2 * IOMMU implementation for Cell Broadband Processor Architecture
4 * (C) Copyright IBM Corporation 2006-2008
6 * Author: Jeremy Kerr <jk@ozlabs.org>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2, or (at your option)
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #include <linux/kernel.h>
26 #include <linux/init.h>
27 #include <linux/interrupt.h>
28 #include <linux/notifier.h>
30 #include <linux/of_platform.h>
31 #include <linux/lmb.h>
34 #include <asm/iommu.h>
35 #include <asm/machdep.h>
36 #include <asm/pci-bridge.h>
38 #include <asm/firmware.h>
39 #include <asm/cell-regs.h>
41 #include "interrupt.h"
43 /* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages
44 * instead of leaving them mapped to some dummy page. This can be
45 * enabled once the appropriate workarounds for spider bugs have
48 #define CELL_IOMMU_REAL_UNMAP
50 /* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of
51 * IO PTEs based on the transfer direction. That can be enabled
52 * once spider-net has been fixed to pass the correct direction
53 * to the DMA mapping functions
55 #define CELL_IOMMU_STRICT_PROTECTION
60 /* IOC mmap registers */
61 #define IOC_Reg_Size 0x2000
63 #define IOC_IOPT_CacheInvd 0x908
64 #define IOC_IOPT_CacheInvd_NE_Mask 0xffe0000000000000ul
65 #define IOC_IOPT_CacheInvd_IOPTE_Mask 0x000003fffffffff8ul
66 #define IOC_IOPT_CacheInvd_Busy 0x0000000000000001ul
68 #define IOC_IOST_Origin 0x918
69 #define IOC_IOST_Origin_E 0x8000000000000000ul
70 #define IOC_IOST_Origin_HW 0x0000000000000800ul
71 #define IOC_IOST_Origin_HL 0x0000000000000400ul
73 #define IOC_IO_ExcpStat 0x920
74 #define IOC_IO_ExcpStat_V 0x8000000000000000ul
75 #define IOC_IO_ExcpStat_SPF_Mask 0x6000000000000000ul
76 #define IOC_IO_ExcpStat_SPF_S 0x6000000000000000ul
77 #define IOC_IO_ExcpStat_SPF_P 0x2000000000000000ul
78 #define IOC_IO_ExcpStat_ADDR_Mask 0x00000007fffff000ul
79 #define IOC_IO_ExcpStat_RW_Mask 0x0000000000000800ul
80 #define IOC_IO_ExcpStat_IOID_Mask 0x00000000000007fful
82 #define IOC_IO_ExcpMask 0x928
83 #define IOC_IO_ExcpMask_SFE 0x4000000000000000ul
84 #define IOC_IO_ExcpMask_PFE 0x2000000000000000ul
86 #define IOC_IOCmd_Offset 0x1000
88 #define IOC_IOCmd_Cfg 0xc00
89 #define IOC_IOCmd_Cfg_TE 0x0000800000000000ul
92 /* Segment table entries */
93 #define IOSTE_V 0x8000000000000000ul /* valid */
94 #define IOSTE_H 0x4000000000000000ul /* cache hint */
95 #define IOSTE_PT_Base_RPN_Mask 0x3ffffffffffff000ul /* base RPN of IOPT */
96 #define IOSTE_NPPT_Mask 0x0000000000000fe0ul /* no. pages in IOPT */
97 #define IOSTE_PS_Mask 0x0000000000000007ul /* page size */
98 #define IOSTE_PS_4K 0x0000000000000001ul /* - 4kB */
99 #define IOSTE_PS_64K 0x0000000000000003ul /* - 64kB */
100 #define IOSTE_PS_1M 0x0000000000000005ul /* - 1MB */
101 #define IOSTE_PS_16M 0x0000000000000007ul /* - 16MB */
103 /* Page table entries */
104 #define IOPTE_PP_W 0x8000000000000000ul /* protection: write */
105 #define IOPTE_PP_R 0x4000000000000000ul /* protection: read */
106 #define IOPTE_M 0x2000000000000000ul /* coherency required */
107 #define IOPTE_SO_R 0x1000000000000000ul /* ordering: writes */
108 #define IOPTE_SO_RW 0x1800000000000000ul /* ordering: r & w */
109 #define IOPTE_RPN_Mask 0x07fffffffffff000ul /* RPN */
110 #define IOPTE_H 0x0000000000000800ul /* cache hint */
111 #define IOPTE_IOID_Mask 0x00000000000007fful /* ioid */
115 #define IO_SEGMENT_SHIFT 28
116 #define IO_PAGENO_BITS(shift) (IO_SEGMENT_SHIFT - (shift))
118 /* The high bit needs to be set on every DMA address */
119 #define SPIDER_DMA_OFFSET 0x80000000ul
121 struct iommu_window {
122 struct list_head list;
123 struct cbe_iommu *iommu;
124 unsigned long offset;
127 struct iommu_table table;
134 void __iomem *xlate_regs;
135 void __iomem *cmd_regs;
139 struct list_head windows;
142 /* Static array of iommus, one per node
143 * each contains a list of windows, keyed from dma_window property
144 * - on bus setup, look for a matching window, or create one
145 * - on dev setup, assign iommu_table ptr
147 static struct cbe_iommu iommus[NR_IOMMUS];
148 static int cbe_nr_iommus;
150 static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
157 reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
160 /* we can invalidate up to 1 << 11 PTEs at once */
161 n = min(n_ptes, 1l << 11);
162 val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask)
163 | (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask)
164 | IOC_IOPT_CacheInvd_Busy;
167 while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
175 static int tce_build_cell(struct iommu_table *tbl, long index, long npages,
176 unsigned long uaddr, enum dma_data_direction direction,
177 struct dma_attrs *attrs)
180 unsigned long *io_pte, base_pte;
181 struct iommu_window *window =
182 container_of(tbl, struct iommu_window, table);
184 /* implementing proper protection causes problems with the spidernet
185 * driver - check mapping directions later, but allow read & write by
187 #ifdef CELL_IOMMU_STRICT_PROTECTION
188 /* to avoid referencing a global, we use a trick here to setup the
189 * protection bit. "prot" is setup to be 3 fields of 4 bits apprended
190 * together for each of the 3 supported direction values. It is then
191 * shifted left so that the fields matching the desired direction
192 * lands on the appropriate bits, and other bits are masked out.
194 const unsigned long prot = 0xc48;
196 ((prot << (52 + 4 * direction)) & (IOPTE_PP_W | IOPTE_PP_R))
197 | IOPTE_M | IOPTE_SO_RW | (window->ioid & IOPTE_IOID_Mask);
199 base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW |
200 (window->ioid & IOPTE_IOID_Mask);
202 if (unlikely(dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs)))
203 base_pte &= ~IOPTE_SO_RW;
205 io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
207 for (i = 0; i < npages; i++, uaddr += IOMMU_PAGE_SIZE)
208 io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask);
212 invalidate_tce_cache(window->iommu, io_pte, npages);
214 pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
215 index, npages, direction, base_pte);
219 static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
223 unsigned long *io_pte, pte;
224 struct iommu_window *window =
225 container_of(tbl, struct iommu_window, table);
227 pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
229 #ifdef CELL_IOMMU_REAL_UNMAP
232 /* spider bridge does PCI reads after freeing - insert a mapping
233 * to a scratch page instead of an invalid entry */
234 pte = IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | __pa(window->iommu->pad_page)
235 | (window->ioid & IOPTE_IOID_Mask);
238 io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
240 for (i = 0; i < npages; i++)
245 invalidate_tce_cache(window->iommu, io_pte, npages);
248 static irqreturn_t ioc_interrupt(int irq, void *data)
250 unsigned long stat, spf;
251 struct cbe_iommu *iommu = data;
253 stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
254 spf = stat & IOC_IO_ExcpStat_SPF_Mask;
256 /* Might want to rate limit it */
257 printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat);
258 printk(KERN_ERR " V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n",
259 !!(stat & IOC_IO_ExcpStat_V),
260 (spf == IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
261 (spf == IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ',
262 (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write",
263 (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask));
264 printk(KERN_ERR " page=0x%016lx\n",
265 stat & IOC_IO_ExcpStat_ADDR_Mask);
267 /* clear interrupt */
268 stat &= ~IOC_IO_ExcpStat_V;
269 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat);
274 static int cell_iommu_find_ioc(int nid, unsigned long *base)
276 struct device_node *np;
281 /* First look for new style /be nodes */
282 for_each_node_by_name(np, "ioc") {
283 if (of_node_to_nid(np) != nid)
285 if (of_address_to_resource(np, 0, &r)) {
286 printk(KERN_ERR "iommu: can't get address for %s\n",
295 /* Ok, let's try the old way */
296 for_each_node_by_type(np, "cpu") {
297 const unsigned int *nidp;
298 const unsigned long *tmp;
300 nidp = of_get_property(np, "node-id", NULL);
301 if (nidp && *nidp == nid) {
302 tmp = of_get_property(np, "ioc-translation", NULL);
314 static void cell_iommu_setup_stab(struct cbe_iommu *iommu,
315 unsigned long dbase, unsigned long dsize,
316 unsigned long fbase, unsigned long fsize)
319 unsigned long segments, stab_size;
321 segments = max(dbase + dsize, fbase + fsize) >> IO_SEGMENT_SHIFT;
323 pr_debug("%s: iommu[%d]: segments: %lu\n",
324 __func__, iommu->nid, segments);
326 /* set up the segment table */
327 stab_size = segments * sizeof(unsigned long);
328 page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(stab_size));
330 iommu->stab = page_address(page);
331 memset(iommu->stab, 0, stab_size);
334 static unsigned long *cell_iommu_alloc_ptab(struct cbe_iommu *iommu,
335 unsigned long base, unsigned long size, unsigned long gap_base,
336 unsigned long gap_size, unsigned long page_shift)
340 unsigned long reg, segments, pages_per_segment, ptab_size,
341 n_pte_pages, start_seg, *ptab;
343 start_seg = base >> IO_SEGMENT_SHIFT;
344 segments = size >> IO_SEGMENT_SHIFT;
345 pages_per_segment = 1ull << IO_PAGENO_BITS(page_shift);
346 /* PTEs for each segment must start on a 4K bounday */
347 pages_per_segment = max(pages_per_segment,
348 (1 << 12) / sizeof(unsigned long));
350 ptab_size = segments * pages_per_segment * sizeof(unsigned long);
351 pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __func__,
352 iommu->nid, ptab_size, get_order(ptab_size));
353 page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size));
356 ptab = page_address(page);
357 memset(ptab, 0, ptab_size);
359 /* number of 4K pages needed for a page table */
360 n_pte_pages = (pages_per_segment * sizeof(unsigned long)) >> 12;
362 pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n",
363 __func__, iommu->nid, iommu->stab, ptab,
366 /* initialise the STEs */
367 reg = IOSTE_V | ((n_pte_pages - 1) << 5);
369 switch (page_shift) {
370 case 12: reg |= IOSTE_PS_4K; break;
371 case 16: reg |= IOSTE_PS_64K; break;
372 case 20: reg |= IOSTE_PS_1M; break;
373 case 24: reg |= IOSTE_PS_16M; break;
377 gap_base = gap_base >> IO_SEGMENT_SHIFT;
378 gap_size = gap_size >> IO_SEGMENT_SHIFT;
380 pr_debug("Setting up IOMMU stab:\n");
381 for (i = start_seg; i < (start_seg + segments); i++) {
382 if (i >= gap_base && i < (gap_base + gap_size)) {
383 pr_debug("\toverlap at %d, skipping\n", i);
386 iommu->stab[i] = reg | (__pa(ptab) + (n_pte_pages << 12) *
388 pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]);
394 static void cell_iommu_enable_hardware(struct cbe_iommu *iommu)
397 unsigned long reg, xlate_base;
400 if (cell_iommu_find_ioc(iommu->nid, &xlate_base))
401 panic("%s: missing IOC register mappings for node %d\n",
402 __func__, iommu->nid);
404 iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size);
405 iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset;
407 /* ensure that the STEs have updated */
410 /* setup interrupts for the iommu. */
411 reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
412 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat,
413 reg & ~IOC_IO_ExcpStat_V);
414 out_be64(iommu->xlate_regs + IOC_IO_ExcpMask,
415 IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE);
417 virq = irq_create_mapping(NULL,
418 IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT));
419 BUG_ON(virq == NO_IRQ);
421 ret = request_irq(virq, ioc_interrupt, IRQF_DISABLED,
425 /* set the IOC segment table origin register (and turn on the iommu) */
426 reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
427 out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
428 in_be64(iommu->xlate_regs + IOC_IOST_Origin);
430 /* turn on IO translation */
431 reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
432 out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
435 static void cell_iommu_setup_hardware(struct cbe_iommu *iommu,
436 unsigned long base, unsigned long size)
438 cell_iommu_setup_stab(iommu, base, size, 0, 0);
439 iommu->ptab = cell_iommu_alloc_ptab(iommu, base, size, 0, 0,
441 cell_iommu_enable_hardware(iommu);
444 #if 0/* Unused for now */
445 static struct iommu_window *find_window(struct cbe_iommu *iommu,
446 unsigned long offset, unsigned long size)
448 struct iommu_window *window;
450 /* todo: check for overlapping (but not equal) windows) */
452 list_for_each_entry(window, &(iommu->windows), list) {
453 if (window->offset == offset && window->size == size)
461 static inline u32 cell_iommu_get_ioid(struct device_node *np)
465 ioid = of_get_property(np, "ioid", NULL);
467 printk(KERN_WARNING "iommu: missing ioid for %s using 0\n",
475 static struct iommu_window * __init
476 cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
477 unsigned long offset, unsigned long size,
478 unsigned long pte_offset)
480 struct iommu_window *window;
484 ioid = cell_iommu_get_ioid(np);
486 window = kmalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
487 BUG_ON(window == NULL);
489 window->offset = offset;
492 window->iommu = iommu;
494 window->table.it_blocksize = 16;
495 window->table.it_base = (unsigned long)iommu->ptab;
496 window->table.it_index = iommu->nid;
497 window->table.it_offset = (offset >> IOMMU_PAGE_SHIFT) + pte_offset;
498 window->table.it_size = size >> IOMMU_PAGE_SHIFT;
500 iommu_init_table(&window->table, iommu->nid);
502 pr_debug("\tioid %d\n", window->ioid);
503 pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
504 pr_debug("\tbase 0x%016lx\n", window->table.it_base);
505 pr_debug("\toffset 0x%lx\n", window->table.it_offset);
506 pr_debug("\tsize %ld\n", window->table.it_size);
508 list_add(&window->list, &iommu->windows);
513 /* We need to map and reserve the first IOMMU page since it's used
514 * by the spider workaround. In theory, we only need to do that when
515 * running on spider but it doesn't really matter.
517 * This code also assumes that we have a window that starts at 0,
518 * which is the case on all spider based blades.
520 page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
522 iommu->pad_page = page_address(page);
523 clear_page(iommu->pad_page);
525 __set_bit(0, window->table.it_map);
526 tce_build_cell(&window->table, window->table.it_offset, 1,
527 (unsigned long)iommu->pad_page, DMA_TO_DEVICE, NULL);
528 window->table.it_hint = window->table.it_blocksize;
533 static struct cbe_iommu *cell_iommu_for_node(int nid)
537 for (i = 0; i < cbe_nr_iommus; i++)
538 if (iommus[i].nid == nid)
543 static unsigned long cell_dma_direct_offset;
545 static unsigned long dma_iommu_fixed_base;
547 /* iommu_fixed_is_weak is set if booted with iommu_fixed=weak */
548 static int iommu_fixed_is_weak;
550 static struct iommu_table *cell_get_iommu_table(struct device *dev)
552 struct iommu_window *window;
553 struct cbe_iommu *iommu;
554 struct dev_archdata *archdata = &dev->archdata;
556 /* Current implementation uses the first window available in that
557 * node's iommu. We -might- do something smarter later though it may
560 iommu = cell_iommu_for_node(dev_to_node(dev));
561 if (iommu == NULL || list_empty(&iommu->windows)) {
562 printk(KERN_ERR "iommu: missing iommu for %s (node %d)\n",
563 archdata->of_node ? archdata->of_node->full_name : "?",
567 window = list_entry(iommu->windows.next, struct iommu_window, list);
569 return &window->table;
572 /* A coherent allocation implies strong ordering */
574 static void *dma_fixed_alloc_coherent(struct device *dev, size_t size,
575 dma_addr_t *dma_handle, gfp_t flag)
577 if (iommu_fixed_is_weak)
578 return iommu_alloc_coherent(dev, cell_get_iommu_table(dev),
580 device_to_mask(dev), flag,
583 return dma_direct_ops.alloc_coherent(dev, size, dma_handle,
587 static void dma_fixed_free_coherent(struct device *dev, size_t size,
588 void *vaddr, dma_addr_t dma_handle)
590 if (iommu_fixed_is_weak)
591 iommu_free_coherent(cell_get_iommu_table(dev), size, vaddr,
594 dma_direct_ops.free_coherent(dev, size, vaddr, dma_handle);
597 static dma_addr_t dma_fixed_map_page(struct device *dev, struct page *page,
598 unsigned long offset, size_t size,
599 enum dma_data_direction direction,
600 struct dma_attrs *attrs)
602 if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
603 return dma_direct_ops.map_page(dev, page, offset, size,
606 return iommu_map_page(dev, cell_get_iommu_table(dev), page,
607 offset, size, device_to_mask(dev),
611 static void dma_fixed_unmap_page(struct device *dev, dma_addr_t dma_addr,
612 size_t size, enum dma_data_direction direction,
613 struct dma_attrs *attrs)
615 if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
616 dma_direct_ops.unmap_page(dev, dma_addr, size, direction,
619 iommu_unmap_page(cell_get_iommu_table(dev), dma_addr, size,
623 static int dma_fixed_map_sg(struct device *dev, struct scatterlist *sg,
624 int nents, enum dma_data_direction direction,
625 struct dma_attrs *attrs)
627 if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
628 return dma_direct_ops.map_sg(dev, sg, nents, direction, attrs);
630 return iommu_map_sg(dev, cell_get_iommu_table(dev), sg, nents,
631 device_to_mask(dev), direction, attrs);
634 static void dma_fixed_unmap_sg(struct device *dev, struct scatterlist *sg,
635 int nents, enum dma_data_direction direction,
636 struct dma_attrs *attrs)
638 if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
639 dma_direct_ops.unmap_sg(dev, sg, nents, direction, attrs);
641 iommu_unmap_sg(cell_get_iommu_table(dev), sg, nents, direction,
645 static int dma_fixed_dma_supported(struct device *dev, u64 mask)
647 return mask == DMA_64BIT_MASK;
650 static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask);
652 struct dma_mapping_ops dma_iommu_fixed_ops = {
653 .alloc_coherent = dma_fixed_alloc_coherent,
654 .free_coherent = dma_fixed_free_coherent,
655 .map_sg = dma_fixed_map_sg,
656 .unmap_sg = dma_fixed_unmap_sg,
657 .dma_supported = dma_fixed_dma_supported,
658 .set_dma_mask = dma_set_mask_and_switch,
659 .map_page = dma_fixed_map_page,
660 .unmap_page = dma_fixed_unmap_page,
663 static void cell_dma_dev_setup_fixed(struct device *dev);
665 static void cell_dma_dev_setup(struct device *dev)
667 struct dev_archdata *archdata = &dev->archdata;
669 /* Order is important here, these are not mutually exclusive */
670 if (get_dma_ops(dev) == &dma_iommu_fixed_ops)
671 cell_dma_dev_setup_fixed(dev);
672 else if (get_pci_dma_ops() == &dma_iommu_ops)
673 archdata->dma_data = cell_get_iommu_table(dev);
674 else if (get_pci_dma_ops() == &dma_direct_ops)
675 archdata->dma_data = (void *)cell_dma_direct_offset;
680 static void cell_pci_dma_dev_setup(struct pci_dev *dev)
682 cell_dma_dev_setup(&dev->dev);
685 static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
688 struct device *dev = data;
690 /* We are only intereted in device addition */
691 if (action != BUS_NOTIFY_ADD_DEVICE)
694 /* We use the PCI DMA ops */
695 dev->archdata.dma_ops = get_pci_dma_ops();
697 cell_dma_dev_setup(dev);
702 static struct notifier_block cell_of_bus_notifier = {
703 .notifier_call = cell_of_bus_notify
706 static int __init cell_iommu_get_window(struct device_node *np,
710 const void *dma_window;
713 /* Use ibm,dma-window if available, else, hard code ! */
714 dma_window = of_get_property(np, "ibm,dma-window", NULL);
715 if (dma_window == NULL) {
721 of_parse_dma_window(np, dma_window, &index, base, size);
725 static struct cbe_iommu * __init cell_iommu_alloc(struct device_node *np)
727 struct cbe_iommu *iommu;
731 nid = of_node_to_nid(np);
733 printk(KERN_ERR "iommu: failed to get node for %s\n",
737 pr_debug("iommu: setting up iommu for node %d (%s)\n",
740 /* XXX todo: If we can have multiple windows on the same IOMMU, which
741 * isn't the case today, we probably want here to check wether the
742 * iommu for that node is already setup.
743 * However, there might be issue with getting the size right so let's
744 * ignore that for now. We might want to completely get rid of the
745 * multiple window support since the cell iommu supports per-page ioids
748 if (cbe_nr_iommus >= NR_IOMMUS) {
749 printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n",
754 /* Init base fields */
759 snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
760 INIT_LIST_HEAD(&iommu->windows);
765 static void __init cell_iommu_init_one(struct device_node *np,
766 unsigned long offset)
768 struct cbe_iommu *iommu;
769 unsigned long base, size;
771 iommu = cell_iommu_alloc(np);
775 /* Obtain a window for it */
776 cell_iommu_get_window(np, &base, &size);
778 pr_debug("\ttranslating window 0x%lx...0x%lx\n",
779 base, base + size - 1);
781 /* Initialize the hardware */
782 cell_iommu_setup_hardware(iommu, base, size);
784 /* Setup the iommu_table */
785 cell_iommu_setup_window(iommu, np, base, size,
786 offset >> IOMMU_PAGE_SHIFT);
789 static void __init cell_disable_iommus(void)
792 unsigned long base, val;
793 void __iomem *xregs, *cregs;
795 /* Make sure IOC translation is disabled on all nodes */
796 for_each_online_node(node) {
797 if (cell_iommu_find_ioc(node, &base))
799 xregs = ioremap(base, IOC_Reg_Size);
802 cregs = xregs + IOC_IOCmd_Offset;
804 pr_debug("iommu: cleaning up iommu on node %d\n", node);
806 out_be64(xregs + IOC_IOST_Origin, 0);
807 (void)in_be64(xregs + IOC_IOST_Origin);
808 val = in_be64(cregs + IOC_IOCmd_Cfg);
809 val &= ~IOC_IOCmd_Cfg_TE;
810 out_be64(cregs + IOC_IOCmd_Cfg, val);
811 (void)in_be64(cregs + IOC_IOCmd_Cfg);
817 static int __init cell_iommu_init_disabled(void)
819 struct device_node *np = NULL;
820 unsigned long base = 0, size;
822 /* When no iommu is present, we use direct DMA ops */
823 set_pci_dma_ops(&dma_direct_ops);
825 /* First make sure all IOC translation is turned off */
826 cell_disable_iommus();
828 /* If we have no Axon, we set up the spider DMA magic offset */
829 if (of_find_node_by_name(NULL, "axon") == NULL)
830 cell_dma_direct_offset = SPIDER_DMA_OFFSET;
832 /* Now we need to check to see where the memory is mapped
833 * in PCI space. We assume that all busses use the same dma
834 * window which is always the case so far on Cell, thus we
835 * pick up the first pci-internal node we can find and check
836 * the DMA window from there.
838 for_each_node_by_name(np, "axon") {
839 if (np->parent == NULL || np->parent->parent != NULL)
841 if (cell_iommu_get_window(np, &base, &size) == 0)
845 for_each_node_by_name(np, "pci-internal") {
846 if (np->parent == NULL || np->parent->parent != NULL)
848 if (cell_iommu_get_window(np, &base, &size) == 0)
854 /* If we found a DMA window, we check if it's big enough to enclose
855 * all of physical memory. If not, we force enable IOMMU
857 if (np && size < lmb_end_of_DRAM()) {
858 printk(KERN_WARNING "iommu: force-enabled, dma window"
859 " (%ldMB) smaller than total memory (%lldMB)\n",
860 size >> 20, lmb_end_of_DRAM() >> 20);
864 cell_dma_direct_offset += base;
866 if (cell_dma_direct_offset != 0)
867 ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
869 printk("iommu: disabled, direct DMA offset is 0x%lx\n",
870 cell_dma_direct_offset);
876 * Fixed IOMMU mapping support
878 * This code adds support for setting up a fixed IOMMU mapping on certain
879 * cell machines. For 64-bit devices this avoids the performance overhead of
880 * mapping and unmapping pages at runtime. 32-bit devices are unable to use
883 * The fixed mapping is established at boot, and maps all of physical memory
884 * 1:1 into device space at some offset. On machines with < 30 GB of memory
885 * we setup the fixed mapping immediately above the normal IOMMU window.
887 * For example a machine with 4GB of memory would end up with the normal
888 * IOMMU window from 0-2GB and the fixed mapping window from 2GB to 6GB. In
889 * this case a 64-bit device wishing to DMA to 1GB would be told to DMA to
890 * 3GB, plus any offset required by firmware. The firmware offset is encoded
891 * in the "dma-ranges" property.
893 * On machines with 30GB or more of memory, we are unable to place the fixed
894 * mapping above the normal IOMMU window as we would run out of address space.
895 * Instead we move the normal IOMMU window to coincide with the hash page
896 * table, this region does not need to be part of the fixed mapping as no
897 * device should ever be DMA'ing to it. We then setup the fixed mapping
901 static u64 cell_iommu_get_fixed_address(struct device *dev)
903 u64 cpu_addr, size, best_size, dev_addr = OF_BAD_ADDR;
904 struct device_node *np;
905 const u32 *ranges = NULL;
906 int i, len, best, naddr, nsize, pna, range_size;
908 np = of_node_get(dev->archdata.of_node);
910 naddr = of_n_addr_cells(np);
911 nsize = of_n_size_cells(np);
912 np = of_get_next_parent(np);
916 ranges = of_get_property(np, "dma-ranges", &len);
918 /* Ignore empty ranges, they imply no translation required */
919 if (ranges && len > 0)
924 dev_dbg(dev, "iommu: no dma-ranges found\n");
930 pna = of_n_addr_cells(np);
931 range_size = naddr + nsize + pna;
933 /* dma-ranges format:
934 * child addr : naddr cells
935 * parent addr : pna cells
938 for (i = 0, best = -1, best_size = 0; i < len; i += range_size) {
939 cpu_addr = of_translate_dma_address(np, ranges + i + naddr);
940 size = of_read_number(ranges + i + naddr + pna, nsize);
942 if (cpu_addr == 0 && size > best_size) {
949 dev_addr = of_read_number(ranges + best, naddr);
951 dev_dbg(dev, "iommu: no suitable range found!\n");
959 static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask)
961 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
964 if (dma_mask == DMA_BIT_MASK(64) &&
965 cell_iommu_get_fixed_address(dev) != OF_BAD_ADDR)
967 dev_dbg(dev, "iommu: 64-bit OK, using fixed ops\n");
968 set_dma_ops(dev, &dma_iommu_fixed_ops);
970 dev_dbg(dev, "iommu: not 64-bit, using default ops\n");
971 set_dma_ops(dev, get_pci_dma_ops());
974 cell_dma_dev_setup(dev);
976 *dev->dma_mask = dma_mask;
981 static void cell_dma_dev_setup_fixed(struct device *dev)
983 struct dev_archdata *archdata = &dev->archdata;
986 addr = cell_iommu_get_fixed_address(dev) + dma_iommu_fixed_base;
987 archdata->dma_data = (void *)addr;
989 dev_dbg(dev, "iommu: fixed addr = %llx\n", addr);
992 static void insert_16M_pte(unsigned long addr, unsigned long *ptab,
993 unsigned long base_pte)
995 unsigned long segment, offset;
997 segment = addr >> IO_SEGMENT_SHIFT;
998 offset = (addr >> 24) - (segment << IO_PAGENO_BITS(24));
999 ptab = ptab + (segment * (1 << 12) / sizeof(unsigned long));
1001 pr_debug("iommu: addr %lx ptab %p segment %lx offset %lx\n",
1002 addr, ptab, segment, offset);
1004 ptab[offset] = base_pte | (__pa(addr) & IOPTE_RPN_Mask);
1007 static void cell_iommu_setup_fixed_ptab(struct cbe_iommu *iommu,
1008 struct device_node *np, unsigned long dbase, unsigned long dsize,
1009 unsigned long fbase, unsigned long fsize)
1011 unsigned long base_pte, uaddr, ioaddr, *ptab;
1013 ptab = cell_iommu_alloc_ptab(iommu, fbase, fsize, dbase, dsize, 24);
1015 dma_iommu_fixed_base = fbase;
1017 pr_debug("iommu: mapping 0x%lx pages from 0x%lx\n", fsize, fbase);
1019 base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M
1020 | (cell_iommu_get_ioid(np) & IOPTE_IOID_Mask);
1022 if (iommu_fixed_is_weak)
1023 pr_info("IOMMU: Using weak ordering for fixed mapping\n");
1025 pr_info("IOMMU: Using strong ordering for fixed mapping\n");
1026 base_pte |= IOPTE_SO_RW;
1029 for (uaddr = 0; uaddr < fsize; uaddr += (1 << 24)) {
1030 /* Don't touch the dynamic region */
1031 ioaddr = uaddr + fbase;
1032 if (ioaddr >= dbase && ioaddr < (dbase + dsize)) {
1033 pr_debug("iommu: fixed/dynamic overlap, skipping\n");
1037 insert_16M_pte(uaddr, ptab, base_pte);
1043 static int __init cell_iommu_fixed_mapping_init(void)
1045 unsigned long dbase, dsize, fbase, fsize, hbase, hend;
1046 struct cbe_iommu *iommu;
1047 struct device_node *np;
1049 /* The fixed mapping is only supported on axon machines */
1050 np = of_find_node_by_name(NULL, "axon");
1052 pr_debug("iommu: fixed mapping disabled, no axons found\n");
1056 /* We must have dma-ranges properties for fixed mapping to work */
1057 np = of_find_node_with_property(NULL, "dma-ranges");
1061 pr_debug("iommu: no dma-ranges found, no fixed mapping\n");
1065 /* The default setup is to have the fixed mapping sit after the
1066 * dynamic region, so find the top of the largest IOMMU window
1067 * on any axon, then add the size of RAM and that's our max value.
1068 * If that is > 32GB we have to do other shennanigans.
1071 for_each_node_by_name(np, "axon") {
1072 cell_iommu_get_window(np, &dbase, &dsize);
1073 fbase = max(fbase, dbase + dsize);
1076 fbase = _ALIGN_UP(fbase, 1 << IO_SEGMENT_SHIFT);
1077 fsize = lmb_phys_mem_size();
1079 if ((fbase + fsize) <= 0x800000000)
1080 hbase = 0; /* use the device tree window */
1082 /* If we're over 32 GB we need to cheat. We can't map all of
1083 * RAM with the fixed mapping, and also fit the dynamic
1084 * region. So try to place the dynamic region where the hash
1085 * table sits, drivers never need to DMA to it, we don't
1086 * need a fixed mapping for that area.
1088 if (!htab_address) {
1089 pr_debug("iommu: htab is NULL, on LPAR? Huh?\n");
1092 hbase = __pa(htab_address);
1093 hend = hbase + htab_size_bytes;
1095 /* The window must start and end on a segment boundary */
1096 if ((hbase != _ALIGN_UP(hbase, 1 << IO_SEGMENT_SHIFT)) ||
1097 (hend != _ALIGN_UP(hend, 1 << IO_SEGMENT_SHIFT))) {
1098 pr_debug("iommu: hash window not segment aligned\n");
1102 /* Check the hash window fits inside the real DMA window */
1103 for_each_node_by_name(np, "axon") {
1104 cell_iommu_get_window(np, &dbase, &dsize);
1106 if (hbase < dbase || (hend > (dbase + dsize))) {
1107 pr_debug("iommu: hash window doesn't fit in"
1108 "real DMA window\n");
1116 /* Setup the dynamic regions */
1117 for_each_node_by_name(np, "axon") {
1118 iommu = cell_iommu_alloc(np);
1122 cell_iommu_get_window(np, &dbase, &dsize);
1125 dsize = htab_size_bytes;
1128 printk(KERN_DEBUG "iommu: node %d, dynamic window 0x%lx-0x%lx "
1129 "fixed window 0x%lx-0x%lx\n", iommu->nid, dbase,
1130 dbase + dsize, fbase, fbase + fsize);
1132 cell_iommu_setup_stab(iommu, dbase, dsize, fbase, fsize);
1133 iommu->ptab = cell_iommu_alloc_ptab(iommu, dbase, dsize, 0, 0,
1135 cell_iommu_setup_fixed_ptab(iommu, np, dbase, dsize,
1137 cell_iommu_enable_hardware(iommu);
1138 cell_iommu_setup_window(iommu, np, dbase, dsize, 0);
1141 dma_iommu_ops.set_dma_mask = dma_set_mask_and_switch;
1142 set_pci_dma_ops(&dma_iommu_ops);
1147 static int iommu_fixed_disabled;
1149 static int __init setup_iommu_fixed(char *str)
1151 struct device_node *pciep;
1153 if (strcmp(str, "off") == 0)
1154 iommu_fixed_disabled = 1;
1156 /* If we can find a pcie-endpoint in the device tree assume that
1157 * we're on a triblade or a CAB so by default the fixed mapping
1158 * should be set to be weakly ordered; but only if the boot
1159 * option WASN'T set for strong ordering
1161 pciep = of_find_node_by_type(NULL, "pcie-endpoint");
1163 if (strcmp(str, "weak") == 0 || (pciep && strcmp(str, "strong") != 0))
1164 iommu_fixed_is_weak = 1;
1170 __setup("iommu_fixed=", setup_iommu_fixed);
1172 static int __init cell_iommu_init(void)
1174 struct device_node *np;
1176 /* If IOMMU is disabled or we have little enough RAM to not need
1177 * to enable it, we setup a direct mapping.
1179 * Note: should we make sure we have the IOMMU actually disabled ?
1182 (!iommu_force_on && lmb_end_of_DRAM() <= 0x80000000ull))
1183 if (cell_iommu_init_disabled() == 0)
1186 /* Setup various ppc_md. callbacks */
1187 ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
1188 ppc_md.tce_build = tce_build_cell;
1189 ppc_md.tce_free = tce_free_cell;
1191 if (!iommu_fixed_disabled && cell_iommu_fixed_mapping_init() == 0)
1194 /* Create an iommu for each /axon node. */
1195 for_each_node_by_name(np, "axon") {
1196 if (np->parent == NULL || np->parent->parent != NULL)
1198 cell_iommu_init_one(np, 0);
1201 /* Create an iommu for each toplevel /pci-internal node for
1202 * old hardware/firmware
1204 for_each_node_by_name(np, "pci-internal") {
1205 if (np->parent == NULL || np->parent->parent != NULL)
1207 cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
1210 /* Setup default PCI iommu ops */
1211 set_pci_dma_ops(&dma_iommu_ops);
1214 /* Register callbacks on OF platform device addition/removal
1215 * to handle linking them to the right DMA operations
1217 bus_register_notifier(&of_platform_bus_type, &cell_of_bus_notifier);
1221 machine_arch_initcall(cell, cell_iommu_init);
1222 machine_arch_initcall(celleb_native, cell_iommu_init);