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>
26 #include <linux/iommu.h>
27 #include <linux/rculist.h>
28 #include <linux/sizes.h>
30 #include <asm/sections.h>
33 #include <asm/pci-bridge.h>
34 #include <asm/machdep.h>
35 #include <asm/msi_bitmap.h>
36 #include <asm/ppc-pci.h>
38 #include <asm/iommu.h>
41 #include <asm/debug.h>
42 #include <asm/firmware.h>
43 #include <asm/pnv-pci.h>
44 #include <asm/mmzone.h>
46 #include <misc/cxl-base.h>
51 #define PNV_IODA1_M64_NUM 16 /* Number of M64 BARs */
52 #define PNV_IODA1_M64_SEGS 8 /* Segments per M64 BAR */
53 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
55 #define POWERNV_IOMMU_DEFAULT_LEVELS 1
56 #define POWERNV_IOMMU_MAX_LEVELS 5
58 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl);
60 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
72 if (pe->flags & PNV_IODA_PE_DEV)
73 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
74 else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
75 sprintf(pfix, "%04x:%02x ",
76 pci_domain_nr(pe->pbus), pe->pbus->number);
78 else if (pe->flags & PNV_IODA_PE_VF)
79 sprintf(pfix, "%04x:%02x:%2x.%d",
80 pci_domain_nr(pe->parent_dev->bus),
81 (pe->rid & 0xff00) >> 8,
82 PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
83 #endif /* CONFIG_PCI_IOV*/
85 printk("%spci %s: [PE# %.3d] %pV",
86 level, pfix, pe->pe_number, &vaf);
91 static bool pnv_iommu_bypass_disabled __read_mostly;
93 static int __init iommu_setup(char *str)
99 if (!strncmp(str, "nobypass", 8)) {
100 pnv_iommu_bypass_disabled = true;
101 pr_info("PowerNV: IOMMU bypass window disabled.\n");
104 str += strcspn(str, ",");
111 early_param("iommu", iommu_setup);
113 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags)
115 return ((flags & (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)) ==
116 (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH));
119 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
121 phb->ioda.pe_array[pe_no].phb = phb;
122 phb->ioda.pe_array[pe_no].pe_number = pe_no;
124 return &phb->ioda.pe_array[pe_no];
127 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
129 if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
130 pr_warn("%s: Invalid PE %d on PHB#%x\n",
131 __func__, pe_no, phb->hose->global_number);
135 if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
136 pr_debug("%s: PE %d was reserved on PHB#%x\n",
137 __func__, pe_no, phb->hose->global_number);
139 pnv_ioda_init_pe(phb, pe_no);
142 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
147 pe = find_next_zero_bit(phb->ioda.pe_alloc,
148 phb->ioda.total_pe_num, 0);
149 if (pe >= phb->ioda.total_pe_num)
151 } while(test_and_set_bit(pe, phb->ioda.pe_alloc));
153 return pnv_ioda_init_pe(phb, pe);
156 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
158 struct pnv_phb *phb = pe->phb;
162 memset(pe, 0, sizeof(struct pnv_ioda_pe));
163 clear_bit(pe->pe_number, phb->ioda.pe_alloc);
166 /* The default M64 BAR is shared by all PEs */
167 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
173 /* Configure the default M64 BAR */
174 rc = opal_pci_set_phb_mem_window(phb->opal_id,
175 OPAL_M64_WINDOW_TYPE,
176 phb->ioda.m64_bar_idx,
180 if (rc != OPAL_SUCCESS) {
181 desc = "configuring";
185 /* Enable the default M64 BAR */
186 rc = opal_pci_phb_mmio_enable(phb->opal_id,
187 OPAL_M64_WINDOW_TYPE,
188 phb->ioda.m64_bar_idx,
189 OPAL_ENABLE_M64_SPLIT);
190 if (rc != OPAL_SUCCESS) {
195 /* Mark the M64 BAR assigned */
196 set_bit(phb->ioda.m64_bar_idx, &phb->ioda.m64_bar_alloc);
199 * Strip off the segment used by the reserved PE, which is
200 * expected to be 0 or last one of PE capabicity.
202 r = &phb->hose->mem_resources[1];
203 if (phb->ioda.reserved_pe_idx == 0)
204 r->start += phb->ioda.m64_segsize;
205 else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
206 r->end -= phb->ioda.m64_segsize;
208 pr_warn(" Cannot strip M64 segment for reserved PE#%d\n",
209 phb->ioda.reserved_pe_idx);
214 pr_warn(" Failure %lld %s M64 BAR#%d\n",
215 rc, desc, phb->ioda.m64_bar_idx);
216 opal_pci_phb_mmio_enable(phb->opal_id,
217 OPAL_M64_WINDOW_TYPE,
218 phb->ioda.m64_bar_idx,
223 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
224 unsigned long *pe_bitmap)
226 struct pci_controller *hose = pci_bus_to_host(pdev->bus);
227 struct pnv_phb *phb = hose->private_data;
229 resource_size_t base, sgsz, start, end;
232 base = phb->ioda.m64_base;
233 sgsz = phb->ioda.m64_segsize;
234 for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
235 r = &pdev->resource[i];
236 if (!r->parent || !pnv_pci_is_mem_pref_64(r->flags))
239 start = _ALIGN_DOWN(r->start - base, sgsz);
240 end = _ALIGN_UP(r->end - base, sgsz);
241 for (segno = start / sgsz; segno < end / sgsz; segno++) {
243 set_bit(segno, pe_bitmap);
245 pnv_ioda_reserve_pe(phb, segno);
250 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
256 * There are 16 M64 BARs, each of which has 8 segments. So
257 * there are as many M64 segments as the maximum number of
260 for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
261 unsigned long base, segsz = phb->ioda.m64_segsize;
264 base = phb->ioda.m64_base +
265 index * PNV_IODA1_M64_SEGS * segsz;
266 rc = opal_pci_set_phb_mem_window(phb->opal_id,
267 OPAL_M64_WINDOW_TYPE, index, base, 0,
268 PNV_IODA1_M64_SEGS * segsz);
269 if (rc != OPAL_SUCCESS) {
270 pr_warn(" Error %lld setting M64 PHB#%d-BAR#%d\n",
271 rc, phb->hose->global_number, index);
275 rc = opal_pci_phb_mmio_enable(phb->opal_id,
276 OPAL_M64_WINDOW_TYPE, index,
277 OPAL_ENABLE_M64_SPLIT);
278 if (rc != OPAL_SUCCESS) {
279 pr_warn(" Error %lld enabling M64 PHB#%d-BAR#%d\n",
280 rc, phb->hose->global_number, index);
286 * Exclude the segment used by the reserved PE, which
287 * is expected to be 0 or last supported PE#.
289 r = &phb->hose->mem_resources[1];
290 if (phb->ioda.reserved_pe_idx == 0)
291 r->start += phb->ioda.m64_segsize;
292 else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
293 r->end -= phb->ioda.m64_segsize;
295 WARN(1, "Wrong reserved PE#%d on PHB#%d\n",
296 phb->ioda.reserved_pe_idx, phb->hose->global_number);
301 for ( ; index >= 0; index--)
302 opal_pci_phb_mmio_enable(phb->opal_id,
303 OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
308 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
309 unsigned long *pe_bitmap,
312 struct pci_dev *pdev;
314 list_for_each_entry(pdev, &bus->devices, bus_list) {
315 pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
317 if (all && pdev->subordinate)
318 pnv_ioda_reserve_m64_pe(pdev->subordinate,
323 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
325 struct pci_controller *hose = pci_bus_to_host(bus);
326 struct pnv_phb *phb = hose->private_data;
327 struct pnv_ioda_pe *master_pe, *pe;
328 unsigned long size, *pe_alloc;
331 /* Root bus shouldn't use M64 */
332 if (pci_is_root_bus(bus))
335 /* Allocate bitmap */
336 size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
337 pe_alloc = kzalloc(size, GFP_KERNEL);
339 pr_warn("%s: Out of memory !\n",
344 /* Figure out reserved PE numbers by the PE */
345 pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
348 * the current bus might not own M64 window and that's all
349 * contributed by its child buses. For the case, we needn't
350 * pick M64 dependent PE#.
352 if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
358 * Figure out the master PE and put all slave PEs to master
359 * PE's list to form compound PE.
363 while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
364 phb->ioda.total_pe_num) {
365 pe = &phb->ioda.pe_array[i];
367 phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
369 pe->flags |= PNV_IODA_PE_MASTER;
370 INIT_LIST_HEAD(&pe->slaves);
373 pe->flags |= PNV_IODA_PE_SLAVE;
374 pe->master = master_pe;
375 list_add_tail(&pe->list, &master_pe->slaves);
379 * P7IOC supports M64DT, which helps mapping M64 segment
380 * to one particular PE#. However, PHB3 has fixed mapping
381 * between M64 segment and PE#. In order to have same logic
382 * for P7IOC and PHB3, we enforce fixed mapping between M64
383 * segment and PE# on P7IOC.
385 if (phb->type == PNV_PHB_IODA1) {
388 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
389 pe->pe_number, OPAL_M64_WINDOW_TYPE,
390 pe->pe_number / PNV_IODA1_M64_SEGS,
391 pe->pe_number % PNV_IODA1_M64_SEGS);
392 if (rc != OPAL_SUCCESS)
393 pr_warn("%s: Error %lld mapping M64 for PHB#%d-PE#%d\n",
394 __func__, rc, phb->hose->global_number,
403 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
405 struct pci_controller *hose = phb->hose;
406 struct device_node *dn = hose->dn;
407 struct resource *res;
411 if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
412 pr_info(" Not support M64 window\n");
416 if (!firmware_has_feature(FW_FEATURE_OPAL)) {
417 pr_info(" Firmware too old to support M64 window\n");
421 r = of_get_property(dn, "ibm,opal-m64-window", NULL);
423 pr_info(" No <ibm,opal-m64-window> on %s\n",
428 res = &hose->mem_resources[1];
429 res->name = dn->full_name;
430 res->start = of_translate_address(dn, r + 2);
431 res->end = res->start + of_read_number(r + 4, 2) - 1;
432 res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
433 pci_addr = of_read_number(r, 2);
434 hose->mem_offset[1] = res->start - pci_addr;
436 phb->ioda.m64_size = resource_size(res);
437 phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
438 phb->ioda.m64_base = pci_addr;
440 pr_info(" MEM64 0x%016llx..0x%016llx -> 0x%016llx\n",
441 res->start, res->end, pci_addr);
443 /* Use last M64 BAR to cover M64 window */
444 phb->ioda.m64_bar_idx = 15;
445 if (phb->type == PNV_PHB_IODA1)
446 phb->init_m64 = pnv_ioda1_init_m64;
448 phb->init_m64 = pnv_ioda2_init_m64;
449 phb->reserve_m64_pe = pnv_ioda_reserve_m64_pe;
450 phb->pick_m64_pe = pnv_ioda_pick_m64_pe;
453 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
455 struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
456 struct pnv_ioda_pe *slave;
459 /* Fetch master PE */
460 if (pe->flags & PNV_IODA_PE_SLAVE) {
462 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
465 pe_no = pe->pe_number;
468 /* Freeze master PE */
469 rc = opal_pci_eeh_freeze_set(phb->opal_id,
471 OPAL_EEH_ACTION_SET_FREEZE_ALL);
472 if (rc != OPAL_SUCCESS) {
473 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
474 __func__, rc, phb->hose->global_number, pe_no);
478 /* Freeze slave PEs */
479 if (!(pe->flags & PNV_IODA_PE_MASTER))
482 list_for_each_entry(slave, &pe->slaves, list) {
483 rc = opal_pci_eeh_freeze_set(phb->opal_id,
485 OPAL_EEH_ACTION_SET_FREEZE_ALL);
486 if (rc != OPAL_SUCCESS)
487 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
488 __func__, rc, phb->hose->global_number,
493 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
495 struct pnv_ioda_pe *pe, *slave;
499 pe = &phb->ioda.pe_array[pe_no];
500 if (pe->flags & PNV_IODA_PE_SLAVE) {
502 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
503 pe_no = pe->pe_number;
506 /* Clear frozen state for master PE */
507 rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
508 if (rc != OPAL_SUCCESS) {
509 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
510 __func__, rc, opt, phb->hose->global_number, pe_no);
514 if (!(pe->flags & PNV_IODA_PE_MASTER))
517 /* Clear frozen state for slave PEs */
518 list_for_each_entry(slave, &pe->slaves, list) {
519 rc = opal_pci_eeh_freeze_clear(phb->opal_id,
522 if (rc != OPAL_SUCCESS) {
523 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
524 __func__, rc, opt, phb->hose->global_number,
533 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
535 struct pnv_ioda_pe *slave, *pe;
540 /* Sanity check on PE number */
541 if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
542 return OPAL_EEH_STOPPED_PERM_UNAVAIL;
545 * Fetch the master PE and the PE instance might be
546 * not initialized yet.
548 pe = &phb->ioda.pe_array[pe_no];
549 if (pe->flags & PNV_IODA_PE_SLAVE) {
551 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
552 pe_no = pe->pe_number;
555 /* Check the master PE */
556 rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
557 &state, &pcierr, NULL);
558 if (rc != OPAL_SUCCESS) {
559 pr_warn("%s: Failure %lld getting "
560 "PHB#%x-PE#%x state\n",
562 phb->hose->global_number, pe_no);
563 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
566 /* Check the slave PE */
567 if (!(pe->flags & PNV_IODA_PE_MASTER))
570 list_for_each_entry(slave, &pe->slaves, list) {
571 rc = opal_pci_eeh_freeze_status(phb->opal_id,
576 if (rc != OPAL_SUCCESS) {
577 pr_warn("%s: Failure %lld getting "
578 "PHB#%x-PE#%x state\n",
580 phb->hose->global_number, slave->pe_number);
581 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
585 * Override the result based on the ascending
595 /* Currently those 2 are only used when MSIs are enabled, this will change
596 * but in the meantime, we need to protect them to avoid warnings
598 #ifdef CONFIG_PCI_MSI
599 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
601 struct pci_controller *hose = pci_bus_to_host(dev->bus);
602 struct pnv_phb *phb = hose->private_data;
603 struct pci_dn *pdn = pci_get_pdn(dev);
607 if (pdn->pe_number == IODA_INVALID_PE)
609 return &phb->ioda.pe_array[pdn->pe_number];
611 #endif /* CONFIG_PCI_MSI */
613 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
614 struct pnv_ioda_pe *parent,
615 struct pnv_ioda_pe *child,
618 const char *desc = is_add ? "adding" : "removing";
619 uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
620 OPAL_REMOVE_PE_FROM_DOMAIN;
621 struct pnv_ioda_pe *slave;
624 /* Parent PE affects child PE */
625 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
626 child->pe_number, op);
627 if (rc != OPAL_SUCCESS) {
628 pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
633 if (!(child->flags & PNV_IODA_PE_MASTER))
636 /* Compound case: parent PE affects slave PEs */
637 list_for_each_entry(slave, &child->slaves, list) {
638 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
639 slave->pe_number, op);
640 if (rc != OPAL_SUCCESS) {
641 pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
650 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
651 struct pnv_ioda_pe *pe,
654 struct pnv_ioda_pe *slave;
655 struct pci_dev *pdev = NULL;
659 * Clear PE frozen state. If it's master PE, we need
660 * clear slave PE frozen state as well.
663 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
664 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
665 if (pe->flags & PNV_IODA_PE_MASTER) {
666 list_for_each_entry(slave, &pe->slaves, list)
667 opal_pci_eeh_freeze_clear(phb->opal_id,
669 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
674 * Associate PE in PELT. We need add the PE into the
675 * corresponding PELT-V as well. Otherwise, the error
676 * originated from the PE might contribute to other
679 ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
683 /* For compound PEs, any one affects all of them */
684 if (pe->flags & PNV_IODA_PE_MASTER) {
685 list_for_each_entry(slave, &pe->slaves, list) {
686 ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
692 if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
693 pdev = pe->pbus->self;
694 else if (pe->flags & PNV_IODA_PE_DEV)
695 pdev = pe->pdev->bus->self;
696 #ifdef CONFIG_PCI_IOV
697 else if (pe->flags & PNV_IODA_PE_VF)
698 pdev = pe->parent_dev;
699 #endif /* CONFIG_PCI_IOV */
701 struct pci_dn *pdn = pci_get_pdn(pdev);
702 struct pnv_ioda_pe *parent;
704 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
705 parent = &phb->ioda.pe_array[pdn->pe_number];
706 ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
711 pdev = pdev->bus->self;
717 #ifdef CONFIG_PCI_IOV
718 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
720 struct pci_dev *parent;
721 uint8_t bcomp, dcomp, fcomp;
725 /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
729 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
730 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
731 parent = pe->pbus->self;
732 if (pe->flags & PNV_IODA_PE_BUS_ALL)
733 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
738 case 1: bcomp = OpalPciBusAll; break;
739 case 2: bcomp = OpalPciBus7Bits; break;
740 case 4: bcomp = OpalPciBus6Bits; break;
741 case 8: bcomp = OpalPciBus5Bits; break;
742 case 16: bcomp = OpalPciBus4Bits; break;
743 case 32: bcomp = OpalPciBus3Bits; break;
745 dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
747 /* Do an exact match only */
748 bcomp = OpalPciBusAll;
750 rid_end = pe->rid + (count << 8);
752 if (pe->flags & PNV_IODA_PE_VF)
753 parent = pe->parent_dev;
755 parent = pe->pdev->bus->self;
756 bcomp = OpalPciBusAll;
757 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
758 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
759 rid_end = pe->rid + 1;
762 /* Clear the reverse map */
763 for (rid = pe->rid; rid < rid_end; rid++)
764 phb->ioda.pe_rmap[rid] = 0;
766 /* Release from all parents PELT-V */
768 struct pci_dn *pdn = pci_get_pdn(parent);
769 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
770 rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
771 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
772 /* XXX What to do in case of error ? */
774 parent = parent->bus->self;
777 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
778 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
780 /* Disassociate PE in PELT */
781 rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
782 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
784 pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
785 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
786 bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
788 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
792 pe->parent_dev = NULL;
796 #endif /* CONFIG_PCI_IOV */
798 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
800 struct pci_dev *parent;
801 uint8_t bcomp, dcomp, fcomp;
802 long rc, rid_end, rid;
804 /* Bus validation ? */
808 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
809 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
810 parent = pe->pbus->self;
811 if (pe->flags & PNV_IODA_PE_BUS_ALL)
812 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
817 case 1: bcomp = OpalPciBusAll; break;
818 case 2: bcomp = OpalPciBus7Bits; break;
819 case 4: bcomp = OpalPciBus6Bits; break;
820 case 8: bcomp = OpalPciBus5Bits; break;
821 case 16: bcomp = OpalPciBus4Bits; break;
822 case 32: bcomp = OpalPciBus3Bits; break;
824 dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
826 /* Do an exact match only */
827 bcomp = OpalPciBusAll;
829 rid_end = pe->rid + (count << 8);
831 #ifdef CONFIG_PCI_IOV
832 if (pe->flags & PNV_IODA_PE_VF)
833 parent = pe->parent_dev;
835 #endif /* CONFIG_PCI_IOV */
836 parent = pe->pdev->bus->self;
837 bcomp = OpalPciBusAll;
838 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
839 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
840 rid_end = pe->rid + 1;
844 * Associate PE in PELT. We need add the PE into the
845 * corresponding PELT-V as well. Otherwise, the error
846 * originated from the PE might contribute to other
849 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
850 bcomp, dcomp, fcomp, OPAL_MAP_PE);
852 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
857 * Configure PELTV. NPUs don't have a PELTV table so skip
858 * configuration on them.
860 if (phb->type != PNV_PHB_NPU)
861 pnv_ioda_set_peltv(phb, pe, true);
863 /* Setup reverse map */
864 for (rid = pe->rid; rid < rid_end; rid++)
865 phb->ioda.pe_rmap[rid] = pe->pe_number;
867 /* Setup one MVTs on IODA1 */
868 if (phb->type != PNV_PHB_IODA1) {
873 pe->mve_number = pe->pe_number;
874 rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
875 if (rc != OPAL_SUCCESS) {
876 pe_err(pe, "OPAL error %ld setting up MVE %d\n",
880 rc = opal_pci_set_mve_enable(phb->opal_id,
881 pe->mve_number, OPAL_ENABLE_MVE);
883 pe_err(pe, "OPAL error %ld enabling MVE %d\n",
893 #ifdef CONFIG_PCI_IOV
894 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
896 struct pci_dn *pdn = pci_get_pdn(dev);
898 struct resource *res, res2;
899 resource_size_t size;
906 * "offset" is in VFs. The M64 windows are sized so that when they
907 * are segmented, each segment is the same size as the IOV BAR.
908 * Each segment is in a separate PE, and the high order bits of the
909 * address are the PE number. Therefore, each VF's BAR is in a
910 * separate PE, and changing the IOV BAR start address changes the
911 * range of PEs the VFs are in.
913 num_vfs = pdn->num_vfs;
914 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
915 res = &dev->resource[i + PCI_IOV_RESOURCES];
916 if (!res->flags || !res->parent)
920 * The actual IOV BAR range is determined by the start address
921 * and the actual size for num_vfs VFs BAR. This check is to
922 * make sure that after shifting, the range will not overlap
923 * with another device.
925 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
926 res2.flags = res->flags;
927 res2.start = res->start + (size * offset);
928 res2.end = res2.start + (size * num_vfs) - 1;
930 if (res2.end > res->end) {
931 dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
932 i, &res2, res, num_vfs, offset);
938 * After doing so, there would be a "hole" in the /proc/iomem when
939 * offset is a positive value. It looks like the device return some
940 * mmio back to the system, which actually no one could use it.
942 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
943 res = &dev->resource[i + PCI_IOV_RESOURCES];
944 if (!res->flags || !res->parent)
947 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
949 res->start += size * offset;
951 dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
952 i, &res2, res, (offset > 0) ? "En" : "Dis",
954 pci_update_resource(dev, i + PCI_IOV_RESOURCES);
958 #endif /* CONFIG_PCI_IOV */
960 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
962 struct pci_controller *hose = pci_bus_to_host(dev->bus);
963 struct pnv_phb *phb = hose->private_data;
964 struct pci_dn *pdn = pci_get_pdn(dev);
965 struct pnv_ioda_pe *pe;
968 pr_err("%s: Device tree node not associated properly\n",
972 if (pdn->pe_number != IODA_INVALID_PE)
975 pe = pnv_ioda_alloc_pe(phb);
977 pr_warning("%s: Not enough PE# available, disabling device\n",
982 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
983 * pointer in the PE data structure, both should be destroyed at the
984 * same time. However, this needs to be looked at more closely again
985 * once we actually start removing things (Hotplug, SR-IOV, ...)
987 * At some point we want to remove the PDN completely anyways
991 pdn->pe_number = pe->pe_number;
992 pe->flags = PNV_IODA_PE_DEV;
996 pe->rid = dev->bus->number << 8 | pdn->devfn;
998 pe_info(pe, "Associated device to PE\n");
1000 if (pnv_ioda_configure_pe(phb, pe)) {
1001 /* XXX What do we do here ? */
1002 pnv_ioda_free_pe(pe);
1003 pdn->pe_number = IODA_INVALID_PE;
1012 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1014 struct pci_dev *dev;
1016 list_for_each_entry(dev, &bus->devices, bus_list) {
1017 struct pci_dn *pdn = pci_get_pdn(dev);
1020 pr_warn("%s: No device node associated with device !\n",
1025 pdn->pe_number = pe->pe_number;
1026 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1027 pnv_ioda_setup_same_PE(dev->subordinate, pe);
1032 * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1033 * single PCI bus. Another one that contains the primary PCI bus and its
1034 * subordinate PCI devices and buses. The second type of PE is normally
1035 * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1037 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1039 struct pci_controller *hose = pci_bus_to_host(bus);
1040 struct pnv_phb *phb = hose->private_data;
1041 struct pnv_ioda_pe *pe = NULL;
1043 /* Check if PE is determined by M64 */
1044 if (phb->pick_m64_pe)
1045 pe = phb->pick_m64_pe(bus, all);
1047 /* The PE number isn't pinned by M64 */
1049 pe = pnv_ioda_alloc_pe(phb);
1052 pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1053 __func__, pci_domain_nr(bus), bus->number);
1057 pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1060 pe->mve_number = -1;
1061 pe->rid = bus->busn_res.start << 8;
1064 pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
1065 bus->busn_res.start, bus->busn_res.end, pe->pe_number);
1067 pe_info(pe, "Secondary bus %d associated with PE#%d\n",
1068 bus->busn_res.start, pe->pe_number);
1070 if (pnv_ioda_configure_pe(phb, pe)) {
1071 /* XXX What do we do here ? */
1072 pnv_ioda_free_pe(pe);
1077 /* Associate it with all child devices */
1078 pnv_ioda_setup_same_PE(bus, pe);
1080 /* Put PE to the list */
1081 list_add_tail(&pe->list, &phb->ioda.pe_list);
1086 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1088 int pe_num, found_pe = false, rc;
1090 struct pnv_ioda_pe *pe;
1091 struct pci_dev *gpu_pdev;
1092 struct pci_dn *npu_pdn;
1093 struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1094 struct pnv_phb *phb = hose->private_data;
1097 * Due to a hardware errata PE#0 on the NPU is reserved for
1098 * error handling. This means we only have three PEs remaining
1099 * which need to be assigned to four links, implying some
1100 * links must share PEs.
1102 * To achieve this we assign PEs such that NPUs linking the
1103 * same GPU get assigned the same PE.
1105 gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1106 for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1107 pe = &phb->ioda.pe_array[pe_num];
1111 if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1113 * This device has the same peer GPU so should
1114 * be assigned the same PE as the existing
1117 dev_info(&npu_pdev->dev,
1118 "Associating to existing PE %d\n", pe_num);
1119 pci_dev_get(npu_pdev);
1120 npu_pdn = pci_get_pdn(npu_pdev);
1121 rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1122 npu_pdn->pcidev = npu_pdev;
1123 npu_pdn->pe_number = pe_num;
1124 phb->ioda.pe_rmap[rid] = pe->pe_number;
1126 /* Map the PE to this link */
1127 rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1129 OPAL_COMPARE_RID_DEVICE_NUMBER,
1130 OPAL_COMPARE_RID_FUNCTION_NUMBER,
1132 WARN_ON(rc != OPAL_SUCCESS);
1140 * Could not find an existing PE so allocate a new
1143 return pnv_ioda_setup_dev_PE(npu_pdev);
1148 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1150 struct pci_dev *pdev;
1152 list_for_each_entry(pdev, &bus->devices, bus_list)
1153 pnv_ioda_setup_npu_PE(pdev);
1156 static void pnv_ioda_setup_PEs(struct pci_bus *bus)
1158 struct pci_dev *dev;
1160 pnv_ioda_setup_bus_PE(bus, false);
1162 list_for_each_entry(dev, &bus->devices, bus_list) {
1163 if (dev->subordinate) {
1164 if (pci_pcie_type(dev) == PCI_EXP_TYPE_PCI_BRIDGE)
1165 pnv_ioda_setup_bus_PE(dev->subordinate, true);
1167 pnv_ioda_setup_PEs(dev->subordinate);
1173 * Configure PEs so that the downstream PCI buses and devices
1174 * could have their associated PE#. Unfortunately, we didn't
1175 * figure out the way to identify the PLX bridge yet. So we
1176 * simply put the PCI bus and the subordinate behind the root
1177 * port to PE# here. The game rule here is expected to be changed
1178 * as soon as we can detected PLX bridge correctly.
1180 static void pnv_pci_ioda_setup_PEs(void)
1182 struct pci_controller *hose, *tmp;
1183 struct pnv_phb *phb;
1185 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1186 phb = hose->private_data;
1188 /* M64 layout might affect PE allocation */
1189 if (phb->reserve_m64_pe)
1190 phb->reserve_m64_pe(hose->bus, NULL, true);
1193 * On NPU PHB, we expect separate PEs for individual PCI
1194 * functions. PCI bus dependent PEs are required for the
1195 * remaining types of PHBs.
1197 if (phb->type == PNV_PHB_NPU) {
1198 /* PE#0 is needed for error reporting */
1199 pnv_ioda_reserve_pe(phb, 0);
1200 pnv_ioda_setup_npu_PEs(hose->bus);
1202 pnv_ioda_setup_PEs(hose->bus);
1206 #ifdef CONFIG_PCI_IOV
1207 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1209 struct pci_bus *bus;
1210 struct pci_controller *hose;
1211 struct pnv_phb *phb;
1217 hose = pci_bus_to_host(bus);
1218 phb = hose->private_data;
1219 pdn = pci_get_pdn(pdev);
1221 if (pdn->m64_single_mode)
1226 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1227 for (j = 0; j < m64_bars; j++) {
1228 if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1230 opal_pci_phb_mmio_enable(phb->opal_id,
1231 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1232 clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1233 pdn->m64_map[j][i] = IODA_INVALID_M64;
1236 kfree(pdn->m64_map);
1240 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1242 struct pci_bus *bus;
1243 struct pci_controller *hose;
1244 struct pnv_phb *phb;
1247 struct resource *res;
1251 resource_size_t size, start;
1256 hose = pci_bus_to_host(bus);
1257 phb = hose->private_data;
1258 pdn = pci_get_pdn(pdev);
1259 total_vfs = pci_sriov_get_totalvfs(pdev);
1261 if (pdn->m64_single_mode)
1266 pdn->m64_map = kmalloc(sizeof(*pdn->m64_map) * m64_bars, GFP_KERNEL);
1269 /* Initialize the m64_map to IODA_INVALID_M64 */
1270 for (i = 0; i < m64_bars ; i++)
1271 for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1272 pdn->m64_map[i][j] = IODA_INVALID_M64;
1275 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1276 res = &pdev->resource[i + PCI_IOV_RESOURCES];
1277 if (!res->flags || !res->parent)
1280 for (j = 0; j < m64_bars; j++) {
1282 win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1283 phb->ioda.m64_bar_idx + 1, 0);
1285 if (win >= phb->ioda.m64_bar_idx + 1)
1287 } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1289 pdn->m64_map[j][i] = win;
1291 if (pdn->m64_single_mode) {
1292 size = pci_iov_resource_size(pdev,
1293 PCI_IOV_RESOURCES + i);
1294 start = res->start + size * j;
1296 size = resource_size(res);
1300 /* Map the M64 here */
1301 if (pdn->m64_single_mode) {
1302 pe_num = pdn->pe_num_map[j];
1303 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1304 pe_num, OPAL_M64_WINDOW_TYPE,
1305 pdn->m64_map[j][i], 0);
1308 rc = opal_pci_set_phb_mem_window(phb->opal_id,
1309 OPAL_M64_WINDOW_TYPE,
1316 if (rc != OPAL_SUCCESS) {
1317 dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1322 if (pdn->m64_single_mode)
1323 rc = opal_pci_phb_mmio_enable(phb->opal_id,
1324 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1326 rc = opal_pci_phb_mmio_enable(phb->opal_id,
1327 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1329 if (rc != OPAL_SUCCESS) {
1330 dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1339 pnv_pci_vf_release_m64(pdev, num_vfs);
1343 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1345 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
1347 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1349 struct iommu_table *tbl;
1352 tbl = pe->table_group.tables[0];
1353 rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1355 pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1357 pnv_pci_ioda2_set_bypass(pe, false);
1358 if (pe->table_group.group) {
1359 iommu_group_put(pe->table_group.group);
1360 BUG_ON(pe->table_group.group);
1362 pnv_pci_ioda2_table_free_pages(tbl);
1363 iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
1366 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1368 struct pci_bus *bus;
1369 struct pci_controller *hose;
1370 struct pnv_phb *phb;
1371 struct pnv_ioda_pe *pe, *pe_n;
1375 hose = pci_bus_to_host(bus);
1376 phb = hose->private_data;
1377 pdn = pci_get_pdn(pdev);
1379 if (!pdev->is_physfn)
1382 list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1383 if (pe->parent_dev != pdev)
1386 pnv_pci_ioda2_release_dma_pe(pdev, pe);
1388 /* Remove from list */
1389 mutex_lock(&phb->ioda.pe_list_mutex);
1390 list_del(&pe->list);
1391 mutex_unlock(&phb->ioda.pe_list_mutex);
1393 pnv_ioda_deconfigure_pe(phb, pe);
1395 pnv_ioda_free_pe(pe);
1399 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1401 struct pci_bus *bus;
1402 struct pci_controller *hose;
1403 struct pnv_phb *phb;
1404 struct pnv_ioda_pe *pe;
1406 struct pci_sriov *iov;
1410 hose = pci_bus_to_host(bus);
1411 phb = hose->private_data;
1412 pdn = pci_get_pdn(pdev);
1414 num_vfs = pdn->num_vfs;
1416 /* Release VF PEs */
1417 pnv_ioda_release_vf_PE(pdev);
1419 if (phb->type == PNV_PHB_IODA2) {
1420 if (!pdn->m64_single_mode)
1421 pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1423 /* Release M64 windows */
1424 pnv_pci_vf_release_m64(pdev, num_vfs);
1426 /* Release PE numbers */
1427 if (pdn->m64_single_mode) {
1428 for (i = 0; i < num_vfs; i++) {
1429 if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1432 pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1433 pnv_ioda_free_pe(pe);
1436 bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1437 /* Releasing pe_num_map */
1438 kfree(pdn->pe_num_map);
1442 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1443 struct pnv_ioda_pe *pe);
1444 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1446 struct pci_bus *bus;
1447 struct pci_controller *hose;
1448 struct pnv_phb *phb;
1449 struct pnv_ioda_pe *pe;
1455 hose = pci_bus_to_host(bus);
1456 phb = hose->private_data;
1457 pdn = pci_get_pdn(pdev);
1459 if (!pdev->is_physfn)
1462 /* Reserve PE for each VF */
1463 for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1464 if (pdn->m64_single_mode)
1465 pe_num = pdn->pe_num_map[vf_index];
1467 pe_num = *pdn->pe_num_map + vf_index;
1469 pe = &phb->ioda.pe_array[pe_num];
1470 pe->pe_number = pe_num;
1472 pe->flags = PNV_IODA_PE_VF;
1474 pe->parent_dev = pdev;
1475 pe->mve_number = -1;
1476 pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1477 pci_iov_virtfn_devfn(pdev, vf_index);
1479 pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%d\n",
1480 hose->global_number, pdev->bus->number,
1481 PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1482 PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1484 if (pnv_ioda_configure_pe(phb, pe)) {
1485 /* XXX What do we do here ? */
1486 pnv_ioda_free_pe(pe);
1491 /* Put PE to the list */
1492 mutex_lock(&phb->ioda.pe_list_mutex);
1493 list_add_tail(&pe->list, &phb->ioda.pe_list);
1494 mutex_unlock(&phb->ioda.pe_list_mutex);
1496 pnv_pci_ioda2_setup_dma_pe(phb, pe);
1500 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1502 struct pci_bus *bus;
1503 struct pci_controller *hose;
1504 struct pnv_phb *phb;
1505 struct pnv_ioda_pe *pe;
1511 hose = pci_bus_to_host(bus);
1512 phb = hose->private_data;
1513 pdn = pci_get_pdn(pdev);
1515 if (phb->type == PNV_PHB_IODA2) {
1516 if (!pdn->vfs_expanded) {
1517 dev_info(&pdev->dev, "don't support this SRIOV device"
1518 " with non 64bit-prefetchable IOV BAR\n");
1523 * When M64 BARs functions in Single PE mode, the number of VFs
1524 * could be enabled must be less than the number of M64 BARs.
1526 if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1527 dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1531 /* Allocating pe_num_map */
1532 if (pdn->m64_single_mode)
1533 pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map) * num_vfs,
1536 pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1538 if (!pdn->pe_num_map)
1541 if (pdn->m64_single_mode)
1542 for (i = 0; i < num_vfs; i++)
1543 pdn->pe_num_map[i] = IODA_INVALID_PE;
1545 /* Calculate available PE for required VFs */
1546 if (pdn->m64_single_mode) {
1547 for (i = 0; i < num_vfs; i++) {
1548 pe = pnv_ioda_alloc_pe(phb);
1554 pdn->pe_num_map[i] = pe->pe_number;
1557 mutex_lock(&phb->ioda.pe_alloc_mutex);
1558 *pdn->pe_num_map = bitmap_find_next_zero_area(
1559 phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1561 if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1562 mutex_unlock(&phb->ioda.pe_alloc_mutex);
1563 dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1564 kfree(pdn->pe_num_map);
1567 bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1568 mutex_unlock(&phb->ioda.pe_alloc_mutex);
1570 pdn->num_vfs = num_vfs;
1572 /* Assign M64 window accordingly */
1573 ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1575 dev_info(&pdev->dev, "Not enough M64 window resources\n");
1580 * When using one M64 BAR to map one IOV BAR, we need to shift
1581 * the IOV BAR according to the PE# allocated to the VFs.
1582 * Otherwise, the PE# for the VF will conflict with others.
1584 if (!pdn->m64_single_mode) {
1585 ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1592 pnv_ioda_setup_vf_PE(pdev, num_vfs);
1597 if (pdn->m64_single_mode) {
1598 for (i = 0; i < num_vfs; i++) {
1599 if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1602 pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1603 pnv_ioda_free_pe(pe);
1606 bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1608 /* Releasing pe_num_map */
1609 kfree(pdn->pe_num_map);
1614 int pcibios_sriov_disable(struct pci_dev *pdev)
1616 pnv_pci_sriov_disable(pdev);
1618 /* Release PCI data */
1619 remove_dev_pci_data(pdev);
1623 int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1625 /* Allocate PCI data */
1626 add_dev_pci_data(pdev);
1628 return pnv_pci_sriov_enable(pdev, num_vfs);
1630 #endif /* CONFIG_PCI_IOV */
1632 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1634 struct pci_dn *pdn = pci_get_pdn(pdev);
1635 struct pnv_ioda_pe *pe;
1638 * The function can be called while the PE#
1639 * hasn't been assigned. Do nothing for the
1642 if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1645 pe = &phb->ioda.pe_array[pdn->pe_number];
1646 WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1647 set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1648 set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1650 * Note: iommu_add_device() will fail here as
1651 * for physical PE: the device is already added by now;
1652 * for virtual PE: sysfs entries are not ready yet and
1653 * tce_iommu_bus_notifier will add the device to a group later.
1657 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1659 struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1660 struct pnv_phb *phb = hose->private_data;
1661 struct pci_dn *pdn = pci_get_pdn(pdev);
1662 struct pnv_ioda_pe *pe;
1664 bool bypass = false;
1666 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1669 pe = &phb->ioda.pe_array[pdn->pe_number];
1670 if (pe->tce_bypass_enabled) {
1671 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1672 bypass = (dma_mask >= top);
1676 dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1677 set_dma_ops(&pdev->dev, &dma_direct_ops);
1679 dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1680 set_dma_ops(&pdev->dev, &dma_iommu_ops);
1682 *pdev->dev.dma_mask = dma_mask;
1684 /* Update peer npu devices */
1685 pnv_npu_try_dma_set_bypass(pdev, bypass);
1690 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev)
1692 struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1693 struct pnv_phb *phb = hose->private_data;
1694 struct pci_dn *pdn = pci_get_pdn(pdev);
1695 struct pnv_ioda_pe *pe;
1698 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1701 pe = &phb->ioda.pe_array[pdn->pe_number];
1702 if (!pe->tce_bypass_enabled)
1703 return __dma_get_required_mask(&pdev->dev);
1706 end = pe->tce_bypass_base + memblock_end_of_DRAM();
1707 mask = 1ULL << (fls64(end) - 1);
1713 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1714 struct pci_bus *bus)
1716 struct pci_dev *dev;
1718 list_for_each_entry(dev, &bus->devices, bus_list) {
1719 set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1720 set_dma_offset(&dev->dev, pe->tce_bypass_base);
1721 iommu_add_device(&dev->dev);
1723 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1724 pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1728 static void pnv_pci_ioda1_tce_invalidate(struct iommu_table *tbl,
1729 unsigned long index, unsigned long npages, bool rm)
1731 struct iommu_table_group_link *tgl = list_first_entry_or_null(
1732 &tbl->it_group_list, struct iommu_table_group_link,
1734 struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1735 struct pnv_ioda_pe, table_group);
1736 __be64 __iomem *invalidate = rm ?
1737 (__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys :
1738 pe->phb->ioda.tce_inval_reg;
1739 unsigned long start, end, inc;
1740 const unsigned shift = tbl->it_page_shift;
1742 start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1743 end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1746 /* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
1747 if (tbl->it_busno) {
1750 inc = 128ull << shift;
1751 start |= tbl->it_busno;
1752 end |= tbl->it_busno;
1753 } else if (tbl->it_type & TCE_PCI_SWINV_PAIR) {
1754 /* p7ioc-style invalidation, 2 TCEs per write */
1755 start |= (1ull << 63);
1756 end |= (1ull << 63);
1759 /* Default (older HW) */
1763 end |= inc - 1; /* round up end to be different than start */
1765 mb(); /* Ensure above stores are visible */
1766 while (start <= end) {
1768 __raw_rm_writeq(cpu_to_be64(start), invalidate);
1770 __raw_writeq(cpu_to_be64(start), invalidate);
1775 * The iommu layer will do another mb() for us on build()
1776 * and we don't care on free()
1780 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1781 long npages, unsigned long uaddr,
1782 enum dma_data_direction direction,
1783 struct dma_attrs *attrs)
1785 int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1788 if (!ret && (tbl->it_type & TCE_PCI_SWINV_CREATE))
1789 pnv_pci_ioda1_tce_invalidate(tbl, index, npages, false);
1794 #ifdef CONFIG_IOMMU_API
1795 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
1796 unsigned long *hpa, enum dma_data_direction *direction)
1798 long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1800 if (!ret && (tbl->it_type &
1801 (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE)))
1802 pnv_pci_ioda1_tce_invalidate(tbl, index, 1, false);
1808 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1811 pnv_tce_free(tbl, index, npages);
1813 if (tbl->it_type & TCE_PCI_SWINV_FREE)
1814 pnv_pci_ioda1_tce_invalidate(tbl, index, npages, false);
1817 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1818 .set = pnv_ioda1_tce_build,
1819 #ifdef CONFIG_IOMMU_API
1820 .exchange = pnv_ioda1_tce_xchg,
1822 .clear = pnv_ioda1_tce_free,
1826 #define TCE_KILL_INVAL_ALL PPC_BIT(0)
1827 #define TCE_KILL_INVAL_PE PPC_BIT(1)
1828 #define TCE_KILL_INVAL_TCE PPC_BIT(2)
1830 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
1832 const unsigned long val = TCE_KILL_INVAL_ALL;
1834 mb(); /* Ensure previous TCE table stores are visible */
1836 __raw_rm_writeq(cpu_to_be64(val),
1838 phb->ioda.tce_inval_reg_phys);
1840 __raw_writeq(cpu_to_be64(val), phb->ioda.tce_inval_reg);
1843 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1845 /* 01xb - invalidate TCEs that match the specified PE# */
1846 unsigned long val = TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1847 struct pnv_phb *phb = pe->phb;
1849 if (!phb->ioda.tce_inval_reg)
1852 mb(); /* Ensure above stores are visible */
1853 __raw_writeq(cpu_to_be64(val), phb->ioda.tce_inval_reg);
1856 static void pnv_pci_ioda2_do_tce_invalidate(unsigned pe_number, bool rm,
1857 __be64 __iomem *invalidate, unsigned shift,
1858 unsigned long index, unsigned long npages)
1860 unsigned long start, end, inc;
1862 /* We'll invalidate DMA address in PE scope */
1863 start = TCE_KILL_INVAL_TCE;
1864 start |= (pe_number & 0xFF);
1867 /* Figure out the start, end and step */
1868 start |= (index << shift);
1869 end |= ((index + npages - 1) << shift);
1870 inc = (0x1ull << shift);
1873 while (start <= end) {
1875 __raw_rm_writeq(cpu_to_be64(start), invalidate);
1877 __raw_writeq(cpu_to_be64(start), invalidate);
1882 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1883 unsigned long index, unsigned long npages, bool rm)
1885 struct iommu_table_group_link *tgl;
1887 list_for_each_entry_rcu(tgl, &tbl->it_group_list, next) {
1888 struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1889 struct pnv_ioda_pe, table_group);
1890 __be64 __iomem *invalidate = rm ?
1891 (__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys :
1892 pe->phb->ioda.tce_inval_reg;
1894 if (pe->phb->type == PNV_PHB_NPU) {
1896 * The NVLink hardware does not support TCE kill
1897 * per TCE entry so we have to invalidate
1898 * the entire cache for it.
1900 pnv_pci_ioda2_tce_invalidate_entire(pe->phb, rm);
1903 pnv_pci_ioda2_do_tce_invalidate(pe->pe_number, rm,
1904 invalidate, tbl->it_page_shift,
1909 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1910 long npages, unsigned long uaddr,
1911 enum dma_data_direction direction,
1912 struct dma_attrs *attrs)
1914 int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1917 if (!ret && (tbl->it_type & TCE_PCI_SWINV_CREATE))
1918 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1923 #ifdef CONFIG_IOMMU_API
1924 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
1925 unsigned long *hpa, enum dma_data_direction *direction)
1927 long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1929 if (!ret && (tbl->it_type &
1930 (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE)))
1931 pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
1937 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1940 pnv_tce_free(tbl, index, npages);
1942 if (tbl->it_type & TCE_PCI_SWINV_FREE)
1943 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1946 static void pnv_ioda2_table_free(struct iommu_table *tbl)
1948 pnv_pci_ioda2_table_free_pages(tbl);
1949 iommu_free_table(tbl, "pnv");
1952 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
1953 .set = pnv_ioda2_tce_build,
1954 #ifdef CONFIG_IOMMU_API
1955 .exchange = pnv_ioda2_tce_xchg,
1957 .clear = pnv_ioda2_tce_free,
1959 .free = pnv_ioda2_table_free,
1962 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
1964 unsigned int *weight = (unsigned int *)data;
1966 /* This is quite simplistic. The "base" weight of a device
1967 * is 10. 0 means no DMA is to be accounted for it.
1969 if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1972 if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
1973 dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
1974 dev->class == PCI_CLASS_SERIAL_USB_EHCI)
1976 else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
1984 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
1986 unsigned int weight = 0;
1988 /* SRIOV VF has same DMA32 weight as its PF */
1989 #ifdef CONFIG_PCI_IOV
1990 if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
1991 pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
1996 if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
1997 pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
1998 } else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
1999 struct pci_dev *pdev;
2001 list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
2002 pnv_pci_ioda_dev_dma_weight(pdev, &weight);
2003 } else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2004 pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2010 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2011 struct pnv_ioda_pe *pe)
2014 struct page *tce_mem = NULL;
2015 struct iommu_table *tbl;
2016 unsigned int weight, total_weight = 0;
2017 unsigned int tce32_segsz, base, segs, avail, i;
2021 /* XXX FIXME: Handle 64-bit only DMA devices */
2022 /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2023 /* XXX FIXME: Allocate multi-level tables on PHB3 */
2024 weight = pnv_pci_ioda_pe_dma_weight(pe);
2028 pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2030 segs = (weight * phb->ioda.dma32_count) / total_weight;
2035 * Allocate contiguous DMA32 segments. We begin with the expected
2036 * number of segments. With one more attempt, the number of DMA32
2037 * segments to be allocated is decreased by one until one segment
2038 * is allocated successfully.
2041 for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2042 for (avail = 0, i = base; i < base + segs; i++) {
2043 if (phb->ioda.dma32_segmap[i] ==
2054 pe_warn(pe, "No available DMA32 segments\n");
2059 tbl = pnv_pci_table_alloc(phb->hose->node);
2060 iommu_register_group(&pe->table_group, phb->hose->global_number,
2062 pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2064 /* Grab a 32-bit TCE table */
2065 pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2066 weight, total_weight, base, segs);
2067 pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2068 base * PNV_IODA1_DMA32_SEGSIZE,
2069 (base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2071 /* XXX Currently, we allocate one big contiguous table for the
2072 * TCEs. We only really need one chunk per 256M of TCE space
2073 * (ie per segment) but that's an optimization for later, it
2074 * requires some added smarts with our get/put_tce implementation
2076 * Each TCE page is 4KB in size and each TCE entry occupies 8
2079 tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2080 tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2081 get_order(tce32_segsz * segs));
2083 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2086 addr = page_address(tce_mem);
2087 memset(addr, 0, tce32_segsz * segs);
2090 for (i = 0; i < segs; i++) {
2091 rc = opal_pci_map_pe_dma_window(phb->opal_id,
2094 __pa(addr) + tce32_segsz * i,
2095 tce32_segsz, IOMMU_PAGE_SIZE_4K);
2097 pe_err(pe, " Failed to configure 32-bit TCE table,"
2103 /* Setup DMA32 segment mapping */
2104 for (i = base; i < base + segs; i++)
2105 phb->ioda.dma32_segmap[i] = pe->pe_number;
2107 /* Setup linux iommu table */
2108 pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2109 base * PNV_IODA1_DMA32_SEGSIZE,
2110 IOMMU_PAGE_SHIFT_4K);
2112 /* OPAL variant of P7IOC SW invalidated TCEs */
2113 if (phb->ioda.tce_inval_reg)
2114 tbl->it_type |= (TCE_PCI_SWINV_CREATE |
2115 TCE_PCI_SWINV_FREE |
2116 TCE_PCI_SWINV_PAIR);
2118 tbl->it_ops = &pnv_ioda1_iommu_ops;
2119 pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2120 pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2121 iommu_init_table(tbl, phb->hose->node);
2123 if (pe->flags & PNV_IODA_PE_DEV) {
2125 * Setting table base here only for carrying iommu_group
2126 * further down to let iommu_add_device() do the job.
2127 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2129 set_iommu_table_base(&pe->pdev->dev, tbl);
2130 iommu_add_device(&pe->pdev->dev);
2131 } else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2132 pnv_ioda_setup_bus_dma(pe, pe->pbus);
2136 /* XXX Failure: Try to fallback to 64-bit only ? */
2138 __free_pages(tce_mem, get_order(tce32_segsz * segs));
2140 pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2141 iommu_free_table(tbl, "pnv");
2145 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2146 int num, struct iommu_table *tbl)
2148 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2150 struct pnv_phb *phb = pe->phb;
2152 const unsigned long size = tbl->it_indirect_levels ?
2153 tbl->it_level_size : tbl->it_size;
2154 const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2155 const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2157 pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num,
2158 start_addr, start_addr + win_size - 1,
2159 IOMMU_PAGE_SIZE(tbl));
2162 * Map TCE table through TVT. The TVE index is the PE number
2163 * shifted by 1 bit for 32-bits DMA space.
2165 rc = opal_pci_map_pe_dma_window(phb->opal_id,
2167 (pe->pe_number << 1) + num,
2168 tbl->it_indirect_levels + 1,
2171 IOMMU_PAGE_SIZE(tbl));
2173 pe_err(pe, "Failed to configure TCE table, err %ld\n", rc);
2177 pnv_pci_link_table_and_group(phb->hose->node, num,
2178 tbl, &pe->table_group);
2179 pnv_pci_ioda2_tce_invalidate_pe(pe);
2184 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2186 uint16_t window_id = (pe->pe_number << 1 ) + 1;
2189 pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2191 phys_addr_t top = memblock_end_of_DRAM();
2193 top = roundup_pow_of_two(top);
2194 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2197 pe->tce_bypass_base,
2200 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2203 pe->tce_bypass_base,
2207 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2209 pe->tce_bypass_enabled = enable;
2212 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2213 __u32 page_shift, __u64 window_size, __u32 levels,
2214 struct iommu_table *tbl);
2216 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2217 int num, __u32 page_shift, __u64 window_size, __u32 levels,
2218 struct iommu_table **ptbl)
2220 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2222 int nid = pe->phb->hose->node;
2223 __u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2225 struct iommu_table *tbl;
2227 tbl = pnv_pci_table_alloc(nid);
2231 ret = pnv_pci_ioda2_table_alloc_pages(nid,
2232 bus_offset, page_shift, window_size,
2235 iommu_free_table(tbl, "pnv");
2239 tbl->it_ops = &pnv_ioda2_iommu_ops;
2240 if (pe->phb->ioda.tce_inval_reg)
2241 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
2248 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2250 struct iommu_table *tbl = NULL;
2254 * crashkernel= specifies the kdump kernel's maximum memory at
2255 * some offset and there is no guaranteed the result is a power
2256 * of 2, which will cause errors later.
2258 const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2261 * In memory constrained environments, e.g. kdump kernel, the
2262 * DMA window can be larger than available memory, which will
2263 * cause errors later.
2265 const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2267 rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2268 IOMMU_PAGE_SHIFT_4K,
2270 POWERNV_IOMMU_DEFAULT_LEVELS, &tbl);
2272 pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2277 iommu_init_table(tbl, pe->phb->hose->node);
2279 rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2281 pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2283 pnv_ioda2_table_free(tbl);
2287 if (!pnv_iommu_bypass_disabled)
2288 pnv_pci_ioda2_set_bypass(pe, true);
2290 /* OPAL variant of PHB3 invalidated TCEs */
2291 if (pe->phb->ioda.tce_inval_reg)
2292 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
2295 * Setting table base here only for carrying iommu_group
2296 * further down to let iommu_add_device() do the job.
2297 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2299 if (pe->flags & PNV_IODA_PE_DEV)
2300 set_iommu_table_base(&pe->pdev->dev, tbl);
2305 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2306 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2309 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2311 struct pnv_phb *phb = pe->phb;
2314 pe_info(pe, "Removing DMA window #%d\n", num);
2316 ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2317 (pe->pe_number << 1) + num,
2318 0/* levels */, 0/* table address */,
2319 0/* table size */, 0/* page size */);
2321 pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2323 pnv_pci_ioda2_tce_invalidate_pe(pe);
2325 pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2331 #ifdef CONFIG_IOMMU_API
2332 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2333 __u64 window_size, __u32 levels)
2335 unsigned long bytes = 0;
2336 const unsigned window_shift = ilog2(window_size);
2337 unsigned entries_shift = window_shift - page_shift;
2338 unsigned table_shift = entries_shift + 3;
2339 unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2340 unsigned long direct_table_size;
2342 if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2343 (window_size > memory_hotplug_max()) ||
2344 !is_power_of_2(window_size))
2347 /* Calculate a direct table size from window_size and levels */
2348 entries_shift = (entries_shift + levels - 1) / levels;
2349 table_shift = entries_shift + 3;
2350 table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2351 direct_table_size = 1UL << table_shift;
2353 for ( ; levels; --levels) {
2354 bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2356 tce_table_size /= direct_table_size;
2357 tce_table_size <<= 3;
2358 tce_table_size = _ALIGN_UP(tce_table_size, direct_table_size);
2364 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2366 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2368 /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2369 struct iommu_table *tbl = pe->table_group.tables[0];
2371 pnv_pci_ioda2_set_bypass(pe, false);
2372 pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2373 pnv_ioda2_table_free(tbl);
2376 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2378 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2381 pnv_pci_ioda2_setup_default_config(pe);
2384 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2385 .get_table_size = pnv_pci_ioda2_get_table_size,
2386 .create_table = pnv_pci_ioda2_create_table,
2387 .set_window = pnv_pci_ioda2_set_window,
2388 .unset_window = pnv_pci_ioda2_unset_window,
2389 .take_ownership = pnv_ioda2_take_ownership,
2390 .release_ownership = pnv_ioda2_release_ownership,
2394 static void pnv_pci_ioda_setup_opal_tce_kill(struct pnv_phb *phb)
2396 const __be64 *swinvp;
2398 /* OPAL variant of PHB3 invalidated TCEs */
2399 swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
2403 phb->ioda.tce_inval_reg_phys = be64_to_cpup(swinvp);
2404 phb->ioda.tce_inval_reg = ioremap(phb->ioda.tce_inval_reg_phys, 8);
2407 static __be64 *pnv_pci_ioda2_table_do_alloc_pages(int nid, unsigned shift,
2408 unsigned levels, unsigned long limit,
2409 unsigned long *current_offset, unsigned long *total_allocated)
2411 struct page *tce_mem = NULL;
2413 unsigned order = max_t(unsigned, shift, PAGE_SHIFT) - PAGE_SHIFT;
2414 unsigned long allocated = 1UL << (order + PAGE_SHIFT);
2415 unsigned entries = 1UL << (shift - 3);
2418 tce_mem = alloc_pages_node(nid, GFP_KERNEL, order);
2420 pr_err("Failed to allocate a TCE memory, order=%d\n", order);
2423 addr = page_address(tce_mem);
2424 memset(addr, 0, allocated);
2425 *total_allocated += allocated;
2429 *current_offset += allocated;
2433 for (i = 0; i < entries; ++i) {
2434 tmp = pnv_pci_ioda2_table_do_alloc_pages(nid, shift,
2435 levels, limit, current_offset, total_allocated);
2439 addr[i] = cpu_to_be64(__pa(tmp) |
2440 TCE_PCI_READ | TCE_PCI_WRITE);
2442 if (*current_offset >= limit)
2449 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2450 unsigned long size, unsigned level);
2452 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2453 __u32 page_shift, __u64 window_size, __u32 levels,
2454 struct iommu_table *tbl)
2457 unsigned long offset = 0, level_shift, total_allocated = 0;
2458 const unsigned window_shift = ilog2(window_size);
2459 unsigned entries_shift = window_shift - page_shift;
2460 unsigned table_shift = max_t(unsigned, entries_shift + 3, PAGE_SHIFT);
2461 const unsigned long tce_table_size = 1UL << table_shift;
2463 if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS))
2466 if ((window_size > memory_hotplug_max()) || !is_power_of_2(window_size))
2469 /* Adjust direct table size from window_size and levels */
2470 entries_shift = (entries_shift + levels - 1) / levels;
2471 level_shift = entries_shift + 3;
2472 level_shift = max_t(unsigned, level_shift, PAGE_SHIFT);
2474 /* Allocate TCE table */
2475 addr = pnv_pci_ioda2_table_do_alloc_pages(nid, level_shift,
2476 levels, tce_table_size, &offset, &total_allocated);
2478 /* addr==NULL means that the first level allocation failed */
2483 * First level was allocated but some lower level failed as
2484 * we did not allocate as much as we wanted,
2485 * release partially allocated table.
2487 if (offset < tce_table_size) {
2488 pnv_pci_ioda2_table_do_free_pages(addr,
2489 1ULL << (level_shift - 3), levels - 1);
2493 /* Setup linux iommu table */
2494 pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, bus_offset,
2496 tbl->it_level_size = 1ULL << (level_shift - 3);
2497 tbl->it_indirect_levels = levels - 1;
2498 tbl->it_allocated_size = total_allocated;
2500 pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2501 window_size, tce_table_size, bus_offset);
2506 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2507 unsigned long size, unsigned level)
2509 const unsigned long addr_ul = (unsigned long) addr &
2510 ~(TCE_PCI_READ | TCE_PCI_WRITE);
2514 u64 *tmp = (u64 *) addr_ul;
2516 for (i = 0; i < size; ++i) {
2517 unsigned long hpa = be64_to_cpu(tmp[i]);
2519 if (!(hpa & (TCE_PCI_READ | TCE_PCI_WRITE)))
2522 pnv_pci_ioda2_table_do_free_pages(__va(hpa), size,
2527 free_pages(addr_ul, get_order(size << 3));
2530 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl)
2532 const unsigned long size = tbl->it_indirect_levels ?
2533 tbl->it_level_size : tbl->it_size;
2538 pnv_pci_ioda2_table_do_free_pages((__be64 *)tbl->it_base, size,
2539 tbl->it_indirect_levels);
2542 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2543 struct pnv_ioda_pe *pe)
2547 /* TVE #1 is selected by PCI address bit 59 */
2548 pe->tce_bypass_base = 1ull << 59;
2550 iommu_register_group(&pe->table_group, phb->hose->global_number,
2553 /* The PE will reserve all possible 32-bits space */
2554 pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2555 phb->ioda.m32_pci_base);
2557 /* Setup linux iommu table */
2558 pe->table_group.tce32_start = 0;
2559 pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2560 pe->table_group.max_dynamic_windows_supported =
2561 IOMMU_TABLE_GROUP_MAX_TABLES;
2562 pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2563 pe->table_group.pgsizes = SZ_4K | SZ_64K | SZ_16M;
2564 #ifdef CONFIG_IOMMU_API
2565 pe->table_group.ops = &pnv_pci_ioda2_ops;
2568 rc = pnv_pci_ioda2_setup_default_config(pe);
2572 if (pe->flags & PNV_IODA_PE_DEV)
2573 iommu_add_device(&pe->pdev->dev);
2574 else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2575 pnv_ioda_setup_bus_dma(pe, pe->pbus);
2578 static void pnv_ioda_setup_dma(struct pnv_phb *phb)
2580 struct pci_controller *hose = phb->hose;
2581 struct pnv_ioda_pe *pe;
2582 unsigned int weight;
2584 /* If we have more PE# than segments available, hand out one
2585 * per PE until we run out and let the rest fail. If not,
2586 * then we assign at least one segment per PE, plus more based
2587 * on the amount of devices under that PE
2589 pr_info("PCI: Domain %04x has %d available 32-bit DMA segments\n",
2590 hose->global_number, phb->ioda.dma32_count);
2592 pnv_pci_ioda_setup_opal_tce_kill(phb);
2594 /* Walk our PE list and configure their DMA segments */
2595 list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2596 weight = pnv_pci_ioda_pe_dma_weight(pe);
2601 * For IODA2 compliant PHB3, we needn't care about the weight.
2602 * The all available 32-bits DMA space will be assigned to
2605 if (phb->type == PNV_PHB_IODA1) {
2606 pnv_pci_ioda1_setup_dma_pe(phb, pe);
2607 } else if (phb->type == PNV_PHB_IODA2) {
2608 pe_info(pe, "Assign DMA32 space\n");
2609 pnv_pci_ioda2_setup_dma_pe(phb, pe);
2610 } else if (phb->type == PNV_PHB_NPU) {
2612 * We initialise the DMA space for an NPU PHB
2613 * after setup of the PHB is complete as we
2614 * point the NPU TVT to the the same location
2621 #ifdef CONFIG_PCI_MSI
2622 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2624 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2625 struct irq_chip *chip = irq_data_get_irq_chip(d);
2626 struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2630 rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
2637 static void set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2639 struct irq_data *idata;
2640 struct irq_chip *ichip;
2642 if (phb->type != PNV_PHB_IODA2)
2645 if (!phb->ioda.irq_chip_init) {
2647 * First time we setup an MSI IRQ, we need to setup the
2648 * corresponding IRQ chip to route correctly.
2650 idata = irq_get_irq_data(virq);
2651 ichip = irq_data_get_irq_chip(idata);
2652 phb->ioda.irq_chip_init = 1;
2653 phb->ioda.irq_chip = *ichip;
2654 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2656 irq_set_chip(virq, &phb->ioda.irq_chip);
2659 #ifdef CONFIG_CXL_BASE
2661 struct device_node *pnv_pci_get_phb_node(struct pci_dev *dev)
2663 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2665 return of_node_get(hose->dn);
2667 EXPORT_SYMBOL(pnv_pci_get_phb_node);
2669 int pnv_phb_to_cxl_mode(struct pci_dev *dev, uint64_t mode)
2671 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2672 struct pnv_phb *phb = hose->private_data;
2673 struct pnv_ioda_pe *pe;
2676 pe = pnv_ioda_get_pe(dev);
2680 pe_info(pe, "Switching PHB to CXL\n");
2682 rc = opal_pci_set_phb_cxl_mode(phb->opal_id, mode, pe->pe_number);
2684 dev_err(&dev->dev, "opal_pci_set_phb_cxl_mode failed: %i\n", rc);
2688 EXPORT_SYMBOL(pnv_phb_to_cxl_mode);
2690 /* Find PHB for cxl dev and allocate MSI hwirqs?
2691 * Returns the absolute hardware IRQ number
2693 int pnv_cxl_alloc_hwirqs(struct pci_dev *dev, int num)
2695 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2696 struct pnv_phb *phb = hose->private_data;
2697 int hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, num);
2700 dev_warn(&dev->dev, "Failed to find a free MSI\n");
2704 return phb->msi_base + hwirq;
2706 EXPORT_SYMBOL(pnv_cxl_alloc_hwirqs);
2708 void pnv_cxl_release_hwirqs(struct pci_dev *dev, int hwirq, int num)
2710 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2711 struct pnv_phb *phb = hose->private_data;
2713 msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq - phb->msi_base, num);
2715 EXPORT_SYMBOL(pnv_cxl_release_hwirqs);
2717 void pnv_cxl_release_hwirq_ranges(struct cxl_irq_ranges *irqs,
2718 struct pci_dev *dev)
2720 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2721 struct pnv_phb *phb = hose->private_data;
2724 for (i = 1; i < CXL_IRQ_RANGES; i++) {
2725 if (!irqs->range[i])
2727 pr_devel("cxl release irq range 0x%x: offset: 0x%lx limit: %ld\n",
2730 hwirq = irqs->offset[i] - phb->msi_base;
2731 msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq,
2735 EXPORT_SYMBOL(pnv_cxl_release_hwirq_ranges);
2737 int pnv_cxl_alloc_hwirq_ranges(struct cxl_irq_ranges *irqs,
2738 struct pci_dev *dev, int num)
2740 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2741 struct pnv_phb *phb = hose->private_data;
2744 memset(irqs, 0, sizeof(struct cxl_irq_ranges));
2746 /* 0 is reserved for the multiplexed PSL DSI interrupt */
2747 for (i = 1; i < CXL_IRQ_RANGES && num; i++) {
2750 hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, try);
2758 irqs->offset[i] = phb->msi_base + hwirq;
2759 irqs->range[i] = try;
2760 pr_devel("cxl alloc irq range 0x%x: offset: 0x%lx limit: %li\n",
2761 i, irqs->offset[i], irqs->range[i]);
2769 pnv_cxl_release_hwirq_ranges(irqs, dev);
2772 EXPORT_SYMBOL(pnv_cxl_alloc_hwirq_ranges);
2774 int pnv_cxl_get_irq_count(struct pci_dev *dev)
2776 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2777 struct pnv_phb *phb = hose->private_data;
2779 return phb->msi_bmp.irq_count;
2781 EXPORT_SYMBOL(pnv_cxl_get_irq_count);
2783 int pnv_cxl_ioda_msi_setup(struct pci_dev *dev, unsigned int hwirq,
2786 struct pci_controller *hose = pci_bus_to_host(dev->bus);
2787 struct pnv_phb *phb = hose->private_data;
2788 unsigned int xive_num = hwirq - phb->msi_base;
2789 struct pnv_ioda_pe *pe;
2792 if (!(pe = pnv_ioda_get_pe(dev)))
2795 /* Assign XIVE to PE */
2796 rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2798 pe_warn(pe, "%s: OPAL error %d setting msi_base 0x%x "
2799 "hwirq 0x%x XIVE 0x%x PE\n",
2800 pci_name(dev), rc, phb->msi_base, hwirq, xive_num);
2803 set_msi_irq_chip(phb, virq);
2807 EXPORT_SYMBOL(pnv_cxl_ioda_msi_setup);
2810 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2811 unsigned int hwirq, unsigned int virq,
2812 unsigned int is_64, struct msi_msg *msg)
2814 struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2815 unsigned int xive_num = hwirq - phb->msi_base;
2819 /* No PE assigned ? bail out ... no MSI for you ! */
2823 /* Check if we have an MVE */
2824 if (pe->mve_number < 0)
2827 /* Force 32-bit MSI on some broken devices */
2828 if (dev->no_64bit_msi)
2831 /* Assign XIVE to PE */
2832 rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2834 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2835 pci_name(dev), rc, xive_num);
2842 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2845 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2849 msg->address_hi = be64_to_cpu(addr64) >> 32;
2850 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2854 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2857 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2861 msg->address_hi = 0;
2862 msg->address_lo = be32_to_cpu(addr32);
2864 msg->data = be32_to_cpu(data);
2866 set_msi_irq_chip(phb, virq);
2868 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2869 " address=%x_%08x data=%x PE# %d\n",
2870 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2871 msg->address_hi, msg->address_lo, data, pe->pe_number);
2876 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2879 const __be32 *prop = of_get_property(phb->hose->dn,
2880 "ibm,opal-msi-ranges", NULL);
2883 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2888 phb->msi_base = be32_to_cpup(prop);
2889 count = be32_to_cpup(prop + 1);
2890 if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2891 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2892 phb->hose->global_number);
2896 phb->msi_setup = pnv_pci_ioda_msi_setup;
2897 phb->msi32_support = 1;
2898 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2899 count, phb->msi_base);
2902 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
2903 #endif /* CONFIG_PCI_MSI */
2905 #ifdef CONFIG_PCI_IOV
2906 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2908 struct pci_controller *hose = pci_bus_to_host(pdev->bus);
2909 struct pnv_phb *phb = hose->private_data;
2910 const resource_size_t gate = phb->ioda.m64_segsize >> 2;
2911 struct resource *res;
2913 resource_size_t size, total_vf_bar_sz;
2917 if (!pdev->is_physfn || pdev->is_added)
2920 pdn = pci_get_pdn(pdev);
2921 pdn->vfs_expanded = 0;
2922 pdn->m64_single_mode = false;
2924 total_vfs = pci_sriov_get_totalvfs(pdev);
2925 mul = phb->ioda.total_pe_num;
2926 total_vf_bar_sz = 0;
2928 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2929 res = &pdev->resource[i + PCI_IOV_RESOURCES];
2930 if (!res->flags || res->parent)
2932 if (!pnv_pci_is_mem_pref_64(res->flags)) {
2933 dev_warn(&pdev->dev, "Don't support SR-IOV with"
2934 " non M64 VF BAR%d: %pR. \n",
2939 total_vf_bar_sz += pci_iov_resource_size(pdev,
2940 i + PCI_IOV_RESOURCES);
2943 * If bigger than quarter of M64 segment size, just round up
2946 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
2947 * with other devices, IOV BAR size is expanded to be
2948 * (total_pe * VF_BAR_size). When VF_BAR_size is half of M64
2949 * segment size , the expanded size would equal to half of the
2950 * whole M64 space size, which will exhaust the M64 Space and
2951 * limit the system flexibility. This is a design decision to
2952 * set the boundary to quarter of the M64 segment size.
2954 if (total_vf_bar_sz > gate) {
2955 mul = roundup_pow_of_two(total_vfs);
2956 dev_info(&pdev->dev,
2957 "VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
2958 total_vf_bar_sz, gate, mul);
2959 pdn->m64_single_mode = true;
2964 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2965 res = &pdev->resource[i + PCI_IOV_RESOURCES];
2966 if (!res->flags || res->parent)
2969 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
2971 * On PHB3, the minimum size alignment of M64 BAR in single
2974 if (pdn->m64_single_mode && (size < SZ_32M))
2976 dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
2977 res->end = res->start + size * mul - 1;
2978 dev_dbg(&pdev->dev, " %pR\n", res);
2979 dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2982 pdn->vfs_expanded = mul;
2987 /* To save MMIO space, IOV BAR is truncated. */
2988 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2989 res = &pdev->resource[i + PCI_IOV_RESOURCES];
2991 res->end = res->start - 1;
2994 #endif /* CONFIG_PCI_IOV */
2996 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2997 struct resource *res)
2999 struct pnv_phb *phb = pe->phb;
3000 struct pci_bus_region region;
3004 if (!res || !res->flags || res->start > res->end)
3007 if (res->flags & IORESOURCE_IO) {
3008 region.start = res->start - phb->ioda.io_pci_base;
3009 region.end = res->end - phb->ioda.io_pci_base;
3010 index = region.start / phb->ioda.io_segsize;
3012 while (index < phb->ioda.total_pe_num &&
3013 region.start <= region.end) {
3014 phb->ioda.io_segmap[index] = pe->pe_number;
3015 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3016 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
3017 if (rc != OPAL_SUCCESS) {
3018 pr_err("%s: Error %lld mapping IO segment#%d to PE#%d\n",
3019 __func__, rc, index, pe->pe_number);
3023 region.start += phb->ioda.io_segsize;
3026 } else if ((res->flags & IORESOURCE_MEM) &&
3027 !pnv_pci_is_mem_pref_64(res->flags)) {
3028 region.start = res->start -
3029 phb->hose->mem_offset[0] -
3030 phb->ioda.m32_pci_base;
3031 region.end = res->end -
3032 phb->hose->mem_offset[0] -
3033 phb->ioda.m32_pci_base;
3034 index = region.start / phb->ioda.m32_segsize;
3036 while (index < phb->ioda.total_pe_num &&
3037 region.start <= region.end) {
3038 phb->ioda.m32_segmap[index] = pe->pe_number;
3039 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3040 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
3041 if (rc != OPAL_SUCCESS) {
3042 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%d",
3043 __func__, rc, index, pe->pe_number);
3047 region.start += phb->ioda.m32_segsize;
3054 * This function is supposed to be called on basis of PE from top
3055 * to bottom style. So the the I/O or MMIO segment assigned to
3056 * parent PE could be overrided by its child PEs if necessary.
3058 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
3060 struct pci_dev *pdev;
3064 * NOTE: We only care PCI bus based PE for now. For PCI
3065 * device based PE, for example SRIOV sensitive VF should
3066 * be figured out later.
3068 BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3070 list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3071 for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3072 pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3075 * If the PE contains all subordinate PCI buses, the
3076 * windows of the child bridges should be mapped to
3079 if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3081 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3082 pnv_ioda_setup_pe_res(pe,
3083 &pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3087 static void pnv_pci_ioda_setup_seg(void)
3089 struct pci_controller *tmp, *hose;
3090 struct pnv_phb *phb;
3091 struct pnv_ioda_pe *pe;
3093 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3094 phb = hose->private_data;
3096 /* NPU PHB does not support IO or MMIO segmentation */
3097 if (phb->type == PNV_PHB_NPU)
3100 list_for_each_entry(pe, &phb->ioda.pe_list, list) {
3101 pnv_ioda_setup_pe_seg(pe);
3106 static void pnv_pci_ioda_setup_DMA(void)
3108 struct pci_controller *hose, *tmp;
3109 struct pnv_phb *phb;
3111 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3112 pnv_ioda_setup_dma(hose->private_data);
3114 /* Mark the PHB initialization done */
3115 phb = hose->private_data;
3116 phb->initialized = 1;
3120 static void pnv_pci_ioda_create_dbgfs(void)
3122 #ifdef CONFIG_DEBUG_FS
3123 struct pci_controller *hose, *tmp;
3124 struct pnv_phb *phb;
3127 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3128 phb = hose->private_data;
3130 sprintf(name, "PCI%04x", hose->global_number);
3131 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3133 pr_warning("%s: Error on creating debugfs on PHB#%x\n",
3134 __func__, hose->global_number);
3136 #endif /* CONFIG_DEBUG_FS */
3139 static void pnv_pci_ioda_fixup(void)
3141 pnv_pci_ioda_setup_PEs();
3142 pnv_pci_ioda_setup_seg();
3143 pnv_pci_ioda_setup_DMA();
3145 pnv_pci_ioda_create_dbgfs();
3149 eeh_addr_cache_build();
3154 * Returns the alignment for I/O or memory windows for P2P
3155 * bridges. That actually depends on how PEs are segmented.
3156 * For now, we return I/O or M32 segment size for PE sensitive
3157 * P2P bridges. Otherwise, the default values (4KiB for I/O,
3158 * 1MiB for memory) will be returned.
3160 * The current PCI bus might be put into one PE, which was
3161 * create against the parent PCI bridge. For that case, we
3162 * needn't enlarge the alignment so that we can save some
3165 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3168 struct pci_dev *bridge;
3169 struct pci_controller *hose = pci_bus_to_host(bus);
3170 struct pnv_phb *phb = hose->private_data;
3171 int num_pci_bridges = 0;
3175 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3177 if (num_pci_bridges >= 2)
3181 bridge = bridge->bus->self;
3184 /* We fail back to M32 if M64 isn't supported */
3185 if (phb->ioda.m64_segsize &&
3186 pnv_pci_is_mem_pref_64(type))
3187 return phb->ioda.m64_segsize;
3188 if (type & IORESOURCE_MEM)
3189 return phb->ioda.m32_segsize;
3191 return phb->ioda.io_segsize;
3194 #ifdef CONFIG_PCI_IOV
3195 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3198 struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3199 struct pnv_phb *phb = hose->private_data;
3200 struct pci_dn *pdn = pci_get_pdn(pdev);
3201 resource_size_t align;
3204 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3205 * SR-IOV. While from hardware perspective, the range mapped by M64
3206 * BAR should be size aligned.
3208 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3209 * powernv-specific hardware restriction is gone. But if just use the
3210 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3211 * in one segment of M64 #15, which introduces the PE conflict between
3212 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3215 * This function returns the total IOV BAR size if M64 BAR is in
3216 * Shared PE mode or just VF BAR size if not.
3217 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3218 * M64 segment size if IOV BAR size is less.
3220 align = pci_iov_resource_size(pdev, resno);
3221 if (!pdn->vfs_expanded)
3223 if (pdn->m64_single_mode)
3224 return max(align, (resource_size_t)phb->ioda.m64_segsize);
3226 return pdn->vfs_expanded * align;
3228 #endif /* CONFIG_PCI_IOV */
3230 /* Prevent enabling devices for which we couldn't properly
3233 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3235 struct pci_controller *hose = pci_bus_to_host(dev->bus);
3236 struct pnv_phb *phb = hose->private_data;
3239 /* The function is probably called while the PEs have
3240 * not be created yet. For example, resource reassignment
3241 * during PCI probe period. We just skip the check if
3244 if (!phb->initialized)
3247 pdn = pci_get_pdn(dev);
3248 if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3254 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3256 struct pnv_phb *phb = hose->private_data;
3258 opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3262 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3263 .dma_dev_setup = pnv_pci_dma_dev_setup,
3264 .dma_bus_setup = pnv_pci_dma_bus_setup,
3265 #ifdef CONFIG_PCI_MSI
3266 .setup_msi_irqs = pnv_setup_msi_irqs,
3267 .teardown_msi_irqs = pnv_teardown_msi_irqs,
3269 .enable_device_hook = pnv_pci_enable_device_hook,
3270 .window_alignment = pnv_pci_window_alignment,
3271 .reset_secondary_bus = pnv_pci_reset_secondary_bus,
3272 .dma_set_mask = pnv_pci_ioda_dma_set_mask,
3273 .dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask,
3274 .shutdown = pnv_pci_ioda_shutdown,
3277 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask)
3279 dev_err_once(&npdev->dev,
3280 "%s operation unsupported for NVLink devices\n",
3285 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3286 .dma_dev_setup = pnv_pci_dma_dev_setup,
3287 #ifdef CONFIG_PCI_MSI
3288 .setup_msi_irqs = pnv_setup_msi_irqs,
3289 .teardown_msi_irqs = pnv_teardown_msi_irqs,
3291 .enable_device_hook = pnv_pci_enable_device_hook,
3292 .window_alignment = pnv_pci_window_alignment,
3293 .reset_secondary_bus = pnv_pci_reset_secondary_bus,
3294 .dma_set_mask = pnv_npu_dma_set_mask,
3295 .shutdown = pnv_pci_ioda_shutdown,
3298 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3299 u64 hub_id, int ioda_type)
3301 struct pci_controller *hose;
3302 struct pnv_phb *phb;
3303 unsigned long size, m64map_off, m32map_off, pemap_off;
3304 unsigned long iomap_off = 0, dma32map_off = 0;
3305 const __be64 *prop64;
3306 const __be32 *prop32;
3313 pr_info("Initializing IODA%d OPAL PHB %s\n", ioda_type, np->full_name);
3315 prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3317 pr_err(" Missing \"ibm,opal-phbid\" property !\n");
3320 phb_id = be64_to_cpup(prop64);
3321 pr_debug(" PHB-ID : 0x%016llx\n", phb_id);
3323 phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
3325 /* Allocate PCI controller */
3326 phb->hose = hose = pcibios_alloc_controller(np);
3328 pr_err(" Can't allocate PCI controller for %s\n",
3330 memblock_free(__pa(phb), sizeof(struct pnv_phb));
3334 spin_lock_init(&phb->lock);
3335 prop32 = of_get_property(np, "bus-range", &len);
3336 if (prop32 && len == 8) {
3337 hose->first_busno = be32_to_cpu(prop32[0]);
3338 hose->last_busno = be32_to_cpu(prop32[1]);
3340 pr_warn(" Broken <bus-range> on %s\n", np->full_name);
3341 hose->first_busno = 0;
3342 hose->last_busno = 0xff;
3344 hose->private_data = phb;
3345 phb->hub_id = hub_id;
3346 phb->opal_id = phb_id;
3347 phb->type = ioda_type;
3348 mutex_init(&phb->ioda.pe_alloc_mutex);
3350 /* Detect specific models for error handling */
3351 if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3352 phb->model = PNV_PHB_MODEL_P7IOC;
3353 else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3354 phb->model = PNV_PHB_MODEL_PHB3;
3355 else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3356 phb->model = PNV_PHB_MODEL_NPU;
3358 phb->model = PNV_PHB_MODEL_UNKNOWN;
3360 /* Parse 32-bit and IO ranges (if any) */
3361 pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3364 phb->regs = of_iomap(np, 0);
3365 if (phb->regs == NULL)
3366 pr_err(" Failed to map registers !\n");
3368 /* Initialize more IODA stuff */
3369 phb->ioda.total_pe_num = 1;
3370 prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3372 phb->ioda.total_pe_num = be32_to_cpup(prop32);
3373 prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3375 phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3377 /* Parse 64-bit MMIO range */
3378 pnv_ioda_parse_m64_window(phb);
3380 phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3381 /* FW Has already off top 64k of M32 space (MSI space) */
3382 phb->ioda.m32_size += 0x10000;
3384 phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3385 phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3386 phb->ioda.io_size = hose->pci_io_size;
3387 phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3388 phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3390 /* Calculate how many 32-bit TCE segments we have */
3391 phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3392 PNV_IODA1_DMA32_SEGSIZE;
3394 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3395 size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
3397 size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3399 size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3400 if (phb->type == PNV_PHB_IODA1) {
3402 size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3403 dma32map_off = size;
3404 size += phb->ioda.dma32_count *
3405 sizeof(phb->ioda.dma32_segmap[0]);
3408 size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3409 aux = memblock_virt_alloc(size, 0);
3410 phb->ioda.pe_alloc = aux;
3411 phb->ioda.m64_segmap = aux + m64map_off;
3412 phb->ioda.m32_segmap = aux + m32map_off;
3413 for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3414 phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3415 phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3417 if (phb->type == PNV_PHB_IODA1) {
3418 phb->ioda.io_segmap = aux + iomap_off;
3419 for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3420 phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3422 phb->ioda.dma32_segmap = aux + dma32map_off;
3423 for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3424 phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3426 phb->ioda.pe_array = aux + pemap_off;
3427 set_bit(phb->ioda.reserved_pe_idx, phb->ioda.pe_alloc);
3429 INIT_LIST_HEAD(&phb->ioda.pe_list);
3430 mutex_init(&phb->ioda.pe_list_mutex);
3432 /* Calculate how many 32-bit TCE segments we have */
3433 phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3434 PNV_IODA1_DMA32_SEGSIZE;
3436 #if 0 /* We should really do that ... */
3437 rc = opal_pci_set_phb_mem_window(opal->phb_id,
3440 starting_real_address,
3441 starting_pci_address,
3445 pr_info(" %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3446 phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3447 phb->ioda.m32_size, phb->ioda.m32_segsize);
3448 if (phb->ioda.m64_size)
3449 pr_info(" M64: 0x%lx [segment=0x%lx]\n",
3450 phb->ioda.m64_size, phb->ioda.m64_segsize);
3451 if (phb->ioda.io_size)
3452 pr_info(" IO: 0x%x [segment=0x%x]\n",
3453 phb->ioda.io_size, phb->ioda.io_segsize);
3456 phb->hose->ops = &pnv_pci_ops;
3457 phb->get_pe_state = pnv_ioda_get_pe_state;
3458 phb->freeze_pe = pnv_ioda_freeze_pe;
3459 phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3461 /* Setup MSI support */
3462 pnv_pci_init_ioda_msis(phb);
3465 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3466 * to let the PCI core do resource assignment. It's supposed
3467 * that the PCI core will do correct I/O and MMIO alignment
3468 * for the P2P bridge bars so that each PCI bus (excluding
3469 * the child P2P bridges) can form individual PE.
3471 ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3473 if (phb->type == PNV_PHB_NPU) {
3474 hose->controller_ops = pnv_npu_ioda_controller_ops;
3476 phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
3477 hose->controller_ops = pnv_pci_ioda_controller_ops;
3480 #ifdef CONFIG_PCI_IOV
3481 ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
3482 ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3485 pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3487 /* Reset IODA tables to a clean state */
3488 rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3490 pr_warning(" OPAL Error %ld performing IODA table reset !\n", rc);
3492 /* If we're running in kdump kerenl, the previous kerenl never
3493 * shutdown PCI devices correctly. We already got IODA table
3494 * cleaned out. So we have to issue PHB reset to stop all PCI
3495 * transactions from previous kerenl.
3497 if (is_kdump_kernel()) {
3498 pr_info(" Issue PHB reset ...\n");
3499 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3500 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3503 /* Remove M64 resource if we can't configure it successfully */
3504 if (!phb->init_m64 || phb->init_m64(phb))
3505 hose->mem_resources[1].flags = 0;
3508 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3510 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3513 void __init pnv_pci_init_npu_phb(struct device_node *np)
3515 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU);
3518 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3520 struct device_node *phbn;
3521 const __be64 *prop64;
3524 pr_info("Probing IODA IO-Hub %s\n", np->full_name);
3526 prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3528 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3531 hub_id = be64_to_cpup(prop64);
3532 pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3534 /* Count child PHBs */
3535 for_each_child_of_node(np, phbn) {
3536 /* Look for IODA1 PHBs */
3537 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3538 pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);