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 * The interaction between virtio and a possible IOMMU is a mess.
122 * On most systems with virtio, physical addresses match bus addresses,
123 * and it doesn't particularly matter whether we use the DMA API.
125 * On some systems, including Xen and any system with a physical device
126 * that speaks virtio behind a physical IOMMU, we must use the DMA API
127 * for virtio DMA to work at all.
129 * On other systems, including SPARC and PPC64, virtio-pci devices are
130 * enumerated as though they are behind an IOMMU, but the virtio host
131 * ignores the IOMMU, so we must either pretend that the IOMMU isn't
132 * there or somehow map everything as the identity.
134 * For the time being, we preserve historic behavior and bypass the DMA
138 static bool vring_use_dma_api(struct virtio_device *vdev)
141 * In theory, it's possible to have a buggy QEMU-supposed
142 * emulated Q35 IOMMU and Xen enabled at the same time. On
143 * such a configuration, virtio has never worked and will
144 * not work without an even larger kludge. Instead, enable
145 * the DMA API if we're a Xen guest, which at least allows
146 * all of the sensible Xen configurations to work correctly.
155 * The DMA ops on various arches are rather gnarly right now, and
156 * making all of the arch DMA ops work on the vring device itself
157 * is a mess. For now, we use the parent device for DMA ops.
159 struct device *vring_dma_dev(const struct vring_virtqueue *vq)
161 return vq->vq.vdev->dev.parent;
164 /* Map one sg entry. */
165 static dma_addr_t vring_map_one_sg(const struct vring_virtqueue *vq,
166 struct scatterlist *sg,
167 enum dma_data_direction direction)
169 if (!vring_use_dma_api(vq->vq.vdev))
170 return (dma_addr_t)sg_phys(sg);
173 * We can't use dma_map_sg, because we don't use scatterlists in
174 * the way it expects (we don't guarantee that the scatterlist
175 * will exist for the lifetime of the mapping).
177 return dma_map_page(vring_dma_dev(vq),
178 sg_page(sg), sg->offset, sg->length,
182 static dma_addr_t vring_map_single(const struct vring_virtqueue *vq,
183 void *cpu_addr, size_t size,
184 enum dma_data_direction direction)
186 if (!vring_use_dma_api(vq->vq.vdev))
187 return (dma_addr_t)virt_to_phys(cpu_addr);
189 return dma_map_single(vring_dma_dev(vq),
190 cpu_addr, size, direction);
193 static void vring_unmap_one(const struct vring_virtqueue *vq,
194 struct vring_desc *desc)
198 if (!vring_use_dma_api(vq->vq.vdev))
201 flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
203 if (flags & VRING_DESC_F_INDIRECT) {
204 dma_unmap_single(vring_dma_dev(vq),
205 virtio64_to_cpu(vq->vq.vdev, desc->addr),
206 virtio32_to_cpu(vq->vq.vdev, desc->len),
207 (flags & VRING_DESC_F_WRITE) ?
208 DMA_FROM_DEVICE : DMA_TO_DEVICE);
210 dma_unmap_page(vring_dma_dev(vq),
211 virtio64_to_cpu(vq->vq.vdev, desc->addr),
212 virtio32_to_cpu(vq->vq.vdev, desc->len),
213 (flags & VRING_DESC_F_WRITE) ?
214 DMA_FROM_DEVICE : DMA_TO_DEVICE);
218 static int vring_mapping_error(const struct vring_virtqueue *vq,
221 if (!vring_use_dma_api(vq->vq.vdev))
224 return dma_mapping_error(vring_dma_dev(vq), addr);
227 static struct vring_desc *alloc_indirect(struct virtqueue *_vq,
228 unsigned int total_sg, gfp_t gfp)
230 struct vring_desc *desc;
234 * We require lowmem mappings for the descriptors because
235 * otherwise virt_to_phys will give us bogus addresses in the
238 gfp &= ~__GFP_HIGHMEM;
240 desc = kmalloc(total_sg * sizeof(struct vring_desc), gfp);
244 for (i = 0; i < total_sg; i++)
245 desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
249 static inline int virtqueue_add(struct virtqueue *_vq,
250 struct scatterlist *sgs[],
251 unsigned int total_sg,
252 unsigned int out_sgs,
257 struct vring_virtqueue *vq = to_vvq(_vq);
258 struct scatterlist *sg;
259 struct vring_desc *desc;
260 unsigned int i, n, avail, descs_used, uninitialized_var(prev), err_idx;
266 BUG_ON(data == NULL);
268 if (unlikely(vq->broken)) {
275 ktime_t now = ktime_get();
277 /* No kick or get, with .1 second between? Warn. */
278 if (vq->last_add_time_valid)
279 WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
281 vq->last_add_time = now;
282 vq->last_add_time_valid = true;
286 BUG_ON(total_sg > vq->vring.num);
287 BUG_ON(total_sg == 0);
289 head = vq->free_head;
291 /* If the host supports indirect descriptor tables, and we have multiple
292 * buffers, then go indirect. FIXME: tune this threshold */
293 if (vq->indirect && total_sg > 1 && vq->vq.num_free)
294 desc = alloc_indirect(_vq, total_sg, gfp);
299 /* Use a single buffer which doesn't continue */
301 /* Set up rest to use this indirect table. */
306 desc = vq->vring.desc;
308 descs_used = total_sg;
311 if (vq->vq.num_free < descs_used) {
312 pr_debug("Can't add buf len %i - avail = %i\n",
313 descs_used, vq->vq.num_free);
314 /* FIXME: for historical reasons, we force a notify here if
315 * there are outgoing parts to the buffer. Presumably the
316 * host should service the ring ASAP. */
323 for (n = 0; n < out_sgs; n++) {
324 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
325 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
326 if (vring_mapping_error(vq, addr))
329 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
330 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
331 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
333 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
336 for (; n < (out_sgs + in_sgs); n++) {
337 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
338 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
339 if (vring_mapping_error(vq, addr))
342 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
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 /* Last one doesn't continue. */
350 desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
353 /* Now that the indirect table is filled in, map it. */
354 dma_addr_t addr = vring_map_single(
355 vq, desc, total_sg * sizeof(struct vring_desc),
357 if (vring_mapping_error(vq, addr))
360 vq->vring.desc[head].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_INDIRECT);
361 vq->vring.desc[head].addr = cpu_to_virtio64(_vq->vdev, addr);
363 vq->vring.desc[head].len = cpu_to_virtio32(_vq->vdev, total_sg * sizeof(struct vring_desc));
366 /* We're using some buffers from the free list. */
367 vq->vq.num_free -= descs_used;
369 /* Update free pointer */
371 vq->free_head = virtio16_to_cpu(_vq->vdev, vq->vring.desc[head].next);
375 /* Store token and indirect buffer state. */
376 vq->desc_state[head].data = data;
378 vq->desc_state[head].indir_desc = desc;
380 /* Put entry in available array (but don't update avail->idx until they
382 avail = vq->avail_idx_shadow & (vq->vring.num - 1);
383 vq->vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
385 /* Descriptors and available array need to be set before we expose the
386 * new available array entries. */
387 virtio_wmb(vq->weak_barriers);
388 vq->avail_idx_shadow++;
389 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
392 pr_debug("Added buffer head %i to %p\n", head, vq);
395 /* This is very unlikely, but theoretically possible. Kick
397 if (unlikely(vq->num_added == (1 << 16) - 1))
406 for (n = 0; n < total_sg; n++) {
409 vring_unmap_one(vq, &desc[i]);
410 i = vq->vring.desc[i].next;
413 vq->vq.num_free += total_sg;
422 * virtqueue_add_sgs - expose buffers to other end
423 * @vq: the struct virtqueue we're talking about.
424 * @sgs: array of terminated scatterlists.
425 * @out_num: the number of scatterlists readable by other side
426 * @in_num: the number of scatterlists which are writable (after readable ones)
427 * @data: the token identifying the buffer.
428 * @gfp: how to do memory allocations (if necessary).
430 * Caller must ensure we don't call this with other virtqueue operations
431 * at the same time (except where noted).
433 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
435 int virtqueue_add_sgs(struct virtqueue *_vq,
436 struct scatterlist *sgs[],
437 unsigned int out_sgs,
442 unsigned int i, total_sg = 0;
444 /* Count them first. */
445 for (i = 0; i < out_sgs + in_sgs; i++) {
446 struct scatterlist *sg;
447 for (sg = sgs[i]; sg; sg = sg_next(sg))
450 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs, data, gfp);
452 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
455 * virtqueue_add_outbuf - expose output buffers to other end
456 * @vq: the struct virtqueue we're talking about.
457 * @sg: scatterlist (must be well-formed and terminated!)
458 * @num: the number of entries in @sg readable by other side
459 * @data: the token identifying the buffer.
460 * @gfp: how to do memory allocations (if necessary).
462 * Caller must ensure we don't call this with other virtqueue operations
463 * at the same time (except where noted).
465 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
467 int virtqueue_add_outbuf(struct virtqueue *vq,
468 struct scatterlist *sg, unsigned int num,
472 return virtqueue_add(vq, &sg, num, 1, 0, data, gfp);
474 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
477 * virtqueue_add_inbuf - expose input buffers to other end
478 * @vq: the struct virtqueue we're talking about.
479 * @sg: scatterlist (must be well-formed and terminated!)
480 * @num: the number of entries in @sg writable by other side
481 * @data: the token identifying the buffer.
482 * @gfp: how to do memory allocations (if necessary).
484 * Caller must ensure we don't call this with other virtqueue operations
485 * at the same time (except where noted).
487 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
489 int virtqueue_add_inbuf(struct virtqueue *vq,
490 struct scatterlist *sg, unsigned int num,
494 return virtqueue_add(vq, &sg, num, 0, 1, data, gfp);
496 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
499 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
500 * @vq: the struct virtqueue
502 * Instead of virtqueue_kick(), you can do:
503 * if (virtqueue_kick_prepare(vq))
504 * virtqueue_notify(vq);
506 * This is sometimes useful because the virtqueue_kick_prepare() needs
507 * to be serialized, but the actual virtqueue_notify() call does not.
509 bool virtqueue_kick_prepare(struct virtqueue *_vq)
511 struct vring_virtqueue *vq = to_vvq(_vq);
516 /* We need to expose available array entries before checking avail
518 virtio_mb(vq->weak_barriers);
520 old = vq->avail_idx_shadow - vq->num_added;
521 new = vq->avail_idx_shadow;
525 if (vq->last_add_time_valid) {
526 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
527 vq->last_add_time)) > 100);
529 vq->last_add_time_valid = false;
533 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev, vring_avail_event(&vq->vring)),
536 needs_kick = !(vq->vring.used->flags & cpu_to_virtio16(_vq->vdev, VRING_USED_F_NO_NOTIFY));
541 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
544 * virtqueue_notify - second half of split virtqueue_kick call.
545 * @vq: the struct virtqueue
547 * This does not need to be serialized.
549 * Returns false if host notify failed or queue is broken, otherwise true.
551 bool virtqueue_notify(struct virtqueue *_vq)
553 struct vring_virtqueue *vq = to_vvq(_vq);
555 if (unlikely(vq->broken))
558 /* Prod other side to tell it about changes. */
559 if (!vq->notify(_vq)) {
565 EXPORT_SYMBOL_GPL(virtqueue_notify);
568 * virtqueue_kick - update after add_buf
569 * @vq: the struct virtqueue
571 * After one or more virtqueue_add_* calls, invoke this to kick
574 * Caller must ensure we don't call this with other virtqueue
575 * operations at the same time (except where noted).
577 * Returns false if kick failed, otherwise true.
579 bool virtqueue_kick(struct virtqueue *vq)
581 if (virtqueue_kick_prepare(vq))
582 return virtqueue_notify(vq);
585 EXPORT_SYMBOL_GPL(virtqueue_kick);
587 static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
590 u16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
592 /* Clear data ptr. */
593 vq->desc_state[head].data = NULL;
595 /* Put back on free list: unmap first-level descriptors and find end */
598 while (vq->vring.desc[i].flags & nextflag) {
599 vring_unmap_one(vq, &vq->vring.desc[i]);
600 i = virtio16_to_cpu(vq->vq.vdev, vq->vring.desc[i].next);
604 vring_unmap_one(vq, &vq->vring.desc[i]);
605 vq->vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev, vq->free_head);
606 vq->free_head = head;
608 /* Plus final descriptor */
611 /* Free the indirect table, if any, now that it's unmapped. */
612 if (vq->desc_state[head].indir_desc) {
613 struct vring_desc *indir_desc = vq->desc_state[head].indir_desc;
614 u32 len = virtio32_to_cpu(vq->vq.vdev, vq->vring.desc[head].len);
616 BUG_ON(!(vq->vring.desc[head].flags &
617 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
618 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
620 for (j = 0; j < len / sizeof(struct vring_desc); j++)
621 vring_unmap_one(vq, &indir_desc[j]);
623 kfree(vq->desc_state[head].indir_desc);
624 vq->desc_state[head].indir_desc = NULL;
628 static inline bool more_used(const struct vring_virtqueue *vq)
630 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev, vq->vring.used->idx);
634 * virtqueue_get_buf - get the next used buffer
635 * @vq: the struct virtqueue we're talking about.
636 * @len: the length written into the buffer
638 * If the driver wrote data into the buffer, @len will be set to the
639 * amount written. This means you don't need to clear the buffer
640 * beforehand to ensure there's no data leakage in the case of short
643 * Caller must ensure we don't call this with other virtqueue
644 * operations at the same time (except where noted).
646 * Returns NULL if there are no used buffers, or the "data" token
647 * handed to virtqueue_add_*().
649 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
651 struct vring_virtqueue *vq = to_vvq(_vq);
658 if (unlikely(vq->broken)) {
663 if (!more_used(vq)) {
664 pr_debug("No more buffers in queue\n");
669 /* Only get used array entries after they have been exposed by host. */
670 virtio_rmb(vq->weak_barriers);
672 last_used = (vq->last_used_idx & (vq->vring.num - 1));
673 i = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].id);
674 *len = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].len);
676 if (unlikely(i >= vq->vring.num)) {
677 BAD_RING(vq, "id %u out of range\n", i);
680 if (unlikely(!vq->desc_state[i].data)) {
681 BAD_RING(vq, "id %u is not a head!\n", i);
685 /* detach_buf clears data, so grab it now. */
686 ret = vq->desc_state[i].data;
689 /* If we expect an interrupt for the next entry, tell host
690 * by writing event index and flush out the write before
691 * the read in the next get_buf call. */
692 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
693 virtio_store_mb(vq->weak_barriers,
694 &vring_used_event(&vq->vring),
695 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
698 vq->last_add_time_valid = false;
704 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
707 * virtqueue_disable_cb - disable callbacks
708 * @vq: the struct virtqueue we're talking about.
710 * Note that this is not necessarily synchronous, hence unreliable and only
711 * useful as an optimization.
713 * Unlike other operations, this need not be serialized.
715 void virtqueue_disable_cb(struct virtqueue *_vq)
717 struct vring_virtqueue *vq = to_vvq(_vq);
719 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
720 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
721 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
725 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
728 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
729 * @vq: the struct virtqueue we're talking about.
731 * This re-enables callbacks; it returns current queue state
732 * in an opaque unsigned value. This value should be later tested by
733 * virtqueue_poll, to detect a possible race between the driver checking for
734 * more work, and enabling callbacks.
736 * Caller must ensure we don't call this with other virtqueue
737 * operations at the same time (except where noted).
739 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
741 struct vring_virtqueue *vq = to_vvq(_vq);
746 /* We optimistically turn back on interrupts, then check if there was
748 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
749 * either clear the flags bit or point the event index at the next
750 * entry. Always do both to keep code simple. */
751 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
752 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
753 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
755 vring_used_event(&vq->vring) = cpu_to_virtio16(_vq->vdev, last_used_idx = vq->last_used_idx);
757 return last_used_idx;
759 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
762 * virtqueue_poll - query pending used buffers
763 * @vq: the struct virtqueue we're talking about.
764 * @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
766 * Returns "true" if there are pending used buffers in the queue.
768 * This does not need to be serialized.
770 bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
772 struct vring_virtqueue *vq = to_vvq(_vq);
774 virtio_mb(vq->weak_barriers);
775 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev, vq->vring.used->idx);
777 EXPORT_SYMBOL_GPL(virtqueue_poll);
780 * virtqueue_enable_cb - restart callbacks after disable_cb.
781 * @vq: the struct virtqueue we're talking about.
783 * This re-enables callbacks; it returns "false" if there are pending
784 * buffers in the queue, to detect a possible race between the driver
785 * checking for more work, and enabling callbacks.
787 * Caller must ensure we don't call this with other virtqueue
788 * operations at the same time (except where noted).
790 bool virtqueue_enable_cb(struct virtqueue *_vq)
792 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
793 return !virtqueue_poll(_vq, last_used_idx);
795 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
798 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
799 * @vq: the struct virtqueue we're talking about.
801 * This re-enables callbacks but hints to the other side to delay
802 * interrupts until most of the available buffers have been processed;
803 * it returns "false" if there are many pending buffers in the queue,
804 * to detect a possible race between the driver checking for more work,
805 * and enabling callbacks.
807 * Caller must ensure we don't call this with other virtqueue
808 * operations at the same time (except where noted).
810 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
812 struct vring_virtqueue *vq = to_vvq(_vq);
817 /* We optimistically turn back on interrupts, then check if there was
819 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
820 * either clear the flags bit or point the event index at the next
821 * entry. Always do both to keep code simple. */
822 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
823 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
824 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
826 /* TODO: tune this threshold */
827 bufs = (u16)(vq->avail_idx_shadow - vq->last_used_idx) * 3 / 4;
829 virtio_store_mb(vq->weak_barriers,
830 &vring_used_event(&vq->vring),
831 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
833 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->vring.used->idx) - vq->last_used_idx) > bufs)) {
841 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
844 * virtqueue_detach_unused_buf - detach first unused buffer
845 * @vq: the struct virtqueue we're talking about.
847 * Returns NULL or the "data" token handed to virtqueue_add_*().
848 * This is not valid on an active queue; it is useful only for device
851 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
853 struct vring_virtqueue *vq = to_vvq(_vq);
859 for (i = 0; i < vq->vring.num; i++) {
860 if (!vq->desc_state[i].data)
862 /* detach_buf clears data, so grab it now. */
863 buf = vq->desc_state[i].data;
865 vq->avail_idx_shadow--;
866 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
870 /* That should have freed everything. */
871 BUG_ON(vq->vq.num_free != vq->vring.num);
876 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
878 irqreturn_t vring_interrupt(int irq, void *_vq)
880 struct vring_virtqueue *vq = to_vvq(_vq);
882 if (!more_used(vq)) {
883 pr_debug("virtqueue interrupt with no work for %p\n", vq);
887 if (unlikely(vq->broken))
890 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
892 vq->vq.callback(&vq->vq);
896 EXPORT_SYMBOL_GPL(vring_interrupt);
898 struct virtqueue *__vring_new_virtqueue(unsigned int index,
900 struct virtio_device *vdev,
902 bool (*notify)(struct virtqueue *),
903 void (*callback)(struct virtqueue *),
907 struct vring_virtqueue *vq;
909 vq = kmalloc(sizeof(*vq) + vring.num * sizeof(struct vring_desc_state),
915 vq->vq.callback = callback;
918 vq->vq.num_free = vring.num;
919 vq->vq.index = index;
920 vq->we_own_ring = false;
921 vq->queue_dma_addr = 0;
922 vq->queue_size_in_bytes = 0;
924 vq->weak_barriers = weak_barriers;
926 vq->last_used_idx = 0;
927 vq->avail_flags_shadow = 0;
928 vq->avail_idx_shadow = 0;
930 list_add_tail(&vq->vq.list, &vdev->vqs);
933 vq->last_add_time_valid = false;
936 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
937 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
939 /* No callback? Tell other side not to bother us. */
941 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
942 vq->vring.avail->flags = cpu_to_virtio16(vdev, vq->avail_flags_shadow);
945 /* Put everything in free lists. */
947 for (i = 0; i < vring.num-1; i++)
948 vq->vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
949 memset(vq->desc_state, 0, vring.num * sizeof(struct vring_desc_state));
953 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
955 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
956 dma_addr_t *dma_handle, gfp_t flag)
958 if (vring_use_dma_api(vdev)) {
959 return dma_alloc_coherent(vdev->dev.parent, size,
962 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
964 phys_addr_t phys_addr = virt_to_phys(queue);
965 *dma_handle = (dma_addr_t)phys_addr;
968 * Sanity check: make sure we dind't truncate
969 * the address. The only arches I can find that
970 * have 64-bit phys_addr_t but 32-bit dma_addr_t
971 * are certain non-highmem MIPS and x86
972 * configurations, but these configurations
973 * should never allocate physical pages above 32
974 * bits, so this is fine. Just in case, throw a
975 * warning and abort if we end up with an
976 * unrepresentable address.
978 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
979 free_pages_exact(queue, PAGE_ALIGN(size));
987 static void vring_free_queue(struct virtio_device *vdev, size_t size,
988 void *queue, dma_addr_t dma_handle)
990 if (vring_use_dma_api(vdev)) {
991 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
993 free_pages_exact(queue, PAGE_ALIGN(size));
997 struct virtqueue *vring_create_virtqueue(
1000 unsigned int vring_align,
1001 struct virtio_device *vdev,
1003 bool may_reduce_num,
1004 bool (*notify)(struct virtqueue *),
1005 void (*callback)(struct virtqueue *),
1008 struct virtqueue *vq;
1010 dma_addr_t dma_addr;
1011 size_t queue_size_in_bytes;
1014 /* We assume num is a power of 2. */
1015 if (num & (num - 1)) {
1016 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
1020 /* TODO: allocate each queue chunk individually */
1021 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
1022 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1024 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1033 /* Try to get a single page. You are my only hope! */
1034 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1035 &dma_addr, GFP_KERNEL|__GFP_ZERO);
1040 queue_size_in_bytes = vring_size(num, vring_align);
1041 vring_init(&vring, num, queue, vring_align);
1043 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers,
1044 notify, callback, name);
1046 vring_free_queue(vdev, queue_size_in_bytes, queue,
1051 to_vvq(vq)->queue_dma_addr = dma_addr;
1052 to_vvq(vq)->queue_size_in_bytes = queue_size_in_bytes;
1053 to_vvq(vq)->we_own_ring = true;
1057 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
1059 struct virtqueue *vring_new_virtqueue(unsigned int index,
1061 unsigned int vring_align,
1062 struct virtio_device *vdev,
1065 bool (*notify)(struct virtqueue *vq),
1066 void (*callback)(struct virtqueue *vq),
1070 vring_init(&vring, num, pages, vring_align);
1071 return __vring_new_virtqueue(index, vring, vdev, weak_barriers,
1072 notify, callback, name);
1074 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
1076 void vring_del_virtqueue(struct virtqueue *_vq)
1078 struct vring_virtqueue *vq = to_vvq(_vq);
1080 if (vq->we_own_ring) {
1081 vring_free_queue(vq->vq.vdev, vq->queue_size_in_bytes,
1082 vq->vring.desc, vq->queue_dma_addr);
1084 list_del(&_vq->list);
1087 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
1089 /* Manipulates transport-specific feature bits. */
1090 void vring_transport_features(struct virtio_device *vdev)
1094 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
1096 case VIRTIO_RING_F_INDIRECT_DESC:
1098 case VIRTIO_RING_F_EVENT_IDX:
1100 case VIRTIO_F_VERSION_1:
1103 /* We don't understand this bit. */
1104 __virtio_clear_bit(vdev, i);
1108 EXPORT_SYMBOL_GPL(vring_transport_features);
1111 * virtqueue_get_vring_size - return the size of the virtqueue's vring
1112 * @vq: the struct virtqueue containing the vring of interest.
1114 * Returns the size of the vring. This is mainly used for boasting to
1115 * userspace. Unlike other operations, this need not be serialized.
1117 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
1120 struct vring_virtqueue *vq = to_vvq(_vq);
1122 return vq->vring.num;
1124 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
1126 bool virtqueue_is_broken(struct virtqueue *_vq)
1128 struct vring_virtqueue *vq = to_vvq(_vq);
1132 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
1135 * This should prevent the device from being used, allowing drivers to
1136 * recover. You may need to grab appropriate locks to flush.
1138 void virtio_break_device(struct virtio_device *dev)
1140 struct virtqueue *_vq;
1142 list_for_each_entry(_vq, &dev->vqs, list) {
1143 struct vring_virtqueue *vq = to_vvq(_vq);
1147 EXPORT_SYMBOL_GPL(virtio_break_device);
1149 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
1151 struct vring_virtqueue *vq = to_vvq(_vq);
1153 BUG_ON(!vq->we_own_ring);
1155 return vq->queue_dma_addr;
1157 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
1159 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
1161 struct vring_virtqueue *vq = to_vvq(_vq);
1163 BUG_ON(!vq->we_own_ring);
1165 return vq->queue_dma_addr +
1166 ((char *)vq->vring.avail - (char *)vq->vring.desc);
1168 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
1170 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
1172 struct vring_virtqueue *vq = to_vvq(_vq);
1174 BUG_ON(!vq->we_own_ring);
1176 return vq->queue_dma_addr +
1177 ((char *)vq->vring.used - (char *)vq->vring.desc);
1179 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
1181 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
1183 return &to_vvq(vq)->vring;
1185 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
1187 MODULE_LICENSE("GPL");