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/delay.h>
18 #include <linux/string.h>
19 #include <linux/init.h>
20 #include <linux/memblock.h>
21 #include <linux/irq.h>
23 #include <linux/msi.h>
24 #include <linux/iommu.h>
25 #include <linux/rculist.h>
26 #include <linux/sizes.h>
28 #include <asm/sections.h>
31 #include <asm/pci-bridge.h>
32 #include <asm/machdep.h>
33 #include <asm/msi_bitmap.h>
34 #include <asm/ppc-pci.h>
36 #include <asm/iommu.h>
39 #include <asm/debugfs.h>
40 #include <asm/firmware.h>
41 #include <asm/pnv-pci.h>
42 #include <asm/mmzone.h>
44 #include <misc/cxl-base.h>
48 #include "../../../../drivers/pci/pci.h"
50 #define PNV_IODA1_M64_NUM 16 /* Number of M64 BARs */
51 #define PNV_IODA1_M64_SEGS 8 /* Segments per M64 BAR */
52 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
54 static const char * const pnv_phb_names
[] = { "IODA1", "IODA2", "NPU_NVLINK",
57 void pe_level_printk(const struct pnv_ioda_pe
*pe
, const char *level
,
69 if (pe
->flags
& PNV_IODA_PE_DEV
)
70 strlcpy(pfix
, dev_name(&pe
->pdev
->dev
), sizeof(pfix
));
71 else if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
72 sprintf(pfix
, "%04x:%02x ",
73 pci_domain_nr(pe
->pbus
), pe
->pbus
->number
);
75 else if (pe
->flags
& PNV_IODA_PE_VF
)
76 sprintf(pfix
, "%04x:%02x:%2x.%d",
77 pci_domain_nr(pe
->parent_dev
->bus
),
78 (pe
->rid
& 0xff00) >> 8,
79 PCI_SLOT(pe
->rid
), PCI_FUNC(pe
->rid
));
80 #endif /* CONFIG_PCI_IOV*/
82 printk("%spci %s: [PE# %.2x] %pV",
83 level
, pfix
, pe
->pe_number
, &vaf
);
88 static bool pnv_iommu_bypass_disabled __read_mostly
;
89 static bool pci_reset_phbs __read_mostly
;
91 static int __init
iommu_setup(char *str
)
97 if (!strncmp(str
, "nobypass", 8)) {
98 pnv_iommu_bypass_disabled
= true;
99 pr_info("PowerNV: IOMMU bypass window disabled.\n");
102 str
+= strcspn(str
, ",");
109 early_param("iommu", iommu_setup
);
111 static int __init
pci_reset_phbs_setup(char *str
)
113 pci_reset_phbs
= true;
117 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup
);
119 static inline bool pnv_pci_is_m64(struct pnv_phb
*phb
, struct resource
*r
)
122 * WARNING: We cannot rely on the resource flags. The Linux PCI
123 * allocation code sometimes decides to put a 64-bit prefetchable
124 * BAR in the 32-bit window, so we have to compare the addresses.
126 * For simplicity we only test resource start.
128 return (r
->start
>= phb
->ioda
.m64_base
&&
129 r
->start
< (phb
->ioda
.m64_base
+ phb
->ioda
.m64_size
));
132 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags
)
134 unsigned long flags
= (IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
);
136 return (resource_flags
& flags
) == flags
;
139 static struct pnv_ioda_pe
*pnv_ioda_init_pe(struct pnv_phb
*phb
, int pe_no
)
143 phb
->ioda
.pe_array
[pe_no
].phb
= phb
;
144 phb
->ioda
.pe_array
[pe_no
].pe_number
= pe_no
;
147 * Clear the PE frozen state as it might be put into frozen state
148 * in the last PCI remove path. It's not harmful to do so when the
149 * PE is already in unfrozen state.
151 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
, pe_no
,
152 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
153 if (rc
!= OPAL_SUCCESS
&& rc
!= OPAL_UNSUPPORTED
)
154 pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
155 __func__
, rc
, phb
->hose
->global_number
, pe_no
);
157 return &phb
->ioda
.pe_array
[pe_no
];
160 static void pnv_ioda_reserve_pe(struct pnv_phb
*phb
, int pe_no
)
162 if (!(pe_no
>= 0 && pe_no
< phb
->ioda
.total_pe_num
)) {
163 pr_warn("%s: Invalid PE %x on PHB#%x\n",
164 __func__
, pe_no
, phb
->hose
->global_number
);
168 if (test_and_set_bit(pe_no
, phb
->ioda
.pe_alloc
))
169 pr_debug("%s: PE %x was reserved on PHB#%x\n",
170 __func__
, pe_no
, phb
->hose
->global_number
);
172 pnv_ioda_init_pe(phb
, pe_no
);
175 static struct pnv_ioda_pe
*pnv_ioda_alloc_pe(struct pnv_phb
*phb
)
179 for (pe
= phb
->ioda
.total_pe_num
- 1; pe
>= 0; pe
--) {
180 if (!test_and_set_bit(pe
, phb
->ioda
.pe_alloc
))
181 return pnv_ioda_init_pe(phb
, pe
);
187 static void pnv_ioda_free_pe(struct pnv_ioda_pe
*pe
)
189 struct pnv_phb
*phb
= pe
->phb
;
190 unsigned int pe_num
= pe
->pe_number
;
193 WARN_ON(pe
->npucomp
); /* NPUs are not supposed to be freed */
195 memset(pe
, 0, sizeof(struct pnv_ioda_pe
));
196 clear_bit(pe_num
, phb
->ioda
.pe_alloc
);
199 /* The default M64 BAR is shared by all PEs */
200 static int pnv_ioda2_init_m64(struct pnv_phb
*phb
)
206 /* Configure the default M64 BAR */
207 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
208 OPAL_M64_WINDOW_TYPE
,
209 phb
->ioda
.m64_bar_idx
,
213 if (rc
!= OPAL_SUCCESS
) {
214 desc
= "configuring";
218 /* Enable the default M64 BAR */
219 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
220 OPAL_M64_WINDOW_TYPE
,
221 phb
->ioda
.m64_bar_idx
,
222 OPAL_ENABLE_M64_SPLIT
);
223 if (rc
!= OPAL_SUCCESS
) {
229 * Exclude the segments for reserved and root bus PE, which
230 * are first or last two PEs.
232 r
= &phb
->hose
->mem_resources
[1];
233 if (phb
->ioda
.reserved_pe_idx
== 0)
234 r
->start
+= (2 * phb
->ioda
.m64_segsize
);
235 else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1))
236 r
->end
-= (2 * phb
->ioda
.m64_segsize
);
238 pr_warn(" Cannot strip M64 segment for reserved PE#%x\n",
239 phb
->ioda
.reserved_pe_idx
);
244 pr_warn(" Failure %lld %s M64 BAR#%d\n",
245 rc
, desc
, phb
->ioda
.m64_bar_idx
);
246 opal_pci_phb_mmio_enable(phb
->opal_id
,
247 OPAL_M64_WINDOW_TYPE
,
248 phb
->ioda
.m64_bar_idx
,
253 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev
*pdev
,
254 unsigned long *pe_bitmap
)
256 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
257 struct pnv_phb
*phb
= hose
->private_data
;
259 resource_size_t base
, sgsz
, start
, end
;
262 base
= phb
->ioda
.m64_base
;
263 sgsz
= phb
->ioda
.m64_segsize
;
264 for (i
= 0; i
<= PCI_ROM_RESOURCE
; i
++) {
265 r
= &pdev
->resource
[i
];
266 if (!r
->parent
|| !pnv_pci_is_m64(phb
, r
))
269 start
= _ALIGN_DOWN(r
->start
- base
, sgsz
);
270 end
= _ALIGN_UP(r
->end
- base
, sgsz
);
271 for (segno
= start
/ sgsz
; segno
< end
/ sgsz
; segno
++) {
273 set_bit(segno
, pe_bitmap
);
275 pnv_ioda_reserve_pe(phb
, segno
);
280 static int pnv_ioda1_init_m64(struct pnv_phb
*phb
)
286 * There are 16 M64 BARs, each of which has 8 segments. So
287 * there are as many M64 segments as the maximum number of
290 for (index
= 0; index
< PNV_IODA1_M64_NUM
; index
++) {
291 unsigned long base
, segsz
= phb
->ioda
.m64_segsize
;
294 base
= phb
->ioda
.m64_base
+
295 index
* PNV_IODA1_M64_SEGS
* segsz
;
296 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
297 OPAL_M64_WINDOW_TYPE
, index
, base
, 0,
298 PNV_IODA1_M64_SEGS
* segsz
);
299 if (rc
!= OPAL_SUCCESS
) {
300 pr_warn(" Error %lld setting M64 PHB#%x-BAR#%d\n",
301 rc
, phb
->hose
->global_number
, index
);
305 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
306 OPAL_M64_WINDOW_TYPE
, index
,
307 OPAL_ENABLE_M64_SPLIT
);
308 if (rc
!= OPAL_SUCCESS
) {
309 pr_warn(" Error %lld enabling M64 PHB#%x-BAR#%d\n",
310 rc
, phb
->hose
->global_number
, index
);
316 * Exclude the segments for reserved and root bus PE, which
317 * are first or last two PEs.
319 r
= &phb
->hose
->mem_resources
[1];
320 if (phb
->ioda
.reserved_pe_idx
== 0)
321 r
->start
+= (2 * phb
->ioda
.m64_segsize
);
322 else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1))
323 r
->end
-= (2 * phb
->ioda
.m64_segsize
);
325 WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
326 phb
->ioda
.reserved_pe_idx
, phb
->hose
->global_number
);
331 for ( ; index
>= 0; index
--)
332 opal_pci_phb_mmio_enable(phb
->opal_id
,
333 OPAL_M64_WINDOW_TYPE
, index
, OPAL_DISABLE_M64
);
338 static void pnv_ioda_reserve_m64_pe(struct pci_bus
*bus
,
339 unsigned long *pe_bitmap
,
342 struct pci_dev
*pdev
;
344 list_for_each_entry(pdev
, &bus
->devices
, bus_list
) {
345 pnv_ioda_reserve_dev_m64_pe(pdev
, pe_bitmap
);
347 if (all
&& pdev
->subordinate
)
348 pnv_ioda_reserve_m64_pe(pdev
->subordinate
,
353 static struct pnv_ioda_pe
*pnv_ioda_pick_m64_pe(struct pci_bus
*bus
, bool all
)
355 struct pci_controller
*hose
= pci_bus_to_host(bus
);
356 struct pnv_phb
*phb
= hose
->private_data
;
357 struct pnv_ioda_pe
*master_pe
, *pe
;
358 unsigned long size
, *pe_alloc
;
361 /* Root bus shouldn't use M64 */
362 if (pci_is_root_bus(bus
))
365 /* Allocate bitmap */
366 size
= _ALIGN_UP(phb
->ioda
.total_pe_num
/ 8, sizeof(unsigned long));
367 pe_alloc
= kzalloc(size
, GFP_KERNEL
);
369 pr_warn("%s: Out of memory !\n",
374 /* Figure out reserved PE numbers by the PE */
375 pnv_ioda_reserve_m64_pe(bus
, pe_alloc
, all
);
378 * the current bus might not own M64 window and that's all
379 * contributed by its child buses. For the case, we needn't
380 * pick M64 dependent PE#.
382 if (bitmap_empty(pe_alloc
, phb
->ioda
.total_pe_num
)) {
388 * Figure out the master PE and put all slave PEs to master
389 * PE's list to form compound PE.
393 while ((i
= find_next_bit(pe_alloc
, phb
->ioda
.total_pe_num
, i
+ 1)) <
394 phb
->ioda
.total_pe_num
) {
395 pe
= &phb
->ioda
.pe_array
[i
];
397 phb
->ioda
.m64_segmap
[pe
->pe_number
] = pe
->pe_number
;
399 pe
->flags
|= PNV_IODA_PE_MASTER
;
400 INIT_LIST_HEAD(&pe
->slaves
);
403 pe
->flags
|= PNV_IODA_PE_SLAVE
;
404 pe
->master
= master_pe
;
405 list_add_tail(&pe
->list
, &master_pe
->slaves
);
409 * P7IOC supports M64DT, which helps mapping M64 segment
410 * to one particular PE#. However, PHB3 has fixed mapping
411 * between M64 segment and PE#. In order to have same logic
412 * for P7IOC and PHB3, we enforce fixed mapping between M64
413 * segment and PE# on P7IOC.
415 if (phb
->type
== PNV_PHB_IODA1
) {
418 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
419 pe
->pe_number
, OPAL_M64_WINDOW_TYPE
,
420 pe
->pe_number
/ PNV_IODA1_M64_SEGS
,
421 pe
->pe_number
% PNV_IODA1_M64_SEGS
);
422 if (rc
!= OPAL_SUCCESS
)
423 pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
424 __func__
, rc
, phb
->hose
->global_number
,
433 static void __init
pnv_ioda_parse_m64_window(struct pnv_phb
*phb
)
435 struct pci_controller
*hose
= phb
->hose
;
436 struct device_node
*dn
= hose
->dn
;
437 struct resource
*res
;
442 if (phb
->type
!= PNV_PHB_IODA1
&& phb
->type
!= PNV_PHB_IODA2
) {
443 pr_info(" Not support M64 window\n");
447 if (!firmware_has_feature(FW_FEATURE_OPAL
)) {
448 pr_info(" Firmware too old to support M64 window\n");
452 r
= of_get_property(dn
, "ibm,opal-m64-window", NULL
);
454 pr_info(" No <ibm,opal-m64-window> on %pOF\n",
460 * Find the available M64 BAR range and pickup the last one for
461 * covering the whole 64-bits space. We support only one range.
463 if (of_property_read_u32_array(dn
, "ibm,opal-available-m64-ranges",
465 /* In absence of the property, assume 0..15 */
469 /* We only support 64 bits in our allocator */
470 if (m64_range
[1] > 63) {
471 pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
472 __func__
, m64_range
[1], phb
->hose
->global_number
);
475 /* Empty range, no m64 */
476 if (m64_range
[1] <= m64_range
[0]) {
477 pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
478 __func__
, phb
->hose
->global_number
);
482 /* Configure M64 informations */
483 res
= &hose
->mem_resources
[1];
484 res
->name
= dn
->full_name
;
485 res
->start
= of_translate_address(dn
, r
+ 2);
486 res
->end
= res
->start
+ of_read_number(r
+ 4, 2) - 1;
487 res
->flags
= (IORESOURCE_MEM
| IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
);
488 pci_addr
= of_read_number(r
, 2);
489 hose
->mem_offset
[1] = res
->start
- pci_addr
;
491 phb
->ioda
.m64_size
= resource_size(res
);
492 phb
->ioda
.m64_segsize
= phb
->ioda
.m64_size
/ phb
->ioda
.total_pe_num
;
493 phb
->ioda
.m64_base
= pci_addr
;
495 /* This lines up nicely with the display from processing OF ranges */
496 pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
497 res
->start
, res
->end
, pci_addr
, m64_range
[0],
498 m64_range
[0] + m64_range
[1] - 1);
500 /* Mark all M64 used up by default */
501 phb
->ioda
.m64_bar_alloc
= (unsigned long)-1;
503 /* Use last M64 BAR to cover M64 window */
505 phb
->ioda
.m64_bar_idx
= m64_range
[0] + m64_range
[1];
507 pr_info(" Using M64 #%d as default window\n", phb
->ioda
.m64_bar_idx
);
509 /* Mark remaining ones free */
510 for (i
= m64_range
[0]; i
< m64_range
[1]; i
++)
511 clear_bit(i
, &phb
->ioda
.m64_bar_alloc
);
514 * Setup init functions for M64 based on IODA version, IODA3 uses
517 if (phb
->type
== PNV_PHB_IODA1
)
518 phb
->init_m64
= pnv_ioda1_init_m64
;
520 phb
->init_m64
= pnv_ioda2_init_m64
;
523 static void pnv_ioda_freeze_pe(struct pnv_phb
*phb
, int pe_no
)
525 struct pnv_ioda_pe
*pe
= &phb
->ioda
.pe_array
[pe_no
];
526 struct pnv_ioda_pe
*slave
;
529 /* Fetch master PE */
530 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
532 if (WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
)))
535 pe_no
= pe
->pe_number
;
538 /* Freeze master PE */
539 rc
= opal_pci_eeh_freeze_set(phb
->opal_id
,
541 OPAL_EEH_ACTION_SET_FREEZE_ALL
);
542 if (rc
!= OPAL_SUCCESS
) {
543 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
544 __func__
, rc
, phb
->hose
->global_number
, pe_no
);
548 /* Freeze slave PEs */
549 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
552 list_for_each_entry(slave
, &pe
->slaves
, list
) {
553 rc
= opal_pci_eeh_freeze_set(phb
->opal_id
,
555 OPAL_EEH_ACTION_SET_FREEZE_ALL
);
556 if (rc
!= OPAL_SUCCESS
)
557 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
558 __func__
, rc
, phb
->hose
->global_number
,
563 static int pnv_ioda_unfreeze_pe(struct pnv_phb
*phb
, int pe_no
, int opt
)
565 struct pnv_ioda_pe
*pe
, *slave
;
569 pe
= &phb
->ioda
.pe_array
[pe_no
];
570 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
572 WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
));
573 pe_no
= pe
->pe_number
;
576 /* Clear frozen state for master PE */
577 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
, pe_no
, opt
);
578 if (rc
!= OPAL_SUCCESS
) {
579 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
580 __func__
, rc
, opt
, phb
->hose
->global_number
, pe_no
);
584 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
587 /* Clear frozen state for slave PEs */
588 list_for_each_entry(slave
, &pe
->slaves
, list
) {
589 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
,
592 if (rc
!= OPAL_SUCCESS
) {
593 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
594 __func__
, rc
, opt
, phb
->hose
->global_number
,
603 static int pnv_ioda_get_pe_state(struct pnv_phb
*phb
, int pe_no
)
605 struct pnv_ioda_pe
*slave
, *pe
;
606 u8 fstate
= 0, state
;
610 /* Sanity check on PE number */
611 if (pe_no
< 0 || pe_no
>= phb
->ioda
.total_pe_num
)
612 return OPAL_EEH_STOPPED_PERM_UNAVAIL
;
615 * Fetch the master PE and the PE instance might be
616 * not initialized yet.
618 pe
= &phb
->ioda
.pe_array
[pe_no
];
619 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
621 WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
));
622 pe_no
= pe
->pe_number
;
625 /* Check the master PE */
626 rc
= opal_pci_eeh_freeze_status(phb
->opal_id
, pe_no
,
627 &state
, &pcierr
, NULL
);
628 if (rc
!= OPAL_SUCCESS
) {
629 pr_warn("%s: Failure %lld getting "
630 "PHB#%x-PE#%x state\n",
632 phb
->hose
->global_number
, pe_no
);
633 return OPAL_EEH_STOPPED_TEMP_UNAVAIL
;
636 /* Check the slave PE */
637 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
640 list_for_each_entry(slave
, &pe
->slaves
, list
) {
641 rc
= opal_pci_eeh_freeze_status(phb
->opal_id
,
646 if (rc
!= OPAL_SUCCESS
) {
647 pr_warn("%s: Failure %lld getting "
648 "PHB#%x-PE#%x state\n",
650 phb
->hose
->global_number
, slave
->pe_number
);
651 return OPAL_EEH_STOPPED_TEMP_UNAVAIL
;
655 * Override the result based on the ascending
665 struct pnv_ioda_pe
*pnv_ioda_get_pe(struct pci_dev
*dev
)
667 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
668 struct pnv_phb
*phb
= hose
->private_data
;
669 struct pci_dn
*pdn
= pci_get_pdn(dev
);
673 if (pdn
->pe_number
== IODA_INVALID_PE
)
675 return &phb
->ioda
.pe_array
[pdn
->pe_number
];
678 static int pnv_ioda_set_one_peltv(struct pnv_phb
*phb
,
679 struct pnv_ioda_pe
*parent
,
680 struct pnv_ioda_pe
*child
,
683 const char *desc
= is_add
? "adding" : "removing";
684 uint8_t op
= is_add
? OPAL_ADD_PE_TO_DOMAIN
:
685 OPAL_REMOVE_PE_FROM_DOMAIN
;
686 struct pnv_ioda_pe
*slave
;
689 /* Parent PE affects child PE */
690 rc
= opal_pci_set_peltv(phb
->opal_id
, parent
->pe_number
,
691 child
->pe_number
, op
);
692 if (rc
!= OPAL_SUCCESS
) {
693 pe_warn(child
, "OPAL error %ld %s to parent PELTV\n",
698 if (!(child
->flags
& PNV_IODA_PE_MASTER
))
701 /* Compound case: parent PE affects slave PEs */
702 list_for_each_entry(slave
, &child
->slaves
, list
) {
703 rc
= opal_pci_set_peltv(phb
->opal_id
, parent
->pe_number
,
704 slave
->pe_number
, op
);
705 if (rc
!= OPAL_SUCCESS
) {
706 pe_warn(slave
, "OPAL error %ld %s to parent PELTV\n",
715 static int pnv_ioda_set_peltv(struct pnv_phb
*phb
,
716 struct pnv_ioda_pe
*pe
,
719 struct pnv_ioda_pe
*slave
;
720 struct pci_dev
*pdev
= NULL
;
724 * Clear PE frozen state. If it's master PE, we need
725 * clear slave PE frozen state as well.
728 opal_pci_eeh_freeze_clear(phb
->opal_id
, pe
->pe_number
,
729 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
730 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
731 list_for_each_entry(slave
, &pe
->slaves
, list
)
732 opal_pci_eeh_freeze_clear(phb
->opal_id
,
734 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
739 * Associate PE in PELT. We need add the PE into the
740 * corresponding PELT-V as well. Otherwise, the error
741 * originated from the PE might contribute to other
744 ret
= pnv_ioda_set_one_peltv(phb
, pe
, pe
, is_add
);
748 /* For compound PEs, any one affects all of them */
749 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
750 list_for_each_entry(slave
, &pe
->slaves
, list
) {
751 ret
= pnv_ioda_set_one_peltv(phb
, slave
, pe
, is_add
);
757 if (pe
->flags
& (PNV_IODA_PE_BUS_ALL
| PNV_IODA_PE_BUS
))
758 pdev
= pe
->pbus
->self
;
759 else if (pe
->flags
& PNV_IODA_PE_DEV
)
760 pdev
= pe
->pdev
->bus
->self
;
761 #ifdef CONFIG_PCI_IOV
762 else if (pe
->flags
& PNV_IODA_PE_VF
)
763 pdev
= pe
->parent_dev
;
764 #endif /* CONFIG_PCI_IOV */
766 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
767 struct pnv_ioda_pe
*parent
;
769 if (pdn
&& pdn
->pe_number
!= IODA_INVALID_PE
) {
770 parent
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
771 ret
= pnv_ioda_set_one_peltv(phb
, parent
, pe
, is_add
);
776 pdev
= pdev
->bus
->self
;
782 static int pnv_ioda_deconfigure_pe(struct pnv_phb
*phb
, struct pnv_ioda_pe
*pe
)
784 struct pci_dev
*parent
;
785 uint8_t bcomp
, dcomp
, fcomp
;
789 /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
793 dcomp
= OPAL_IGNORE_RID_DEVICE_NUMBER
;
794 fcomp
= OPAL_IGNORE_RID_FUNCTION_NUMBER
;
795 parent
= pe
->pbus
->self
;
796 if (pe
->flags
& PNV_IODA_PE_BUS_ALL
)
797 count
= pe
->pbus
->busn_res
.end
- pe
->pbus
->busn_res
.start
+ 1;
802 case 1: bcomp
= OpalPciBusAll
; break;
803 case 2: bcomp
= OpalPciBus7Bits
; break;
804 case 4: bcomp
= OpalPciBus6Bits
; break;
805 case 8: bcomp
= OpalPciBus5Bits
; break;
806 case 16: bcomp
= OpalPciBus4Bits
; break;
807 case 32: bcomp
= OpalPciBus3Bits
; break;
809 dev_err(&pe
->pbus
->dev
, "Number of subordinate buses %d unsupported\n",
811 /* Do an exact match only */
812 bcomp
= OpalPciBusAll
;
814 rid_end
= pe
->rid
+ (count
<< 8);
816 #ifdef CONFIG_PCI_IOV
817 if (pe
->flags
& PNV_IODA_PE_VF
)
818 parent
= pe
->parent_dev
;
821 parent
= pe
->pdev
->bus
->self
;
822 bcomp
= OpalPciBusAll
;
823 dcomp
= OPAL_COMPARE_RID_DEVICE_NUMBER
;
824 fcomp
= OPAL_COMPARE_RID_FUNCTION_NUMBER
;
825 rid_end
= pe
->rid
+ 1;
828 /* Clear the reverse map */
829 for (rid
= pe
->rid
; rid
< rid_end
; rid
++)
830 phb
->ioda
.pe_rmap
[rid
] = IODA_INVALID_PE
;
832 /* Release from all parents PELT-V */
834 struct pci_dn
*pdn
= pci_get_pdn(parent
);
835 if (pdn
&& pdn
->pe_number
!= IODA_INVALID_PE
) {
836 rc
= opal_pci_set_peltv(phb
->opal_id
, pdn
->pe_number
,
837 pe
->pe_number
, OPAL_REMOVE_PE_FROM_DOMAIN
);
838 /* XXX What to do in case of error ? */
840 parent
= parent
->bus
->self
;
843 opal_pci_eeh_freeze_clear(phb
->opal_id
, pe
->pe_number
,
844 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
846 /* Disassociate PE in PELT */
847 rc
= opal_pci_set_peltv(phb
->opal_id
, pe
->pe_number
,
848 pe
->pe_number
, OPAL_REMOVE_PE_FROM_DOMAIN
);
850 pe_warn(pe
, "OPAL error %ld remove self from PELTV\n", rc
);
851 rc
= opal_pci_set_pe(phb
->opal_id
, pe
->pe_number
, pe
->rid
,
852 bcomp
, dcomp
, fcomp
, OPAL_UNMAP_PE
);
854 pe_err(pe
, "OPAL error %ld trying to setup PELT table\n", rc
);
858 #ifdef CONFIG_PCI_IOV
859 pe
->parent_dev
= NULL
;
865 static int pnv_ioda_configure_pe(struct pnv_phb
*phb
, struct pnv_ioda_pe
*pe
)
867 struct pci_dev
*parent
;
868 uint8_t bcomp
, dcomp
, fcomp
;
869 long rc
, rid_end
, rid
;
871 /* Bus validation ? */
875 dcomp
= OPAL_IGNORE_RID_DEVICE_NUMBER
;
876 fcomp
= OPAL_IGNORE_RID_FUNCTION_NUMBER
;
877 parent
= pe
->pbus
->self
;
878 if (pe
->flags
& PNV_IODA_PE_BUS_ALL
)
879 count
= pe
->pbus
->busn_res
.end
- pe
->pbus
->busn_res
.start
+ 1;
884 case 1: bcomp
= OpalPciBusAll
; break;
885 case 2: bcomp
= OpalPciBus7Bits
; break;
886 case 4: bcomp
= OpalPciBus6Bits
; break;
887 case 8: bcomp
= OpalPciBus5Bits
; break;
888 case 16: bcomp
= OpalPciBus4Bits
; break;
889 case 32: bcomp
= OpalPciBus3Bits
; break;
891 dev_err(&pe
->pbus
->dev
, "Number of subordinate buses %d unsupported\n",
893 /* Do an exact match only */
894 bcomp
= OpalPciBusAll
;
896 rid_end
= pe
->rid
+ (count
<< 8);
898 #ifdef CONFIG_PCI_IOV
899 if (pe
->flags
& PNV_IODA_PE_VF
)
900 parent
= pe
->parent_dev
;
902 #endif /* CONFIG_PCI_IOV */
903 parent
= pe
->pdev
->bus
->self
;
904 bcomp
= OpalPciBusAll
;
905 dcomp
= OPAL_COMPARE_RID_DEVICE_NUMBER
;
906 fcomp
= OPAL_COMPARE_RID_FUNCTION_NUMBER
;
907 rid_end
= pe
->rid
+ 1;
911 * Associate PE in PELT. We need add the PE into the
912 * corresponding PELT-V as well. Otherwise, the error
913 * originated from the PE might contribute to other
916 rc
= opal_pci_set_pe(phb
->opal_id
, pe
->pe_number
, pe
->rid
,
917 bcomp
, dcomp
, fcomp
, OPAL_MAP_PE
);
919 pe_err(pe
, "OPAL error %ld trying to setup PELT table\n", rc
);
924 * Configure PELTV. NPUs don't have a PELTV table so skip
925 * configuration on them.
927 if (phb
->type
!= PNV_PHB_NPU_NVLINK
&& phb
->type
!= PNV_PHB_NPU_OCAPI
)
928 pnv_ioda_set_peltv(phb
, pe
, true);
930 /* Setup reverse map */
931 for (rid
= pe
->rid
; rid
< rid_end
; rid
++)
932 phb
->ioda
.pe_rmap
[rid
] = pe
->pe_number
;
934 /* Setup one MVTs on IODA1 */
935 if (phb
->type
!= PNV_PHB_IODA1
) {
940 pe
->mve_number
= pe
->pe_number
;
941 rc
= opal_pci_set_mve(phb
->opal_id
, pe
->mve_number
, pe
->pe_number
);
942 if (rc
!= OPAL_SUCCESS
) {
943 pe_err(pe
, "OPAL error %ld setting up MVE %x\n",
947 rc
= opal_pci_set_mve_enable(phb
->opal_id
,
948 pe
->mve_number
, OPAL_ENABLE_MVE
);
950 pe_err(pe
, "OPAL error %ld enabling MVE %x\n",
960 #ifdef CONFIG_PCI_IOV
961 static int pnv_pci_vf_resource_shift(struct pci_dev
*dev
, int offset
)
963 struct pci_dn
*pdn
= pci_get_pdn(dev
);
965 struct resource
*res
, res2
;
966 resource_size_t size
;
973 * "offset" is in VFs. The M64 windows are sized so that when they
974 * are segmented, each segment is the same size as the IOV BAR.
975 * Each segment is in a separate PE, and the high order bits of the
976 * address are the PE number. Therefore, each VF's BAR is in a
977 * separate PE, and changing the IOV BAR start address changes the
978 * range of PEs the VFs are in.
980 num_vfs
= pdn
->num_vfs
;
981 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
982 res
= &dev
->resource
[i
+ PCI_IOV_RESOURCES
];
983 if (!res
->flags
|| !res
->parent
)
987 * The actual IOV BAR range is determined by the start address
988 * and the actual size for num_vfs VFs BAR. This check is to
989 * make sure that after shifting, the range will not overlap
990 * with another device.
992 size
= pci_iov_resource_size(dev
, i
+ PCI_IOV_RESOURCES
);
993 res2
.flags
= res
->flags
;
994 res2
.start
= res
->start
+ (size
* offset
);
995 res2
.end
= res2
.start
+ (size
* num_vfs
) - 1;
997 if (res2
.end
> res
->end
) {
998 dev_err(&dev
->dev
, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
999 i
, &res2
, res
, num_vfs
, offset
);
1005 * Since M64 BAR shares segments among all possible 256 PEs,
1006 * we have to shift the beginning of PF IOV BAR to make it start from
1007 * the segment which belongs to the PE number assigned to the first VF.
1008 * This creates a "hole" in the /proc/iomem which could be used for
1009 * allocating other resources so we reserve this area below and
1010 * release when IOV is released.
1012 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
1013 res
= &dev
->resource
[i
+ PCI_IOV_RESOURCES
];
1014 if (!res
->flags
|| !res
->parent
)
1017 size
= pci_iov_resource_size(dev
, i
+ PCI_IOV_RESOURCES
);
1019 res
->start
+= size
* offset
;
1021 dev_info(&dev
->dev
, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
1022 i
, &res2
, res
, (offset
> 0) ? "En" : "Dis",
1026 devm_release_resource(&dev
->dev
, &pdn
->holes
[i
]);
1027 memset(&pdn
->holes
[i
], 0, sizeof(pdn
->holes
[i
]));
1030 pci_update_resource(dev
, i
+ PCI_IOV_RESOURCES
);
1033 pdn
->holes
[i
].start
= res2
.start
;
1034 pdn
->holes
[i
].end
= res2
.start
+ size
* offset
- 1;
1035 pdn
->holes
[i
].flags
= IORESOURCE_BUS
;
1036 pdn
->holes
[i
].name
= "pnv_iov_reserved";
1037 devm_request_resource(&dev
->dev
, res
->parent
,
1043 #endif /* CONFIG_PCI_IOV */
1045 static struct pnv_ioda_pe
*pnv_ioda_setup_dev_PE(struct pci_dev
*dev
)
1047 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
1048 struct pnv_phb
*phb
= hose
->private_data
;
1049 struct pci_dn
*pdn
= pci_get_pdn(dev
);
1050 struct pnv_ioda_pe
*pe
;
1053 pr_err("%s: Device tree node not associated properly\n",
1057 if (pdn
->pe_number
!= IODA_INVALID_PE
)
1060 pe
= pnv_ioda_alloc_pe(phb
);
1062 pr_warn("%s: Not enough PE# available, disabling device\n",
1067 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1068 * pointer in the PE data structure, both should be destroyed at the
1069 * same time. However, this needs to be looked at more closely again
1070 * once we actually start removing things (Hotplug, SR-IOV, ...)
1072 * At some point we want to remove the PDN completely anyways
1075 pdn
->pe_number
= pe
->pe_number
;
1076 pe
->flags
= PNV_IODA_PE_DEV
;
1079 pe
->mve_number
= -1;
1080 pe
->rid
= dev
->bus
->number
<< 8 | pdn
->devfn
;
1082 pe_info(pe
, "Associated device to PE\n");
1084 if (pnv_ioda_configure_pe(phb
, pe
)) {
1085 /* XXX What do we do here ? */
1086 pnv_ioda_free_pe(pe
);
1087 pdn
->pe_number
= IODA_INVALID_PE
;
1093 /* Put PE to the list */
1094 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1099 static void pnv_ioda_setup_same_PE(struct pci_bus
*bus
, struct pnv_ioda_pe
*pe
)
1101 struct pci_dev
*dev
;
1103 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
1104 struct pci_dn
*pdn
= pci_get_pdn(dev
);
1107 pr_warn("%s: No device node associated with device !\n",
1113 * In partial hotplug case, the PCI device might be still
1114 * associated with the PE and needn't attach it to the PE
1117 if (pdn
->pe_number
!= IODA_INVALID_PE
)
1121 pdn
->pe_number
= pe
->pe_number
;
1122 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
1123 pnv_ioda_setup_same_PE(dev
->subordinate
, pe
);
1128 * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1129 * single PCI bus. Another one that contains the primary PCI bus and its
1130 * subordinate PCI devices and buses. The second type of PE is normally
1131 * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1133 static struct pnv_ioda_pe
*pnv_ioda_setup_bus_PE(struct pci_bus
*bus
, bool all
)
1135 struct pci_controller
*hose
= pci_bus_to_host(bus
);
1136 struct pnv_phb
*phb
= hose
->private_data
;
1137 struct pnv_ioda_pe
*pe
= NULL
;
1138 unsigned int pe_num
;
1141 * In partial hotplug case, the PE instance might be still alive.
1142 * We should reuse it instead of allocating a new one.
1144 pe_num
= phb
->ioda
.pe_rmap
[bus
->number
<< 8];
1145 if (pe_num
!= IODA_INVALID_PE
) {
1146 pe
= &phb
->ioda
.pe_array
[pe_num
];
1147 pnv_ioda_setup_same_PE(bus
, pe
);
1151 /* PE number for root bus should have been reserved */
1152 if (pci_is_root_bus(bus
) &&
1153 phb
->ioda
.root_pe_idx
!= IODA_INVALID_PE
)
1154 pe
= &phb
->ioda
.pe_array
[phb
->ioda
.root_pe_idx
];
1156 /* Check if PE is determined by M64 */
1158 pe
= pnv_ioda_pick_m64_pe(bus
, all
);
1160 /* The PE number isn't pinned by M64 */
1162 pe
= pnv_ioda_alloc_pe(phb
);
1165 pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1166 __func__
, pci_domain_nr(bus
), bus
->number
);
1170 pe
->flags
|= (all
? PNV_IODA_PE_BUS_ALL
: PNV_IODA_PE_BUS
);
1173 pe
->mve_number
= -1;
1174 pe
->rid
= bus
->busn_res
.start
<< 8;
1177 pe_info(pe
, "Secondary bus %d..%d associated with PE#%x\n",
1178 bus
->busn_res
.start
, bus
->busn_res
.end
, pe
->pe_number
);
1180 pe_info(pe
, "Secondary bus %d associated with PE#%x\n",
1181 bus
->busn_res
.start
, pe
->pe_number
);
1183 if (pnv_ioda_configure_pe(phb
, pe
)) {
1184 /* XXX What do we do here ? */
1185 pnv_ioda_free_pe(pe
);
1190 /* Associate it with all child devices */
1191 pnv_ioda_setup_same_PE(bus
, pe
);
1193 /* Put PE to the list */
1194 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1199 static struct pnv_ioda_pe
*pnv_ioda_setup_npu_PE(struct pci_dev
*npu_pdev
)
1201 int pe_num
, found_pe
= false, rc
;
1203 struct pnv_ioda_pe
*pe
;
1204 struct pci_dev
*gpu_pdev
;
1205 struct pci_dn
*npu_pdn
;
1206 struct pci_controller
*hose
= pci_bus_to_host(npu_pdev
->bus
);
1207 struct pnv_phb
*phb
= hose
->private_data
;
1210 * Due to a hardware errata PE#0 on the NPU is reserved for
1211 * error handling. This means we only have three PEs remaining
1212 * which need to be assigned to four links, implying some
1213 * links must share PEs.
1215 * To achieve this we assign PEs such that NPUs linking the
1216 * same GPU get assigned the same PE.
1218 gpu_pdev
= pnv_pci_get_gpu_dev(npu_pdev
);
1219 for (pe_num
= 0; pe_num
< phb
->ioda
.total_pe_num
; pe_num
++) {
1220 pe
= &phb
->ioda
.pe_array
[pe_num
];
1224 if (pnv_pci_get_gpu_dev(pe
->pdev
) == gpu_pdev
) {
1226 * This device has the same peer GPU so should
1227 * be assigned the same PE as the existing
1230 dev_info(&npu_pdev
->dev
,
1231 "Associating to existing PE %x\n", pe_num
);
1232 pci_dev_get(npu_pdev
);
1233 npu_pdn
= pci_get_pdn(npu_pdev
);
1234 rid
= npu_pdev
->bus
->number
<< 8 | npu_pdn
->devfn
;
1235 npu_pdn
->pe_number
= pe_num
;
1236 phb
->ioda
.pe_rmap
[rid
] = pe
->pe_number
;
1238 /* Map the PE to this link */
1239 rc
= opal_pci_set_pe(phb
->opal_id
, pe_num
, rid
,
1241 OPAL_COMPARE_RID_DEVICE_NUMBER
,
1242 OPAL_COMPARE_RID_FUNCTION_NUMBER
,
1244 WARN_ON(rc
!= OPAL_SUCCESS
);
1252 * Could not find an existing PE so allocate a new
1255 return pnv_ioda_setup_dev_PE(npu_pdev
);
1260 static void pnv_ioda_setup_npu_PEs(struct pci_bus
*bus
)
1262 struct pci_dev
*pdev
;
1264 list_for_each_entry(pdev
, &bus
->devices
, bus_list
)
1265 pnv_ioda_setup_npu_PE(pdev
);
1268 static void pnv_pci_ioda_setup_PEs(void)
1270 struct pci_controller
*hose
;
1271 struct pnv_phb
*phb
;
1272 struct pci_bus
*bus
;
1273 struct pci_dev
*pdev
;
1274 struct pnv_ioda_pe
*pe
;
1276 list_for_each_entry(hose
, &hose_list
, list_node
) {
1277 phb
= hose
->private_data
;
1278 if (phb
->type
== PNV_PHB_NPU_NVLINK
) {
1279 /* PE#0 is needed for error reporting */
1280 pnv_ioda_reserve_pe(phb
, 0);
1281 pnv_ioda_setup_npu_PEs(hose
->bus
);
1282 if (phb
->model
== PNV_PHB_MODEL_NPU2
)
1283 WARN_ON_ONCE(pnv_npu2_init(hose
));
1285 if (phb
->type
== PNV_PHB_NPU_OCAPI
) {
1287 list_for_each_entry(pdev
, &bus
->devices
, bus_list
)
1288 pnv_ioda_setup_dev_PE(pdev
);
1291 list_for_each_entry(hose
, &hose_list
, list_node
) {
1292 phb
= hose
->private_data
;
1293 if (phb
->type
!= PNV_PHB_IODA2
)
1296 list_for_each_entry(pe
, &phb
->ioda
.pe_list
, list
)
1297 pnv_npu2_map_lpar(pe
, MSR_DR
| MSR_PR
| MSR_HV
);
1301 #ifdef CONFIG_PCI_IOV
1302 static int pnv_pci_vf_release_m64(struct pci_dev
*pdev
, u16 num_vfs
)
1304 struct pci_bus
*bus
;
1305 struct pci_controller
*hose
;
1306 struct pnv_phb
*phb
;
1312 hose
= pci_bus_to_host(bus
);
1313 phb
= hose
->private_data
;
1314 pdn
= pci_get_pdn(pdev
);
1316 if (pdn
->m64_single_mode
)
1321 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++)
1322 for (j
= 0; j
< m64_bars
; j
++) {
1323 if (pdn
->m64_map
[j
][i
] == IODA_INVALID_M64
)
1325 opal_pci_phb_mmio_enable(phb
->opal_id
,
1326 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 0);
1327 clear_bit(pdn
->m64_map
[j
][i
], &phb
->ioda
.m64_bar_alloc
);
1328 pdn
->m64_map
[j
][i
] = IODA_INVALID_M64
;
1331 kfree(pdn
->m64_map
);
1335 static int pnv_pci_vf_assign_m64(struct pci_dev
*pdev
, u16 num_vfs
)
1337 struct pci_bus
*bus
;
1338 struct pci_controller
*hose
;
1339 struct pnv_phb
*phb
;
1342 struct resource
*res
;
1346 resource_size_t size
, start
;
1351 hose
= pci_bus_to_host(bus
);
1352 phb
= hose
->private_data
;
1353 pdn
= pci_get_pdn(pdev
);
1354 total_vfs
= pci_sriov_get_totalvfs(pdev
);
1356 if (pdn
->m64_single_mode
)
1361 pdn
->m64_map
= kmalloc_array(m64_bars
,
1362 sizeof(*pdn
->m64_map
),
1366 /* Initialize the m64_map to IODA_INVALID_M64 */
1367 for (i
= 0; i
< m64_bars
; i
++)
1368 for (j
= 0; j
< PCI_SRIOV_NUM_BARS
; j
++)
1369 pdn
->m64_map
[i
][j
] = IODA_INVALID_M64
;
1372 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
1373 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
1374 if (!res
->flags
|| !res
->parent
)
1377 for (j
= 0; j
< m64_bars
; j
++) {
1379 win
= find_next_zero_bit(&phb
->ioda
.m64_bar_alloc
,
1380 phb
->ioda
.m64_bar_idx
+ 1, 0);
1382 if (win
>= phb
->ioda
.m64_bar_idx
+ 1)
1384 } while (test_and_set_bit(win
, &phb
->ioda
.m64_bar_alloc
));
1386 pdn
->m64_map
[j
][i
] = win
;
1388 if (pdn
->m64_single_mode
) {
1389 size
= pci_iov_resource_size(pdev
,
1390 PCI_IOV_RESOURCES
+ i
);
1391 start
= res
->start
+ size
* j
;
1393 size
= resource_size(res
);
1397 /* Map the M64 here */
1398 if (pdn
->m64_single_mode
) {
1399 pe_num
= pdn
->pe_num_map
[j
];
1400 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
1401 pe_num
, OPAL_M64_WINDOW_TYPE
,
1402 pdn
->m64_map
[j
][i
], 0);
1405 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
1406 OPAL_M64_WINDOW_TYPE
,
1413 if (rc
!= OPAL_SUCCESS
) {
1414 dev_err(&pdev
->dev
, "Failed to map M64 window #%d: %lld\n",
1419 if (pdn
->m64_single_mode
)
1420 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
1421 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 2);
1423 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
1424 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 1);
1426 if (rc
!= OPAL_SUCCESS
) {
1427 dev_err(&pdev
->dev
, "Failed to enable M64 window #%d: %llx\n",
1436 pnv_pci_vf_release_m64(pdev
, num_vfs
);
1440 static long pnv_pci_ioda2_unset_window(struct iommu_table_group
*table_group
,
1443 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev
*dev
, struct pnv_ioda_pe
*pe
)
1445 struct iommu_table
*tbl
;
1448 tbl
= pe
->table_group
.tables
[0];
1449 rc
= pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
1451 pe_warn(pe
, "OPAL error %ld release DMA window\n", rc
);
1453 pnv_pci_ioda2_set_bypass(pe
, false);
1454 if (pe
->table_group
.group
) {
1455 iommu_group_put(pe
->table_group
.group
);
1456 BUG_ON(pe
->table_group
.group
);
1458 iommu_tce_table_put(tbl
);
1461 static void pnv_ioda_release_vf_PE(struct pci_dev
*pdev
)
1463 struct pci_bus
*bus
;
1464 struct pci_controller
*hose
;
1465 struct pnv_phb
*phb
;
1466 struct pnv_ioda_pe
*pe
, *pe_n
;
1470 hose
= pci_bus_to_host(bus
);
1471 phb
= hose
->private_data
;
1472 pdn
= pci_get_pdn(pdev
);
1474 if (!pdev
->is_physfn
)
1477 list_for_each_entry_safe(pe
, pe_n
, &phb
->ioda
.pe_list
, list
) {
1478 if (pe
->parent_dev
!= pdev
)
1481 pnv_pci_ioda2_release_dma_pe(pdev
, pe
);
1483 /* Remove from list */
1484 mutex_lock(&phb
->ioda
.pe_list_mutex
);
1485 list_del(&pe
->list
);
1486 mutex_unlock(&phb
->ioda
.pe_list_mutex
);
1488 pnv_ioda_deconfigure_pe(phb
, pe
);
1490 pnv_ioda_free_pe(pe
);
1494 void pnv_pci_sriov_disable(struct pci_dev
*pdev
)
1496 struct pci_bus
*bus
;
1497 struct pci_controller
*hose
;
1498 struct pnv_phb
*phb
;
1499 struct pnv_ioda_pe
*pe
;
1504 hose
= pci_bus_to_host(bus
);
1505 phb
= hose
->private_data
;
1506 pdn
= pci_get_pdn(pdev
);
1507 num_vfs
= pdn
->num_vfs
;
1509 /* Release VF PEs */
1510 pnv_ioda_release_vf_PE(pdev
);
1512 if (phb
->type
== PNV_PHB_IODA2
) {
1513 if (!pdn
->m64_single_mode
)
1514 pnv_pci_vf_resource_shift(pdev
, -*pdn
->pe_num_map
);
1516 /* Release M64 windows */
1517 pnv_pci_vf_release_m64(pdev
, num_vfs
);
1519 /* Release PE numbers */
1520 if (pdn
->m64_single_mode
) {
1521 for (i
= 0; i
< num_vfs
; i
++) {
1522 if (pdn
->pe_num_map
[i
] == IODA_INVALID_PE
)
1525 pe
= &phb
->ioda
.pe_array
[pdn
->pe_num_map
[i
]];
1526 pnv_ioda_free_pe(pe
);
1529 bitmap_clear(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1530 /* Releasing pe_num_map */
1531 kfree(pdn
->pe_num_map
);
1535 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb
*phb
,
1536 struct pnv_ioda_pe
*pe
);
1537 #ifdef CONFIG_IOMMU_API
1538 static void pnv_ioda_setup_bus_iommu_group(struct pnv_ioda_pe
*pe
,
1539 struct iommu_table_group
*table_group
, struct pci_bus
*bus
);
1542 static void pnv_ioda_setup_vf_PE(struct pci_dev
*pdev
, u16 num_vfs
)
1544 struct pci_bus
*bus
;
1545 struct pci_controller
*hose
;
1546 struct pnv_phb
*phb
;
1547 struct pnv_ioda_pe
*pe
;
1553 hose
= pci_bus_to_host(bus
);
1554 phb
= hose
->private_data
;
1555 pdn
= pci_get_pdn(pdev
);
1557 if (!pdev
->is_physfn
)
1560 /* Reserve PE for each VF */
1561 for (vf_index
= 0; vf_index
< num_vfs
; vf_index
++) {
1562 if (pdn
->m64_single_mode
)
1563 pe_num
= pdn
->pe_num_map
[vf_index
];
1565 pe_num
= *pdn
->pe_num_map
+ vf_index
;
1567 pe
= &phb
->ioda
.pe_array
[pe_num
];
1568 pe
->pe_number
= pe_num
;
1570 pe
->flags
= PNV_IODA_PE_VF
;
1572 pe
->parent_dev
= pdev
;
1573 pe
->mve_number
= -1;
1574 pe
->rid
= (pci_iov_virtfn_bus(pdev
, vf_index
) << 8) |
1575 pci_iov_virtfn_devfn(pdev
, vf_index
);
1577 pe_info(pe
, "VF %04d:%02d:%02d.%d associated with PE#%x\n",
1578 hose
->global_number
, pdev
->bus
->number
,
1579 PCI_SLOT(pci_iov_virtfn_devfn(pdev
, vf_index
)),
1580 PCI_FUNC(pci_iov_virtfn_devfn(pdev
, vf_index
)), pe_num
);
1582 if (pnv_ioda_configure_pe(phb
, pe
)) {
1583 /* XXX What do we do here ? */
1584 pnv_ioda_free_pe(pe
);
1589 /* Put PE to the list */
1590 mutex_lock(&phb
->ioda
.pe_list_mutex
);
1591 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1592 mutex_unlock(&phb
->ioda
.pe_list_mutex
);
1594 pnv_pci_ioda2_setup_dma_pe(phb
, pe
);
1595 #ifdef CONFIG_IOMMU_API
1596 iommu_register_group(&pe
->table_group
,
1597 pe
->phb
->hose
->global_number
, pe
->pe_number
);
1598 pnv_ioda_setup_bus_iommu_group(pe
, &pe
->table_group
, NULL
);
1603 int pnv_pci_sriov_enable(struct pci_dev
*pdev
, u16 num_vfs
)
1605 struct pci_bus
*bus
;
1606 struct pci_controller
*hose
;
1607 struct pnv_phb
*phb
;
1608 struct pnv_ioda_pe
*pe
;
1614 hose
= pci_bus_to_host(bus
);
1615 phb
= hose
->private_data
;
1616 pdn
= pci_get_pdn(pdev
);
1618 if (phb
->type
== PNV_PHB_IODA2
) {
1619 if (!pdn
->vfs_expanded
) {
1620 dev_info(&pdev
->dev
, "don't support this SRIOV device"
1621 " with non 64bit-prefetchable IOV BAR\n");
1626 * When M64 BARs functions in Single PE mode, the number of VFs
1627 * could be enabled must be less than the number of M64 BARs.
1629 if (pdn
->m64_single_mode
&& num_vfs
> phb
->ioda
.m64_bar_idx
) {
1630 dev_info(&pdev
->dev
, "Not enough M64 BAR for VFs\n");
1634 /* Allocating pe_num_map */
1635 if (pdn
->m64_single_mode
)
1636 pdn
->pe_num_map
= kmalloc_array(num_vfs
,
1637 sizeof(*pdn
->pe_num_map
),
1640 pdn
->pe_num_map
= kmalloc(sizeof(*pdn
->pe_num_map
), GFP_KERNEL
);
1642 if (!pdn
->pe_num_map
)
1645 if (pdn
->m64_single_mode
)
1646 for (i
= 0; i
< num_vfs
; i
++)
1647 pdn
->pe_num_map
[i
] = IODA_INVALID_PE
;
1649 /* Calculate available PE for required VFs */
1650 if (pdn
->m64_single_mode
) {
1651 for (i
= 0; i
< num_vfs
; i
++) {
1652 pe
= pnv_ioda_alloc_pe(phb
);
1658 pdn
->pe_num_map
[i
] = pe
->pe_number
;
1661 mutex_lock(&phb
->ioda
.pe_alloc_mutex
);
1662 *pdn
->pe_num_map
= bitmap_find_next_zero_area(
1663 phb
->ioda
.pe_alloc
, phb
->ioda
.total_pe_num
,
1665 if (*pdn
->pe_num_map
>= phb
->ioda
.total_pe_num
) {
1666 mutex_unlock(&phb
->ioda
.pe_alloc_mutex
);
1667 dev_info(&pdev
->dev
, "Failed to enable VF%d\n", num_vfs
);
1668 kfree(pdn
->pe_num_map
);
1671 bitmap_set(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1672 mutex_unlock(&phb
->ioda
.pe_alloc_mutex
);
1674 pdn
->num_vfs
= num_vfs
;
1676 /* Assign M64 window accordingly */
1677 ret
= pnv_pci_vf_assign_m64(pdev
, num_vfs
);
1679 dev_info(&pdev
->dev
, "Not enough M64 window resources\n");
1684 * When using one M64 BAR to map one IOV BAR, we need to shift
1685 * the IOV BAR according to the PE# allocated to the VFs.
1686 * Otherwise, the PE# for the VF will conflict with others.
1688 if (!pdn
->m64_single_mode
) {
1689 ret
= pnv_pci_vf_resource_shift(pdev
, *pdn
->pe_num_map
);
1696 pnv_ioda_setup_vf_PE(pdev
, num_vfs
);
1701 if (pdn
->m64_single_mode
) {
1702 for (i
= 0; i
< num_vfs
; i
++) {
1703 if (pdn
->pe_num_map
[i
] == IODA_INVALID_PE
)
1706 pe
= &phb
->ioda
.pe_array
[pdn
->pe_num_map
[i
]];
1707 pnv_ioda_free_pe(pe
);
1710 bitmap_clear(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1712 /* Releasing pe_num_map */
1713 kfree(pdn
->pe_num_map
);
1718 int pnv_pcibios_sriov_disable(struct pci_dev
*pdev
)
1720 pnv_pci_sriov_disable(pdev
);
1722 /* Release PCI data */
1723 remove_dev_pci_data(pdev
);
1727 int pnv_pcibios_sriov_enable(struct pci_dev
*pdev
, u16 num_vfs
)
1729 /* Allocate PCI data */
1730 add_dev_pci_data(pdev
);
1732 return pnv_pci_sriov_enable(pdev
, num_vfs
);
1734 #endif /* CONFIG_PCI_IOV */
1736 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb
*phb
, struct pci_dev
*pdev
)
1738 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1739 struct pnv_ioda_pe
*pe
;
1742 * The function can be called while the PE#
1743 * hasn't been assigned. Do nothing for the
1746 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
1749 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1750 WARN_ON(get_dma_ops(&pdev
->dev
) != &dma_iommu_ops
);
1751 pdev
->dev
.archdata
.dma_offset
= pe
->tce_bypass_base
;
1752 set_iommu_table_base(&pdev
->dev
, pe
->table_group
.tables
[0]);
1754 * Note: iommu_add_device() will fail here as
1755 * for physical PE: the device is already added by now;
1756 * for virtual PE: sysfs entries are not ready yet and
1757 * tce_iommu_bus_notifier will add the device to a group later.
1762 * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1764 * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1765 * Devices can only access more than that if bit 59 of the PCI address is set
1766 * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1767 * Many PCI devices are not capable of addressing that many bits, and as a
1768 * result are limited to the 4GB of virtual memory made available to 32-bit
1771 * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1772 * devices by configuring the virtual memory past the first 4GB inaccessible
1773 * by 64-bit DMAs. This should only be used by devices that want more than
1774 * 4GB, and only on PEs that have no 32-bit devices.
1776 * Currently this will only work on PHB3 (POWER8).
1778 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe
*pe
)
1780 u64 window_size
, table_size
, tce_count
, addr
;
1781 struct page
*table_pages
;
1782 u64 tce_order
= 28; /* 256MB TCEs */
1787 * Window size needs to be a power of two, but needs to account for
1788 * shifting memory by the 4GB offset required to skip 32bit space.
1790 window_size
= roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1791 tce_count
= window_size
>> tce_order
;
1792 table_size
= tce_count
<< 3;
1794 if (table_size
< PAGE_SIZE
)
1795 table_size
= PAGE_SIZE
;
1797 table_pages
= alloc_pages_node(pe
->phb
->hose
->node
, GFP_KERNEL
,
1798 get_order(table_size
));
1802 tces
= page_address(table_pages
);
1806 memset(tces
, 0, table_size
);
1808 for (addr
= 0; addr
< memory_hotplug_max(); addr
+= (1 << tce_order
)) {
1809 tces
[(addr
+ (1ULL << 32)) >> tce_order
] =
1810 cpu_to_be64(addr
| TCE_PCI_READ
| TCE_PCI_WRITE
);
1813 rc
= opal_pci_map_pe_dma_window(pe
->phb
->opal_id
,
1815 /* reconfigure window 0 */
1816 (pe
->pe_number
<< 1) + 0,
1821 if (rc
== OPAL_SUCCESS
) {
1822 pe_info(pe
, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1826 pe_err(pe
, "Error configuring 64-bit DMA bypass\n");
1830 static bool pnv_pci_ioda_iommu_bypass_supported(struct pci_dev
*pdev
,
1833 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
1834 struct pnv_phb
*phb
= hose
->private_data
;
1835 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1836 struct pnv_ioda_pe
*pe
;
1838 if (WARN_ON(!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
))
1841 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1842 if (pe
->tce_bypass_enabled
) {
1843 u64 top
= pe
->tce_bypass_base
+ memblock_end_of_DRAM() - 1;
1844 if (dma_mask
>= top
)
1849 * If the device can't set the TCE bypass bit but still wants
1850 * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1851 * bypass the 32-bit region and be usable for 64-bit DMAs.
1852 * The device needs to be able to address all of this space.
1854 if (dma_mask
>> 32 &&
1855 dma_mask
> (memory_hotplug_max() + (1ULL << 32)) &&
1856 /* pe->pdev should be set if it's a single device, pe->pbus if not */
1857 (pe
->device_count
== 1 || !pe
->pbus
) &&
1858 phb
->model
== PNV_PHB_MODEL_PHB3
) {
1859 /* Configure the bypass mode */
1860 s64 rc
= pnv_pci_ioda_dma_64bit_bypass(pe
);
1863 /* 4GB offset bypasses 32-bit space */
1864 pdev
->dev
.archdata
.dma_offset
= (1ULL << 32);
1871 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe
*pe
, struct pci_bus
*bus
)
1873 struct pci_dev
*dev
;
1875 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
1876 set_iommu_table_base(&dev
->dev
, pe
->table_group
.tables
[0]);
1877 dev
->dev
.archdata
.dma_offset
= pe
->tce_bypass_base
;
1879 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
1880 pnv_ioda_setup_bus_dma(pe
, dev
->subordinate
);
1884 static inline __be64 __iomem
*pnv_ioda_get_inval_reg(struct pnv_phb
*phb
,
1887 return real_mode
? (__be64 __iomem
*)(phb
->regs_phys
+ 0x210) :
1888 (phb
->regs
+ 0x210);
1891 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table
*tbl
,
1892 unsigned long index
, unsigned long npages
, bool rm
)
1894 struct iommu_table_group_link
*tgl
= list_first_entry_or_null(
1895 &tbl
->it_group_list
, struct iommu_table_group_link
,
1897 struct pnv_ioda_pe
*pe
= container_of(tgl
->table_group
,
1898 struct pnv_ioda_pe
, table_group
);
1899 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(pe
->phb
, rm
);
1900 unsigned long start
, end
, inc
;
1902 start
= __pa(((__be64
*)tbl
->it_base
) + index
- tbl
->it_offset
);
1903 end
= __pa(((__be64
*)tbl
->it_base
) + index
- tbl
->it_offset
+
1906 /* p7ioc-style invalidation, 2 TCEs per write */
1907 start
|= (1ull << 63);
1908 end
|= (1ull << 63);
1910 end
|= inc
- 1; /* round up end to be different than start */
1912 mb(); /* Ensure above stores are visible */
1913 while (start
<= end
) {
1915 __raw_rm_writeq_be(start
, invalidate
);
1917 __raw_writeq_be(start
, invalidate
);
1923 * The iommu layer will do another mb() for us on build()
1924 * and we don't care on free()
1928 static int pnv_ioda1_tce_build(struct iommu_table
*tbl
, long index
,
1929 long npages
, unsigned long uaddr
,
1930 enum dma_data_direction direction
,
1931 unsigned long attrs
)
1933 int ret
= pnv_tce_build(tbl
, index
, npages
, uaddr
, direction
,
1937 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, npages
, false);
1942 #ifdef CONFIG_IOMMU_API
1943 static int pnv_ioda1_tce_xchg(struct iommu_table
*tbl
, long index
,
1944 unsigned long *hpa
, enum dma_data_direction
*direction
)
1946 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
, true);
1949 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, 1, false);
1954 static int pnv_ioda1_tce_xchg_rm(struct iommu_table
*tbl
, long index
,
1955 unsigned long *hpa
, enum dma_data_direction
*direction
)
1957 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
, false);
1960 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, 1, true);
1966 static void pnv_ioda1_tce_free(struct iommu_table
*tbl
, long index
,
1969 pnv_tce_free(tbl
, index
, npages
);
1971 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, npages
, false);
1974 static struct iommu_table_ops pnv_ioda1_iommu_ops
= {
1975 .set
= pnv_ioda1_tce_build
,
1976 #ifdef CONFIG_IOMMU_API
1977 .exchange
= pnv_ioda1_tce_xchg
,
1978 .exchange_rm
= pnv_ioda1_tce_xchg_rm
,
1979 .useraddrptr
= pnv_tce_useraddrptr
,
1981 .clear
= pnv_ioda1_tce_free
,
1985 #define PHB3_TCE_KILL_INVAL_ALL PPC_BIT(0)
1986 #define PHB3_TCE_KILL_INVAL_PE PPC_BIT(1)
1987 #define PHB3_TCE_KILL_INVAL_ONE PPC_BIT(2)
1989 static void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb
*phb
, bool rm
)
1991 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(phb
, rm
);
1992 const unsigned long val
= PHB3_TCE_KILL_INVAL_ALL
;
1994 mb(); /* Ensure previous TCE table stores are visible */
1996 __raw_rm_writeq_be(val
, invalidate
);
1998 __raw_writeq_be(val
, invalidate
);
2001 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe
*pe
)
2003 /* 01xb - invalidate TCEs that match the specified PE# */
2004 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(pe
->phb
, false);
2005 unsigned long val
= PHB3_TCE_KILL_INVAL_PE
| (pe
->pe_number
& 0xFF);
2007 mb(); /* Ensure above stores are visible */
2008 __raw_writeq_be(val
, invalidate
);
2011 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe
*pe
, bool rm
,
2012 unsigned shift
, unsigned long index
,
2013 unsigned long npages
)
2015 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(pe
->phb
, rm
);
2016 unsigned long start
, end
, inc
;
2018 /* We'll invalidate DMA address in PE scope */
2019 start
= PHB3_TCE_KILL_INVAL_ONE
;
2020 start
|= (pe
->pe_number
& 0xFF);
2023 /* Figure out the start, end and step */
2024 start
|= (index
<< shift
);
2025 end
|= ((index
+ npages
- 1) << shift
);
2026 inc
= (0x1ull
<< shift
);
2029 while (start
<= end
) {
2031 __raw_rm_writeq_be(start
, invalidate
);
2033 __raw_writeq_be(start
, invalidate
);
2038 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe
*pe
)
2040 struct pnv_phb
*phb
= pe
->phb
;
2042 if (phb
->model
== PNV_PHB_MODEL_PHB3
&& phb
->regs
)
2043 pnv_pci_phb3_tce_invalidate_pe(pe
);
2045 opal_pci_tce_kill(phb
->opal_id
, OPAL_PCI_TCE_KILL_PE
,
2046 pe
->pe_number
, 0, 0, 0);
2049 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table
*tbl
,
2050 unsigned long index
, unsigned long npages
, bool rm
)
2052 struct iommu_table_group_link
*tgl
;
2054 list_for_each_entry_lockless(tgl
, &tbl
->it_group_list
, next
) {
2055 struct pnv_ioda_pe
*pe
= container_of(tgl
->table_group
,
2056 struct pnv_ioda_pe
, table_group
);
2057 struct pnv_phb
*phb
= pe
->phb
;
2058 unsigned int shift
= tbl
->it_page_shift
;
2061 * NVLink1 can use the TCE kill register directly as
2062 * it's the same as PHB3. NVLink2 is different and
2063 * should go via the OPAL call.
2065 if (phb
->model
== PNV_PHB_MODEL_NPU
) {
2067 * The NVLink hardware does not support TCE kill
2068 * per TCE entry so we have to invalidate
2069 * the entire cache for it.
2071 pnv_pci_phb3_tce_invalidate_entire(phb
, rm
);
2074 if (phb
->model
== PNV_PHB_MODEL_PHB3
&& phb
->regs
)
2075 pnv_pci_phb3_tce_invalidate(pe
, rm
, shift
,
2078 opal_pci_tce_kill(phb
->opal_id
,
2079 OPAL_PCI_TCE_KILL_PAGES
,
2080 pe
->pe_number
, 1u << shift
,
2081 index
<< shift
, npages
);
2085 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb
*phb
, bool rm
)
2087 if (phb
->model
== PNV_PHB_MODEL_NPU
|| phb
->model
== PNV_PHB_MODEL_PHB3
)
2088 pnv_pci_phb3_tce_invalidate_entire(phb
, rm
);
2090 opal_pci_tce_kill(phb
->opal_id
, OPAL_PCI_TCE_KILL
, 0, 0, 0, 0);
2093 static int pnv_ioda2_tce_build(struct iommu_table
*tbl
, long index
,
2094 long npages
, unsigned long uaddr
,
2095 enum dma_data_direction direction
,
2096 unsigned long attrs
)
2098 int ret
= pnv_tce_build(tbl
, index
, npages
, uaddr
, direction
,
2102 pnv_pci_ioda2_tce_invalidate(tbl
, index
, npages
, false);
2107 #ifdef CONFIG_IOMMU_API
2108 static int pnv_ioda2_tce_xchg(struct iommu_table
*tbl
, long index
,
2109 unsigned long *hpa
, enum dma_data_direction
*direction
)
2111 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
, true);
2114 pnv_pci_ioda2_tce_invalidate(tbl
, index
, 1, false);
2119 static int pnv_ioda2_tce_xchg_rm(struct iommu_table
*tbl
, long index
,
2120 unsigned long *hpa
, enum dma_data_direction
*direction
)
2122 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
, false);
2125 pnv_pci_ioda2_tce_invalidate(tbl
, index
, 1, true);
2131 static void pnv_ioda2_tce_free(struct iommu_table
*tbl
, long index
,
2134 pnv_tce_free(tbl
, index
, npages
);
2136 pnv_pci_ioda2_tce_invalidate(tbl
, index
, npages
, false);
2139 static struct iommu_table_ops pnv_ioda2_iommu_ops
= {
2140 .set
= pnv_ioda2_tce_build
,
2141 #ifdef CONFIG_IOMMU_API
2142 .exchange
= pnv_ioda2_tce_xchg
,
2143 .exchange_rm
= pnv_ioda2_tce_xchg_rm
,
2144 .useraddrptr
= pnv_tce_useraddrptr
,
2146 .clear
= pnv_ioda2_tce_free
,
2148 .free
= pnv_pci_ioda2_table_free_pages
,
2151 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev
*dev
, void *data
)
2153 unsigned int *weight
= (unsigned int *)data
;
2155 /* This is quite simplistic. The "base" weight of a device
2156 * is 10. 0 means no DMA is to be accounted for it.
2158 if (dev
->hdr_type
!= PCI_HEADER_TYPE_NORMAL
)
2161 if (dev
->class == PCI_CLASS_SERIAL_USB_UHCI
||
2162 dev
->class == PCI_CLASS_SERIAL_USB_OHCI
||
2163 dev
->class == PCI_CLASS_SERIAL_USB_EHCI
)
2165 else if ((dev
->class >> 8) == PCI_CLASS_STORAGE_RAID
)
2173 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe
*pe
)
2175 unsigned int weight
= 0;
2177 /* SRIOV VF has same DMA32 weight as its PF */
2178 #ifdef CONFIG_PCI_IOV
2179 if ((pe
->flags
& PNV_IODA_PE_VF
) && pe
->parent_dev
) {
2180 pnv_pci_ioda_dev_dma_weight(pe
->parent_dev
, &weight
);
2185 if ((pe
->flags
& PNV_IODA_PE_DEV
) && pe
->pdev
) {
2186 pnv_pci_ioda_dev_dma_weight(pe
->pdev
, &weight
);
2187 } else if ((pe
->flags
& PNV_IODA_PE_BUS
) && pe
->pbus
) {
2188 struct pci_dev
*pdev
;
2190 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
)
2191 pnv_pci_ioda_dev_dma_weight(pdev
, &weight
);
2192 } else if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && pe
->pbus
) {
2193 pci_walk_bus(pe
->pbus
, pnv_pci_ioda_dev_dma_weight
, &weight
);
2199 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb
*phb
,
2200 struct pnv_ioda_pe
*pe
)
2203 struct page
*tce_mem
= NULL
;
2204 struct iommu_table
*tbl
;
2205 unsigned int weight
, total_weight
= 0;
2206 unsigned int tce32_segsz
, base
, segs
, avail
, i
;
2210 /* XXX FIXME: Handle 64-bit only DMA devices */
2211 /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2212 /* XXX FIXME: Allocate multi-level tables on PHB3 */
2213 weight
= pnv_pci_ioda_pe_dma_weight(pe
);
2217 pci_walk_bus(phb
->hose
->bus
, pnv_pci_ioda_dev_dma_weight
,
2219 segs
= (weight
* phb
->ioda
.dma32_count
) / total_weight
;
2224 * Allocate contiguous DMA32 segments. We begin with the expected
2225 * number of segments. With one more attempt, the number of DMA32
2226 * segments to be allocated is decreased by one until one segment
2227 * is allocated successfully.
2230 for (base
= 0; base
<= phb
->ioda
.dma32_count
- segs
; base
++) {
2231 for (avail
= 0, i
= base
; i
< base
+ segs
; i
++) {
2232 if (phb
->ioda
.dma32_segmap
[i
] ==
2243 pe_warn(pe
, "No available DMA32 segments\n");
2248 tbl
= pnv_pci_table_alloc(phb
->hose
->node
);
2252 iommu_register_group(&pe
->table_group
, phb
->hose
->global_number
,
2254 pnv_pci_link_table_and_group(phb
->hose
->node
, 0, tbl
, &pe
->table_group
);
2256 /* Grab a 32-bit TCE table */
2257 pe_info(pe
, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2258 weight
, total_weight
, base
, segs
);
2259 pe_info(pe
, " Setting up 32-bit TCE table at %08x..%08x\n",
2260 base
* PNV_IODA1_DMA32_SEGSIZE
,
2261 (base
+ segs
) * PNV_IODA1_DMA32_SEGSIZE
- 1);
2263 /* XXX Currently, we allocate one big contiguous table for the
2264 * TCEs. We only really need one chunk per 256M of TCE space
2265 * (ie per segment) but that's an optimization for later, it
2266 * requires some added smarts with our get/put_tce implementation
2268 * Each TCE page is 4KB in size and each TCE entry occupies 8
2271 tce32_segsz
= PNV_IODA1_DMA32_SEGSIZE
>> (IOMMU_PAGE_SHIFT_4K
- 3);
2272 tce_mem
= alloc_pages_node(phb
->hose
->node
, GFP_KERNEL
,
2273 get_order(tce32_segsz
* segs
));
2275 pe_err(pe
, " Failed to allocate a 32-bit TCE memory\n");
2278 addr
= page_address(tce_mem
);
2279 memset(addr
, 0, tce32_segsz
* segs
);
2282 for (i
= 0; i
< segs
; i
++) {
2283 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
,
2286 __pa(addr
) + tce32_segsz
* i
,
2287 tce32_segsz
, IOMMU_PAGE_SIZE_4K
);
2289 pe_err(pe
, " Failed to configure 32-bit TCE table,"
2295 /* Setup DMA32 segment mapping */
2296 for (i
= base
; i
< base
+ segs
; i
++)
2297 phb
->ioda
.dma32_segmap
[i
] = pe
->pe_number
;
2299 /* Setup linux iommu table */
2300 pnv_pci_setup_iommu_table(tbl
, addr
, tce32_segsz
* segs
,
2301 base
* PNV_IODA1_DMA32_SEGSIZE
,
2302 IOMMU_PAGE_SHIFT_4K
);
2304 tbl
->it_ops
= &pnv_ioda1_iommu_ops
;
2305 pe
->table_group
.tce32_start
= tbl
->it_offset
<< tbl
->it_page_shift
;
2306 pe
->table_group
.tce32_size
= tbl
->it_size
<< tbl
->it_page_shift
;
2307 iommu_init_table(tbl
, phb
->hose
->node
);
2309 if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
2310 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
);
2314 /* XXX Failure: Try to fallback to 64-bit only ? */
2316 __free_pages(tce_mem
, get_order(tce32_segsz
* segs
));
2318 pnv_pci_unlink_table_and_group(tbl
, &pe
->table_group
);
2319 iommu_tce_table_put(tbl
);
2323 static long pnv_pci_ioda2_set_window(struct iommu_table_group
*table_group
,
2324 int num
, struct iommu_table
*tbl
)
2326 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2328 struct pnv_phb
*phb
= pe
->phb
;
2330 const unsigned long size
= tbl
->it_indirect_levels
?
2331 tbl
->it_level_size
: tbl
->it_size
;
2332 const __u64 start_addr
= tbl
->it_offset
<< tbl
->it_page_shift
;
2333 const __u64 win_size
= tbl
->it_size
<< tbl
->it_page_shift
;
2335 pe_info(pe
, "Setting up window#%d %llx..%llx pg=%x\n", num
,
2336 start_addr
, start_addr
+ win_size
- 1,
2337 IOMMU_PAGE_SIZE(tbl
));
2340 * Map TCE table through TVT. The TVE index is the PE number
2341 * shifted by 1 bit for 32-bits DMA space.
2343 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
,
2345 (pe
->pe_number
<< 1) + num
,
2346 tbl
->it_indirect_levels
+ 1,
2349 IOMMU_PAGE_SIZE(tbl
));
2351 pe_err(pe
, "Failed to configure TCE table, err %ld\n", rc
);
2355 pnv_pci_link_table_and_group(phb
->hose
->node
, num
,
2356 tbl
, &pe
->table_group
);
2357 pnv_pci_ioda2_tce_invalidate_pe(pe
);
2362 void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe
*pe
, bool enable
)
2364 uint16_t window_id
= (pe
->pe_number
<< 1 ) + 1;
2367 pe_info(pe
, "%sabling 64-bit DMA bypass\n", enable
? "En" : "Dis");
2369 phys_addr_t top
= memblock_end_of_DRAM();
2371 top
= roundup_pow_of_two(top
);
2372 rc
= opal_pci_map_pe_dma_window_real(pe
->phb
->opal_id
,
2375 pe
->tce_bypass_base
,
2378 rc
= opal_pci_map_pe_dma_window_real(pe
->phb
->opal_id
,
2381 pe
->tce_bypass_base
,
2385 pe_err(pe
, "OPAL error %lld configuring bypass window\n", rc
);
2387 pe
->tce_bypass_enabled
= enable
;
2390 static long pnv_pci_ioda2_create_table(struct iommu_table_group
*table_group
,
2391 int num
, __u32 page_shift
, __u64 window_size
, __u32 levels
,
2392 bool alloc_userspace_copy
, struct iommu_table
**ptbl
)
2394 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2396 int nid
= pe
->phb
->hose
->node
;
2397 __u64 bus_offset
= num
? pe
->tce_bypass_base
: table_group
->tce32_start
;
2399 struct iommu_table
*tbl
;
2401 tbl
= pnv_pci_table_alloc(nid
);
2405 tbl
->it_ops
= &pnv_ioda2_iommu_ops
;
2407 ret
= pnv_pci_ioda2_table_alloc_pages(nid
,
2408 bus_offset
, page_shift
, window_size
,
2409 levels
, alloc_userspace_copy
, tbl
);
2411 iommu_tce_table_put(tbl
);
2420 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe
*pe
)
2422 struct iommu_table
*tbl
= NULL
;
2426 * crashkernel= specifies the kdump kernel's maximum memory at
2427 * some offset and there is no guaranteed the result is a power
2428 * of 2, which will cause errors later.
2430 const u64 max_memory
= __rounddown_pow_of_two(memory_hotplug_max());
2433 * In memory constrained environments, e.g. kdump kernel, the
2434 * DMA window can be larger than available memory, which will
2435 * cause errors later.
2437 const u64 window_size
= min((u64
)pe
->table_group
.tce32_size
, max_memory
);
2439 rc
= pnv_pci_ioda2_create_table(&pe
->table_group
, 0,
2440 IOMMU_PAGE_SHIFT_4K
,
2442 POWERNV_IOMMU_DEFAULT_LEVELS
, false, &tbl
);
2444 pe_err(pe
, "Failed to create 32-bit TCE table, err %ld",
2449 iommu_init_table(tbl
, pe
->phb
->hose
->node
);
2451 rc
= pnv_pci_ioda2_set_window(&pe
->table_group
, 0, tbl
);
2453 pe_err(pe
, "Failed to configure 32-bit TCE table, err %ld\n",
2455 iommu_tce_table_put(tbl
);
2459 if (!pnv_iommu_bypass_disabled
)
2460 pnv_pci_ioda2_set_bypass(pe
, true);
2465 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2466 static long pnv_pci_ioda2_unset_window(struct iommu_table_group
*table_group
,
2469 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2471 struct pnv_phb
*phb
= pe
->phb
;
2474 pe_info(pe
, "Removing DMA window #%d\n", num
);
2476 ret
= opal_pci_map_pe_dma_window(phb
->opal_id
, pe
->pe_number
,
2477 (pe
->pe_number
<< 1) + num
,
2478 0/* levels */, 0/* table address */,
2479 0/* table size */, 0/* page size */);
2481 pe_warn(pe
, "Unmapping failed, ret = %ld\n", ret
);
2483 pnv_pci_ioda2_tce_invalidate_pe(pe
);
2485 pnv_pci_unlink_table_and_group(table_group
->tables
[num
], table_group
);
2491 #ifdef CONFIG_IOMMU_API
2492 unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift
,
2493 __u64 window_size
, __u32 levels
)
2495 unsigned long bytes
= 0;
2496 const unsigned window_shift
= ilog2(window_size
);
2497 unsigned entries_shift
= window_shift
- page_shift
;
2498 unsigned table_shift
= entries_shift
+ 3;
2499 unsigned long tce_table_size
= max(0x1000UL
, 1UL << table_shift
);
2500 unsigned long direct_table_size
;
2502 if (!levels
|| (levels
> POWERNV_IOMMU_MAX_LEVELS
) ||
2503 !is_power_of_2(window_size
))
2506 /* Calculate a direct table size from window_size and levels */
2507 entries_shift
= (entries_shift
+ levels
- 1) / levels
;
2508 table_shift
= entries_shift
+ 3;
2509 table_shift
= max_t(unsigned, table_shift
, PAGE_SHIFT
);
2510 direct_table_size
= 1UL << table_shift
;
2512 for ( ; levels
; --levels
) {
2513 bytes
+= _ALIGN_UP(tce_table_size
, direct_table_size
);
2515 tce_table_size
/= direct_table_size
;
2516 tce_table_size
<<= 3;
2517 tce_table_size
= max_t(unsigned long,
2518 tce_table_size
, direct_table_size
);
2521 return bytes
+ bytes
; /* one for HW table, one for userspace copy */
2524 static long pnv_pci_ioda2_create_table_userspace(
2525 struct iommu_table_group
*table_group
,
2526 int num
, __u32 page_shift
, __u64 window_size
, __u32 levels
,
2527 struct iommu_table
**ptbl
)
2529 long ret
= pnv_pci_ioda2_create_table(table_group
,
2530 num
, page_shift
, window_size
, levels
, true, ptbl
);
2533 (*ptbl
)->it_allocated_size
= pnv_pci_ioda2_get_table_size(
2534 page_shift
, window_size
, levels
);
2538 static void pnv_ioda2_take_ownership(struct iommu_table_group
*table_group
)
2540 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2542 /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2543 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
2545 pnv_pci_ioda2_set_bypass(pe
, false);
2546 pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
2548 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
);
2549 iommu_tce_table_put(tbl
);
2552 static void pnv_ioda2_release_ownership(struct iommu_table_group
*table_group
)
2554 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2557 pnv_pci_ioda2_setup_default_config(pe
);
2559 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
);
2562 static struct iommu_table_group_ops pnv_pci_ioda2_ops
= {
2563 .get_table_size
= pnv_pci_ioda2_get_table_size
,
2564 .create_table
= pnv_pci_ioda2_create_table_userspace
,
2565 .set_window
= pnv_pci_ioda2_set_window
,
2566 .unset_window
= pnv_pci_ioda2_unset_window
,
2567 .take_ownership
= pnv_ioda2_take_ownership
,
2568 .release_ownership
= pnv_ioda2_release_ownership
,
2571 static void pnv_ioda_setup_bus_iommu_group_add_devices(struct pnv_ioda_pe
*pe
,
2572 struct iommu_table_group
*table_group
,
2573 struct pci_bus
*bus
)
2575 struct pci_dev
*dev
;
2577 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
2578 iommu_add_device(table_group
, &dev
->dev
);
2580 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
2581 pnv_ioda_setup_bus_iommu_group_add_devices(pe
,
2582 table_group
, dev
->subordinate
);
2586 static void pnv_ioda_setup_bus_iommu_group(struct pnv_ioda_pe
*pe
,
2587 struct iommu_table_group
*table_group
, struct pci_bus
*bus
)
2590 if (pe
->flags
& PNV_IODA_PE_DEV
)
2591 iommu_add_device(table_group
, &pe
->pdev
->dev
);
2593 if ((pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
)) || bus
)
2594 pnv_ioda_setup_bus_iommu_group_add_devices(pe
, table_group
,
2598 static unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb
*phb
);
2600 static void pnv_pci_ioda_setup_iommu_api(void)
2602 struct pci_controller
*hose
;
2603 struct pnv_phb
*phb
;
2604 struct pnv_ioda_pe
*pe
;
2607 * There are 4 types of PEs:
2608 * - PNV_IODA_PE_BUS: a downstream port with an adapter,
2609 * created from pnv_pci_setup_bridge();
2610 * - PNV_IODA_PE_BUS_ALL: a PCI-PCIX bridge with devices behind it,
2611 * created from pnv_pci_setup_bridge();
2612 * - PNV_IODA_PE_VF: a SRIOV virtual function,
2613 * created from pnv_pcibios_sriov_enable();
2614 * - PNV_IODA_PE_DEV: an NPU or OCAPI device,
2615 * created from pnv_pci_ioda_fixup().
2617 * Normally a PE is represented by an IOMMU group, however for
2618 * devices with side channels the groups need to be more strict.
2620 list_for_each_entry(hose
, &hose_list
, list_node
) {
2621 phb
= hose
->private_data
;
2623 if (phb
->type
== PNV_PHB_NPU_NVLINK
||
2624 phb
->type
== PNV_PHB_NPU_OCAPI
)
2627 list_for_each_entry(pe
, &phb
->ioda
.pe_list
, list
) {
2628 struct iommu_table_group
*table_group
;
2630 table_group
= pnv_try_setup_npu_table_group(pe
);
2632 if (!pnv_pci_ioda_pe_dma_weight(pe
))
2635 table_group
= &pe
->table_group
;
2636 iommu_register_group(&pe
->table_group
,
2637 pe
->phb
->hose
->global_number
,
2640 pnv_ioda_setup_bus_iommu_group(pe
, table_group
,
2646 * Now we have all PHBs discovered, time to add NPU devices to
2647 * the corresponding IOMMU groups.
2649 list_for_each_entry(hose
, &hose_list
, list_node
) {
2650 unsigned long pgsizes
;
2652 phb
= hose
->private_data
;
2654 if (phb
->type
!= PNV_PHB_NPU_NVLINK
)
2657 pgsizes
= pnv_ioda_parse_tce_sizes(phb
);
2658 list_for_each_entry(pe
, &phb
->ioda
.pe_list
, list
) {
2660 * IODA2 bridges get this set up from
2661 * pci_controller_ops::setup_bridge but NPU bridges
2662 * do not have this hook defined so we do it here.
2664 pe
->table_group
.pgsizes
= pgsizes
;
2665 pnv_npu_compound_attach(pe
);
2669 #else /* !CONFIG_IOMMU_API */
2670 static void pnv_pci_ioda_setup_iommu_api(void) { };
2673 static unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb
*phb
)
2675 struct pci_controller
*hose
= phb
->hose
;
2676 struct device_node
*dn
= hose
->dn
;
2677 unsigned long mask
= 0;
2681 count
= of_property_count_u32_elems(dn
, "ibm,supported-tce-sizes");
2683 mask
= SZ_4K
| SZ_64K
;
2684 /* Add 16M for POWER8 by default */
2685 if (cpu_has_feature(CPU_FTR_ARCH_207S
) &&
2686 !cpu_has_feature(CPU_FTR_ARCH_300
))
2687 mask
|= SZ_16M
| SZ_256M
;
2691 for (i
= 0; i
< count
; i
++) {
2692 rc
= of_property_read_u32_index(dn
, "ibm,supported-tce-sizes",
2695 mask
|= 1ULL << val
;
2701 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb
*phb
,
2702 struct pnv_ioda_pe
*pe
)
2706 if (!pnv_pci_ioda_pe_dma_weight(pe
))
2709 /* TVE #1 is selected by PCI address bit 59 */
2710 pe
->tce_bypass_base
= 1ull << 59;
2712 /* The PE will reserve all possible 32-bits space */
2713 pe_info(pe
, "Setting up 32-bit TCE table at 0..%08x\n",
2714 phb
->ioda
.m32_pci_base
);
2716 /* Setup linux iommu table */
2717 pe
->table_group
.tce32_start
= 0;
2718 pe
->table_group
.tce32_size
= phb
->ioda
.m32_pci_base
;
2719 pe
->table_group
.max_dynamic_windows_supported
=
2720 IOMMU_TABLE_GROUP_MAX_TABLES
;
2721 pe
->table_group
.max_levels
= POWERNV_IOMMU_MAX_LEVELS
;
2722 pe
->table_group
.pgsizes
= pnv_ioda_parse_tce_sizes(phb
);
2723 #ifdef CONFIG_IOMMU_API
2724 pe
->table_group
.ops
= &pnv_pci_ioda2_ops
;
2727 rc
= pnv_pci_ioda2_setup_default_config(pe
);
2731 if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
2732 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
);
2735 int64_t pnv_opal_pci_msi_eoi(struct irq_chip
*chip
, unsigned int hw_irq
)
2737 struct pnv_phb
*phb
= container_of(chip
, struct pnv_phb
,
2740 return opal_pci_msi_eoi(phb
->opal_id
, hw_irq
);
2743 static void pnv_ioda2_msi_eoi(struct irq_data
*d
)
2746 unsigned int hw_irq
= (unsigned int)irqd_to_hwirq(d
);
2747 struct irq_chip
*chip
= irq_data_get_irq_chip(d
);
2749 rc
= pnv_opal_pci_msi_eoi(chip
, hw_irq
);
2756 void pnv_set_msi_irq_chip(struct pnv_phb
*phb
, unsigned int virq
)
2758 struct irq_data
*idata
;
2759 struct irq_chip
*ichip
;
2761 /* The MSI EOI OPAL call is only needed on PHB3 */
2762 if (phb
->model
!= PNV_PHB_MODEL_PHB3
)
2765 if (!phb
->ioda
.irq_chip_init
) {
2767 * First time we setup an MSI IRQ, we need to setup the
2768 * corresponding IRQ chip to route correctly.
2770 idata
= irq_get_irq_data(virq
);
2771 ichip
= irq_data_get_irq_chip(idata
);
2772 phb
->ioda
.irq_chip_init
= 1;
2773 phb
->ioda
.irq_chip
= *ichip
;
2774 phb
->ioda
.irq_chip
.irq_eoi
= pnv_ioda2_msi_eoi
;
2776 irq_set_chip(virq
, &phb
->ioda
.irq_chip
);
2780 * Returns true iff chip is something that we could call
2781 * pnv_opal_pci_msi_eoi for.
2783 bool is_pnv_opal_msi(struct irq_chip
*chip
)
2785 return chip
->irq_eoi
== pnv_ioda2_msi_eoi
;
2787 EXPORT_SYMBOL_GPL(is_pnv_opal_msi
);
2789 static int pnv_pci_ioda_msi_setup(struct pnv_phb
*phb
, struct pci_dev
*dev
,
2790 unsigned int hwirq
, unsigned int virq
,
2791 unsigned int is_64
, struct msi_msg
*msg
)
2793 struct pnv_ioda_pe
*pe
= pnv_ioda_get_pe(dev
);
2794 unsigned int xive_num
= hwirq
- phb
->msi_base
;
2798 /* No PE assigned ? bail out ... no MSI for you ! */
2802 /* Check if we have an MVE */
2803 if (pe
->mve_number
< 0)
2806 /* Force 32-bit MSI on some broken devices */
2807 if (dev
->no_64bit_msi
)
2810 /* Assign XIVE to PE */
2811 rc
= opal_pci_set_xive_pe(phb
->opal_id
, pe
->pe_number
, xive_num
);
2813 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2814 pci_name(dev
), rc
, xive_num
);
2821 rc
= opal_get_msi_64(phb
->opal_id
, pe
->mve_number
, xive_num
, 1,
2824 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2828 msg
->address_hi
= be64_to_cpu(addr64
) >> 32;
2829 msg
->address_lo
= be64_to_cpu(addr64
) & 0xfffffffful
;
2833 rc
= opal_get_msi_32(phb
->opal_id
, pe
->mve_number
, xive_num
, 1,
2836 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2840 msg
->address_hi
= 0;
2841 msg
->address_lo
= be32_to_cpu(addr32
);
2843 msg
->data
= be32_to_cpu(data
);
2845 pnv_set_msi_irq_chip(phb
, virq
);
2847 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2848 " address=%x_%08x data=%x PE# %x\n",
2849 pci_name(dev
), is_64
? "64" : "32", hwirq
, xive_num
,
2850 msg
->address_hi
, msg
->address_lo
, data
, pe
->pe_number
);
2855 static void pnv_pci_init_ioda_msis(struct pnv_phb
*phb
)
2858 const __be32
*prop
= of_get_property(phb
->hose
->dn
,
2859 "ibm,opal-msi-ranges", NULL
);
2862 prop
= of_get_property(phb
->hose
->dn
, "msi-ranges", NULL
);
2867 phb
->msi_base
= be32_to_cpup(prop
);
2868 count
= be32_to_cpup(prop
+ 1);
2869 if (msi_bitmap_alloc(&phb
->msi_bmp
, count
, phb
->hose
->dn
)) {
2870 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2871 phb
->hose
->global_number
);
2875 phb
->msi_setup
= pnv_pci_ioda_msi_setup
;
2876 phb
->msi32_support
= 1;
2877 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2878 count
, phb
->msi_base
);
2881 #ifdef CONFIG_PCI_IOV
2882 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev
*pdev
)
2884 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
2885 struct pnv_phb
*phb
= hose
->private_data
;
2886 const resource_size_t gate
= phb
->ioda
.m64_segsize
>> 2;
2887 struct resource
*res
;
2889 resource_size_t size
, total_vf_bar_sz
;
2893 if (!pdev
->is_physfn
|| pci_dev_is_added(pdev
))
2896 pdn
= pci_get_pdn(pdev
);
2897 pdn
->vfs_expanded
= 0;
2898 pdn
->m64_single_mode
= false;
2900 total_vfs
= pci_sriov_get_totalvfs(pdev
);
2901 mul
= phb
->ioda
.total_pe_num
;
2902 total_vf_bar_sz
= 0;
2904 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
2905 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
2906 if (!res
->flags
|| res
->parent
)
2908 if (!pnv_pci_is_m64_flags(res
->flags
)) {
2909 dev_warn(&pdev
->dev
, "Don't support SR-IOV with"
2910 " non M64 VF BAR%d: %pR. \n",
2915 total_vf_bar_sz
+= pci_iov_resource_size(pdev
,
2916 i
+ PCI_IOV_RESOURCES
);
2919 * If bigger than quarter of M64 segment size, just round up
2922 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
2923 * with other devices, IOV BAR size is expanded to be
2924 * (total_pe * VF_BAR_size). When VF_BAR_size is half of M64
2925 * segment size , the expanded size would equal to half of the
2926 * whole M64 space size, which will exhaust the M64 Space and
2927 * limit the system flexibility. This is a design decision to
2928 * set the boundary to quarter of the M64 segment size.
2930 if (total_vf_bar_sz
> gate
) {
2931 mul
= roundup_pow_of_two(total_vfs
);
2932 dev_info(&pdev
->dev
,
2933 "VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
2934 total_vf_bar_sz
, gate
, mul
);
2935 pdn
->m64_single_mode
= true;
2940 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
2941 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
2942 if (!res
->flags
|| res
->parent
)
2945 size
= pci_iov_resource_size(pdev
, i
+ PCI_IOV_RESOURCES
);
2947 * On PHB3, the minimum size alignment of M64 BAR in single
2950 if (pdn
->m64_single_mode
&& (size
< SZ_32M
))
2952 dev_dbg(&pdev
->dev
, " Fixing VF BAR%d: %pR to\n", i
, res
);
2953 res
->end
= res
->start
+ size
* mul
- 1;
2954 dev_dbg(&pdev
->dev
, " %pR\n", res
);
2955 dev_info(&pdev
->dev
, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2958 pdn
->vfs_expanded
= mul
;
2963 /* To save MMIO space, IOV BAR is truncated. */
2964 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
2965 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
2967 res
->end
= res
->start
- 1;
2970 #endif /* CONFIG_PCI_IOV */
2972 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe
*pe
,
2973 struct resource
*res
)
2975 struct pnv_phb
*phb
= pe
->phb
;
2976 struct pci_bus_region region
;
2980 if (!res
|| !res
->flags
|| res
->start
> res
->end
)
2983 if (res
->flags
& IORESOURCE_IO
) {
2984 region
.start
= res
->start
- phb
->ioda
.io_pci_base
;
2985 region
.end
= res
->end
- phb
->ioda
.io_pci_base
;
2986 index
= region
.start
/ phb
->ioda
.io_segsize
;
2988 while (index
< phb
->ioda
.total_pe_num
&&
2989 region
.start
<= region
.end
) {
2990 phb
->ioda
.io_segmap
[index
] = pe
->pe_number
;
2991 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
2992 pe
->pe_number
, OPAL_IO_WINDOW_TYPE
, 0, index
);
2993 if (rc
!= OPAL_SUCCESS
) {
2994 pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
2995 __func__
, rc
, index
, pe
->pe_number
);
2999 region
.start
+= phb
->ioda
.io_segsize
;
3002 } else if ((res
->flags
& IORESOURCE_MEM
) &&
3003 !pnv_pci_is_m64(phb
, res
)) {
3004 region
.start
= res
->start
-
3005 phb
->hose
->mem_offset
[0] -
3006 phb
->ioda
.m32_pci_base
;
3007 region
.end
= res
->end
-
3008 phb
->hose
->mem_offset
[0] -
3009 phb
->ioda
.m32_pci_base
;
3010 index
= region
.start
/ phb
->ioda
.m32_segsize
;
3012 while (index
< phb
->ioda
.total_pe_num
&&
3013 region
.start
<= region
.end
) {
3014 phb
->ioda
.m32_segmap
[index
] = pe
->pe_number
;
3015 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3016 pe
->pe_number
, OPAL_M32_WINDOW_TYPE
, 0, index
);
3017 if (rc
!= OPAL_SUCCESS
) {
3018 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
3019 __func__
, rc
, index
, pe
->pe_number
);
3023 region
.start
+= phb
->ioda
.m32_segsize
;
3030 * This function is supposed to be called on basis of PE from top
3031 * to bottom style. So the the I/O or MMIO segment assigned to
3032 * parent PE could be overridden by its child PEs if necessary.
3034 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe
*pe
)
3036 struct pci_dev
*pdev
;
3040 * NOTE: We only care PCI bus based PE for now. For PCI
3041 * device based PE, for example SRIOV sensitive VF should
3042 * be figured out later.
3044 BUG_ON(!(pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
)));
3046 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
) {
3047 for (i
= 0; i
<= PCI_ROM_RESOURCE
; i
++)
3048 pnv_ioda_setup_pe_res(pe
, &pdev
->resource
[i
]);
3051 * If the PE contains all subordinate PCI buses, the
3052 * windows of the child bridges should be mapped to
3055 if (!(pe
->flags
& PNV_IODA_PE_BUS_ALL
) || !pci_is_bridge(pdev
))
3057 for (i
= 0; i
< PCI_BRIDGE_RESOURCE_NUM
; i
++)
3058 pnv_ioda_setup_pe_res(pe
,
3059 &pdev
->resource
[PCI_BRIDGE_RESOURCES
+ i
]);
3063 #ifdef CONFIG_DEBUG_FS
3064 static int pnv_pci_diag_data_set(void *data
, u64 val
)
3066 struct pci_controller
*hose
;
3067 struct pnv_phb
*phb
;
3073 hose
= (struct pci_controller
*)data
;
3074 if (!hose
|| !hose
->private_data
)
3077 phb
= hose
->private_data
;
3079 /* Retrieve the diag data from firmware */
3080 ret
= opal_pci_get_phb_diag_data2(phb
->opal_id
, phb
->diag_data
,
3081 phb
->diag_data_size
);
3082 if (ret
!= OPAL_SUCCESS
)
3085 /* Print the diag data to the kernel log */
3086 pnv_pci_dump_phb_diag_data(phb
->hose
, phb
->diag_data
);
3090 DEFINE_SIMPLE_ATTRIBUTE(pnv_pci_diag_data_fops
, NULL
,
3091 pnv_pci_diag_data_set
, "%llu\n");
3093 #endif /* CONFIG_DEBUG_FS */
3095 static void pnv_pci_ioda_create_dbgfs(void)
3097 #ifdef CONFIG_DEBUG_FS
3098 struct pci_controller
*hose
, *tmp
;
3099 struct pnv_phb
*phb
;
3102 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
3103 phb
= hose
->private_data
;
3105 /* Notify initialization of PHB done */
3106 phb
->initialized
= 1;
3108 sprintf(name
, "PCI%04x", hose
->global_number
);
3109 phb
->dbgfs
= debugfs_create_dir(name
, powerpc_debugfs_root
);
3111 pr_warn("%s: Error on creating debugfs on PHB#%x\n",
3112 __func__
, hose
->global_number
);
3116 debugfs_create_file("dump_diag_regs", 0200, phb
->dbgfs
, hose
,
3117 &pnv_pci_diag_data_fops
);
3119 #endif /* CONFIG_DEBUG_FS */
3122 static void pnv_pci_enable_bridge(struct pci_bus
*bus
)
3124 struct pci_dev
*dev
= bus
->self
;
3125 struct pci_bus
*child
;
3127 /* Empty bus ? bail */
3128 if (list_empty(&bus
->devices
))
3132 * If there's a bridge associated with that bus enable it. This works
3133 * around races in the generic code if the enabling is done during
3134 * parallel probing. This can be removed once those races have been
3138 int rc
= pci_enable_device(dev
);
3140 pci_err(dev
, "Error enabling bridge (%d)\n", rc
);
3141 pci_set_master(dev
);
3144 /* Perform the same to child busses */
3145 list_for_each_entry(child
, &bus
->children
, node
)
3146 pnv_pci_enable_bridge(child
);
3149 static void pnv_pci_enable_bridges(void)
3151 struct pci_controller
*hose
;
3153 list_for_each_entry(hose
, &hose_list
, list_node
)
3154 pnv_pci_enable_bridge(hose
->bus
);
3157 static void pnv_pci_ioda_fixup(void)
3159 pnv_pci_ioda_setup_PEs();
3160 pnv_pci_ioda_setup_iommu_api();
3161 pnv_pci_ioda_create_dbgfs();
3163 pnv_pci_enable_bridges();
3166 pnv_eeh_post_init();
3171 * Returns the alignment for I/O or memory windows for P2P
3172 * bridges. That actually depends on how PEs are segmented.
3173 * For now, we return I/O or M32 segment size for PE sensitive
3174 * P2P bridges. Otherwise, the default values (4KiB for I/O,
3175 * 1MiB for memory) will be returned.
3177 * The current PCI bus might be put into one PE, which was
3178 * create against the parent PCI bridge. For that case, we
3179 * needn't enlarge the alignment so that we can save some
3182 static resource_size_t
pnv_pci_window_alignment(struct pci_bus
*bus
,
3185 struct pci_dev
*bridge
;
3186 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3187 struct pnv_phb
*phb
= hose
->private_data
;
3188 int num_pci_bridges
= 0;
3192 if (pci_pcie_type(bridge
) == PCI_EXP_TYPE_PCI_BRIDGE
) {
3194 if (num_pci_bridges
>= 2)
3198 bridge
= bridge
->bus
->self
;
3202 * We fall back to M32 if M64 isn't supported. We enforce the M64
3203 * alignment for any 64-bit resource, PCIe doesn't care and
3204 * bridges only do 64-bit prefetchable anyway.
3206 if (phb
->ioda
.m64_segsize
&& pnv_pci_is_m64_flags(type
))
3207 return phb
->ioda
.m64_segsize
;
3208 if (type
& IORESOURCE_MEM
)
3209 return phb
->ioda
.m32_segsize
;
3211 return phb
->ioda
.io_segsize
;
3215 * We are updating root port or the upstream port of the
3216 * bridge behind the root port with PHB's windows in order
3217 * to accommodate the changes on required resources during
3218 * PCI (slot) hotplug, which is connected to either root
3219 * port or the downstream ports of PCIe switch behind the
3222 static void pnv_pci_fixup_bridge_resources(struct pci_bus
*bus
,
3225 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3226 struct pnv_phb
*phb
= hose
->private_data
;
3227 struct pci_dev
*bridge
= bus
->self
;
3228 struct resource
*r
, *w
;
3229 bool msi_region
= false;
3232 /* Check if we need apply fixup to the bridge's windows */
3233 if (!pci_is_root_bus(bridge
->bus
) &&
3234 !pci_is_root_bus(bridge
->bus
->self
->bus
))
3237 /* Fixup the resources */
3238 for (i
= 0; i
< PCI_BRIDGE_RESOURCE_NUM
; i
++) {
3239 r
= &bridge
->resource
[PCI_BRIDGE_RESOURCES
+ i
];
3240 if (!r
->flags
|| !r
->parent
)
3244 if (r
->flags
& type
& IORESOURCE_IO
)
3245 w
= &hose
->io_resource
;
3246 else if (pnv_pci_is_m64(phb
, r
) &&
3247 (type
& IORESOURCE_PREFETCH
) &&
3248 phb
->ioda
.m64_segsize
)
3249 w
= &hose
->mem_resources
[1];
3250 else if (r
->flags
& type
& IORESOURCE_MEM
) {
3251 w
= &hose
->mem_resources
[0];
3255 r
->start
= w
->start
;
3258 /* The 64KB 32-bits MSI region shouldn't be included in
3259 * the 32-bits bridge window. Otherwise, we can see strange
3260 * issues. One of them is EEH error observed on Garrison.
3262 * Exclude top 1MB region which is the minimal alignment of
3263 * 32-bits bridge window.
3272 static void pnv_pci_setup_bridge(struct pci_bus
*bus
, unsigned long type
)
3274 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3275 struct pnv_phb
*phb
= hose
->private_data
;
3276 struct pci_dev
*bridge
= bus
->self
;
3277 struct pnv_ioda_pe
*pe
;
3278 bool all
= (pci_pcie_type(bridge
) == PCI_EXP_TYPE_PCI_BRIDGE
);
3280 /* Extend bridge's windows if necessary */
3281 pnv_pci_fixup_bridge_resources(bus
, type
);
3283 /* The PE for root bus should be realized before any one else */
3284 if (!phb
->ioda
.root_pe_populated
) {
3285 pe
= pnv_ioda_setup_bus_PE(phb
->hose
->bus
, false);
3287 phb
->ioda
.root_pe_idx
= pe
->pe_number
;
3288 phb
->ioda
.root_pe_populated
= true;
3292 /* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3293 if (list_empty(&bus
->devices
))
3296 /* Reserve PEs according to used M64 resources */
3297 pnv_ioda_reserve_m64_pe(bus
, NULL
, all
);
3300 * Assign PE. We might run here because of partial hotplug.
3301 * For the case, we just pick up the existing PE and should
3302 * not allocate resources again.
3304 pe
= pnv_ioda_setup_bus_PE(bus
, all
);
3308 pnv_ioda_setup_pe_seg(pe
);
3309 switch (phb
->type
) {
3311 pnv_pci_ioda1_setup_dma_pe(phb
, pe
);
3314 pnv_pci_ioda2_setup_dma_pe(phb
, pe
);
3317 pr_warn("%s: No DMA for PHB#%x (type %d)\n",
3318 __func__
, phb
->hose
->global_number
, phb
->type
);
3322 static resource_size_t
pnv_pci_default_alignment(void)
3327 #ifdef CONFIG_PCI_IOV
3328 static resource_size_t
pnv_pci_iov_resource_alignment(struct pci_dev
*pdev
,
3331 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3332 struct pnv_phb
*phb
= hose
->private_data
;
3333 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
3334 resource_size_t align
;
3337 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3338 * SR-IOV. While from hardware perspective, the range mapped by M64
3339 * BAR should be size aligned.
3341 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3342 * powernv-specific hardware restriction is gone. But if just use the
3343 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3344 * in one segment of M64 #15, which introduces the PE conflict between
3345 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3348 * This function returns the total IOV BAR size if M64 BAR is in
3349 * Shared PE mode or just VF BAR size if not.
3350 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3351 * M64 segment size if IOV BAR size is less.
3353 align
= pci_iov_resource_size(pdev
, resno
);
3354 if (!pdn
->vfs_expanded
)
3356 if (pdn
->m64_single_mode
)
3357 return max(align
, (resource_size_t
)phb
->ioda
.m64_segsize
);
3359 return pdn
->vfs_expanded
* align
;
3361 #endif /* CONFIG_PCI_IOV */
3363 /* Prevent enabling devices for which we couldn't properly
3366 static bool pnv_pci_enable_device_hook(struct pci_dev
*dev
)
3368 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
3369 struct pnv_phb
*phb
= hose
->private_data
;
3372 /* The function is probably called while the PEs have
3373 * not be created yet. For example, resource reassignment
3374 * during PCI probe period. We just skip the check if
3377 if (!phb
->initialized
)
3380 pdn
= pci_get_pdn(dev
);
3381 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
3387 static long pnv_pci_ioda1_unset_window(struct iommu_table_group
*table_group
,
3390 struct pnv_ioda_pe
*pe
= container_of(table_group
,
3391 struct pnv_ioda_pe
, table_group
);
3392 struct pnv_phb
*phb
= pe
->phb
;
3396 pe_info(pe
, "Removing DMA window #%d\n", num
);
3397 for (idx
= 0; idx
< phb
->ioda
.dma32_count
; idx
++) {
3398 if (phb
->ioda
.dma32_segmap
[idx
] != pe
->pe_number
)
3401 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
, pe
->pe_number
,
3402 idx
, 0, 0ul, 0ul, 0ul);
3403 if (rc
!= OPAL_SUCCESS
) {
3404 pe_warn(pe
, "Failure %ld unmapping DMA32 segment#%d\n",
3409 phb
->ioda
.dma32_segmap
[idx
] = IODA_INVALID_PE
;
3412 pnv_pci_unlink_table_and_group(table_group
->tables
[num
], table_group
);
3413 return OPAL_SUCCESS
;
3416 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe
*pe
)
3418 unsigned int weight
= pnv_pci_ioda_pe_dma_weight(pe
);
3419 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
3425 rc
= pnv_pci_ioda1_unset_window(&pe
->table_group
, 0);
3426 if (rc
!= OPAL_SUCCESS
)
3429 pnv_pci_p7ioc_tce_invalidate(tbl
, tbl
->it_offset
, tbl
->it_size
, false);
3430 if (pe
->table_group
.group
) {
3431 iommu_group_put(pe
->table_group
.group
);
3432 WARN_ON(pe
->table_group
.group
);
3435 free_pages(tbl
->it_base
, get_order(tbl
->it_size
<< 3));
3436 iommu_tce_table_put(tbl
);
3439 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe
*pe
)
3441 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
3442 unsigned int weight
= pnv_pci_ioda_pe_dma_weight(pe
);
3443 #ifdef CONFIG_IOMMU_API
3450 #ifdef CONFIG_IOMMU_API
3451 rc
= pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
3453 pe_warn(pe
, "OPAL error %ld release DMA window\n", rc
);
3456 pnv_pci_ioda2_set_bypass(pe
, false);
3457 if (pe
->table_group
.group
) {
3458 iommu_group_put(pe
->table_group
.group
);
3459 WARN_ON(pe
->table_group
.group
);
3462 iommu_tce_table_put(tbl
);
3465 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe
*pe
,
3469 struct pnv_phb
*phb
= pe
->phb
;
3473 for (idx
= 0; idx
< phb
->ioda
.total_pe_num
; idx
++) {
3474 if (map
[idx
] != pe
->pe_number
)
3477 if (win
== OPAL_M64_WINDOW_TYPE
)
3478 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3479 phb
->ioda
.reserved_pe_idx
, win
,
3480 idx
/ PNV_IODA1_M64_SEGS
,
3481 idx
% PNV_IODA1_M64_SEGS
);
3483 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3484 phb
->ioda
.reserved_pe_idx
, win
, 0, idx
);
3486 if (rc
!= OPAL_SUCCESS
)
3487 pe_warn(pe
, "Error %ld unmapping (%d) segment#%d\n",
3490 map
[idx
] = IODA_INVALID_PE
;
3494 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe
*pe
)
3496 struct pnv_phb
*phb
= pe
->phb
;
3498 if (phb
->type
== PNV_PHB_IODA1
) {
3499 pnv_ioda_free_pe_seg(pe
, OPAL_IO_WINDOW_TYPE
,
3500 phb
->ioda
.io_segmap
);
3501 pnv_ioda_free_pe_seg(pe
, OPAL_M32_WINDOW_TYPE
,
3502 phb
->ioda
.m32_segmap
);
3503 pnv_ioda_free_pe_seg(pe
, OPAL_M64_WINDOW_TYPE
,
3504 phb
->ioda
.m64_segmap
);
3505 } else if (phb
->type
== PNV_PHB_IODA2
) {
3506 pnv_ioda_free_pe_seg(pe
, OPAL_M32_WINDOW_TYPE
,
3507 phb
->ioda
.m32_segmap
);
3511 static void pnv_ioda_release_pe(struct pnv_ioda_pe
*pe
)
3513 struct pnv_phb
*phb
= pe
->phb
;
3514 struct pnv_ioda_pe
*slave
, *tmp
;
3516 list_del(&pe
->list
);
3517 switch (phb
->type
) {
3519 pnv_pci_ioda1_release_pe_dma(pe
);
3522 pnv_pci_ioda2_release_pe_dma(pe
);
3528 pnv_ioda_release_pe_seg(pe
);
3529 pnv_ioda_deconfigure_pe(pe
->phb
, pe
);
3531 /* Release slave PEs in the compound PE */
3532 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
3533 list_for_each_entry_safe(slave
, tmp
, &pe
->slaves
, list
) {
3534 list_del(&slave
->list
);
3535 pnv_ioda_free_pe(slave
);
3540 * The PE for root bus can be removed because of hotplug in EEH
3541 * recovery for fenced PHB error. We need to mark the PE dead so
3542 * that it can be populated again in PCI hot add path. The PE
3543 * shouldn't be destroyed as it's the global reserved resource.
3545 if (phb
->ioda
.root_pe_populated
&&
3546 phb
->ioda
.root_pe_idx
== pe
->pe_number
)
3547 phb
->ioda
.root_pe_populated
= false;
3549 pnv_ioda_free_pe(pe
);
3552 static void pnv_pci_release_device(struct pci_dev
*pdev
)
3554 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3555 struct pnv_phb
*phb
= hose
->private_data
;
3556 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
3557 struct pnv_ioda_pe
*pe
;
3559 if (pdev
->is_virtfn
)
3562 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
3566 * PCI hotplug can happen as part of EEH error recovery. The @pdn
3567 * isn't removed and added afterwards in this scenario. We should
3568 * set the PE number in @pdn to an invalid one. Otherwise, the PE's
3569 * device count is decreased on removing devices while failing to
3570 * be increased on adding devices. It leads to unbalanced PE's device
3571 * count and eventually make normal PCI hotplug path broken.
3573 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
3574 pdn
->pe_number
= IODA_INVALID_PE
;
3576 WARN_ON(--pe
->device_count
< 0);
3577 if (pe
->device_count
== 0)
3578 pnv_ioda_release_pe(pe
);
3581 static void pnv_npu_disable_device(struct pci_dev
*pdev
)
3583 struct eeh_dev
*edev
= pci_dev_to_eeh_dev(pdev
);
3584 struct eeh_pe
*eehpe
= edev
? edev
->pe
: NULL
;
3586 if (eehpe
&& eeh_ops
&& eeh_ops
->reset
)
3587 eeh_ops
->reset(eehpe
, EEH_RESET_HOT
);
3590 static void pnv_pci_ioda_shutdown(struct pci_controller
*hose
)
3592 struct pnv_phb
*phb
= hose
->private_data
;
3594 opal_pci_reset(phb
->opal_id
, OPAL_RESET_PCI_IODA_TABLE
,
3598 static const struct pci_controller_ops pnv_pci_ioda_controller_ops
= {
3599 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3600 .dma_bus_setup
= pnv_pci_dma_bus_setup
,
3601 .iommu_bypass_supported
= pnv_pci_ioda_iommu_bypass_supported
,
3602 .setup_msi_irqs
= pnv_setup_msi_irqs
,
3603 .teardown_msi_irqs
= pnv_teardown_msi_irqs
,
3604 .enable_device_hook
= pnv_pci_enable_device_hook
,
3605 .release_device
= pnv_pci_release_device
,
3606 .window_alignment
= pnv_pci_window_alignment
,
3607 .setup_bridge
= pnv_pci_setup_bridge
,
3608 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3609 .shutdown
= pnv_pci_ioda_shutdown
,
3612 static const struct pci_controller_ops pnv_npu_ioda_controller_ops
= {
3613 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3614 .setup_msi_irqs
= pnv_setup_msi_irqs
,
3615 .teardown_msi_irqs
= pnv_teardown_msi_irqs
,
3616 .enable_device_hook
= pnv_pci_enable_device_hook
,
3617 .window_alignment
= pnv_pci_window_alignment
,
3618 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3619 .shutdown
= pnv_pci_ioda_shutdown
,
3620 .disable_device
= pnv_npu_disable_device
,
3623 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops
= {
3624 .enable_device_hook
= pnv_pci_enable_device_hook
,
3625 .window_alignment
= pnv_pci_window_alignment
,
3626 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3627 .shutdown
= pnv_pci_ioda_shutdown
,
3630 static void __init
pnv_pci_init_ioda_phb(struct device_node
*np
,
3631 u64 hub_id
, int ioda_type
)
3633 struct pci_controller
*hose
;
3634 struct pnv_phb
*phb
;
3635 unsigned long size
, m64map_off
, m32map_off
, pemap_off
;
3636 unsigned long iomap_off
= 0, dma32map_off
= 0;
3638 const __be64
*prop64
;
3639 const __be32
*prop32
;
3646 if (!of_device_is_available(np
))
3649 pr_info("Initializing %s PHB (%pOF)\n", pnv_phb_names
[ioda_type
], np
);
3651 prop64
= of_get_property(np
, "ibm,opal-phbid", NULL
);
3653 pr_err(" Missing \"ibm,opal-phbid\" property !\n");
3656 phb_id
= be64_to_cpup(prop64
);
3657 pr_debug(" PHB-ID : 0x%016llx\n", phb_id
);
3659 phb
= memblock_alloc(sizeof(*phb
), SMP_CACHE_BYTES
);
3661 panic("%s: Failed to allocate %zu bytes\n", __func__
,
3664 /* Allocate PCI controller */
3665 phb
->hose
= hose
= pcibios_alloc_controller(np
);
3667 pr_err(" Can't allocate PCI controller for %pOF\n",
3669 memblock_free(__pa(phb
), sizeof(struct pnv_phb
));
3673 spin_lock_init(&phb
->lock
);
3674 prop32
= of_get_property(np
, "bus-range", &len
);
3675 if (prop32
&& len
== 8) {
3676 hose
->first_busno
= be32_to_cpu(prop32
[0]);
3677 hose
->last_busno
= be32_to_cpu(prop32
[1]);
3679 pr_warn(" Broken <bus-range> on %pOF\n", np
);
3680 hose
->first_busno
= 0;
3681 hose
->last_busno
= 0xff;
3683 hose
->private_data
= phb
;
3684 phb
->hub_id
= hub_id
;
3685 phb
->opal_id
= phb_id
;
3686 phb
->type
= ioda_type
;
3687 mutex_init(&phb
->ioda
.pe_alloc_mutex
);
3689 /* Detect specific models for error handling */
3690 if (of_device_is_compatible(np
, "ibm,p7ioc-pciex"))
3691 phb
->model
= PNV_PHB_MODEL_P7IOC
;
3692 else if (of_device_is_compatible(np
, "ibm,power8-pciex"))
3693 phb
->model
= PNV_PHB_MODEL_PHB3
;
3694 else if (of_device_is_compatible(np
, "ibm,power8-npu-pciex"))
3695 phb
->model
= PNV_PHB_MODEL_NPU
;
3696 else if (of_device_is_compatible(np
, "ibm,power9-npu-pciex"))
3697 phb
->model
= PNV_PHB_MODEL_NPU2
;
3699 phb
->model
= PNV_PHB_MODEL_UNKNOWN
;
3701 /* Initialize diagnostic data buffer */
3702 prop32
= of_get_property(np
, "ibm,phb-diag-data-size", NULL
);
3704 phb
->diag_data_size
= be32_to_cpup(prop32
);
3706 phb
->diag_data_size
= PNV_PCI_DIAG_BUF_SIZE
;
3708 phb
->diag_data
= memblock_alloc(phb
->diag_data_size
, SMP_CACHE_BYTES
);
3709 if (!phb
->diag_data
)
3710 panic("%s: Failed to allocate %u bytes\n", __func__
,
3711 phb
->diag_data_size
);
3713 /* Parse 32-bit and IO ranges (if any) */
3714 pci_process_bridge_OF_ranges(hose
, np
, !hose
->global_number
);
3717 if (!of_address_to_resource(np
, 0, &r
)) {
3718 phb
->regs_phys
= r
.start
;
3719 phb
->regs
= ioremap(r
.start
, resource_size(&r
));
3720 if (phb
->regs
== NULL
)
3721 pr_err(" Failed to map registers !\n");
3724 /* Initialize more IODA stuff */
3725 phb
->ioda
.total_pe_num
= 1;
3726 prop32
= of_get_property(np
, "ibm,opal-num-pes", NULL
);
3728 phb
->ioda
.total_pe_num
= be32_to_cpup(prop32
);
3729 prop32
= of_get_property(np
, "ibm,opal-reserved-pe", NULL
);
3731 phb
->ioda
.reserved_pe_idx
= be32_to_cpup(prop32
);
3733 /* Invalidate RID to PE# mapping */
3734 for (segno
= 0; segno
< ARRAY_SIZE(phb
->ioda
.pe_rmap
); segno
++)
3735 phb
->ioda
.pe_rmap
[segno
] = IODA_INVALID_PE
;
3737 /* Parse 64-bit MMIO range */
3738 pnv_ioda_parse_m64_window(phb
);
3740 phb
->ioda
.m32_size
= resource_size(&hose
->mem_resources
[0]);
3741 /* FW Has already off top 64k of M32 space (MSI space) */
3742 phb
->ioda
.m32_size
+= 0x10000;
3744 phb
->ioda
.m32_segsize
= phb
->ioda
.m32_size
/ phb
->ioda
.total_pe_num
;
3745 phb
->ioda
.m32_pci_base
= hose
->mem_resources
[0].start
- hose
->mem_offset
[0];
3746 phb
->ioda
.io_size
= hose
->pci_io_size
;
3747 phb
->ioda
.io_segsize
= phb
->ioda
.io_size
/ phb
->ioda
.total_pe_num
;
3748 phb
->ioda
.io_pci_base
= 0; /* XXX calculate this ? */
3750 /* Calculate how many 32-bit TCE segments we have */
3751 phb
->ioda
.dma32_count
= phb
->ioda
.m32_pci_base
/
3752 PNV_IODA1_DMA32_SEGSIZE
;
3754 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3755 size
= _ALIGN_UP(max_t(unsigned, phb
->ioda
.total_pe_num
, 8) / 8,
3756 sizeof(unsigned long));
3758 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.m64_segmap
[0]);
3760 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.m32_segmap
[0]);
3761 if (phb
->type
== PNV_PHB_IODA1
) {
3763 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.io_segmap
[0]);
3764 dma32map_off
= size
;
3765 size
+= phb
->ioda
.dma32_count
*
3766 sizeof(phb
->ioda
.dma32_segmap
[0]);
3769 size
+= phb
->ioda
.total_pe_num
* sizeof(struct pnv_ioda_pe
);
3770 aux
= memblock_alloc(size
, SMP_CACHE_BYTES
);
3772 panic("%s: Failed to allocate %lu bytes\n", __func__
, size
);
3773 phb
->ioda
.pe_alloc
= aux
;
3774 phb
->ioda
.m64_segmap
= aux
+ m64map_off
;
3775 phb
->ioda
.m32_segmap
= aux
+ m32map_off
;
3776 for (segno
= 0; segno
< phb
->ioda
.total_pe_num
; segno
++) {
3777 phb
->ioda
.m64_segmap
[segno
] = IODA_INVALID_PE
;
3778 phb
->ioda
.m32_segmap
[segno
] = IODA_INVALID_PE
;
3780 if (phb
->type
== PNV_PHB_IODA1
) {
3781 phb
->ioda
.io_segmap
= aux
+ iomap_off
;
3782 for (segno
= 0; segno
< phb
->ioda
.total_pe_num
; segno
++)
3783 phb
->ioda
.io_segmap
[segno
] = IODA_INVALID_PE
;
3785 phb
->ioda
.dma32_segmap
= aux
+ dma32map_off
;
3786 for (segno
= 0; segno
< phb
->ioda
.dma32_count
; segno
++)
3787 phb
->ioda
.dma32_segmap
[segno
] = IODA_INVALID_PE
;
3789 phb
->ioda
.pe_array
= aux
+ pemap_off
;
3792 * Choose PE number for root bus, which shouldn't have
3793 * M64 resources consumed by its child devices. To pick
3794 * the PE number adjacent to the reserved one if possible.
3796 pnv_ioda_reserve_pe(phb
, phb
->ioda
.reserved_pe_idx
);
3797 if (phb
->ioda
.reserved_pe_idx
== 0) {
3798 phb
->ioda
.root_pe_idx
= 1;
3799 pnv_ioda_reserve_pe(phb
, phb
->ioda
.root_pe_idx
);
3800 } else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1)) {
3801 phb
->ioda
.root_pe_idx
= phb
->ioda
.reserved_pe_idx
- 1;
3802 pnv_ioda_reserve_pe(phb
, phb
->ioda
.root_pe_idx
);
3804 phb
->ioda
.root_pe_idx
= IODA_INVALID_PE
;
3807 INIT_LIST_HEAD(&phb
->ioda
.pe_list
);
3808 mutex_init(&phb
->ioda
.pe_list_mutex
);
3810 /* Calculate how many 32-bit TCE segments we have */
3811 phb
->ioda
.dma32_count
= phb
->ioda
.m32_pci_base
/
3812 PNV_IODA1_DMA32_SEGSIZE
;
3814 #if 0 /* We should really do that ... */
3815 rc
= opal_pci_set_phb_mem_window(opal
->phb_id
,
3818 starting_real_address
,
3819 starting_pci_address
,
3823 pr_info(" %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3824 phb
->ioda
.total_pe_num
, phb
->ioda
.reserved_pe_idx
,
3825 phb
->ioda
.m32_size
, phb
->ioda
.m32_segsize
);
3826 if (phb
->ioda
.m64_size
)
3827 pr_info(" M64: 0x%lx [segment=0x%lx]\n",
3828 phb
->ioda
.m64_size
, phb
->ioda
.m64_segsize
);
3829 if (phb
->ioda
.io_size
)
3830 pr_info(" IO: 0x%x [segment=0x%x]\n",
3831 phb
->ioda
.io_size
, phb
->ioda
.io_segsize
);
3834 phb
->hose
->ops
= &pnv_pci_ops
;
3835 phb
->get_pe_state
= pnv_ioda_get_pe_state
;
3836 phb
->freeze_pe
= pnv_ioda_freeze_pe
;
3837 phb
->unfreeze_pe
= pnv_ioda_unfreeze_pe
;
3839 /* Setup MSI support */
3840 pnv_pci_init_ioda_msis(phb
);
3843 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3844 * to let the PCI core do resource assignment. It's supposed
3845 * that the PCI core will do correct I/O and MMIO alignment
3846 * for the P2P bridge bars so that each PCI bus (excluding
3847 * the child P2P bridges) can form individual PE.
3849 ppc_md
.pcibios_fixup
= pnv_pci_ioda_fixup
;
3851 switch (phb
->type
) {
3852 case PNV_PHB_NPU_NVLINK
:
3853 hose
->controller_ops
= pnv_npu_ioda_controller_ops
;
3855 case PNV_PHB_NPU_OCAPI
:
3856 hose
->controller_ops
= pnv_npu_ocapi_ioda_controller_ops
;
3859 phb
->dma_dev_setup
= pnv_pci_ioda_dma_dev_setup
;
3860 hose
->controller_ops
= pnv_pci_ioda_controller_ops
;
3863 ppc_md
.pcibios_default_alignment
= pnv_pci_default_alignment
;
3865 #ifdef CONFIG_PCI_IOV
3866 ppc_md
.pcibios_fixup_sriov
= pnv_pci_ioda_fixup_iov_resources
;
3867 ppc_md
.pcibios_iov_resource_alignment
= pnv_pci_iov_resource_alignment
;
3868 ppc_md
.pcibios_sriov_enable
= pnv_pcibios_sriov_enable
;
3869 ppc_md
.pcibios_sriov_disable
= pnv_pcibios_sriov_disable
;
3872 pci_add_flags(PCI_REASSIGN_ALL_RSRC
);
3874 /* Reset IODA tables to a clean state */
3875 rc
= opal_pci_reset(phb_id
, OPAL_RESET_PCI_IODA_TABLE
, OPAL_ASSERT_RESET
);
3877 pr_warn(" OPAL Error %ld performing IODA table reset !\n", rc
);
3880 * If we're running in kdump kernel, the previous kernel never
3881 * shutdown PCI devices correctly. We already got IODA table
3882 * cleaned out. So we have to issue PHB reset to stop all PCI
3883 * transactions from previous kernel. The ppc_pci_reset_phbs
3884 * kernel parameter will force this reset too. Additionally,
3885 * if the IODA reset above failed then use a bigger hammer.
3886 * This can happen if we get a PHB fatal error in very early
3889 if (is_kdump_kernel() || pci_reset_phbs
|| rc
) {
3890 pr_info(" Issue PHB reset ...\n");
3891 pnv_eeh_phb_reset(hose
, EEH_RESET_FUNDAMENTAL
);
3892 pnv_eeh_phb_reset(hose
, EEH_RESET_DEACTIVATE
);
3895 /* Remove M64 resource if we can't configure it successfully */
3896 if (!phb
->init_m64
|| phb
->init_m64(phb
))
3897 hose
->mem_resources
[1].flags
= 0;
3900 void __init
pnv_pci_init_ioda2_phb(struct device_node
*np
)
3902 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_IODA2
);
3905 void __init
pnv_pci_init_npu_phb(struct device_node
*np
)
3907 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_NPU_NVLINK
);
3910 void __init
pnv_pci_init_npu2_opencapi_phb(struct device_node
*np
)
3912 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_NPU_OCAPI
);
3915 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev
*dev
)
3917 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
3918 struct pnv_phb
*phb
= hose
->private_data
;
3920 if (!machine_is(powernv
))
3923 if (phb
->type
== PNV_PHB_NPU_OCAPI
)
3924 dev
->cfg_size
= PCI_CFG_SPACE_EXP_SIZE
;
3926 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID
, PCI_ANY_ID
, pnv_npu2_opencapi_cfg_size_fixup
);
3928 void __init
pnv_pci_init_ioda_hub(struct device_node
*np
)
3930 struct device_node
*phbn
;
3931 const __be64
*prop64
;
3934 pr_info("Probing IODA IO-Hub %pOF\n", np
);
3936 prop64
= of_get_property(np
, "ibm,opal-hubid", NULL
);
3938 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3941 hub_id
= be64_to_cpup(prop64
);
3942 pr_devel(" HUB-ID : 0x%016llx\n", hub_id
);
3944 /* Count child PHBs */
3945 for_each_child_of_node(np
, phbn
) {
3946 /* Look for IODA1 PHBs */
3947 if (of_device_is_compatible(phbn
, "ibm,ioda-phb"))
3948 pnv_pci_init_ioda_phb(phbn
, hub_id
, PNV_PHB_IODA1
);