1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
7 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
8 * Copyright (C) 2006 Olof Johansson <olof@lixom.net>
10 * Dynamic DMA mapping support, pSeries-specific parts, both SMP and LPAR.
13 #include <linux/init.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
17 #include <linux/memblock.h>
18 #include <linux/spinlock.h>
19 #include <linux/string.h>
20 #include <linux/pci.h>
21 #include <linux/dma-mapping.h>
22 #include <linux/crash_dump.h>
23 #include <linux/memory.h>
25 #include <linux/of_address.h>
26 #include <linux/iommu.h>
27 #include <linux/rculist.h>
31 #include <asm/iommu.h>
32 #include <asm/pci-bridge.h>
33 #include <asm/machdep.h>
34 #include <asm/firmware.h>
36 #include <asm/ppc-pci.h>
38 #include <asm/mmzone.h>
39 #include <asm/plpar_wrappers.h>
44 DDW_QUERY_PE_DMA_WIN = 0,
45 DDW_CREATE_PE_DMA_WIN = 1,
46 DDW_REMOVE_PE_DMA_WIN = 2,
53 DDW_EXT_RESET_DMA_WIN = 1,
54 DDW_EXT_QUERY_OUT_SIZE = 2
57 static struct iommu_table *iommu_pseries_alloc_table(int node)
59 struct iommu_table *tbl;
61 tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL, node);
65 INIT_LIST_HEAD_RCU(&tbl->it_group_list);
66 kref_init(&tbl->it_kref);
70 static struct iommu_table_group *iommu_pseries_alloc_group(int node)
72 struct iommu_table_group *table_group;
74 table_group = kzalloc_node(sizeof(*table_group), GFP_KERNEL, node);
78 #ifdef CONFIG_IOMMU_API
79 table_group->ops = &spapr_tce_table_group_ops;
80 table_group->pgsizes = SZ_4K;
83 table_group->tables[0] = iommu_pseries_alloc_table(node);
84 if (table_group->tables[0])
91 static void iommu_pseries_free_group(struct iommu_table_group *table_group,
92 const char *node_name)
97 #ifdef CONFIG_IOMMU_API
98 if (table_group->group) {
99 iommu_group_put(table_group->group);
100 BUG_ON(table_group->group);
104 /* Default DMA window table is at index 0, while DDW at 1. SR-IOV
105 * adapters only have table on index 1.
107 if (table_group->tables[0])
108 iommu_tce_table_put(table_group->tables[0]);
110 if (table_group->tables[1])
111 iommu_tce_table_put(table_group->tables[1]);
116 static int tce_build_pSeries(struct iommu_table *tbl, long index,
117 long npages, unsigned long uaddr,
118 enum dma_data_direction direction,
124 const unsigned long tceshift = tbl->it_page_shift;
125 const unsigned long pagesize = IOMMU_PAGE_SIZE(tbl);
127 proto_tce = TCE_PCI_READ; // Read allowed
129 if (direction != DMA_TO_DEVICE)
130 proto_tce |= TCE_PCI_WRITE;
132 tcep = ((__be64 *)tbl->it_base) + index;
135 /* can't move this out since we might cross MEMBLOCK boundary */
136 rpn = __pa(uaddr) >> tceshift;
137 *tcep = cpu_to_be64(proto_tce | rpn << tceshift);
146 static void tce_free_pSeries(struct iommu_table *tbl, long index, long npages)
150 tcep = ((__be64 *)tbl->it_base) + index;
156 static unsigned long tce_get_pseries(struct iommu_table *tbl, long index)
160 tcep = ((__be64 *)tbl->it_base) + index;
162 return be64_to_cpu(*tcep);
165 static void tce_free_pSeriesLP(unsigned long liobn, long, long, long);
166 static void tce_freemulti_pSeriesLP(struct iommu_table*, long, long);
168 static int tce_build_pSeriesLP(unsigned long liobn, long tcenum, long tceshift,
169 long npages, unsigned long uaddr,
170 enum dma_data_direction direction,
177 long tcenum_start = tcenum, npages_start = npages;
179 rpn = __pa(uaddr) >> tceshift;
180 proto_tce = TCE_PCI_READ;
181 if (direction != DMA_TO_DEVICE)
182 proto_tce |= TCE_PCI_WRITE;
185 tce = proto_tce | rpn << tceshift;
186 rc = plpar_tce_put((u64)liobn, (u64)tcenum << tceshift, tce);
188 if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
190 tce_free_pSeriesLP(liobn, tcenum_start, tceshift,
191 (npages_start - (npages + 1)));
195 if (rc && printk_ratelimit()) {
196 printk("tce_build_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
197 printk("\tindex = 0x%llx\n", (u64)liobn);
198 printk("\ttcenum = 0x%llx\n", (u64)tcenum);
199 printk("\ttce val = 0x%llx\n", tce );
209 static DEFINE_PER_CPU(__be64 *, tce_page);
211 static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
212 long npages, unsigned long uaddr,
213 enum dma_data_direction direction,
221 long tcenum_start = tcenum, npages_start = npages;
224 const unsigned long tceshift = tbl->it_page_shift;
226 if ((npages == 1) || !firmware_has_feature(FW_FEATURE_PUT_TCE_IND)) {
227 return tce_build_pSeriesLP(tbl->it_index, tcenum,
228 tceshift, npages, uaddr,
232 local_irq_save(flags); /* to protect tcep and the page behind it */
234 tcep = __this_cpu_read(tce_page);
236 /* This is safe to do since interrupts are off when we're called
237 * from iommu_alloc{,_sg}()
240 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
241 /* If allocation fails, fall back to the loop implementation */
243 local_irq_restore(flags);
244 return tce_build_pSeriesLP(tbl->it_index, tcenum,
246 npages, uaddr, direction, attrs);
248 __this_cpu_write(tce_page, tcep);
251 rpn = __pa(uaddr) >> tceshift;
252 proto_tce = TCE_PCI_READ;
253 if (direction != DMA_TO_DEVICE)
254 proto_tce |= TCE_PCI_WRITE;
256 /* We can map max one pageful of TCEs at a time */
259 * Set up the page with TCE data, looping through and setting
262 limit = min_t(long, npages, 4096 / TCE_ENTRY_SIZE);
264 for (l = 0; l < limit; l++) {
265 tcep[l] = cpu_to_be64(proto_tce | rpn << tceshift);
269 rc = plpar_tce_put_indirect((u64)tbl->it_index,
270 (u64)tcenum << tceshift,
276 } while (npages > 0 && !rc);
278 local_irq_restore(flags);
280 if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
282 tce_freemulti_pSeriesLP(tbl, tcenum_start,
283 (npages_start - (npages + limit)));
287 if (rc && printk_ratelimit()) {
288 printk("tce_buildmulti_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
289 printk("\tindex = 0x%llx\n", (u64)tbl->it_index);
290 printk("\tnpages = 0x%llx\n", (u64)npages);
291 printk("\ttce[0] val = 0x%llx\n", tcep[0]);
297 static void tce_free_pSeriesLP(unsigned long liobn, long tcenum, long tceshift,
303 rc = plpar_tce_put((u64)liobn, (u64)tcenum << tceshift, 0);
305 if (rc && printk_ratelimit()) {
306 printk("tce_free_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
307 printk("\tindex = 0x%llx\n", (u64)liobn);
308 printk("\ttcenum = 0x%llx\n", (u64)tcenum);
317 static void tce_freemulti_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
320 long rpages = npages;
323 if (!firmware_has_feature(FW_FEATURE_STUFF_TCE))
324 return tce_free_pSeriesLP(tbl->it_index, tcenum,
325 tbl->it_page_shift, npages);
328 limit = min_t(unsigned long, rpages, 512);
330 rc = plpar_tce_stuff((u64)tbl->it_index,
331 (u64)tcenum << tbl->it_page_shift, 0, limit);
335 } while (rpages > 0 && !rc);
337 if (rc && printk_ratelimit()) {
338 printk("tce_freemulti_pSeriesLP: plpar_tce_stuff failed\n");
339 printk("\trc = %lld\n", rc);
340 printk("\tindex = 0x%llx\n", (u64)tbl->it_index);
341 printk("\tnpages = 0x%llx\n", (u64)npages);
346 static unsigned long tce_get_pSeriesLP(struct iommu_table *tbl, long tcenum)
349 unsigned long tce_ret;
351 rc = plpar_tce_get((u64)tbl->it_index,
352 (u64)tcenum << tbl->it_page_shift, &tce_ret);
354 if (rc && printk_ratelimit()) {
355 printk("tce_get_pSeriesLP: plpar_tce_get failed. rc=%lld\n", rc);
356 printk("\tindex = 0x%llx\n", (u64)tbl->it_index);
357 printk("\ttcenum = 0x%llx\n", (u64)tcenum);
364 /* this is compatible with cells for the device tree property */
365 struct dynamic_dma_window_prop {
366 __be32 liobn; /* tce table number */
367 __be64 dma_base; /* address hi,lo */
368 __be32 tce_shift; /* ilog2(tce_page_size) */
369 __be32 window_shift; /* ilog2(tce_window_size) */
373 struct device_node *device;
374 const struct dynamic_dma_window_prop *prop;
375 struct list_head list;
378 /* Dynamic DMA Window support */
379 struct ddw_query_response {
380 u32 windows_available;
381 u64 largest_available_block;
383 u32 migration_capable;
386 struct ddw_create_response {
392 static LIST_HEAD(dma_win_list);
393 /* prevents races between memory on/offline and window creation */
394 static DEFINE_SPINLOCK(dma_win_list_lock);
395 /* protects initializing window twice for same device */
396 static DEFINE_MUTEX(dma_win_init_mutex);
397 #define DIRECT64_PROPNAME "linux,direct64-ddr-window-info"
398 #define DMA64_PROPNAME "linux,dma64-ddr-window-info"
400 static int tce_clearrange_multi_pSeriesLP(unsigned long start_pfn,
401 unsigned long num_pfn, const void *arg)
403 const struct dynamic_dma_window_prop *maprange = arg;
405 u64 tce_size, num_tce, dma_offset, next;
409 tce_shift = be32_to_cpu(maprange->tce_shift);
410 tce_size = 1ULL << tce_shift;
411 next = start_pfn << PAGE_SHIFT;
412 num_tce = num_pfn << PAGE_SHIFT;
414 /* round back to the beginning of the tce page size */
415 num_tce += next & (tce_size - 1);
416 next &= ~(tce_size - 1);
418 /* covert to number of tces */
419 num_tce |= tce_size - 1;
420 num_tce >>= tce_shift;
424 * Set up the page with TCE data, looping through and setting
427 limit = min_t(long, num_tce, 512);
428 dma_offset = next + be64_to_cpu(maprange->dma_base);
430 rc = plpar_tce_stuff((u64)be32_to_cpu(maprange->liobn),
433 next += limit * tce_size;
435 } while (num_tce > 0 && !rc);
440 static int tce_setrange_multi_pSeriesLP(unsigned long start_pfn,
441 unsigned long num_pfn, const void *arg)
443 const struct dynamic_dma_window_prop *maprange = arg;
444 u64 tce_size, num_tce, dma_offset, next, proto_tce, liobn;
450 if (!firmware_has_feature(FW_FEATURE_PUT_TCE_IND)) {
451 unsigned long tceshift = be32_to_cpu(maprange->tce_shift);
452 unsigned long dmastart = (start_pfn << PAGE_SHIFT) +
453 be64_to_cpu(maprange->dma_base);
454 unsigned long tcenum = dmastart >> tceshift;
455 unsigned long npages = num_pfn << PAGE_SHIFT >> tceshift;
456 void *uaddr = __va(start_pfn << PAGE_SHIFT);
458 return tce_build_pSeriesLP(be32_to_cpu(maprange->liobn),
459 tcenum, tceshift, npages, (unsigned long) uaddr,
460 DMA_BIDIRECTIONAL, 0);
463 local_irq_disable(); /* to protect tcep and the page behind it */
464 tcep = __this_cpu_read(tce_page);
467 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
472 __this_cpu_write(tce_page, tcep);
475 proto_tce = TCE_PCI_READ | TCE_PCI_WRITE;
477 liobn = (u64)be32_to_cpu(maprange->liobn);
478 tce_shift = be32_to_cpu(maprange->tce_shift);
479 tce_size = 1ULL << tce_shift;
480 next = start_pfn << PAGE_SHIFT;
481 num_tce = num_pfn << PAGE_SHIFT;
483 /* round back to the beginning of the tce page size */
484 num_tce += next & (tce_size - 1);
485 next &= ~(tce_size - 1);
487 /* covert to number of tces */
488 num_tce |= tce_size - 1;
489 num_tce >>= tce_shift;
491 /* We can map max one pageful of TCEs at a time */
494 * Set up the page with TCE data, looping through and setting
497 limit = min_t(long, num_tce, 4096 / TCE_ENTRY_SIZE);
498 dma_offset = next + be64_to_cpu(maprange->dma_base);
500 for (l = 0; l < limit; l++) {
501 tcep[l] = cpu_to_be64(proto_tce | next);
505 rc = plpar_tce_put_indirect(liobn,
511 } while (num_tce > 0 && !rc);
513 /* error cleanup: caller will clear whole range */
519 static int tce_setrange_multi_pSeriesLP_walk(unsigned long start_pfn,
520 unsigned long num_pfn, void *arg)
522 return tce_setrange_multi_pSeriesLP(start_pfn, num_pfn, arg);
525 static void iommu_table_setparms_common(struct iommu_table *tbl, unsigned long busno,
526 unsigned long liobn, unsigned long win_addr,
527 unsigned long window_size, unsigned long page_shift,
528 void *base, struct iommu_table_ops *table_ops)
530 tbl->it_busno = busno;
531 tbl->it_index = liobn;
532 tbl->it_offset = win_addr >> page_shift;
533 tbl->it_size = window_size >> page_shift;
534 tbl->it_page_shift = page_shift;
535 tbl->it_base = (unsigned long)base;
536 tbl->it_blocksize = 16;
537 tbl->it_type = TCE_PCI;
538 tbl->it_ops = table_ops;
541 struct iommu_table_ops iommu_table_pseries_ops;
543 static void iommu_table_setparms(struct pci_controller *phb,
544 struct device_node *dn,
545 struct iommu_table *tbl)
547 struct device_node *node;
548 const unsigned long *basep;
551 /* Test if we are going over 2GB of DMA space */
552 if (phb->dma_window_base_cur + phb->dma_window_size > SZ_2G) {
553 udbg_printf("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
554 panic("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
558 basep = of_get_property(node, "linux,tce-base", NULL);
559 sizep = of_get_property(node, "linux,tce-size", NULL);
560 if (basep == NULL || sizep == NULL) {
561 printk(KERN_ERR "PCI_DMA: iommu_table_setparms: %pOF has "
562 "missing tce entries !\n", dn);
566 iommu_table_setparms_common(tbl, phb->bus->number, 0, phb->dma_window_base_cur,
567 phb->dma_window_size, IOMMU_PAGE_SHIFT_4K,
568 __va(*basep), &iommu_table_pseries_ops);
570 if (!is_kdump_kernel())
571 memset((void *)tbl->it_base, 0, *sizep);
573 phb->dma_window_base_cur += phb->dma_window_size;
576 struct iommu_table_ops iommu_table_lpar_multi_ops;
579 * iommu_table_setparms_lpar
581 * Function: On pSeries LPAR systems, return TCE table info, given a pci bus.
583 static void iommu_table_setparms_lpar(struct pci_controller *phb,
584 struct device_node *dn,
585 struct iommu_table *tbl,
586 struct iommu_table_group *table_group,
587 const __be32 *dma_window)
589 unsigned long offset, size, liobn;
591 of_parse_dma_window(dn, dma_window, &liobn, &offset, &size);
593 iommu_table_setparms_common(tbl, phb->bus->number, liobn, offset, size, IOMMU_PAGE_SHIFT_4K, NULL,
594 &iommu_table_lpar_multi_ops);
597 table_group->tce32_start = offset;
598 table_group->tce32_size = size;
601 struct iommu_table_ops iommu_table_pseries_ops = {
602 .set = tce_build_pSeries,
603 .clear = tce_free_pSeries,
604 .get = tce_get_pseries
607 static void pci_dma_bus_setup_pSeries(struct pci_bus *bus)
609 struct device_node *dn;
610 struct iommu_table *tbl;
611 struct device_node *isa_dn, *isa_dn_orig;
612 struct device_node *tmp;
616 dn = pci_bus_to_OF_node(bus);
618 pr_debug("pci_dma_bus_setup_pSeries: setting up bus %pOF\n", dn);
621 /* This is not a root bus, any setup will be done for the
622 * device-side of the bridge in iommu_dev_setup_pSeries().
628 /* Check if the ISA bus on the system is under
631 isa_dn = isa_dn_orig = of_find_node_by_type(NULL, "isa");
633 while (isa_dn && isa_dn != dn)
634 isa_dn = isa_dn->parent;
636 of_node_put(isa_dn_orig);
638 /* Count number of direct PCI children of the PHB. */
639 for (children = 0, tmp = dn->child; tmp; tmp = tmp->sibling)
642 pr_debug("Children: %d\n", children);
644 /* Calculate amount of DMA window per slot. Each window must be
645 * a power of two (due to pci_alloc_consistent requirements).
647 * Keep 256MB aside for PHBs with ISA.
651 /* No ISA/IDE - just set window size and return */
652 pci->phb->dma_window_size = 0x80000000ul; /* To be divided */
654 while (pci->phb->dma_window_size * children > 0x80000000ul)
655 pci->phb->dma_window_size >>= 1;
656 pr_debug("No ISA/IDE, window size is 0x%llx\n",
657 pci->phb->dma_window_size);
658 pci->phb->dma_window_base_cur = 0;
663 /* If we have ISA, then we probably have an IDE
664 * controller too. Allocate a 128MB table but
665 * skip the first 128MB to avoid stepping on ISA
668 pci->phb->dma_window_size = 0x8000000ul;
669 pci->phb->dma_window_base_cur = 0x8000000ul;
671 pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
672 tbl = pci->table_group->tables[0];
674 iommu_table_setparms(pci->phb, dn, tbl);
676 if (!iommu_init_table(tbl, pci->phb->node, 0, 0))
677 panic("Failed to initialize iommu table");
679 /* Divide the rest (1.75GB) among the children */
680 pci->phb->dma_window_size = 0x80000000ul;
681 while (pci->phb->dma_window_size * children > 0x70000000ul)
682 pci->phb->dma_window_size >>= 1;
684 pr_debug("ISA/IDE, window size is 0x%llx\n", pci->phb->dma_window_size);
687 #ifdef CONFIG_IOMMU_API
688 static int tce_exchange_pseries(struct iommu_table *tbl, long index, unsigned
689 long *tce, enum dma_data_direction *direction)
692 unsigned long ioba = (unsigned long) index << tbl->it_page_shift;
693 unsigned long flags, oldtce = 0;
694 u64 proto_tce = iommu_direction_to_tce_perm(*direction);
695 unsigned long newtce = *tce | proto_tce;
697 spin_lock_irqsave(&tbl->large_pool.lock, flags);
699 rc = plpar_tce_get((u64)tbl->it_index, ioba, &oldtce);
701 rc = plpar_tce_put((u64)tbl->it_index, ioba, newtce);
704 *direction = iommu_tce_direction(oldtce);
705 *tce = oldtce & ~(TCE_PCI_READ | TCE_PCI_WRITE);
708 spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
714 struct iommu_table_ops iommu_table_lpar_multi_ops = {
715 .set = tce_buildmulti_pSeriesLP,
716 #ifdef CONFIG_IOMMU_API
717 .xchg_no_kill = tce_exchange_pseries,
719 .clear = tce_freemulti_pSeriesLP,
720 .get = tce_get_pSeriesLP
724 * Find nearest ibm,dma-window (default DMA window) or direct DMA window or
725 * dynamic 64bit DMA window, walking up the device tree.
727 static struct device_node *pci_dma_find(struct device_node *dn,
728 const __be32 **dma_window)
730 const __be32 *dw = NULL;
732 for ( ; dn && PCI_DN(dn); dn = dn->parent) {
733 dw = of_get_property(dn, "ibm,dma-window", NULL);
739 dw = of_get_property(dn, DIRECT64_PROPNAME, NULL);
742 dw = of_get_property(dn, DMA64_PROPNAME, NULL);
750 static void pci_dma_bus_setup_pSeriesLP(struct pci_bus *bus)
752 struct iommu_table *tbl;
753 struct device_node *dn, *pdn;
755 const __be32 *dma_window = NULL;
757 dn = pci_bus_to_OF_node(bus);
759 pr_debug("pci_dma_bus_setup_pSeriesLP: setting up bus %pOF\n",
762 pdn = pci_dma_find(dn, &dma_window);
764 if (dma_window == NULL)
765 pr_debug(" no ibm,dma-window property !\n");
769 pr_debug(" parent is %pOF, iommu_table: 0x%p\n",
770 pdn, ppci->table_group);
772 if (!ppci->table_group) {
773 ppci->table_group = iommu_pseries_alloc_group(ppci->phb->node);
774 tbl = ppci->table_group->tables[0];
776 iommu_table_setparms_lpar(ppci->phb, pdn, tbl,
777 ppci->table_group, dma_window);
779 if (!iommu_init_table(tbl, ppci->phb->node, 0, 0))
780 panic("Failed to initialize iommu table");
782 iommu_register_group(ppci->table_group,
783 pci_domain_nr(bus), 0);
784 pr_debug(" created table: %p\n", ppci->table_group);
789 static void pci_dma_dev_setup_pSeries(struct pci_dev *dev)
791 struct device_node *dn;
792 struct iommu_table *tbl;
794 pr_debug("pci_dma_dev_setup_pSeries: %s\n", pci_name(dev));
796 dn = dev->dev.of_node;
798 /* If we're the direct child of a root bus, then we need to allocate
799 * an iommu table ourselves. The bus setup code should have setup
800 * the window sizes already.
802 if (!dev->bus->self) {
803 struct pci_controller *phb = PCI_DN(dn)->phb;
805 pr_debug(" --> first child, no bridge. Allocating iommu table.\n");
806 PCI_DN(dn)->table_group = iommu_pseries_alloc_group(phb->node);
807 tbl = PCI_DN(dn)->table_group->tables[0];
808 iommu_table_setparms(phb, dn, tbl);
810 if (!iommu_init_table(tbl, phb->node, 0, 0))
811 panic("Failed to initialize iommu table");
813 set_iommu_table_base(&dev->dev, tbl);
817 /* If this device is further down the bus tree, search upwards until
818 * an already allocated iommu table is found and use that.
821 while (dn && PCI_DN(dn) && PCI_DN(dn)->table_group == NULL)
824 if (dn && PCI_DN(dn))
825 set_iommu_table_base(&dev->dev,
826 PCI_DN(dn)->table_group->tables[0]);
828 printk(KERN_WARNING "iommu: Device %s has no iommu table\n",
832 static int __read_mostly disable_ddw;
834 static int __init disable_ddw_setup(char *str)
837 printk(KERN_INFO "ppc iommu: disabling ddw.\n");
842 early_param("disable_ddw", disable_ddw_setup);
844 static void clean_dma_window(struct device_node *np, struct dynamic_dma_window_prop *dwp)
848 ret = tce_clearrange_multi_pSeriesLP(0,
849 1ULL << (be32_to_cpu(dwp->window_shift) - PAGE_SHIFT), dwp);
851 pr_warn("%pOF failed to clear tces in window.\n",
854 pr_debug("%pOF successfully cleared tces in window.\n",
859 * Call only if DMA window is clean.
861 static void __remove_dma_window(struct device_node *np, u32 *ddw_avail, u64 liobn)
865 ret = rtas_call(ddw_avail[DDW_REMOVE_PE_DMA_WIN], 1, 1, NULL, liobn);
867 pr_warn("%pOF: failed to remove DMA window: rtas returned "
868 "%d to ibm,remove-pe-dma-window(%x) %llx\n",
869 np, ret, ddw_avail[DDW_REMOVE_PE_DMA_WIN], liobn);
871 pr_debug("%pOF: successfully removed DMA window: rtas returned "
872 "%d to ibm,remove-pe-dma-window(%x) %llx\n",
873 np, ret, ddw_avail[DDW_REMOVE_PE_DMA_WIN], liobn);
876 static void remove_dma_window(struct device_node *np, u32 *ddw_avail,
877 struct property *win)
879 struct dynamic_dma_window_prop *dwp;
883 liobn = (u64)be32_to_cpu(dwp->liobn);
885 clean_dma_window(np, dwp);
886 __remove_dma_window(np, ddw_avail, liobn);
889 static int remove_ddw(struct device_node *np, bool remove_prop, const char *win_name)
891 struct property *win;
892 u32 ddw_avail[DDW_APPLICABLE_SIZE];
895 win = of_find_property(np, win_name, NULL);
899 ret = of_property_read_u32_array(np, "ibm,ddw-applicable",
900 &ddw_avail[0], DDW_APPLICABLE_SIZE);
905 if (win->length >= sizeof(struct dynamic_dma_window_prop))
906 remove_dma_window(np, ddw_avail, win);
911 ret = of_remove_property(np, win);
913 pr_warn("%pOF: failed to remove DMA window property: %d\n",
918 static bool find_existing_ddw(struct device_node *pdn, u64 *dma_addr, int *window_shift)
920 struct dma_win *window;
921 const struct dynamic_dma_window_prop *dma64;
924 spin_lock(&dma_win_list_lock);
925 /* check if we already created a window and dupe that config if so */
926 list_for_each_entry(window, &dma_win_list, list) {
927 if (window->device == pdn) {
928 dma64 = window->prop;
929 *dma_addr = be64_to_cpu(dma64->dma_base);
930 *window_shift = be32_to_cpu(dma64->window_shift);
935 spin_unlock(&dma_win_list_lock);
940 static struct dma_win *ddw_list_new_entry(struct device_node *pdn,
941 const struct dynamic_dma_window_prop *dma64)
943 struct dma_win *window;
945 window = kzalloc(sizeof(*window), GFP_KERNEL);
949 window->device = pdn;
950 window->prop = dma64;
955 static void find_existing_ddw_windows_named(const char *name)
958 struct device_node *pdn;
959 struct dma_win *window;
960 const struct dynamic_dma_window_prop *dma64;
962 for_each_node_with_property(pdn, name) {
963 dma64 = of_get_property(pdn, name, &len);
964 if (!dma64 || len < sizeof(*dma64)) {
965 remove_ddw(pdn, true, name);
969 window = ddw_list_new_entry(pdn, dma64);
975 spin_lock(&dma_win_list_lock);
976 list_add(&window->list, &dma_win_list);
977 spin_unlock(&dma_win_list_lock);
981 static int find_existing_ddw_windows(void)
983 if (!firmware_has_feature(FW_FEATURE_LPAR))
986 find_existing_ddw_windows_named(DIRECT64_PROPNAME);
987 find_existing_ddw_windows_named(DMA64_PROPNAME);
991 machine_arch_initcall(pseries, find_existing_ddw_windows);
994 * ddw_read_ext - Get the value of an DDW extension
995 * @np: device node from which the extension value is to be read.
996 * @extnum: index number of the extension.
997 * @value: pointer to return value, modified when extension is available.
999 * Checks if "ibm,ddw-extensions" exists for this node, and get the value
1000 * on index 'extnum'.
1001 * It can be used only to check if a property exists, passing value == NULL.
1004 * 0 if extension successfully read
1005 * -EINVAL if the "ibm,ddw-extensions" does not exist,
1006 * -ENODATA if "ibm,ddw-extensions" does not have a value, and
1007 * -EOVERFLOW if "ibm,ddw-extensions" does not contain this extension.
1009 static inline int ddw_read_ext(const struct device_node *np, int extnum,
1012 static const char propname[] = "ibm,ddw-extensions";
1016 ret = of_property_read_u32_index(np, propname, DDW_EXT_SIZE, &count);
1026 return of_property_read_u32_index(np, propname, extnum, value);
1029 static int query_ddw(struct pci_dev *dev, const u32 *ddw_avail,
1030 struct ddw_query_response *query,
1031 struct device_node *parent)
1033 struct device_node *dn;
1035 u32 cfg_addr, ext_query, query_out[5];
1040 * From LoPAR level 2.8, "ibm,ddw-extensions" index 3 can rule how many
1041 * output parameters ibm,query-pe-dma-windows will have, ranging from
1044 ret = ddw_read_ext(parent, DDW_EXT_QUERY_OUT_SIZE, &ext_query);
1045 if (!ret && ext_query == 1)
1051 * Get the config address and phb buid of the PE window.
1052 * Rely on eeh to retrieve this for us.
1053 * Retrieve them from the pci device, not the node with the
1054 * dma-window property
1056 dn = pci_device_to_OF_node(dev);
1058 buid = pdn->phb->buid;
1059 cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
1061 ret = rtas_call(ddw_avail[DDW_QUERY_PE_DMA_WIN], 3, out_sz, query_out,
1062 cfg_addr, BUID_HI(buid), BUID_LO(buid));
1066 query->windows_available = query_out[0];
1067 query->largest_available_block = query_out[1];
1068 query->page_size = query_out[2];
1069 query->migration_capable = query_out[3];
1072 query->windows_available = query_out[0];
1073 query->largest_available_block = ((u64)query_out[1] << 32) |
1075 query->page_size = query_out[3];
1076 query->migration_capable = query_out[4];
1080 dev_info(&dev->dev, "ibm,query-pe-dma-windows(%x) %x %x %x returned %d, lb=%llx ps=%x wn=%d\n",
1081 ddw_avail[DDW_QUERY_PE_DMA_WIN], cfg_addr, BUID_HI(buid),
1082 BUID_LO(buid), ret, query->largest_available_block,
1083 query->page_size, query->windows_available);
1088 static int create_ddw(struct pci_dev *dev, const u32 *ddw_avail,
1089 struct ddw_create_response *create, int page_shift,
1092 struct device_node *dn;
1099 * Get the config address and phb buid of the PE window.
1100 * Rely on eeh to retrieve this for us.
1101 * Retrieve them from the pci device, not the node with the
1102 * dma-window property
1104 dn = pci_device_to_OF_node(dev);
1106 buid = pdn->phb->buid;
1107 cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
1110 /* extra outputs are LIOBN and dma-addr (hi, lo) */
1111 ret = rtas_call(ddw_avail[DDW_CREATE_PE_DMA_WIN], 5, 4,
1112 (u32 *)create, cfg_addr, BUID_HI(buid),
1113 BUID_LO(buid), page_shift, window_shift);
1114 } while (rtas_busy_delay(ret));
1116 "ibm,create-pe-dma-window(%x) %x %x %x %x %x returned %d "
1117 "(liobn = 0x%x starting addr = %x %x)\n",
1118 ddw_avail[DDW_CREATE_PE_DMA_WIN], cfg_addr, BUID_HI(buid),
1119 BUID_LO(buid), page_shift, window_shift, ret, create->liobn,
1120 create->addr_hi, create->addr_lo);
1125 struct failed_ddw_pdn {
1126 struct device_node *pdn;
1127 struct list_head list;
1130 static LIST_HEAD(failed_ddw_pdn_list);
1132 static phys_addr_t ddw_memory_hotplug_max(void)
1134 resource_size_t max_addr = memory_hotplug_max();
1135 struct device_node *memory;
1137 for_each_node_by_type(memory, "memory") {
1138 struct resource res;
1140 if (of_address_to_resource(memory, 0, &res))
1143 max_addr = max_t(resource_size_t, max_addr, res.end + 1);
1150 * Platforms supporting the DDW option starting with LoPAR level 2.7 implement
1151 * ibm,ddw-extensions, which carries the rtas token for
1152 * ibm,reset-pe-dma-windows.
1153 * That rtas-call can be used to restore the default DMA window for the device.
1155 static void reset_dma_window(struct pci_dev *dev, struct device_node *par_dn)
1158 u32 cfg_addr, reset_dma_win;
1160 struct device_node *dn;
1163 ret = ddw_read_ext(par_dn, DDW_EXT_RESET_DMA_WIN, &reset_dma_win);
1167 dn = pci_device_to_OF_node(dev);
1169 buid = pdn->phb->buid;
1170 cfg_addr = (pdn->busno << 16) | (pdn->devfn << 8);
1172 ret = rtas_call(reset_dma_win, 3, 1, NULL, cfg_addr, BUID_HI(buid),
1176 "ibm,reset-pe-dma-windows(%x) %x %x %x returned %d ",
1177 reset_dma_win, cfg_addr, BUID_HI(buid), BUID_LO(buid),
1181 /* Return largest page shift based on "IO Page Sizes" output of ibm,query-pe-dma-window. */
1182 static int iommu_get_page_shift(u32 query_page_size)
1184 /* Supported IO page-sizes according to LoPAR, note that 2M is out of order */
1185 const int shift[] = {
1186 __builtin_ctzll(SZ_4K), __builtin_ctzll(SZ_64K), __builtin_ctzll(SZ_16M),
1187 __builtin_ctzll(SZ_32M), __builtin_ctzll(SZ_64M), __builtin_ctzll(SZ_128M),
1188 __builtin_ctzll(SZ_256M), __builtin_ctzll(SZ_16G), __builtin_ctzll(SZ_2M)
1191 int i = ARRAY_SIZE(shift) - 1;
1195 * On LoPAR, ibm,query-pe-dma-window outputs "IO Page Sizes" using a bit field:
1196 * - bit 31 means 4k pages are supported,
1197 * - bit 30 means 64k pages are supported, and so on.
1198 * Larger pagesizes map more memory with the same amount of TCEs, so start probing them.
1200 for (; i >= 0 ; i--) {
1201 if (query_page_size & (1 << i))
1202 ret = max(ret, shift[i]);
1208 static struct property *ddw_property_create(const char *propname, u32 liobn, u64 dma_addr,
1209 u32 page_shift, u32 window_shift)
1211 struct dynamic_dma_window_prop *ddwprop;
1212 struct property *win64;
1214 win64 = kzalloc(sizeof(*win64), GFP_KERNEL);
1218 win64->name = kstrdup(propname, GFP_KERNEL);
1219 ddwprop = kzalloc(sizeof(*ddwprop), GFP_KERNEL);
1220 win64->value = ddwprop;
1221 win64->length = sizeof(*ddwprop);
1222 if (!win64->name || !win64->value) {
1224 kfree(win64->value);
1229 ddwprop->liobn = cpu_to_be32(liobn);
1230 ddwprop->dma_base = cpu_to_be64(dma_addr);
1231 ddwprop->tce_shift = cpu_to_be32(page_shift);
1232 ddwprop->window_shift = cpu_to_be32(window_shift);
1238 * If the PE supports dynamic dma windows, and there is space for a table
1239 * that can map all pages in a linear offset, then setup such a table,
1240 * and record the dma-offset in the struct device.
1242 * dev: the pci device we are checking
1243 * pdn: the parent pe node with the ibm,dma_window property
1244 * Future: also check if we can remap the base window for our base page size
1246 * returns true if can map all pages (direct mapping), false otherwise..
1248 static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
1251 int max_ram_len = order_base_2(ddw_memory_hotplug_max());
1252 struct ddw_query_response query;
1253 struct ddw_create_response create;
1256 const char *win_name;
1257 struct device_node *dn;
1258 u32 ddw_avail[DDW_APPLICABLE_SIZE];
1259 struct dma_win *window;
1260 struct property *win64;
1261 struct failed_ddw_pdn *fpdn;
1262 bool default_win_removed = false, direct_mapping = false;
1264 struct pci_dn *pci = PCI_DN(pdn);
1265 struct property *default_win = NULL;
1267 dn = of_find_node_by_type(NULL, "ibm,pmemory");
1268 pmem_present = dn != NULL;
1271 mutex_lock(&dma_win_init_mutex);
1273 if (find_existing_ddw(pdn, &dev->dev.archdata.dma_offset, &len)) {
1274 direct_mapping = (len >= max_ram_len);
1279 * If we already went through this for a previous function of
1280 * the same device and failed, we don't want to muck with the
1281 * DMA window again, as it will race with in-flight operations
1282 * and can lead to EEHs. The above mutex protects access to the
1285 list_for_each_entry(fpdn, &failed_ddw_pdn_list, list) {
1286 if (fpdn->pdn == pdn)
1291 * the ibm,ddw-applicable property holds the tokens for:
1292 * ibm,query-pe-dma-window
1293 * ibm,create-pe-dma-window
1294 * ibm,remove-pe-dma-window
1295 * for the given node in that order.
1296 * the property is actually in the parent, not the PE
1298 ret = of_property_read_u32_array(pdn, "ibm,ddw-applicable",
1299 &ddw_avail[0], DDW_APPLICABLE_SIZE);
1304 * Query if there is a second window of size to map the
1305 * whole partition. Query returns number of windows, largest
1306 * block assigned to PE (partition endpoint), and two bitmasks
1307 * of page sizes: supported and supported for migrate-dma.
1309 dn = pci_device_to_OF_node(dev);
1310 ret = query_ddw(dev, ddw_avail, &query, pdn);
1315 * If there is no window available, remove the default DMA window,
1316 * if it's present. This will make all the resources available to the
1318 * If anything fails after this, we need to restore it, so also check
1319 * for extensions presence.
1321 if (query.windows_available == 0) {
1324 /* DDW + IOMMU on single window may fail if there is any allocation */
1325 if (iommu_table_in_use(pci->table_group->tables[0])) {
1326 dev_warn(&dev->dev, "current IOMMU table in use, can't be replaced.\n");
1330 default_win = of_find_property(pdn, "ibm,dma-window", NULL);
1334 reset_win_ext = ddw_read_ext(pdn, DDW_EXT_RESET_DMA_WIN, NULL);
1338 remove_dma_window(pdn, ddw_avail, default_win);
1339 default_win_removed = true;
1341 /* Query again, to check if the window is available */
1342 ret = query_ddw(dev, ddw_avail, &query, pdn);
1346 if (query.windows_available == 0) {
1347 /* no windows are available for this device. */
1348 dev_dbg(&dev->dev, "no free dynamic windows");
1353 page_shift = iommu_get_page_shift(query.page_size);
1355 dev_dbg(&dev->dev, "no supported page size in mask %x",
1362 * The "ibm,pmemory" can appear anywhere in the address space.
1363 * Assuming it is still backed by page structs, try MAX_PHYSMEM_BITS
1364 * for the upper limit and fallback to max RAM otherwise but this
1365 * disables device::dma_ops_bypass.
1369 if (query.largest_available_block >=
1370 (1ULL << (MAX_PHYSMEM_BITS - page_shift)))
1371 len = MAX_PHYSMEM_BITS;
1373 dev_info(&dev->dev, "Skipping ibm,pmemory");
1376 /* check if the available block * number of ptes will map everything */
1377 if (query.largest_available_block < (1ULL << (len - page_shift))) {
1379 "can't map partition max 0x%llx with %llu %llu-sized pages\n",
1381 query.largest_available_block,
1382 1ULL << page_shift);
1384 len = order_base_2(query.largest_available_block << page_shift);
1385 win_name = DMA64_PROPNAME;
1387 direct_mapping = !default_win_removed ||
1388 (len == MAX_PHYSMEM_BITS) ||
1389 (!pmem_present && (len == max_ram_len));
1390 win_name = direct_mapping ? DIRECT64_PROPNAME : DMA64_PROPNAME;
1393 ret = create_ddw(dev, ddw_avail, &create, page_shift, len);
1397 dev_dbg(&dev->dev, "created tce table LIOBN 0x%x for %pOF\n",
1400 win_addr = ((u64)create.addr_hi << 32) | create.addr_lo;
1401 win64 = ddw_property_create(win_name, create.liobn, win_addr, page_shift, len);
1405 "couldn't allocate property, property name, or value\n");
1406 goto out_remove_win;
1409 ret = of_add_property(pdn, win64);
1411 dev_err(&dev->dev, "unable to add DMA window property for %pOF: %d",
1416 window = ddw_list_new_entry(pdn, win64->value);
1420 if (direct_mapping) {
1421 /* DDW maps the whole partition, so enable direct DMA mapping */
1422 ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
1423 win64->value, tce_setrange_multi_pSeriesLP_walk);
1425 dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
1428 /* Make sure to clean DDW if any TCE was set*/
1429 clean_dma_window(pdn, win64->value);
1433 struct iommu_table *newtbl;
1435 unsigned long start = 0, end = 0;
1437 for (i = 0; i < ARRAY_SIZE(pci->phb->mem_resources); i++) {
1438 const unsigned long mask = IORESOURCE_MEM_64 | IORESOURCE_MEM;
1440 /* Look for MMIO32 */
1441 if ((pci->phb->mem_resources[i].flags & mask) == IORESOURCE_MEM) {
1442 start = pci->phb->mem_resources[i].start;
1443 end = pci->phb->mem_resources[i].end;
1448 /* New table for using DDW instead of the default DMA window */
1449 newtbl = iommu_pseries_alloc_table(pci->phb->node);
1451 dev_dbg(&dev->dev, "couldn't create new IOMMU table\n");
1455 iommu_table_setparms_common(newtbl, pci->phb->bus->number, create.liobn, win_addr,
1456 1UL << len, page_shift, NULL, &iommu_table_lpar_multi_ops);
1457 iommu_init_table(newtbl, pci->phb->node, start, end);
1459 pci->table_group->tables[1] = newtbl;
1461 set_iommu_table_base(&dev->dev, newtbl);
1464 if (default_win_removed) {
1465 iommu_tce_table_put(pci->table_group->tables[0]);
1466 pci->table_group->tables[0] = NULL;
1468 /* default_win is valid here because default_win_removed == true */
1469 of_remove_property(pdn, default_win);
1470 dev_info(&dev->dev, "Removed default DMA window for %pOF\n", pdn);
1473 spin_lock(&dma_win_list_lock);
1474 list_add(&window->list, &dma_win_list);
1475 spin_unlock(&dma_win_list_lock);
1477 dev->dev.archdata.dma_offset = win_addr;
1484 of_remove_property(pdn, win64);
1488 kfree(win64->value);
1492 /* DDW is clean, so it's ok to call this directly. */
1493 __remove_dma_window(pdn, ddw_avail, create.liobn);
1496 if (default_win_removed)
1497 reset_dma_window(dev, pdn);
1499 fpdn = kzalloc(sizeof(*fpdn), GFP_KERNEL);
1503 list_add(&fpdn->list, &failed_ddw_pdn_list);
1506 mutex_unlock(&dma_win_init_mutex);
1509 * If we have persistent memory and the window size is only as big
1510 * as RAM, then we failed to create a window to cover persistent
1511 * memory and need to set the DMA limit.
1513 if (pmem_present && direct_mapping && len == max_ram_len)
1514 dev->dev.bus_dma_limit = dev->dev.archdata.dma_offset + (1ULL << len);
1516 return direct_mapping;
1519 static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
1521 struct device_node *pdn, *dn;
1522 struct iommu_table *tbl;
1523 const __be32 *dma_window = NULL;
1526 pr_debug("pci_dma_dev_setup_pSeriesLP: %s\n", pci_name(dev));
1528 /* dev setup for LPAR is a little tricky, since the device tree might
1529 * contain the dma-window properties per-device and not necessarily
1530 * for the bus. So we need to search upwards in the tree until we
1531 * either hit a dma-window property, OR find a parent with a table
1532 * already allocated.
1534 dn = pci_device_to_OF_node(dev);
1535 pr_debug(" node is %pOF\n", dn);
1537 pdn = pci_dma_find(dn, &dma_window);
1538 if (!pdn || !PCI_DN(pdn)) {
1539 printk(KERN_WARNING "pci_dma_dev_setup_pSeriesLP: "
1540 "no DMA window found for pci dev=%s dn=%pOF\n",
1544 pr_debug(" parent is %pOF\n", pdn);
1547 if (!pci->table_group) {
1548 pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
1549 tbl = pci->table_group->tables[0];
1550 iommu_table_setparms_lpar(pci->phb, pdn, tbl,
1551 pci->table_group, dma_window);
1553 iommu_init_table(tbl, pci->phb->node, 0, 0);
1554 iommu_register_group(pci->table_group,
1555 pci_domain_nr(pci->phb->bus), 0);
1556 pr_debug(" created table: %p\n", pci->table_group);
1558 pr_debug(" found DMA window, table: %p\n", pci->table_group);
1561 set_iommu_table_base(&dev->dev, pci->table_group->tables[0]);
1562 iommu_add_device(pci->table_group, &dev->dev);
1565 static bool iommu_bypass_supported_pSeriesLP(struct pci_dev *pdev, u64 dma_mask)
1567 struct device_node *dn = pci_device_to_OF_node(pdev), *pdn;
1569 /* only attempt to use a new window if 64-bit DMA is requested */
1570 if (dma_mask < DMA_BIT_MASK(64))
1573 dev_dbg(&pdev->dev, "node is %pOF\n", dn);
1576 * the device tree might contain the dma-window properties
1577 * per-device and not necessarily for the bus. So we need to
1578 * search upwards in the tree until we either hit a dma-window
1579 * property, OR find a parent with a table already allocated.
1581 pdn = pci_dma_find(dn, NULL);
1582 if (pdn && PCI_DN(pdn))
1583 return enable_ddw(pdev, pdn);
1588 static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
1591 struct dma_win *window;
1592 struct memory_notify *arg = data;
1596 case MEM_GOING_ONLINE:
1597 spin_lock(&dma_win_list_lock);
1598 list_for_each_entry(window, &dma_win_list, list) {
1599 ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
1600 arg->nr_pages, window->prop);
1603 spin_unlock(&dma_win_list_lock);
1605 case MEM_CANCEL_ONLINE:
1607 spin_lock(&dma_win_list_lock);
1608 list_for_each_entry(window, &dma_win_list, list) {
1609 ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
1610 arg->nr_pages, window->prop);
1613 spin_unlock(&dma_win_list_lock);
1618 if (ret && action != MEM_CANCEL_ONLINE)
1624 static struct notifier_block iommu_mem_nb = {
1625 .notifier_call = iommu_mem_notifier,
1628 static int iommu_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *data)
1630 int err = NOTIFY_OK;
1631 struct of_reconfig_data *rd = data;
1632 struct device_node *np = rd->dn;
1633 struct pci_dn *pci = PCI_DN(np);
1634 struct dma_win *window;
1637 case OF_RECONFIG_DETACH_NODE:
1639 * Removing the property will invoke the reconfig
1640 * notifier again, which causes dead-lock on the
1641 * read-write semaphore of the notifier chain. So
1642 * we have to remove the property when releasing
1645 if (remove_ddw(np, false, DIRECT64_PROPNAME))
1646 remove_ddw(np, false, DMA64_PROPNAME);
1648 if (pci && pci->table_group)
1649 iommu_pseries_free_group(pci->table_group,
1652 spin_lock(&dma_win_list_lock);
1653 list_for_each_entry(window, &dma_win_list, list) {
1654 if (window->device == np) {
1655 list_del(&window->list);
1660 spin_unlock(&dma_win_list_lock);
1669 static struct notifier_block iommu_reconfig_nb = {
1670 .notifier_call = iommu_reconfig_notifier,
1673 /* These are called very early. */
1674 void __init iommu_init_early_pSeries(void)
1676 if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
1679 if (firmware_has_feature(FW_FEATURE_LPAR)) {
1680 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeriesLP;
1681 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeriesLP;
1683 pseries_pci_controller_ops.iommu_bypass_supported =
1684 iommu_bypass_supported_pSeriesLP;
1686 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeries;
1687 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeries;
1691 of_reconfig_notifier_register(&iommu_reconfig_nb);
1692 register_memory_notifier(&iommu_mem_nb);
1694 set_pci_dma_ops(&dma_iommu_ops);
1697 static int __init disable_multitce(char *str)
1699 if (strcmp(str, "off") == 0 &&
1700 firmware_has_feature(FW_FEATURE_LPAR) &&
1701 (firmware_has_feature(FW_FEATURE_PUT_TCE_IND) ||
1702 firmware_has_feature(FW_FEATURE_STUFF_TCE))) {
1703 printk(KERN_INFO "Disabling MULTITCE firmware feature\n");
1704 powerpc_firmware_features &=
1705 ~(FW_FEATURE_PUT_TCE_IND | FW_FEATURE_STUFF_TCE);
1710 __setup("multitce=", disable_multitce);
1712 #ifdef CONFIG_SPAPR_TCE_IOMMU
1713 struct iommu_group *pSeries_pci_device_group(struct pci_controller *hose,
1714 struct pci_dev *pdev)
1716 struct device_node *pdn, *dn = pdev->dev.of_node;
1717 struct iommu_group *grp;
1720 pdn = pci_dma_find(dn, NULL);
1721 if (!pdn || !PCI_DN(pdn))
1722 return ERR_PTR(-ENODEV);
1725 if (!pci->table_group)
1726 return ERR_PTR(-ENODEV);
1728 grp = pci->table_group->group;
1730 return ERR_PTR(-ENODEV);
1732 return iommu_group_ref_get(grp);