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/dma-mapping.h>
30 /* For development, we want to crash whenever the ring is screwed. */
31 #define BAD_RING(_vq, fmt, args...) \
33 dev_err(&(_vq)->vq.vdev->dev, \
34 "%s:"fmt, (_vq)->vq.name, ##args); \
37 /* Caller is supposed to guarantee no reentry. */
38 #define START_USE(_vq) \
41 panic("%s:in_use = %i\n", \
42 (_vq)->vq.name, (_vq)->in_use); \
43 (_vq)->in_use = __LINE__; \
45 #define END_USE(_vq) \
46 do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
48 #define BAD_RING(_vq, fmt, args...) \
50 dev_err(&_vq->vq.vdev->dev, \
51 "%s:"fmt, (_vq)->vq.name, ##args); \
52 (_vq)->broken = true; \
58 struct vring_desc_state {
59 void *data; /* Data for callback. */
60 struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
63 struct vring_virtqueue {
66 /* Actual memory layout for this queue */
69 /* Can we use weak barriers? */
72 /* Other side has made a mess, don't try any more. */
75 /* Host supports indirect buffers */
78 /* Host publishes avail event idx */
81 /* Head of free buffer list. */
82 unsigned int free_head;
83 /* Number we've added since last sync. */
84 unsigned int num_added;
86 /* Last used index we've seen. */
89 /* Last written value to avail->flags */
90 u16 avail_flags_shadow;
92 /* Last written value to avail->idx in guest byte order */
95 /* How to notify other side. FIXME: commonalize hcalls! */
96 bool (*notify)(struct virtqueue *vq);
98 /* DMA, allocation, and size information */
100 size_t queue_size_in_bytes;
101 dma_addr_t queue_dma_addr;
104 /* They're supposed to lock for us. */
107 /* Figure out if their kicks are too delayed. */
108 bool last_add_time_valid;
109 ktime_t last_add_time;
112 /* Per-descriptor state. */
113 struct vring_desc_state desc_state[];
116 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
119 * Modern virtio devices have feature bits to specify whether they need a
120 * quirk and bypass the IOMMU. If not there, just use the DMA API.
122 * If there, the interaction between virtio and DMA API is messy.
124 * On most systems with virtio, physical addresses match bus addresses,
125 * and it doesn't particularly matter whether we use the DMA API.
127 * On some systems, including Xen and any system with a physical device
128 * that speaks virtio behind a physical IOMMU, we must use the DMA API
129 * for virtio DMA to work at all.
131 * On other systems, including SPARC and PPC64, virtio-pci devices are
132 * enumerated as though they are behind an IOMMU, but the virtio host
133 * ignores the IOMMU, so we must either pretend that the IOMMU isn't
134 * there or somehow map everything as the identity.
136 * For the time being, we preserve historic behavior and bypass the DMA
139 * TODO: install a per-device DMA ops structure that does the right thing
140 * taking into account all the above quirks, and use the DMA API
141 * unconditionally on data path.
144 static bool vring_use_dma_api(struct virtio_device *vdev)
146 if (!virtio_has_iommu_quirk(vdev))
149 /* Otherwise, we are left to guess. */
151 * In theory, it's possible to have a buggy QEMU-supposed
152 * emulated Q35 IOMMU and Xen enabled at the same time. On
153 * such a configuration, virtio has never worked and will
154 * not work without an even larger kludge. Instead, enable
155 * the DMA API if we're a Xen guest, which at least allows
156 * all of the sensible Xen configurations to work correctly.
165 * The DMA ops on various arches are rather gnarly right now, and
166 * making all of the arch DMA ops work on the vring device itself
167 * is a mess. For now, we use the parent device for DMA ops.
169 static inline struct device *vring_dma_dev(const struct vring_virtqueue *vq)
171 return vq->vq.vdev->dev.parent;
174 /* Map one sg entry. */
175 static dma_addr_t vring_map_one_sg(const struct vring_virtqueue *vq,
176 struct scatterlist *sg,
177 enum dma_data_direction direction)
179 if (!vring_use_dma_api(vq->vq.vdev))
180 return (dma_addr_t)sg_phys(sg);
183 * We can't use dma_map_sg, because we don't use scatterlists in
184 * the way it expects (we don't guarantee that the scatterlist
185 * will exist for the lifetime of the mapping).
187 return dma_map_page(vring_dma_dev(vq),
188 sg_page(sg), sg->offset, sg->length,
192 static dma_addr_t vring_map_single(const struct vring_virtqueue *vq,
193 void *cpu_addr, size_t size,
194 enum dma_data_direction direction)
196 if (!vring_use_dma_api(vq->vq.vdev))
197 return (dma_addr_t)virt_to_phys(cpu_addr);
199 return dma_map_single(vring_dma_dev(vq),
200 cpu_addr, size, direction);
203 static void vring_unmap_one(const struct vring_virtqueue *vq,
204 struct vring_desc *desc)
208 if (!vring_use_dma_api(vq->vq.vdev))
211 flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
213 if (flags & VRING_DESC_F_INDIRECT) {
214 dma_unmap_single(vring_dma_dev(vq),
215 virtio64_to_cpu(vq->vq.vdev, desc->addr),
216 virtio32_to_cpu(vq->vq.vdev, desc->len),
217 (flags & VRING_DESC_F_WRITE) ?
218 DMA_FROM_DEVICE : DMA_TO_DEVICE);
220 dma_unmap_page(vring_dma_dev(vq),
221 virtio64_to_cpu(vq->vq.vdev, desc->addr),
222 virtio32_to_cpu(vq->vq.vdev, desc->len),
223 (flags & VRING_DESC_F_WRITE) ?
224 DMA_FROM_DEVICE : DMA_TO_DEVICE);
228 static int vring_mapping_error(const struct vring_virtqueue *vq,
231 if (!vring_use_dma_api(vq->vq.vdev))
234 return dma_mapping_error(vring_dma_dev(vq), addr);
237 static struct vring_desc *alloc_indirect(struct virtqueue *_vq,
238 unsigned int total_sg, gfp_t gfp)
240 struct vring_desc *desc;
244 * We require lowmem mappings for the descriptors because
245 * otherwise virt_to_phys will give us bogus addresses in the
248 gfp &= ~__GFP_HIGHMEM;
250 desc = kmalloc_array(total_sg, sizeof(struct vring_desc), gfp);
254 for (i = 0; i < total_sg; i++)
255 desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
259 static inline int virtqueue_add(struct virtqueue *_vq,
260 struct scatterlist *sgs[],
261 unsigned int total_sg,
262 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);
278 BUG_ON(ctx && vq->indirect);
280 if (unlikely(vq->broken)) {
287 ktime_t now = ktime_get();
289 /* No kick or get, with .1 second between? Warn. */
290 if (vq->last_add_time_valid)
291 WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
293 vq->last_add_time = now;
294 vq->last_add_time_valid = true;
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);
308 WARN_ON_ONCE(total_sg > vq->vring.num && !vq->indirect);
312 /* Use a single buffer which doesn't continue */
314 /* Set up rest to use this indirect table. */
319 desc = vq->vring.desc;
321 descs_used = total_sg;
324 if (vq->vq.num_free < descs_used) {
325 pr_debug("Can't add buf len %i - avail = %i\n",
326 descs_used, vq->vq.num_free);
327 /* FIXME: for historical reasons, we force a notify here if
328 * there are outgoing parts to the buffer. Presumably the
329 * host should service the ring ASAP. */
338 for (n = 0; n < out_sgs; n++) {
339 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
340 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
341 if (vring_mapping_error(vq, addr))
344 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
345 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
346 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
348 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
351 for (; n < (out_sgs + in_sgs); n++) {
352 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
353 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
354 if (vring_mapping_error(vq, addr))
357 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
358 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
359 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
361 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
364 /* Last one doesn't continue. */
365 desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
368 /* Now that the indirect table is filled in, map it. */
369 dma_addr_t addr = vring_map_single(
370 vq, desc, total_sg * sizeof(struct vring_desc),
372 if (vring_mapping_error(vq, addr))
375 vq->vring.desc[head].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_INDIRECT);
376 vq->vring.desc[head].addr = cpu_to_virtio64(_vq->vdev, addr);
378 vq->vring.desc[head].len = cpu_to_virtio32(_vq->vdev, total_sg * sizeof(struct vring_desc));
381 /* We're using some buffers from the free list. */
382 vq->vq.num_free -= descs_used;
384 /* Update free pointer */
386 vq->free_head = virtio16_to_cpu(_vq->vdev, vq->vring.desc[head].next);
390 /* Store token and indirect buffer state. */
391 vq->desc_state[head].data = data;
393 vq->desc_state[head].indir_desc = desc;
395 vq->desc_state[head].indir_desc = ctx;
397 /* Put entry in available array (but don't update avail->idx until they
399 avail = vq->avail_idx_shadow & (vq->vring.num - 1);
400 vq->vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
402 /* Descriptors and available array need to be set before we expose the
403 * new available array entries. */
404 virtio_wmb(vq->weak_barriers);
405 vq->avail_idx_shadow++;
406 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
409 pr_debug("Added buffer head %i to %p\n", head, vq);
412 /* This is very unlikely, but theoretically possible. Kick
414 if (unlikely(vq->num_added == (1 << 16) - 1))
423 for (n = 0; n < total_sg; n++) {
426 vring_unmap_one(vq, &desc[i]);
427 i = virtio16_to_cpu(_vq->vdev, vq->vring.desc[i].next);
438 * virtqueue_add_sgs - expose buffers to other end
439 * @vq: the struct virtqueue we're talking about.
440 * @sgs: array of terminated scatterlists.
441 * @out_num: the number of scatterlists readable by other side
442 * @in_num: the number of scatterlists which are writable (after readable ones)
443 * @data: the token identifying the buffer.
444 * @gfp: how to do memory allocations (if necessary).
446 * Caller must ensure we don't call this with other virtqueue operations
447 * at the same time (except where noted).
449 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
451 int virtqueue_add_sgs(struct virtqueue *_vq,
452 struct scatterlist *sgs[],
453 unsigned int out_sgs,
458 unsigned int i, total_sg = 0;
460 /* Count them first. */
461 for (i = 0; i < out_sgs + in_sgs; i++) {
462 struct scatterlist *sg;
463 for (sg = sgs[i]; sg; sg = sg_next(sg))
466 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs,
469 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
472 * virtqueue_add_outbuf - expose output buffers to other end
473 * @vq: the struct virtqueue we're talking about.
474 * @sg: scatterlist (must be well-formed and terminated!)
475 * @num: the number of entries in @sg readable by other side
476 * @data: the token identifying the buffer.
477 * @gfp: how to do memory allocations (if necessary).
479 * Caller must ensure we don't call this with other virtqueue operations
480 * at the same time (except where noted).
482 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
484 int virtqueue_add_outbuf(struct virtqueue *vq,
485 struct scatterlist *sg, unsigned int num,
489 return virtqueue_add(vq, &sg, num, 1, 0, data, NULL, gfp);
491 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
494 * virtqueue_add_inbuf - expose input buffers to other end
495 * @vq: the struct virtqueue we're talking about.
496 * @sg: scatterlist (must be well-formed and terminated!)
497 * @num: the number of entries in @sg writable by other side
498 * @data: the token identifying the buffer.
499 * @gfp: how to do memory allocations (if necessary).
501 * Caller must ensure we don't call this with other virtqueue operations
502 * at the same time (except where noted).
504 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
506 int virtqueue_add_inbuf(struct virtqueue *vq,
507 struct scatterlist *sg, unsigned int num,
511 return virtqueue_add(vq, &sg, num, 0, 1, data, NULL, gfp);
513 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
516 * virtqueue_add_inbuf_ctx - expose input buffers to other end
517 * @vq: the struct virtqueue we're talking about.
518 * @sg: scatterlist (must be well-formed and terminated!)
519 * @num: the number of entries in @sg writable by other side
520 * @data: the token identifying the buffer.
521 * @ctx: extra context for the token
522 * @gfp: how to do memory allocations (if necessary).
524 * Caller must ensure we don't call this with other virtqueue operations
525 * at the same time (except where noted).
527 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
529 int virtqueue_add_inbuf_ctx(struct virtqueue *vq,
530 struct scatterlist *sg, unsigned int num,
535 return virtqueue_add(vq, &sg, num, 0, 1, data, ctx, gfp);
537 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf_ctx);
540 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
541 * @vq: the struct virtqueue
543 * Instead of virtqueue_kick(), you can do:
544 * if (virtqueue_kick_prepare(vq))
545 * virtqueue_notify(vq);
547 * This is sometimes useful because the virtqueue_kick_prepare() needs
548 * to be serialized, but the actual virtqueue_notify() call does not.
550 bool virtqueue_kick_prepare(struct virtqueue *_vq)
552 struct vring_virtqueue *vq = to_vvq(_vq);
557 /* We need to expose available array entries before checking avail
559 virtio_mb(vq->weak_barriers);
561 old = vq->avail_idx_shadow - vq->num_added;
562 new = vq->avail_idx_shadow;
566 if (vq->last_add_time_valid) {
567 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
568 vq->last_add_time)) > 100);
570 vq->last_add_time_valid = false;
574 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev, vring_avail_event(&vq->vring)),
577 needs_kick = !(vq->vring.used->flags & cpu_to_virtio16(_vq->vdev, VRING_USED_F_NO_NOTIFY));
582 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
585 * virtqueue_notify - second half of split virtqueue_kick call.
586 * @vq: the struct virtqueue
588 * This does not need to be serialized.
590 * Returns false if host notify failed or queue is broken, otherwise true.
592 bool virtqueue_notify(struct virtqueue *_vq)
594 struct vring_virtqueue *vq = to_vvq(_vq);
596 if (unlikely(vq->broken))
599 /* Prod other side to tell it about changes. */
600 if (!vq->notify(_vq)) {
606 EXPORT_SYMBOL_GPL(virtqueue_notify);
609 * virtqueue_kick - update after add_buf
610 * @vq: the struct virtqueue
612 * After one or more virtqueue_add_* calls, invoke this to kick
615 * Caller must ensure we don't call this with other virtqueue
616 * operations at the same time (except where noted).
618 * Returns false if kick failed, otherwise true.
620 bool virtqueue_kick(struct virtqueue *vq)
622 if (virtqueue_kick_prepare(vq))
623 return virtqueue_notify(vq);
626 EXPORT_SYMBOL_GPL(virtqueue_kick);
628 static void detach_buf(struct vring_virtqueue *vq, unsigned int head,
632 __virtio16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
634 /* Clear data ptr. */
635 vq->desc_state[head].data = NULL;
637 /* Put back on free list: unmap first-level descriptors and find end */
640 while (vq->vring.desc[i].flags & nextflag) {
641 vring_unmap_one(vq, &vq->vring.desc[i]);
642 i = virtio16_to_cpu(vq->vq.vdev, vq->vring.desc[i].next);
646 vring_unmap_one(vq, &vq->vring.desc[i]);
647 vq->vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev, vq->free_head);
648 vq->free_head = head;
650 /* Plus final descriptor */
654 struct vring_desc *indir_desc = vq->desc_state[head].indir_desc;
657 /* Free the indirect table, if any, now that it's unmapped. */
661 len = virtio32_to_cpu(vq->vq.vdev, vq->vring.desc[head].len);
663 BUG_ON(!(vq->vring.desc[head].flags &
664 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
665 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
667 for (j = 0; j < len / sizeof(struct vring_desc); j++)
668 vring_unmap_one(vq, &indir_desc[j]);
671 vq->desc_state[head].indir_desc = NULL;
673 *ctx = vq->desc_state[head].indir_desc;
677 static inline bool more_used(const struct vring_virtqueue *vq)
679 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev, vq->vring.used->idx);
683 * virtqueue_get_buf - get the next used buffer
684 * @vq: the struct virtqueue we're talking about.
685 * @len: the length written into the buffer
687 * If the device wrote data into the buffer, @len will be set to the
688 * amount written. This means you don't need to clear the buffer
689 * beforehand to ensure there's no data leakage in the case of short
692 * Caller must ensure we don't call this with other virtqueue
693 * operations at the same time (except where noted).
695 * Returns NULL if there are no used buffers, or the "data" token
696 * handed to virtqueue_add_*().
698 void *virtqueue_get_buf_ctx(struct virtqueue *_vq, unsigned int *len,
701 struct vring_virtqueue *vq = to_vvq(_vq);
708 if (unlikely(vq->broken)) {
713 if (!more_used(vq)) {
714 pr_debug("No more buffers in queue\n");
719 /* Only get used array entries after they have been exposed by host. */
720 virtio_rmb(vq->weak_barriers);
722 last_used = (vq->last_used_idx & (vq->vring.num - 1));
723 i = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].id);
724 *len = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].len);
726 if (unlikely(i >= vq->vring.num)) {
727 BAD_RING(vq, "id %u out of range\n", i);
730 if (unlikely(!vq->desc_state[i].data)) {
731 BAD_RING(vq, "id %u is not a head!\n", i);
735 /* detach_buf clears data, so grab it now. */
736 ret = vq->desc_state[i].data;
737 detach_buf(vq, i, ctx);
739 /* If we expect an interrupt for the next entry, tell host
740 * by writing event index and flush out the write before
741 * the read in the next get_buf call. */
742 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
743 virtio_store_mb(vq->weak_barriers,
744 &vring_used_event(&vq->vring),
745 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
748 vq->last_add_time_valid = false;
754 EXPORT_SYMBOL_GPL(virtqueue_get_buf_ctx);
756 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
758 return virtqueue_get_buf_ctx(_vq, len, NULL);
760 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
762 * virtqueue_disable_cb - disable callbacks
763 * @vq: the struct virtqueue we're talking about.
765 * Note that this is not necessarily synchronous, hence unreliable and only
766 * useful as an optimization.
768 * Unlike other operations, this need not be serialized.
770 void virtqueue_disable_cb(struct virtqueue *_vq)
772 struct vring_virtqueue *vq = to_vvq(_vq);
774 if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
775 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
777 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
781 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
784 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
785 * @vq: the struct virtqueue we're talking about.
787 * This re-enables callbacks; it returns current queue state
788 * in an opaque unsigned value. This value should be later tested by
789 * virtqueue_poll, to detect a possible race between the driver checking for
790 * more work, and enabling callbacks.
792 * Caller must ensure we don't call this with other virtqueue
793 * operations at the same time (except where noted).
795 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
797 struct vring_virtqueue *vq = to_vvq(_vq);
802 /* We optimistically turn back on interrupts, then check if there was
804 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
805 * either clear the flags bit or point the event index at the next
806 * entry. Always do both to keep code simple. */
807 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
808 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
810 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
812 vring_used_event(&vq->vring) = cpu_to_virtio16(_vq->vdev, last_used_idx = vq->last_used_idx);
814 return last_used_idx;
816 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
819 * virtqueue_poll - query pending used buffers
820 * @vq: the struct virtqueue we're talking about.
821 * @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
823 * Returns "true" if there are pending used buffers in the queue.
825 * This does not need to be serialized.
827 bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
829 struct vring_virtqueue *vq = to_vvq(_vq);
831 if (unlikely(vq->broken))
834 virtio_mb(vq->weak_barriers);
835 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev, vq->vring.used->idx);
837 EXPORT_SYMBOL_GPL(virtqueue_poll);
840 * virtqueue_enable_cb - restart callbacks after disable_cb.
841 * @vq: the struct virtqueue we're talking about.
843 * This re-enables callbacks; it returns "false" if there are pending
844 * buffers in the queue, to detect a possible race between the driver
845 * checking for more work, and enabling callbacks.
847 * Caller must ensure we don't call this with other virtqueue
848 * operations at the same time (except where noted).
850 bool virtqueue_enable_cb(struct virtqueue *_vq)
852 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
853 return !virtqueue_poll(_vq, last_used_idx);
855 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
858 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
859 * @vq: the struct virtqueue we're talking about.
861 * This re-enables callbacks but hints to the other side to delay
862 * interrupts until most of the available buffers have been processed;
863 * it returns "false" if there are many pending buffers in the queue,
864 * to detect a possible race between the driver checking for more work,
865 * and enabling callbacks.
867 * Caller must ensure we don't call this with other virtqueue
868 * operations at the same time (except where noted).
870 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
872 struct vring_virtqueue *vq = to_vvq(_vq);
877 /* We optimistically turn back on interrupts, then check if there was
879 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
880 * either clear the flags bit or point the event index at the next
881 * entry. Always update the event index to keep code simple. */
882 if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
883 vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
885 vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
887 /* TODO: tune this threshold */
888 bufs = (u16)(vq->avail_idx_shadow - vq->last_used_idx) * 3 / 4;
890 virtio_store_mb(vq->weak_barriers,
891 &vring_used_event(&vq->vring),
892 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
894 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->vring.used->idx) - vq->last_used_idx) > bufs)) {
902 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
905 * virtqueue_detach_unused_buf - detach first unused buffer
906 * @vq: the struct virtqueue we're talking about.
908 * Returns NULL or the "data" token handed to virtqueue_add_*().
909 * This is not valid on an active queue; it is useful only for device
912 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
914 struct vring_virtqueue *vq = to_vvq(_vq);
920 for (i = 0; i < vq->vring.num; i++) {
921 if (!vq->desc_state[i].data)
923 /* detach_buf clears data, so grab it now. */
924 buf = vq->desc_state[i].data;
925 detach_buf(vq, i, NULL);
926 vq->avail_idx_shadow--;
927 vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
931 /* That should have freed everything. */
932 BUG_ON(vq->vq.num_free != vq->vring.num);
937 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
939 irqreturn_t vring_interrupt(int irq, void *_vq)
941 struct vring_virtqueue *vq = to_vvq(_vq);
943 if (!more_used(vq)) {
944 pr_debug("virtqueue interrupt with no work for %p\n", vq);
948 if (unlikely(vq->broken))
951 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
953 vq->vq.callback(&vq->vq);
957 EXPORT_SYMBOL_GPL(vring_interrupt);
959 struct virtqueue *__vring_new_virtqueue(unsigned int index,
961 struct virtio_device *vdev,
964 bool (*notify)(struct virtqueue *),
965 void (*callback)(struct virtqueue *),
969 struct vring_virtqueue *vq;
971 vq = kmalloc(sizeof(*vq) + vring.num * sizeof(struct vring_desc_state),
977 vq->vq.callback = callback;
980 vq->vq.num_free = vring.num;
981 vq->vq.index = index;
982 vq->we_own_ring = false;
983 vq->queue_dma_addr = 0;
984 vq->queue_size_in_bytes = 0;
986 vq->weak_barriers = weak_barriers;
988 vq->last_used_idx = 0;
989 vq->avail_flags_shadow = 0;
990 vq->avail_idx_shadow = 0;
992 list_add_tail(&vq->vq.list, &vdev->vqs);
995 vq->last_add_time_valid = false;
998 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
1000 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
1002 /* No callback? Tell other side not to bother us. */
1004 vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
1006 vq->vring.avail->flags = cpu_to_virtio16(vdev, vq->avail_flags_shadow);
1009 /* Put everything in free lists. */
1011 for (i = 0; i < vring.num-1; i++)
1012 vq->vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
1013 memset(vq->desc_state, 0, vring.num * sizeof(struct vring_desc_state));
1017 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
1019 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
1020 dma_addr_t *dma_handle, gfp_t flag)
1022 if (vring_use_dma_api(vdev)) {
1023 return dma_alloc_coherent(vdev->dev.parent, size,
1026 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
1028 phys_addr_t phys_addr = virt_to_phys(queue);
1029 *dma_handle = (dma_addr_t)phys_addr;
1032 * Sanity check: make sure we dind't truncate
1033 * the address. The only arches I can find that
1034 * have 64-bit phys_addr_t but 32-bit dma_addr_t
1035 * are certain non-highmem MIPS and x86
1036 * configurations, but these configurations
1037 * should never allocate physical pages above 32
1038 * bits, so this is fine. Just in case, throw a
1039 * warning and abort if we end up with an
1040 * unrepresentable address.
1042 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
1043 free_pages_exact(queue, PAGE_ALIGN(size));
1051 static void vring_free_queue(struct virtio_device *vdev, size_t size,
1052 void *queue, dma_addr_t dma_handle)
1054 if (vring_use_dma_api(vdev)) {
1055 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
1057 free_pages_exact(queue, PAGE_ALIGN(size));
1061 struct virtqueue *vring_create_virtqueue(
1064 unsigned int vring_align,
1065 struct virtio_device *vdev,
1067 bool may_reduce_num,
1069 bool (*notify)(struct virtqueue *),
1070 void (*callback)(struct virtqueue *),
1073 struct virtqueue *vq;
1075 dma_addr_t dma_addr;
1076 size_t queue_size_in_bytes;
1079 /* We assume num is a power of 2. */
1080 if (num & (num - 1)) {
1081 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
1085 /* TODO: allocate each queue chunk individually */
1086 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
1087 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1089 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1092 if (!may_reduce_num)
1100 /* Try to get a single page. You are my only hope! */
1101 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
1102 &dma_addr, GFP_KERNEL|__GFP_ZERO);
1107 queue_size_in_bytes = vring_size(num, vring_align);
1108 vring_init(&vring, num, queue, vring_align);
1110 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
1111 notify, callback, name);
1113 vring_free_queue(vdev, queue_size_in_bytes, queue,
1118 to_vvq(vq)->queue_dma_addr = dma_addr;
1119 to_vvq(vq)->queue_size_in_bytes = queue_size_in_bytes;
1120 to_vvq(vq)->we_own_ring = true;
1124 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
1126 struct virtqueue *vring_new_virtqueue(unsigned int index,
1128 unsigned int vring_align,
1129 struct virtio_device *vdev,
1133 bool (*notify)(struct virtqueue *vq),
1134 void (*callback)(struct virtqueue *vq),
1138 vring_init(&vring, num, pages, vring_align);
1139 return __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
1140 notify, callback, name);
1142 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
1144 void vring_del_virtqueue(struct virtqueue *_vq)
1146 struct vring_virtqueue *vq = to_vvq(_vq);
1148 if (vq->we_own_ring) {
1149 vring_free_queue(vq->vq.vdev, vq->queue_size_in_bytes,
1150 vq->vring.desc, vq->queue_dma_addr);
1152 list_del(&_vq->list);
1155 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
1157 /* Manipulates transport-specific feature bits. */
1158 void vring_transport_features(struct virtio_device *vdev)
1162 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
1164 case VIRTIO_RING_F_INDIRECT_DESC:
1166 case VIRTIO_RING_F_EVENT_IDX:
1168 case VIRTIO_F_VERSION_1:
1170 case VIRTIO_F_IOMMU_PLATFORM:
1173 /* We don't understand this bit. */
1174 __virtio_clear_bit(vdev, i);
1178 EXPORT_SYMBOL_GPL(vring_transport_features);
1181 * virtqueue_get_vring_size - return the size of the virtqueue's vring
1182 * @vq: the struct virtqueue containing the vring of interest.
1184 * Returns the size of the vring. This is mainly used for boasting to
1185 * userspace. Unlike other operations, this need not be serialized.
1187 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
1190 struct vring_virtqueue *vq = to_vvq(_vq);
1192 return vq->vring.num;
1194 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
1196 bool virtqueue_is_broken(struct virtqueue *_vq)
1198 struct vring_virtqueue *vq = to_vvq(_vq);
1200 return READ_ONCE(vq->broken);
1202 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
1205 * This should prevent the device from being used, allowing drivers to
1206 * recover. You may need to grab appropriate locks to flush.
1208 void virtio_break_device(struct virtio_device *dev)
1210 struct virtqueue *_vq;
1212 list_for_each_entry(_vq, &dev->vqs, list) {
1213 struct vring_virtqueue *vq = to_vvq(_vq);
1215 /* Pairs with READ_ONCE() in virtqueue_is_broken(). */
1216 WRITE_ONCE(vq->broken, true);
1219 EXPORT_SYMBOL_GPL(virtio_break_device);
1221 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
1223 struct vring_virtqueue *vq = to_vvq(_vq);
1225 BUG_ON(!vq->we_own_ring);
1227 return vq->queue_dma_addr;
1229 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
1231 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
1233 struct vring_virtqueue *vq = to_vvq(_vq);
1235 BUG_ON(!vq->we_own_ring);
1237 return vq->queue_dma_addr +
1238 ((char *)vq->vring.avail - (char *)vq->vring.desc);
1240 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
1242 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
1244 struct vring_virtqueue *vq = to_vvq(_vq);
1246 BUG_ON(!vq->we_own_ring);
1248 return vq->queue_dma_addr +
1249 ((char *)vq->vring.used - (char *)vq->vring.desc);
1251 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
1253 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
1255 return &to_vvq(vq)->vring;
1257 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
1259 MODULE_LICENSE("GPL");