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 | { | |
ae9b9403 | 53 | return iommu->cap & (1UL << IOMMU_CAP_NPCACHE); |
4da70b9e JR |
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 | { | |
519c31ba JR |
204 | int ret, ready = 0; |
205 | unsigned status = 0; | |
d6449536 | 206 | struct iommu_cmd cmd; |
136f78a1 | 207 | unsigned long 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 | ||
215 | ret = iommu_queue_command(iommu, &cmd); | |
216 | ||
217 | if (ret) | |
218 | return ret; | |
219 | ||
136f78a1 JR |
220 | while (!ready && (i < EXIT_LOOP_COUNT)) { |
221 | ++i; | |
519c31ba JR |
222 | /* wait for the bit to become one */ |
223 | status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); | |
224 | ready = status & MMIO_STATUS_COM_WAIT_INT_MASK; | |
136f78a1 JR |
225 | } |
226 | ||
519c31ba JR |
227 | /* set bit back to zero */ |
228 | status &= ~MMIO_STATUS_COM_WAIT_INT_MASK; | |
229 | writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET); | |
230 | ||
136f78a1 JR |
231 | if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit())) |
232 | printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n"); | |
a19ae1ec JR |
233 | |
234 | return 0; | |
235 | } | |
236 | ||
431b2a20 JR |
237 | /* |
238 | * Command send function for invalidating a device table entry | |
239 | */ | |
a19ae1ec JR |
240 | static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid) |
241 | { | |
d6449536 | 242 | struct iommu_cmd cmd; |
a19ae1ec JR |
243 | |
244 | BUG_ON(iommu == NULL); | |
245 | ||
246 | memset(&cmd, 0, sizeof(cmd)); | |
247 | CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY); | |
248 | cmd.data[0] = devid; | |
249 | ||
250 | iommu->need_sync = 1; | |
251 | ||
252 | return iommu_queue_command(iommu, &cmd); | |
253 | } | |
254 | ||
431b2a20 JR |
255 | /* |
256 | * Generic command send function for invalidaing TLB entries | |
257 | */ | |
a19ae1ec JR |
258 | static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu, |
259 | u64 address, u16 domid, int pde, int s) | |
260 | { | |
d6449536 | 261 | struct iommu_cmd cmd; |
a19ae1ec JR |
262 | |
263 | memset(&cmd, 0, sizeof(cmd)); | |
264 | address &= PAGE_MASK; | |
265 | CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES); | |
266 | cmd.data[1] |= domid; | |
8a456695 | 267 | cmd.data[2] = lower_32_bits(address); |
8ea80d78 | 268 | cmd.data[3] = upper_32_bits(address); |
431b2a20 | 269 | if (s) /* size bit - we flush more than one 4kb page */ |
a19ae1ec | 270 | cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK; |
431b2a20 | 271 | if (pde) /* PDE bit - we wan't flush everything not only the PTEs */ |
a19ae1ec JR |
272 | cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK; |
273 | ||
274 | iommu->need_sync = 1; | |
275 | ||
276 | return iommu_queue_command(iommu, &cmd); | |
277 | } | |
278 | ||
431b2a20 JR |
279 | /* |
280 | * TLB invalidation function which is called from the mapping functions. | |
281 | * It invalidates a single PTE if the range to flush is within a single | |
282 | * page. Otherwise it flushes the whole TLB of the IOMMU. | |
283 | */ | |
a19ae1ec JR |
284 | static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid, |
285 | u64 address, size_t size) | |
286 | { | |
999ba417 | 287 | int s = 0; |
a8132e5f | 288 | unsigned pages = iommu_num_pages(address, size); |
a19ae1ec JR |
289 | |
290 | address &= PAGE_MASK; | |
291 | ||
999ba417 JR |
292 | if (pages > 1) { |
293 | /* | |
294 | * If we have to flush more than one page, flush all | |
295 | * TLB entries for this domain | |
296 | */ | |
297 | address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS; | |
298 | s = 1; | |
a19ae1ec JR |
299 | } |
300 | ||
999ba417 JR |
301 | iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s); |
302 | ||
a19ae1ec JR |
303 | return 0; |
304 | } | |
b6c02715 | 305 | |
1c655773 JR |
306 | /* Flush the whole IO/TLB for a given protection domain */ |
307 | static void iommu_flush_tlb(struct amd_iommu *iommu, u16 domid) | |
308 | { | |
309 | u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS; | |
310 | ||
311 | iommu_queue_inv_iommu_pages(iommu, address, domid, 0, 1); | |
312 | } | |
313 | ||
431b2a20 JR |
314 | /**************************************************************************** |
315 | * | |
316 | * The functions below are used the create the page table mappings for | |
317 | * unity mapped regions. | |
318 | * | |
319 | ****************************************************************************/ | |
320 | ||
321 | /* | |
322 | * Generic mapping functions. It maps a physical address into a DMA | |
323 | * address space. It allocates the page table pages if necessary. | |
324 | * In the future it can be extended to a generic mapping function | |
325 | * supporting all features of AMD IOMMU page tables like level skipping | |
326 | * and full 64 bit address spaces. | |
327 | */ | |
bd0e5211 JR |
328 | static int iommu_map(struct protection_domain *dom, |
329 | unsigned long bus_addr, | |
330 | unsigned long phys_addr, | |
331 | int prot) | |
332 | { | |
333 | u64 __pte, *pte, *page; | |
334 | ||
335 | bus_addr = PAGE_ALIGN(bus_addr); | |
336 | phys_addr = PAGE_ALIGN(bus_addr); | |
337 | ||
338 | /* only support 512GB address spaces for now */ | |
339 | if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK)) | |
340 | return -EINVAL; | |
341 | ||
342 | pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)]; | |
343 | ||
344 | if (!IOMMU_PTE_PRESENT(*pte)) { | |
345 | page = (u64 *)get_zeroed_page(GFP_KERNEL); | |
346 | if (!page) | |
347 | return -ENOMEM; | |
348 | *pte = IOMMU_L2_PDE(virt_to_phys(page)); | |
349 | } | |
350 | ||
351 | pte = IOMMU_PTE_PAGE(*pte); | |
352 | pte = &pte[IOMMU_PTE_L1_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_L1_PDE(virt_to_phys(page)); | |
359 | } | |
360 | ||
361 | pte = IOMMU_PTE_PAGE(*pte); | |
362 | pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)]; | |
363 | ||
364 | if (IOMMU_PTE_PRESENT(*pte)) | |
365 | return -EBUSY; | |
366 | ||
367 | __pte = phys_addr | IOMMU_PTE_P; | |
368 | if (prot & IOMMU_PROT_IR) | |
369 | __pte |= IOMMU_PTE_IR; | |
370 | if (prot & IOMMU_PROT_IW) | |
371 | __pte |= IOMMU_PTE_IW; | |
372 | ||
373 | *pte = __pte; | |
374 | ||
375 | return 0; | |
376 | } | |
377 | ||
431b2a20 JR |
378 | /* |
379 | * This function checks if a specific unity mapping entry is needed for | |
380 | * this specific IOMMU. | |
381 | */ | |
bd0e5211 JR |
382 | static int iommu_for_unity_map(struct amd_iommu *iommu, |
383 | struct unity_map_entry *entry) | |
384 | { | |
385 | u16 bdf, i; | |
386 | ||
387 | for (i = entry->devid_start; i <= entry->devid_end; ++i) { | |
388 | bdf = amd_iommu_alias_table[i]; | |
389 | if (amd_iommu_rlookup_table[bdf] == iommu) | |
390 | return 1; | |
391 | } | |
392 | ||
393 | return 0; | |
394 | } | |
395 | ||
431b2a20 JR |
396 | /* |
397 | * Init the unity mappings for a specific IOMMU in the system | |
398 | * | |
399 | * Basically iterates over all unity mapping entries and applies them to | |
400 | * the default domain DMA of that IOMMU if necessary. | |
401 | */ | |
bd0e5211 JR |
402 | static int iommu_init_unity_mappings(struct amd_iommu *iommu) |
403 | { | |
404 | struct unity_map_entry *entry; | |
405 | int ret; | |
406 | ||
407 | list_for_each_entry(entry, &amd_iommu_unity_map, list) { | |
408 | if (!iommu_for_unity_map(iommu, entry)) | |
409 | continue; | |
410 | ret = dma_ops_unity_map(iommu->default_dom, entry); | |
411 | if (ret) | |
412 | return ret; | |
413 | } | |
414 | ||
415 | return 0; | |
416 | } | |
417 | ||
431b2a20 JR |
418 | /* |
419 | * This function actually applies the mapping to the page table of the | |
420 | * dma_ops domain. | |
421 | */ | |
bd0e5211 JR |
422 | static int dma_ops_unity_map(struct dma_ops_domain *dma_dom, |
423 | struct unity_map_entry *e) | |
424 | { | |
425 | u64 addr; | |
426 | int ret; | |
427 | ||
428 | for (addr = e->address_start; addr < e->address_end; | |
429 | addr += PAGE_SIZE) { | |
430 | ret = iommu_map(&dma_dom->domain, addr, addr, e->prot); | |
431 | if (ret) | |
432 | return ret; | |
433 | /* | |
434 | * if unity mapping is in aperture range mark the page | |
435 | * as allocated in the aperture | |
436 | */ | |
437 | if (addr < dma_dom->aperture_size) | |
438 | __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap); | |
439 | } | |
440 | ||
441 | return 0; | |
442 | } | |
443 | ||
431b2a20 JR |
444 | /* |
445 | * Inits the unity mappings required for a specific device | |
446 | */ | |
bd0e5211 JR |
447 | static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom, |
448 | u16 devid) | |
449 | { | |
450 | struct unity_map_entry *e; | |
451 | int ret; | |
452 | ||
453 | list_for_each_entry(e, &amd_iommu_unity_map, list) { | |
454 | if (!(devid >= e->devid_start && devid <= e->devid_end)) | |
455 | continue; | |
456 | ret = dma_ops_unity_map(dma_dom, e); | |
457 | if (ret) | |
458 | return ret; | |
459 | } | |
460 | ||
461 | return 0; | |
462 | } | |
463 | ||
431b2a20 JR |
464 | /**************************************************************************** |
465 | * | |
466 | * The next functions belong to the address allocator for the dma_ops | |
467 | * interface functions. They work like the allocators in the other IOMMU | |
468 | * drivers. Its basically a bitmap which marks the allocated pages in | |
469 | * the aperture. Maybe it could be enhanced in the future to a more | |
470 | * efficient allocator. | |
471 | * | |
472 | ****************************************************************************/ | |
d3086444 | 473 | |
431b2a20 JR |
474 | /* |
475 | * The address allocator core function. | |
476 | * | |
477 | * called with domain->lock held | |
478 | */ | |
d3086444 JR |
479 | static unsigned long dma_ops_alloc_addresses(struct device *dev, |
480 | struct dma_ops_domain *dom, | |
6d4f343f | 481 | unsigned int pages, |
832a90c3 JR |
482 | unsigned long align_mask, |
483 | u64 dma_mask) | |
d3086444 | 484 | { |
40becd8d | 485 | unsigned long limit; |
d3086444 | 486 | unsigned long address; |
d3086444 JR |
487 | unsigned long boundary_size; |
488 | ||
489 | boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1, | |
490 | PAGE_SIZE) >> PAGE_SHIFT; | |
40becd8d FT |
491 | limit = iommu_device_max_index(dom->aperture_size >> PAGE_SHIFT, 0, |
492 | dma_mask >> PAGE_SHIFT); | |
d3086444 | 493 | |
1c655773 | 494 | if (dom->next_bit >= limit) { |
d3086444 | 495 | dom->next_bit = 0; |
1c655773 JR |
496 | dom->need_flush = true; |
497 | } | |
d3086444 JR |
498 | |
499 | address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages, | |
6d4f343f | 500 | 0 , boundary_size, align_mask); |
1c655773 | 501 | if (address == -1) { |
d3086444 | 502 | address = iommu_area_alloc(dom->bitmap, limit, 0, pages, |
6d4f343f | 503 | 0, boundary_size, align_mask); |
1c655773 JR |
504 | dom->need_flush = true; |
505 | } | |
d3086444 JR |
506 | |
507 | if (likely(address != -1)) { | |
d3086444 JR |
508 | dom->next_bit = address + pages; |
509 | address <<= PAGE_SHIFT; | |
510 | } else | |
511 | address = bad_dma_address; | |
512 | ||
513 | WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size); | |
514 | ||
515 | return address; | |
516 | } | |
517 | ||
431b2a20 JR |
518 | /* |
519 | * The address free function. | |
520 | * | |
521 | * called with domain->lock held | |
522 | */ | |
d3086444 JR |
523 | static void dma_ops_free_addresses(struct dma_ops_domain *dom, |
524 | unsigned long address, | |
525 | unsigned int pages) | |
526 | { | |
527 | address >>= PAGE_SHIFT; | |
528 | iommu_area_free(dom->bitmap, address, pages); | |
529 | } | |
530 | ||
431b2a20 JR |
531 | /**************************************************************************** |
532 | * | |
533 | * The next functions belong to the domain allocation. A domain is | |
534 | * allocated for every IOMMU as the default domain. If device isolation | |
535 | * is enabled, every device get its own domain. The most important thing | |
536 | * about domains is the page table mapping the DMA address space they | |
537 | * contain. | |
538 | * | |
539 | ****************************************************************************/ | |
540 | ||
ec487d1a JR |
541 | static u16 domain_id_alloc(void) |
542 | { | |
543 | unsigned long flags; | |
544 | int id; | |
545 | ||
546 | write_lock_irqsave(&amd_iommu_devtable_lock, flags); | |
547 | id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID); | |
548 | BUG_ON(id == 0); | |
549 | if (id > 0 && id < MAX_DOMAIN_ID) | |
550 | __set_bit(id, amd_iommu_pd_alloc_bitmap); | |
551 | else | |
552 | id = 0; | |
553 | write_unlock_irqrestore(&amd_iommu_devtable_lock, flags); | |
554 | ||
555 | return id; | |
556 | } | |
557 | ||
431b2a20 JR |
558 | /* |
559 | * Used to reserve address ranges in the aperture (e.g. for exclusion | |
560 | * ranges. | |
561 | */ | |
ec487d1a JR |
562 | static void dma_ops_reserve_addresses(struct dma_ops_domain *dom, |
563 | unsigned long start_page, | |
564 | unsigned int pages) | |
565 | { | |
566 | unsigned int last_page = dom->aperture_size >> PAGE_SHIFT; | |
567 | ||
568 | if (start_page + pages > last_page) | |
569 | pages = last_page - start_page; | |
570 | ||
d26dbc5c | 571 | iommu_area_reserve(dom->bitmap, start_page, pages); |
ec487d1a JR |
572 | } |
573 | ||
574 | static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom) | |
575 | { | |
576 | int i, j; | |
577 | u64 *p1, *p2, *p3; | |
578 | ||
579 | p1 = dma_dom->domain.pt_root; | |
580 | ||
581 | if (!p1) | |
582 | return; | |
583 | ||
584 | for (i = 0; i < 512; ++i) { | |
585 | if (!IOMMU_PTE_PRESENT(p1[i])) | |
586 | continue; | |
587 | ||
588 | p2 = IOMMU_PTE_PAGE(p1[i]); | |
589 | for (j = 0; j < 512; ++i) { | |
590 | if (!IOMMU_PTE_PRESENT(p2[j])) | |
591 | continue; | |
592 | p3 = IOMMU_PTE_PAGE(p2[j]); | |
593 | free_page((unsigned long)p3); | |
594 | } | |
595 | ||
596 | free_page((unsigned long)p2); | |
597 | } | |
598 | ||
599 | free_page((unsigned long)p1); | |
600 | } | |
601 | ||
431b2a20 JR |
602 | /* |
603 | * Free a domain, only used if something went wrong in the | |
604 | * allocation path and we need to free an already allocated page table | |
605 | */ | |
ec487d1a JR |
606 | static void dma_ops_domain_free(struct dma_ops_domain *dom) |
607 | { | |
608 | if (!dom) | |
609 | return; | |
610 | ||
611 | dma_ops_free_pagetable(dom); | |
612 | ||
613 | kfree(dom->pte_pages); | |
614 | ||
615 | kfree(dom->bitmap); | |
616 | ||
617 | kfree(dom); | |
618 | } | |
619 | ||
431b2a20 JR |
620 | /* |
621 | * Allocates a new protection domain usable for the dma_ops functions. | |
622 | * It also intializes the page table and the address allocator data | |
623 | * structures required for the dma_ops interface | |
624 | */ | |
ec487d1a JR |
625 | static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu, |
626 | unsigned order) | |
627 | { | |
628 | struct dma_ops_domain *dma_dom; | |
629 | unsigned i, num_pte_pages; | |
630 | u64 *l2_pde; | |
631 | u64 address; | |
632 | ||
633 | /* | |
634 | * Currently the DMA aperture must be between 32 MB and 1GB in size | |
635 | */ | |
636 | if ((order < 25) || (order > 30)) | |
637 | return NULL; | |
638 | ||
639 | dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL); | |
640 | if (!dma_dom) | |
641 | return NULL; | |
642 | ||
643 | spin_lock_init(&dma_dom->domain.lock); | |
644 | ||
645 | dma_dom->domain.id = domain_id_alloc(); | |
646 | if (dma_dom->domain.id == 0) | |
647 | goto free_dma_dom; | |
648 | dma_dom->domain.mode = PAGE_MODE_3_LEVEL; | |
649 | dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL); | |
650 | dma_dom->domain.priv = dma_dom; | |
651 | if (!dma_dom->domain.pt_root) | |
652 | goto free_dma_dom; | |
653 | dma_dom->aperture_size = (1ULL << order); | |
654 | dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8), | |
655 | GFP_KERNEL); | |
656 | if (!dma_dom->bitmap) | |
657 | goto free_dma_dom; | |
658 | /* | |
659 | * mark the first page as allocated so we never return 0 as | |
660 | * a valid dma-address. So we can use 0 as error value | |
661 | */ | |
662 | dma_dom->bitmap[0] = 1; | |
663 | dma_dom->next_bit = 0; | |
664 | ||
1c655773 | 665 | dma_dom->need_flush = false; |
bd60b735 | 666 | dma_dom->target_dev = 0xffff; |
1c655773 | 667 | |
431b2a20 | 668 | /* Intialize the exclusion range if necessary */ |
ec487d1a JR |
669 | if (iommu->exclusion_start && |
670 | iommu->exclusion_start < dma_dom->aperture_size) { | |
671 | unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT; | |
a8132e5f JR |
672 | int pages = iommu_num_pages(iommu->exclusion_start, |
673 | iommu->exclusion_length); | |
ec487d1a JR |
674 | dma_ops_reserve_addresses(dma_dom, startpage, pages); |
675 | } | |
676 | ||
431b2a20 JR |
677 | /* |
678 | * At the last step, build the page tables so we don't need to | |
679 | * allocate page table pages in the dma_ops mapping/unmapping | |
680 | * path. | |
681 | */ | |
ec487d1a JR |
682 | num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512); |
683 | dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *), | |
684 | GFP_KERNEL); | |
685 | if (!dma_dom->pte_pages) | |
686 | goto free_dma_dom; | |
687 | ||
688 | l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL); | |
689 | if (l2_pde == NULL) | |
690 | goto free_dma_dom; | |
691 | ||
692 | dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde)); | |
693 | ||
694 | for (i = 0; i < num_pte_pages; ++i) { | |
695 | dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL); | |
696 | if (!dma_dom->pte_pages[i]) | |
697 | goto free_dma_dom; | |
698 | address = virt_to_phys(dma_dom->pte_pages[i]); | |
699 | l2_pde[i] = IOMMU_L1_PDE(address); | |
700 | } | |
701 | ||
702 | return dma_dom; | |
703 | ||
704 | free_dma_dom: | |
705 | dma_ops_domain_free(dma_dom); | |
706 | ||
707 | return NULL; | |
708 | } | |
709 | ||
431b2a20 JR |
710 | /* |
711 | * Find out the protection domain structure for a given PCI device. This | |
712 | * will give us the pointer to the page table root for example. | |
713 | */ | |
b20ac0d4 JR |
714 | static struct protection_domain *domain_for_device(u16 devid) |
715 | { | |
716 | struct protection_domain *dom; | |
717 | unsigned long flags; | |
718 | ||
719 | read_lock_irqsave(&amd_iommu_devtable_lock, flags); | |
720 | dom = amd_iommu_pd_table[devid]; | |
721 | read_unlock_irqrestore(&amd_iommu_devtable_lock, flags); | |
722 | ||
723 | return dom; | |
724 | } | |
725 | ||
431b2a20 JR |
726 | /* |
727 | * If a device is not yet associated with a domain, this function does | |
728 | * assigns it visible for the hardware | |
729 | */ | |
b20ac0d4 JR |
730 | static void set_device_domain(struct amd_iommu *iommu, |
731 | struct protection_domain *domain, | |
732 | u16 devid) | |
733 | { | |
734 | unsigned long flags; | |
735 | ||
736 | u64 pte_root = virt_to_phys(domain->pt_root); | |
737 | ||
38ddf41b JR |
738 | pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK) |
739 | << DEV_ENTRY_MODE_SHIFT; | |
740 | pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV; | |
b20ac0d4 JR |
741 | |
742 | write_lock_irqsave(&amd_iommu_devtable_lock, flags); | |
38ddf41b JR |
743 | amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root); |
744 | amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root); | |
b20ac0d4 JR |
745 | amd_iommu_dev_table[devid].data[2] = domain->id; |
746 | ||
747 | amd_iommu_pd_table[devid] = domain; | |
748 | write_unlock_irqrestore(&amd_iommu_devtable_lock, flags); | |
749 | ||
750 | iommu_queue_inv_dev_entry(iommu, devid); | |
751 | ||
752 | iommu->need_sync = 1; | |
753 | } | |
754 | ||
431b2a20 JR |
755 | /***************************************************************************** |
756 | * | |
757 | * The next functions belong to the dma_ops mapping/unmapping code. | |
758 | * | |
759 | *****************************************************************************/ | |
760 | ||
dbcc112e JR |
761 | /* |
762 | * This function checks if the driver got a valid device from the caller to | |
763 | * avoid dereferencing invalid pointers. | |
764 | */ | |
765 | static bool check_device(struct device *dev) | |
766 | { | |
767 | if (!dev || !dev->dma_mask) | |
768 | return false; | |
769 | ||
770 | return true; | |
771 | } | |
772 | ||
bd60b735 JR |
773 | /* |
774 | * In this function the list of preallocated protection domains is traversed to | |
775 | * find the domain for a specific device | |
776 | */ | |
777 | static struct dma_ops_domain *find_protection_domain(u16 devid) | |
778 | { | |
779 | struct dma_ops_domain *entry, *ret = NULL; | |
780 | unsigned long flags; | |
781 | ||
782 | if (list_empty(&iommu_pd_list)) | |
783 | return NULL; | |
784 | ||
785 | spin_lock_irqsave(&iommu_pd_list_lock, flags); | |
786 | ||
787 | list_for_each_entry(entry, &iommu_pd_list, list) { | |
788 | if (entry->target_dev == devid) { | |
789 | ret = entry; | |
790 | list_del(&ret->list); | |
791 | break; | |
792 | } | |
793 | } | |
794 | ||
795 | spin_unlock_irqrestore(&iommu_pd_list_lock, flags); | |
796 | ||
797 | return ret; | |
798 | } | |
799 | ||
431b2a20 JR |
800 | /* |
801 | * In the dma_ops path we only have the struct device. This function | |
802 | * finds the corresponding IOMMU, the protection domain and the | |
803 | * requestor id for a given device. | |
804 | * If the device is not yet associated with a domain this is also done | |
805 | * in this function. | |
806 | */ | |
b20ac0d4 JR |
807 | static int get_device_resources(struct device *dev, |
808 | struct amd_iommu **iommu, | |
809 | struct protection_domain **domain, | |
810 | u16 *bdf) | |
811 | { | |
812 | struct dma_ops_domain *dma_dom; | |
813 | struct pci_dev *pcidev; | |
814 | u16 _bdf; | |
815 | ||
dbcc112e JR |
816 | *iommu = NULL; |
817 | *domain = NULL; | |
818 | *bdf = 0xffff; | |
819 | ||
820 | if (dev->bus != &pci_bus_type) | |
821 | return 0; | |
b20ac0d4 JR |
822 | |
823 | pcidev = to_pci_dev(dev); | |
d591b0a3 | 824 | _bdf = calc_devid(pcidev->bus->number, pcidev->devfn); |
b20ac0d4 | 825 | |
431b2a20 | 826 | /* device not translated by any IOMMU in the system? */ |
dbcc112e | 827 | if (_bdf > amd_iommu_last_bdf) |
b20ac0d4 | 828 | return 0; |
b20ac0d4 JR |
829 | |
830 | *bdf = amd_iommu_alias_table[_bdf]; | |
831 | ||
832 | *iommu = amd_iommu_rlookup_table[*bdf]; | |
833 | if (*iommu == NULL) | |
834 | return 0; | |
b20ac0d4 JR |
835 | *domain = domain_for_device(*bdf); |
836 | if (*domain == NULL) { | |
bd60b735 JR |
837 | dma_dom = find_protection_domain(*bdf); |
838 | if (!dma_dom) | |
839 | dma_dom = (*iommu)->default_dom; | |
b20ac0d4 JR |
840 | *domain = &dma_dom->domain; |
841 | set_device_domain(*iommu, *domain, *bdf); | |
842 | printk(KERN_INFO "AMD IOMMU: Using protection domain %d for " | |
843 | "device ", (*domain)->id); | |
844 | print_devid(_bdf, 1); | |
845 | } | |
846 | ||
847 | return 1; | |
848 | } | |
849 | ||
431b2a20 JR |
850 | /* |
851 | * This is the generic map function. It maps one 4kb page at paddr to | |
852 | * the given address in the DMA address space for the domain. | |
853 | */ | |
cb76c322 JR |
854 | static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu, |
855 | struct dma_ops_domain *dom, | |
856 | unsigned long address, | |
857 | phys_addr_t paddr, | |
858 | int direction) | |
859 | { | |
860 | u64 *pte, __pte; | |
861 | ||
862 | WARN_ON(address > dom->aperture_size); | |
863 | ||
864 | paddr &= PAGE_MASK; | |
865 | ||
866 | pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)]; | |
867 | pte += IOMMU_PTE_L0_INDEX(address); | |
868 | ||
869 | __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC; | |
870 | ||
871 | if (direction == DMA_TO_DEVICE) | |
872 | __pte |= IOMMU_PTE_IR; | |
873 | else if (direction == DMA_FROM_DEVICE) | |
874 | __pte |= IOMMU_PTE_IW; | |
875 | else if (direction == DMA_BIDIRECTIONAL) | |
876 | __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW; | |
877 | ||
878 | WARN_ON(*pte); | |
879 | ||
880 | *pte = __pte; | |
881 | ||
882 | return (dma_addr_t)address; | |
883 | } | |
884 | ||
431b2a20 JR |
885 | /* |
886 | * The generic unmapping function for on page in the DMA address space. | |
887 | */ | |
cb76c322 JR |
888 | static void dma_ops_domain_unmap(struct amd_iommu *iommu, |
889 | struct dma_ops_domain *dom, | |
890 | unsigned long address) | |
891 | { | |
892 | u64 *pte; | |
893 | ||
894 | if (address >= dom->aperture_size) | |
895 | return; | |
896 | ||
897 | WARN_ON(address & 0xfffULL || address > dom->aperture_size); | |
898 | ||
899 | pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)]; | |
900 | pte += IOMMU_PTE_L0_INDEX(address); | |
901 | ||
902 | WARN_ON(!*pte); | |
903 | ||
904 | *pte = 0ULL; | |
905 | } | |
906 | ||
431b2a20 JR |
907 | /* |
908 | * This function contains common code for mapping of a physically | |
909 | * contiguous memory region into DMA address space. It is uses by all | |
910 | * mapping functions provided by this IOMMU driver. | |
911 | * Must be called with the domain lock held. | |
912 | */ | |
cb76c322 JR |
913 | static dma_addr_t __map_single(struct device *dev, |
914 | struct amd_iommu *iommu, | |
915 | struct dma_ops_domain *dma_dom, | |
916 | phys_addr_t paddr, | |
917 | size_t size, | |
6d4f343f | 918 | int dir, |
832a90c3 JR |
919 | bool align, |
920 | u64 dma_mask) | |
cb76c322 JR |
921 | { |
922 | dma_addr_t offset = paddr & ~PAGE_MASK; | |
923 | dma_addr_t address, start; | |
924 | unsigned int pages; | |
6d4f343f | 925 | unsigned long align_mask = 0; |
cb76c322 JR |
926 | int i; |
927 | ||
a8132e5f | 928 | pages = iommu_num_pages(paddr, size); |
cb76c322 JR |
929 | paddr &= PAGE_MASK; |
930 | ||
6d4f343f JR |
931 | if (align) |
932 | align_mask = (1UL << get_order(size)) - 1; | |
933 | ||
832a90c3 JR |
934 | address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask, |
935 | dma_mask); | |
cb76c322 JR |
936 | if (unlikely(address == bad_dma_address)) |
937 | goto out; | |
938 | ||
939 | start = address; | |
940 | for (i = 0; i < pages; ++i) { | |
941 | dma_ops_domain_map(iommu, dma_dom, start, paddr, dir); | |
942 | paddr += PAGE_SIZE; | |
943 | start += PAGE_SIZE; | |
944 | } | |
945 | address += offset; | |
946 | ||
afa9fdc2 | 947 | if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) { |
1c655773 JR |
948 | iommu_flush_tlb(iommu, dma_dom->domain.id); |
949 | dma_dom->need_flush = false; | |
950 | } else if (unlikely(iommu_has_npcache(iommu))) | |
270cab24 JR |
951 | iommu_flush_pages(iommu, dma_dom->domain.id, address, size); |
952 | ||
cb76c322 JR |
953 | out: |
954 | return address; | |
955 | } | |
956 | ||
431b2a20 JR |
957 | /* |
958 | * Does the reverse of the __map_single function. Must be called with | |
959 | * the domain lock held too | |
960 | */ | |
cb76c322 JR |
961 | static void __unmap_single(struct amd_iommu *iommu, |
962 | struct dma_ops_domain *dma_dom, | |
963 | dma_addr_t dma_addr, | |
964 | size_t size, | |
965 | int dir) | |
966 | { | |
967 | dma_addr_t i, start; | |
968 | unsigned int pages; | |
969 | ||
970 | if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size)) | |
971 | return; | |
972 | ||
a8132e5f | 973 | pages = iommu_num_pages(dma_addr, size); |
cb76c322 JR |
974 | dma_addr &= PAGE_MASK; |
975 | start = dma_addr; | |
976 | ||
977 | for (i = 0; i < pages; ++i) { | |
978 | dma_ops_domain_unmap(iommu, dma_dom, start); | |
979 | start += PAGE_SIZE; | |
980 | } | |
981 | ||
982 | dma_ops_free_addresses(dma_dom, dma_addr, pages); | |
270cab24 | 983 | |
afa9fdc2 | 984 | if (amd_iommu_unmap_flush) |
1c655773 | 985 | iommu_flush_pages(iommu, dma_dom->domain.id, dma_addr, size); |
cb76c322 JR |
986 | } |
987 | ||
431b2a20 JR |
988 | /* |
989 | * The exported map_single function for dma_ops. | |
990 | */ | |
4da70b9e JR |
991 | static dma_addr_t map_single(struct device *dev, phys_addr_t paddr, |
992 | size_t size, int dir) | |
993 | { | |
994 | unsigned long flags; | |
995 | struct amd_iommu *iommu; | |
996 | struct protection_domain *domain; | |
997 | u16 devid; | |
998 | dma_addr_t addr; | |
832a90c3 | 999 | u64 dma_mask; |
4da70b9e | 1000 | |
dbcc112e JR |
1001 | if (!check_device(dev)) |
1002 | return bad_dma_address; | |
1003 | ||
832a90c3 JR |
1004 | dma_mask = *dev->dma_mask; |
1005 | ||
4da70b9e JR |
1006 | get_device_resources(dev, &iommu, &domain, &devid); |
1007 | ||
1008 | if (iommu == NULL || domain == NULL) | |
431b2a20 | 1009 | /* device not handled by any AMD IOMMU */ |
4da70b9e JR |
1010 | return (dma_addr_t)paddr; |
1011 | ||
1012 | spin_lock_irqsave(&domain->lock, flags); | |
832a90c3 JR |
1013 | addr = __map_single(dev, iommu, domain->priv, paddr, size, dir, false, |
1014 | dma_mask); | |
4da70b9e JR |
1015 | if (addr == bad_dma_address) |
1016 | goto out; | |
1017 | ||
5507eef8 | 1018 | if (unlikely(iommu->need_sync)) |
4da70b9e JR |
1019 | iommu_completion_wait(iommu); |
1020 | ||
1021 | out: | |
1022 | spin_unlock_irqrestore(&domain->lock, flags); | |
1023 | ||
1024 | return addr; | |
1025 | } | |
1026 | ||
431b2a20 JR |
1027 | /* |
1028 | * The exported unmap_single function for dma_ops. | |
1029 | */ | |
4da70b9e JR |
1030 | static void unmap_single(struct device *dev, dma_addr_t dma_addr, |
1031 | size_t size, int dir) | |
1032 | { | |
1033 | unsigned long flags; | |
1034 | struct amd_iommu *iommu; | |
1035 | struct protection_domain *domain; | |
1036 | u16 devid; | |
1037 | ||
dbcc112e JR |
1038 | if (!check_device(dev) || |
1039 | !get_device_resources(dev, &iommu, &domain, &devid)) | |
431b2a20 | 1040 | /* device not handled by any AMD IOMMU */ |
4da70b9e JR |
1041 | return; |
1042 | ||
1043 | spin_lock_irqsave(&domain->lock, flags); | |
1044 | ||
1045 | __unmap_single(iommu, domain->priv, dma_addr, size, dir); | |
1046 | ||
5507eef8 | 1047 | if (unlikely(iommu->need_sync)) |
4da70b9e JR |
1048 | iommu_completion_wait(iommu); |
1049 | ||
1050 | spin_unlock_irqrestore(&domain->lock, flags); | |
1051 | } | |
1052 | ||
431b2a20 JR |
1053 | /* |
1054 | * This is a special map_sg function which is used if we should map a | |
1055 | * device which is not handled by an AMD IOMMU in the system. | |
1056 | */ | |
65b050ad JR |
1057 | static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist, |
1058 | int nelems, int dir) | |
1059 | { | |
1060 | struct scatterlist *s; | |
1061 | int i; | |
1062 | ||
1063 | for_each_sg(sglist, s, nelems, i) { | |
1064 | s->dma_address = (dma_addr_t)sg_phys(s); | |
1065 | s->dma_length = s->length; | |
1066 | } | |
1067 | ||
1068 | return nelems; | |
1069 | } | |
1070 | ||
431b2a20 JR |
1071 | /* |
1072 | * The exported map_sg function for dma_ops (handles scatter-gather | |
1073 | * lists). | |
1074 | */ | |
65b050ad JR |
1075 | static int map_sg(struct device *dev, struct scatterlist *sglist, |
1076 | int nelems, int dir) | |
1077 | { | |
1078 | unsigned long flags; | |
1079 | struct amd_iommu *iommu; | |
1080 | struct protection_domain *domain; | |
1081 | u16 devid; | |
1082 | int i; | |
1083 | struct scatterlist *s; | |
1084 | phys_addr_t paddr; | |
1085 | int mapped_elems = 0; | |
832a90c3 | 1086 | u64 dma_mask; |
65b050ad | 1087 | |
dbcc112e JR |
1088 | if (!check_device(dev)) |
1089 | return 0; | |
1090 | ||
832a90c3 JR |
1091 | dma_mask = *dev->dma_mask; |
1092 | ||
65b050ad JR |
1093 | get_device_resources(dev, &iommu, &domain, &devid); |
1094 | ||
1095 | if (!iommu || !domain) | |
1096 | return map_sg_no_iommu(dev, sglist, nelems, dir); | |
1097 | ||
1098 | spin_lock_irqsave(&domain->lock, flags); | |
1099 | ||
1100 | for_each_sg(sglist, s, nelems, i) { | |
1101 | paddr = sg_phys(s); | |
1102 | ||
1103 | s->dma_address = __map_single(dev, iommu, domain->priv, | |
832a90c3 JR |
1104 | paddr, s->length, dir, false, |
1105 | dma_mask); | |
65b050ad JR |
1106 | |
1107 | if (s->dma_address) { | |
1108 | s->dma_length = s->length; | |
1109 | mapped_elems++; | |
1110 | } else | |
1111 | goto unmap; | |
65b050ad JR |
1112 | } |
1113 | ||
5507eef8 | 1114 | if (unlikely(iommu->need_sync)) |
65b050ad JR |
1115 | iommu_completion_wait(iommu); |
1116 | ||
1117 | out: | |
1118 | spin_unlock_irqrestore(&domain->lock, flags); | |
1119 | ||
1120 | return mapped_elems; | |
1121 | unmap: | |
1122 | for_each_sg(sglist, s, mapped_elems, i) { | |
1123 | if (s->dma_address) | |
1124 | __unmap_single(iommu, domain->priv, s->dma_address, | |
1125 | s->dma_length, dir); | |
1126 | s->dma_address = s->dma_length = 0; | |
1127 | } | |
1128 | ||
1129 | mapped_elems = 0; | |
1130 | ||
1131 | goto out; | |
1132 | } | |
1133 | ||
431b2a20 JR |
1134 | /* |
1135 | * The exported map_sg function for dma_ops (handles scatter-gather | |
1136 | * lists). | |
1137 | */ | |
65b050ad JR |
1138 | static void unmap_sg(struct device *dev, struct scatterlist *sglist, |
1139 | int nelems, int dir) | |
1140 | { | |
1141 | unsigned long flags; | |
1142 | struct amd_iommu *iommu; | |
1143 | struct protection_domain *domain; | |
1144 | struct scatterlist *s; | |
1145 | u16 devid; | |
1146 | int i; | |
1147 | ||
dbcc112e JR |
1148 | if (!check_device(dev) || |
1149 | !get_device_resources(dev, &iommu, &domain, &devid)) | |
65b050ad JR |
1150 | return; |
1151 | ||
1152 | spin_lock_irqsave(&domain->lock, flags); | |
1153 | ||
1154 | for_each_sg(sglist, s, nelems, i) { | |
1155 | __unmap_single(iommu, domain->priv, s->dma_address, | |
1156 | s->dma_length, dir); | |
65b050ad JR |
1157 | s->dma_address = s->dma_length = 0; |
1158 | } | |
1159 | ||
5507eef8 | 1160 | if (unlikely(iommu->need_sync)) |
65b050ad JR |
1161 | iommu_completion_wait(iommu); |
1162 | ||
1163 | spin_unlock_irqrestore(&domain->lock, flags); | |
1164 | } | |
1165 | ||
431b2a20 JR |
1166 | /* |
1167 | * The exported alloc_coherent function for dma_ops. | |
1168 | */ | |
5d8b53cf JR |
1169 | static void *alloc_coherent(struct device *dev, size_t size, |
1170 | dma_addr_t *dma_addr, gfp_t flag) | |
1171 | { | |
1172 | unsigned long flags; | |
1173 | void *virt_addr; | |
1174 | struct amd_iommu *iommu; | |
1175 | struct protection_domain *domain; | |
1176 | u16 devid; | |
1177 | phys_addr_t paddr; | |
832a90c3 | 1178 | u64 dma_mask = dev->coherent_dma_mask; |
5d8b53cf | 1179 | |
dbcc112e JR |
1180 | if (!check_device(dev)) |
1181 | return NULL; | |
1182 | ||
13d9fead FT |
1183 | if (!get_device_resources(dev, &iommu, &domain, &devid)) |
1184 | flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32); | |
1185 | ||
c97ac535 | 1186 | flag |= __GFP_ZERO; |
5d8b53cf JR |
1187 | virt_addr = (void *)__get_free_pages(flag, get_order(size)); |
1188 | if (!virt_addr) | |
1189 | return 0; | |
1190 | ||
5d8b53cf JR |
1191 | paddr = virt_to_phys(virt_addr); |
1192 | ||
5d8b53cf JR |
1193 | if (!iommu || !domain) { |
1194 | *dma_addr = (dma_addr_t)paddr; | |
1195 | return virt_addr; | |
1196 | } | |
1197 | ||
832a90c3 JR |
1198 | if (!dma_mask) |
1199 | dma_mask = *dev->dma_mask; | |
1200 | ||
5d8b53cf JR |
1201 | spin_lock_irqsave(&domain->lock, flags); |
1202 | ||
1203 | *dma_addr = __map_single(dev, iommu, domain->priv, paddr, | |
832a90c3 | 1204 | size, DMA_BIDIRECTIONAL, true, dma_mask); |
5d8b53cf JR |
1205 | |
1206 | if (*dma_addr == bad_dma_address) { | |
1207 | free_pages((unsigned long)virt_addr, get_order(size)); | |
1208 | virt_addr = NULL; | |
1209 | goto out; | |
1210 | } | |
1211 | ||
5507eef8 | 1212 | if (unlikely(iommu->need_sync)) |
5d8b53cf JR |
1213 | iommu_completion_wait(iommu); |
1214 | ||
1215 | out: | |
1216 | spin_unlock_irqrestore(&domain->lock, flags); | |
1217 | ||
1218 | return virt_addr; | |
1219 | } | |
1220 | ||
431b2a20 JR |
1221 | /* |
1222 | * The exported free_coherent function for dma_ops. | |
431b2a20 | 1223 | */ |
5d8b53cf JR |
1224 | static void free_coherent(struct device *dev, size_t size, |
1225 | void *virt_addr, dma_addr_t dma_addr) | |
1226 | { | |
1227 | unsigned long flags; | |
1228 | struct amd_iommu *iommu; | |
1229 | struct protection_domain *domain; | |
1230 | u16 devid; | |
1231 | ||
dbcc112e JR |
1232 | if (!check_device(dev)) |
1233 | return; | |
1234 | ||
5d8b53cf JR |
1235 | get_device_resources(dev, &iommu, &domain, &devid); |
1236 | ||
1237 | if (!iommu || !domain) | |
1238 | goto free_mem; | |
1239 | ||
1240 | spin_lock_irqsave(&domain->lock, flags); | |
1241 | ||
1242 | __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL); | |
5d8b53cf | 1243 | |
5507eef8 | 1244 | if (unlikely(iommu->need_sync)) |
5d8b53cf JR |
1245 | iommu_completion_wait(iommu); |
1246 | ||
1247 | spin_unlock_irqrestore(&domain->lock, flags); | |
1248 | ||
1249 | free_mem: | |
1250 | free_pages((unsigned long)virt_addr, get_order(size)); | |
1251 | } | |
1252 | ||
b39ba6ad JR |
1253 | /* |
1254 | * This function is called by the DMA layer to find out if we can handle a | |
1255 | * particular device. It is part of the dma_ops. | |
1256 | */ | |
1257 | static int amd_iommu_dma_supported(struct device *dev, u64 mask) | |
1258 | { | |
1259 | u16 bdf; | |
1260 | struct pci_dev *pcidev; | |
1261 | ||
1262 | /* No device or no PCI device */ | |
1263 | if (!dev || dev->bus != &pci_bus_type) | |
1264 | return 0; | |
1265 | ||
1266 | pcidev = to_pci_dev(dev); | |
1267 | ||
1268 | bdf = calc_devid(pcidev->bus->number, pcidev->devfn); | |
1269 | ||
1270 | /* Out of our scope? */ | |
1271 | if (bdf > amd_iommu_last_bdf) | |
1272 | return 0; | |
1273 | ||
1274 | return 1; | |
1275 | } | |
1276 | ||
c432f3df | 1277 | /* |
431b2a20 JR |
1278 | * The function for pre-allocating protection domains. |
1279 | * | |
c432f3df JR |
1280 | * If the driver core informs the DMA layer if a driver grabs a device |
1281 | * we don't need to preallocate the protection domains anymore. | |
1282 | * For now we have to. | |
1283 | */ | |
1284 | void prealloc_protection_domains(void) | |
1285 | { | |
1286 | struct pci_dev *dev = NULL; | |
1287 | struct dma_ops_domain *dma_dom; | |
1288 | struct amd_iommu *iommu; | |
1289 | int order = amd_iommu_aperture_order; | |
1290 | u16 devid; | |
1291 | ||
1292 | while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) { | |
1293 | devid = (dev->bus->number << 8) | dev->devfn; | |
3a61ec38 | 1294 | if (devid > amd_iommu_last_bdf) |
c432f3df JR |
1295 | continue; |
1296 | devid = amd_iommu_alias_table[devid]; | |
1297 | if (domain_for_device(devid)) | |
1298 | continue; | |
1299 | iommu = amd_iommu_rlookup_table[devid]; | |
1300 | if (!iommu) | |
1301 | continue; | |
1302 | dma_dom = dma_ops_domain_alloc(iommu, order); | |
1303 | if (!dma_dom) | |
1304 | continue; | |
1305 | init_unity_mappings_for_device(dma_dom, devid); | |
bd60b735 JR |
1306 | dma_dom->target_dev = devid; |
1307 | ||
1308 | list_add_tail(&dma_dom->list, &iommu_pd_list); | |
c432f3df JR |
1309 | } |
1310 | } | |
1311 | ||
6631ee9d JR |
1312 | static struct dma_mapping_ops amd_iommu_dma_ops = { |
1313 | .alloc_coherent = alloc_coherent, | |
1314 | .free_coherent = free_coherent, | |
1315 | .map_single = map_single, | |
1316 | .unmap_single = unmap_single, | |
1317 | .map_sg = map_sg, | |
1318 | .unmap_sg = unmap_sg, | |
b39ba6ad | 1319 | .dma_supported = amd_iommu_dma_supported, |
6631ee9d JR |
1320 | }; |
1321 | ||
431b2a20 JR |
1322 | /* |
1323 | * The function which clues the AMD IOMMU driver into dma_ops. | |
1324 | */ | |
6631ee9d JR |
1325 | int __init amd_iommu_init_dma_ops(void) |
1326 | { | |
1327 | struct amd_iommu *iommu; | |
1328 | int order = amd_iommu_aperture_order; | |
1329 | int ret; | |
1330 | ||
431b2a20 JR |
1331 | /* |
1332 | * first allocate a default protection domain for every IOMMU we | |
1333 | * found in the system. Devices not assigned to any other | |
1334 | * protection domain will be assigned to the default one. | |
1335 | */ | |
6631ee9d JR |
1336 | list_for_each_entry(iommu, &amd_iommu_list, list) { |
1337 | iommu->default_dom = dma_ops_domain_alloc(iommu, order); | |
1338 | if (iommu->default_dom == NULL) | |
1339 | return -ENOMEM; | |
1340 | ret = iommu_init_unity_mappings(iommu); | |
1341 | if (ret) | |
1342 | goto free_domains; | |
1343 | } | |
1344 | ||
431b2a20 JR |
1345 | /* |
1346 | * If device isolation is enabled, pre-allocate the protection | |
1347 | * domains for each device. | |
1348 | */ | |
6631ee9d JR |
1349 | if (amd_iommu_isolate) |
1350 | prealloc_protection_domains(); | |
1351 | ||
1352 | iommu_detected = 1; | |
1353 | force_iommu = 1; | |
1354 | bad_dma_address = 0; | |
92af4e29 | 1355 | #ifdef CONFIG_GART_IOMMU |
6631ee9d JR |
1356 | gart_iommu_aperture_disabled = 1; |
1357 | gart_iommu_aperture = 0; | |
92af4e29 | 1358 | #endif |
6631ee9d | 1359 | |
431b2a20 | 1360 | /* Make the driver finally visible to the drivers */ |
6631ee9d JR |
1361 | dma_ops = &amd_iommu_dma_ops; |
1362 | ||
1363 | return 0; | |
1364 | ||
1365 | free_domains: | |
1366 | ||
1367 | list_for_each_entry(iommu, &amd_iommu_list, list) { | |
1368 | if (iommu->default_dom) | |
1369 | dma_ops_domain_free(iommu->default_dom); | |
1370 | } | |
1371 | ||
1372 | return ret; | |
1373 | } |