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/debugfs.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 
26 #include <asm/sections.h>
27 #include <asm/io.h>
28 #include <asm/prom.h>
29 #include <asm/pci-bridge.h>
30 #include <asm/machdep.h>
31 #include <asm/msi_bitmap.h>
32 #include <asm/ppc-pci.h>
33 #include <asm/opal.h>
34 #include <asm/iommu.h>
35 #include <asm/tce.h>
36 #include <asm/xics.h>
37 #include <asm/debug.h>
38 
39 #include "powernv.h"
40 #include "pci.h"
41 
42 #define define_pe_printk_level(func, kern_level)		\
43 static int func(const struct pnv_ioda_pe *pe, const char *fmt, ...)	\
44 {								\
45 	struct va_format vaf;					\
46 	va_list args;						\
47 	char pfix[32];						\
48 	int r;							\
49 								\
50 	va_start(args, fmt);					\
51 								\
52 	vaf.fmt = fmt;						\
53 	vaf.va = &args;						\
54 								\
55 	if (pe->pdev)						\
56 		strlcpy(pfix, dev_name(&pe->pdev->dev),		\
57 			sizeof(pfix));				\
58 	else							\
59 		sprintf(pfix, "%04x:%02x     ",			\
60 			pci_domain_nr(pe->pbus),		\
61 			pe->pbus->number);			\
62 	r = printk(kern_level "pci %s: [PE# %.3d] %pV",		\
63 		   pfix, pe->pe_number, &vaf);			\
64 								\
65 	va_end(args);						\
66 								\
67 	return r;						\
68 }								\
69 
70 define_pe_printk_level(pe_err, KERN_ERR);
71 define_pe_printk_level(pe_warn, KERN_WARNING);
72 define_pe_printk_level(pe_info, KERN_INFO);
73 
74 /*
75  * stdcix is only supposed to be used in hypervisor real mode as per
76  * the architecture spec
77  */
78 static inline void __raw_rm_writeq(u64 val, volatile void __iomem *paddr)
79 {
80 	__asm__ __volatile__("stdcix %0,0,%1"
81 		: : "r" (val), "r" (paddr) : "memory");
82 }
83 
84 static int pnv_ioda_alloc_pe(struct pnv_phb *phb)
85 {
86 	unsigned long pe;
87 
88 	do {
89 		pe = find_next_zero_bit(phb->ioda.pe_alloc,
90 					phb->ioda.total_pe, 0);
91 		if (pe >= phb->ioda.total_pe)
92 			return IODA_INVALID_PE;
93 	} while(test_and_set_bit(pe, phb->ioda.pe_alloc));
94 
95 	phb->ioda.pe_array[pe].phb = phb;
96 	phb->ioda.pe_array[pe].pe_number = pe;
97 	return pe;
98 }
99 
100 static void pnv_ioda_free_pe(struct pnv_phb *phb, int pe)
101 {
102 	WARN_ON(phb->ioda.pe_array[pe].pdev);
103 
104 	memset(&phb->ioda.pe_array[pe], 0, sizeof(struct pnv_ioda_pe));
105 	clear_bit(pe, phb->ioda.pe_alloc);
106 }
107 
108 /* Currently those 2 are only used when MSIs are enabled, this will change
109  * but in the meantime, we need to protect them to avoid warnings
110  */
111 #ifdef CONFIG_PCI_MSI
112 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
113 {
114 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
115 	struct pnv_phb *phb = hose->private_data;
116 	struct pci_dn *pdn = pci_get_pdn(dev);
117 
118 	if (!pdn)
119 		return NULL;
120 	if (pdn->pe_number == IODA_INVALID_PE)
121 		return NULL;
122 	return &phb->ioda.pe_array[pdn->pe_number];
123 }
124 #endif /* CONFIG_PCI_MSI */
125 
126 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
127 {
128 	struct pci_dev *parent;
129 	uint8_t bcomp, dcomp, fcomp;
130 	long rc, rid_end, rid;
131 
132 	/* Bus validation ? */
133 	if (pe->pbus) {
134 		int count;
135 
136 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
137 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
138 		parent = pe->pbus->self;
139 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
140 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
141 		else
142 			count = 1;
143 
144 		switch(count) {
145 		case  1: bcomp = OpalPciBusAll;		break;
146 		case  2: bcomp = OpalPciBus7Bits;	break;
147 		case  4: bcomp = OpalPciBus6Bits;	break;
148 		case  8: bcomp = OpalPciBus5Bits;	break;
149 		case 16: bcomp = OpalPciBus4Bits;	break;
150 		case 32: bcomp = OpalPciBus3Bits;	break;
151 		default:
152 			pr_err("%s: Number of subordinate busses %d"
153 			       " unsupported\n",
154 			       pci_name(pe->pbus->self), count);
155 			/* Do an exact match only */
156 			bcomp = OpalPciBusAll;
157 		}
158 		rid_end = pe->rid + (count << 8);
159 	} else {
160 		parent = pe->pdev->bus->self;
161 		bcomp = OpalPciBusAll;
162 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
163 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
164 		rid_end = pe->rid + 1;
165 	}
166 
167 	/*
168 	 * Associate PE in PELT. We need add the PE into the
169 	 * corresponding PELT-V as well. Otherwise, the error
170 	 * originated from the PE might contribute to other
171 	 * PEs.
172 	 */
173 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
174 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
175 	if (rc) {
176 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
177 		return -ENXIO;
178 	}
179 
180 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
181 				pe->pe_number, OPAL_ADD_PE_TO_DOMAIN);
182 	if (rc)
183 		pe_warn(pe, "OPAL error %d adding self to PELTV\n", rc);
184 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
185 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
186 
187 	/* Add to all parents PELT-V */
188 	while (parent) {
189 		struct pci_dn *pdn = pci_get_pdn(parent);
190 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
191 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
192 						pe->pe_number, OPAL_ADD_PE_TO_DOMAIN);
193 			/* XXX What to do in case of error ? */
194 		}
195 		parent = parent->bus->self;
196 	}
197 	/* Setup reverse map */
198 	for (rid = pe->rid; rid < rid_end; rid++)
199 		phb->ioda.pe_rmap[rid] = pe->pe_number;
200 
201 	/* Setup one MVTs on IODA1 */
202 	if (phb->type == PNV_PHB_IODA1) {
203 		pe->mve_number = pe->pe_number;
204 		rc = opal_pci_set_mve(phb->opal_id, pe->mve_number,
205 				      pe->pe_number);
206 		if (rc) {
207 			pe_err(pe, "OPAL error %ld setting up MVE %d\n",
208 			       rc, pe->mve_number);
209 			pe->mve_number = -1;
210 		} else {
211 			rc = opal_pci_set_mve_enable(phb->opal_id,
212 						     pe->mve_number, OPAL_ENABLE_MVE);
213 			if (rc) {
214 				pe_err(pe, "OPAL error %ld enabling MVE %d\n",
215 				       rc, pe->mve_number);
216 				pe->mve_number = -1;
217 			}
218 		}
219 	} else if (phb->type == PNV_PHB_IODA2)
220 		pe->mve_number = 0;
221 
222 	return 0;
223 }
224 
225 static void pnv_ioda_link_pe_by_weight(struct pnv_phb *phb,
226 				       struct pnv_ioda_pe *pe)
227 {
228 	struct pnv_ioda_pe *lpe;
229 
230 	list_for_each_entry(lpe, &phb->ioda.pe_dma_list, dma_link) {
231 		if (lpe->dma_weight < pe->dma_weight) {
232 			list_add_tail(&pe->dma_link, &lpe->dma_link);
233 			return;
234 		}
235 	}
236 	list_add_tail(&pe->dma_link, &phb->ioda.pe_dma_list);
237 }
238 
239 static unsigned int pnv_ioda_dma_weight(struct pci_dev *dev)
240 {
241 	/* This is quite simplistic. The "base" weight of a device
242 	 * is 10. 0 means no DMA is to be accounted for it.
243 	 */
244 
245 	/* If it's a bridge, no DMA */
246 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
247 		return 0;
248 
249 	/* Reduce the weight of slow USB controllers */
250 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
251 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
252 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
253 		return 3;
254 
255 	/* Increase the weight of RAID (includes Obsidian) */
256 	if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
257 		return 15;
258 
259 	/* Default */
260 	return 10;
261 }
262 
263 #if 0
264 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
265 {
266 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
267 	struct pnv_phb *phb = hose->private_data;
268 	struct pci_dn *pdn = pci_get_pdn(dev);
269 	struct pnv_ioda_pe *pe;
270 	int pe_num;
271 
272 	if (!pdn) {
273 		pr_err("%s: Device tree node not associated properly\n",
274 			   pci_name(dev));
275 		return NULL;
276 	}
277 	if (pdn->pe_number != IODA_INVALID_PE)
278 		return NULL;
279 
280 	/* PE#0 has been pre-set */
281 	if (dev->bus->number == 0)
282 		pe_num = 0;
283 	else
284 		pe_num = pnv_ioda_alloc_pe(phb);
285 	if (pe_num == IODA_INVALID_PE) {
286 		pr_warning("%s: Not enough PE# available, disabling device\n",
287 			   pci_name(dev));
288 		return NULL;
289 	}
290 
291 	/* NOTE: We get only one ref to the pci_dev for the pdn, not for the
292 	 * pointer in the PE data structure, both should be destroyed at the
293 	 * same time. However, this needs to be looked at more closely again
294 	 * once we actually start removing things (Hotplug, SR-IOV, ...)
295 	 *
296 	 * At some point we want to remove the PDN completely anyways
297 	 */
298 	pe = &phb->ioda.pe_array[pe_num];
299 	pci_dev_get(dev);
300 	pdn->pcidev = dev;
301 	pdn->pe_number = pe_num;
302 	pe->pdev = dev;
303 	pe->pbus = NULL;
304 	pe->tce32_seg = -1;
305 	pe->mve_number = -1;
306 	pe->rid = dev->bus->number << 8 | pdn->devfn;
307 
308 	pe_info(pe, "Associated device to PE\n");
309 
310 	if (pnv_ioda_configure_pe(phb, pe)) {
311 		/* XXX What do we do here ? */
312 		if (pe_num)
313 			pnv_ioda_free_pe(phb, pe_num);
314 		pdn->pe_number = IODA_INVALID_PE;
315 		pe->pdev = NULL;
316 		pci_dev_put(dev);
317 		return NULL;
318 	}
319 
320 	/* Assign a DMA weight to the device */
321 	pe->dma_weight = pnv_ioda_dma_weight(dev);
322 	if (pe->dma_weight != 0) {
323 		phb->ioda.dma_weight += pe->dma_weight;
324 		phb->ioda.dma_pe_count++;
325 	}
326 
327 	/* Link the PE */
328 	pnv_ioda_link_pe_by_weight(phb, pe);
329 
330 	return pe;
331 }
332 #endif /* Useful for SRIOV case */
333 
334 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
335 {
336 	struct pci_dev *dev;
337 
338 	list_for_each_entry(dev, &bus->devices, bus_list) {
339 		struct pci_dn *pdn = pci_get_pdn(dev);
340 
341 		if (pdn == NULL) {
342 			pr_warn("%s: No device node associated with device !\n",
343 				pci_name(dev));
344 			continue;
345 		}
346 		pci_dev_get(dev);
347 		pdn->pcidev = dev;
348 		pdn->pe_number = pe->pe_number;
349 		pe->dma_weight += pnv_ioda_dma_weight(dev);
350 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
351 			pnv_ioda_setup_same_PE(dev->subordinate, pe);
352 	}
353 }
354 
355 /*
356  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
357  * single PCI bus. Another one that contains the primary PCI bus and its
358  * subordinate PCI devices and buses. The second type of PE is normally
359  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
360  */
361 static void pnv_ioda_setup_bus_PE(struct pci_bus *bus, int all)
362 {
363 	struct pci_controller *hose = pci_bus_to_host(bus);
364 	struct pnv_phb *phb = hose->private_data;
365 	struct pnv_ioda_pe *pe;
366 	int pe_num;
367 
368 	pe_num = pnv_ioda_alloc_pe(phb);
369 	if (pe_num == IODA_INVALID_PE) {
370 		pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
371 			__func__, pci_domain_nr(bus), bus->number);
372 		return;
373 	}
374 
375 	pe = &phb->ioda.pe_array[pe_num];
376 	pe->flags = (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
377 	pe->pbus = bus;
378 	pe->pdev = NULL;
379 	pe->tce32_seg = -1;
380 	pe->mve_number = -1;
381 	pe->rid = bus->busn_res.start << 8;
382 	pe->dma_weight = 0;
383 
384 	if (all)
385 		pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
386 			bus->busn_res.start, bus->busn_res.end, pe_num);
387 	else
388 		pe_info(pe, "Secondary bus %d associated with PE#%d\n",
389 			bus->busn_res.start, pe_num);
390 
391 	if (pnv_ioda_configure_pe(phb, pe)) {
392 		/* XXX What do we do here ? */
393 		if (pe_num)
394 			pnv_ioda_free_pe(phb, pe_num);
395 		pe->pbus = NULL;
396 		return;
397 	}
398 
399 	/* Associate it with all child devices */
400 	pnv_ioda_setup_same_PE(bus, pe);
401 
402 	/* Put PE to the list */
403 	list_add_tail(&pe->list, &phb->ioda.pe_list);
404 
405 	/* Account for one DMA PE if at least one DMA capable device exist
406 	 * below the bridge
407 	 */
408 	if (pe->dma_weight != 0) {
409 		phb->ioda.dma_weight += pe->dma_weight;
410 		phb->ioda.dma_pe_count++;
411 	}
412 
413 	/* Link the PE */
414 	pnv_ioda_link_pe_by_weight(phb, pe);
415 }
416 
417 static void pnv_ioda_setup_PEs(struct pci_bus *bus)
418 {
419 	struct pci_dev *dev;
420 
421 	pnv_ioda_setup_bus_PE(bus, 0);
422 
423 	list_for_each_entry(dev, &bus->devices, bus_list) {
424 		if (dev->subordinate) {
425 			if (pci_pcie_type(dev) == PCI_EXP_TYPE_PCI_BRIDGE)
426 				pnv_ioda_setup_bus_PE(dev->subordinate, 1);
427 			else
428 				pnv_ioda_setup_PEs(dev->subordinate);
429 		}
430 	}
431 }
432 
433 /*
434  * Configure PEs so that the downstream PCI buses and devices
435  * could have their associated PE#. Unfortunately, we didn't
436  * figure out the way to identify the PLX bridge yet. So we
437  * simply put the PCI bus and the subordinate behind the root
438  * port to PE# here. The game rule here is expected to be changed
439  * as soon as we can detected PLX bridge correctly.
440  */
441 static void pnv_pci_ioda_setup_PEs(void)
442 {
443 	struct pci_controller *hose, *tmp;
444 
445 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
446 		pnv_ioda_setup_PEs(hose->bus);
447 	}
448 }
449 
450 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
451 {
452 	struct pci_dn *pdn = pci_get_pdn(pdev);
453 	struct pnv_ioda_pe *pe;
454 
455 	/*
456 	 * The function can be called while the PE#
457 	 * hasn't been assigned. Do nothing for the
458 	 * case.
459 	 */
460 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
461 		return;
462 
463 	pe = &phb->ioda.pe_array[pdn->pe_number];
464 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
465 	set_iommu_table_base_and_group(&pdev->dev, &pe->tce32_table);
466 }
467 
468 static int pnv_pci_ioda_dma_set_mask(struct pnv_phb *phb,
469 				     struct pci_dev *pdev, u64 dma_mask)
470 {
471 	struct pci_dn *pdn = pci_get_pdn(pdev);
472 	struct pnv_ioda_pe *pe;
473 	uint64_t top;
474 	bool bypass = false;
475 
476 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
477 		return -ENODEV;;
478 
479 	pe = &phb->ioda.pe_array[pdn->pe_number];
480 	if (pe->tce_bypass_enabled) {
481 		top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
482 		bypass = (dma_mask >= top);
483 	}
484 
485 	if (bypass) {
486 		dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
487 		set_dma_ops(&pdev->dev, &dma_direct_ops);
488 		set_dma_offset(&pdev->dev, pe->tce_bypass_base);
489 	} else {
490 		dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
491 		set_dma_ops(&pdev->dev, &dma_iommu_ops);
492 		set_iommu_table_base(&pdev->dev, &pe->tce32_table);
493 	}
494 	return 0;
495 }
496 
497 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus)
498 {
499 	struct pci_dev *dev;
500 
501 	list_for_each_entry(dev, &bus->devices, bus_list) {
502 		set_iommu_table_base_and_group(&dev->dev, &pe->tce32_table);
503 		if (dev->subordinate)
504 			pnv_ioda_setup_bus_dma(pe, dev->subordinate);
505 	}
506 }
507 
508 static void pnv_pci_ioda1_tce_invalidate(struct pnv_ioda_pe *pe,
509 					 struct iommu_table *tbl,
510 					 __be64 *startp, __be64 *endp, bool rm)
511 {
512 	__be64 __iomem *invalidate = rm ?
513 		(__be64 __iomem *)pe->tce_inval_reg_phys :
514 		(__be64 __iomem *)tbl->it_index;
515 	unsigned long start, end, inc;
516 
517 	start = __pa(startp);
518 	end = __pa(endp);
519 
520 	/* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
521 	if (tbl->it_busno) {
522 		start <<= 12;
523 		end <<= 12;
524 		inc = 128 << 12;
525 		start |= tbl->it_busno;
526 		end |= tbl->it_busno;
527 	} else if (tbl->it_type & TCE_PCI_SWINV_PAIR) {
528 		/* p7ioc-style invalidation, 2 TCEs per write */
529 		start |= (1ull << 63);
530 		end |= (1ull << 63);
531 		inc = 16;
532         } else {
533 		/* Default (older HW) */
534                 inc = 128;
535 	}
536 
537         end |= inc - 1;	/* round up end to be different than start */
538 
539         mb(); /* Ensure above stores are visible */
540         while (start <= end) {
541 		if (rm)
542 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
543 		else
544 			__raw_writeq(cpu_to_be64(start), invalidate);
545                 start += inc;
546         }
547 
548 	/*
549 	 * The iommu layer will do another mb() for us on build()
550 	 * and we don't care on free()
551 	 */
552 }
553 
554 static void pnv_pci_ioda2_tce_invalidate(struct pnv_ioda_pe *pe,
555 					 struct iommu_table *tbl,
556 					 __be64 *startp, __be64 *endp, bool rm)
557 {
558 	unsigned long start, end, inc;
559 	__be64 __iomem *invalidate = rm ?
560 		(__be64 __iomem *)pe->tce_inval_reg_phys :
561 		(__be64 __iomem *)tbl->it_index;
562 
563 	/* We'll invalidate DMA address in PE scope */
564 	start = 0x2ul << 60;
565 	start |= (pe->pe_number & 0xFF);
566 	end = start;
567 
568 	/* Figure out the start, end and step */
569 	inc = tbl->it_offset + (((u64)startp - tbl->it_base) / sizeof(u64));
570 	start |= (inc << 12);
571 	inc = tbl->it_offset + (((u64)endp - tbl->it_base) / sizeof(u64));
572 	end |= (inc << 12);
573 	inc = (0x1ul << 12);
574 	mb();
575 
576 	while (start <= end) {
577 		if (rm)
578 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
579 		else
580 			__raw_writeq(cpu_to_be64(start), invalidate);
581 		start += inc;
582 	}
583 }
584 
585 void pnv_pci_ioda_tce_invalidate(struct iommu_table *tbl,
586 				 __be64 *startp, __be64 *endp, bool rm)
587 {
588 	struct pnv_ioda_pe *pe = container_of(tbl, struct pnv_ioda_pe,
589 					      tce32_table);
590 	struct pnv_phb *phb = pe->phb;
591 
592 	if (phb->type == PNV_PHB_IODA1)
593 		pnv_pci_ioda1_tce_invalidate(pe, tbl, startp, endp, rm);
594 	else
595 		pnv_pci_ioda2_tce_invalidate(pe, tbl, startp, endp, rm);
596 }
597 
598 static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,
599 				      struct pnv_ioda_pe *pe, unsigned int base,
600 				      unsigned int segs)
601 {
602 
603 	struct page *tce_mem = NULL;
604 	const __be64 *swinvp;
605 	struct iommu_table *tbl;
606 	unsigned int i;
607 	int64_t rc;
608 	void *addr;
609 
610 	/* 256M DMA window, 4K TCE pages, 8 bytes TCE */
611 #define TCE32_TABLE_SIZE	((0x10000000 / 0x1000) * 8)
612 
613 	/* XXX FIXME: Handle 64-bit only DMA devices */
614 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
615 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
616 
617 	/* We shouldn't already have a 32-bit DMA associated */
618 	if (WARN_ON(pe->tce32_seg >= 0))
619 		return;
620 
621 	/* Grab a 32-bit TCE table */
622 	pe->tce32_seg = base;
623 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
624 		(base << 28), ((base + segs) << 28) - 1);
625 
626 	/* XXX Currently, we allocate one big contiguous table for the
627 	 * TCEs. We only really need one chunk per 256M of TCE space
628 	 * (ie per segment) but that's an optimization for later, it
629 	 * requires some added smarts with our get/put_tce implementation
630 	 */
631 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
632 				   get_order(TCE32_TABLE_SIZE * segs));
633 	if (!tce_mem) {
634 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
635 		goto fail;
636 	}
637 	addr = page_address(tce_mem);
638 	memset(addr, 0, TCE32_TABLE_SIZE * segs);
639 
640 	/* Configure HW */
641 	for (i = 0; i < segs; i++) {
642 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
643 					      pe->pe_number,
644 					      base + i, 1,
645 					      __pa(addr) + TCE32_TABLE_SIZE * i,
646 					      TCE32_TABLE_SIZE, 0x1000);
647 		if (rc) {
648 			pe_err(pe, " Failed to configure 32-bit TCE table,"
649 			       " err %ld\n", rc);
650 			goto fail;
651 		}
652 	}
653 
654 	/* Setup linux iommu table */
655 	tbl = &pe->tce32_table;
656 	pnv_pci_setup_iommu_table(tbl, addr, TCE32_TABLE_SIZE * segs,
657 				  base << 28);
658 
659 	/* OPAL variant of P7IOC SW invalidated TCEs */
660 	swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
661 	if (swinvp) {
662 		/* We need a couple more fields -- an address and a data
663 		 * to or.  Since the bus is only printed out on table free
664 		 * errors, and on the first pass the data will be a relative
665 		 * bus number, print that out instead.
666 		 */
667 		tbl->it_busno = 0;
668 		pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
669 		tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
670 				8);
671 		tbl->it_type = TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE |
672 			       TCE_PCI_SWINV_PAIR;
673 	}
674 	iommu_init_table(tbl, phb->hose->node);
675 	iommu_register_group(tbl, pci_domain_nr(pe->pbus), pe->pe_number);
676 
677 	if (pe->pdev)
678 		set_iommu_table_base_and_group(&pe->pdev->dev, tbl);
679 	else
680 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
681 
682 	return;
683  fail:
684 	/* XXX Failure: Try to fallback to 64-bit only ? */
685 	if (pe->tce32_seg >= 0)
686 		pe->tce32_seg = -1;
687 	if (tce_mem)
688 		__free_pages(tce_mem, get_order(TCE32_TABLE_SIZE * segs));
689 }
690 
691 static void pnv_pci_ioda2_set_bypass(struct iommu_table *tbl, bool enable)
692 {
693 	struct pnv_ioda_pe *pe = container_of(tbl, struct pnv_ioda_pe,
694 					      tce32_table);
695 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
696 	int64_t rc;
697 
698 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
699 	if (enable) {
700 		phys_addr_t top = memblock_end_of_DRAM();
701 
702 		top = roundup_pow_of_two(top);
703 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
704 						     pe->pe_number,
705 						     window_id,
706 						     pe->tce_bypass_base,
707 						     top);
708 	} else {
709 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
710 						     pe->pe_number,
711 						     window_id,
712 						     pe->tce_bypass_base,
713 						     0);
714 
715 		/*
716 		 * We might want to reset the DMA ops of all devices on
717 		 * this PE. However in theory, that shouldn't be necessary
718 		 * as this is used for VFIO/KVM pass-through and the device
719 		 * hasn't yet been returned to its kernel driver
720 		 */
721 	}
722 	if (rc)
723 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
724 	else
725 		pe->tce_bypass_enabled = enable;
726 }
727 
728 static void pnv_pci_ioda2_setup_bypass_pe(struct pnv_phb *phb,
729 					  struct pnv_ioda_pe *pe)
730 {
731 	/* TVE #1 is selected by PCI address bit 59 */
732 	pe->tce_bypass_base = 1ull << 59;
733 
734 	/* Install set_bypass callback for VFIO */
735 	pe->tce32_table.set_bypass = pnv_pci_ioda2_set_bypass;
736 
737 	/* Enable bypass by default */
738 	pnv_pci_ioda2_set_bypass(&pe->tce32_table, true);
739 }
740 
741 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
742 				       struct pnv_ioda_pe *pe)
743 {
744 	struct page *tce_mem = NULL;
745 	void *addr;
746 	const __be64 *swinvp;
747 	struct iommu_table *tbl;
748 	unsigned int tce_table_size, end;
749 	int64_t rc;
750 
751 	/* We shouldn't already have a 32-bit DMA associated */
752 	if (WARN_ON(pe->tce32_seg >= 0))
753 		return;
754 
755 	/* The PE will reserve all possible 32-bits space */
756 	pe->tce32_seg = 0;
757 	end = (1 << ilog2(phb->ioda.m32_pci_base));
758 	tce_table_size = (end / 0x1000) * 8;
759 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
760 		end);
761 
762 	/* Allocate TCE table */
763 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
764 				   get_order(tce_table_size));
765 	if (!tce_mem) {
766 		pe_err(pe, "Failed to allocate a 32-bit TCE memory\n");
767 		goto fail;
768 	}
769 	addr = page_address(tce_mem);
770 	memset(addr, 0, tce_table_size);
771 
772 	/*
773 	 * Map TCE table through TVT. The TVE index is the PE number
774 	 * shifted by 1 bit for 32-bits DMA space.
775 	 */
776 	rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
777 					pe->pe_number << 1, 1, __pa(addr),
778 					tce_table_size, 0x1000);
779 	if (rc) {
780 		pe_err(pe, "Failed to configure 32-bit TCE table,"
781 		       " err %ld\n", rc);
782 		goto fail;
783 	}
784 
785 	/* Setup linux iommu table */
786 	tbl = &pe->tce32_table;
787 	pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, 0);
788 
789 	/* OPAL variant of PHB3 invalidated TCEs */
790 	swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
791 	if (swinvp) {
792 		/* We need a couple more fields -- an address and a data
793 		 * to or.  Since the bus is only printed out on table free
794 		 * errors, and on the first pass the data will be a relative
795 		 * bus number, print that out instead.
796 		 */
797 		tbl->it_busno = 0;
798 		pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
799 		tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
800 				8);
801 		tbl->it_type = TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE;
802 	}
803 	iommu_init_table(tbl, phb->hose->node);
804 	iommu_register_group(tbl, pci_domain_nr(pe->pbus), pe->pe_number);
805 
806 	if (pe->pdev)
807 		set_iommu_table_base_and_group(&pe->pdev->dev, tbl);
808 	else
809 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
810 
811 	/* Also create a bypass window */
812 	pnv_pci_ioda2_setup_bypass_pe(phb, pe);
813 	return;
814 fail:
815 	if (pe->tce32_seg >= 0)
816 		pe->tce32_seg = -1;
817 	if (tce_mem)
818 		__free_pages(tce_mem, get_order(tce_table_size));
819 }
820 
821 static void pnv_ioda_setup_dma(struct pnv_phb *phb)
822 {
823 	struct pci_controller *hose = phb->hose;
824 	unsigned int residual, remaining, segs, tw, base;
825 	struct pnv_ioda_pe *pe;
826 
827 	/* If we have more PE# than segments available, hand out one
828 	 * per PE until we run out and let the rest fail. If not,
829 	 * then we assign at least one segment per PE, plus more based
830 	 * on the amount of devices under that PE
831 	 */
832 	if (phb->ioda.dma_pe_count > phb->ioda.tce32_count)
833 		residual = 0;
834 	else
835 		residual = phb->ioda.tce32_count -
836 			phb->ioda.dma_pe_count;
837 
838 	pr_info("PCI: Domain %04x has %ld available 32-bit DMA segments\n",
839 		hose->global_number, phb->ioda.tce32_count);
840 	pr_info("PCI: %d PE# for a total weight of %d\n",
841 		phb->ioda.dma_pe_count, phb->ioda.dma_weight);
842 
843 	/* Walk our PE list and configure their DMA segments, hand them
844 	 * out one base segment plus any residual segments based on
845 	 * weight
846 	 */
847 	remaining = phb->ioda.tce32_count;
848 	tw = phb->ioda.dma_weight;
849 	base = 0;
850 	list_for_each_entry(pe, &phb->ioda.pe_dma_list, dma_link) {
851 		if (!pe->dma_weight)
852 			continue;
853 		if (!remaining) {
854 			pe_warn(pe, "No DMA32 resources available\n");
855 			continue;
856 		}
857 		segs = 1;
858 		if (residual) {
859 			segs += ((pe->dma_weight * residual)  + (tw / 2)) / tw;
860 			if (segs > remaining)
861 				segs = remaining;
862 		}
863 
864 		/*
865 		 * For IODA2 compliant PHB3, we needn't care about the weight.
866 		 * The all available 32-bits DMA space will be assigned to
867 		 * the specific PE.
868 		 */
869 		if (phb->type == PNV_PHB_IODA1) {
870 			pe_info(pe, "DMA weight %d, assigned %d DMA32 segments\n",
871 				pe->dma_weight, segs);
872 			pnv_pci_ioda_setup_dma_pe(phb, pe, base, segs);
873 		} else {
874 			pe_info(pe, "Assign DMA32 space\n");
875 			segs = 0;
876 			pnv_pci_ioda2_setup_dma_pe(phb, pe);
877 		}
878 
879 		remaining -= segs;
880 		base += segs;
881 	}
882 }
883 
884 #ifdef CONFIG_PCI_MSI
885 static void pnv_ioda2_msi_eoi(struct irq_data *d)
886 {
887 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
888 	struct irq_chip *chip = irq_data_get_irq_chip(d);
889 	struct pnv_phb *phb = container_of(chip, struct pnv_phb,
890 					   ioda.irq_chip);
891 	int64_t rc;
892 
893 	rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
894 	WARN_ON_ONCE(rc);
895 
896 	icp_native_eoi(d);
897 }
898 
899 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
900 				  unsigned int hwirq, unsigned int virq,
901 				  unsigned int is_64, struct msi_msg *msg)
902 {
903 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
904 	struct pci_dn *pdn = pci_get_pdn(dev);
905 	struct irq_data *idata;
906 	struct irq_chip *ichip;
907 	unsigned int xive_num = hwirq - phb->msi_base;
908 	__be32 data;
909 	int rc;
910 
911 	/* No PE assigned ? bail out ... no MSI for you ! */
912 	if (pe == NULL)
913 		return -ENXIO;
914 
915 	/* Check if we have an MVE */
916 	if (pe->mve_number < 0)
917 		return -ENXIO;
918 
919 	/* Force 32-bit MSI on some broken devices */
920 	if (pdn && pdn->force_32bit_msi)
921 		is_64 = 0;
922 
923 	/* Assign XIVE to PE */
924 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
925 	if (rc) {
926 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
927 			pci_name(dev), rc, xive_num);
928 		return -EIO;
929 	}
930 
931 	if (is_64) {
932 		__be64 addr64;
933 
934 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
935 				     &addr64, &data);
936 		if (rc) {
937 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
938 				pci_name(dev), rc);
939 			return -EIO;
940 		}
941 		msg->address_hi = be64_to_cpu(addr64) >> 32;
942 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
943 	} else {
944 		__be32 addr32;
945 
946 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
947 				     &addr32, &data);
948 		if (rc) {
949 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
950 				pci_name(dev), rc);
951 			return -EIO;
952 		}
953 		msg->address_hi = 0;
954 		msg->address_lo = be32_to_cpu(addr32);
955 	}
956 	msg->data = be32_to_cpu(data);
957 
958 	/*
959 	 * Change the IRQ chip for the MSI interrupts on PHB3.
960 	 * The corresponding IRQ chip should be populated for
961 	 * the first time.
962 	 */
963 	if (phb->type == PNV_PHB_IODA2) {
964 		if (!phb->ioda.irq_chip_init) {
965 			idata = irq_get_irq_data(virq);
966 			ichip = irq_data_get_irq_chip(idata);
967 			phb->ioda.irq_chip_init = 1;
968 			phb->ioda.irq_chip = *ichip;
969 			phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
970 		}
971 
972 		irq_set_chip(virq, &phb->ioda.irq_chip);
973 	}
974 
975 	pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
976 		 " address=%x_%08x data=%x PE# %d\n",
977 		 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
978 		 msg->address_hi, msg->address_lo, data, pe->pe_number);
979 
980 	return 0;
981 }
982 
983 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
984 {
985 	unsigned int count;
986 	const __be32 *prop = of_get_property(phb->hose->dn,
987 					     "ibm,opal-msi-ranges", NULL);
988 	if (!prop) {
989 		/* BML Fallback */
990 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
991 	}
992 	if (!prop)
993 		return;
994 
995 	phb->msi_base = be32_to_cpup(prop);
996 	count = be32_to_cpup(prop + 1);
997 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
998 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
999 		       phb->hose->global_number);
1000 		return;
1001 	}
1002 
1003 	phb->msi_setup = pnv_pci_ioda_msi_setup;
1004 	phb->msi32_support = 1;
1005 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
1006 		count, phb->msi_base);
1007 }
1008 #else
1009 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
1010 #endif /* CONFIG_PCI_MSI */
1011 
1012 /*
1013  * This function is supposed to be called on basis of PE from top
1014  * to bottom style. So the the I/O or MMIO segment assigned to
1015  * parent PE could be overrided by its child PEs if necessary.
1016  */
1017 static void pnv_ioda_setup_pe_seg(struct pci_controller *hose,
1018 				  struct pnv_ioda_pe *pe)
1019 {
1020 	struct pnv_phb *phb = hose->private_data;
1021 	struct pci_bus_region region;
1022 	struct resource *res;
1023 	int i, index;
1024 	int rc;
1025 
1026 	/*
1027 	 * NOTE: We only care PCI bus based PE for now. For PCI
1028 	 * device based PE, for example SRIOV sensitive VF should
1029 	 * be figured out later.
1030 	 */
1031 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
1032 
1033 	pci_bus_for_each_resource(pe->pbus, res, i) {
1034 		if (!res || !res->flags ||
1035 		    res->start > res->end)
1036 			continue;
1037 
1038 		if (res->flags & IORESOURCE_IO) {
1039 			region.start = res->start - phb->ioda.io_pci_base;
1040 			region.end   = res->end - phb->ioda.io_pci_base;
1041 			index = region.start / phb->ioda.io_segsize;
1042 
1043 			while (index < phb->ioda.total_pe &&
1044 			       region.start <= region.end) {
1045 				phb->ioda.io_segmap[index] = pe->pe_number;
1046 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1047 					pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
1048 				if (rc != OPAL_SUCCESS) {
1049 					pr_err("%s: OPAL error %d when mapping IO "
1050 					       "segment #%d to PE#%d\n",
1051 					       __func__, rc, index, pe->pe_number);
1052 					break;
1053 				}
1054 
1055 				region.start += phb->ioda.io_segsize;
1056 				index++;
1057 			}
1058 		} else if (res->flags & IORESOURCE_MEM) {
1059 			/* WARNING: Assumes M32 is mem region 0 in PHB. We need to
1060 			 * harden that algorithm when we start supporting M64
1061 			 */
1062 			region.start = res->start -
1063 				       hose->mem_offset[0] -
1064 				       phb->ioda.m32_pci_base;
1065 			region.end   = res->end -
1066 				       hose->mem_offset[0] -
1067 				       phb->ioda.m32_pci_base;
1068 			index = region.start / phb->ioda.m32_segsize;
1069 
1070 			while (index < phb->ioda.total_pe &&
1071 			       region.start <= region.end) {
1072 				phb->ioda.m32_segmap[index] = pe->pe_number;
1073 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1074 					pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
1075 				if (rc != OPAL_SUCCESS) {
1076 					pr_err("%s: OPAL error %d when mapping M32 "
1077 					       "segment#%d to PE#%d",
1078 					       __func__, rc, index, pe->pe_number);
1079 					break;
1080 				}
1081 
1082 				region.start += phb->ioda.m32_segsize;
1083 				index++;
1084 			}
1085 		}
1086 	}
1087 }
1088 
1089 static void pnv_pci_ioda_setup_seg(void)
1090 {
1091 	struct pci_controller *tmp, *hose;
1092 	struct pnv_phb *phb;
1093 	struct pnv_ioda_pe *pe;
1094 
1095 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1096 		phb = hose->private_data;
1097 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
1098 			pnv_ioda_setup_pe_seg(hose, pe);
1099 		}
1100 	}
1101 }
1102 
1103 static void pnv_pci_ioda_setup_DMA(void)
1104 {
1105 	struct pci_controller *hose, *tmp;
1106 	struct pnv_phb *phb;
1107 
1108 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1109 		pnv_ioda_setup_dma(hose->private_data);
1110 
1111 		/* Mark the PHB initialization done */
1112 		phb = hose->private_data;
1113 		phb->initialized = 1;
1114 	}
1115 }
1116 
1117 static void pnv_pci_ioda_create_dbgfs(void)
1118 {
1119 #ifdef CONFIG_DEBUG_FS
1120 	struct pci_controller *hose, *tmp;
1121 	struct pnv_phb *phb;
1122 	char name[16];
1123 
1124 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1125 		phb = hose->private_data;
1126 
1127 		sprintf(name, "PCI%04x", hose->global_number);
1128 		phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
1129 		if (!phb->dbgfs)
1130 			pr_warning("%s: Error on creating debugfs on PHB#%x\n",
1131 				__func__, hose->global_number);
1132 	}
1133 #endif /* CONFIG_DEBUG_FS */
1134 }
1135 
1136 static void pnv_pci_ioda_fixup(void)
1137 {
1138 	pnv_pci_ioda_setup_PEs();
1139 	pnv_pci_ioda_setup_seg();
1140 	pnv_pci_ioda_setup_DMA();
1141 
1142 	pnv_pci_ioda_create_dbgfs();
1143 
1144 #ifdef CONFIG_EEH
1145 	eeh_probe_mode_set(EEH_PROBE_MODE_DEV);
1146 	eeh_addr_cache_build();
1147 	eeh_init();
1148 #endif
1149 }
1150 
1151 /*
1152  * Returns the alignment for I/O or memory windows for P2P
1153  * bridges. That actually depends on how PEs are segmented.
1154  * For now, we return I/O or M32 segment size for PE sensitive
1155  * P2P bridges. Otherwise, the default values (4KiB for I/O,
1156  * 1MiB for memory) will be returned.
1157  *
1158  * The current PCI bus might be put into one PE, which was
1159  * create against the parent PCI bridge. For that case, we
1160  * needn't enlarge the alignment so that we can save some
1161  * resources.
1162  */
1163 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
1164 						unsigned long type)
1165 {
1166 	struct pci_dev *bridge;
1167 	struct pci_controller *hose = pci_bus_to_host(bus);
1168 	struct pnv_phb *phb = hose->private_data;
1169 	int num_pci_bridges = 0;
1170 
1171 	bridge = bus->self;
1172 	while (bridge) {
1173 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
1174 			num_pci_bridges++;
1175 			if (num_pci_bridges >= 2)
1176 				return 1;
1177 		}
1178 
1179 		bridge = bridge->bus->self;
1180 	}
1181 
1182 	/* We need support prefetchable memory window later */
1183 	if (type & IORESOURCE_MEM)
1184 		return phb->ioda.m32_segsize;
1185 
1186 	return phb->ioda.io_segsize;
1187 }
1188 
1189 /* Prevent enabling devices for which we couldn't properly
1190  * assign a PE
1191  */
1192 static int pnv_pci_enable_device_hook(struct pci_dev *dev)
1193 {
1194 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
1195 	struct pnv_phb *phb = hose->private_data;
1196 	struct pci_dn *pdn;
1197 
1198 	/* The function is probably called while the PEs have
1199 	 * not be created yet. For example, resource reassignment
1200 	 * during PCI probe period. We just skip the check if
1201 	 * PEs isn't ready.
1202 	 */
1203 	if (!phb->initialized)
1204 		return 0;
1205 
1206 	pdn = pci_get_pdn(dev);
1207 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1208 		return -EINVAL;
1209 
1210 	return 0;
1211 }
1212 
1213 static u32 pnv_ioda_bdfn_to_pe(struct pnv_phb *phb, struct pci_bus *bus,
1214 			       u32 devfn)
1215 {
1216 	return phb->ioda.pe_rmap[(bus->number << 8) | devfn];
1217 }
1218 
1219 static void pnv_pci_ioda_shutdown(struct pnv_phb *phb)
1220 {
1221 	opal_pci_reset(phb->opal_id, OPAL_PCI_IODA_TABLE_RESET,
1222 		       OPAL_ASSERT_RESET);
1223 }
1224 
1225 void __init pnv_pci_init_ioda_phb(struct device_node *np,
1226 				  u64 hub_id, int ioda_type)
1227 {
1228 	struct pci_controller *hose;
1229 	struct pnv_phb *phb;
1230 	unsigned long size, m32map_off, pemap_off, iomap_off = 0;
1231 	const __be64 *prop64;
1232 	const __be32 *prop32;
1233 	int len;
1234 	u64 phb_id;
1235 	void *aux;
1236 	long rc;
1237 
1238 	pr_info("Initializing IODA%d OPAL PHB %s\n", ioda_type, np->full_name);
1239 
1240 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
1241 	if (!prop64) {
1242 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
1243 		return;
1244 	}
1245 	phb_id = be64_to_cpup(prop64);
1246 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
1247 
1248 	phb = alloc_bootmem(sizeof(struct pnv_phb));
1249 	if (!phb) {
1250 		pr_err("  Out of memory !\n");
1251 		return;
1252 	}
1253 
1254 	/* Allocate PCI controller */
1255 	memset(phb, 0, sizeof(struct pnv_phb));
1256 	phb->hose = hose = pcibios_alloc_controller(np);
1257 	if (!phb->hose) {
1258 		pr_err("  Can't allocate PCI controller for %s\n",
1259 		       np->full_name);
1260 		free_bootmem((unsigned long)phb, sizeof(struct pnv_phb));
1261 		return;
1262 	}
1263 
1264 	spin_lock_init(&phb->lock);
1265 	prop32 = of_get_property(np, "bus-range", &len);
1266 	if (prop32 && len == 8) {
1267 		hose->first_busno = be32_to_cpu(prop32[0]);
1268 		hose->last_busno = be32_to_cpu(prop32[1]);
1269 	} else {
1270 		pr_warn("  Broken <bus-range> on %s\n", np->full_name);
1271 		hose->first_busno = 0;
1272 		hose->last_busno = 0xff;
1273 	}
1274 	hose->private_data = phb;
1275 	phb->hub_id = hub_id;
1276 	phb->opal_id = phb_id;
1277 	phb->type = ioda_type;
1278 
1279 	/* Detect specific models for error handling */
1280 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
1281 		phb->model = PNV_PHB_MODEL_P7IOC;
1282 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
1283 		phb->model = PNV_PHB_MODEL_PHB3;
1284 	else
1285 		phb->model = PNV_PHB_MODEL_UNKNOWN;
1286 
1287 	/* Parse 32-bit and IO ranges (if any) */
1288 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
1289 
1290 	/* Get registers */
1291 	phb->regs = of_iomap(np, 0);
1292 	if (phb->regs == NULL)
1293 		pr_err("  Failed to map registers !\n");
1294 
1295 	/* Initialize more IODA stuff */
1296 	phb->ioda.total_pe = 1;
1297 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
1298 	if (prop32)
1299 		phb->ioda.total_pe = be32_to_cpup(prop32);
1300 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
1301 	if (prop32)
1302 		phb->ioda.reserved_pe = be32_to_cpup(prop32);
1303 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
1304 	/* FW Has already off top 64k of M32 space (MSI space) */
1305 	phb->ioda.m32_size += 0x10000;
1306 
1307 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe;
1308 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
1309 	phb->ioda.io_size = hose->pci_io_size;
1310 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe;
1311 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
1312 
1313 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
1314 	size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long));
1315 	m32map_off = size;
1316 	size += phb->ioda.total_pe * sizeof(phb->ioda.m32_segmap[0]);
1317 	if (phb->type == PNV_PHB_IODA1) {
1318 		iomap_off = size;
1319 		size += phb->ioda.total_pe * sizeof(phb->ioda.io_segmap[0]);
1320 	}
1321 	pemap_off = size;
1322 	size += phb->ioda.total_pe * sizeof(struct pnv_ioda_pe);
1323 	aux = alloc_bootmem(size);
1324 	memset(aux, 0, size);
1325 	phb->ioda.pe_alloc = aux;
1326 	phb->ioda.m32_segmap = aux + m32map_off;
1327 	if (phb->type == PNV_PHB_IODA1)
1328 		phb->ioda.io_segmap = aux + iomap_off;
1329 	phb->ioda.pe_array = aux + pemap_off;
1330 	set_bit(phb->ioda.reserved_pe, phb->ioda.pe_alloc);
1331 
1332 	INIT_LIST_HEAD(&phb->ioda.pe_dma_list);
1333 	INIT_LIST_HEAD(&phb->ioda.pe_list);
1334 
1335 	/* Calculate how many 32-bit TCE segments we have */
1336 	phb->ioda.tce32_count = phb->ioda.m32_pci_base >> 28;
1337 
1338 	/* Clear unusable m64 */
1339 	hose->mem_resources[1].flags = 0;
1340 	hose->mem_resources[1].start = 0;
1341 	hose->mem_resources[1].end = 0;
1342 	hose->mem_resources[2].flags = 0;
1343 	hose->mem_resources[2].start = 0;
1344 	hose->mem_resources[2].end = 0;
1345 
1346 #if 0 /* We should really do that ... */
1347 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
1348 					 window_type,
1349 					 window_num,
1350 					 starting_real_address,
1351 					 starting_pci_address,
1352 					 segment_size);
1353 #endif
1354 
1355 	pr_info("  %d (%d) PE's M32: 0x%x [segment=0x%x]"
1356 		" IO: 0x%x [segment=0x%x]\n",
1357 		phb->ioda.total_pe,
1358 		phb->ioda.reserved_pe,
1359 		phb->ioda.m32_size, phb->ioda.m32_segsize,
1360 		phb->ioda.io_size, phb->ioda.io_segsize);
1361 
1362 	phb->hose->ops = &pnv_pci_ops;
1363 #ifdef CONFIG_EEH
1364 	phb->eeh_ops = &ioda_eeh_ops;
1365 #endif
1366 
1367 	/* Setup RID -> PE mapping function */
1368 	phb->bdfn_to_pe = pnv_ioda_bdfn_to_pe;
1369 
1370 	/* Setup TCEs */
1371 	phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
1372 	phb->dma_set_mask = pnv_pci_ioda_dma_set_mask;
1373 
1374 	/* Setup shutdown function for kexec */
1375 	phb->shutdown = pnv_pci_ioda_shutdown;
1376 
1377 	/* Setup MSI support */
1378 	pnv_pci_init_ioda_msis(phb);
1379 
1380 	/*
1381 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
1382 	 * to let the PCI core do resource assignment. It's supposed
1383 	 * that the PCI core will do correct I/O and MMIO alignment
1384 	 * for the P2P bridge bars so that each PCI bus (excluding
1385 	 * the child P2P bridges) can form individual PE.
1386 	 */
1387 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
1388 	ppc_md.pcibios_enable_device_hook = pnv_pci_enable_device_hook;
1389 	ppc_md.pcibios_window_alignment = pnv_pci_window_alignment;
1390 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
1391 
1392 	/* Reset IODA tables to a clean state */
1393 	rc = opal_pci_reset(phb_id, OPAL_PCI_IODA_TABLE_RESET, OPAL_ASSERT_RESET);
1394 	if (rc)
1395 		pr_warning("  OPAL Error %ld performing IODA table reset !\n", rc);
1396 }
1397 
1398 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
1399 {
1400 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
1401 }
1402 
1403 void __init pnv_pci_init_ioda_hub(struct device_node *np)
1404 {
1405 	struct device_node *phbn;
1406 	const __be64 *prop64;
1407 	u64 hub_id;
1408 
1409 	pr_info("Probing IODA IO-Hub %s\n", np->full_name);
1410 
1411 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
1412 	if (!prop64) {
1413 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
1414 		return;
1415 	}
1416 	hub_id = be64_to_cpup(prop64);
1417 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
1418 
1419 	/* Count child PHBs */
1420 	for_each_child_of_node(np, phbn) {
1421 		/* Look for IODA1 PHBs */
1422 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
1423 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
1424 	}
1425 }
1426