]> git.proxmox.com Git - mirror_qemu.git/blob - dma-helpers.c
qmp: add NBD server commands
[mirror_qemu.git] / dma-helpers.c
1 /*
2 * DMA helper functions
3 *
4 * Copyright (c) 2009 Red Hat
5 *
6 * This work is licensed under the terms of the GNU General Public License
7 * (GNU GPL), version 2 or later.
8 */
9
10 #include "dma.h"
11 #include "trace.h"
12 #include "range.h"
13 #include "qemu-thread.h"
14
15 /* #define DEBUG_IOMMU */
16
17 static void do_dma_memory_set(AddressSpace *as,
18 dma_addr_t addr, uint8_t c, dma_addr_t len)
19 {
20 #define FILLBUF_SIZE 512
21 uint8_t fillbuf[FILLBUF_SIZE];
22 int l;
23
24 memset(fillbuf, c, FILLBUF_SIZE);
25 while (len > 0) {
26 l = len < FILLBUF_SIZE ? len : FILLBUF_SIZE;
27 address_space_rw(as, addr, fillbuf, l, true);
28 len -= l;
29 addr += l;
30 }
31 }
32
33 int dma_memory_set(DMAContext *dma, dma_addr_t addr, uint8_t c, dma_addr_t len)
34 {
35 dma_barrier(dma, DMA_DIRECTION_FROM_DEVICE);
36
37 if (dma_has_iommu(dma)) {
38 return iommu_dma_memory_set(dma, addr, c, len);
39 }
40 do_dma_memory_set(dma->as, addr, c, len);
41
42 return 0;
43 }
44
45 void qemu_sglist_init(QEMUSGList *qsg, int alloc_hint, DMAContext *dma)
46 {
47 qsg->sg = g_malloc(alloc_hint * sizeof(ScatterGatherEntry));
48 qsg->nsg = 0;
49 qsg->nalloc = alloc_hint;
50 qsg->size = 0;
51 qsg->dma = dma;
52 }
53
54 void qemu_sglist_add(QEMUSGList *qsg, dma_addr_t base, dma_addr_t len)
55 {
56 if (qsg->nsg == qsg->nalloc) {
57 qsg->nalloc = 2 * qsg->nalloc + 1;
58 qsg->sg = g_realloc(qsg->sg, qsg->nalloc * sizeof(ScatterGatherEntry));
59 }
60 qsg->sg[qsg->nsg].base = base;
61 qsg->sg[qsg->nsg].len = len;
62 qsg->size += len;
63 ++qsg->nsg;
64 }
65
66 void qemu_sglist_destroy(QEMUSGList *qsg)
67 {
68 g_free(qsg->sg);
69 memset(qsg, 0, sizeof(*qsg));
70 }
71
72 typedef struct {
73 BlockDriverAIOCB common;
74 BlockDriverState *bs;
75 BlockDriverAIOCB *acb;
76 QEMUSGList *sg;
77 uint64_t sector_num;
78 DMADirection dir;
79 bool in_cancel;
80 int sg_cur_index;
81 dma_addr_t sg_cur_byte;
82 QEMUIOVector iov;
83 QEMUBH *bh;
84 DMAIOFunc *io_func;
85 } DMAAIOCB;
86
87 static void dma_bdrv_cb(void *opaque, int ret);
88
89 static void reschedule_dma(void *opaque)
90 {
91 DMAAIOCB *dbs = (DMAAIOCB *)opaque;
92
93 qemu_bh_delete(dbs->bh);
94 dbs->bh = NULL;
95 dma_bdrv_cb(dbs, 0);
96 }
97
98 static void continue_after_map_failure(void *opaque)
99 {
100 DMAAIOCB *dbs = (DMAAIOCB *)opaque;
101
102 dbs->bh = qemu_bh_new(reschedule_dma, dbs);
103 qemu_bh_schedule(dbs->bh);
104 }
105
106 static void dma_bdrv_unmap(DMAAIOCB *dbs)
107 {
108 int i;
109
110 for (i = 0; i < dbs->iov.niov; ++i) {
111 dma_memory_unmap(dbs->sg->dma, dbs->iov.iov[i].iov_base,
112 dbs->iov.iov[i].iov_len, dbs->dir,
113 dbs->iov.iov[i].iov_len);
114 }
115 qemu_iovec_reset(&dbs->iov);
116 }
117
118 static void dma_complete(DMAAIOCB *dbs, int ret)
119 {
120 trace_dma_complete(dbs, ret, dbs->common.cb);
121
122 dma_bdrv_unmap(dbs);
123 if (dbs->common.cb) {
124 dbs->common.cb(dbs->common.opaque, ret);
125 }
126 qemu_iovec_destroy(&dbs->iov);
127 if (dbs->bh) {
128 qemu_bh_delete(dbs->bh);
129 dbs->bh = NULL;
130 }
131 if (!dbs->in_cancel) {
132 /* Requests may complete while dma_aio_cancel is in progress. In
133 * this case, the AIOCB should not be released because it is still
134 * referenced by dma_aio_cancel. */
135 qemu_aio_release(dbs);
136 }
137 }
138
139 static void dma_bdrv_cb(void *opaque, int ret)
140 {
141 DMAAIOCB *dbs = (DMAAIOCB *)opaque;
142 dma_addr_t cur_addr, cur_len;
143 void *mem;
144
145 trace_dma_bdrv_cb(dbs, ret);
146
147 dbs->acb = NULL;
148 dbs->sector_num += dbs->iov.size / 512;
149 dma_bdrv_unmap(dbs);
150
151 if (dbs->sg_cur_index == dbs->sg->nsg || ret < 0) {
152 dma_complete(dbs, ret);
153 return;
154 }
155
156 while (dbs->sg_cur_index < dbs->sg->nsg) {
157 cur_addr = dbs->sg->sg[dbs->sg_cur_index].base + dbs->sg_cur_byte;
158 cur_len = dbs->sg->sg[dbs->sg_cur_index].len - dbs->sg_cur_byte;
159 mem = dma_memory_map(dbs->sg->dma, cur_addr, &cur_len, dbs->dir);
160 if (!mem)
161 break;
162 qemu_iovec_add(&dbs->iov, mem, cur_len);
163 dbs->sg_cur_byte += cur_len;
164 if (dbs->sg_cur_byte == dbs->sg->sg[dbs->sg_cur_index].len) {
165 dbs->sg_cur_byte = 0;
166 ++dbs->sg_cur_index;
167 }
168 }
169
170 if (dbs->iov.size == 0) {
171 trace_dma_map_wait(dbs);
172 cpu_register_map_client(dbs, continue_after_map_failure);
173 return;
174 }
175
176 dbs->acb = dbs->io_func(dbs->bs, dbs->sector_num, &dbs->iov,
177 dbs->iov.size / 512, dma_bdrv_cb, dbs);
178 assert(dbs->acb);
179 }
180
181 static void dma_aio_cancel(BlockDriverAIOCB *acb)
182 {
183 DMAAIOCB *dbs = container_of(acb, DMAAIOCB, common);
184
185 trace_dma_aio_cancel(dbs);
186
187 if (dbs->acb) {
188 BlockDriverAIOCB *acb = dbs->acb;
189 dbs->acb = NULL;
190 dbs->in_cancel = true;
191 bdrv_aio_cancel(acb);
192 dbs->in_cancel = false;
193 }
194 dbs->common.cb = NULL;
195 dma_complete(dbs, 0);
196 }
197
198 static AIOPool dma_aio_pool = {
199 .aiocb_size = sizeof(DMAAIOCB),
200 .cancel = dma_aio_cancel,
201 };
202
203 BlockDriverAIOCB *dma_bdrv_io(
204 BlockDriverState *bs, QEMUSGList *sg, uint64_t sector_num,
205 DMAIOFunc *io_func, BlockDriverCompletionFunc *cb,
206 void *opaque, DMADirection dir)
207 {
208 DMAAIOCB *dbs = qemu_aio_get(&dma_aio_pool, bs, cb, opaque);
209
210 trace_dma_bdrv_io(dbs, bs, sector_num, (dir == DMA_DIRECTION_TO_DEVICE));
211
212 dbs->acb = NULL;
213 dbs->bs = bs;
214 dbs->sg = sg;
215 dbs->sector_num = sector_num;
216 dbs->sg_cur_index = 0;
217 dbs->sg_cur_byte = 0;
218 dbs->dir = dir;
219 dbs->io_func = io_func;
220 dbs->bh = NULL;
221 qemu_iovec_init(&dbs->iov, sg->nsg);
222 dma_bdrv_cb(dbs, 0);
223 return &dbs->common;
224 }
225
226
227 BlockDriverAIOCB *dma_bdrv_read(BlockDriverState *bs,
228 QEMUSGList *sg, uint64_t sector,
229 void (*cb)(void *opaque, int ret), void *opaque)
230 {
231 return dma_bdrv_io(bs, sg, sector, bdrv_aio_readv, cb, opaque,
232 DMA_DIRECTION_FROM_DEVICE);
233 }
234
235 BlockDriverAIOCB *dma_bdrv_write(BlockDriverState *bs,
236 QEMUSGList *sg, uint64_t sector,
237 void (*cb)(void *opaque, int ret), void *opaque)
238 {
239 return dma_bdrv_io(bs, sg, sector, bdrv_aio_writev, cb, opaque,
240 DMA_DIRECTION_TO_DEVICE);
241 }
242
243
244 static uint64_t dma_buf_rw(uint8_t *ptr, int32_t len, QEMUSGList *sg,
245 DMADirection dir)
246 {
247 uint64_t resid;
248 int sg_cur_index;
249
250 resid = sg->size;
251 sg_cur_index = 0;
252 len = MIN(len, resid);
253 while (len > 0) {
254 ScatterGatherEntry entry = sg->sg[sg_cur_index++];
255 int32_t xfer = MIN(len, entry.len);
256 dma_memory_rw(sg->dma, entry.base, ptr, xfer, dir);
257 ptr += xfer;
258 len -= xfer;
259 resid -= xfer;
260 }
261
262 return resid;
263 }
264
265 uint64_t dma_buf_read(uint8_t *ptr, int32_t len, QEMUSGList *sg)
266 {
267 return dma_buf_rw(ptr, len, sg, DMA_DIRECTION_FROM_DEVICE);
268 }
269
270 uint64_t dma_buf_write(uint8_t *ptr, int32_t len, QEMUSGList *sg)
271 {
272 return dma_buf_rw(ptr, len, sg, DMA_DIRECTION_TO_DEVICE);
273 }
274
275 void dma_acct_start(BlockDriverState *bs, BlockAcctCookie *cookie,
276 QEMUSGList *sg, enum BlockAcctType type)
277 {
278 bdrv_acct_start(bs, cookie, sg->size, type);
279 }
280
281 bool iommu_dma_memory_valid(DMAContext *dma, dma_addr_t addr, dma_addr_t len,
282 DMADirection dir)
283 {
284 target_phys_addr_t paddr, plen;
285
286 #ifdef DEBUG_IOMMU
287 fprintf(stderr, "dma_memory_check context=%p addr=0x" DMA_ADDR_FMT
288 " len=0x" DMA_ADDR_FMT " dir=%d\n", dma, addr, len, dir);
289 #endif
290
291 while (len) {
292 if (dma->translate(dma, addr, &paddr, &plen, dir) != 0) {
293 return false;
294 }
295
296 /* The translation might be valid for larger regions. */
297 if (plen > len) {
298 plen = len;
299 }
300
301 len -= plen;
302 addr += plen;
303 }
304
305 return true;
306 }
307
308 int iommu_dma_memory_rw(DMAContext *dma, dma_addr_t addr,
309 void *buf, dma_addr_t len, DMADirection dir)
310 {
311 target_phys_addr_t paddr, plen;
312 int err;
313
314 #ifdef DEBUG_IOMMU
315 fprintf(stderr, "dma_memory_rw context=%p addr=0x" DMA_ADDR_FMT " len=0x"
316 DMA_ADDR_FMT " dir=%d\n", dma, addr, len, dir);
317 #endif
318
319 while (len) {
320 err = dma->translate(dma, addr, &paddr, &plen, dir);
321 if (err) {
322 /*
323 * In case of failure on reads from the guest, we clean the
324 * destination buffer so that a device that doesn't test
325 * for errors will not expose qemu internal memory.
326 */
327 memset(buf, 0, len);
328 return -1;
329 }
330
331 /* The translation might be valid for larger regions. */
332 if (plen > len) {
333 plen = len;
334 }
335
336 address_space_rw(dma->as, paddr, buf, plen, dir == DMA_DIRECTION_FROM_DEVICE);
337
338 len -= plen;
339 addr += plen;
340 buf += plen;
341 }
342
343 return 0;
344 }
345
346 int iommu_dma_memory_set(DMAContext *dma, dma_addr_t addr, uint8_t c,
347 dma_addr_t len)
348 {
349 target_phys_addr_t paddr, plen;
350 int err;
351
352 #ifdef DEBUG_IOMMU
353 fprintf(stderr, "dma_memory_set context=%p addr=0x" DMA_ADDR_FMT
354 " len=0x" DMA_ADDR_FMT "\n", dma, addr, len);
355 #endif
356
357 while (len) {
358 err = dma->translate(dma, addr, &paddr, &plen,
359 DMA_DIRECTION_FROM_DEVICE);
360 if (err) {
361 return err;
362 }
363
364 /* The translation might be valid for larger regions. */
365 if (plen > len) {
366 plen = len;
367 }
368
369 do_dma_memory_set(dma->as, paddr, c, plen);
370
371 len -= plen;
372 addr += plen;
373 }
374
375 return 0;
376 }
377
378 void dma_context_init(DMAContext *dma, AddressSpace *as, DMATranslateFunc translate,
379 DMAMapFunc map, DMAUnmapFunc unmap)
380 {
381 #ifdef DEBUG_IOMMU
382 fprintf(stderr, "dma_context_init(%p, %p, %p, %p)\n",
383 dma, translate, map, unmap);
384 #endif
385 dma->as = as;
386 dma->translate = translate;
387 dma->map = map;
388 dma->unmap = unmap;
389 }
390
391 void *iommu_dma_memory_map(DMAContext *dma, dma_addr_t addr, dma_addr_t *len,
392 DMADirection dir)
393 {
394 int err;
395 target_phys_addr_t paddr, plen;
396 void *buf;
397
398 if (dma->map) {
399 return dma->map(dma, addr, len, dir);
400 }
401
402 plen = *len;
403 err = dma->translate(dma, addr, &paddr, &plen, dir);
404 if (err) {
405 return NULL;
406 }
407
408 /*
409 * If this is true, the virtual region is contiguous,
410 * but the translated physical region isn't. We just
411 * clamp *len, much like address_space_map() does.
412 */
413 if (plen < *len) {
414 *len = plen;
415 }
416
417 buf = address_space_map(dma->as, paddr, &plen, dir == DMA_DIRECTION_FROM_DEVICE);
418 *len = plen;
419
420 return buf;
421 }
422
423 void iommu_dma_memory_unmap(DMAContext *dma, void *buffer, dma_addr_t len,
424 DMADirection dir, dma_addr_t access_len)
425 {
426 if (dma->unmap) {
427 dma->unmap(dma, buffer, len, dir, access_len);
428 return;
429 }
430
431 address_space_unmap(dma->as, buffer, len, dir == DMA_DIRECTION_FROM_DEVICE,
432 access_len);
433
434 }