4 * Copyright (c) 2009 Red Hat
6 * This work is licensed under the terms of the GNU General Public License
7 * (GNU GPL), version 2 or later.
13 #include "qemu-thread.h"
15 /* #define DEBUG_IOMMU */
17 static void do_dma_memory_set(dma_addr_t addr, uint8_t c, dma_addr_t len)
19 #define FILLBUF_SIZE 512
20 uint8_t fillbuf[FILLBUF_SIZE];
23 memset(fillbuf, c, FILLBUF_SIZE);
25 l = len < FILLBUF_SIZE ? len : FILLBUF_SIZE;
26 cpu_physical_memory_rw(addr, fillbuf, l, true);
32 int dma_memory_set(DMAContext *dma, dma_addr_t addr, uint8_t c, dma_addr_t len)
34 dma_barrier(dma, DMA_DIRECTION_FROM_DEVICE);
36 if (dma_has_iommu(dma)) {
37 return iommu_dma_memory_set(dma, addr, c, len);
39 do_dma_memory_set(addr, c, len);
44 void qemu_sglist_init(QEMUSGList *qsg, int alloc_hint, DMAContext *dma)
46 qsg->sg = g_malloc(alloc_hint * sizeof(ScatterGatherEntry));
48 qsg->nalloc = alloc_hint;
53 void qemu_sglist_add(QEMUSGList *qsg, dma_addr_t base, dma_addr_t len)
55 if (qsg->nsg == qsg->nalloc) {
56 qsg->nalloc = 2 * qsg->nalloc + 1;
57 qsg->sg = g_realloc(qsg->sg, qsg->nalloc * sizeof(ScatterGatherEntry));
59 qsg->sg[qsg->nsg].base = base;
60 qsg->sg[qsg->nsg].len = len;
65 void qemu_sglist_destroy(QEMUSGList *qsg)
68 memset(qsg, 0, sizeof(*qsg));
72 BlockDriverAIOCB common;
74 BlockDriverAIOCB *acb;
80 dma_addr_t sg_cur_byte;
86 static void dma_bdrv_cb(void *opaque, int ret);
88 static void reschedule_dma(void *opaque)
90 DMAAIOCB *dbs = (DMAAIOCB *)opaque;
92 qemu_bh_delete(dbs->bh);
97 static void continue_after_map_failure(void *opaque)
99 DMAAIOCB *dbs = (DMAAIOCB *)opaque;
101 dbs->bh = qemu_bh_new(reschedule_dma, dbs);
102 qemu_bh_schedule(dbs->bh);
105 static void dma_bdrv_unmap(DMAAIOCB *dbs)
109 for (i = 0; i < dbs->iov.niov; ++i) {
110 dma_memory_unmap(dbs->sg->dma, dbs->iov.iov[i].iov_base,
111 dbs->iov.iov[i].iov_len, dbs->dir,
112 dbs->iov.iov[i].iov_len);
114 qemu_iovec_reset(&dbs->iov);
117 static void dma_complete(DMAAIOCB *dbs, int ret)
119 trace_dma_complete(dbs, ret, dbs->common.cb);
122 if (dbs->common.cb) {
123 dbs->common.cb(dbs->common.opaque, ret);
125 qemu_iovec_destroy(&dbs->iov);
127 qemu_bh_delete(dbs->bh);
130 if (!dbs->in_cancel) {
131 /* Requests may complete while dma_aio_cancel is in progress. In
132 * this case, the AIOCB should not be released because it is still
133 * referenced by dma_aio_cancel. */
134 qemu_aio_release(dbs);
138 static void dma_bdrv_cb(void *opaque, int ret)
140 DMAAIOCB *dbs = (DMAAIOCB *)opaque;
141 dma_addr_t cur_addr, cur_len;
144 trace_dma_bdrv_cb(dbs, ret);
147 dbs->sector_num += dbs->iov.size / 512;
150 if (dbs->sg_cur_index == dbs->sg->nsg || ret < 0) {
151 dma_complete(dbs, ret);
155 while (dbs->sg_cur_index < dbs->sg->nsg) {
156 cur_addr = dbs->sg->sg[dbs->sg_cur_index].base + dbs->sg_cur_byte;
157 cur_len = dbs->sg->sg[dbs->sg_cur_index].len - dbs->sg_cur_byte;
158 mem = dma_memory_map(dbs->sg->dma, cur_addr, &cur_len, dbs->dir);
161 qemu_iovec_add(&dbs->iov, mem, cur_len);
162 dbs->sg_cur_byte += cur_len;
163 if (dbs->sg_cur_byte == dbs->sg->sg[dbs->sg_cur_index].len) {
164 dbs->sg_cur_byte = 0;
169 if (dbs->iov.size == 0) {
170 trace_dma_map_wait(dbs);
171 cpu_register_map_client(dbs, continue_after_map_failure);
175 dbs->acb = dbs->io_func(dbs->bs, dbs->sector_num, &dbs->iov,
176 dbs->iov.size / 512, dma_bdrv_cb, dbs);
180 static void dma_aio_cancel(BlockDriverAIOCB *acb)
182 DMAAIOCB *dbs = container_of(acb, DMAAIOCB, common);
184 trace_dma_aio_cancel(dbs);
187 BlockDriverAIOCB *acb = dbs->acb;
189 dbs->in_cancel = true;
190 bdrv_aio_cancel(acb);
191 dbs->in_cancel = false;
193 dbs->common.cb = NULL;
194 dma_complete(dbs, 0);
197 static AIOPool dma_aio_pool = {
198 .aiocb_size = sizeof(DMAAIOCB),
199 .cancel = dma_aio_cancel,
202 BlockDriverAIOCB *dma_bdrv_io(
203 BlockDriverState *bs, QEMUSGList *sg, uint64_t sector_num,
204 DMAIOFunc *io_func, BlockDriverCompletionFunc *cb,
205 void *opaque, DMADirection dir)
207 DMAAIOCB *dbs = qemu_aio_get(&dma_aio_pool, bs, cb, opaque);
209 trace_dma_bdrv_io(dbs, bs, sector_num, (dir == DMA_DIRECTION_TO_DEVICE));
214 dbs->sector_num = sector_num;
215 dbs->sg_cur_index = 0;
216 dbs->sg_cur_byte = 0;
218 dbs->io_func = io_func;
220 qemu_iovec_init(&dbs->iov, sg->nsg);
226 BlockDriverAIOCB *dma_bdrv_read(BlockDriverState *bs,
227 QEMUSGList *sg, uint64_t sector,
228 void (*cb)(void *opaque, int ret), void *opaque)
230 return dma_bdrv_io(bs, sg, sector, bdrv_aio_readv, cb, opaque,
231 DMA_DIRECTION_FROM_DEVICE);
234 BlockDriverAIOCB *dma_bdrv_write(BlockDriverState *bs,
235 QEMUSGList *sg, uint64_t sector,
236 void (*cb)(void *opaque, int ret), void *opaque)
238 return dma_bdrv_io(bs, sg, sector, bdrv_aio_writev, cb, opaque,
239 DMA_DIRECTION_TO_DEVICE);
243 static uint64_t dma_buf_rw(uint8_t *ptr, int32_t len, QEMUSGList *sg,
251 len = MIN(len, resid);
253 ScatterGatherEntry entry = sg->sg[sg_cur_index++];
254 int32_t xfer = MIN(len, entry.len);
255 dma_memory_rw(sg->dma, entry.base, ptr, xfer, dir);
264 uint64_t dma_buf_read(uint8_t *ptr, int32_t len, QEMUSGList *sg)
266 return dma_buf_rw(ptr, len, sg, DMA_DIRECTION_FROM_DEVICE);
269 uint64_t dma_buf_write(uint8_t *ptr, int32_t len, QEMUSGList *sg)
271 return dma_buf_rw(ptr, len, sg, DMA_DIRECTION_TO_DEVICE);
274 void dma_acct_start(BlockDriverState *bs, BlockAcctCookie *cookie,
275 QEMUSGList *sg, enum BlockAcctType type)
277 bdrv_acct_start(bs, cookie, sg->size, type);
280 bool iommu_dma_memory_valid(DMAContext *dma, dma_addr_t addr, dma_addr_t len,
283 target_phys_addr_t paddr, plen;
286 fprintf(stderr, "dma_memory_check context=%p addr=0x" DMA_ADDR_FMT
287 " len=0x" DMA_ADDR_FMT " dir=%d\n", dma, addr, len, dir);
291 if (dma->translate(dma, addr, &paddr, &plen, dir) != 0) {
295 /* The translation might be valid for larger regions. */
307 int iommu_dma_memory_rw(DMAContext *dma, dma_addr_t addr,
308 void *buf, dma_addr_t len, DMADirection dir)
310 target_phys_addr_t paddr, plen;
314 fprintf(stderr, "dma_memory_rw context=%p addr=0x" DMA_ADDR_FMT " len=0x"
315 DMA_ADDR_FMT " dir=%d\n", dma, addr, len, dir);
319 err = dma->translate(dma, addr, &paddr, &plen, dir);
322 * In case of failure on reads from the guest, we clean the
323 * destination buffer so that a device that doesn't test
324 * for errors will not expose qemu internal memory.
330 /* The translation might be valid for larger regions. */
335 cpu_physical_memory_rw(paddr, buf, plen,
336 dir == DMA_DIRECTION_FROM_DEVICE);
346 int iommu_dma_memory_set(DMAContext *dma, dma_addr_t addr, uint8_t c,
349 target_phys_addr_t paddr, plen;
353 fprintf(stderr, "dma_memory_set context=%p addr=0x" DMA_ADDR_FMT
354 " len=0x" DMA_ADDR_FMT "\n", dma, addr, len);
358 err = dma->translate(dma, addr, &paddr, &plen,
359 DMA_DIRECTION_FROM_DEVICE);
364 /* The translation might be valid for larger regions. */
369 do_dma_memory_set(paddr, c, plen);
378 void dma_context_init(DMAContext *dma, DMATranslateFunc translate,
379 DMAMapFunc map, DMAUnmapFunc unmap)
382 fprintf(stderr, "dma_context_init(%p, %p, %p, %p)\n",
383 dma, translate, map, unmap);
385 dma->translate = translate;
390 void *iommu_dma_memory_map(DMAContext *dma, dma_addr_t addr, dma_addr_t *len,
394 target_phys_addr_t paddr, plen;
398 return dma->map(dma, addr, len, dir);
402 err = dma->translate(dma, addr, &paddr, &plen, dir);
408 * If this is true, the virtual region is contiguous,
409 * but the translated physical region isn't. We just
410 * clamp *len, much like cpu_physical_memory_map() does.
416 buf = cpu_physical_memory_map(paddr, &plen,
417 dir == DMA_DIRECTION_FROM_DEVICE);
423 void iommu_dma_memory_unmap(DMAContext *dma, void *buffer, dma_addr_t len,
424 DMADirection dir, dma_addr_t access_len)
427 dma->unmap(dma, buffer, len, dir, access_len);
431 cpu_physical_memory_unmap(buffer, len,
432 dir == DMA_DIRECTION_FROM_DEVICE,