1 /* Virtio ring implementation.
3 * Copyright 2007 Rusty Russell IBM Corporation
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
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.
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., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 #include <linux/virtio.h>
20 #include <linux/virtio_ring.h>
21 #include <linux/virtio_config.h>
22 #include <linux/device.h>
23 #include <linux/slab.h>
24 #include <linux/module.h>
25 #include <linux/hrtimer.h>
26 #include <linux/kmemleak.h>
27 #include <linux/dma-mapping.h>
31 /* For development, we want to crash whenever the ring is screwed. */
32 #define BAD_RING(_vq, fmt, args...) \
34 dev_err(&(_vq)->vq.vdev->dev, \
35 "%s:"fmt, (_vq)->vq.name, ##args); \
38 /* Caller is supposed to guarantee no reentry. */
39 #define START_USE(_vq) \
42 panic("%s:in_use = %i\n", \
43 (_vq)->vq.name, (_vq)->in_use); \
44 (_vq)->in_use = __LINE__; \
46 #define END_USE(_vq) \
47 do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
49 #define BAD_RING(_vq, fmt, args...) \
51 dev_err(&_vq->vq.vdev->dev, \
52 "%s:"fmt, (_vq)->vq.name, ##args); \
53 (_vq)->broken = true; \
59 struct vring_desc_state {
60 void *data; /* Data for callback. */
61 struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
64 struct vring_virtqueue {
67 /* Actual memory layout for this queue */
70 /* Can we use weak barriers? */
73 /* Other side has made a mess, don't try any more. */
76 /* Host supports indirect buffers */
79 /* Host publishes avail event idx */
82 /* Head of free buffer list. */
83 unsigned int free_head;
84 /* Number we've added since last sync. */
85 unsigned int num_added;
87 /* Last used index we've seen. */
90 /* Last written value to avail->flags */
91 u16 avail_flags_shadow;
93 /* Last written value to avail->idx in guest byte order */
96 /* How to notify other side. FIXME: commonalize hcalls! */
97 bool (*notify)(struct virtqueue *vq);
99 /* DMA, allocation, and size information */
101 size_t queue_size_in_bytes;
102 dma_addr_t queue_dma_addr;
105 /* They're supposed to lock for us. */
108 /* Figure out if their kicks are too delayed. */
109 bool last_add_time_valid;
110 ktime_t last_add_time;
113 /* Per-descriptor state. */
114 struct vring_desc_state desc_state[];
117 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
120 * Modern virtio devices have feature bits to specify whether they need a
121 * quirk and bypass the IOMMU. If not there, just use the DMA API.
123 * If there, the interaction between virtio and DMA API is messy.
125 * On most systems with virtio, physical addresses match bus addresses,
126 * and it doesn't particularly matter whether we use the DMA API.
128 * On some systems, including Xen and any system with a physical device
129 * that speaks virtio behind a physical IOMMU, we must use the DMA API
130 * for virtio DMA to work at all.
132 * On other systems, including SPARC and PPC64, virtio-pci devices are
133 * enumerated as though they are behind an IOMMU, but the virtio host
134 * ignores the IOMMU, so we must either pretend that the IOMMU isn't
135 * there or somehow map everything as the identity.
137 * For the time being, we preserve historic behavior and bypass the DMA
140 * TODO: install a per-device DMA ops structure that does the right thing
141 * taking into account all the above quirks, and use the DMA API
142 * unconditionally on data path.
145 static bool vring_use_dma_api(struct virtio_device *vdev)
147 if (!virtio_has_iommu_quirk(vdev))
150 /* Otherwise, we are left to guess. */
152 * In theory, it's possible to have a buggy QEMU-supposed
153 * emulated Q35 IOMMU and Xen enabled at the same time. On
154 * such a configuration, virtio has never worked and will
155 * not work without an even larger kludge. Instead, enable
156 * the DMA API if we're a Xen guest, which at least allows
157 * all of the sensible Xen configurations to work correctly.
166 * The DMA ops on various arches are rather gnarly right now, and
167 * making all of the arch DMA ops work on the vring device itself
168 * is a mess. For now, we use the parent device for DMA ops.
170 static inline struct device *vring_dma_dev(const struct vring_virtqueue *vq)
172 return vq->vq.vdev->dev.parent;
175 /* Map one sg entry. */
176 static dma_addr_t vring_map_one_sg(const struct vring_virtqueue *vq,
177 struct scatterlist *sg,
178 enum dma_data_direction direction)
180 if (!vring_use_dma_api(vq->vq.vdev))
181 return (dma_addr_t)sg_phys(sg);
184 * We can't use dma_map_sg, because we don't use scatterlists in
185 * the way it expects (we don't guarantee that the scatterlist
186 * will exist for the lifetime of the mapping).
188 return dma_map_page(vring_dma_dev(vq),
189 sg_page(sg), sg->offset, sg->length,
193 static dma_addr_t vring_map_single(const struct vring_virtqueue *vq,
194 void *cpu_addr, size_t size,
195 enum dma_data_direction direction)
197 if (!vring_use_dma_api(vq->vq.vdev))
198 return (dma_addr_t)virt_to_phys(cpu_addr);
200 return dma_map_single(vring_dma_dev(vq),
201 cpu_addr, size, direction);
204 static void vring_unmap_one(const struct vring_virtqueue *vq,
205 struct vring_desc *desc)
209 if (!vring_use_dma_api(vq->vq.vdev))
212 flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
214 if (flags & VRING_DESC_F_INDIRECT) {
215 dma_unmap_single(vring_dma_dev(vq),
216 virtio64_to_cpu(vq->vq.vdev, desc->addr),
217 virtio32_to_cpu(vq->vq.vdev, desc->len),
218 (flags & VRING_DESC_F_WRITE) ?
219 DMA_FROM_DEVICE : DMA_TO_DEVICE);
221 dma_unmap_page(vring_dma_dev(vq),
222 virtio64_to_cpu(vq->vq.vdev, desc->addr),
223 virtio32_to_cpu(vq->vq.vdev, desc->len),
224 (flags & VRING_DESC_F_WRITE) ?
225 DMA_FROM_DEVICE : DMA_TO_DEVICE);
229 static int vring_mapping_error(const struct vring_virtqueue *vq,
232 if (!vring_use_dma_api(vq->vq.vdev))
235 return dma_mapping_error(vring_dma_dev(vq), addr);
238 static struct vring_desc *alloc_indirect(struct virtqueue *_vq,
239 unsigned int total_sg, gfp_t gfp)
241 struct vring_desc *desc;
245 * We require lowmem mappings for the descriptors because
246 * otherwise virt_to_phys will give us bogus addresses in the
249 gfp &= ~__GFP_HIGHMEM;
251 desc = kmalloc(total_sg * sizeof(struct vring_desc), gfp);
255 for (i = 0; i < total_sg; i++)
256 desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
260 static inline int virtqueue_add(struct virtqueue *_vq,
261 struct scatterlist *sgs[],
262 unsigned int total_sg,
263 unsigned int out_sgs,
269 struct vring_virtqueue *vq = to_vvq(_vq);
270 struct scatterlist *sg;
271 struct vring_desc *desc;
272 unsigned int i, n, avail, descs_used, prev, err_idx;
278 BUG_ON(data == NULL);
279 BUG_ON(ctx && vq->indirect);
281 if (unlikely(vq->broken)) {
288 ktime_t now = ktime_get();
290 /* No kick or get, with .1 second between? Warn. */
291 if (vq->last_add_time_valid)
292 WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
294 vq->last_add_time = now;
295 vq->last_add_time_valid = true;
299 BUG_ON(total_sg == 0);
301 head = vq->free_head;
303 /* If the host supports indirect descriptor tables, and we have multiple
304 * buffers, then go indirect. FIXME: tune this threshold */
305 if (vq->indirect && total_sg > 1 && vq->vq.num_free)
306 desc = alloc_indirect(_vq, total_sg, gfp);
309 WARN_ON_ONCE(total_sg > vq->vring.num && !vq->indirect);
313 /* Use a single buffer which doesn't continue */
315 /* Set up rest to use this indirect table. */
320 desc = vq->vring.desc;
322 descs_used = total_sg;
325 if (vq->vq.num_free < descs_used) {
326 pr_debug("Can't add buf len %i - avail = %i\n",
327 descs_used, vq->vq.num_free);
328 /* FIXME: for historical reasons, we force a notify here if
329 * there are outgoing parts to the buffer. Presumably the
330 * host should service the ring ASAP. */
339 for (n = 0; n < out_sgs; n++) {
340 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
341 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
342 if (vring_mapping_error(vq, addr))
345 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
346 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
347 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
349 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
352 for (; n < (out_sgs + in_sgs); n++) {
353 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
354 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
355 if (vring_mapping_error(vq, addr))
358 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
359 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
360 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
362 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
365 /* Last one doesn't continue. */
366 desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
369 /* Now that the indirect table is filled in, map it. */
370 dma_addr_t addr = vring_map_single(
371 vq, desc, total_sg * sizeof(struct vring_desc),
373 if (vring_mapping_error(vq, addr))
376 vq->vring.desc[head].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_INDIRECT);
377 vq->vring.desc[head].addr = cpu_to_virtio64(_vq->vdev, addr);
379 vq->vring.desc[head].len = cpu_to_virtio32(_vq->vdev, total_sg * sizeof(struct vring_desc));
382 /* We're using some buffers from the free list. */
383 vq->vq.num_free -= descs_used;
385 /* Update free pointer */
387 vq->free_head = virtio16_to_cpu(_vq->vdev, vq->vring.desc[head].next);
391 /* Store token and indirect buffer state. */
392 vq->desc_state[head].data = data;
394 vq->desc_state[head].indir_desc = desc;
396 vq->desc_state[head].indir_desc = ctx;
398 /* Put entry in available array (but don't update avail->idx until they
400 avail = vq->avail_idx_shadow & (vq->vring.num - 1);
401 vq->vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
403 /* Descriptors and available array need to be set before we expose the
404 * new available array entries. */
405 virtio_wmb(vq->weak_barriers);
406 vq->avail_idx_shadow++;
407 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
410 pr_debug("Added buffer head %i to %p\n", head, vq);
413 /* This is very unlikely, but theoretically possible. Kick
415 if (unlikely(vq->num_added == (1 << 16) - 1))
424 for (n = 0; n < total_sg; n++) {
427 vring_unmap_one(vq, &desc[i]);
428 i = virtio16_to_cpu(_vq->vdev, vq->vring.desc[i].next);
439 * virtqueue_add_sgs - expose buffers to other end
440 * @vq: the struct virtqueue we're talking about.
441 * @sgs: array of terminated scatterlists.
442 * @out_num: the number of scatterlists readable by other side
443 * @in_num: the number of scatterlists which are writable (after readable ones)
444 * @data: the token identifying the buffer.
445 * @gfp: how to do memory allocations (if necessary).
447 * Caller must ensure we don't call this with other virtqueue operations
448 * at the same time (except where noted).
450 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
452 int virtqueue_add_sgs(struct virtqueue *_vq,
453 struct scatterlist *sgs[],
454 unsigned int out_sgs,
459 unsigned int i, total_sg = 0;
461 /* Count them first. */
462 for (i = 0; i < out_sgs + in_sgs; i++) {
463 struct scatterlist *sg;
464 for (sg = sgs[i]; sg; sg = sg_next(sg))
467 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs,
470 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
473 * virtqueue_add_outbuf - expose output buffers to other end
474 * @vq: the struct virtqueue we're talking about.
475 * @sg: scatterlist (must be well-formed and terminated!)
476 * @num: the number of entries in @sg readable by other side
477 * @data: the token identifying the buffer.
478 * @gfp: how to do memory allocations (if necessary).
480 * Caller must ensure we don't call this with other virtqueue operations
481 * at the same time (except where noted).
483 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
485 int virtqueue_add_outbuf(struct virtqueue *vq,
486 struct scatterlist *sg, unsigned int num,
490 return virtqueue_add(vq, &sg, num, 1, 0, data, NULL, gfp);
492 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
495 * virtqueue_add_inbuf - expose input buffers to other end
496 * @vq: the struct virtqueue we're talking about.
497 * @sg: scatterlist (must be well-formed and terminated!)
498 * @num: the number of entries in @sg writable by other side
499 * @data: the token identifying the buffer.
500 * @gfp: how to do memory allocations (if necessary).
502 * Caller must ensure we don't call this with other virtqueue operations
503 * at the same time (except where noted).
505 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
507 int virtqueue_add_inbuf(struct virtqueue *vq,
508 struct scatterlist *sg, unsigned int num,
512 return virtqueue_add(vq, &sg, num, 0, 1, data, NULL, gfp);
514 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
517 * virtqueue_add_inbuf_ctx - expose input buffers to other end
518 * @vq: the struct virtqueue we're talking about.
519 * @sg: scatterlist (must be well-formed and terminated!)
520 * @num: the number of entries in @sg writable by other side
521 * @data: the token identifying the buffer.
522 * @ctx: extra context for the token
523 * @gfp: how to do memory allocations (if necessary).
525 * Caller must ensure we don't call this with other virtqueue operations
526 * at the same time (except where noted).
528 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
530 int virtqueue_add_inbuf_ctx(struct virtqueue *vq,
531 struct scatterlist *sg, unsigned int num,
536 return virtqueue_add(vq, &sg, num, 0, 1, data, ctx, gfp);
538 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf_ctx);
541 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
542 * @vq: the struct virtqueue
544 * Instead of virtqueue_kick(), you can do:
545 * if (virtqueue_kick_prepare(vq))
546 * virtqueue_notify(vq);
548 * This is sometimes useful because the virtqueue_kick_prepare() needs
549 * to be serialized, but the actual virtqueue_notify() call does not.
551 bool virtqueue_kick_prepare(struct virtqueue *_vq)
553 struct vring_virtqueue *vq = to_vvq(_vq);
558 /* We need to expose available array entries before checking avail
560 virtio_mb(vq->weak_barriers);
562 old = vq->avail_idx_shadow - vq->num_added;
563 new = vq->avail_idx_shadow;
567 if (vq->last_add_time_valid) {
568 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
569 vq->last_add_time)) > 100);
571 vq->last_add_time_valid = false;
575 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev, vring_avail_event(&vq->vring)),
578 needs_kick = !(vq->vring.used->flags & cpu_to_virtio16(_vq->vdev, VRING_USED_F_NO_NOTIFY));
583 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
586 * virtqueue_notify - second half of split virtqueue_kick call.
587 * @vq: the struct virtqueue
589 * This does not need to be serialized.
591 * Returns false if host notify failed or queue is broken, otherwise true.
593 bool virtqueue_notify(struct virtqueue *_vq)
595 struct vring_virtqueue *vq = to_vvq(_vq);
597 if (unlikely(vq->broken))
600 /* Prod other side to tell it about changes. */
601 if (!vq->notify(_vq)) {
607 EXPORT_SYMBOL_GPL(virtqueue_notify);
610 * virtqueue_kick - update after add_buf
611 * @vq: the struct virtqueue
613 * After one or more virtqueue_add_* calls, invoke this to kick
616 * Caller must ensure we don't call this with other virtqueue
617 * operations at the same time (except where noted).
619 * Returns false if kick failed, otherwise true.
621 bool virtqueue_kick(struct virtqueue *vq)
623 if (virtqueue_kick_prepare(vq))
624 return virtqueue_notify(vq);
627 EXPORT_SYMBOL_GPL(virtqueue_kick);
629 static void detach_buf(struct vring_virtqueue *vq, unsigned int head,
633 __virtio16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
635 /* Clear data ptr. */
636 vq->desc_state[head].data = NULL;
638 /* Put back on free list: unmap first-level descriptors and find end */
641 while (vq->vring.desc[i].flags & nextflag) {
642 vring_unmap_one(vq, &vq->vring.desc[i]);
643 i = virtio16_to_cpu(vq->vq.vdev, vq->vring.desc[i].next);
647 vring_unmap_one(vq, &vq->vring.desc[i]);
648 vq->vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev, vq->free_head);
649 vq->free_head = head;
651 /* Plus final descriptor */
655 struct vring_desc *indir_desc = vq->desc_state[head].indir_desc;
658 /* Free the indirect table, if any, now that it's unmapped. */
662 len = virtio32_to_cpu(vq->vq.vdev, vq->vring.desc[head].len);
664 BUG_ON(!(vq->vring.desc[head].flags &
665 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
666 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
668 for (j = 0; j < len / sizeof(struct vring_desc); j++)
669 vring_unmap_one(vq, &indir_desc[j]);
672 vq->desc_state[head].indir_desc = NULL;
674 *ctx = vq->desc_state[head].indir_desc;
678 static inline bool more_used(const struct vring_virtqueue *vq)
680 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev, vq->vring.used->idx);
684 * virtqueue_get_buf - get the next used buffer
685 * @vq: the struct virtqueue we're talking about.
686 * @len: the length written into the buffer
688 * If the device wrote data into the buffer, @len will be set to the
689 * amount written. This means you don't need to clear the buffer
690 * beforehand to ensure there's no data leakage in the case of short
693 * Caller must ensure we don't call this with other virtqueue
694 * operations at the same time (except where noted).
696 * Returns NULL if there are no used buffers, or the "data" token
697 * handed to virtqueue_add_*().
699 void *virtqueue_get_buf_ctx(struct virtqueue *_vq, unsigned int *len,
702 struct vring_virtqueue *vq = to_vvq(_vq);
709 if (unlikely(vq->broken)) {
714 if (!more_used(vq)) {
715 pr_debug("No more buffers in queue\n");
720 /* Only get used array entries after they have been exposed by host. */
721 virtio_rmb(vq->weak_barriers);
723 last_used = (vq->last_used_idx & (vq->vring.num - 1));
724 i = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].id);
725 *len = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].len);
727 if (unlikely(i >= vq->vring.num)) {
728 BAD_RING(vq, "id %u out of range\n", i);
731 if (unlikely(!vq->desc_state[i].data)) {
732 BAD_RING(vq, "id %u is not a head!\n", i);
736 /* detach_buf clears data, so grab it now. */
737 ret = vq->desc_state[i].data;
738 detach_buf(vq, i, ctx);
740 /* If we expect an interrupt for the next entry, tell host
741 * by writing event index and flush out the write before
742 * the read in the next get_buf call. */
743 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
744 virtio_store_mb(vq->weak_barriers,
745 &vring_used_event(&vq->vring),
746 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
749 vq->last_add_time_valid = false;
755 EXPORT_SYMBOL_GPL(virtqueue_get_buf_ctx);
757 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
759 return virtqueue_get_buf_ctx(_vq, len, NULL);
761 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
763 * virtqueue_disable_cb - disable callbacks
764 * @vq: the struct virtqueue we're talking about.
766 * Note that this is not necessarily synchronous, hence unreliable and only
767 * useful as an optimization.
769 * Unlike other operations, this need not be serialized.
771 void virtqueue_disable_cb(struct virtqueue *_vq)
773 struct vring_virtqueue *vq = to_vvq(_vq);
775 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
776 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
778 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
782 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
785 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
786 * @vq: the struct virtqueue we're talking about.
788 * This re-enables callbacks; it returns current queue state
789 * in an opaque unsigned value. This value should be later tested by
790 * virtqueue_poll, to detect a possible race between the driver checking for
791 * more work, and enabling callbacks.
793 * Caller must ensure we don't call this with other virtqueue
794 * operations at the same time (except where noted).
796 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
798 struct vring_virtqueue *vq = to_vvq(_vq);
803 /* We optimistically turn back on interrupts, then check if there was
805 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
806 * either clear the flags bit or point the event index at the next
807 * entry. Always do both to keep code simple. */
808 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
809 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
811 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
813 vring_used_event(&vq->vring) = cpu_to_virtio16(_vq->vdev, last_used_idx = vq->last_used_idx);
815 return last_used_idx;
817 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
820 * virtqueue_poll - query pending used buffers
821 * @vq: the struct virtqueue we're talking about.
822 * @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
824 * Returns "true" if there are pending used buffers in the queue.
826 * This does not need to be serialized.
828 bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
830 struct vring_virtqueue *vq = to_vvq(_vq);
832 if (unlikely(vq->broken))
835 virtio_mb(vq->weak_barriers);
836 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev, vq->vring.used->idx);
838 EXPORT_SYMBOL_GPL(virtqueue_poll);
841 * virtqueue_enable_cb - restart callbacks after disable_cb.
842 * @vq: the struct virtqueue we're talking about.
844 * This re-enables callbacks; it returns "false" if there are pending
845 * buffers in the queue, to detect a possible race between the driver
846 * checking for more work, and enabling callbacks.
848 * Caller must ensure we don't call this with other virtqueue
849 * operations at the same time (except where noted).
851 bool virtqueue_enable_cb(struct virtqueue *_vq)
853 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
854 return !virtqueue_poll(_vq, last_used_idx);
856 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
859 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
860 * @vq: the struct virtqueue we're talking about.
862 * This re-enables callbacks but hints to the other side to delay
863 * interrupts until most of the available buffers have been processed;
864 * it returns "false" if there are many pending buffers in the queue,
865 * to detect a possible race between the driver checking for more work,
866 * and enabling callbacks.
868 * Caller must ensure we don't call this with other virtqueue
869 * operations at the same time (except where noted).
871 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
873 struct vring_virtqueue *vq = to_vvq(_vq);
878 /* We optimistically turn back on interrupts, then check if there was
880 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
881 * either clear the flags bit or point the event index at the next
882 * entry. Always update the event index to keep code simple. */
883 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
884 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
886 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
888 /* TODO: tune this threshold */
889 bufs = (u16)(vq->avail_idx_shadow - vq->last_used_idx) * 3 / 4;
891 virtio_store_mb(vq->weak_barriers,
892 &vring_used_event(&vq->vring),
893 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
895 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->vring.used->idx) - vq->last_used_idx) > bufs)) {
903 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
906 * virtqueue_detach_unused_buf - detach first unused buffer
907 * @vq: the struct virtqueue we're talking about.
909 * Returns NULL or the "data" token handed to virtqueue_add_*().
910 * This is not valid on an active queue; it is useful only for device
913 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
915 struct vring_virtqueue *vq = to_vvq(_vq);
921 for (i = 0; i < vq->vring.num; i++) {
922 if (!vq->desc_state[i].data)
924 /* detach_buf clears data, so grab it now. */
925 buf = vq->desc_state[i].data;
926 detach_buf(vq, i, NULL);
927 vq->avail_idx_shadow--;
928 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
932 /* That should have freed everything. */
933 BUG_ON(vq->vq.num_free != vq->vring.num);
938 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
940 irqreturn_t vring_interrupt(int irq, void *_vq)
942 struct vring_virtqueue *vq = to_vvq(_vq);
944 if (!more_used(vq)) {
945 pr_debug("virtqueue interrupt with no work for %p\n", vq);
949 if (unlikely(vq->broken))
952 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
954 vq->vq.callback(&vq->vq);
958 EXPORT_SYMBOL_GPL(vring_interrupt);
960 struct virtqueue *__vring_new_virtqueue(unsigned int index,
962 struct virtio_device *vdev,
965 bool (*notify)(struct virtqueue *),
966 void (*callback)(struct virtqueue *),
970 struct vring_virtqueue *vq;
972 vq = kmalloc(sizeof(*vq) + vring.num * sizeof(struct vring_desc_state),
978 vq->vq.callback = callback;
981 vq->vq.num_free = vring.num;
982 vq->vq.index = index;
983 vq->we_own_ring = false;
984 vq->queue_dma_addr = 0;
985 vq->queue_size_in_bytes = 0;
987 vq->weak_barriers = weak_barriers;
989 vq->last_used_idx = 0;
990 vq->avail_flags_shadow = 0;
991 vq->avail_idx_shadow = 0;
993 list_add_tail(&vq->vq.list, &vdev->vqs);
996 vq->last_add_time_valid = false;
999 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
1001 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
1003 /* No callback? Tell other side not to bother us. */
1005 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
1007 vq->vring.avail->flags = cpu_to_virtio16(vdev, vq->avail_flags_shadow);
1010 /* Put everything in free lists. */
1012 for (i = 0; i < vring.num-1; i++)
1013 vq->vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
1014 memset(vq->desc_state, 0, vring.num * sizeof(struct vring_desc_state));
1018 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
1020 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
1021 dma_addr_t *dma_handle, gfp_t flag)
1023 if (vring_use_dma_api(vdev)) {
1024 return dma_alloc_coherent(vdev->dev.parent, size,
1027 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
1029 phys_addr_t phys_addr = virt_to_phys(queue);
1030 *dma_handle = (dma_addr_t)phys_addr;
1033 * Sanity check: make sure we dind't truncate
1034 * the address. The only arches I can find that
1035 * have 64-bit phys_addr_t but 32-bit dma_addr_t
1036 * are certain non-highmem MIPS and x86
1037 * configurations, but these configurations
1038 * should never allocate physical pages above 32
1039 * bits, so this is fine. Just in case, throw a
1040 * warning and abort if we end up with an
1041 * unrepresentable address.
1043 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
1044 free_pages_exact(queue, PAGE_ALIGN(size));
1052 static void vring_free_queue(struct virtio_device *vdev, size_t size,
1053 void *queue, dma_addr_t dma_handle)
1055 if (vring_use_dma_api(vdev)) {
1056 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
1058 free_pages_exact(queue, PAGE_ALIGN(size));
1062 struct virtqueue *vring_create_virtqueue(
1065 unsigned int vring_align,
1066 struct virtio_device *vdev,
1068 bool may_reduce_num,
1070 bool (*notify)(struct virtqueue *),
1071 void (*callback)(struct virtqueue *),
1074 struct virtqueue *vq;
1076 dma_addr_t dma_addr;
1077 size_t queue_size_in_bytes;
1080 /* We assume num is a power of 2. */
1081 if (num & (num - 1)) {
1082 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
1086 /* TODO: allocate each queue chunk individually */
1087 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
1088 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1090 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1093 if (!may_reduce_num)
1101 /* Try to get a single page. You are my only hope! */
1102 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1103 &dma_addr, GFP_KERNEL|__GFP_ZERO);
1108 queue_size_in_bytes = vring_size(num, vring_align);
1109 vring_init(&vring, num, queue, vring_align);
1111 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
1112 notify, callback, name);
1114 vring_free_queue(vdev, queue_size_in_bytes, queue,
1119 to_vvq(vq)->queue_dma_addr = dma_addr;
1120 to_vvq(vq)->queue_size_in_bytes = queue_size_in_bytes;
1121 to_vvq(vq)->we_own_ring = true;
1125 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
1127 struct virtqueue *vring_new_virtqueue(unsigned int index,
1129 unsigned int vring_align,
1130 struct virtio_device *vdev,
1134 bool (*notify)(struct virtqueue *vq),
1135 void (*callback)(struct virtqueue *vq),
1139 vring_init(&vring, num, pages, vring_align);
1140 return __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
1141 notify, callback, name);
1143 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
1145 void vring_del_virtqueue(struct virtqueue *_vq)
1147 struct vring_virtqueue *vq = to_vvq(_vq);
1149 if (vq->we_own_ring) {
1150 vring_free_queue(vq->vq.vdev, vq->queue_size_in_bytes,
1151 vq->vring.desc, vq->queue_dma_addr);
1153 list_del(&_vq->list);
1156 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
1158 /* Manipulates transport-specific feature bits. */
1159 void vring_transport_features(struct virtio_device *vdev)
1163 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
1165 case VIRTIO_RING_F_INDIRECT_DESC:
1167 case VIRTIO_RING_F_EVENT_IDX:
1169 case VIRTIO_F_VERSION_1:
1171 case VIRTIO_F_IOMMU_PLATFORM:
1174 /* We don't understand this bit. */
1175 __virtio_clear_bit(vdev, i);
1179 EXPORT_SYMBOL_GPL(vring_transport_features);
1182 * virtqueue_get_vring_size - return the size of the virtqueue's vring
1183 * @vq: the struct virtqueue containing the vring of interest.
1185 * Returns the size of the vring. This is mainly used for boasting to
1186 * userspace. Unlike other operations, this need not be serialized.
1188 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
1191 struct vring_virtqueue *vq = to_vvq(_vq);
1193 return vq->vring.num;
1195 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
1197 bool virtqueue_is_broken(struct virtqueue *_vq)
1199 struct vring_virtqueue *vq = to_vvq(_vq);
1201 return READ_ONCE(vq->broken);
1203 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
1206 * This should prevent the device from being used, allowing drivers to
1207 * recover. You may need to grab appropriate locks to flush.
1209 void virtio_break_device(struct virtio_device *dev)
1211 struct virtqueue *_vq;
1213 list_for_each_entry(_vq, &dev->vqs, list) {
1214 struct vring_virtqueue *vq = to_vvq(_vq);
1216 /* Pairs with READ_ONCE() in virtqueue_is_broken(). */
1217 WRITE_ONCE(vq->broken, true);
1220 EXPORT_SYMBOL_GPL(virtio_break_device);
1222 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
1224 struct vring_virtqueue *vq = to_vvq(_vq);
1226 BUG_ON(!vq->we_own_ring);
1228 return vq->queue_dma_addr;
1230 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
1232 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
1234 struct vring_virtqueue *vq = to_vvq(_vq);
1236 BUG_ON(!vq->we_own_ring);
1238 return vq->queue_dma_addr +
1239 ((char *)vq->vring.avail - (char *)vq->vring.desc);
1241 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
1243 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
1245 struct vring_virtqueue *vq = to_vvq(_vq);
1247 BUG_ON(!vq->we_own_ring);
1249 return vq->queue_dma_addr +
1250 ((char *)vq->vring.used - (char *)vq->vring.desc);
1252 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
1254 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
1256 return &to_vvq(vq)->vring;
1258 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
1260 MODULE_LICENSE("GPL");