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 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,
268 struct vring_virtqueue *vq = to_vvq(_vq);
269 struct scatterlist *sg;
270 struct vring_desc *desc;
271 unsigned int i, n, avail, descs_used, uninitialized_var(prev), err_idx;
277 BUG_ON(data == NULL);
279 if (unlikely(vq->broken)) {
286 ktime_t now = ktime_get();
288 /* No kick or get, with .1 second between? Warn. */
289 if (vq->last_add_time_valid)
290 WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
292 vq->last_add_time = now;
293 vq->last_add_time_valid = true;
297 BUG_ON(total_sg > vq->vring.num);
298 BUG_ON(total_sg == 0);
300 head = vq->free_head;
302 /* If the host supports indirect descriptor tables, and we have multiple
303 * buffers, then go indirect. FIXME: tune this threshold */
304 if (vq->indirect && total_sg > 1 && vq->vq.num_free)
305 desc = alloc_indirect(_vq, total_sg, gfp);
310 /* Use a single buffer which doesn't continue */
312 /* Set up rest to use this indirect table. */
317 desc = vq->vring.desc;
319 descs_used = total_sg;
322 if (vq->vq.num_free < descs_used) {
323 pr_debug("Can't add buf len %i - avail = %i\n",
324 descs_used, vq->vq.num_free);
325 /* FIXME: for historical reasons, we force a notify here if
326 * there are outgoing parts to the buffer. Presumably the
327 * host should service the ring ASAP. */
336 for (n = 0; n < out_sgs; n++) {
337 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
338 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
339 if (vring_mapping_error(vq, addr))
342 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
343 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
344 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
346 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
349 for (; n < (out_sgs + in_sgs); n++) {
350 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
351 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
352 if (vring_mapping_error(vq, addr))
355 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
356 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
357 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
359 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
362 /* Last one doesn't continue. */
363 desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
366 /* Now that the indirect table is filled in, map it. */
367 dma_addr_t addr = vring_map_single(
368 vq, desc, total_sg * sizeof(struct vring_desc),
370 if (vring_mapping_error(vq, addr))
373 vq->vring.desc[head].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_INDIRECT);
374 vq->vring.desc[head].addr = cpu_to_virtio64(_vq->vdev, addr);
376 vq->vring.desc[head].len = cpu_to_virtio32(_vq->vdev, total_sg * sizeof(struct vring_desc));
379 /* We're using some buffers from the free list. */
380 vq->vq.num_free -= descs_used;
382 /* Update free pointer */
384 vq->free_head = virtio16_to_cpu(_vq->vdev, vq->vring.desc[head].next);
388 /* Store token and indirect buffer state. */
389 vq->desc_state[head].data = data;
391 vq->desc_state[head].indir_desc = desc;
393 /* Put entry in available array (but don't update avail->idx until they
395 avail = vq->avail_idx_shadow & (vq->vring.num - 1);
396 vq->vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
398 /* Descriptors and available array need to be set before we expose the
399 * new available array entries. */
400 virtio_wmb(vq->weak_barriers);
401 vq->avail_idx_shadow++;
402 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
405 pr_debug("Added buffer head %i to %p\n", head, vq);
408 /* This is very unlikely, but theoretically possible. Kick
410 if (unlikely(vq->num_added == (1 << 16) - 1))
419 for (n = 0; n < total_sg; n++) {
422 vring_unmap_one(vq, &desc[i]);
423 i = vq->vring.desc[i].next;
426 vq->vq.num_free += total_sg;
436 * virtqueue_add_sgs - expose buffers to other end
437 * @vq: the struct virtqueue we're talking about.
438 * @sgs: array of terminated scatterlists.
439 * @out_num: the number of scatterlists readable by other side
440 * @in_num: the number of scatterlists which are writable (after readable ones)
441 * @data: the token identifying the buffer.
442 * @gfp: how to do memory allocations (if necessary).
444 * Caller must ensure we don't call this with other virtqueue operations
445 * at the same time (except where noted).
447 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
449 int virtqueue_add_sgs(struct virtqueue *_vq,
450 struct scatterlist *sgs[],
451 unsigned int out_sgs,
456 unsigned int i, total_sg = 0;
458 /* Count them first. */
459 for (i = 0; i < out_sgs + in_sgs; i++) {
460 struct scatterlist *sg;
461 for (sg = sgs[i]; sg; sg = sg_next(sg))
464 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs, data, gfp);
466 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
469 * virtqueue_add_outbuf - expose output buffers to other end
470 * @vq: the struct virtqueue we're talking about.
471 * @sg: scatterlist (must be well-formed and terminated!)
472 * @num: the number of entries in @sg readable by other side
473 * @data: the token identifying the buffer.
474 * @gfp: how to do memory allocations (if necessary).
476 * Caller must ensure we don't call this with other virtqueue operations
477 * at the same time (except where noted).
479 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
481 int virtqueue_add_outbuf(struct virtqueue *vq,
482 struct scatterlist *sg, unsigned int num,
486 return virtqueue_add(vq, &sg, num, 1, 0, data, gfp);
488 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
491 * virtqueue_add_inbuf - expose input buffers to other end
492 * @vq: the struct virtqueue we're talking about.
493 * @sg: scatterlist (must be well-formed and terminated!)
494 * @num: the number of entries in @sg writable by other side
495 * @data: the token identifying the buffer.
496 * @gfp: how to do memory allocations (if necessary).
498 * Caller must ensure we don't call this with other virtqueue operations
499 * at the same time (except where noted).
501 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
503 int virtqueue_add_inbuf(struct virtqueue *vq,
504 struct scatterlist *sg, unsigned int num,
508 return virtqueue_add(vq, &sg, num, 0, 1, data, gfp);
510 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
513 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
514 * @vq: the struct virtqueue
516 * Instead of virtqueue_kick(), you can do:
517 * if (virtqueue_kick_prepare(vq))
518 * virtqueue_notify(vq);
520 * This is sometimes useful because the virtqueue_kick_prepare() needs
521 * to be serialized, but the actual virtqueue_notify() call does not.
523 bool virtqueue_kick_prepare(struct virtqueue *_vq)
525 struct vring_virtqueue *vq = to_vvq(_vq);
530 /* We need to expose available array entries before checking avail
532 virtio_mb(vq->weak_barriers);
534 old = vq->avail_idx_shadow - vq->num_added;
535 new = vq->avail_idx_shadow;
539 if (vq->last_add_time_valid) {
540 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
541 vq->last_add_time)) > 100);
543 vq->last_add_time_valid = false;
547 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev, vring_avail_event(&vq->vring)),
550 needs_kick = !(vq->vring.used->flags & cpu_to_virtio16(_vq->vdev, VRING_USED_F_NO_NOTIFY));
555 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
558 * virtqueue_notify - second half of split virtqueue_kick call.
559 * @vq: the struct virtqueue
561 * This does not need to be serialized.
563 * Returns false if host notify failed or queue is broken, otherwise true.
565 bool virtqueue_notify(struct virtqueue *_vq)
567 struct vring_virtqueue *vq = to_vvq(_vq);
569 if (unlikely(vq->broken))
572 /* Prod other side to tell it about changes. */
573 if (!vq->notify(_vq)) {
579 EXPORT_SYMBOL_GPL(virtqueue_notify);
582 * virtqueue_kick - update after add_buf
583 * @vq: the struct virtqueue
585 * After one or more virtqueue_add_* calls, invoke this to kick
588 * Caller must ensure we don't call this with other virtqueue
589 * operations at the same time (except where noted).
591 * Returns false if kick failed, otherwise true.
593 bool virtqueue_kick(struct virtqueue *vq)
595 if (virtqueue_kick_prepare(vq))
596 return virtqueue_notify(vq);
599 EXPORT_SYMBOL_GPL(virtqueue_kick);
601 static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
604 u16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
606 /* Clear data ptr. */
607 vq->desc_state[head].data = NULL;
609 /* Put back on free list: unmap first-level descriptors and find end */
612 while (vq->vring.desc[i].flags & nextflag) {
613 vring_unmap_one(vq, &vq->vring.desc[i]);
614 i = virtio16_to_cpu(vq->vq.vdev, vq->vring.desc[i].next);
618 vring_unmap_one(vq, &vq->vring.desc[i]);
619 vq->vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev, vq->free_head);
620 vq->free_head = head;
622 /* Plus final descriptor */
625 /* Free the indirect table, if any, now that it's unmapped. */
626 if (vq->desc_state[head].indir_desc) {
627 struct vring_desc *indir_desc = vq->desc_state[head].indir_desc;
628 u32 len = virtio32_to_cpu(vq->vq.vdev, vq->vring.desc[head].len);
630 BUG_ON(!(vq->vring.desc[head].flags &
631 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
632 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
634 for (j = 0; j < len / sizeof(struct vring_desc); j++)
635 vring_unmap_one(vq, &indir_desc[j]);
637 kfree(vq->desc_state[head].indir_desc);
638 vq->desc_state[head].indir_desc = NULL;
642 static inline bool more_used(const struct vring_virtqueue *vq)
644 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev, vq->vring.used->idx);
648 * virtqueue_get_buf - get the next used buffer
649 * @vq: the struct virtqueue we're talking about.
650 * @len: the length written into the buffer
652 * If the driver wrote data into the buffer, @len will be set to the
653 * amount written. This means you don't need to clear the buffer
654 * beforehand to ensure there's no data leakage in the case of short
657 * Caller must ensure we don't call this with other virtqueue
658 * operations at the same time (except where noted).
660 * Returns NULL if there are no used buffers, or the "data" token
661 * handed to virtqueue_add_*().
663 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
665 struct vring_virtqueue *vq = to_vvq(_vq);
672 if (unlikely(vq->broken)) {
677 if (!more_used(vq)) {
678 pr_debug("No more buffers in queue\n");
683 /* Only get used array entries after they have been exposed by host. */
684 virtio_rmb(vq->weak_barriers);
686 last_used = (vq->last_used_idx & (vq->vring.num - 1));
687 i = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].id);
688 *len = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].len);
690 if (unlikely(i >= vq->vring.num)) {
691 BAD_RING(vq, "id %u out of range\n", i);
694 if (unlikely(!vq->desc_state[i].data)) {
695 BAD_RING(vq, "id %u is not a head!\n", i);
699 /* detach_buf clears data, so grab it now. */
700 ret = vq->desc_state[i].data;
703 /* If we expect an interrupt for the next entry, tell host
704 * by writing event index and flush out the write before
705 * the read in the next get_buf call. */
706 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
707 virtio_store_mb(vq->weak_barriers,
708 &vring_used_event(&vq->vring),
709 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
712 vq->last_add_time_valid = false;
718 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
721 * virtqueue_disable_cb - disable callbacks
722 * @vq: the struct virtqueue we're talking about.
724 * Note that this is not necessarily synchronous, hence unreliable and only
725 * useful as an optimization.
727 * Unlike other operations, this need not be serialized.
729 void virtqueue_disable_cb(struct virtqueue *_vq)
731 struct vring_virtqueue *vq = to_vvq(_vq);
733 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
734 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
735 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
739 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
742 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
743 * @vq: the struct virtqueue we're talking about.
745 * This re-enables callbacks; it returns current queue state
746 * in an opaque unsigned value. This value should be later tested by
747 * virtqueue_poll, to detect a possible race between the driver checking for
748 * more work, and enabling callbacks.
750 * Caller must ensure we don't call this with other virtqueue
751 * operations at the same time (except where noted).
753 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
755 struct vring_virtqueue *vq = to_vvq(_vq);
760 /* We optimistically turn back on interrupts, then check if there was
762 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
763 * either clear the flags bit or point the event index at the next
764 * entry. Always do both to keep code simple. */
765 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
766 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
767 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
769 vring_used_event(&vq->vring) = cpu_to_virtio16(_vq->vdev, last_used_idx = vq->last_used_idx);
771 return last_used_idx;
773 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
776 * virtqueue_poll - query pending used buffers
777 * @vq: the struct virtqueue we're talking about.
778 * @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
780 * Returns "true" if there are pending used buffers in the queue.
782 * This does not need to be serialized.
784 bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
786 struct vring_virtqueue *vq = to_vvq(_vq);
788 virtio_mb(vq->weak_barriers);
789 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev, vq->vring.used->idx);
791 EXPORT_SYMBOL_GPL(virtqueue_poll);
794 * virtqueue_enable_cb - restart callbacks after disable_cb.
795 * @vq: the struct virtqueue we're talking about.
797 * This re-enables callbacks; it returns "false" if there are pending
798 * buffers in the queue, to detect a possible race between the driver
799 * checking for more work, and enabling callbacks.
801 * Caller must ensure we don't call this with other virtqueue
802 * operations at the same time (except where noted).
804 bool virtqueue_enable_cb(struct virtqueue *_vq)
806 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
807 return !virtqueue_poll(_vq, last_used_idx);
809 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
812 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
813 * @vq: the struct virtqueue we're talking about.
815 * This re-enables callbacks but hints to the other side to delay
816 * interrupts until most of the available buffers have been processed;
817 * it returns "false" if there are many pending buffers in the queue,
818 * to detect a possible race between the driver checking for more work,
819 * and enabling callbacks.
821 * Caller must ensure we don't call this with other virtqueue
822 * operations at the same time (except where noted).
824 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
826 struct vring_virtqueue *vq = to_vvq(_vq);
831 /* We optimistically turn back on interrupts, then check if there was
833 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
834 * either clear the flags bit or point the event index at the next
835 * entry. Always do both to keep code simple. */
836 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
837 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
838 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
840 /* TODO: tune this threshold */
841 bufs = (u16)(vq->avail_idx_shadow - vq->last_used_idx) * 3 / 4;
843 virtio_store_mb(vq->weak_barriers,
844 &vring_used_event(&vq->vring),
845 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
847 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->vring.used->idx) - vq->last_used_idx) > bufs)) {
855 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
858 * virtqueue_detach_unused_buf - detach first unused buffer
859 * @vq: the struct virtqueue we're talking about.
861 * Returns NULL or the "data" token handed to virtqueue_add_*().
862 * This is not valid on an active queue; it is useful only for device
865 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
867 struct vring_virtqueue *vq = to_vvq(_vq);
873 for (i = 0; i < vq->vring.num; i++) {
874 if (!vq->desc_state[i].data)
876 /* detach_buf clears data, so grab it now. */
877 buf = vq->desc_state[i].data;
879 vq->avail_idx_shadow--;
880 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
884 /* That should have freed everything. */
885 BUG_ON(vq->vq.num_free != vq->vring.num);
890 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
892 irqreturn_t vring_interrupt(int irq, void *_vq)
894 struct vring_virtqueue *vq = to_vvq(_vq);
896 if (!more_used(vq)) {
897 pr_debug("virtqueue interrupt with no work for %p\n", vq);
901 if (unlikely(vq->broken))
904 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
906 vq->vq.callback(&vq->vq);
910 EXPORT_SYMBOL_GPL(vring_interrupt);
912 struct virtqueue *__vring_new_virtqueue(unsigned int index,
914 struct virtio_device *vdev,
916 bool (*notify)(struct virtqueue *),
917 void (*callback)(struct virtqueue *),
921 struct vring_virtqueue *vq;
923 vq = kmalloc(sizeof(*vq) + vring.num * sizeof(struct vring_desc_state),
929 vq->vq.callback = callback;
932 vq->vq.num_free = vring.num;
933 vq->vq.index = index;
934 vq->we_own_ring = false;
935 vq->queue_dma_addr = 0;
936 vq->queue_size_in_bytes = 0;
938 vq->weak_barriers = weak_barriers;
940 vq->last_used_idx = 0;
941 vq->avail_flags_shadow = 0;
942 vq->avail_idx_shadow = 0;
944 list_add_tail(&vq->vq.list, &vdev->vqs);
947 vq->last_add_time_valid = false;
950 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
951 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
953 /* No callback? Tell other side not to bother us. */
955 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
956 vq->vring.avail->flags = cpu_to_virtio16(vdev, vq->avail_flags_shadow);
959 /* Put everything in free lists. */
961 for (i = 0; i < vring.num-1; i++)
962 vq->vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
963 memset(vq->desc_state, 0, vring.num * sizeof(struct vring_desc_state));
967 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
969 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
970 dma_addr_t *dma_handle, gfp_t flag)
972 if (vring_use_dma_api(vdev)) {
973 return dma_alloc_coherent(vdev->dev.parent, size,
976 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
978 phys_addr_t phys_addr = virt_to_phys(queue);
979 *dma_handle = (dma_addr_t)phys_addr;
982 * Sanity check: make sure we dind't truncate
983 * the address. The only arches I can find that
984 * have 64-bit phys_addr_t but 32-bit dma_addr_t
985 * are certain non-highmem MIPS and x86
986 * configurations, but these configurations
987 * should never allocate physical pages above 32
988 * bits, so this is fine. Just in case, throw a
989 * warning and abort if we end up with an
990 * unrepresentable address.
992 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
993 free_pages_exact(queue, PAGE_ALIGN(size));
1001 static void vring_free_queue(struct virtio_device *vdev, size_t size,
1002 void *queue, dma_addr_t dma_handle)
1004 if (vring_use_dma_api(vdev)) {
1005 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
1007 free_pages_exact(queue, PAGE_ALIGN(size));
1011 struct virtqueue *vring_create_virtqueue(
1014 unsigned int vring_align,
1015 struct virtio_device *vdev,
1017 bool may_reduce_num,
1018 bool (*notify)(struct virtqueue *),
1019 void (*callback)(struct virtqueue *),
1022 struct virtqueue *vq;
1024 dma_addr_t dma_addr;
1025 size_t queue_size_in_bytes;
1028 /* We assume num is a power of 2. */
1029 if (num & (num - 1)) {
1030 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
1034 /* TODO: allocate each queue chunk individually */
1035 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
1036 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1038 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1047 /* Try to get a single page. You are my only hope! */
1048 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1049 &dma_addr, GFP_KERNEL|__GFP_ZERO);
1054 queue_size_in_bytes = vring_size(num, vring_align);
1055 vring_init(&vring, num, queue, vring_align);
1057 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers,
1058 notify, callback, name);
1060 vring_free_queue(vdev, queue_size_in_bytes, queue,
1065 to_vvq(vq)->queue_dma_addr = dma_addr;
1066 to_vvq(vq)->queue_size_in_bytes = queue_size_in_bytes;
1067 to_vvq(vq)->we_own_ring = true;
1071 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
1073 struct virtqueue *vring_new_virtqueue(unsigned int index,
1075 unsigned int vring_align,
1076 struct virtio_device *vdev,
1079 bool (*notify)(struct virtqueue *vq),
1080 void (*callback)(struct virtqueue *vq),
1084 vring_init(&vring, num, pages, vring_align);
1085 return __vring_new_virtqueue(index, vring, vdev, weak_barriers,
1086 notify, callback, name);
1088 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
1090 void vring_del_virtqueue(struct virtqueue *_vq)
1092 struct vring_virtqueue *vq = to_vvq(_vq);
1094 if (vq->we_own_ring) {
1095 vring_free_queue(vq->vq.vdev, vq->queue_size_in_bytes,
1096 vq->vring.desc, vq->queue_dma_addr);
1098 list_del(&_vq->list);
1101 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
1103 /* Manipulates transport-specific feature bits. */
1104 void vring_transport_features(struct virtio_device *vdev)
1108 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
1110 case VIRTIO_RING_F_INDIRECT_DESC:
1112 case VIRTIO_RING_F_EVENT_IDX:
1114 case VIRTIO_F_VERSION_1:
1116 case VIRTIO_F_IOMMU_PLATFORM:
1119 /* We don't understand this bit. */
1120 __virtio_clear_bit(vdev, i);
1124 EXPORT_SYMBOL_GPL(vring_transport_features);
1127 * virtqueue_get_vring_size - return the size of the virtqueue's vring
1128 * @vq: the struct virtqueue containing the vring of interest.
1130 * Returns the size of the vring. This is mainly used for boasting to
1131 * userspace. Unlike other operations, this need not be serialized.
1133 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
1136 struct vring_virtqueue *vq = to_vvq(_vq);
1138 return vq->vring.num;
1140 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
1142 bool virtqueue_is_broken(struct virtqueue *_vq)
1144 struct vring_virtqueue *vq = to_vvq(_vq);
1148 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
1151 * This should prevent the device from being used, allowing drivers to
1152 * recover. You may need to grab appropriate locks to flush.
1154 void virtio_break_device(struct virtio_device *dev)
1156 struct virtqueue *_vq;
1158 list_for_each_entry(_vq, &dev->vqs, list) {
1159 struct vring_virtqueue *vq = to_vvq(_vq);
1163 EXPORT_SYMBOL_GPL(virtio_break_device);
1165 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
1167 struct vring_virtqueue *vq = to_vvq(_vq);
1169 BUG_ON(!vq->we_own_ring);
1171 return vq->queue_dma_addr;
1173 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
1175 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
1177 struct vring_virtqueue *vq = to_vvq(_vq);
1179 BUG_ON(!vq->we_own_ring);
1181 return vq->queue_dma_addr +
1182 ((char *)vq->vring.avail - (char *)vq->vring.desc);
1184 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
1186 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
1188 struct vring_virtqueue *vq = to_vvq(_vq);
1190 BUG_ON(!vq->we_own_ring);
1192 return vq->queue_dma_addr +
1193 ((char *)vq->vring.used - (char *)vq->vring.desc);
1195 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
1197 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
1199 return &to_vvq(vq)->vring;
1201 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
1203 MODULE_LICENSE("GPL");