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,
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;
434 * virtqueue_add_sgs - expose buffers to other end
435 * @vq: the struct virtqueue we're talking about.
436 * @sgs: array of terminated scatterlists.
437 * @out_num: the number of scatterlists readable by other side
438 * @in_num: the number of scatterlists which are writable (after readable ones)
439 * @data: the token identifying the buffer.
440 * @gfp: how to do memory allocations (if necessary).
442 * Caller must ensure we don't call this with other virtqueue operations
443 * at the same time (except where noted).
445 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
447 int virtqueue_add_sgs(struct virtqueue *_vq,
448 struct scatterlist *sgs[],
449 unsigned int out_sgs,
454 unsigned int i, total_sg = 0;
456 /* Count them first. */
457 for (i = 0; i < out_sgs + in_sgs; i++) {
458 struct scatterlist *sg;
459 for (sg = sgs[i]; sg; sg = sg_next(sg))
462 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs, data, gfp);
464 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
467 * virtqueue_add_outbuf - expose output buffers to other end
468 * @vq: the struct virtqueue we're talking about.
469 * @sg: scatterlist (must be well-formed and terminated!)
470 * @num: the number of entries in @sg readable by other side
471 * @data: the token identifying the buffer.
472 * @gfp: how to do memory allocations (if necessary).
474 * Caller must ensure we don't call this with other virtqueue operations
475 * at the same time (except where noted).
477 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
479 int virtqueue_add_outbuf(struct virtqueue *vq,
480 struct scatterlist *sg, unsigned int num,
484 return virtqueue_add(vq, &sg, num, 1, 0, data, gfp);
486 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
489 * virtqueue_add_inbuf - expose input buffers to other end
490 * @vq: the struct virtqueue we're talking about.
491 * @sg: scatterlist (must be well-formed and terminated!)
492 * @num: the number of entries in @sg writable by other side
493 * @data: the token identifying the buffer.
494 * @gfp: how to do memory allocations (if necessary).
496 * Caller must ensure we don't call this with other virtqueue operations
497 * at the same time (except where noted).
499 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
501 int virtqueue_add_inbuf(struct virtqueue *vq,
502 struct scatterlist *sg, unsigned int num,
506 return virtqueue_add(vq, &sg, num, 0, 1, data, gfp);
508 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
511 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
512 * @vq: the struct virtqueue
514 * Instead of virtqueue_kick(), you can do:
515 * if (virtqueue_kick_prepare(vq))
516 * virtqueue_notify(vq);
518 * This is sometimes useful because the virtqueue_kick_prepare() needs
519 * to be serialized, but the actual virtqueue_notify() call does not.
521 bool virtqueue_kick_prepare(struct virtqueue *_vq)
523 struct vring_virtqueue *vq = to_vvq(_vq);
528 /* We need to expose available array entries before checking avail
530 virtio_mb(vq->weak_barriers);
532 old = vq->avail_idx_shadow - vq->num_added;
533 new = vq->avail_idx_shadow;
537 if (vq->last_add_time_valid) {
538 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
539 vq->last_add_time)) > 100);
541 vq->last_add_time_valid = false;
545 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev, vring_avail_event(&vq->vring)),
548 needs_kick = !(vq->vring.used->flags & cpu_to_virtio16(_vq->vdev, VRING_USED_F_NO_NOTIFY));
553 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
556 * virtqueue_notify - second half of split virtqueue_kick call.
557 * @vq: the struct virtqueue
559 * This does not need to be serialized.
561 * Returns false if host notify failed or queue is broken, otherwise true.
563 bool virtqueue_notify(struct virtqueue *_vq)
565 struct vring_virtqueue *vq = to_vvq(_vq);
567 if (unlikely(vq->broken))
570 /* Prod other side to tell it about changes. */
571 if (!vq->notify(_vq)) {
577 EXPORT_SYMBOL_GPL(virtqueue_notify);
580 * virtqueue_kick - update after add_buf
581 * @vq: the struct virtqueue
583 * After one or more virtqueue_add_* calls, invoke this to kick
586 * Caller must ensure we don't call this with other virtqueue
587 * operations at the same time (except where noted).
589 * Returns false if kick failed, otherwise true.
591 bool virtqueue_kick(struct virtqueue *vq)
593 if (virtqueue_kick_prepare(vq))
594 return virtqueue_notify(vq);
597 EXPORT_SYMBOL_GPL(virtqueue_kick);
599 static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
602 u16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
604 /* Clear data ptr. */
605 vq->desc_state[head].data = NULL;
607 /* Put back on free list: unmap first-level descriptors and find end */
610 while (vq->vring.desc[i].flags & nextflag) {
611 vring_unmap_one(vq, &vq->vring.desc[i]);
612 i = virtio16_to_cpu(vq->vq.vdev, vq->vring.desc[i].next);
616 vring_unmap_one(vq, &vq->vring.desc[i]);
617 vq->vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev, vq->free_head);
618 vq->free_head = head;
620 /* Plus final descriptor */
623 /* Free the indirect table, if any, now that it's unmapped. */
624 if (vq->desc_state[head].indir_desc) {
625 struct vring_desc *indir_desc = vq->desc_state[head].indir_desc;
626 u32 len = virtio32_to_cpu(vq->vq.vdev, vq->vring.desc[head].len);
628 BUG_ON(!(vq->vring.desc[head].flags &
629 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
630 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
632 for (j = 0; j < len / sizeof(struct vring_desc); j++)
633 vring_unmap_one(vq, &indir_desc[j]);
635 kfree(vq->desc_state[head].indir_desc);
636 vq->desc_state[head].indir_desc = NULL;
640 static inline bool more_used(const struct vring_virtqueue *vq)
642 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev, vq->vring.used->idx);
646 * virtqueue_get_buf - get the next used buffer
647 * @vq: the struct virtqueue we're talking about.
648 * @len: the length written into the buffer
650 * If the driver wrote data into the buffer, @len will be set to the
651 * amount written. This means you don't need to clear the buffer
652 * beforehand to ensure there's no data leakage in the case of short
655 * Caller must ensure we don't call this with other virtqueue
656 * operations at the same time (except where noted).
658 * Returns NULL if there are no used buffers, or the "data" token
659 * handed to virtqueue_add_*().
661 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
663 struct vring_virtqueue *vq = to_vvq(_vq);
670 if (unlikely(vq->broken)) {
675 if (!more_used(vq)) {
676 pr_debug("No more buffers in queue\n");
681 /* Only get used array entries after they have been exposed by host. */
682 virtio_rmb(vq->weak_barriers);
684 last_used = (vq->last_used_idx & (vq->vring.num - 1));
685 i = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].id);
686 *len = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].len);
688 if (unlikely(i >= vq->vring.num)) {
689 BAD_RING(vq, "id %u out of range\n", i);
692 if (unlikely(!vq->desc_state[i].data)) {
693 BAD_RING(vq, "id %u is not a head!\n", i);
697 /* detach_buf clears data, so grab it now. */
698 ret = vq->desc_state[i].data;
701 /* If we expect an interrupt for the next entry, tell host
702 * by writing event index and flush out the write before
703 * the read in the next get_buf call. */
704 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
705 virtio_store_mb(vq->weak_barriers,
706 &vring_used_event(&vq->vring),
707 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
710 vq->last_add_time_valid = false;
716 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
719 * virtqueue_disable_cb - disable callbacks
720 * @vq: the struct virtqueue we're talking about.
722 * Note that this is not necessarily synchronous, hence unreliable and only
723 * useful as an optimization.
725 * Unlike other operations, this need not be serialized.
727 void virtqueue_disable_cb(struct virtqueue *_vq)
729 struct vring_virtqueue *vq = to_vvq(_vq);
731 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
732 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
734 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
738 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
741 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
742 * @vq: the struct virtqueue we're talking about.
744 * This re-enables callbacks; it returns current queue state
745 * in an opaque unsigned value. This value should be later tested by
746 * virtqueue_poll, to detect a possible race between the driver checking for
747 * more work, and enabling callbacks.
749 * Caller must ensure we don't call this with other virtqueue
750 * operations at the same time (except where noted).
752 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
754 struct vring_virtqueue *vq = to_vvq(_vq);
759 /* We optimistically turn back on interrupts, then check if there was
761 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
762 * either clear the flags bit or point the event index at the next
763 * entry. Always do both to keep code simple. */
764 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
765 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 if (unlikely(vq->broken))
791 virtio_mb(vq->weak_barriers);
792 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev, vq->vring.used->idx);
794 EXPORT_SYMBOL_GPL(virtqueue_poll);
797 * virtqueue_enable_cb - restart callbacks after disable_cb.
798 * @vq: the struct virtqueue we're talking about.
800 * This re-enables callbacks; it returns "false" if there are pending
801 * buffers in the queue, to detect a possible race between the driver
802 * checking for more work, and enabling callbacks.
804 * Caller must ensure we don't call this with other virtqueue
805 * operations at the same time (except where noted).
807 bool virtqueue_enable_cb(struct virtqueue *_vq)
809 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
810 return !virtqueue_poll(_vq, last_used_idx);
812 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
815 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
816 * @vq: the struct virtqueue we're talking about.
818 * This re-enables callbacks but hints to the other side to delay
819 * interrupts until most of the available buffers have been processed;
820 * it returns "false" if there are many pending buffers in the queue,
821 * to detect a possible race between the driver checking for more work,
822 * and enabling callbacks.
824 * Caller must ensure we don't call this with other virtqueue
825 * operations at the same time (except where noted).
827 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
829 struct vring_virtqueue *vq = to_vvq(_vq);
834 /* We optimistically turn back on interrupts, then check if there was
836 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
837 * either clear the flags bit or point the event index at the next
838 * entry. Always update the event index to keep code simple. */
839 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
840 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
842 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
844 /* TODO: tune this threshold */
845 bufs = (u16)(vq->avail_idx_shadow - vq->last_used_idx) * 3 / 4;
847 virtio_store_mb(vq->weak_barriers,
848 &vring_used_event(&vq->vring),
849 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
851 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->vring.used->idx) - vq->last_used_idx) > bufs)) {
859 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
862 * virtqueue_detach_unused_buf - detach first unused buffer
863 * @vq: the struct virtqueue we're talking about.
865 * Returns NULL or the "data" token handed to virtqueue_add_*().
866 * This is not valid on an active queue; it is useful only for device
869 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
871 struct vring_virtqueue *vq = to_vvq(_vq);
877 for (i = 0; i < vq->vring.num; i++) {
878 if (!vq->desc_state[i].data)
880 /* detach_buf clears data, so grab it now. */
881 buf = vq->desc_state[i].data;
883 vq->avail_idx_shadow--;
884 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
888 /* That should have freed everything. */
889 BUG_ON(vq->vq.num_free != vq->vring.num);
894 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
896 irqreturn_t vring_interrupt(int irq, void *_vq)
898 struct vring_virtqueue *vq = to_vvq(_vq);
900 if (!more_used(vq)) {
901 pr_debug("virtqueue interrupt with no work for %p\n", vq);
905 if (unlikely(vq->broken))
908 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
910 vq->vq.callback(&vq->vq);
914 EXPORT_SYMBOL_GPL(vring_interrupt);
916 struct virtqueue *__vring_new_virtqueue(unsigned int index,
918 struct virtio_device *vdev,
920 bool (*notify)(struct virtqueue *),
921 void (*callback)(struct virtqueue *),
925 struct vring_virtqueue *vq;
927 vq = kmalloc(sizeof(*vq) + vring.num * sizeof(struct vring_desc_state),
933 vq->vq.callback = callback;
936 vq->vq.num_free = vring.num;
937 vq->vq.index = index;
938 vq->we_own_ring = false;
939 vq->queue_dma_addr = 0;
940 vq->queue_size_in_bytes = 0;
942 vq->weak_barriers = weak_barriers;
944 vq->last_used_idx = 0;
945 vq->avail_flags_shadow = 0;
946 vq->avail_idx_shadow = 0;
948 list_add_tail(&vq->vq.list, &vdev->vqs);
951 vq->last_add_time_valid = false;
954 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
955 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
957 /* No callback? Tell other side not to bother us. */
959 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
961 vq->vring.avail->flags = cpu_to_virtio16(vdev, vq->avail_flags_shadow);
964 /* Put everything in free lists. */
966 for (i = 0; i < vring.num-1; i++)
967 vq->vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
968 memset(vq->desc_state, 0, vring.num * sizeof(struct vring_desc_state));
972 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
974 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
975 dma_addr_t *dma_handle, gfp_t flag)
977 if (vring_use_dma_api(vdev)) {
978 return dma_alloc_coherent(vdev->dev.parent, size,
981 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
983 phys_addr_t phys_addr = virt_to_phys(queue);
984 *dma_handle = (dma_addr_t)phys_addr;
987 * Sanity check: make sure we dind't truncate
988 * the address. The only arches I can find that
989 * have 64-bit phys_addr_t but 32-bit dma_addr_t
990 * are certain non-highmem MIPS and x86
991 * configurations, but these configurations
992 * should never allocate physical pages above 32
993 * bits, so this is fine. Just in case, throw a
994 * warning and abort if we end up with an
995 * unrepresentable address.
997 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
998 free_pages_exact(queue, PAGE_ALIGN(size));
1006 static void vring_free_queue(struct virtio_device *vdev, size_t size,
1007 void *queue, dma_addr_t dma_handle)
1009 if (vring_use_dma_api(vdev)) {
1010 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
1012 free_pages_exact(queue, PAGE_ALIGN(size));
1016 struct virtqueue *vring_create_virtqueue(
1019 unsigned int vring_align,
1020 struct virtio_device *vdev,
1022 bool may_reduce_num,
1023 bool (*notify)(struct virtqueue *),
1024 void (*callback)(struct virtqueue *),
1027 struct virtqueue *vq;
1029 dma_addr_t dma_addr;
1030 size_t queue_size_in_bytes;
1033 /* We assume num is a power of 2. */
1034 if (num & (num - 1)) {
1035 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
1039 /* TODO: allocate each queue chunk individually */
1040 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
1041 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1043 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1046 if (!may_reduce_num)
1054 /* Try to get a single page. You are my only hope! */
1055 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1056 &dma_addr, GFP_KERNEL|__GFP_ZERO);
1061 queue_size_in_bytes = vring_size(num, vring_align);
1062 vring_init(&vring, num, queue, vring_align);
1064 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers,
1065 notify, callback, name);
1067 vring_free_queue(vdev, queue_size_in_bytes, queue,
1072 to_vvq(vq)->queue_dma_addr = dma_addr;
1073 to_vvq(vq)->queue_size_in_bytes = queue_size_in_bytes;
1074 to_vvq(vq)->we_own_ring = true;
1078 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
1080 struct virtqueue *vring_new_virtqueue(unsigned int index,
1082 unsigned int vring_align,
1083 struct virtio_device *vdev,
1086 bool (*notify)(struct virtqueue *vq),
1087 void (*callback)(struct virtqueue *vq),
1091 vring_init(&vring, num, pages, vring_align);
1092 return __vring_new_virtqueue(index, vring, vdev, weak_barriers,
1093 notify, callback, name);
1095 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
1097 void vring_del_virtqueue(struct virtqueue *_vq)
1099 struct vring_virtqueue *vq = to_vvq(_vq);
1101 if (vq->we_own_ring) {
1102 vring_free_queue(vq->vq.vdev, vq->queue_size_in_bytes,
1103 vq->vring.desc, vq->queue_dma_addr);
1105 list_del(&_vq->list);
1108 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
1110 /* Manipulates transport-specific feature bits. */
1111 void vring_transport_features(struct virtio_device *vdev)
1115 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
1117 case VIRTIO_RING_F_INDIRECT_DESC:
1119 case VIRTIO_RING_F_EVENT_IDX:
1121 case VIRTIO_F_VERSION_1:
1123 case VIRTIO_F_IOMMU_PLATFORM:
1126 /* We don't understand this bit. */
1127 __virtio_clear_bit(vdev, i);
1131 EXPORT_SYMBOL_GPL(vring_transport_features);
1134 * virtqueue_get_vring_size - return the size of the virtqueue's vring
1135 * @vq: the struct virtqueue containing the vring of interest.
1137 * Returns the size of the vring. This is mainly used for boasting to
1138 * userspace. Unlike other operations, this need not be serialized.
1140 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
1143 struct vring_virtqueue *vq = to_vvq(_vq);
1145 return vq->vring.num;
1147 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
1149 bool virtqueue_is_broken(struct virtqueue *_vq)
1151 struct vring_virtqueue *vq = to_vvq(_vq);
1153 return READ_ONCE(vq->broken);
1155 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
1158 * This should prevent the device from being used, allowing drivers to
1159 * recover. You may need to grab appropriate locks to flush.
1161 void virtio_break_device(struct virtio_device *dev)
1163 struct virtqueue *_vq;
1165 list_for_each_entry(_vq, &dev->vqs, list) {
1166 struct vring_virtqueue *vq = to_vvq(_vq);
1168 /* Pairs with READ_ONCE() in virtqueue_is_broken(). */
1169 WRITE_ONCE(vq->broken, true);
1172 EXPORT_SYMBOL_GPL(virtio_break_device);
1174 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
1176 struct vring_virtqueue *vq = to_vvq(_vq);
1178 BUG_ON(!vq->we_own_ring);
1180 return vq->queue_dma_addr;
1182 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
1184 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
1186 struct vring_virtqueue *vq = to_vvq(_vq);
1188 BUG_ON(!vq->we_own_ring);
1190 return vq->queue_dma_addr +
1191 ((char *)vq->vring.avail - (char *)vq->vring.desc);
1193 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
1195 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
1197 struct vring_virtqueue *vq = to_vvq(_vq);
1199 BUG_ON(!vq->we_own_ring);
1201 return vq->queue_dma_addr +
1202 ((char *)vq->vring.used - (char *)vq->vring.desc);
1204 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
1206 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
1208 return &to_vvq(vq)->vring;
1210 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
1212 MODULE_LICENSE("GPL");