Commit | Line | Data |
---|---|---|
b6c02715 JR |
1 | /* |
2 | * Copyright (C) 2007-2008 Advanced Micro Devices, Inc. | |
3 | * Author: Joerg Roedel <joerg.roedel@amd.com> | |
4 | * Leo Duran <leo.duran@amd.com> | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify it | |
7 | * under the terms of the GNU General Public License version 2 as published | |
8 | * by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, | |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 | * GNU General Public License for more details. | |
14 | * | |
15 | * You should have received a copy of the GNU General Public License | |
16 | * along with this program; if not, write to the Free Software | |
17 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
18 | */ | |
19 | ||
20 | #include <linux/pci.h> | |
21 | #include <linux/gfp.h> | |
22 | #include <linux/bitops.h> | |
23 | #include <linux/scatterlist.h> | |
24 | #include <linux/iommu-helper.h> | |
25 | #include <asm/proto.h> | |
46a7fa27 | 26 | #include <asm/iommu.h> |
b6c02715 | 27 | #include <asm/amd_iommu_types.h> |
c6da992e | 28 | #include <asm/amd_iommu.h> |
b6c02715 JR |
29 | |
30 | #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28)) | |
31 | ||
136f78a1 JR |
32 | #define EXIT_LOOP_COUNT 10000000 |
33 | ||
b6c02715 JR |
34 | static DEFINE_RWLOCK(amd_iommu_devtable_lock); |
35 | ||
bd60b735 JR |
36 | /* A list of preallocated protection domains */ |
37 | static LIST_HEAD(iommu_pd_list); | |
38 | static DEFINE_SPINLOCK(iommu_pd_list_lock); | |
39 | ||
431b2a20 JR |
40 | /* |
41 | * general struct to manage commands send to an IOMMU | |
42 | */ | |
d6449536 | 43 | struct iommu_cmd { |
b6c02715 JR |
44 | u32 data[4]; |
45 | }; | |
46 | ||
bd0e5211 JR |
47 | static int dma_ops_unity_map(struct dma_ops_domain *dma_dom, |
48 | struct unity_map_entry *e); | |
49 | ||
431b2a20 | 50 | /* returns !0 if the IOMMU is caching non-present entries in its TLB */ |
4da70b9e JR |
51 | static int iommu_has_npcache(struct amd_iommu *iommu) |
52 | { | |
53 | return iommu->cap & IOMMU_CAP_NPCACHE; | |
54 | } | |
55 | ||
a80dc3e0 JR |
56 | /**************************************************************************** |
57 | * | |
58 | * Interrupt handling functions | |
59 | * | |
60 | ****************************************************************************/ | |
61 | ||
90008ee4 JR |
62 | static void iommu_print_event(void *__evt) |
63 | { | |
64 | u32 *event = __evt; | |
65 | int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK; | |
66 | int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK; | |
67 | int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK; | |
68 | int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK; | |
69 | u64 address = (u64)(((u64)event[3]) << 32) | event[2]; | |
70 | ||
71 | printk(KERN_ERR "AMD IOMMU: Event logged ["); | |
72 | ||
73 | switch (type) { | |
74 | case EVENT_TYPE_ILL_DEV: | |
75 | printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x " | |
76 | "address=0x%016llx flags=0x%04x]\n", | |
77 | PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid), | |
78 | address, flags); | |
79 | break; | |
80 | case EVENT_TYPE_IO_FAULT: | |
81 | printk("IO_PAGE_FAULT device=%02x:%02x.%x " | |
82 | "domain=0x%04x address=0x%016llx flags=0x%04x]\n", | |
83 | PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid), | |
84 | domid, address, flags); | |
85 | break; | |
86 | case EVENT_TYPE_DEV_TAB_ERR: | |
87 | printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x " | |
88 | "address=0x%016llx flags=0x%04x]\n", | |
89 | PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid), | |
90 | address, flags); | |
91 | break; | |
92 | case EVENT_TYPE_PAGE_TAB_ERR: | |
93 | printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x " | |
94 | "domain=0x%04x address=0x%016llx flags=0x%04x]\n", | |
95 | PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid), | |
96 | domid, address, flags); | |
97 | break; | |
98 | case EVENT_TYPE_ILL_CMD: | |
99 | printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address); | |
100 | break; | |
101 | case EVENT_TYPE_CMD_HARD_ERR: | |
102 | printk("COMMAND_HARDWARE_ERROR address=0x%016llx " | |
103 | "flags=0x%04x]\n", address, flags); | |
104 | break; | |
105 | case EVENT_TYPE_IOTLB_INV_TO: | |
106 | printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x " | |
107 | "address=0x%016llx]\n", | |
108 | PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid), | |
109 | address); | |
110 | break; | |
111 | case EVENT_TYPE_INV_DEV_REQ: | |
112 | printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x " | |
113 | "address=0x%016llx flags=0x%04x]\n", | |
114 | PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid), | |
115 | address, flags); | |
116 | break; | |
117 | default: | |
118 | printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type); | |
119 | } | |
120 | } | |
121 | ||
122 | static void iommu_poll_events(struct amd_iommu *iommu) | |
123 | { | |
124 | u32 head, tail; | |
125 | unsigned long flags; | |
126 | ||
127 | spin_lock_irqsave(&iommu->lock, flags); | |
128 | ||
129 | head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET); | |
130 | tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET); | |
131 | ||
132 | while (head != tail) { | |
133 | iommu_print_event(iommu->evt_buf + head); | |
134 | head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size; | |
135 | } | |
136 | ||
137 | writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET); | |
138 | ||
139 | spin_unlock_irqrestore(&iommu->lock, flags); | |
140 | } | |
141 | ||
a80dc3e0 JR |
142 | irqreturn_t amd_iommu_int_handler(int irq, void *data) |
143 | { | |
90008ee4 JR |
144 | struct amd_iommu *iommu; |
145 | ||
146 | list_for_each_entry(iommu, &amd_iommu_list, list) | |
147 | iommu_poll_events(iommu); | |
148 | ||
149 | return IRQ_HANDLED; | |
a80dc3e0 JR |
150 | } |
151 | ||
431b2a20 JR |
152 | /**************************************************************************** |
153 | * | |
154 | * IOMMU command queuing functions | |
155 | * | |
156 | ****************************************************************************/ | |
157 | ||
158 | /* | |
159 | * Writes the command to the IOMMUs command buffer and informs the | |
160 | * hardware about the new command. Must be called with iommu->lock held. | |
161 | */ | |
d6449536 | 162 | static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd) |
a19ae1ec JR |
163 | { |
164 | u32 tail, head; | |
165 | u8 *target; | |
166 | ||
167 | tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET); | |
8a7c5ef3 | 168 | target = iommu->cmd_buf + tail; |
a19ae1ec JR |
169 | memcpy_toio(target, cmd, sizeof(*cmd)); |
170 | tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size; | |
171 | head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET); | |
172 | if (tail == head) | |
173 | return -ENOMEM; | |
174 | writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET); | |
175 | ||
176 | return 0; | |
177 | } | |
178 | ||
431b2a20 JR |
179 | /* |
180 | * General queuing function for commands. Takes iommu->lock and calls | |
181 | * __iommu_queue_command(). | |
182 | */ | |
d6449536 | 183 | static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd) |
a19ae1ec JR |
184 | { |
185 | unsigned long flags; | |
186 | int ret; | |
187 | ||
188 | spin_lock_irqsave(&iommu->lock, flags); | |
189 | ret = __iommu_queue_command(iommu, cmd); | |
190 | spin_unlock_irqrestore(&iommu->lock, flags); | |
191 | ||
192 | return ret; | |
193 | } | |
194 | ||
431b2a20 JR |
195 | /* |
196 | * This function is called whenever we need to ensure that the IOMMU has | |
197 | * completed execution of all commands we sent. It sends a | |
198 | * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs | |
199 | * us about that by writing a value to a physical address we pass with | |
200 | * the command. | |
201 | */ | |
a19ae1ec JR |
202 | static int iommu_completion_wait(struct amd_iommu *iommu) |
203 | { | |
7e4f88da | 204 | int ret = 0, ready = 0; |
519c31ba | 205 | unsigned status = 0; |
d6449536 | 206 | struct iommu_cmd cmd; |
7e4f88da | 207 | unsigned long flags, i = 0; |
a19ae1ec JR |
208 | |
209 | memset(&cmd, 0, sizeof(cmd)); | |
519c31ba | 210 | cmd.data[0] = CMD_COMPL_WAIT_INT_MASK; |
a19ae1ec JR |
211 | CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT); |
212 | ||
213 | iommu->need_sync = 0; | |
214 | ||
7e4f88da JR |
215 | spin_lock_irqsave(&iommu->lock, flags); |
216 | ||
217 | ret = __iommu_queue_command(iommu, &cmd); | |
a19ae1ec JR |
218 | |
219 | if (ret) | |
7e4f88da | 220 | goto out; |
a19ae1ec | 221 | |
136f78a1 JR |
222 | while (!ready && (i < EXIT_LOOP_COUNT)) { |
223 | ++i; | |
519c31ba JR |
224 | /* wait for the bit to become one */ |
225 | status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); | |
226 | ready = status & MMIO_STATUS_COM_WAIT_INT_MASK; | |
136f78a1 JR |
227 | } |
228 | ||
519c31ba JR |
229 | /* set bit back to zero */ |
230 | status &= ~MMIO_STATUS_COM_WAIT_INT_MASK; | |
231 | writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET); | |
232 | ||
136f78a1 JR |
233 | if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit())) |
234 | printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n"); | |
7e4f88da JR |
235 | out: |
236 | spin_unlock_irqrestore(&iommu->lock, flags); | |
a19ae1ec JR |
237 | |
238 | return 0; | |
239 | } | |
240 | ||
431b2a20 JR |
241 | /* |
242 | * Command send function for invalidating a device table entry | |
243 | */ | |
a19ae1ec JR |
244 | static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid) |
245 | { | |
d6449536 | 246 | struct iommu_cmd cmd; |
ee2fa743 | 247 | int ret; |
a19ae1ec JR |
248 | |
249 | BUG_ON(iommu == NULL); | |
250 | ||
251 | memset(&cmd, 0, sizeof(cmd)); | |
252 | CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY); | |
253 | cmd.data[0] = devid; | |
254 | ||
ee2fa743 JR |
255 | ret = iommu_queue_command(iommu, &cmd); |
256 | ||
a19ae1ec JR |
257 | iommu->need_sync = 1; |
258 | ||
ee2fa743 | 259 | return ret; |
a19ae1ec JR |
260 | } |
261 | ||
431b2a20 JR |
262 | /* |
263 | * Generic command send function for invalidaing TLB entries | |
264 | */ | |
a19ae1ec JR |
265 | static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu, |
266 | u64 address, u16 domid, int pde, int s) | |
267 | { | |
d6449536 | 268 | struct iommu_cmd cmd; |
ee2fa743 | 269 | int ret; |
a19ae1ec JR |
270 | |
271 | memset(&cmd, 0, sizeof(cmd)); | |
272 | address &= PAGE_MASK; | |
273 | CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES); | |
274 | cmd.data[1] |= domid; | |
8a456695 | 275 | cmd.data[2] = lower_32_bits(address); |
8ea80d78 | 276 | cmd.data[3] = upper_32_bits(address); |
431b2a20 | 277 | if (s) /* size bit - we flush more than one 4kb page */ |
a19ae1ec | 278 | cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK; |
431b2a20 | 279 | if (pde) /* PDE bit - we wan't flush everything not only the PTEs */ |
a19ae1ec JR |
280 | cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK; |
281 | ||
ee2fa743 JR |
282 | ret = iommu_queue_command(iommu, &cmd); |
283 | ||
a19ae1ec JR |
284 | iommu->need_sync = 1; |
285 | ||
ee2fa743 | 286 | return ret; |
a19ae1ec JR |
287 | } |
288 | ||
431b2a20 JR |
289 | /* |
290 | * TLB invalidation function which is called from the mapping functions. | |
291 | * It invalidates a single PTE if the range to flush is within a single | |
292 | * page. Otherwise it flushes the whole TLB of the IOMMU. | |
293 | */ | |
a19ae1ec JR |
294 | static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid, |
295 | u64 address, size_t size) | |
296 | { | |
999ba417 | 297 | int s = 0; |
e3c449f5 | 298 | unsigned pages = iommu_num_pages(address, size, PAGE_SIZE); |
a19ae1ec JR |
299 | |
300 | address &= PAGE_MASK; | |
301 | ||
999ba417 JR |
302 | if (pages > 1) { |
303 | /* | |
304 | * If we have to flush more than one page, flush all | |
305 | * TLB entries for this domain | |
306 | */ | |
307 | address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS; | |
308 | s = 1; | |
a19ae1ec JR |
309 | } |
310 | ||
999ba417 JR |
311 | iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s); |
312 | ||
a19ae1ec JR |
313 | return 0; |
314 | } | |
b6c02715 | 315 | |
1c655773 JR |
316 | /* Flush the whole IO/TLB for a given protection domain */ |
317 | static void iommu_flush_tlb(struct amd_iommu *iommu, u16 domid) | |
318 | { | |
319 | u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS; | |
320 | ||
321 | iommu_queue_inv_iommu_pages(iommu, address, domid, 0, 1); | |
322 | } | |
323 | ||
431b2a20 JR |
324 | /**************************************************************************** |
325 | * | |
326 | * The functions below are used the create the page table mappings for | |
327 | * unity mapped regions. | |
328 | * | |
329 | ****************************************************************************/ | |
330 | ||
331 | /* | |
332 | * Generic mapping functions. It maps a physical address into a DMA | |
333 | * address space. It allocates the page table pages if necessary. | |
334 | * In the future it can be extended to a generic mapping function | |
335 | * supporting all features of AMD IOMMU page tables like level skipping | |
336 | * and full 64 bit address spaces. | |
337 | */ | |
bd0e5211 JR |
338 | static int iommu_map(struct protection_domain *dom, |
339 | unsigned long bus_addr, | |
340 | unsigned long phys_addr, | |
341 | int prot) | |
342 | { | |
343 | u64 __pte, *pte, *page; | |
344 | ||
345 | bus_addr = PAGE_ALIGN(bus_addr); | |
346 | phys_addr = PAGE_ALIGN(bus_addr); | |
347 | ||
348 | /* only support 512GB address spaces for now */ | |
349 | if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK)) | |
350 | return -EINVAL; | |
351 | ||
352 | pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)]; | |
353 | ||
354 | if (!IOMMU_PTE_PRESENT(*pte)) { | |
355 | page = (u64 *)get_zeroed_page(GFP_KERNEL); | |
356 | if (!page) | |
357 | return -ENOMEM; | |
358 | *pte = IOMMU_L2_PDE(virt_to_phys(page)); | |
359 | } | |
360 | ||
361 | pte = IOMMU_PTE_PAGE(*pte); | |
362 | pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)]; | |
363 | ||
364 | if (!IOMMU_PTE_PRESENT(*pte)) { | |
365 | page = (u64 *)get_zeroed_page(GFP_KERNEL); | |
366 | if (!page) | |
367 | return -ENOMEM; | |
368 | *pte = IOMMU_L1_PDE(virt_to_phys(page)); | |
369 | } | |
370 | ||
371 | pte = IOMMU_PTE_PAGE(*pte); | |
372 | pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)]; | |
373 | ||
374 | if (IOMMU_PTE_PRESENT(*pte)) | |
375 | return -EBUSY; | |
376 | ||
377 | __pte = phys_addr | IOMMU_PTE_P; | |
378 | if (prot & IOMMU_PROT_IR) | |
379 | __pte |= IOMMU_PTE_IR; | |
380 | if (prot & IOMMU_PROT_IW) | |
381 | __pte |= IOMMU_PTE_IW; | |
382 | ||
383 | *pte = __pte; | |
384 | ||
385 | return 0; | |
386 | } | |
387 | ||
431b2a20 JR |
388 | /* |
389 | * This function checks if a specific unity mapping entry is needed for | |
390 | * this specific IOMMU. | |
391 | */ | |
bd0e5211 JR |
392 | static int iommu_for_unity_map(struct amd_iommu *iommu, |
393 | struct unity_map_entry *entry) | |
394 | { | |
395 | u16 bdf, i; | |
396 | ||
397 | for (i = entry->devid_start; i <= entry->devid_end; ++i) { | |
398 | bdf = amd_iommu_alias_table[i]; | |
399 | if (amd_iommu_rlookup_table[bdf] == iommu) | |
400 | return 1; | |
401 | } | |
402 | ||
403 | return 0; | |
404 | } | |
405 | ||
431b2a20 JR |
406 | /* |
407 | * Init the unity mappings for a specific IOMMU in the system | |
408 | * | |
409 | * Basically iterates over all unity mapping entries and applies them to | |
410 | * the default domain DMA of that IOMMU if necessary. | |
411 | */ | |
bd0e5211 JR |
412 | static int iommu_init_unity_mappings(struct amd_iommu *iommu) |
413 | { | |
414 | struct unity_map_entry *entry; | |
415 | int ret; | |
416 | ||
417 | list_for_each_entry(entry, &amd_iommu_unity_map, list) { | |
418 | if (!iommu_for_unity_map(iommu, entry)) | |
419 | continue; | |
420 | ret = dma_ops_unity_map(iommu->default_dom, entry); | |
421 | if (ret) | |
422 | return ret; | |
423 | } | |
424 | ||
425 | return 0; | |
426 | } | |
427 | ||
431b2a20 JR |
428 | /* |
429 | * This function actually applies the mapping to the page table of the | |
430 | * dma_ops domain. | |
431 | */ | |
bd0e5211 JR |
432 | static int dma_ops_unity_map(struct dma_ops_domain *dma_dom, |
433 | struct unity_map_entry *e) | |
434 | { | |
435 | u64 addr; | |
436 | int ret; | |
437 | ||
438 | for (addr = e->address_start; addr < e->address_end; | |
439 | addr += PAGE_SIZE) { | |
440 | ret = iommu_map(&dma_dom->domain, addr, addr, e->prot); | |
441 | if (ret) | |
442 | return ret; | |
443 | /* | |
444 | * if unity mapping is in aperture range mark the page | |
445 | * as allocated in the aperture | |
446 | */ | |
447 | if (addr < dma_dom->aperture_size) | |
448 | __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap); | |
449 | } | |
450 | ||
451 | return 0; | |
452 | } | |
453 | ||
431b2a20 JR |
454 | /* |
455 | * Inits the unity mappings required for a specific device | |
456 | */ | |
bd0e5211 JR |
457 | static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom, |
458 | u16 devid) | |
459 | { | |
460 | struct unity_map_entry *e; | |
461 | int ret; | |
462 | ||
463 | list_for_each_entry(e, &amd_iommu_unity_map, list) { | |
464 | if (!(devid >= e->devid_start && devid <= e->devid_end)) | |
465 | continue; | |
466 | ret = dma_ops_unity_map(dma_dom, e); | |
467 | if (ret) | |
468 | return ret; | |
469 | } | |
470 | ||
471 | return 0; | |
472 | } | |
473 | ||
431b2a20 JR |
474 | /**************************************************************************** |
475 | * | |
476 | * The next functions belong to the address allocator for the dma_ops | |
477 | * interface functions. They work like the allocators in the other IOMMU | |
478 | * drivers. Its basically a bitmap which marks the allocated pages in | |
479 | * the aperture. Maybe it could be enhanced in the future to a more | |
480 | * efficient allocator. | |
481 | * | |
482 | ****************************************************************************/ | |
d3086444 | 483 | |
431b2a20 JR |
484 | /* |
485 | * The address allocator core function. | |
486 | * | |
487 | * called with domain->lock held | |
488 | */ | |
d3086444 JR |
489 | static unsigned long dma_ops_alloc_addresses(struct device *dev, |
490 | struct dma_ops_domain *dom, | |
6d4f343f | 491 | unsigned int pages, |
832a90c3 JR |
492 | unsigned long align_mask, |
493 | u64 dma_mask) | |
d3086444 | 494 | { |
40becd8d | 495 | unsigned long limit; |
d3086444 | 496 | unsigned long address; |
d3086444 JR |
497 | unsigned long boundary_size; |
498 | ||
499 | boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1, | |
500 | PAGE_SIZE) >> PAGE_SHIFT; | |
40becd8d FT |
501 | limit = iommu_device_max_index(dom->aperture_size >> PAGE_SHIFT, 0, |
502 | dma_mask >> PAGE_SHIFT); | |
d3086444 | 503 | |
1c655773 | 504 | if (dom->next_bit >= limit) { |
d3086444 | 505 | dom->next_bit = 0; |
1c655773 JR |
506 | dom->need_flush = true; |
507 | } | |
d3086444 JR |
508 | |
509 | address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages, | |
6d4f343f | 510 | 0 , boundary_size, align_mask); |
1c655773 | 511 | if (address == -1) { |
d3086444 | 512 | address = iommu_area_alloc(dom->bitmap, limit, 0, pages, |
6d4f343f | 513 | 0, boundary_size, align_mask); |
1c655773 JR |
514 | dom->need_flush = true; |
515 | } | |
d3086444 JR |
516 | |
517 | if (likely(address != -1)) { | |
d3086444 JR |
518 | dom->next_bit = address + pages; |
519 | address <<= PAGE_SHIFT; | |
520 | } else | |
521 | address = bad_dma_address; | |
522 | ||
523 | WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size); | |
524 | ||
525 | return address; | |
526 | } | |
527 | ||
431b2a20 JR |
528 | /* |
529 | * The address free function. | |
530 | * | |
531 | * called with domain->lock held | |
532 | */ | |
d3086444 JR |
533 | static void dma_ops_free_addresses(struct dma_ops_domain *dom, |
534 | unsigned long address, | |
535 | unsigned int pages) | |
536 | { | |
537 | address >>= PAGE_SHIFT; | |
538 | iommu_area_free(dom->bitmap, address, pages); | |
539 | } | |
540 | ||
431b2a20 JR |
541 | /**************************************************************************** |
542 | * | |
543 | * The next functions belong to the domain allocation. A domain is | |
544 | * allocated for every IOMMU as the default domain. If device isolation | |
545 | * is enabled, every device get its own domain. The most important thing | |
546 | * about domains is the page table mapping the DMA address space they | |
547 | * contain. | |
548 | * | |
549 | ****************************************************************************/ | |
550 | ||
ec487d1a JR |
551 | static u16 domain_id_alloc(void) |
552 | { | |
553 | unsigned long flags; | |
554 | int id; | |
555 | ||
556 | write_lock_irqsave(&amd_iommu_devtable_lock, flags); | |
557 | id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID); | |
558 | BUG_ON(id == 0); | |
559 | if (id > 0 && id < MAX_DOMAIN_ID) | |
560 | __set_bit(id, amd_iommu_pd_alloc_bitmap); | |
561 | else | |
562 | id = 0; | |
563 | write_unlock_irqrestore(&amd_iommu_devtable_lock, flags); | |
564 | ||
565 | return id; | |
566 | } | |
567 | ||
431b2a20 JR |
568 | /* |
569 | * Used to reserve address ranges in the aperture (e.g. for exclusion | |
570 | * ranges. | |
571 | */ | |
ec487d1a JR |
572 | static void dma_ops_reserve_addresses(struct dma_ops_domain *dom, |
573 | unsigned long start_page, | |
574 | unsigned int pages) | |
575 | { | |
576 | unsigned int last_page = dom->aperture_size >> PAGE_SHIFT; | |
577 | ||
578 | if (start_page + pages > last_page) | |
579 | pages = last_page - start_page; | |
580 | ||
d26dbc5c | 581 | iommu_area_reserve(dom->bitmap, start_page, pages); |
ec487d1a JR |
582 | } |
583 | ||
584 | static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom) | |
585 | { | |
586 | int i, j; | |
587 | u64 *p1, *p2, *p3; | |
588 | ||
589 | p1 = dma_dom->domain.pt_root; | |
590 | ||
591 | if (!p1) | |
592 | return; | |
593 | ||
594 | for (i = 0; i < 512; ++i) { | |
595 | if (!IOMMU_PTE_PRESENT(p1[i])) | |
596 | continue; | |
597 | ||
598 | p2 = IOMMU_PTE_PAGE(p1[i]); | |
599 | for (j = 0; j < 512; ++i) { | |
600 | if (!IOMMU_PTE_PRESENT(p2[j])) | |
601 | continue; | |
602 | p3 = IOMMU_PTE_PAGE(p2[j]); | |
603 | free_page((unsigned long)p3); | |
604 | } | |
605 | ||
606 | free_page((unsigned long)p2); | |
607 | } | |
608 | ||
609 | free_page((unsigned long)p1); | |
610 | } | |
611 | ||
431b2a20 JR |
612 | /* |
613 | * Free a domain, only used if something went wrong in the | |
614 | * allocation path and we need to free an already allocated page table | |
615 | */ | |
ec487d1a JR |
616 | static void dma_ops_domain_free(struct dma_ops_domain *dom) |
617 | { | |
618 | if (!dom) | |
619 | return; | |
620 | ||
621 | dma_ops_free_pagetable(dom); | |
622 | ||
623 | kfree(dom->pte_pages); | |
624 | ||
625 | kfree(dom->bitmap); | |
626 | ||
627 | kfree(dom); | |
628 | } | |
629 | ||
431b2a20 JR |
630 | /* |
631 | * Allocates a new protection domain usable for the dma_ops functions. | |
632 | * It also intializes the page table and the address allocator data | |
633 | * structures required for the dma_ops interface | |
634 | */ | |
ec487d1a JR |
635 | static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu, |
636 | unsigned order) | |
637 | { | |
638 | struct dma_ops_domain *dma_dom; | |
639 | unsigned i, num_pte_pages; | |
640 | u64 *l2_pde; | |
641 | u64 address; | |
642 | ||
643 | /* | |
644 | * Currently the DMA aperture must be between 32 MB and 1GB in size | |
645 | */ | |
646 | if ((order < 25) || (order > 30)) | |
647 | return NULL; | |
648 | ||
649 | dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL); | |
650 | if (!dma_dom) | |
651 | return NULL; | |
652 | ||
653 | spin_lock_init(&dma_dom->domain.lock); | |
654 | ||
655 | dma_dom->domain.id = domain_id_alloc(); | |
656 | if (dma_dom->domain.id == 0) | |
657 | goto free_dma_dom; | |
658 | dma_dom->domain.mode = PAGE_MODE_3_LEVEL; | |
659 | dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL); | |
660 | dma_dom->domain.priv = dma_dom; | |
661 | if (!dma_dom->domain.pt_root) | |
662 | goto free_dma_dom; | |
663 | dma_dom->aperture_size = (1ULL << order); | |
664 | dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8), | |
665 | GFP_KERNEL); | |
666 | if (!dma_dom->bitmap) | |
667 | goto free_dma_dom; | |
668 | /* | |
669 | * mark the first page as allocated so we never return 0 as | |
670 | * a valid dma-address. So we can use 0 as error value | |
671 | */ | |
672 | dma_dom->bitmap[0] = 1; | |
673 | dma_dom->next_bit = 0; | |
674 | ||
1c655773 | 675 | dma_dom->need_flush = false; |
bd60b735 | 676 | dma_dom->target_dev = 0xffff; |
1c655773 | 677 | |
431b2a20 | 678 | /* Intialize the exclusion range if necessary */ |
ec487d1a JR |
679 | if (iommu->exclusion_start && |
680 | iommu->exclusion_start < dma_dom->aperture_size) { | |
681 | unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT; | |
e3c449f5 JR |
682 | int pages = iommu_num_pages(iommu->exclusion_start, |
683 | iommu->exclusion_length, | |
684 | PAGE_SIZE); | |
ec487d1a JR |
685 | dma_ops_reserve_addresses(dma_dom, startpage, pages); |
686 | } | |
687 | ||
431b2a20 JR |
688 | /* |
689 | * At the last step, build the page tables so we don't need to | |
690 | * allocate page table pages in the dma_ops mapping/unmapping | |
691 | * path. | |
692 | */ | |
ec487d1a JR |
693 | num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512); |
694 | dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *), | |
695 | GFP_KERNEL); | |
696 | if (!dma_dom->pte_pages) | |
697 | goto free_dma_dom; | |
698 | ||
699 | l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL); | |
700 | if (l2_pde == NULL) | |
701 | goto free_dma_dom; | |
702 | ||
703 | dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde)); | |
704 | ||
705 | for (i = 0; i < num_pte_pages; ++i) { | |
706 | dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL); | |
707 | if (!dma_dom->pte_pages[i]) | |
708 | goto free_dma_dom; | |
709 | address = virt_to_phys(dma_dom->pte_pages[i]); | |
710 | l2_pde[i] = IOMMU_L1_PDE(address); | |
711 | } | |
712 | ||
713 | return dma_dom; | |
714 | ||
715 | free_dma_dom: | |
716 | dma_ops_domain_free(dma_dom); | |
717 | ||
718 | return NULL; | |
719 | } | |
720 | ||
431b2a20 JR |
721 | /* |
722 | * Find out the protection domain structure for a given PCI device. This | |
723 | * will give us the pointer to the page table root for example. | |
724 | */ | |
b20ac0d4 JR |
725 | static struct protection_domain *domain_for_device(u16 devid) |
726 | { | |
727 | struct protection_domain *dom; | |
728 | unsigned long flags; | |
729 | ||
730 | read_lock_irqsave(&amd_iommu_devtable_lock, flags); | |
731 | dom = amd_iommu_pd_table[devid]; | |
732 | read_unlock_irqrestore(&amd_iommu_devtable_lock, flags); | |
733 | ||
734 | return dom; | |
735 | } | |
736 | ||
431b2a20 JR |
737 | /* |
738 | * If a device is not yet associated with a domain, this function does | |
739 | * assigns it visible for the hardware | |
740 | */ | |
b20ac0d4 JR |
741 | static void set_device_domain(struct amd_iommu *iommu, |
742 | struct protection_domain *domain, | |
743 | u16 devid) | |
744 | { | |
745 | unsigned long flags; | |
746 | ||
747 | u64 pte_root = virt_to_phys(domain->pt_root); | |
748 | ||
38ddf41b JR |
749 | pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK) |
750 | << DEV_ENTRY_MODE_SHIFT; | |
751 | pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV; | |
b20ac0d4 JR |
752 | |
753 | write_lock_irqsave(&amd_iommu_devtable_lock, flags); | |
38ddf41b JR |
754 | amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root); |
755 | amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root); | |
b20ac0d4 JR |
756 | amd_iommu_dev_table[devid].data[2] = domain->id; |
757 | ||
758 | amd_iommu_pd_table[devid] = domain; | |
759 | write_unlock_irqrestore(&amd_iommu_devtable_lock, flags); | |
760 | ||
761 | iommu_queue_inv_dev_entry(iommu, devid); | |
762 | ||
763 | iommu->need_sync = 1; | |
764 | } | |
765 | ||
431b2a20 JR |
766 | /***************************************************************************** |
767 | * | |
768 | * The next functions belong to the dma_ops mapping/unmapping code. | |
769 | * | |
770 | *****************************************************************************/ | |
771 | ||
dbcc112e JR |
772 | /* |
773 | * This function checks if the driver got a valid device from the caller to | |
774 | * avoid dereferencing invalid pointers. | |
775 | */ | |
776 | static bool check_device(struct device *dev) | |
777 | { | |
778 | if (!dev || !dev->dma_mask) | |
779 | return false; | |
780 | ||
781 | return true; | |
782 | } | |
783 | ||
bd60b735 JR |
784 | /* |
785 | * In this function the list of preallocated protection domains is traversed to | |
786 | * find the domain for a specific device | |
787 | */ | |
788 | static struct dma_ops_domain *find_protection_domain(u16 devid) | |
789 | { | |
790 | struct dma_ops_domain *entry, *ret = NULL; | |
791 | unsigned long flags; | |
792 | ||
793 | if (list_empty(&iommu_pd_list)) | |
794 | return NULL; | |
795 | ||
796 | spin_lock_irqsave(&iommu_pd_list_lock, flags); | |
797 | ||
798 | list_for_each_entry(entry, &iommu_pd_list, list) { | |
799 | if (entry->target_dev == devid) { | |
800 | ret = entry; | |
801 | list_del(&ret->list); | |
802 | break; | |
803 | } | |
804 | } | |
805 | ||
806 | spin_unlock_irqrestore(&iommu_pd_list_lock, flags); | |
807 | ||
808 | return ret; | |
809 | } | |
810 | ||
431b2a20 JR |
811 | /* |
812 | * In the dma_ops path we only have the struct device. This function | |
813 | * finds the corresponding IOMMU, the protection domain and the | |
814 | * requestor id for a given device. | |
815 | * If the device is not yet associated with a domain this is also done | |
816 | * in this function. | |
817 | */ | |
b20ac0d4 JR |
818 | static int get_device_resources(struct device *dev, |
819 | struct amd_iommu **iommu, | |
820 | struct protection_domain **domain, | |
821 | u16 *bdf) | |
822 | { | |
823 | struct dma_ops_domain *dma_dom; | |
824 | struct pci_dev *pcidev; | |
825 | u16 _bdf; | |
826 | ||
dbcc112e JR |
827 | *iommu = NULL; |
828 | *domain = NULL; | |
829 | *bdf = 0xffff; | |
830 | ||
831 | if (dev->bus != &pci_bus_type) | |
832 | return 0; | |
b20ac0d4 JR |
833 | |
834 | pcidev = to_pci_dev(dev); | |
d591b0a3 | 835 | _bdf = calc_devid(pcidev->bus->number, pcidev->devfn); |
b20ac0d4 | 836 | |
431b2a20 | 837 | /* device not translated by any IOMMU in the system? */ |
dbcc112e | 838 | if (_bdf > amd_iommu_last_bdf) |
b20ac0d4 | 839 | return 0; |
b20ac0d4 JR |
840 | |
841 | *bdf = amd_iommu_alias_table[_bdf]; | |
842 | ||
843 | *iommu = amd_iommu_rlookup_table[*bdf]; | |
844 | if (*iommu == NULL) | |
845 | return 0; | |
b20ac0d4 JR |
846 | *domain = domain_for_device(*bdf); |
847 | if (*domain == NULL) { | |
bd60b735 JR |
848 | dma_dom = find_protection_domain(*bdf); |
849 | if (!dma_dom) | |
850 | dma_dom = (*iommu)->default_dom; | |
b20ac0d4 JR |
851 | *domain = &dma_dom->domain; |
852 | set_device_domain(*iommu, *domain, *bdf); | |
853 | printk(KERN_INFO "AMD IOMMU: Using protection domain %d for " | |
854 | "device ", (*domain)->id); | |
855 | print_devid(_bdf, 1); | |
856 | } | |
857 | ||
858 | return 1; | |
859 | } | |
860 | ||
431b2a20 JR |
861 | /* |
862 | * This is the generic map function. It maps one 4kb page at paddr to | |
863 | * the given address in the DMA address space for the domain. | |
864 | */ | |
cb76c322 JR |
865 | static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu, |
866 | struct dma_ops_domain *dom, | |
867 | unsigned long address, | |
868 | phys_addr_t paddr, | |
869 | int direction) | |
870 | { | |
871 | u64 *pte, __pte; | |
872 | ||
873 | WARN_ON(address > dom->aperture_size); | |
874 | ||
875 | paddr &= PAGE_MASK; | |
876 | ||
877 | pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)]; | |
878 | pte += IOMMU_PTE_L0_INDEX(address); | |
879 | ||
880 | __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC; | |
881 | ||
882 | if (direction == DMA_TO_DEVICE) | |
883 | __pte |= IOMMU_PTE_IR; | |
884 | else if (direction == DMA_FROM_DEVICE) | |
885 | __pte |= IOMMU_PTE_IW; | |
886 | else if (direction == DMA_BIDIRECTIONAL) | |
887 | __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW; | |
888 | ||
889 | WARN_ON(*pte); | |
890 | ||
891 | *pte = __pte; | |
892 | ||
893 | return (dma_addr_t)address; | |
894 | } | |
895 | ||
431b2a20 JR |
896 | /* |
897 | * The generic unmapping function for on page in the DMA address space. | |
898 | */ | |
cb76c322 JR |
899 | static void dma_ops_domain_unmap(struct amd_iommu *iommu, |
900 | struct dma_ops_domain *dom, | |
901 | unsigned long address) | |
902 | { | |
903 | u64 *pte; | |
904 | ||
905 | if (address >= dom->aperture_size) | |
906 | return; | |
907 | ||
908 | WARN_ON(address & 0xfffULL || address > dom->aperture_size); | |
909 | ||
910 | pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)]; | |
911 | pte += IOMMU_PTE_L0_INDEX(address); | |
912 | ||
913 | WARN_ON(!*pte); | |
914 | ||
915 | *pte = 0ULL; | |
916 | } | |
917 | ||
431b2a20 JR |
918 | /* |
919 | * This function contains common code for mapping of a physically | |
920 | * contiguous memory region into DMA address space. It is uses by all | |
921 | * mapping functions provided by this IOMMU driver. | |
922 | * Must be called with the domain lock held. | |
923 | */ | |
cb76c322 JR |
924 | static dma_addr_t __map_single(struct device *dev, |
925 | struct amd_iommu *iommu, | |
926 | struct dma_ops_domain *dma_dom, | |
927 | phys_addr_t paddr, | |
928 | size_t size, | |
6d4f343f | 929 | int dir, |
832a90c3 JR |
930 | bool align, |
931 | u64 dma_mask) | |
cb76c322 JR |
932 | { |
933 | dma_addr_t offset = paddr & ~PAGE_MASK; | |
934 | dma_addr_t address, start; | |
935 | unsigned int pages; | |
6d4f343f | 936 | unsigned long align_mask = 0; |
cb76c322 JR |
937 | int i; |
938 | ||
e3c449f5 | 939 | pages = iommu_num_pages(paddr, size, PAGE_SIZE); |
cb76c322 JR |
940 | paddr &= PAGE_MASK; |
941 | ||
6d4f343f JR |
942 | if (align) |
943 | align_mask = (1UL << get_order(size)) - 1; | |
944 | ||
832a90c3 JR |
945 | address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask, |
946 | dma_mask); | |
cb76c322 JR |
947 | if (unlikely(address == bad_dma_address)) |
948 | goto out; | |
949 | ||
950 | start = address; | |
951 | for (i = 0; i < pages; ++i) { | |
952 | dma_ops_domain_map(iommu, dma_dom, start, paddr, dir); | |
953 | paddr += PAGE_SIZE; | |
954 | start += PAGE_SIZE; | |
955 | } | |
956 | address += offset; | |
957 | ||
afa9fdc2 | 958 | if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) { |
1c655773 JR |
959 | iommu_flush_tlb(iommu, dma_dom->domain.id); |
960 | dma_dom->need_flush = false; | |
961 | } else if (unlikely(iommu_has_npcache(iommu))) | |
270cab24 JR |
962 | iommu_flush_pages(iommu, dma_dom->domain.id, address, size); |
963 | ||
cb76c322 JR |
964 | out: |
965 | return address; | |
966 | } | |
967 | ||
431b2a20 JR |
968 | /* |
969 | * Does the reverse of the __map_single function. Must be called with | |
970 | * the domain lock held too | |
971 | */ | |
cb76c322 JR |
972 | static void __unmap_single(struct amd_iommu *iommu, |
973 | struct dma_ops_domain *dma_dom, | |
974 | dma_addr_t dma_addr, | |
975 | size_t size, | |
976 | int dir) | |
977 | { | |
978 | dma_addr_t i, start; | |
979 | unsigned int pages; | |
980 | ||
981 | if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size)) | |
982 | return; | |
983 | ||
e3c449f5 | 984 | pages = iommu_num_pages(dma_addr, size, PAGE_SIZE); |
cb76c322 JR |
985 | dma_addr &= PAGE_MASK; |
986 | start = dma_addr; | |
987 | ||
988 | for (i = 0; i < pages; ++i) { | |
989 | dma_ops_domain_unmap(iommu, dma_dom, start); | |
990 | start += PAGE_SIZE; | |
991 | } | |
992 | ||
993 | dma_ops_free_addresses(dma_dom, dma_addr, pages); | |
270cab24 | 994 | |
afa9fdc2 | 995 | if (amd_iommu_unmap_flush) |
1c655773 | 996 | iommu_flush_pages(iommu, dma_dom->domain.id, dma_addr, size); |
cb76c322 JR |
997 | } |
998 | ||
431b2a20 JR |
999 | /* |
1000 | * The exported map_single function for dma_ops. | |
1001 | */ | |
4da70b9e JR |
1002 | static dma_addr_t map_single(struct device *dev, phys_addr_t paddr, |
1003 | size_t size, int dir) | |
1004 | { | |
1005 | unsigned long flags; | |
1006 | struct amd_iommu *iommu; | |
1007 | struct protection_domain *domain; | |
1008 | u16 devid; | |
1009 | dma_addr_t addr; | |
832a90c3 | 1010 | u64 dma_mask; |
4da70b9e | 1011 | |
dbcc112e JR |
1012 | if (!check_device(dev)) |
1013 | return bad_dma_address; | |
1014 | ||
832a90c3 | 1015 | dma_mask = *dev->dma_mask; |
4da70b9e JR |
1016 | |
1017 | get_device_resources(dev, &iommu, &domain, &devid); | |
1018 | ||
1019 | if (iommu == NULL || domain == NULL) | |
431b2a20 | 1020 | /* device not handled by any AMD IOMMU */ |
4da70b9e JR |
1021 | return (dma_addr_t)paddr; |
1022 | ||
1023 | spin_lock_irqsave(&domain->lock, flags); | |
832a90c3 JR |
1024 | addr = __map_single(dev, iommu, domain->priv, paddr, size, dir, false, |
1025 | dma_mask); | |
4da70b9e JR |
1026 | if (addr == bad_dma_address) |
1027 | goto out; | |
1028 | ||
5507eef8 | 1029 | if (unlikely(iommu->need_sync)) |
4da70b9e JR |
1030 | iommu_completion_wait(iommu); |
1031 | ||
1032 | out: | |
1033 | spin_unlock_irqrestore(&domain->lock, flags); | |
1034 | ||
1035 | return addr; | |
1036 | } | |
1037 | ||
431b2a20 JR |
1038 | /* |
1039 | * The exported unmap_single function for dma_ops. | |
1040 | */ | |
4da70b9e JR |
1041 | static void unmap_single(struct device *dev, dma_addr_t dma_addr, |
1042 | size_t size, int dir) | |
1043 | { | |
1044 | unsigned long flags; | |
1045 | struct amd_iommu *iommu; | |
1046 | struct protection_domain *domain; | |
1047 | u16 devid; | |
1048 | ||
dbcc112e JR |
1049 | if (!check_device(dev) || |
1050 | !get_device_resources(dev, &iommu, &domain, &devid)) | |
431b2a20 | 1051 | /* device not handled by any AMD IOMMU */ |
4da70b9e JR |
1052 | return; |
1053 | ||
1054 | spin_lock_irqsave(&domain->lock, flags); | |
1055 | ||
1056 | __unmap_single(iommu, domain->priv, dma_addr, size, dir); | |
1057 | ||
5507eef8 | 1058 | if (unlikely(iommu->need_sync)) |
4da70b9e JR |
1059 | iommu_completion_wait(iommu); |
1060 | ||
1061 | spin_unlock_irqrestore(&domain->lock, flags); | |
1062 | } | |
1063 | ||
431b2a20 JR |
1064 | /* |
1065 | * This is a special map_sg function which is used if we should map a | |
1066 | * device which is not handled by an AMD IOMMU in the system. | |
1067 | */ | |
65b050ad JR |
1068 | static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist, |
1069 | int nelems, int dir) | |
1070 | { | |
1071 | struct scatterlist *s; | |
1072 | int i; | |
1073 | ||
1074 | for_each_sg(sglist, s, nelems, i) { | |
1075 | s->dma_address = (dma_addr_t)sg_phys(s); | |
1076 | s->dma_length = s->length; | |
1077 | } | |
1078 | ||
1079 | return nelems; | |
1080 | } | |
1081 | ||
431b2a20 JR |
1082 | /* |
1083 | * The exported map_sg function for dma_ops (handles scatter-gather | |
1084 | * lists). | |
1085 | */ | |
65b050ad JR |
1086 | static int map_sg(struct device *dev, struct scatterlist *sglist, |
1087 | int nelems, int dir) | |
1088 | { | |
1089 | unsigned long flags; | |
1090 | struct amd_iommu *iommu; | |
1091 | struct protection_domain *domain; | |
1092 | u16 devid; | |
1093 | int i; | |
1094 | struct scatterlist *s; | |
1095 | phys_addr_t paddr; | |
1096 | int mapped_elems = 0; | |
832a90c3 | 1097 | u64 dma_mask; |
65b050ad | 1098 | |
dbcc112e JR |
1099 | if (!check_device(dev)) |
1100 | return 0; | |
1101 | ||
832a90c3 | 1102 | dma_mask = *dev->dma_mask; |
65b050ad JR |
1103 | |
1104 | get_device_resources(dev, &iommu, &domain, &devid); | |
1105 | ||
1106 | if (!iommu || !domain) | |
1107 | return map_sg_no_iommu(dev, sglist, nelems, dir); | |
1108 | ||
1109 | spin_lock_irqsave(&domain->lock, flags); | |
1110 | ||
1111 | for_each_sg(sglist, s, nelems, i) { | |
1112 | paddr = sg_phys(s); | |
1113 | ||
1114 | s->dma_address = __map_single(dev, iommu, domain->priv, | |
832a90c3 JR |
1115 | paddr, s->length, dir, false, |
1116 | dma_mask); | |
65b050ad JR |
1117 | |
1118 | if (s->dma_address) { | |
1119 | s->dma_length = s->length; | |
1120 | mapped_elems++; | |
1121 | } else | |
1122 | goto unmap; | |
65b050ad JR |
1123 | } |
1124 | ||
5507eef8 | 1125 | if (unlikely(iommu->need_sync)) |
65b050ad JR |
1126 | iommu_completion_wait(iommu); |
1127 | ||
1128 | out: | |
1129 | spin_unlock_irqrestore(&domain->lock, flags); | |
1130 | ||
1131 | return mapped_elems; | |
1132 | unmap: | |
1133 | for_each_sg(sglist, s, mapped_elems, i) { | |
1134 | if (s->dma_address) | |
1135 | __unmap_single(iommu, domain->priv, s->dma_address, | |
1136 | s->dma_length, dir); | |
1137 | s->dma_address = s->dma_length = 0; | |
1138 | } | |
1139 | ||
1140 | mapped_elems = 0; | |
1141 | ||
1142 | goto out; | |
1143 | } | |
1144 | ||
431b2a20 JR |
1145 | /* |
1146 | * The exported map_sg function for dma_ops (handles scatter-gather | |
1147 | * lists). | |
1148 | */ | |
65b050ad JR |
1149 | static void unmap_sg(struct device *dev, struct scatterlist *sglist, |
1150 | int nelems, int dir) | |
1151 | { | |
1152 | unsigned long flags; | |
1153 | struct amd_iommu *iommu; | |
1154 | struct protection_domain *domain; | |
1155 | struct scatterlist *s; | |
1156 | u16 devid; | |
1157 | int i; | |
1158 | ||
dbcc112e JR |
1159 | if (!check_device(dev) || |
1160 | !get_device_resources(dev, &iommu, &domain, &devid)) | |
65b050ad JR |
1161 | return; |
1162 | ||
1163 | spin_lock_irqsave(&domain->lock, flags); | |
1164 | ||
1165 | for_each_sg(sglist, s, nelems, i) { | |
1166 | __unmap_single(iommu, domain->priv, s->dma_address, | |
1167 | s->dma_length, dir); | |
65b050ad JR |
1168 | s->dma_address = s->dma_length = 0; |
1169 | } | |
1170 | ||
5507eef8 | 1171 | if (unlikely(iommu->need_sync)) |
65b050ad JR |
1172 | iommu_completion_wait(iommu); |
1173 | ||
1174 | spin_unlock_irqrestore(&domain->lock, flags); | |
1175 | } | |
1176 | ||
431b2a20 JR |
1177 | /* |
1178 | * The exported alloc_coherent function for dma_ops. | |
1179 | */ | |
5d8b53cf JR |
1180 | static void *alloc_coherent(struct device *dev, size_t size, |
1181 | dma_addr_t *dma_addr, gfp_t flag) | |
1182 | { | |
1183 | unsigned long flags; | |
1184 | void *virt_addr; | |
1185 | struct amd_iommu *iommu; | |
1186 | struct protection_domain *domain; | |
1187 | u16 devid; | |
1188 | phys_addr_t paddr; | |
832a90c3 | 1189 | u64 dma_mask = dev->coherent_dma_mask; |
5d8b53cf | 1190 | |
dbcc112e JR |
1191 | if (!check_device(dev)) |
1192 | return NULL; | |
5d8b53cf | 1193 | |
13d9fead FT |
1194 | if (!get_device_resources(dev, &iommu, &domain, &devid)) |
1195 | flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32); | |
5d8b53cf | 1196 | |
c97ac535 | 1197 | flag |= __GFP_ZERO; |
5d8b53cf JR |
1198 | virt_addr = (void *)__get_free_pages(flag, get_order(size)); |
1199 | if (!virt_addr) | |
1200 | return 0; | |
1201 | ||
5d8b53cf JR |
1202 | paddr = virt_to_phys(virt_addr); |
1203 | ||
5d8b53cf JR |
1204 | if (!iommu || !domain) { |
1205 | *dma_addr = (dma_addr_t)paddr; | |
1206 | return virt_addr; | |
1207 | } | |
1208 | ||
832a90c3 JR |
1209 | if (!dma_mask) |
1210 | dma_mask = *dev->dma_mask; | |
1211 | ||
5d8b53cf JR |
1212 | spin_lock_irqsave(&domain->lock, flags); |
1213 | ||
1214 | *dma_addr = __map_single(dev, iommu, domain->priv, paddr, | |
832a90c3 | 1215 | size, DMA_BIDIRECTIONAL, true, dma_mask); |
5d8b53cf JR |
1216 | |
1217 | if (*dma_addr == bad_dma_address) { | |
1218 | free_pages((unsigned long)virt_addr, get_order(size)); | |
1219 | virt_addr = NULL; | |
1220 | goto out; | |
1221 | } | |
1222 | ||
5507eef8 | 1223 | if (unlikely(iommu->need_sync)) |
5d8b53cf JR |
1224 | iommu_completion_wait(iommu); |
1225 | ||
1226 | out: | |
1227 | spin_unlock_irqrestore(&domain->lock, flags); | |
1228 | ||
1229 | return virt_addr; | |
1230 | } | |
1231 | ||
431b2a20 JR |
1232 | /* |
1233 | * The exported free_coherent function for dma_ops. | |
431b2a20 | 1234 | */ |
5d8b53cf JR |
1235 | static void free_coherent(struct device *dev, size_t size, |
1236 | void *virt_addr, dma_addr_t dma_addr) | |
1237 | { | |
1238 | unsigned long flags; | |
1239 | struct amd_iommu *iommu; | |
1240 | struct protection_domain *domain; | |
1241 | u16 devid; | |
1242 | ||
dbcc112e JR |
1243 | if (!check_device(dev)) |
1244 | return; | |
1245 | ||
5d8b53cf JR |
1246 | get_device_resources(dev, &iommu, &domain, &devid); |
1247 | ||
1248 | if (!iommu || !domain) | |
1249 | goto free_mem; | |
1250 | ||
1251 | spin_lock_irqsave(&domain->lock, flags); | |
1252 | ||
1253 | __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL); | |
5d8b53cf | 1254 | |
5507eef8 | 1255 | if (unlikely(iommu->need_sync)) |
5d8b53cf JR |
1256 | iommu_completion_wait(iommu); |
1257 | ||
1258 | spin_unlock_irqrestore(&domain->lock, flags); | |
1259 | ||
1260 | free_mem: | |
1261 | free_pages((unsigned long)virt_addr, get_order(size)); | |
1262 | } | |
1263 | ||
b39ba6ad JR |
1264 | /* |
1265 | * This function is called by the DMA layer to find out if we can handle a | |
1266 | * particular device. It is part of the dma_ops. | |
1267 | */ | |
1268 | static int amd_iommu_dma_supported(struct device *dev, u64 mask) | |
1269 | { | |
1270 | u16 bdf; | |
1271 | struct pci_dev *pcidev; | |
1272 | ||
1273 | /* No device or no PCI device */ | |
1274 | if (!dev || dev->bus != &pci_bus_type) | |
1275 | return 0; | |
1276 | ||
1277 | pcidev = to_pci_dev(dev); | |
1278 | ||
1279 | bdf = calc_devid(pcidev->bus->number, pcidev->devfn); | |
1280 | ||
1281 | /* Out of our scope? */ | |
1282 | if (bdf > amd_iommu_last_bdf) | |
1283 | return 0; | |
1284 | ||
1285 | return 1; | |
1286 | } | |
1287 | ||
c432f3df | 1288 | /* |
431b2a20 JR |
1289 | * The function for pre-allocating protection domains. |
1290 | * | |
c432f3df JR |
1291 | * If the driver core informs the DMA layer if a driver grabs a device |
1292 | * we don't need to preallocate the protection domains anymore. | |
1293 | * For now we have to. | |
1294 | */ | |
1295 | void prealloc_protection_domains(void) | |
1296 | { | |
1297 | struct pci_dev *dev = NULL; | |
1298 | struct dma_ops_domain *dma_dom; | |
1299 | struct amd_iommu *iommu; | |
1300 | int order = amd_iommu_aperture_order; | |
1301 | u16 devid; | |
1302 | ||
1303 | while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) { | |
1304 | devid = (dev->bus->number << 8) | dev->devfn; | |
3a61ec38 | 1305 | if (devid > amd_iommu_last_bdf) |
c432f3df JR |
1306 | continue; |
1307 | devid = amd_iommu_alias_table[devid]; | |
1308 | if (domain_for_device(devid)) | |
1309 | continue; | |
1310 | iommu = amd_iommu_rlookup_table[devid]; | |
1311 | if (!iommu) | |
1312 | continue; | |
1313 | dma_dom = dma_ops_domain_alloc(iommu, order); | |
1314 | if (!dma_dom) | |
1315 | continue; | |
1316 | init_unity_mappings_for_device(dma_dom, devid); | |
bd60b735 JR |
1317 | dma_dom->target_dev = devid; |
1318 | ||
1319 | list_add_tail(&dma_dom->list, &iommu_pd_list); | |
c432f3df JR |
1320 | } |
1321 | } | |
1322 | ||
6631ee9d JR |
1323 | static struct dma_mapping_ops amd_iommu_dma_ops = { |
1324 | .alloc_coherent = alloc_coherent, | |
1325 | .free_coherent = free_coherent, | |
1326 | .map_single = map_single, | |
1327 | .unmap_single = unmap_single, | |
1328 | .map_sg = map_sg, | |
1329 | .unmap_sg = unmap_sg, | |
b39ba6ad | 1330 | .dma_supported = amd_iommu_dma_supported, |
6631ee9d JR |
1331 | }; |
1332 | ||
431b2a20 JR |
1333 | /* |
1334 | * The function which clues the AMD IOMMU driver into dma_ops. | |
1335 | */ | |
6631ee9d JR |
1336 | int __init amd_iommu_init_dma_ops(void) |
1337 | { | |
1338 | struct amd_iommu *iommu; | |
1339 | int order = amd_iommu_aperture_order; | |
1340 | int ret; | |
1341 | ||
431b2a20 JR |
1342 | /* |
1343 | * first allocate a default protection domain for every IOMMU we | |
1344 | * found in the system. Devices not assigned to any other | |
1345 | * protection domain will be assigned to the default one. | |
1346 | */ | |
6631ee9d JR |
1347 | list_for_each_entry(iommu, &amd_iommu_list, list) { |
1348 | iommu->default_dom = dma_ops_domain_alloc(iommu, order); | |
1349 | if (iommu->default_dom == NULL) | |
1350 | return -ENOMEM; | |
1351 | ret = iommu_init_unity_mappings(iommu); | |
1352 | if (ret) | |
1353 | goto free_domains; | |
1354 | } | |
1355 | ||
431b2a20 JR |
1356 | /* |
1357 | * If device isolation is enabled, pre-allocate the protection | |
1358 | * domains for each device. | |
1359 | */ | |
6631ee9d JR |
1360 | if (amd_iommu_isolate) |
1361 | prealloc_protection_domains(); | |
1362 | ||
1363 | iommu_detected = 1; | |
1364 | force_iommu = 1; | |
1365 | bad_dma_address = 0; | |
92af4e29 | 1366 | #ifdef CONFIG_GART_IOMMU |
6631ee9d JR |
1367 | gart_iommu_aperture_disabled = 1; |
1368 | gart_iommu_aperture = 0; | |
92af4e29 | 1369 | #endif |
6631ee9d | 1370 | |
431b2a20 | 1371 | /* Make the driver finally visible to the drivers */ |
6631ee9d JR |
1372 | dma_ops = &amd_iommu_dma_ops; |
1373 | ||
1374 | return 0; | |
1375 | ||
1376 | free_domains: | |
1377 | ||
1378 | list_for_each_entry(iommu, &amd_iommu_list, list) { | |
1379 | if (iommu->default_dom) | |
1380 | dma_ops_domain_free(iommu->default_dom); | |
1381 | } | |
1382 | ||
1383 | return ret; | |
1384 | } |