2 * Support PCI/PCIe on PowerNV platforms
4 * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
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.
14 #include <linux/kernel.h>
15 #include <linux/pci.h>
16 #include <linux/crash_dump.h>
17 #include <linux/debugfs.h>
18 #include <linux/delay.h>
19 #include <linux/string.h>
20 #include <linux/init.h>
21 #include <linux/bootmem.h>
22 #include <linux/irq.h>
24 #include <linux/msi.h>
25 #include <linux/memblock.h>
27 #include <asm/sections.h>
30 #include <asm/pci-bridge.h>
31 #include <asm/machdep.h>
32 #include <asm/msi_bitmap.h>
33 #include <asm/ppc-pci.h>
35 #include <asm/iommu.h>
38 #include <asm/debug.h>
43 #define define_pe_printk_level(func, kern_level) \
44 static int func(const struct pnv_ioda_pe *pe, const char *fmt, ...) \
46 struct va_format vaf; \
51 va_start(args, fmt); \
57 strlcpy(pfix, dev_name(&pe->pdev->dev), \
60 sprintf(pfix, "%04x:%02x ", \
61 pci_domain_nr(pe->pbus), \
63 r = printk(kern_level "pci %s: [PE# %.3d] %pV", \
64 pfix, pe->pe_number, &vaf); \
71 define_pe_printk_level(pe_err, KERN_ERR);
72 define_pe_printk_level(pe_warn, KERN_WARNING);
73 define_pe_printk_level(pe_info, KERN_INFO);
76 * stdcix is only supposed to be used in hypervisor real mode as per
77 * the architecture spec
79 static inline void __raw_rm_writeq(u64 val, volatile void __iomem *paddr)
81 __asm__ __volatile__("stdcix %0,0,%1"
82 : : "r" (val), "r" (paddr) : "memory");
85 static int pnv_ioda_alloc_pe(struct pnv_phb *phb)
90 pe = find_next_zero_bit(phb->ioda.pe_alloc,
91 phb->ioda.total_pe, 0);
92 if (pe >= phb->ioda.total_pe)
93 return IODA_INVALID_PE;
94 } while(test_and_set_bit(pe, phb->ioda.pe_alloc));
96 phb->ioda.pe_array[pe].phb = phb;
97 phb->ioda.pe_array[pe].pe_number = pe;
101 static void pnv_ioda_free_pe(struct pnv_phb *phb, int pe)
103 WARN_ON(phb->ioda.pe_array[pe].pdev);
105 memset(&phb->ioda.pe_array[pe], 0, sizeof(struct pnv_ioda_pe));
106 clear_bit(pe, phb->ioda.pe_alloc);
109 /* Currently those 2 are only used when MSIs are enabled, this will change
110 * but in the meantime, we need to protect them to avoid warnings
112 #ifdef CONFIG_PCI_MSI
113 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
115 struct pci_controller *hose = pci_bus_to_host(dev->bus);
116 struct pnv_phb *phb = hose->private_data;
117 struct pci_dn *pdn = pci_get_pdn(dev);
121 if (pdn->pe_number == IODA_INVALID_PE)
123 return &phb->ioda.pe_array[pdn->pe_number];
125 #endif /* CONFIG_PCI_MSI */
127 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
129 struct pci_dev *parent;
130 uint8_t bcomp, dcomp, fcomp;
131 long rc, rid_end, rid;
133 /* Bus validation ? */
137 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
138 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
139 parent = pe->pbus->self;
140 if (pe->flags & PNV_IODA_PE_BUS_ALL)
141 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
146 case 1: bcomp = OpalPciBusAll; break;
147 case 2: bcomp = OpalPciBus7Bits; break;
148 case 4: bcomp = OpalPciBus6Bits; break;
149 case 8: bcomp = OpalPciBus5Bits; break;
150 case 16: bcomp = OpalPciBus4Bits; break;
151 case 32: bcomp = OpalPciBus3Bits; break;
153 pr_err("%s: Number of subordinate busses %d"
155 pci_name(pe->pbus->self), count);
156 /* Do an exact match only */
157 bcomp = OpalPciBusAll;
159 rid_end = pe->rid + (count << 8);
161 parent = pe->pdev->bus->self;
162 bcomp = OpalPciBusAll;
163 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
164 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
165 rid_end = pe->rid + 1;
169 * Associate PE in PELT. We need add the PE into the
170 * corresponding PELT-V as well. Otherwise, the error
171 * originated from the PE might contribute to other
174 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
175 bcomp, dcomp, fcomp, OPAL_MAP_PE);
177 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
181 rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
182 pe->pe_number, OPAL_ADD_PE_TO_DOMAIN);
184 pe_warn(pe, "OPAL error %d adding self to PELTV\n", rc);
185 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
186 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
188 /* Add to all parents PELT-V */
190 struct pci_dn *pdn = pci_get_pdn(parent);
191 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
192 rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
193 pe->pe_number, OPAL_ADD_PE_TO_DOMAIN);
194 /* XXX What to do in case of error ? */
196 parent = parent->bus->self;
198 /* Setup reverse map */
199 for (rid = pe->rid; rid < rid_end; rid++)
200 phb->ioda.pe_rmap[rid] = pe->pe_number;
202 /* Setup one MVTs on IODA1 */
203 if (phb->type == PNV_PHB_IODA1) {
204 pe->mve_number = pe->pe_number;
205 rc = opal_pci_set_mve(phb->opal_id, pe->mve_number,
208 pe_err(pe, "OPAL error %ld setting up MVE %d\n",
212 rc = opal_pci_set_mve_enable(phb->opal_id,
213 pe->mve_number, OPAL_ENABLE_MVE);
215 pe_err(pe, "OPAL error %ld enabling MVE %d\n",
220 } else if (phb->type == PNV_PHB_IODA2)
226 static void pnv_ioda_link_pe_by_weight(struct pnv_phb *phb,
227 struct pnv_ioda_pe *pe)
229 struct pnv_ioda_pe *lpe;
231 list_for_each_entry(lpe, &phb->ioda.pe_dma_list, dma_link) {
232 if (lpe->dma_weight < pe->dma_weight) {
233 list_add_tail(&pe->dma_link, &lpe->dma_link);
237 list_add_tail(&pe->dma_link, &phb->ioda.pe_dma_list);
240 static unsigned int pnv_ioda_dma_weight(struct pci_dev *dev)
242 /* This is quite simplistic. The "base" weight of a device
243 * is 10. 0 means no DMA is to be accounted for it.
246 /* If it's a bridge, no DMA */
247 if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
250 /* Reduce the weight of slow USB controllers */
251 if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
252 dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
253 dev->class == PCI_CLASS_SERIAL_USB_EHCI)
256 /* Increase the weight of RAID (includes Obsidian) */
257 if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
265 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
267 struct pci_controller *hose = pci_bus_to_host(dev->bus);
268 struct pnv_phb *phb = hose->private_data;
269 struct pci_dn *pdn = pci_get_pdn(dev);
270 struct pnv_ioda_pe *pe;
274 pr_err("%s: Device tree node not associated properly\n",
278 if (pdn->pe_number != IODA_INVALID_PE)
281 /* PE#0 has been pre-set */
282 if (dev->bus->number == 0)
285 pe_num = pnv_ioda_alloc_pe(phb);
286 if (pe_num == IODA_INVALID_PE) {
287 pr_warning("%s: Not enough PE# available, disabling device\n",
292 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
293 * pointer in the PE data structure, both should be destroyed at the
294 * same time. However, this needs to be looked at more closely again
295 * once we actually start removing things (Hotplug, SR-IOV, ...)
297 * At some point we want to remove the PDN completely anyways
299 pe = &phb->ioda.pe_array[pe_num];
302 pdn->pe_number = pe_num;
307 pe->rid = dev->bus->number << 8 | pdn->devfn;
309 pe_info(pe, "Associated device to PE\n");
311 if (pnv_ioda_configure_pe(phb, pe)) {
312 /* XXX What do we do here ? */
314 pnv_ioda_free_pe(phb, pe_num);
315 pdn->pe_number = IODA_INVALID_PE;
321 /* Assign a DMA weight to the device */
322 pe->dma_weight = pnv_ioda_dma_weight(dev);
323 if (pe->dma_weight != 0) {
324 phb->ioda.dma_weight += pe->dma_weight;
325 phb->ioda.dma_pe_count++;
329 pnv_ioda_link_pe_by_weight(phb, pe);
333 #endif /* Useful for SRIOV case */
335 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
339 list_for_each_entry(dev, &bus->devices, bus_list) {
340 struct pci_dn *pdn = pci_get_pdn(dev);
343 pr_warn("%s: No device node associated with device !\n",
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);
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.
361 static void pnv_ioda_setup_bus_PE(struct pci_bus *bus, int all)
363 struct pci_controller *hose = pci_bus_to_host(bus);
364 struct pnv_phb *phb = hose->private_data;
365 struct pnv_ioda_pe *pe;
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);
375 pe = &phb->ioda.pe_array[pe_num];
376 pe->flags = (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
381 pe->rid = bus->busn_res.start << 8;
385 pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
386 bus->busn_res.start, bus->busn_res.end, pe_num);
388 pe_info(pe, "Secondary bus %d associated with PE#%d\n",
389 bus->busn_res.start, pe_num);
391 if (pnv_ioda_configure_pe(phb, pe)) {
392 /* XXX What do we do here ? */
394 pnv_ioda_free_pe(phb, pe_num);
399 /* Associate it with all child devices */
400 pnv_ioda_setup_same_PE(bus, pe);
402 /* Put PE to the list */
403 list_add_tail(&pe->list, &phb->ioda.pe_list);
405 /* Account for one DMA PE if at least one DMA capable device exist
408 if (pe->dma_weight != 0) {
409 phb->ioda.dma_weight += pe->dma_weight;
410 phb->ioda.dma_pe_count++;
414 pnv_ioda_link_pe_by_weight(phb, pe);
417 static void pnv_ioda_setup_PEs(struct pci_bus *bus)
421 pnv_ioda_setup_bus_PE(bus, 0);
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);
428 pnv_ioda_setup_PEs(dev->subordinate);
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.
441 static void pnv_pci_ioda_setup_PEs(void)
443 struct pci_controller *hose, *tmp;
445 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
446 pnv_ioda_setup_PEs(hose->bus);
450 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
452 struct pci_dn *pdn = pci_get_pdn(pdev);
453 struct pnv_ioda_pe *pe;
456 * The function can be called while the PE#
457 * hasn't been assigned. Do nothing for the
460 if (!pdn || pdn->pe_number == IODA_INVALID_PE)
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(&pdev->dev, &pe->tce32_table);
468 static int pnv_pci_ioda_dma_set_mask(struct pnv_phb *phb,
469 struct pci_dev *pdev, u64 dma_mask)
471 struct pci_dn *pdn = pci_get_pdn(pdev);
472 struct pnv_ioda_pe *pe;
476 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
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);
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);
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);
497 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus)
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);
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)
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;
517 start = __pa(startp);
520 /* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
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);
533 /* Default (older HW) */
537 end |= inc - 1; /* round up end to be different than start */
539 mb(); /* Ensure above stores are visible */
540 while (start <= end) {
542 __raw_rm_writeq(cpu_to_be64(start), invalidate);
544 __raw_writeq(cpu_to_be64(start), invalidate);
549 * The iommu layer will do another mb() for us on build()
550 * and we don't care on free()
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)
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;
563 /* We'll invalidate DMA address in PE scope */
565 start |= (pe->pe_number & 0xFF);
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));
576 while (start <= end) {
578 __raw_rm_writeq(cpu_to_be64(start), invalidate);
580 __raw_writeq(cpu_to_be64(start), invalidate);
585 void pnv_pci_ioda_tce_invalidate(struct iommu_table *tbl,
586 __be64 *startp, __be64 *endp, bool rm)
588 struct pnv_ioda_pe *pe = container_of(tbl, struct pnv_ioda_pe,
590 struct pnv_phb *phb = pe->phb;
592 if (phb->type == PNV_PHB_IODA1)
593 pnv_pci_ioda1_tce_invalidate(pe, tbl, startp, endp, rm);
595 pnv_pci_ioda2_tce_invalidate(pe, tbl, startp, endp, rm);
598 static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,
599 struct pnv_ioda_pe *pe, unsigned int base,
603 struct page *tce_mem = NULL;
604 const __be64 *swinvp;
605 struct iommu_table *tbl;
610 /* 256M DMA window, 4K TCE pages, 8 bytes TCE */
611 #define TCE32_TABLE_SIZE ((0x10000000 / 0x1000) * 8)
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 */
617 /* We shouldn't already have a 32-bit DMA associated */
618 if (WARN_ON(pe->tce32_seg >= 0))
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);
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
631 tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
632 get_order(TCE32_TABLE_SIZE * segs));
634 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
637 addr = page_address(tce_mem);
638 memset(addr, 0, TCE32_TABLE_SIZE * segs);
641 for (i = 0; i < segs; i++) {
642 rc = opal_pci_map_pe_dma_window(phb->opal_id,
645 __pa(addr) + TCE32_TABLE_SIZE * i,
646 TCE32_TABLE_SIZE, 0x1000);
648 pe_err(pe, " Failed to configure 32-bit TCE table,"
654 /* Setup linux iommu table */
655 tbl = &pe->tce32_table;
656 pnv_pci_setup_iommu_table(tbl, addr, TCE32_TABLE_SIZE * segs,
659 /* OPAL variant of P7IOC SW invalidated TCEs */
660 swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
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.
667 pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
668 tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
670 tbl->it_type |= (TCE_PCI_SWINV_CREATE |
674 iommu_init_table(tbl, phb->hose->node);
675 iommu_register_group(tbl, phb->hose->global_number, pe->pe_number);
678 set_iommu_table_base_and_group(&pe->pdev->dev, tbl);
680 pnv_ioda_setup_bus_dma(pe, pe->pbus);
684 /* XXX Failure: Try to fallback to 64-bit only ? */
685 if (pe->tce32_seg >= 0)
688 __free_pages(tce_mem, get_order(TCE32_TABLE_SIZE * segs));
691 static void pnv_pci_ioda2_set_bypass(struct iommu_table *tbl, bool enable)
693 struct pnv_ioda_pe *pe = container_of(tbl, struct pnv_ioda_pe,
695 uint16_t window_id = (pe->pe_number << 1 ) + 1;
698 pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
700 phys_addr_t top = memblock_end_of_DRAM();
702 top = roundup_pow_of_two(top);
703 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
709 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
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
723 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
725 pe->tce_bypass_enabled = enable;
728 static void pnv_pci_ioda2_setup_bypass_pe(struct pnv_phb *phb,
729 struct pnv_ioda_pe *pe)
731 /* TVE #1 is selected by PCI address bit 59 */
732 pe->tce_bypass_base = 1ull << 59;
734 /* Install set_bypass callback for VFIO */
735 pe->tce32_table.set_bypass = pnv_pci_ioda2_set_bypass;
737 /* Enable bypass by default */
738 pnv_pci_ioda2_set_bypass(&pe->tce32_table, true);
741 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
742 struct pnv_ioda_pe *pe)
744 struct page *tce_mem = NULL;
746 const __be64 *swinvp;
747 struct iommu_table *tbl;
748 unsigned int tce_table_size, end;
751 /* We shouldn't already have a 32-bit DMA associated */
752 if (WARN_ON(pe->tce32_seg >= 0))
755 /* The PE will reserve all possible 32-bits space */
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",
762 /* Allocate TCE table */
763 tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
764 get_order(tce_table_size));
766 pe_err(pe, "Failed to allocate a 32-bit TCE memory\n");
769 addr = page_address(tce_mem);
770 memset(addr, 0, tce_table_size);
773 * Map TCE table through TVT. The TVE index is the PE number
774 * shifted by 1 bit for 32-bits DMA space.
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);
780 pe_err(pe, "Failed to configure 32-bit TCE table,"
785 /* Setup linux iommu table */
786 tbl = &pe->tce32_table;
787 pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, 0);
789 /* OPAL variant of PHB3 invalidated TCEs */
790 swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
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.
797 pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
798 tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
800 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
802 iommu_init_table(tbl, phb->hose->node);
803 iommu_register_group(tbl, phb->hose->global_number, pe->pe_number);
806 set_iommu_table_base_and_group(&pe->pdev->dev, tbl);
808 pnv_ioda_setup_bus_dma(pe, pe->pbus);
810 /* Also create a bypass window */
811 pnv_pci_ioda2_setup_bypass_pe(phb, pe);
814 if (pe->tce32_seg >= 0)
817 __free_pages(tce_mem, get_order(tce_table_size));
820 static void pnv_ioda_setup_dma(struct pnv_phb *phb)
822 struct pci_controller *hose = phb->hose;
823 unsigned int residual, remaining, segs, tw, base;
824 struct pnv_ioda_pe *pe;
826 /* If we have more PE# than segments available, hand out one
827 * per PE until we run out and let the rest fail. If not,
828 * then we assign at least one segment per PE, plus more based
829 * on the amount of devices under that PE
831 if (phb->ioda.dma_pe_count > phb->ioda.tce32_count)
834 residual = phb->ioda.tce32_count -
835 phb->ioda.dma_pe_count;
837 pr_info("PCI: Domain %04x has %ld available 32-bit DMA segments\n",
838 hose->global_number, phb->ioda.tce32_count);
839 pr_info("PCI: %d PE# for a total weight of %d\n",
840 phb->ioda.dma_pe_count, phb->ioda.dma_weight);
842 /* Walk our PE list and configure their DMA segments, hand them
843 * out one base segment plus any residual segments based on
846 remaining = phb->ioda.tce32_count;
847 tw = phb->ioda.dma_weight;
849 list_for_each_entry(pe, &phb->ioda.pe_dma_list, dma_link) {
853 pe_warn(pe, "No DMA32 resources available\n");
858 segs += ((pe->dma_weight * residual) + (tw / 2)) / tw;
859 if (segs > remaining)
864 * For IODA2 compliant PHB3, we needn't care about the weight.
865 * The all available 32-bits DMA space will be assigned to
868 if (phb->type == PNV_PHB_IODA1) {
869 pe_info(pe, "DMA weight %d, assigned %d DMA32 segments\n",
870 pe->dma_weight, segs);
871 pnv_pci_ioda_setup_dma_pe(phb, pe, base, segs);
873 pe_info(pe, "Assign DMA32 space\n");
875 pnv_pci_ioda2_setup_dma_pe(phb, pe);
883 #ifdef CONFIG_PCI_MSI
884 static void pnv_ioda2_msi_eoi(struct irq_data *d)
886 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
887 struct irq_chip *chip = irq_data_get_irq_chip(d);
888 struct pnv_phb *phb = container_of(chip, struct pnv_phb,
892 rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
898 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
899 unsigned int hwirq, unsigned int virq,
900 unsigned int is_64, struct msi_msg *msg)
902 struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
903 struct pci_dn *pdn = pci_get_pdn(dev);
904 struct irq_data *idata;
905 struct irq_chip *ichip;
906 unsigned int xive_num = hwirq - phb->msi_base;
910 /* No PE assigned ? bail out ... no MSI for you ! */
914 /* Check if we have an MVE */
915 if (pe->mve_number < 0)
918 /* Force 32-bit MSI on some broken devices */
919 if (pdn && pdn->force_32bit_msi)
922 /* Assign XIVE to PE */
923 rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
925 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
926 pci_name(dev), rc, xive_num);
933 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
936 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
940 msg->address_hi = be64_to_cpu(addr64) >> 32;
941 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
945 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
948 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
953 msg->address_lo = be32_to_cpu(addr32);
955 msg->data = be32_to_cpu(data);
958 * Change the IRQ chip for the MSI interrupts on PHB3.
959 * The corresponding IRQ chip should be populated for
962 if (phb->type == PNV_PHB_IODA2) {
963 if (!phb->ioda.irq_chip_init) {
964 idata = irq_get_irq_data(virq);
965 ichip = irq_data_get_irq_chip(idata);
966 phb->ioda.irq_chip_init = 1;
967 phb->ioda.irq_chip = *ichip;
968 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
971 irq_set_chip(virq, &phb->ioda.irq_chip);
974 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
975 " address=%x_%08x data=%x PE# %d\n",
976 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
977 msg->address_hi, msg->address_lo, data, pe->pe_number);
982 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
985 const __be32 *prop = of_get_property(phb->hose->dn,
986 "ibm,opal-msi-ranges", NULL);
989 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
994 phb->msi_base = be32_to_cpup(prop);
995 count = be32_to_cpup(prop + 1);
996 if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
997 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
998 phb->hose->global_number);
1002 phb->msi_setup = pnv_pci_ioda_msi_setup;
1003 phb->msi32_support = 1;
1004 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
1005 count, phb->msi_base);
1008 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
1009 #endif /* CONFIG_PCI_MSI */
1012 * This function is supposed to be called on basis of PE from top
1013 * to bottom style. So the the I/O or MMIO segment assigned to
1014 * parent PE could be overrided by its child PEs if necessary.
1016 static void pnv_ioda_setup_pe_seg(struct pci_controller *hose,
1017 struct pnv_ioda_pe *pe)
1019 struct pnv_phb *phb = hose->private_data;
1020 struct pci_bus_region region;
1021 struct resource *res;
1026 * NOTE: We only care PCI bus based PE for now. For PCI
1027 * device based PE, for example SRIOV sensitive VF should
1028 * be figured out later.
1030 BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
1032 pci_bus_for_each_resource(pe->pbus, res, i) {
1033 if (!res || !res->flags ||
1034 res->start > res->end)
1037 if (res->flags & IORESOURCE_IO) {
1038 region.start = res->start - phb->ioda.io_pci_base;
1039 region.end = res->end - phb->ioda.io_pci_base;
1040 index = region.start / phb->ioda.io_segsize;
1042 while (index < phb->ioda.total_pe &&
1043 region.start <= region.end) {
1044 phb->ioda.io_segmap[index] = pe->pe_number;
1045 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1046 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
1047 if (rc != OPAL_SUCCESS) {
1048 pr_err("%s: OPAL error %d when mapping IO "
1049 "segment #%d to PE#%d\n",
1050 __func__, rc, index, pe->pe_number);
1054 region.start += phb->ioda.io_segsize;
1057 } else if (res->flags & IORESOURCE_MEM) {
1058 /* WARNING: Assumes M32 is mem region 0 in PHB. We need to
1059 * harden that algorithm when we start supporting M64
1061 region.start = res->start -
1062 hose->mem_offset[0] -
1063 phb->ioda.m32_pci_base;
1064 region.end = res->end -
1065 hose->mem_offset[0] -
1066 phb->ioda.m32_pci_base;
1067 index = region.start / phb->ioda.m32_segsize;
1069 while (index < phb->ioda.total_pe &&
1070 region.start <= region.end) {
1071 phb->ioda.m32_segmap[index] = pe->pe_number;
1072 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1073 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
1074 if (rc != OPAL_SUCCESS) {
1075 pr_err("%s: OPAL error %d when mapping M32 "
1076 "segment#%d to PE#%d",
1077 __func__, rc, index, pe->pe_number);
1081 region.start += phb->ioda.m32_segsize;
1088 static void pnv_pci_ioda_setup_seg(void)
1090 struct pci_controller *tmp, *hose;
1091 struct pnv_phb *phb;
1092 struct pnv_ioda_pe *pe;
1094 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1095 phb = hose->private_data;
1096 list_for_each_entry(pe, &phb->ioda.pe_list, list) {
1097 pnv_ioda_setup_pe_seg(hose, pe);
1102 static void pnv_pci_ioda_setup_DMA(void)
1104 struct pci_controller *hose, *tmp;
1105 struct pnv_phb *phb;
1107 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1108 pnv_ioda_setup_dma(hose->private_data);
1110 /* Mark the PHB initialization done */
1111 phb = hose->private_data;
1112 phb->initialized = 1;
1116 static void pnv_pci_ioda_create_dbgfs(void)
1118 #ifdef CONFIG_DEBUG_FS
1119 struct pci_controller *hose, *tmp;
1120 struct pnv_phb *phb;
1123 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1124 phb = hose->private_data;
1126 sprintf(name, "PCI%04x", hose->global_number);
1127 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
1129 pr_warning("%s: Error on creating debugfs on PHB#%x\n",
1130 __func__, hose->global_number);
1132 #endif /* CONFIG_DEBUG_FS */
1135 static void pnv_pci_ioda_fixup(void)
1137 pnv_pci_ioda_setup_PEs();
1138 pnv_pci_ioda_setup_seg();
1139 pnv_pci_ioda_setup_DMA();
1141 pnv_pci_ioda_create_dbgfs();
1144 eeh_probe_mode_set(EEH_PROBE_MODE_DEV);
1145 eeh_addr_cache_build();
1151 * Returns the alignment for I/O or memory windows for P2P
1152 * bridges. That actually depends on how PEs are segmented.
1153 * For now, we return I/O or M32 segment size for PE sensitive
1154 * P2P bridges. Otherwise, the default values (4KiB for I/O,
1155 * 1MiB for memory) will be returned.
1157 * The current PCI bus might be put into one PE, which was
1158 * create against the parent PCI bridge. For that case, we
1159 * needn't enlarge the alignment so that we can save some
1162 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
1165 struct pci_dev *bridge;
1166 struct pci_controller *hose = pci_bus_to_host(bus);
1167 struct pnv_phb *phb = hose->private_data;
1168 int num_pci_bridges = 0;
1172 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
1174 if (num_pci_bridges >= 2)
1178 bridge = bridge->bus->self;
1181 /* We need support prefetchable memory window later */
1182 if (type & IORESOURCE_MEM)
1183 return phb->ioda.m32_segsize;
1185 return phb->ioda.io_segsize;
1188 /* Prevent enabling devices for which we couldn't properly
1191 static int pnv_pci_enable_device_hook(struct pci_dev *dev)
1193 struct pci_controller *hose = pci_bus_to_host(dev->bus);
1194 struct pnv_phb *phb = hose->private_data;
1197 /* The function is probably called while the PEs have
1198 * not be created yet. For example, resource reassignment
1199 * during PCI probe period. We just skip the check if
1202 if (!phb->initialized)
1205 pdn = pci_get_pdn(dev);
1206 if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1212 static u32 pnv_ioda_bdfn_to_pe(struct pnv_phb *phb, struct pci_bus *bus,
1215 return phb->ioda.pe_rmap[(bus->number << 8) | devfn];
1218 static void pnv_pci_ioda_shutdown(struct pnv_phb *phb)
1220 opal_pci_reset(phb->opal_id, OPAL_PCI_IODA_TABLE_RESET,
1224 void __init pnv_pci_init_ioda_phb(struct device_node *np,
1225 u64 hub_id, int ioda_type)
1227 struct pci_controller *hose;
1228 struct pnv_phb *phb;
1229 unsigned long size, m32map_off, pemap_off, iomap_off = 0;
1230 const __be64 *prop64;
1231 const __be32 *prop32;
1237 pr_info("Initializing IODA%d OPAL PHB %s\n", ioda_type, np->full_name);
1239 prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
1241 pr_err(" Missing \"ibm,opal-phbid\" property !\n");
1244 phb_id = be64_to_cpup(prop64);
1245 pr_debug(" PHB-ID : 0x%016llx\n", phb_id);
1247 phb = alloc_bootmem(sizeof(struct pnv_phb));
1249 pr_err(" Out of memory !\n");
1253 /* Allocate PCI controller */
1254 memset(phb, 0, sizeof(struct pnv_phb));
1255 phb->hose = hose = pcibios_alloc_controller(np);
1257 pr_err(" Can't allocate PCI controller for %s\n",
1259 free_bootmem((unsigned long)phb, sizeof(struct pnv_phb));
1263 spin_lock_init(&phb->lock);
1264 prop32 = of_get_property(np, "bus-range", &len);
1265 if (prop32 && len == 8) {
1266 hose->first_busno = be32_to_cpu(prop32[0]);
1267 hose->last_busno = be32_to_cpu(prop32[1]);
1269 pr_warn(" Broken <bus-range> on %s\n", np->full_name);
1270 hose->first_busno = 0;
1271 hose->last_busno = 0xff;
1273 hose->private_data = phb;
1274 phb->hub_id = hub_id;
1275 phb->opal_id = phb_id;
1276 phb->type = ioda_type;
1278 /* Detect specific models for error handling */
1279 if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
1280 phb->model = PNV_PHB_MODEL_P7IOC;
1281 else if (of_device_is_compatible(np, "ibm,power8-pciex"))
1282 phb->model = PNV_PHB_MODEL_PHB3;
1284 phb->model = PNV_PHB_MODEL_UNKNOWN;
1286 /* Parse 32-bit and IO ranges (if any) */
1287 pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
1290 phb->regs = of_iomap(np, 0);
1291 if (phb->regs == NULL)
1292 pr_err(" Failed to map registers !\n");
1294 /* Initialize more IODA stuff */
1295 phb->ioda.total_pe = 1;
1296 prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
1298 phb->ioda.total_pe = be32_to_cpup(prop32);
1299 prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
1301 phb->ioda.reserved_pe = be32_to_cpup(prop32);
1302 phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
1303 /* FW Has already off top 64k of M32 space (MSI space) */
1304 phb->ioda.m32_size += 0x10000;
1306 phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe;
1307 phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
1308 phb->ioda.io_size = hose->pci_io_size;
1309 phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe;
1310 phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
1312 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
1313 size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long));
1315 size += phb->ioda.total_pe * sizeof(phb->ioda.m32_segmap[0]);
1316 if (phb->type == PNV_PHB_IODA1) {
1318 size += phb->ioda.total_pe * sizeof(phb->ioda.io_segmap[0]);
1321 size += phb->ioda.total_pe * sizeof(struct pnv_ioda_pe);
1322 aux = alloc_bootmem(size);
1323 memset(aux, 0, size);
1324 phb->ioda.pe_alloc = aux;
1325 phb->ioda.m32_segmap = aux + m32map_off;
1326 if (phb->type == PNV_PHB_IODA1)
1327 phb->ioda.io_segmap = aux + iomap_off;
1328 phb->ioda.pe_array = aux + pemap_off;
1329 set_bit(phb->ioda.reserved_pe, phb->ioda.pe_alloc);
1331 INIT_LIST_HEAD(&phb->ioda.pe_dma_list);
1332 INIT_LIST_HEAD(&phb->ioda.pe_list);
1334 /* Calculate how many 32-bit TCE segments we have */
1335 phb->ioda.tce32_count = phb->ioda.m32_pci_base >> 28;
1337 /* Clear unusable m64 */
1338 hose->mem_resources[1].flags = 0;
1339 hose->mem_resources[1].start = 0;
1340 hose->mem_resources[1].end = 0;
1341 hose->mem_resources[2].flags = 0;
1342 hose->mem_resources[2].start = 0;
1343 hose->mem_resources[2].end = 0;
1345 #if 0 /* We should really do that ... */
1346 rc = opal_pci_set_phb_mem_window(opal->phb_id,
1349 starting_real_address,
1350 starting_pci_address,
1354 pr_info(" %d (%d) PE's M32: 0x%x [segment=0x%x]"
1355 " IO: 0x%x [segment=0x%x]\n",
1357 phb->ioda.reserved_pe,
1358 phb->ioda.m32_size, phb->ioda.m32_segsize,
1359 phb->ioda.io_size, phb->ioda.io_segsize);
1361 phb->hose->ops = &pnv_pci_ops;
1363 phb->eeh_ops = &ioda_eeh_ops;
1366 /* Setup RID -> PE mapping function */
1367 phb->bdfn_to_pe = pnv_ioda_bdfn_to_pe;
1370 phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
1371 phb->dma_set_mask = pnv_pci_ioda_dma_set_mask;
1373 /* Setup shutdown function for kexec */
1374 phb->shutdown = pnv_pci_ioda_shutdown;
1376 /* Setup MSI support */
1377 pnv_pci_init_ioda_msis(phb);
1380 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
1381 * to let the PCI core do resource assignment. It's supposed
1382 * that the PCI core will do correct I/O and MMIO alignment
1383 * for the P2P bridge bars so that each PCI bus (excluding
1384 * the child P2P bridges) can form individual PE.
1386 ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
1387 ppc_md.pcibios_enable_device_hook = pnv_pci_enable_device_hook;
1388 ppc_md.pcibios_window_alignment = pnv_pci_window_alignment;
1389 ppc_md.pcibios_reset_secondary_bus = pnv_pci_reset_secondary_bus;
1390 pci_add_flags(PCI_REASSIGN_ALL_RSRC);
1392 /* Reset IODA tables to a clean state */
1393 rc = opal_pci_reset(phb_id, OPAL_PCI_IODA_TABLE_RESET, OPAL_ASSERT_RESET);
1395 pr_warning(" OPAL Error %ld performing IODA table reset !\n", rc);
1397 /* If we're running in kdump kerenl, the previous kerenl never
1398 * shutdown PCI devices correctly. We already got IODA table
1399 * cleaned out. So we have to issue PHB reset to stop all PCI
1400 * transactions from previous kerenl.
1402 if (is_kdump_kernel()) {
1403 pr_info(" Issue PHB reset ...\n");
1404 ioda_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
1405 ioda_eeh_phb_reset(hose, OPAL_DEASSERT_RESET);
1409 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
1411 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
1414 void __init pnv_pci_init_ioda_hub(struct device_node *np)
1416 struct device_node *phbn;
1417 const __be64 *prop64;
1420 pr_info("Probing IODA IO-Hub %s\n", np->full_name);
1422 prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
1424 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
1427 hub_id = be64_to_cpup(prop64);
1428 pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
1430 /* Count child PHBs */
1431 for_each_child_of_node(np, phbn) {
1432 /* Look for IODA1 PHBs */
1433 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
1434 pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);