2 * IOMMU implementation for Cell Broadband Processor Architecture
4 * (C) Copyright IBM Corporation 2006
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>
31 #include <asm/iommu.h>
32 #include <asm/machdep.h>
33 #include <asm/pci-bridge.h>
35 #include <asm/of_platform.h>
39 #include "interrupt.h"
41 /* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages
42 * instead of leaving them mapped to some dummy page. This can be
43 * enabled once the appropriate workarounds for spider bugs have
46 #define CELL_IOMMU_REAL_UNMAP
48 /* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of
49 * IO PTEs based on the transfer direction. That can be enabled
50 * once spider-net has been fixed to pass the correct direction
51 * to the DMA mapping functions
53 #define CELL_IOMMU_STRICT_PROTECTION
58 /* IOC mmap registers */
59 #define IOC_Reg_Size 0x2000
61 #define IOC_IOPT_CacheInvd 0x908
62 #define IOC_IOPT_CacheInvd_NE_Mask 0xffe0000000000000ul
63 #define IOC_IOPT_CacheInvd_IOPTE_Mask 0x000003fffffffff8ul
64 #define IOC_IOPT_CacheInvd_Busy 0x0000000000000001ul
66 #define IOC_IOST_Origin 0x918
67 #define IOC_IOST_Origin_E 0x8000000000000000ul
68 #define IOC_IOST_Origin_HW 0x0000000000000800ul
69 #define IOC_IOST_Origin_HL 0x0000000000000400ul
71 #define IOC_IO_ExcpStat 0x920
72 #define IOC_IO_ExcpStat_V 0x8000000000000000ul
73 #define IOC_IO_ExcpStat_SPF_Mask 0x6000000000000000ul
74 #define IOC_IO_ExcpStat_SPF_S 0x6000000000000000ul
75 #define IOC_IO_ExcpStat_SPF_P 0x4000000000000000ul
76 #define IOC_IO_ExcpStat_ADDR_Mask 0x00000007fffff000ul
77 #define IOC_IO_ExcpStat_RW_Mask 0x0000000000000800ul
78 #define IOC_IO_ExcpStat_IOID_Mask 0x00000000000007fful
80 #define IOC_IO_ExcpMask 0x928
81 #define IOC_IO_ExcpMask_SFE 0x4000000000000000ul
82 #define IOC_IO_ExcpMask_PFE 0x2000000000000000ul
84 #define IOC_IOCmd_Offset 0x1000
86 #define IOC_IOCmd_Cfg 0xc00
87 #define IOC_IOCmd_Cfg_TE 0x0000800000000000ul
90 /* Segment table entries */
91 #define IOSTE_V 0x8000000000000000ul /* valid */
92 #define IOSTE_H 0x4000000000000000ul /* cache hint */
93 #define IOSTE_PT_Base_RPN_Mask 0x3ffffffffffff000ul /* base RPN of IOPT */
94 #define IOSTE_NPPT_Mask 0x0000000000000fe0ul /* no. pages in IOPT */
95 #define IOSTE_PS_Mask 0x0000000000000007ul /* page size */
96 #define IOSTE_PS_4K 0x0000000000000001ul /* - 4kB */
97 #define IOSTE_PS_64K 0x0000000000000003ul /* - 64kB */
98 #define IOSTE_PS_1M 0x0000000000000005ul /* - 1MB */
99 #define IOSTE_PS_16M 0x0000000000000007ul /* - 16MB */
101 /* Page table entries */
102 #define IOPTE_PP_W 0x8000000000000000ul /* protection: write */
103 #define IOPTE_PP_R 0x4000000000000000ul /* protection: read */
104 #define IOPTE_M 0x2000000000000000ul /* coherency required */
105 #define IOPTE_SO_R 0x1000000000000000ul /* ordering: writes */
106 #define IOPTE_SO_RW 0x1800000000000000ul /* ordering: r & w */
107 #define IOPTE_RPN_Mask 0x07fffffffffff000ul /* RPN */
108 #define IOPTE_H 0x0000000000000800ul /* cache hint */
109 #define IOPTE_IOID_Mask 0x00000000000007fful /* ioid */
113 #define IO_SEGMENT_SHIFT 28
114 #define IO_PAGENO_BITS (IO_SEGMENT_SHIFT - IOMMU_PAGE_SHIFT)
116 /* The high bit needs to be set on every DMA address */
117 #define SPIDER_DMA_OFFSET 0x80000000ul
119 struct iommu_window {
120 struct list_head list;
121 struct cbe_iommu *iommu;
122 unsigned long offset;
124 unsigned long pte_offset;
126 struct iommu_table table;
133 void __iomem *xlate_regs;
134 void __iomem *cmd_regs;
138 struct list_head windows;
141 /* Static array of iommus, one per node
142 * each contains a list of windows, keyed from dma_window property
143 * - on bus setup, look for a matching window, or create one
144 * - on dev setup, assign iommu_table ptr
146 static struct cbe_iommu iommus[NR_IOMMUS];
147 static int cbe_nr_iommus;
149 static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
152 unsigned long __iomem *reg;
156 reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
159 /* we can invalidate up to 1 << 11 PTEs at once */
160 n = min(n_ptes, 1l << 11);
161 val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask)
162 | (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask)
163 | IOC_IOPT_CacheInvd_Busy;
166 while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
174 static void tce_build_cell(struct iommu_table *tbl, long index, long npages,
175 unsigned long uaddr, enum dma_data_direction direction)
178 unsigned long *io_pte, base_pte;
179 struct iommu_window *window =
180 container_of(tbl, struct iommu_window, table);
182 /* implementing proper protection causes problems with the spidernet
183 * driver - check mapping directions later, but allow read & write by
185 #ifdef CELL_IOMMU_STRICT_PROTECTION
186 /* to avoid referencing a global, we use a trick here to setup the
187 * protection bit. "prot" is setup to be 3 fields of 4 bits apprended
188 * together for each of the 3 supported direction values. It is then
189 * shifted left so that the fields matching the desired direction
190 * lands on the appropriate bits, and other bits are masked out.
192 const unsigned long prot = 0xc48;
194 ((prot << (52 + 4 * direction)) & (IOPTE_PP_W | IOPTE_PP_R))
195 | IOPTE_M | IOPTE_SO_RW | (window->ioid & IOPTE_IOID_Mask);
197 base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW |
198 (window->ioid & IOPTE_IOID_Mask);
201 io_pte = (unsigned long *)tbl->it_base + (index - window->pte_offset);
203 for (i = 0; i < npages; i++, uaddr += IOMMU_PAGE_SIZE)
204 io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask);
208 invalidate_tce_cache(window->iommu, io_pte, npages);
210 pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
211 index, npages, direction, base_pte);
214 static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
218 unsigned long *io_pte, pte;
219 struct iommu_window *window =
220 container_of(tbl, struct iommu_window, table);
222 pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
224 #ifdef CELL_IOMMU_REAL_UNMAP
227 /* spider bridge does PCI reads after freeing - insert a mapping
228 * to a scratch page instead of an invalid entry */
229 pte = IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | __pa(window->iommu->pad_page)
230 | (window->ioid & IOPTE_IOID_Mask);
233 io_pte = (unsigned long *)tbl->it_base + (index - window->pte_offset);
235 for (i = 0; i < npages; i++)
240 invalidate_tce_cache(window->iommu, io_pte, npages);
243 static irqreturn_t ioc_interrupt(int irq, void *data)
246 struct cbe_iommu *iommu = data;
248 stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
250 /* Might want to rate limit it */
251 printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat);
252 printk(KERN_ERR " V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n",
253 !!(stat & IOC_IO_ExcpStat_V),
254 (stat & IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
255 (stat & IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ',
256 (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write",
257 (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask));
258 printk(KERN_ERR " page=0x%016lx\n",
259 stat & IOC_IO_ExcpStat_ADDR_Mask);
261 /* clear interrupt */
262 stat &= ~IOC_IO_ExcpStat_V;
263 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat);
268 static int cell_iommu_find_ioc(int nid, unsigned long *base)
270 struct device_node *np;
275 /* First look for new style /be nodes */
276 for_each_node_by_name(np, "ioc") {
277 if (of_node_to_nid(np) != nid)
279 if (of_address_to_resource(np, 0, &r)) {
280 printk(KERN_ERR "iommu: can't get address for %s\n",
289 /* Ok, let's try the old way */
290 for_each_node_by_type(np, "cpu") {
291 const unsigned int *nidp;
292 const unsigned long *tmp;
294 nidp = of_get_property(np, "node-id", NULL);
295 if (nidp && *nidp == nid) {
296 tmp = of_get_property(np, "ioc-translation", NULL);
308 static void cell_iommu_setup_hardware(struct cbe_iommu *iommu, unsigned long size)
312 unsigned long reg, segments, pages_per_segment, ptab_size, n_pte_pages;
313 unsigned long xlate_base;
316 if (cell_iommu_find_ioc(iommu->nid, &xlate_base))
317 panic("%s: missing IOC register mappings for node %d\n",
318 __FUNCTION__, iommu->nid);
320 iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size);
321 iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset;
323 segments = size >> IO_SEGMENT_SHIFT;
324 pages_per_segment = 1ull << IO_PAGENO_BITS;
326 pr_debug("%s: iommu[%d]: segments: %lu, pages per segment: %lu\n",
327 __FUNCTION__, iommu->nid, segments, pages_per_segment);
329 /* set up the segment table */
330 page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
332 iommu->stab = page_address(page);
333 clear_page(iommu->stab);
335 /* ... and the page tables. Since these are contiguous, we can treat
336 * the page tables as one array of ptes, like pSeries does.
338 ptab_size = segments * pages_per_segment * sizeof(unsigned long);
339 pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __FUNCTION__,
340 iommu->nid, ptab_size, get_order(ptab_size));
341 page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size));
344 iommu->ptab = page_address(page);
345 memset(iommu->ptab, 0, ptab_size);
347 /* allocate a bogus page for the end of each mapping */
348 page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
350 iommu->pad_page = page_address(page);
351 clear_page(iommu->pad_page);
353 /* number of pages needed for a page table */
354 n_pte_pages = (pages_per_segment *
355 sizeof(unsigned long)) >> IOMMU_PAGE_SHIFT;
357 pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n",
358 __FUNCTION__, iommu->nid, iommu->stab, iommu->ptab,
361 /* initialise the STEs */
362 reg = IOSTE_V | ((n_pte_pages - 1) << 5);
364 if (IOMMU_PAGE_SIZE == 0x1000)
366 else if (IOMMU_PAGE_SIZE == 0x10000)
369 extern void __unknown_page_size_error(void);
370 __unknown_page_size_error();
373 pr_debug("Setting up IOMMU stab:\n");
374 for (i = 0; i * (1ul << IO_SEGMENT_SHIFT) < size; i++) {
375 iommu->stab[i] = reg |
376 (__pa(iommu->ptab) + n_pte_pages * IOMMU_PAGE_SIZE * i);
377 pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]);
380 /* ensure that the STEs have updated */
383 /* setup interrupts for the iommu. */
384 reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
385 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat,
386 reg & ~IOC_IO_ExcpStat_V);
387 out_be64(iommu->xlate_regs + IOC_IO_ExcpMask,
388 IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE);
390 virq = irq_create_mapping(NULL,
391 IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT));
392 BUG_ON(virq == NO_IRQ);
394 ret = request_irq(virq, ioc_interrupt, IRQF_DISABLED,
398 /* set the IOC segment table origin register (and turn on the iommu) */
399 reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
400 out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
401 in_be64(iommu->xlate_regs + IOC_IOST_Origin);
403 /* turn on IO translation */
404 reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
405 out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
408 #if 0/* Unused for now */
409 static struct iommu_window *find_window(struct cbe_iommu *iommu,
410 unsigned long offset, unsigned long size)
412 struct iommu_window *window;
414 /* todo: check for overlapping (but not equal) windows) */
416 list_for_each_entry(window, &(iommu->windows), list) {
417 if (window->offset == offset && window->size == size)
425 static struct iommu_window * __init
426 cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
427 unsigned long offset, unsigned long size,
428 unsigned long pte_offset)
430 struct iommu_window *window;
431 const unsigned int *ioid;
433 ioid = of_get_property(np, "ioid", NULL);
435 printk(KERN_WARNING "iommu: missing ioid for %s using 0\n",
438 window = kmalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
439 BUG_ON(window == NULL);
441 window->offset = offset;
443 window->ioid = ioid ? *ioid : 0;
444 window->iommu = iommu;
445 window->pte_offset = pte_offset;
447 window->table.it_blocksize = 16;
448 window->table.it_base = (unsigned long)iommu->ptab;
449 window->table.it_index = iommu->nid;
450 window->table.it_offset = (offset >> IOMMU_PAGE_SHIFT) +
452 window->table.it_size = size >> IOMMU_PAGE_SHIFT;
454 iommu_init_table(&window->table, iommu->nid);
456 pr_debug("\tioid %d\n", window->ioid);
457 pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
458 pr_debug("\tbase 0x%016lx\n", window->table.it_base);
459 pr_debug("\toffset 0x%lx\n", window->table.it_offset);
460 pr_debug("\tsize %ld\n", window->table.it_size);
462 list_add(&window->list, &iommu->windows);
467 /* We need to map and reserve the first IOMMU page since it's used
468 * by the spider workaround. In theory, we only need to do that when
469 * running on spider but it doesn't really matter.
471 * This code also assumes that we have a window that starts at 0,
472 * which is the case on all spider based blades.
474 __set_bit(0, window->table.it_map);
475 tce_build_cell(&window->table, window->table.it_offset, 1,
476 (unsigned long)iommu->pad_page, DMA_TO_DEVICE);
477 window->table.it_hint = window->table.it_blocksize;
482 static struct cbe_iommu *cell_iommu_for_node(int nid)
486 for (i = 0; i < cbe_nr_iommus; i++)
487 if (iommus[i].nid == nid)
492 static void cell_dma_dev_setup(struct device *dev)
494 struct iommu_window *window;
495 struct cbe_iommu *iommu;
496 struct dev_archdata *archdata = &dev->archdata;
498 /* If we run without iommu, no need to do anything */
499 if (get_pci_dma_ops() == &dma_direct_ops)
502 /* Current implementation uses the first window available in that
503 * node's iommu. We -might- do something smarter later though it may
506 iommu = cell_iommu_for_node(archdata->numa_node);
507 if (iommu == NULL || list_empty(&iommu->windows)) {
508 printk(KERN_ERR "iommu: missing iommu for %s (node %d)\n",
509 archdata->of_node ? archdata->of_node->full_name : "?",
510 archdata->numa_node);
513 window = list_entry(iommu->windows.next, struct iommu_window, list);
515 archdata->dma_data = &window->table;
518 static void cell_pci_dma_dev_setup(struct pci_dev *dev)
520 cell_dma_dev_setup(&dev->dev);
523 static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
526 struct device *dev = data;
528 /* We are only intereted in device addition */
529 if (action != BUS_NOTIFY_ADD_DEVICE)
532 /* We use the PCI DMA ops */
533 dev->archdata.dma_ops = get_pci_dma_ops();
535 cell_dma_dev_setup(dev);
540 static struct notifier_block cell_of_bus_notifier = {
541 .notifier_call = cell_of_bus_notify
544 static int __init cell_iommu_get_window(struct device_node *np,
548 const void *dma_window;
551 /* Use ibm,dma-window if available, else, hard code ! */
552 dma_window = of_get_property(np, "ibm,dma-window", NULL);
553 if (dma_window == NULL) {
559 of_parse_dma_window(np, dma_window, &index, base, size);
563 static void __init cell_iommu_init_one(struct device_node *np, unsigned long offset)
565 struct cbe_iommu *iommu;
566 unsigned long base, size;
570 nid = of_node_to_nid(np);
572 printk(KERN_ERR "iommu: failed to get node for %s\n",
576 pr_debug("iommu: setting up iommu for node %d (%s)\n",
579 /* XXX todo: If we can have multiple windows on the same IOMMU, which
580 * isn't the case today, we probably want here to check wether the
581 * iommu for that node is already setup.
582 * However, there might be issue with getting the size right so let's
583 * ignore that for now. We might want to completely get rid of the
584 * multiple window support since the cell iommu supports per-page ioids
587 if (cbe_nr_iommus >= NR_IOMMUS) {
588 printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n",
593 /* Init base fields */
598 snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
599 INIT_LIST_HEAD(&iommu->windows);
601 /* Obtain a window for it */
602 cell_iommu_get_window(np, &base, &size);
604 pr_debug("\ttranslating window 0x%lx...0x%lx\n",
605 base, base + size - 1);
607 /* Initialize the hardware */
608 cell_iommu_setup_hardware(iommu, size);
610 /* Setup the iommu_table */
611 cell_iommu_setup_window(iommu, np, base, size,
612 offset >> IOMMU_PAGE_SHIFT);
615 static void __init cell_disable_iommus(void)
618 unsigned long base, val;
619 void __iomem *xregs, *cregs;
621 /* Make sure IOC translation is disabled on all nodes */
622 for_each_online_node(node) {
623 if (cell_iommu_find_ioc(node, &base))
625 xregs = ioremap(base, IOC_Reg_Size);
628 cregs = xregs + IOC_IOCmd_Offset;
630 pr_debug("iommu: cleaning up iommu on node %d\n", node);
632 out_be64(xregs + IOC_IOST_Origin, 0);
633 (void)in_be64(xregs + IOC_IOST_Origin);
634 val = in_be64(cregs + IOC_IOCmd_Cfg);
635 val &= ~IOC_IOCmd_Cfg_TE;
636 out_be64(cregs + IOC_IOCmd_Cfg, val);
637 (void)in_be64(cregs + IOC_IOCmd_Cfg);
643 static int __init cell_iommu_init_disabled(void)
645 struct device_node *np = NULL;
646 unsigned long base = 0, size;
648 /* When no iommu is present, we use direct DMA ops */
649 set_pci_dma_ops(&dma_direct_ops);
651 /* First make sure all IOC translation is turned off */
652 cell_disable_iommus();
654 /* If we have no Axon, we set up the spider DMA magic offset */
655 if (of_find_node_by_name(NULL, "axon") == NULL)
656 dma_direct_offset = SPIDER_DMA_OFFSET;
658 /* Now we need to check to see where the memory is mapped
659 * in PCI space. We assume that all busses use the same dma
660 * window which is always the case so far on Cell, thus we
661 * pick up the first pci-internal node we can find and check
662 * the DMA window from there.
664 for_each_node_by_name(np, "axon") {
665 if (np->parent == NULL || np->parent->parent != NULL)
667 if (cell_iommu_get_window(np, &base, &size) == 0)
671 for_each_node_by_name(np, "pci-internal") {
672 if (np->parent == NULL || np->parent->parent != NULL)
674 if (cell_iommu_get_window(np, &base, &size) == 0)
680 /* If we found a DMA window, we check if it's big enough to enclose
681 * all of physical memory. If not, we force enable IOMMU
683 if (np && size < lmb_end_of_DRAM()) {
684 printk(KERN_WARNING "iommu: force-enabled, dma window"
685 " (%ldMB) smaller than total memory (%ldMB)\n",
686 size >> 20, lmb_end_of_DRAM() >> 20);
690 dma_direct_offset += base;
692 printk("iommu: disabled, direct DMA offset is 0x%lx\n",
698 static int __init cell_iommu_init(void)
700 struct device_node *np;
702 if (!machine_is(cell))
705 /* If IOMMU is disabled or we have little enough RAM to not need
706 * to enable it, we setup a direct mapping.
708 * Note: should we make sure we have the IOMMU actually disabled ?
711 (!iommu_force_on && lmb_end_of_DRAM() <= 0x80000000ull))
712 if (cell_iommu_init_disabled() == 0)
715 /* Setup various ppc_md. callbacks */
716 ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
717 ppc_md.tce_build = tce_build_cell;
718 ppc_md.tce_free = tce_free_cell;
720 /* Create an iommu for each /axon node. */
721 for_each_node_by_name(np, "axon") {
722 if (np->parent == NULL || np->parent->parent != NULL)
724 cell_iommu_init_one(np, 0);
727 /* Create an iommu for each toplevel /pci-internal node for
728 * old hardware/firmware
730 for_each_node_by_name(np, "pci-internal") {
731 if (np->parent == NULL || np->parent->parent != NULL)
733 cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
736 /* Setup default PCI iommu ops */
737 set_pci_dma_ops(&dma_iommu_ops);
740 /* Register callbacks on OF platform device addition/removal
741 * to handle linking them to the right DMA operations
743 bus_register_notifier(&of_platform_bus_type, &cell_of_bus_notifier);
747 arch_initcall(cell_iommu_init);