMU groups.
2736 	 */
2737 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2738 		phb = hose->private_data;
2739 
2740 		if (phb->type != PNV_PHB_NPU_NVLINK)
2741 			continue;
2742 
2743 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2744 			gpe = pnv_pci_npu_setup_iommu(pe);
2745 			if (gpe)
2746 				gpe->table_group.ops = &pnv_pci_ioda2_npu_ops;
2747 		}
2748 	}
2749 }
2750 #else /* !CONFIG_IOMMU_API */
2751 static void pnv_pci_ioda_setup_iommu_api(void) { };
2752 #endif
2753 
2754 static __be64 *pnv_pci_ioda2_table_do_alloc_pages(int nid, unsigned shift,
2755 		unsigned levels, unsigned long limit,
2756 		unsigned long *current_offset, unsigned long *total_allocated)
2757 {
2758 	struct page *tce_mem = NULL;
2759 	__be64 *addr, *tmp;
2760 	unsigned order = max_t(unsigned, shift, PAGE_SHIFT) - PAGE_SHIFT;
2761 	unsigned long allocated = 1UL << (order + PAGE_SHIFT);
2762 	unsigned entries = 1UL << (shift - 3);
2763 	long i;
2764 
2765 	tce_mem = alloc_pages_node(nid, GFP_KERNEL, order);
2766 	if (!tce_mem) {
2767 		pr_err("Failed to allocate a TCE memory, order=%d\n", order);
2768 		return NULL;
2769 	}
2770 	addr = page_address(tce_mem);
2771 	memset(addr, 0, allocated);
2772 	*total_allocated += allocated;
2773 
2774 	--levels;
2775 	if (!levels) {
2776 		*current_offset += allocated;
2777 		return addr;
2778 	}
2779 
2780 	for (i = 0; i < entries; ++i) {
2781 		tmp = pnv_pci_ioda2_table_do_alloc_pages(nid, shift,
2782 				levels, limit, current_offset, total_allocated);
2783 		if (!tmp)
2784 			break;
2785 
2786 		addr[i] = cpu_to_be64(__pa(tmp) |
2787 				TCE_PCI_READ | TCE_PCI_WRITE);
2788 
2789 		if (*current_offset >= limit)
2790 			break;
2791 	}
2792 
2793 	return addr;
2794 }
2795 
2796 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2797 		unsigned long size, unsigned level);
2798 
2799 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2800 		__u32 page_shift, __u64 window_size, __u32 levels,
2801 		struct iommu_table *tbl)
2802 {
2803 	void *addr;
2804 	unsigned long offset = 0, level_shift, total_allocated = 0;
2805 	const unsigned window_shift = ilog2(window_size);
2806 	unsigned entries_shift = window_shift - page_shift;
2807 	unsigned table_shift = max_t(unsigned, entries_shift + 3, PAGE_SHIFT);
2808 	const unsigned long tce_table_size = 1UL << table_shift;
2809 
2810 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS))
2811 		return -EINVAL;
2812 
2813 	if (!is_power_of_2(window_size))
2814 		return -EINVAL;
2815 
2816 	/* Adjust direct table size from window_size and levels */
2817 	entries_shift = (entries_shift + levels - 1) / levels;
2818 	level_shift = entries_shift + 3;
2819 	level_shift = max_t(unsigned, level_shift, PAGE_SHIFT);
2820 
2821 	if ((level_shift - 3) * levels + page_shift >= 60)
2822 		return -EINVAL;
2823 
2824 	/* Allocate TCE table */
2825 	addr = pnv_pci_ioda2_table_do_alloc_pages(nid, level_shift,
2826 			levels, tce_table_size, &offset, &total_allocated);
2827 
2828 	/* addr==NULL means that the first level allocation failed */
2829 	if (!addr)
2830 		return -ENOMEM;
2831 
2832 	/*
2833 	 * First level was allocated but some lower level failed as
2834 	 * we did not allocate as much as we wanted,
2835 	 * release partially allocated table.
2836 	 */
2837 	if (offset < tce_table_size) {
2838 		pnv_pci_ioda2_table_do_free_pages(addr,
2839 				1ULL << (level_shift - 3), levels - 1);
2840 		return -ENOMEM;
2841 	}
2842 
2843 	/* Setup linux iommu table */
2844 	pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, bus_offset,
2845 			page_shift);
2846 	tbl->it_level_size = 1ULL << (level_shift - 3);
2847 	tbl->it_indirect_levels = levels - 1;
2848 	tbl->it_allocated_size = total_allocated;
2849 
2850 	pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2851 			window_size, tce_table_size, bus_offset);
2852 
2853 	return 0;
2854 }
2855 
2856 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2857 		unsigned long size, unsigned level)
2858 {
2859 	const unsigned long addr_ul = (unsigned long) addr &
2860 			~(TCE_PCI_READ | TCE_PCI_WRITE);
2861 
2862 	if (level) {
2863 		long i;
2864 		u64 *tmp = (u64 *) addr_ul;
2865 
2866 		for (i = 0; i < size; ++i) {
2867 			unsigned long hpa = be64_to_cpu(tmp[i]);
2868 
2869 			if (!(hpa & (TCE_PCI_READ | TCE_PCI_WRITE)))
2870 				continue;
2871 
2872 			pnv_pci_ioda2_table_do_free_pages(__va(hpa), size,
2873 					level - 1);
2874 		}
2875 	}
2876 
2877 	free_pages(addr_ul, get_order(size << 3));
2878 }
2879 
2880 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl)
2881 {
2882 	const unsigned long size = tbl->it_indirect_levels ?
2883 			tbl->it_level_size : tbl->it_size;
2884 
2885 	if (!tbl->it_size)
2886 		return;
2887 
2888 	pnv_pci_ioda2_table_do_free_pages((__be64 *)tbl->it_base, size,
2889 			tbl->it_indirect_levels);
2890 }
2891 
2892 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2893 				       struct pnv_ioda_pe *pe)
2894 {
2895 	int64_t rc;
2896 
2897 	if (!pnv_pci_ioda_pe_dma_weight(pe))
2898 		return;
2899 
2900 	/* TVE #1 is selected by PCI address bit 59 */
2901 	pe->tce_bypass_base = 1ull << 59;
2902 
2903 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2904 			pe->pe_number);
2905 
2906 	/* The PE will reserve all possible 32-bits space */
2907 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2908 		phb->ioda.m32_pci_base);
2909 
2910 	/* Setup linux iommu table */
2911 	pe->table_group.tce32_start = 0;
2912 	pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2913 	pe->table_group.max_dynamic_windows_supported =
2914 			IOMMU_TABLE_GROUP_MAX_TABLES;
2915 	pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2916 	pe->table_group.pgsizes = SZ_4K | SZ_64K | SZ_16M;
2917 #ifdef CONFIG_IOMMU_API
2918 	pe->table_group.ops = &pnv_pci_ioda2_ops;
2919 #endif
2920 
2921 	rc = pnv_pci_ioda2_setup_default_config(pe);
2922 	if (rc)
2923 		return;
2924 
2925 	if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2926 		pnv_ioda_setup_bus_dma(pe, pe->pbus, true);
2927 }
2928 
2929 #ifdef CONFIG_PCI_MSI
2930 int64_t pnv_opal_pci_msi_eoi(struct irq_chip *chip, unsigned int hw_irq)
2931 {
2932 	struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2933 					   ioda.irq_chip);
2934 
2935 	return opal_pci_msi_eoi(phb->opal_id, hw_irq);
2936 }
2937 
2938 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2939 {
2940 	int64_t rc;
2941 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2942 	struct irq_chip *chip = irq_data_get_irq_chip(d);
2943 
2944 	rc = pnv_opal_pci_msi_eoi(chip, hw_irq);
2945 	WARN_ON_ONCE(rc);
2946 
2947 	icp_native_eoi(d);
2948 }
2949 
2950 
2951 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2952 {
2953 	struct irq_data *idata;
2954 	struct irq_chip *ichip;
2955 
2956 	/* The MSI EOI OPAL call is only needed on PHB3 */
2957 	if (phb->model != PNV_PHB_MODEL_PHB3)
2958 		return;
2959 
2960 	if (!phb->ioda.irq_chip_init) {
2961 		/*
2962 		 * First time we setup an MSI IRQ, we need to setup the
2963 		 * corresponding IRQ chip to route correctly.
2964 		 */
2965 		idata = irq_get_irq_data(virq);
2966 		ichip = irq_data_get_irq_chip(idata);
2967 		phb->ioda.irq_chip_init = 1;
2968 		phb->ioda.irq_chip = *ichip;
2969 		phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2970 	}
2971 	irq_set_chip(virq, &phb->ioda.irq_chip);
2972 }
2973 
2974 /*
2975  * Returns true iff chip is something that we could call
2976  * pnv_opal_pci_msi_eoi for.
2977  */
2978 bool is_pnv_opal_msi(struct irq_chip *chip)
2979 {
2980 	return chip->irq_eoi == pnv_ioda2_msi_eoi;
2981 }
2982 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2983 
2984 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2985 				  unsigned int hwirq, unsigned int virq,
2986 				  unsigned int is_64, struct msi_msg *msg)
2987 {
2988 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2989 	unsigned int xive_num = hwirq - phb->msi_base;
2990 	__be32 data;
2991 	int rc;
2992 
2993 	/* No PE assigned ? bail out ... no MSI for you ! */
2994 	if (pe == NULL)
2995 		return -ENXIO;
2996 
2997 	/* Check if we have an MVE */
2998 	if (pe->mve_number < 0)
2999 		return -ENXIO;
3000 
3001 	/* Force 32-bit MSI on some broken devices */
3002 	if (dev->no_64bit_msi)
3003 		is_64 = 0;
3004 
3005 	/* Assign XIVE to PE */
3006 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
3007 	if (rc) {
3008 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
3009 			pci_name(dev), rc, xive_num);
3010 		return -EIO;
3011 	}
3012 
3013 	if (is_64) {
3014 		__be64 addr64;
3015 
3016 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
3017 				     &addr64, &data);
3018 		if (rc) {
3019 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
3020 				pci_name(dev), rc);
3021 			return -EIO;
3022 		}
3023 		msg->address_hi = be64_to_cpu(addr64) >> 32;
3024 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
3025 	} else {
3026 		__be32 addr32;
3027 
3028 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
3029 				     &addr32, &data);
3030 		if (rc) {
3031 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
3032 				pci_name(dev), rc);
3033 			return -EIO;
3034 		}
3035 		msg->address_hi = 0;
3036 		msg->address_lo = be32_to_cpu(addr32);
3037 	}
3038 	msg->data = be32_to_cpu(data);
3039 
3040 	pnv_set_msi_irq_chip(phb, virq);
3041 
3042 	pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
3043 		 " address=%x_%08x data=%x PE# %x\n",
3044 		 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
3045 		 msg->address_hi, msg->address_lo, data, pe->pe_number);
3046 
3047 	return 0;
3048 }
3049 
3050 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
3051 {
3052 	unsigned int count;
3053 	const __be32 *prop = of_get_property(phb->hose->dn,
3054 					     "ibm,opal-msi-ranges", NULL);
3055 	if (!prop) {
3056 		/* BML Fallback */
3057 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
3058 	}
3059 	if (!prop)
3060 		return;
3061 
3062 	phb->msi_base = be32_to_cpup(prop);
3063 	count = be32_to_cpup(prop + 1);
3064 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
3065 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
3066 		       phb->hose->global_number);
3067 		return;
3068 	}
3069 
3070 	phb->msi_setup = pnv_pci_ioda_msi_setup;
3071 	phb->msi32_support = 1;
3072 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
3073 		count, phb->msi_base);
3074 }
3075 #else
3076 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
3077 #endif /* CONFIG_PCI_MSI */
3078 
3079 #ifdef CONFIG_PCI_IOV
3080 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
3081 {
3082 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3083 	struct pnv_phb *phb = hose->private_data;
3084 	const resource_size_t gate = phb->ioda.m64_segsize >> 2;
3085 	struct resource *res;
3086 	int i;
3087 	resource_size_t size, total_vf_bar_sz;
3088 	struct pci_dn *pdn;
3089 	int mul, total_vfs;
3090 
3091 	if (!pdev->is_physfn || pdev->is_added)
3092 		return;
3093 
3094 	pdn = pci_get_pdn(pdev);
3095 	pdn->vfs_expanded = 0;
3096 	pdn->m64_single_mode = false;
3097 
3098 	total_vfs = pci_sriov_get_totalvfs(pdev);
3099 	mul = phb->ioda.total_pe_num;
3100 	total_vf_bar_sz = 0;
3101 
3102 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3103 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
3104 		if (!res->flags || res->parent)
3105 			continue;
3106 		if (!pnv_pci_is_m64_flags(res->flags)) {
3107 			dev_warn(&pdev->dev, "Don't support SR-IOV with"
3108 					" non M64 VF BAR%d: %pR. \n",
3109 				 i, res);
3110 			goto truncate_iov;
3111 		}
3112 
3113 		total_vf_bar_sz += pci_iov_resource_size(pdev,
3114 				i + PCI_IOV_RESOURCES);
3115 
3116 		/*
3117 		 * If bigger than quarter of M64 segment size, just round up
3118 		 * power of two.
3119 		 *
3120 		 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
3121 		 * with other devices, IOV BAR size is expanded to be
3122 		 * (total_pe * VF_BAR_size).  When VF_BAR_size is half of M64
3123 		 * segment size , the expanded size would equal to half of the
3124 		 * whole M64 space size, which will exhaust the M64 Space and
3125 		 * limit the system flexibility.  This is a design decision to
3126 		 * set the boundary to quarter of the M64 segment size.
3127 		 */
3128 		if (total_vf_bar_sz > gate) {
3129 			mul = roundup_pow_of_two(total_vfs);
3130 			dev_info(&pdev->dev,
3131 				"VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
3132 				total_vf_bar_sz, gate, mul);
3133 			pdn->m64_single_mode = true;
3134 			break;
3135 		}
3136 	}
3137 
3138 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3139 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
3140 		if (!res->flags || res->parent)
3141 			continue;
3142 
3143 		size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
3144 		/*
3145 		 * On PHB3, the minimum size alignment of M64 BAR in single
3146 		 * mode is 32MB.
3147 		 */
3148 		if (pdn->m64_single_mode && (size < SZ_32M))
3149 			goto truncate_iov;
3150 		dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
3151 		res->end = res->start + size * mul - 1;
3152 		dev_dbg(&pdev->dev, "                       %pR\n", res);
3153 		dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
3154 			 i, res, mul);
3155 	}
3156 	pdn->vfs_expanded = mul;
3157 
3158 	return;
3159 
3160 truncate_iov:
3161 	/* To save MMIO space, IOV BAR is truncated. */
3162 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3163 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
3164 		res->flags = 0;
3165 		res->end = res->start - 1;
3166 	}
3167 }
3168 #endif /* CONFIG_PCI_IOV */
3169 
3170 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
3171 				  struct resource *res)
3172 {
3173 	struct pnv_phb *phb = pe->phb;
3174 	struct pci_bus_region region;
3175 	int index;
3176 	int64_t rc;
3177 
3178 	if (!res || !res->flags || res->start > res->end)
3179 		return;
3180 
3181 	if (res->flags & IORESOURCE_IO) {
3182 		region.start = res->start - phb->ioda.io_pci_base;
3183 		region.end   = res->end - phb->ioda.io_pci_base;
3184 		index = region.start / phb->ioda.io_segsize;
3185 
3186 		while (index < phb->ioda.total_pe_num &&
3187 		       region.start <= region.end) {
3188 			phb->ioda.io_segmap[index] = pe->pe_number;
3189 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3190 				pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
3191 			if (rc != OPAL_SUCCESS) {
3192 				pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
3193 				       __func__, rc, index, pe->pe_number);
3194 				break;
3195 			}
3196 
3197 			region.start += phb->ioda.io_segsize;
3198 			index++;
3199 		}
3200 	} else if ((res->flags & IORESOURCE_MEM) &&
3201 		   !pnv_pci_is_m64(phb, res)) {
3202 		region.start = res->start -
3203 			       phb->hose->mem_offset[0] -
3204 			       phb->ioda.m32_pci_base;
3205 		region.end   = res->end -
3206 			       phb->hose->mem_offset[0] -
3207 			       phb->ioda.m32_pci_base;
3208 		index = region.start / phb->ioda.m32_segsize;
3209 
3210 		while (index < phb->ioda.total_pe_num &&
3211 		       region.start <= region.end) {
3212 			phb->ioda.m32_segmap[index] = pe->pe_number;
3213 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3214 				pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
3215 			if (rc != OPAL_SUCCESS) {
3216 				pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
3217 				       __func__, rc, index, pe->pe_number);
3218 				break;
3219 			}
3220 
3221 			region.start += phb->ioda.m32_segsize;
3222 			index++;
3223 		}
3224 	}
3225 }
3226 
3227 /*
3228  * This function is supposed to be called on basis of PE from top
3229  * to bottom style. So the the I/O or MMIO segment assigned to
3230  * parent PE could be overridden by its child PEs if necessary.
3231  */
3232 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
3233 {
3234 	struct pci_dev *pdev;
3235 	int i;
3236 
3237 	/*
3238 	 * NOTE: We only care PCI bus based PE for now. For PCI
3239 	 * device based PE, for example SRIOV sensitive VF should
3240 	 * be figured out later.
3241 	 */
3242 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3243 
3244 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3245 		for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3246 			pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3247 
3248 		/*
3249 		 * If the PE contains all subordinate PCI buses, the
3250 		 * windows of the child bridges should be mapped to
3251 		 * the PE as well.
3252 		 */
3253 		if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3254 			continue;
3255 		for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3256 			pnv_ioda_setup_pe_res(pe,
3257 				&pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3258 	}
3259 }
3260 
3261 #ifdef CONFIG_DEBUG_FS
3262 static int pnv_pci_diag_data_set(void *data, u64 val)
3263 {
3264 	struct pci_controller *hose;
3265 	struct pnv_phb *phb;
3266 	s64 ret;
3267 
3268 	if (val != 1ULL)
3269 		return -EINVAL;
3270 
3271 	hose = (struct pci_controller *)data;
3272 	if (!hose || !hose->private_data)
3273 		return -ENODEV;
3274 
3275 	phb = hose->private_data;
3276 
3277 	/* Retrieve the diag data from firmware */
3278 	ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data,
3279 					  phb->diag_data_size);
3280 	if (ret != OPAL_SUCCESS)
3281 		return -EIO;
3282 
3283 	/* Print the diag data to the kernel log */
3284 	pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data);
3285 	return 0;
3286 }
3287 
3288 DEFINE_SIMPLE_ATTRIBUTE(pnv_pci_diag_data_fops, NULL,
3289 			pnv_pci_diag_data_set, "%llu\n");
3290 
3291 #endif /* CONFIG_DEBUG_FS */
3292 
3293 static void pnv_pci_ioda_create_dbgfs(void)
3294 {
3295 #ifdef CONFIG_DEBUG_FS
3296 	struct pci_controller *hose, *tmp;
3297 	struct pnv_phb *phb;
3298 	char name[16];
3299 
3300 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3301 		phb = hose->private_data;
3302 
3303 		/* Notify initialization of PHB done */
3304 		phb->initialized = 1;
3305 
3306 		sprintf(name, "PCI%04x", hose->global_number);
3307 		phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3308 		if (!phb->dbgfs) {
3309 			pr_warn("%s: Error on creating debugfs on PHB#%x\n",
3310 				__func__, hose->global_number);
3311 			continue;
3312 		}
3313 
3314 		debugfs_create_file("dump_diag_regs", 0200, phb->dbgfs, hose,
3315 				    &pnv_pci_diag_data_fops);
3316 	}
3317 #endif /* CONFIG_DEBUG_FS */
3318 }
3319 
3320 static void pnv_pci_ioda_fixup(void)
3321 {
3322 	pnv_pci_ioda_setup_PEs();
3323 	pnv_pci_ioda_setup_iommu_api();
3324 	pnv_pci_ioda_create_dbgfs();
3325 
3326 #ifdef CONFIG_EEH
3327 	pnv_eeh_post_init();
3328 #endif
3329 }
3330 
3331 /*
3332  * Returns the alignment for I/O or memory windows for P2P
3333  * bridges. That actually depends on how PEs are segmented.
3334  * For now, we return I/O or M32 segment size for PE sensitive
3335  * P2P bridges. Otherwise, the default values (4KiB for I/O,
3336  * 1MiB for memory) will be returned.
3337  *
3338  * The current PCI bus might be put into one PE, which was
3339  * create against the parent PCI bridge. For that case, we
3340  * needn't enlarge the alignment so that we can save some
3341  * resources.
3342  */
3343 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3344 						unsigned long type)
3345 {
3346 	struct pci_dev *bridge;
3347 	struct pci_controller *hose = pci_bus_to_host(bus);
3348 	struct pnv_phb *phb = hose->private_data;
3349 	int num_pci_bridges = 0;
3350 
3351 	bridge = bus->self;
3352 	while (bridge) {
3353 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3354 			num_pci_bridges++;
3355 			if (num_pci_bridges >= 2)
3356 				return 1;
3357 		}
3358 
3359 		bridge = bridge->bus->self;
3360 	}
3361 
3362 	/*
3363 	 * We fall back to M32 if M64 isn't supported. We enforce the M64
3364 	 * alignment for any 64-bit resource, PCIe doesn't care and
3365 	 * bridges only do 64-bit prefetchable anyway.
3366 	 */
3367 	if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
3368 		return phb->ioda.m64_segsize;
3369 	if (type & IORESOURCE_MEM)
3370 		return phb->ioda.m32_segsize;
3371 
3372 	return phb->ioda.io_segsize;
3373 }
3374 
3375 /*
3376  * We are updating root port or the upstream port of the
3377  * bridge behind the root port with PHB's windows in order
3378  * to accommodate the changes on required resources during
3379  * PCI (slot) hotplug, which is connected to either root
3380  * port or the downstream ports of PCIe switch behind the
3381  * root port.
3382  */
3383 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
3384 					   unsigned long type)
3385 {
3386 	struct pci_controller *hose = pci_bus_to_host(bus);
3387 	struct pnv_phb *phb = hose->private_data;
3388 	struct pci_dev *bridge = bus->self;
3389 	struct resource *r, *w;
3390 	bool msi_region = false;
3391 	int i;
3392 
3393 	/* Check if we need apply fixup to the bridge's windows */
3394 	if (!pci_is_root_bus(bridge->bus) &&
3395 	    !pci_is_root_bus(bridge->bus->self->bus))
3396 		return;
3397 
3398 	/* Fixup the resources */
3399 	for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
3400 		r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
3401 		if (!r->flags || !r->parent)
3402 			continue;
3403 
3404 		w = NULL;
3405 		if (r->flags & type & IORESOURCE_IO)
3406 			w = &hose->io_resource;
3407 		else if (pnv_pci_is_m64(phb, r) &&
3408 			 (type & IORESOURCE_PREFETCH) &&
3409 			 phb->ioda.m64_segsize)
3410 			w = &hose->mem_resources[1];
3411 		else if (r->flags & type & IORESOURCE_MEM) {
3412 			w = &hose->mem_resources[0];
3413 			msi_region = true;
3414 		}
3415 
3416 		r->start = w->start;
3417 		r->end = w->end;
3418 
3419 		/* The 64KB 32-bits MSI region shouldn't be included in
3420 		 * the 32-bits bridge window. Otherwise, we can see strange
3421 		 * issues. One of them is EEH error observed on Garrison.
3422 		 *
3423 		 * Exclude top 1MB region which is the minimal alignment of
3424 		 * 32-bits bridge window.
3425 		 */
3426 		if (msi_region) {
3427 			r->end += 0x10000;
3428 			r->end -= 0x100000;
3429 		}
3430 	}
3431 }
3432 
3433 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type)
3434 {
3435 	struct pci_controller *hose = pci_bus_to_host(bus);
3436 	struct pnv_phb *phb = hose->private_data;
3437 	struct pci_dev *bridge = bus->self;
3438 	struct pnv_ioda_pe *pe;
3439 	bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
3440 
3441 	/* Extend bridge's windows if necessary */
3442 	pnv_pci_fixup_bridge_resources(bus, type);
3443 
3444 	/* The PE for root bus should be realized before any one else */
3445 	if (!phb->ioda.root_pe_populated) {
3446 		pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false);
3447 		if (pe) {
3448 			phb->ioda.root_pe_idx = pe->pe_number;
3449 			phb->ioda.root_pe_populated = true;
3450 		}
3451 	}
3452 
3453 	/* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3454 	if (list_empty(&bus->devices))
3455 		return;
3456 
3457 	/* Reserve PEs according to used M64 resources */
3458 	if (phb->reserve_m64_pe)
3459 		phb->reserve_m64_pe(bus, NULL, all);
3460 
3461 	/*
3462 	 * Assign PE. We might run here because of partial hotplug.
3463 	 * For the case, we just pick up the existing PE and should
3464 	 * not allocate resources again.
3465 	 */
3466 	pe = pnv_ioda_setup_bus_PE(bus, all);
3467 	if (!pe)
3468 		return;
3469 
3470 	pnv_ioda_setup_pe_seg(pe);
3471 	switch (phb->type) {
3472 	case PNV_PHB_IODA1:
3473 		pnv_pci_ioda1_setup_dma_pe(phb, pe);
3474 		break;
3475 	case PNV_PHB_IODA2:
3476 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
3477 		break;
3478 	default:
3479 		pr_warn("%s: No DMA for PHB#%x (type %d)\n",
3480 			__func__, phb->hose->global_number, phb->type);
3481 	}
3482 }
3483 
3484 static resource_size_t pnv_pci_default_alignment(void)
3485 {
3486 	return PAGE_SIZE;
3487 }
3488 
3489 #ifdef CONFIG_PCI_IOV
3490 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3491 						      int resno)
3492 {
3493 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3494 	struct pnv_phb *phb = hose->private_data;
3495 	struct pci_dn *pdn = pci_get_pdn(pdev);
3496 	resource_size_t align;
3497 
3498 	/*
3499 	 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3500 	 * SR-IOV. While from hardware perspective, the range mapped by M64
3501 	 * BAR should be size aligned.
3502 	 *
3503 	 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3504 	 * powernv-specific hardware restriction is gone. But if just use the
3505 	 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3506 	 * in one segment of M64 #15, which introduces the PE conflict between
3507 	 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3508 	 * m64_segsize.
3509 	 *
3510 	 * This function returns the total IOV BAR size if M64 BAR is in
3511 	 * Shared PE mode or just VF BAR size if not.
3512 	 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3513 	 * M64 segment size if IOV BAR size is less.
3514 	 */
3515 	align = pci_iov_resource_size(pdev, resno);
3516 	if (!pdn->vfs_expanded)
3517 		return align;
3518 	if (pdn->m64_single_mode)
3519 		return max(align, (resource_size_t)phb->ioda.m64_segsize);
3520 
3521 	return pdn->vfs_expanded * align;
3522 }
3523 #endif /* CONFIG_PCI_IOV */
3524 
3525 /* Prevent enabling devices for which we couldn't properly
3526  * assign a PE
3527  */
3528 bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3529 {
3530 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
3531 	struct pnv_phb *phb = hose->private_data;
3532 	struct pci_dn *pdn;
3533 
3534 	/* The function is probably called while the PEs have
3535 	 * not be created yet. For example, resource reassignment
3536 	 * during PCI probe period. We just skip the check if
3537 	 * PEs isn't ready.
3538 	 */
3539 	if (!phb->initialized)
3540 		return true;
3541 
3542 	pdn = pci_get_pdn(dev);
3543 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3544 		return false;
3545 
3546 	return true;
3547 }
3548 
3549 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
3550 				       int num)
3551 {
3552 	struct pnv_ioda_pe *pe = container_of(table_group,
3553 					      struct pnv_ioda_pe, table_group);
3554 	struct pnv_phb *phb = pe->phb;
3555 	unsigned int idx;
3556 	long rc;
3557 
3558 	pe_info(pe, "Removing DMA window #%d\n", num);
3559 	for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
3560 		if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
3561 			continue;
3562 
3563 		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
3564 						idx, 0, 0ul, 0ul, 0ul);
3565 		if (rc != OPAL_SUCCESS) {
3566 			pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
3567 				rc, idx);
3568 			return rc;
3569 		}
3570 
3571 		phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
3572 	}
3573 
3574 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
3575 	return OPAL_SUCCESS;
3576 }
3577 
3578 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
3579 {
3580 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3581 	struct iommu_table *tbl = pe->table_group.tables[0];
3582 	int64_t rc;
3583 
3584 	if (!weight)
3585 		return;
3586 
3587 	rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
3588 	if (rc != OPAL_SUCCESS)
3589 		return;
3590 
3591 	pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
3592 	if (pe->table_group.group) {
3593 		iommu_group_put(pe->table_group.group);
3594 		WARN_ON(pe->table_group.group);
3595 	}
3596 
3597 	free_pages(tbl->it_base, get_order(tbl->it_size << 3));
3598 	iommu_tce_table_put(tbl);
3599 }
3600 
3601 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
3602 {
3603 	struct iommu_table *tbl = pe->table_group.tables[0];
3604 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3605 #ifdef CONFIG_IOMMU_API
3606 	int64_t rc;
3607 #endif
3608 
3609 	if (!weight)
3610 		return;
3611 
3612 #ifdef CONFIG_IOMMU_API
3613 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
3614 	if (rc)
3615 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
3616 #endif
3617 
3618 	pnv_pci_ioda2_set_bypass(pe, false);
3619 	if (pe->table_group.group) {
3620 		iommu_group_put(pe->table_group.group);
3621 		WARN_ON(pe->table_group.group);
3622 	}
3623 
3624 	pnv_pci_ioda2_table_free_pages(tbl);
3625 	iommu_tce_table_put(tbl);
3626 }
3627 
3628 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
3629 				 unsigned short win,
3630 				 unsigned int *map)
3631 {
3632 	struct pnv_phb *phb = pe->phb;
3633 	int idx;
3634 	int64_t rc;
3635 
3636 	for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
3637 		if (map[idx] != pe->pe_number)
3638 			continue;
3639 
3640 		if (win == OPAL_M64_WINDOW_TYPE)
3641 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3642 					phb->ioda.reserved_pe_idx, win,
3643 					idx / PNV_IODA1_M64_SEGS,
3644 					idx % PNV_IODA1_M64_SEGS);
3645 		else
3646 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3647 					phb->ioda.reserved_pe_idx, win, 0, idx);
3648 
3649 		if (rc != OPAL_SUCCESS)
3650 			pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
3651 				rc, win, idx);
3652 
3653 		map[idx] = IODA_INVALID_PE;
3654 	}
3655 }
3656 
3657 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
3658 {
3659 	struct pnv_phb *phb = pe->phb;
3660 
3661 	if (phb->type == PNV_PHB_IODA1) {
3662 		pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
3663 				     phb->ioda.io_segmap);
3664 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3665 				     phb->ioda.m32_segmap);
3666 		pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE,
3667 				     phb->ioda.m64_segmap);
3668 	} else if (phb->type == PNV_PHB_IODA2) {
3669 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3670 				     phb->ioda.m32_segmap);
3671 	}
3672 }
3673 
3674 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
3675 {
3676 	struct pnv_phb *phb = pe->phb;
3677 	struct pnv_ioda_pe *slave, *tmp;
3678 
3679 	list_del(&pe->list);
3680 	switch (phb->type) {
3681 	case PNV_PHB_IODA1:
3682 		pnv_pci_ioda1_release_pe_dma(pe);
3683 		break;
3684 	case PNV_PHB_IODA2:
3685 		pnv_pci_ioda2_release_pe_dma(pe);
3686 		break;
3687 	default:
3688 		WARN_ON(1);
3689 	}
3690 
3691 	pnv_ioda_release_pe_seg(pe);
3692 	pnv_ioda_deconfigure_pe(pe->phb, pe);
3693 
3694 	/* Release slave PEs in the compound PE */
3695 	if (pe->flags & PNV_IODA_PE_MASTER) {
3696 		list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
3697 			list_del(&slave->list);
3698 			pnv_ioda_free_pe(slave);
3699 		}
3700 	}
3701 
3702 	/*
3703 	 * The PE for root bus can be removed because of hotplug in EEH
3704 	 * recovery for fenced PHB error. We need to mark the PE dead so
3705 	 * that it can be populated again in PCI hot add path. The PE
3706 	 * shouldn't be destroyed as it's the global reserved resource.
3707 	 */
3708 	if (phb->ioda.root_pe_populated &&
3709 	    phb->ioda.root_pe_idx == pe->pe_number)
3710 		phb->ioda.root_pe_populated = false;
3711 	else
3712 		pnv_ioda_free_pe(pe);
3713 }
3714 
3715 static void pnv_pci_release_device(struct pci_dev *pdev)
3716 {
3717 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3718 	struct pnv_phb *phb = hose->private_data;
3719 	struct pci_dn *pdn = pci_get_pdn(pdev);
3720 	struct pnv_ioda_pe *pe;
3721 
3722 	if (pdev->is_virtfn)
3723 		return;
3724 
3725 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3726 		return;
3727 
3728 	/*
3729 	 * PCI hotplug can happen as part of EEH error recovery. The @pdn
3730 	 * isn't removed and added afterwards in this scenario. We should
3731 	 * set the PE number in @pdn to an invalid one. Otherwise, the PE's
3732 	 * device count is decreased on removing devices while failing to
3733 	 * be increased on adding devices. It leads to unbalanced PE's device
3734 	 * count and eventually make normal PCI hotplug path broken.
3735 	 */
3736 	pe = &phb->ioda.pe_array[pdn->pe_number];
3737 	pdn->pe_number = IODA_INVALID_PE;
3738 
3739 	WARN_ON(--pe->device_count < 0);
3740 	if (pe->device_count == 0)
3741 		pnv_ioda_release_pe(pe);
3742 }
3743 
3744 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3745 {
3746 	struct pnv_phb *phb = hose->private_data;
3747 
3748 	opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3749 		       OPAL_ASSERT_RESET);
3750 }
3751 
3752 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3753 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3754 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3755 #ifdef CONFIG_PCI_MSI
3756 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3757 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3758 #endif
3759 	.enable_device_hook	= pnv_pci_enable_device_hook,
3760 	.release_device		= pnv_pci_release_device,
3761 	.window_alignment	= pnv_pci_window_alignment,
3762 	.setup_bridge		= pnv_pci_setup_bridge,
3763 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3764 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3765 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3766 	.shutdown		= pnv_pci_ioda_shutdown,
3767 };
3768 
3769 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask)
3770 {
3771 	dev_err_once(&npdev->dev,
3772 			"%s operation unsupported for NVLink devices\n",
3773 			__func__);
3774 	return -EPERM;
3775 }
3776 
3777 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3778 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3779 #ifdef CONFIG_PCI_MSI
3780 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3781 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3782 #endif
3783 	.enable_device_hook	= pnv_pci_enable_device_hook,
3784 	.window_alignment	= pnv_pci_window_alignment,
3785 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3786 	.dma_set_mask		= pnv_npu_dma_set_mask,
3787 	.shutdown		= pnv_pci_ioda_shutdown,
3788 };
3789 
3790 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = {
3791 	.enable_device_hook	= pnv_pci_enable_device_hook,
3792 	.window_alignment	= pnv_pci_window_alignment,
3793 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3794 	.shutdown		= pnv_pci_ioda_shutdown,
3795 };
3796 
3797 #ifdef CONFIG_CXL_BASE
3798 const struct pci_controller_ops pnv_cxl_cx4_ioda_controller_ops = {
3799 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3800 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3801 #ifdef CONFIG_PCI_MSI
3802 	.setup_msi_irqs		= pnv_cxl_cx4_setup_msi_irqs,
3803 	.teardown_msi_irqs	= pnv_cxl_cx4_teardown_msi_irqs,
3804 #endif
3805 	.enable_device_hook	= pnv_cxl_enable_device_hook,
3806 	.disable_device		= pnv_cxl_disable_device,
3807 	.release_device		= pnv_pci_release_device,
3808 	.window_alignment	= pnv_pci_window_alignment,
3809 	.setup_bridge		= pnv_pci_setup_bridge,
3810 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3811 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3812 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3813 	.shutdown		= pnv_pci_ioda_shutdown,
3814 };
3815 #endif
3816 
3817 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3818 					 u64 hub_id, int ioda_type)
3819 {
3820 	struct pci_controller *hose;
3821 	struct pnv_phb *phb;
3822 	unsigned long size, m64map_off, m32map_off, pemap_off;
3823 	unsigned long iomap_off = 0, dma32map_off = 0;
3824 	struct resource r;
3825 	const __be64 *prop64;
3826 	const __be32 *prop32;
3827 	int len;
3828 	unsigned int segno;
3829 	u64 phb_id;
3830 	void *aux;
3831 	long rc;
3832 
3833 	if (!of_device_is_available(np))
3834 		return;
3835 
3836 	pr_info("Initializing %s PHB (%pOF)\n",	pnv_phb_names[ioda_type], np);
3837 
3838 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3839 	if (!prop64) {
3840 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3841 		return;
3842 	}
3843 	phb_id = be64_to_cpup(prop64);
3844 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3845 
3846 	phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
3847 
3848 	/* Allocate PCI controller */
3849 	phb->hose = hose = pcibios_alloc_controller(np);
3850 	if (!phb->hose) {
3851 		pr_err("  Can't allocate PCI controller for %pOF\n",
3852 		       np);
3853 		memblock_free(__pa(phb), sizeof(struct pnv_phb));
3854 		return;
3855 	}
3856 
3857 	spin_lock_init(&phb->lock);
3858 	prop32 = of_get_property(np, "bus-range", &len);
3859 	if (prop32 && len == 8) {
3860 		hose->first_busno = be32_to_cpu(prop32[0]);
3861 		hose->last_busno = be32_to_cpu(prop32[1]);
3862 	} else {
3863 		pr_warn("  Broken <bus-range> on %pOF\n", np);
3864 		hose->first_busno = 0;
3865 		hose->last_busno = 0xff;
3866 	}
3867 	hose->private_data = phb;
3868 	phb->hub_id = hub_id;
3869 	phb->opal_id = phb_id;
3870 	phb->type = ioda_type;
3871 	mutex_init(&phb->ioda.pe_alloc_mutex);
3872 
3873 	/* Detect specific models for error handling */
3874 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3875 		phb->model = PNV_PHB_MODEL_P7IOC;
3876 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3877 		phb->model = PNV_PHB_MODEL_PHB3;
3878 	else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3879 		phb->model = PNV_PHB_MODEL_NPU;
3880 	else if (of_device_is_compatible(np, "ibm,power9-npu-pciex"))
3881 		phb->model = PNV_PHB_MODEL_NPU2;
3882 	else
3883 		phb->model = PNV_PHB_MODEL_UNKNOWN;
3884 
3885 	/* Initialize diagnostic data buffer */
3886 	prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL);
3887 	if (prop32)
3888 		phb->diag_data_size = be32_to_cpup(prop32);
3889 	else
3890 		phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE;
3891 
3892 	phb->diag_data = memblock_virt_alloc(phb->diag_data_size, 0);
3893 
3894 	/* Parse 32-bit and IO ranges (if any) */
3895 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3896 
3897 	/* Get registers */
3898 	if (!of_address_to_resource(np, 0, &r)) {
3899 		phb->regs_phys = r.start;
3900 		phb->regs = ioremap(r.start, resource_size(&r));
3901 		if (phb->regs == NULL)
3902 			pr_err("  Failed to map registers !\n");
3903 	}
3904 
3905 	/* Initialize more IODA stuff */
3906 	phb->ioda.total_pe_num = 1;
3907 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3908 	if (prop32)
3909 		phb->ioda.total_pe_num = be32_to_cpup(prop32);
3910 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3911 	if (prop32)
3912 		phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3913 
3914 	/* Invalidate RID to PE# mapping */
3915 	for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3916 		phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3917 
3918 	/* Parse 64-bit MMIO range */
3919 	pnv_ioda_parse_m64_window(phb);
3920 
3921 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3922 	/* FW Has already off top 64k of M32 space (MSI space) */
3923 	phb->ioda.m32_size += 0x10000;
3924 
3925 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3926 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3927 	phb->ioda.io_size = hose->pci_io_size;
3928 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3929 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3930 
3931 	/* Calculate how many 32-bit TCE segments we have */
3932 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3933 				PNV_IODA1_DMA32_SEGSIZE;
3934 
3935 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3936 	size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3937 			sizeof(unsigned long));
3938 	m64map_off = size;
3939 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3940 	m32map_off = size;
3941 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3942 	if (phb->type == PNV_PHB_IODA1) {
3943 		iomap_off = size;
3944 		size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3945 		dma32map_off = size;
3946 		size += phb->ioda.dma32_count *
3947 			sizeof(phb->ioda.dma32_segmap[0]);
3948 	}
3949 	pemap_off = size;
3950 	size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3951 	aux = memblock_virt_alloc(size, 0);
3952 	phb->ioda.pe_alloc = aux;
3953 	phb->ioda.m64_segmap = aux + m64map_off;
3954 	phb->ioda.m32_segmap = aux + m32map_off;
3955 	for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3956 		phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3957 		phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3958 	}
3959 	if (phb->type == PNV_PHB_IODA1) {
3960 		phb->ioda.io_segmap = aux + iomap_off;
3961 		for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3962 			phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3963 
3964 		phb->ioda.dma32_segmap = aux + dma32map_off;
3965 		for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3966 			phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3967 	}
3968 	phb->ioda.pe_array = aux + pemap_off;
3969 
3970 	/*
3971 	 * Choose PE number for root bus, which shouldn't have
3972 	 * M64 resources consumed by its child devices. To pick
3973 	 * the PE number adjacent to the reserved one if possible.
3974 	 */
3975 	pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3976 	if (phb->ioda.reserved_pe_idx == 0) {
3977 		phb->ioda.root_pe_idx = 1;
3978 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3979 	} else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3980 		phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3981 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3982 	} else {
3983 		phb->ioda.root_pe_idx = IODA_INVALID_PE;
3984 	}
3985 
3986 	INIT_LIST_HEAD(&phb->ioda.pe_list);
3987 	mutex_init(&phb->ioda.pe_list_mutex);
3988 
3989 	/* Calculate how many 32-bit TCE segments we have */
3990 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3991 				PNV_IODA1_DMA32_SEGSIZE;
3992 
3993 #if 0 /* We should really do that ... */
3994 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
3995 					 window_type,
3996 					 window_num,
3997 					 starting_real_address,
3998 					 starting_pci_address,
3999 					 segment_size);
4000 #endif
4001 
4002 	pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
4003 		phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
4004 		phb->ioda.m32_size, phb->ioda.m32_segsize);
4005 	if (phb->ioda.m64_size)
4006 		pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
4007 			phb->ioda.m64_size, phb->ioda.m64_segsize);
4008 	if (phb->ioda.io_size)
4009 		pr_info("                  IO: 0x%x [segment=0x%x]\n",
4010 			phb->ioda.io_size, phb->ioda.io_segsize);
4011 
4012 
4013 	phb->hose->ops = &pnv_pci_ops;
4014 	phb->get_pe_state = pnv_ioda_get_pe_state;
4015 	phb->freeze_pe = pnv_ioda_freeze_pe;
4016 	phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
4017 
4018 	/* Setup MSI support */
4019 	pnv_pci_init_ioda_msis(phb);
4020 
4021 	/*
4022 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
4023 	 * to let the PCI core do resource assignment. It's supposed
4024 	 * that the PCI core will do correct I/O and MMIO alignment
4025 	 * for the P2P bridge bars so that each PCI bus (excluding
4026 	 * the child P2P bridges) can form individual PE.
4027 	 */
4028 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
4029 
4030 	switch (phb->type) {
4031 	case PNV_PHB_NPU_NVLINK:
4032 		hose->controller_ops = pnv_npu_ioda_controller_ops;
4033 		break;
4034 	case PNV_PHB_NPU_OCAPI:
4035 		hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops;
4036 		break;
4037 	default:
4038 		phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
4039 		hose->controller_ops = pnv_pci_ioda_controller_ops;
4040 	}
4041 
4042 	ppc_md.pcibios_default_alignment = pnv_pci_default_alignment;
4043 
4044 #ifdef CONFIG_PCI_IOV
4045 	ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
4046 	ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
4047 	ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable;
4048 	ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable;
4049 #endif
4050 
4051 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
4052 
4053 	/* Reset IODA tables to a clean state */
4054 	rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
4055 	if (rc)
4056 		pr_warn("  OPAL Error %ld performing IODA table reset !\n", rc);
4057 
4058 	/*
4059 	 * If we're running in kdump kernel, the previous kernel never
4060 	 * shutdown PCI devices correctly. We already got IODA table
4061 	 * cleaned out. So we have to issue PHB reset to stop all PCI
4062 	 * transactions from previous kernel. The ppc_pci_reset_phbs
4063 	 * kernel parameter will force this reset too.
4064 	 */
4065 	if (is_kdump_kernel() || pci_reset_phbs) {
4066 		pr_info("  Issue PHB reset ...\n");
4067 		pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
4068 		pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
4069 	}
4070 
4071 	/* Remove M64 resource if we can't configure it successfully */
4072 	if (!phb->init_m64 || phb->init_m64(phb))
4073 		hose->mem_resources[1].flags = 0;
4074 }
4075 
4076 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
4077 {
4078 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
4079 }
4080 
4081 void __init pnv_pci_init_npu_phb(struct device_node *np)
4082 {
4083 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_NVLINK);
4084 }
4085 
4086 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np)
4087 {
4088 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI);
4089 }
4090 
4091 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev)
4092 {
4093 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
4094 	struct pnv_phb *phb = hose->private_data;
4095 
4096 	if (!machine_is(powernv))
4097 		return;
4098 
4099 	if (phb->type == PNV_PHB_NPU_OCAPI)
4100 		dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE;
4101 }
4102 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup);
4103 
4104 void __init pnv_pci_init_ioda_hub(struct device_node *np)
4105 {
4106 	struct device_node *phbn;
4107 	const __be64 *prop64;
4108 	u64 hub_id;
4109 
4110 	pr_info("Probing IODA IO-Hub %pOF\n", np);
4111 
4112 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
4113 	if (!prop64) {
4114 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
4115 		return;
4116 	}
4117 	hub_id = be64_to_cpup(prop64);
4118 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
4119 
4120 	/* Count child PHBs */
4121 	for_each_child_of_node(np, phbn) {
4122 		/* Look for IODA1 PHBs */
4123 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
4124 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
4125 	}
4126 }
4127