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)
47 #define LAST_ADD_TIME_UPDATE(_vq) \
49 ktime_t now = ktime_get(); \
51 /* No kick or get, with .1 second between? Warn. */ \
52 if ((_vq)->last_add_time_valid) \
53 WARN_ON(ktime_to_ms(ktime_sub(now, \
54 (_vq)->last_add_time)) > 100); \
55 (_vq)->last_add_time = now; \
56 (_vq)->last_add_time_valid = true; \
58 #define LAST_ADD_TIME_CHECK(_vq) \
60 if ((_vq)->last_add_time_valid) { \
61 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(), \
62 (_vq)->last_add_time)) > 100); \
65 #define LAST_ADD_TIME_INVALID(_vq) \
66 ((_vq)->last_add_time_valid = false)
68 #define BAD_RING(_vq, fmt, args...) \
70 dev_err(&_vq->vq.vdev->dev, \
71 "%s:"fmt, (_vq)->vq.name, ##args); \
72 (_vq)->broken = true; \
76 #define LAST_ADD_TIME_UPDATE(vq)
77 #define LAST_ADD_TIME_CHECK(vq)
78 #define LAST_ADD_TIME_INVALID(vq)
81 struct vring_desc_state_split {
82 void *data; /* Data for callback. */
83 struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
86 struct vring_desc_state_packed {
87 void *data; /* Data for callback. */
88 struct vring_packed_desc *indir_desc; /* Indirect descriptor, if any. */
89 u16 num; /* Descriptor list length. */
90 u16 next; /* The next desc state in a list. */
91 u16 last; /* The last desc state in a list. */
94 struct vring_desc_extra_packed {
95 dma_addr_t addr; /* Buffer DMA addr. */
96 u32 len; /* Buffer length. */
97 u16 flags; /* Descriptor flags. */
100 struct vring_virtqueue {
103 /* Is this a packed ring? */
106 /* Is DMA API used? */
109 /* Can we use weak barriers? */
112 /* Other side has made a mess, don't try any more. */
115 /* Host supports indirect buffers */
118 /* Host publishes avail event idx */
121 /* Head of free buffer list. */
122 unsigned int free_head;
123 /* Number we've added since last sync. */
124 unsigned int num_added;
126 /* Last used index we've seen. */
130 /* Available for split ring */
132 /* Actual memory layout for this queue. */
135 /* Last written value to avail->flags */
136 u16 avail_flags_shadow;
139 * Last written value to avail->idx in
142 u16 avail_idx_shadow;
144 /* Per-descriptor state. */
145 struct vring_desc_state_split *desc_state;
147 /* DMA address and size information */
148 dma_addr_t queue_dma_addr;
149 size_t queue_size_in_bytes;
152 /* Available for packed ring */
154 /* Actual memory layout for this queue. */
157 struct vring_packed_desc *desc;
158 struct vring_packed_desc_event *driver;
159 struct vring_packed_desc_event *device;
162 /* Driver ring wrap counter. */
163 bool avail_wrap_counter;
165 /* Device ring wrap counter. */
166 bool used_wrap_counter;
168 /* Avail used flags. */
169 u16 avail_used_flags;
171 /* Index of the next avail descriptor. */
175 * Last written value to driver->flags in
178 u16 event_flags_shadow;
180 /* Per-descriptor state. */
181 struct vring_desc_state_packed *desc_state;
182 struct vring_desc_extra_packed *desc_extra;
184 /* DMA address and size information */
185 dma_addr_t ring_dma_addr;
186 dma_addr_t driver_event_dma_addr;
187 dma_addr_t device_event_dma_addr;
188 size_t ring_size_in_bytes;
189 size_t event_size_in_bytes;
193 /* How to notify other side. FIXME: commonalize hcalls! */
194 bool (*notify)(struct virtqueue *vq);
196 /* DMA, allocation, and size information */
200 /* They're supposed to lock for us. */
203 /* Figure out if their kicks are too delayed. */
204 bool last_add_time_valid;
205 ktime_t last_add_time;
214 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
216 static inline bool virtqueue_use_indirect(struct virtqueue *_vq,
217 unsigned int total_sg)
219 struct vring_virtqueue *vq = to_vvq(_vq);
222 * If the host supports indirect descriptor tables, and we have multiple
223 * buffers, then go indirect. FIXME: tune this threshold
225 return (vq->indirect && total_sg > 1 && vq->vq.num_free);
229 * Modern virtio devices have feature bits to specify whether they need a
230 * quirk and bypass the IOMMU. If not there, just use the DMA API.
232 * If there, the interaction between virtio and DMA API is messy.
234 * On most systems with virtio, physical addresses match bus addresses,
235 * and it doesn't particularly matter whether we use the DMA API.
237 * On some systems, including Xen and any system with a physical device
238 * that speaks virtio behind a physical IOMMU, we must use the DMA API
239 * for virtio DMA to work at all.
241 * On other systems, including SPARC and PPC64, virtio-pci devices are
242 * enumerated as though they are behind an IOMMU, but the virtio host
243 * ignores the IOMMU, so we must either pretend that the IOMMU isn't
244 * there or somehow map everything as the identity.
246 * For the time being, we preserve historic behavior and bypass the DMA
249 * TODO: install a per-device DMA ops structure that does the right thing
250 * taking into account all the above quirks, and use the DMA API
251 * unconditionally on data path.
254 static bool vring_use_dma_api(struct virtio_device *vdev)
256 if (!virtio_has_iommu_quirk(vdev))
259 /* Otherwise, we are left to guess. */
261 * In theory, it's possible to have a buggy QEMU-supposed
262 * emulated Q35 IOMMU and Xen enabled at the same time. On
263 * such a configuration, virtio has never worked and will
264 * not work without an even larger kludge. Instead, enable
265 * the DMA API if we're a Xen guest, which at least allows
266 * all of the sensible Xen configurations to work correctly.
274 size_t virtio_max_dma_size(struct virtio_device *vdev)
276 size_t max_segment_size = SIZE_MAX;
278 if (vring_use_dma_api(vdev))
279 max_segment_size = dma_max_mapping_size(&vdev->dev);
281 return max_segment_size;
283 EXPORT_SYMBOL_GPL(virtio_max_dma_size);
285 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
286 dma_addr_t *dma_handle, gfp_t flag)
288 if (vring_use_dma_api(vdev)) {
289 return dma_alloc_coherent(vdev->dev.parent, size,
292 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
295 phys_addr_t phys_addr = virt_to_phys(queue);
296 *dma_handle = (dma_addr_t)phys_addr;
299 * Sanity check: make sure we dind't truncate
300 * the address. The only arches I can find that
301 * have 64-bit phys_addr_t but 32-bit dma_addr_t
302 * are certain non-highmem MIPS and x86
303 * configurations, but these configurations
304 * should never allocate physical pages above 32
305 * bits, so this is fine. Just in case, throw a
306 * warning and abort if we end up with an
307 * unrepresentable address.
309 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
310 free_pages_exact(queue, PAGE_ALIGN(size));
318 static void vring_free_queue(struct virtio_device *vdev, size_t size,
319 void *queue, dma_addr_t dma_handle)
321 if (vring_use_dma_api(vdev))
322 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
324 free_pages_exact(queue, PAGE_ALIGN(size));
328 * The DMA ops on various arches are rather gnarly right now, and
329 * making all of the arch DMA ops work on the vring device itself
330 * is a mess. For now, we use the parent device for DMA ops.
332 static inline struct device *vring_dma_dev(const struct vring_virtqueue *vq)
334 return vq->vq.vdev->dev.parent;
337 /* Map one sg entry. */
338 static dma_addr_t vring_map_one_sg(const struct vring_virtqueue *vq,
339 struct scatterlist *sg,
340 enum dma_data_direction direction)
342 if (!vq->use_dma_api)
343 return (dma_addr_t)sg_phys(sg);
346 * We can't use dma_map_sg, because we don't use scatterlists in
347 * the way it expects (we don't guarantee that the scatterlist
348 * will exist for the lifetime of the mapping).
350 return dma_map_page(vring_dma_dev(vq),
351 sg_page(sg), sg->offset, sg->length,
355 static dma_addr_t vring_map_single(const struct vring_virtqueue *vq,
356 void *cpu_addr, size_t size,
357 enum dma_data_direction direction)
359 if (!vq->use_dma_api)
360 return (dma_addr_t)virt_to_phys(cpu_addr);
362 return dma_map_single(vring_dma_dev(vq),
363 cpu_addr, size, direction);
366 static int vring_mapping_error(const struct vring_virtqueue *vq,
369 if (!vq->use_dma_api)
372 return dma_mapping_error(vring_dma_dev(vq), addr);
377 * Split ring specific functions - *_split().
380 static void vring_unmap_one_split(const struct vring_virtqueue *vq,
381 struct vring_desc *desc)
385 if (!vq->use_dma_api)
388 flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
390 if (flags & VRING_DESC_F_INDIRECT) {
391 dma_unmap_single(vring_dma_dev(vq),
392 virtio64_to_cpu(vq->vq.vdev, desc->addr),
393 virtio32_to_cpu(vq->vq.vdev, desc->len),
394 (flags & VRING_DESC_F_WRITE) ?
395 DMA_FROM_DEVICE : DMA_TO_DEVICE);
397 dma_unmap_page(vring_dma_dev(vq),
398 virtio64_to_cpu(vq->vq.vdev, desc->addr),
399 virtio32_to_cpu(vq->vq.vdev, desc->len),
400 (flags & VRING_DESC_F_WRITE) ?
401 DMA_FROM_DEVICE : DMA_TO_DEVICE);
405 static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
406 unsigned int total_sg,
409 struct vring_desc *desc;
413 * We require lowmem mappings for the descriptors because
414 * otherwise virt_to_phys will give us bogus addresses in the
417 gfp &= ~__GFP_HIGHMEM;
419 desc = kmalloc_array(total_sg, sizeof(struct vring_desc), gfp);
423 for (i = 0; i < total_sg; i++)
424 desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
428 static inline int virtqueue_add_split(struct virtqueue *_vq,
429 struct scatterlist *sgs[],
430 unsigned int total_sg,
431 unsigned int out_sgs,
437 struct vring_virtqueue *vq = to_vvq(_vq);
438 struct scatterlist *sg;
439 struct vring_desc *desc;
440 unsigned int i, n, avail, descs_used, uninitialized_var(prev), err_idx;
446 BUG_ON(data == NULL);
447 BUG_ON(ctx && vq->indirect);
449 if (unlikely(vq->broken)) {
454 LAST_ADD_TIME_UPDATE(vq);
456 BUG_ON(total_sg == 0);
458 head = vq->free_head;
460 if (virtqueue_use_indirect(_vq, total_sg))
461 desc = alloc_indirect_split(_vq, total_sg, gfp);
464 WARN_ON_ONCE(total_sg > vq->split.vring.num && !vq->indirect);
468 /* Use a single buffer which doesn't continue */
470 /* Set up rest to use this indirect table. */
475 desc = vq->split.vring.desc;
477 descs_used = total_sg;
480 if (vq->vq.num_free < descs_used) {
481 pr_debug("Can't add buf len %i - avail = %i\n",
482 descs_used, vq->vq.num_free);
483 /* FIXME: for historical reasons, we force a notify here if
484 * there are outgoing parts to the buffer. Presumably the
485 * host should service the ring ASAP. */
494 for (n = 0; n < out_sgs; n++) {
495 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
496 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
497 if (vring_mapping_error(vq, addr))
500 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
501 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
502 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
504 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
507 for (; n < (out_sgs + in_sgs); n++) {
508 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
509 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
510 if (vring_mapping_error(vq, addr))
513 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
514 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
515 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
517 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
520 /* Last one doesn't continue. */
521 desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
524 /* Now that the indirect table is filled in, map it. */
525 dma_addr_t addr = vring_map_single(
526 vq, desc, total_sg * sizeof(struct vring_desc),
528 if (vring_mapping_error(vq, addr))
531 vq->split.vring.desc[head].flags = cpu_to_virtio16(_vq->vdev,
532 VRING_DESC_F_INDIRECT);
533 vq->split.vring.desc[head].addr = cpu_to_virtio64(_vq->vdev,
536 vq->split.vring.desc[head].len = cpu_to_virtio32(_vq->vdev,
537 total_sg * sizeof(struct vring_desc));
540 /* We're using some buffers from the free list. */
541 vq->vq.num_free -= descs_used;
543 /* Update free pointer */
545 vq->free_head = virtio16_to_cpu(_vq->vdev,
546 vq->split.vring.desc[head].next);
550 /* Store token and indirect buffer state. */
551 vq->split.desc_state[head].data = data;
553 vq->split.desc_state[head].indir_desc = desc;
555 vq->split.desc_state[head].indir_desc = ctx;
557 /* Put entry in available array (but don't update avail->idx until they
559 avail = vq->split.avail_idx_shadow & (vq->split.vring.num - 1);
560 vq->split.vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
562 /* Descriptors and available array need to be set before we expose the
563 * new available array entries. */
564 virtio_wmb(vq->weak_barriers);
565 vq->split.avail_idx_shadow++;
566 vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
567 vq->split.avail_idx_shadow);
570 pr_debug("Added buffer head %i to %p\n", head, vq);
573 /* This is very unlikely, but theoretically possible. Kick
575 if (unlikely(vq->num_added == (1 << 16) - 1))
584 for (n = 0; n < total_sg; n++) {
587 vring_unmap_one_split(vq, &desc[i]);
588 i = virtio16_to_cpu(_vq->vdev, vq->split.vring.desc[i].next);
598 static bool virtqueue_kick_prepare_split(struct virtqueue *_vq)
600 struct vring_virtqueue *vq = to_vvq(_vq);
605 /* We need to expose available array entries before checking avail
607 virtio_mb(vq->weak_barriers);
609 old = vq->split.avail_idx_shadow - vq->num_added;
610 new = vq->split.avail_idx_shadow;
613 LAST_ADD_TIME_CHECK(vq);
614 LAST_ADD_TIME_INVALID(vq);
617 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev,
618 vring_avail_event(&vq->split.vring)),
621 needs_kick = !(vq->split.vring.used->flags &
622 cpu_to_virtio16(_vq->vdev,
623 VRING_USED_F_NO_NOTIFY));
629 static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
633 __virtio16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
635 /* Clear data ptr. */
636 vq->split.desc_state[head].data = NULL;
638 /* Put back on free list: unmap first-level descriptors and find end */
641 while (vq->split.vring.desc[i].flags & nextflag) {
642 vring_unmap_one_split(vq, &vq->split.vring.desc[i]);
643 i = virtio16_to_cpu(vq->vq.vdev, vq->split.vring.desc[i].next);
647 vring_unmap_one_split(vq, &vq->split.vring.desc[i]);
648 vq->split.vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev,
650 vq->free_head = head;
652 /* Plus final descriptor */
656 struct vring_desc *indir_desc =
657 vq->split.desc_state[head].indir_desc;
660 /* Free the indirect table, if any, now that it's unmapped. */
664 len = virtio32_to_cpu(vq->vq.vdev,
665 vq->split.vring.desc[head].len);
667 BUG_ON(!(vq->split.vring.desc[head].flags &
668 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
669 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
671 for (j = 0; j < len / sizeof(struct vring_desc); j++)
672 vring_unmap_one_split(vq, &indir_desc[j]);
675 vq->split.desc_state[head].indir_desc = NULL;
677 *ctx = vq->split.desc_state[head].indir_desc;
681 static inline bool more_used_split(const struct vring_virtqueue *vq)
683 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev,
684 vq->split.vring.used->idx);
687 static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq,
691 struct vring_virtqueue *vq = to_vvq(_vq);
698 if (unlikely(vq->broken)) {
703 if (!more_used_split(vq)) {
704 pr_debug("No more buffers in queue\n");
709 /* Only get used array entries after they have been exposed by host. */
710 virtio_rmb(vq->weak_barriers);
712 last_used = (vq->last_used_idx & (vq->split.vring.num - 1));
713 i = virtio32_to_cpu(_vq->vdev,
714 vq->split.vring.used->ring[last_used].id);
715 *len = virtio32_to_cpu(_vq->vdev,
716 vq->split.vring.used->ring[last_used].len);
718 if (unlikely(i >= vq->split.vring.num)) {
719 BAD_RING(vq, "id %u out of range\n", i);
722 if (unlikely(!vq->split.desc_state[i].data)) {
723 BAD_RING(vq, "id %u is not a head!\n", i);
727 /* detach_buf_split clears data, so grab it now. */
728 ret = vq->split.desc_state[i].data;
729 detach_buf_split(vq, i, ctx);
731 /* If we expect an interrupt for the next entry, tell host
732 * by writing event index and flush out the write before
733 * the read in the next get_buf call. */
734 if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
735 virtio_store_mb(vq->weak_barriers,
736 &vring_used_event(&vq->split.vring),
737 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
739 LAST_ADD_TIME_INVALID(vq);
745 static void virtqueue_disable_cb_split(struct virtqueue *_vq)
747 struct vring_virtqueue *vq = to_vvq(_vq);
749 if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
750 vq->split.avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
752 vq->split.vring.avail->flags =
753 cpu_to_virtio16(_vq->vdev,
754 vq->split.avail_flags_shadow);
758 static unsigned virtqueue_enable_cb_prepare_split(struct virtqueue *_vq)
760 struct vring_virtqueue *vq = to_vvq(_vq);
765 /* We optimistically turn back on interrupts, then check if there was
767 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
768 * either clear the flags bit or point the event index at the next
769 * entry. Always do both to keep code simple. */
770 if (vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
771 vq->split.avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
773 vq->split.vring.avail->flags =
774 cpu_to_virtio16(_vq->vdev,
775 vq->split.avail_flags_shadow);
777 vring_used_event(&vq->split.vring) = cpu_to_virtio16(_vq->vdev,
778 last_used_idx = vq->last_used_idx);
780 return last_used_idx;
783 static bool virtqueue_poll_split(struct virtqueue *_vq, unsigned last_used_idx)
785 struct vring_virtqueue *vq = to_vvq(_vq);
787 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev,
788 vq->split.vring.used->idx);
791 static bool virtqueue_enable_cb_delayed_split(struct virtqueue *_vq)
793 struct vring_virtqueue *vq = to_vvq(_vq);
798 /* We optimistically turn back on interrupts, then check if there was
800 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
801 * either clear the flags bit or point the event index at the next
802 * entry. Always update the event index to keep code simple. */
803 if (vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
804 vq->split.avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
806 vq->split.vring.avail->flags =
807 cpu_to_virtio16(_vq->vdev,
808 vq->split.avail_flags_shadow);
810 /* TODO: tune this threshold */
811 bufs = (u16)(vq->split.avail_idx_shadow - vq->last_used_idx) * 3 / 4;
813 virtio_store_mb(vq->weak_barriers,
814 &vring_used_event(&vq->split.vring),
815 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
817 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->split.vring.used->idx)
818 - vq->last_used_idx) > bufs)) {
827 static void *virtqueue_detach_unused_buf_split(struct virtqueue *_vq)
829 struct vring_virtqueue *vq = to_vvq(_vq);
835 for (i = 0; i < vq->split.vring.num; i++) {
836 if (!vq->split.desc_state[i].data)
838 /* detach_buf_split clears data, so grab it now. */
839 buf = vq->split.desc_state[i].data;
840 detach_buf_split(vq, i, NULL);
841 vq->split.avail_idx_shadow--;
842 vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
843 vq->split.avail_idx_shadow);
847 /* That should have freed everything. */
848 BUG_ON(vq->vq.num_free != vq->split.vring.num);
854 static struct virtqueue *vring_create_virtqueue_split(
857 unsigned int vring_align,
858 struct virtio_device *vdev,
862 bool (*notify)(struct virtqueue *),
863 void (*callback)(struct virtqueue *),
866 struct virtqueue *vq;
869 size_t queue_size_in_bytes;
872 /* We assume num is a power of 2. */
873 if (num & (num - 1)) {
874 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
878 /* TODO: allocate each queue chunk individually */
879 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
880 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
882 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
891 /* Try to get a single page. You are my only hope! */
892 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
893 &dma_addr, GFP_KERNEL|__GFP_ZERO);
898 queue_size_in_bytes = vring_size(num, vring_align);
899 vring_init(&vring, num, queue, vring_align);
901 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
902 notify, callback, name);
904 vring_free_queue(vdev, queue_size_in_bytes, queue,
909 to_vvq(vq)->split.queue_dma_addr = dma_addr;
910 to_vvq(vq)->split.queue_size_in_bytes = queue_size_in_bytes;
911 to_vvq(vq)->we_own_ring = true;
918 * Packed ring specific functions - *_packed().
921 static void vring_unmap_state_packed(const struct vring_virtqueue *vq,
922 struct vring_desc_extra_packed *state)
926 if (!vq->use_dma_api)
929 flags = state->flags;
931 if (flags & VRING_DESC_F_INDIRECT) {
932 dma_unmap_single(vring_dma_dev(vq),
933 state->addr, state->len,
934 (flags & VRING_DESC_F_WRITE) ?
935 DMA_FROM_DEVICE : DMA_TO_DEVICE);
937 dma_unmap_page(vring_dma_dev(vq),
938 state->addr, state->len,
939 (flags & VRING_DESC_F_WRITE) ?
940 DMA_FROM_DEVICE : DMA_TO_DEVICE);
944 static void vring_unmap_desc_packed(const struct vring_virtqueue *vq,
945 struct vring_packed_desc *desc)
949 if (!vq->use_dma_api)
952 flags = le16_to_cpu(desc->flags);
954 if (flags & VRING_DESC_F_INDIRECT) {
955 dma_unmap_single(vring_dma_dev(vq),
956 le64_to_cpu(desc->addr),
957 le32_to_cpu(desc->len),
958 (flags & VRING_DESC_F_WRITE) ?
959 DMA_FROM_DEVICE : DMA_TO_DEVICE);
961 dma_unmap_page(vring_dma_dev(vq),
962 le64_to_cpu(desc->addr),
963 le32_to_cpu(desc->len),
964 (flags & VRING_DESC_F_WRITE) ?
965 DMA_FROM_DEVICE : DMA_TO_DEVICE);
969 static struct vring_packed_desc *alloc_indirect_packed(unsigned int total_sg,
972 struct vring_packed_desc *desc;
975 * We require lowmem mappings for the descriptors because
976 * otherwise virt_to_phys will give us bogus addresses in the
979 gfp &= ~__GFP_HIGHMEM;
981 desc = kmalloc_array(total_sg, sizeof(struct vring_packed_desc), gfp);
986 static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
987 struct scatterlist *sgs[],
988 unsigned int total_sg,
989 unsigned int out_sgs,
994 struct vring_packed_desc *desc;
995 struct scatterlist *sg;
996 unsigned int i, n, err_idx;
1000 head = vq->packed.next_avail_idx;
1001 desc = alloc_indirect_packed(total_sg, gfp);
1003 if (unlikely(vq->vq.num_free < 1)) {
1004 pr_debug("Can't add buf len 1 - avail = 0\n");
1011 BUG_ON(id == vq->packed.vring.num);
1013 for (n = 0; n < out_sgs + in_sgs; n++) {
1014 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
1015 addr = vring_map_one_sg(vq, sg, n < out_sgs ?
1016 DMA_TO_DEVICE : DMA_FROM_DEVICE);
1017 if (vring_mapping_error(vq, addr))
1020 desc[i].flags = cpu_to_le16(n < out_sgs ?
1021 0 : VRING_DESC_F_WRITE);
1022 desc[i].addr = cpu_to_le64(addr);
1023 desc[i].len = cpu_to_le32(sg->length);
1028 /* Now that the indirect table is filled in, map it. */
1029 addr = vring_map_single(vq, desc,
1030 total_sg * sizeof(struct vring_packed_desc),
1032 if (vring_mapping_error(vq, addr))
1035 vq->packed.vring.desc[head].addr = cpu_to_le64(addr);
1036 vq->packed.vring.desc[head].len = cpu_to_le32(total_sg *
1037 sizeof(struct vring_packed_desc));
1038 vq->packed.vring.desc[head].id = cpu_to_le16(id);
1040 if (vq->use_dma_api) {
1041 vq->packed.desc_extra[id].addr = addr;
1042 vq->packed.desc_extra[id].len = total_sg *
1043 sizeof(struct vring_packed_desc);
1044 vq->packed.desc_extra[id].flags = VRING_DESC_F_INDIRECT |
1045 vq->packed.avail_used_flags;
1049 * A driver MUST NOT make the first descriptor in the list
1050 * available before all subsequent descriptors comprising
1051 * the list are made available.
1053 virtio_wmb(vq->weak_barriers);
1054 vq->packed.vring.desc[head].flags = cpu_to_le16(VRING_DESC_F_INDIRECT |
1055 vq->packed.avail_used_flags);
1057 /* We're using some buffers from the free list. */
1058 vq->vq.num_free -= 1;
1060 /* Update free pointer */
1062 if (n >= vq->packed.vring.num) {
1064 vq->packed.avail_wrap_counter ^= 1;
1065 vq->packed.avail_used_flags ^=
1066 1 << VRING_PACKED_DESC_F_AVAIL |
1067 1 << VRING_PACKED_DESC_F_USED;
1069 vq->packed.next_avail_idx = n;
1070 vq->free_head = vq->packed.desc_state[id].next;
1072 /* Store token and indirect buffer state. */
1073 vq->packed.desc_state[id].num = 1;
1074 vq->packed.desc_state[id].data = data;
1075 vq->packed.desc_state[id].indir_desc = desc;
1076 vq->packed.desc_state[id].last = id;
1080 pr_debug("Added buffer head %i to %p\n", head, vq);
1088 for (i = 0; i < err_idx; i++)
1089 vring_unmap_desc_packed(vq, &desc[i]);
1097 static inline int virtqueue_add_packed(struct virtqueue *_vq,
1098 struct scatterlist *sgs[],
1099 unsigned int total_sg,
1100 unsigned int out_sgs,
1101 unsigned int in_sgs,
1106 struct vring_virtqueue *vq = to_vvq(_vq);
1107 struct vring_packed_desc *desc;
1108 struct scatterlist *sg;
1109 unsigned int i, n, c, descs_used, err_idx;
1110 __le16 uninitialized_var(head_flags), flags;
1111 u16 head, id, uninitialized_var(prev), curr, avail_used_flags;
1115 BUG_ON(data == NULL);
1116 BUG_ON(ctx && vq->indirect);
1118 if (unlikely(vq->broken)) {
1123 LAST_ADD_TIME_UPDATE(vq);
1125 BUG_ON(total_sg == 0);
1127 if (virtqueue_use_indirect(_vq, total_sg))
1128 return virtqueue_add_indirect_packed(vq, sgs, total_sg,
1129 out_sgs, in_sgs, data, gfp);
1131 head = vq->packed.next_avail_idx;
1132 avail_used_flags = vq->packed.avail_used_flags;
1134 WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
1136 desc = vq->packed.vring.desc;
1138 descs_used = total_sg;
1140 if (unlikely(vq->vq.num_free < descs_used)) {
1141 pr_debug("Can't add buf len %i - avail = %i\n",
1142 descs_used, vq->vq.num_free);
1148 BUG_ON(id == vq->packed.vring.num);
1152 for (n = 0; n < out_sgs + in_sgs; n++) {
1153 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
1154 dma_addr_t addr = vring_map_one_sg(vq, sg, n < out_sgs ?
1155 DMA_TO_DEVICE : DMA_FROM_DEVICE);
1156 if (vring_mapping_error(vq, addr))
1159 flags = cpu_to_le16(vq->packed.avail_used_flags |
1160 (++c == total_sg ? 0 : VRING_DESC_F_NEXT) |
1161 (n < out_sgs ? 0 : VRING_DESC_F_WRITE));
1165 desc[i].flags = flags;
1167 desc[i].addr = cpu_to_le64(addr);
1168 desc[i].len = cpu_to_le32(sg->length);
1169 desc[i].id = cpu_to_le16(id);
1171 if (unlikely(vq->use_dma_api)) {
1172 vq->packed.desc_extra[curr].addr = addr;
1173 vq->packed.desc_extra[curr].len = sg->length;
1174 vq->packed.desc_extra[curr].flags =
1178 curr = vq->packed.desc_state[curr].next;
1180 if ((unlikely(++i >= vq->packed.vring.num))) {
1182 vq->packed.avail_used_flags ^=
1183 1 << VRING_PACKED_DESC_F_AVAIL |
1184 1 << VRING_PACKED_DESC_F_USED;
1190 vq->packed.avail_wrap_counter ^= 1;
1192 /* We're using some buffers from the free list. */
1193 vq->vq.num_free -= descs_used;
1195 /* Update free pointer */
1196 vq->packed.next_avail_idx = i;
1197 vq->free_head = curr;
1200 vq->packed.desc_state[id].num = descs_used;
1201 vq->packed.desc_state[id].data = data;
1202 vq->packed.desc_state[id].indir_desc = ctx;
1203 vq->packed.desc_state[id].last = prev;
1206 * A driver MUST NOT make the first descriptor in the list
1207 * available before all subsequent descriptors comprising
1208 * the list are made available.
1210 virtio_wmb(vq->weak_barriers);
1211 vq->packed.vring.desc[head].flags = head_flags;
1212 vq->num_added += descs_used;
1214 pr_debug("Added buffer head %i to %p\n", head, vq);
1223 vq->packed.avail_used_flags = avail_used_flags;
1225 for (n = 0; n < total_sg; n++) {
1228 vring_unmap_desc_packed(vq, &desc[i]);
1230 if (i >= vq->packed.vring.num)
1238 static bool virtqueue_kick_prepare_packed(struct virtqueue *_vq)
1240 struct vring_virtqueue *vq = to_vvq(_vq);
1241 u16 new, old, off_wrap, flags, wrap_counter, event_idx;
1254 * We need to expose the new flags value before checking notification
1257 virtio_mb(vq->weak_barriers);
1259 old = vq->packed.next_avail_idx - vq->num_added;
1260 new = vq->packed.next_avail_idx;
1263 snapshot.u32 = *(u32 *)vq->packed.vring.device;
1264 flags = le16_to_cpu(snapshot.flags);
1266 LAST_ADD_TIME_CHECK(vq);
1267 LAST_ADD_TIME_INVALID(vq);
1269 if (flags != VRING_PACKED_EVENT_FLAG_DESC) {
1270 needs_kick = (flags != VRING_PACKED_EVENT_FLAG_DISABLE);
1274 off_wrap = le16_to_cpu(snapshot.off_wrap);
1276 wrap_counter = off_wrap >> VRING_PACKED_EVENT_F_WRAP_CTR;
1277 event_idx = off_wrap & ~(1 << VRING_PACKED_EVENT_F_WRAP_CTR);
1278 if (wrap_counter != vq->packed.avail_wrap_counter)
1279 event_idx -= vq->packed.vring.num;
1281 needs_kick = vring_need_event(event_idx, new, old);
1287 static void detach_buf_packed(struct vring_virtqueue *vq,
1288 unsigned int id, void **ctx)
1290 struct vring_desc_state_packed *state = NULL;
1291 struct vring_packed_desc *desc;
1292 unsigned int i, curr;
1294 state = &vq->packed.desc_state[id];
1296 /* Clear data ptr. */
1299 vq->packed.desc_state[state->last].next = vq->free_head;
1301 vq->vq.num_free += state->num;
1303 if (unlikely(vq->use_dma_api)) {
1305 for (i = 0; i < state->num; i++) {
1306 vring_unmap_state_packed(vq,
1307 &vq->packed.desc_extra[curr]);
1308 curr = vq->packed.desc_state[curr].next;
1315 /* Free the indirect table, if any, now that it's unmapped. */
1316 desc = state->indir_desc;
1320 if (vq->use_dma_api) {
1321 len = vq->packed.desc_extra[id].len;
1322 for (i = 0; i < len / sizeof(struct vring_packed_desc);
1324 vring_unmap_desc_packed(vq, &desc[i]);
1327 state->indir_desc = NULL;
1329 *ctx = state->indir_desc;
1333 static inline bool is_used_desc_packed(const struct vring_virtqueue *vq,
1334 u16 idx, bool used_wrap_counter)
1339 flags = le16_to_cpu(vq->packed.vring.desc[idx].flags);
1340 avail = !!(flags & (1 << VRING_PACKED_DESC_F_AVAIL));
1341 used = !!(flags & (1 << VRING_PACKED_DESC_F_USED));
1343 return avail == used && used == used_wrap_counter;
1346 static inline bool more_used_packed(const struct vring_virtqueue *vq)
1348 return is_used_desc_packed(vq, vq->last_used_idx,
1349 vq->packed.used_wrap_counter);
1352 static void *virtqueue_get_buf_ctx_packed(struct virtqueue *_vq,
1356 struct vring_virtqueue *vq = to_vvq(_vq);
1362 if (unlikely(vq->broken)) {
1367 if (!more_used_packed(vq)) {
1368 pr_debug("No more buffers in queue\n");
1373 /* Only get used elements after they have been exposed by host. */
1374 virtio_rmb(vq->weak_barriers);
1376 last_used = vq->last_used_idx;
1377 id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
1378 *len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
1380 if (unlikely(id >= vq->packed.vring.num)) {
1381 BAD_RING(vq, "id %u out of range\n", id);
1384 if (unlikely(!vq->packed.desc_state[id].data)) {
1385 BAD_RING(vq, "id %u is not a head!\n", id);
1389 /* detach_buf_packed clears data, so grab it now. */
1390 ret = vq->packed.desc_state[id].data;
1391 detach_buf_packed(vq, id, ctx);
1393 vq->last_used_idx += vq->packed.desc_state[id].num;
1394 if (unlikely(vq->last_used_idx >= vq->packed.vring.num)) {
1395 vq->last_used_idx -= vq->packed.vring.num;
1396 vq->packed.used_wrap_counter ^= 1;
1400 * If we expect an interrupt for the next entry, tell host
1401 * by writing event index and flush out the write before
1402 * the read in the next get_buf call.
1404 if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC)
1405 virtio_store_mb(vq->weak_barriers,
1406 &vq->packed.vring.driver->off_wrap,
1407 cpu_to_le16(vq->last_used_idx |
1408 (vq->packed.used_wrap_counter <<
1409 VRING_PACKED_EVENT_F_WRAP_CTR)));
1411 LAST_ADD_TIME_INVALID(vq);
1417 static void virtqueue_disable_cb_packed(struct virtqueue *_vq)
1419 struct vring_virtqueue *vq = to_vvq(_vq);
1421 if (vq->packed.event_flags_shadow != VRING_PACKED_EVENT_FLAG_DISABLE) {
1422 vq->packed.event_flags_shadow = VRING_PACKED_EVENT_FLAG_DISABLE;
1423 vq->packed.vring.driver->flags =
1424 cpu_to_le16(vq->packed.event_flags_shadow);
1428 static unsigned virtqueue_enable_cb_prepare_packed(struct virtqueue *_vq)
1430 struct vring_virtqueue *vq = to_vvq(_vq);
1435 * We optimistically turn back on interrupts, then check if there was
1440 vq->packed.vring.driver->off_wrap =
1441 cpu_to_le16(vq->last_used_idx |
1442 (vq->packed.used_wrap_counter <<
1443 VRING_PACKED_EVENT_F_WRAP_CTR));
1445 * We need to update event offset and event wrap
1446 * counter first before updating event flags.
1448 virtio_wmb(vq->weak_barriers);
1451 if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DISABLE) {
1452 vq->packed.event_flags_shadow = vq->event ?
1453 VRING_PACKED_EVENT_FLAG_DESC :
1454 VRING_PACKED_EVENT_FLAG_ENABLE;
1455 vq->packed.vring.driver->flags =
1456 cpu_to_le16(vq->packed.event_flags_shadow);
1460 return vq->last_used_idx | ((u16)vq->packed.used_wrap_counter <<
1461 VRING_PACKED_EVENT_F_WRAP_CTR);
1464 static bool virtqueue_poll_packed(struct virtqueue *_vq, u16 off_wrap)
1466 struct vring_virtqueue *vq = to_vvq(_vq);
1470 wrap_counter = off_wrap >> VRING_PACKED_EVENT_F_WRAP_CTR;
1471 used_idx = off_wrap & ~(1 << VRING_PACKED_EVENT_F_WRAP_CTR);
1473 return is_used_desc_packed(vq, used_idx, wrap_counter);
1476 static bool virtqueue_enable_cb_delayed_packed(struct virtqueue *_vq)
1478 struct vring_virtqueue *vq = to_vvq(_vq);
1479 u16 used_idx, wrap_counter;
1485 * We optimistically turn back on interrupts, then check if there was
1490 /* TODO: tune this threshold */
1491 bufs = (vq->packed.vring.num - vq->vq.num_free) * 3 / 4;
1492 wrap_counter = vq->packed.used_wrap_counter;
1494 used_idx = vq->last_used_idx + bufs;
1495 if (used_idx >= vq->packed.vring.num) {
1496 used_idx -= vq->packed.vring.num;
1500 vq->packed.vring.driver->off_wrap = cpu_to_le16(used_idx |
1501 (wrap_counter << VRING_PACKED_EVENT_F_WRAP_CTR));
1504 * We need to update event offset and event wrap
1505 * counter first before updating event flags.
1507 virtio_wmb(vq->weak_barriers);
1509 used_idx = vq->last_used_idx;
1510 wrap_counter = vq->packed.used_wrap_counter;
1513 if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DISABLE) {
1514 vq->packed.event_flags_shadow = vq->event ?
1515 VRING_PACKED_EVENT_FLAG_DESC :
1516 VRING_PACKED_EVENT_FLAG_ENABLE;
1517 vq->packed.vring.driver->flags =
1518 cpu_to_le16(vq->packed.event_flags_shadow);
1522 * We need to update event suppression structure first
1523 * before re-checking for more used buffers.
1525 virtio_mb(vq->weak_barriers);
1527 if (is_used_desc_packed(vq, used_idx, wrap_counter)) {
1536 static void *virtqueue_detach_unused_buf_packed(struct virtqueue *_vq)
1538 struct vring_virtqueue *vq = to_vvq(_vq);
1544 for (i = 0; i < vq->packed.vring.num; i++) {
1545 if (!vq->packed.desc_state[i].data)
1547 /* detach_buf clears data, so grab it now. */
1548 buf = vq->packed.desc_state[i].data;
1549 detach_buf_packed(vq, i, NULL);
1553 /* That should have freed everything. */
1554 BUG_ON(vq->vq.num_free != vq->packed.vring.num);
1560 static struct virtqueue *vring_create_virtqueue_packed(
1563 unsigned int vring_align,
1564 struct virtio_device *vdev,
1566 bool may_reduce_num,
1568 bool (*notify)(struct virtqueue *),
1569 void (*callback)(struct virtqueue *),
1572 struct vring_virtqueue *vq;
1573 struct vring_packed_desc *ring;
1574 struct vring_packed_desc_event *driver, *device;
1575 dma_addr_t ring_dma_addr, driver_event_dma_addr, device_event_dma_addr;
1576 size_t ring_size_in_bytes, event_size_in_bytes;
1579 ring_size_in_bytes = num * sizeof(struct vring_packed_desc);
1581 ring = vring_alloc_queue(vdev, ring_size_in_bytes,
1583 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1587 event_size_in_bytes = sizeof(struct vring_packed_desc_event);
1589 driver = vring_alloc_queue(vdev, event_size_in_bytes,
1590 &driver_event_dma_addr,
1591 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1595 device = vring_alloc_queue(vdev, event_size_in_bytes,
1596 &device_event_dma_addr,
1597 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1601 vq = kmalloc(sizeof(*vq), GFP_KERNEL);
1605 vq->vq.callback = callback;
1608 vq->vq.num_free = num;
1609 vq->vq.index = index;
1610 vq->we_own_ring = true;
1611 vq->notify = notify;
1612 vq->weak_barriers = weak_barriers;
1614 vq->last_used_idx = 0;
1616 vq->packed_ring = true;
1617 vq->use_dma_api = vring_use_dma_api(vdev);
1618 list_add_tail(&vq->vq.list, &vdev->vqs);
1621 vq->last_add_time_valid = false;
1624 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
1626 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
1628 if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
1629 vq->weak_barriers = false;
1631 vq->packed.ring_dma_addr = ring_dma_addr;
1632 vq->packed.driver_event_dma_addr = driver_event_dma_addr;
1633 vq->packed.device_event_dma_addr = device_event_dma_addr;
1635 vq->packed.ring_size_in_bytes = ring_size_in_bytes;
1636 vq->packed.event_size_in_bytes = event_size_in_bytes;
1638 vq->packed.vring.num = num;
1639 vq->packed.vring.desc = ring;
1640 vq->packed.vring.driver = driver;
1641 vq->packed.vring.device = device;
1643 vq->packed.next_avail_idx = 0;
1644 vq->packed.avail_wrap_counter = 1;
1645 vq->packed.used_wrap_counter = 1;
1646 vq->packed.event_flags_shadow = 0;
1647 vq->packed.avail_used_flags = 1 << VRING_PACKED_DESC_F_AVAIL;
1649 vq->packed.desc_state = kmalloc_array(num,
1650 sizeof(struct vring_desc_state_packed),
1652 if (!vq->packed.desc_state)
1653 goto err_desc_state;
1655 memset(vq->packed.desc_state, 0,
1656 num * sizeof(struct vring_desc_state_packed));
1658 /* Put everything in free lists. */
1660 for (i = 0; i < num-1; i++)
1661 vq->packed.desc_state[i].next = i + 1;
1663 vq->packed.desc_extra = kmalloc_array(num,
1664 sizeof(struct vring_desc_extra_packed),
1666 if (!vq->packed.desc_extra)
1667 goto err_desc_extra;
1669 memset(vq->packed.desc_extra, 0,
1670 num * sizeof(struct vring_desc_extra_packed));
1672 /* No callback? Tell other side not to bother us. */
1674 vq->packed.event_flags_shadow = VRING_PACKED_EVENT_FLAG_DISABLE;
1675 vq->packed.vring.driver->flags =
1676 cpu_to_le16(vq->packed.event_flags_shadow);
1682 kfree(vq->packed.desc_state);
1686 vring_free_queue(vdev, event_size_in_bytes, device, ring_dma_addr);
1688 vring_free_queue(vdev, event_size_in_bytes, driver, ring_dma_addr);
1690 vring_free_queue(vdev, ring_size_in_bytes, ring, ring_dma_addr);
1697 * Generic functions and exported symbols.
1700 static inline int virtqueue_add(struct virtqueue *_vq,
1701 struct scatterlist *sgs[],
1702 unsigned int total_sg,
1703 unsigned int out_sgs,
1704 unsigned int in_sgs,
1709 struct vring_virtqueue *vq = to_vvq(_vq);
1711 return vq->packed_ring ? virtqueue_add_packed(_vq, sgs, total_sg,
1712 out_sgs, in_sgs, data, ctx, gfp) :
1713 virtqueue_add_split(_vq, sgs, total_sg,
1714 out_sgs, in_sgs, data, ctx, gfp);
1718 * virtqueue_add_sgs - expose buffers to other end
1719 * @vq: the struct virtqueue we're talking about.
1720 * @sgs: array of terminated scatterlists.
1721 * @out_num: the number of scatterlists readable by other side
1722 * @in_num: the number of scatterlists which are writable (after readable ones)
1723 * @data: the token identifying the buffer.
1724 * @gfp: how to do memory allocations (if necessary).
1726 * Caller must ensure we don't call this with other virtqueue operations
1727 * at the same time (except where noted).
1729 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1731 int virtqueue_add_sgs(struct virtqueue *_vq,
1732 struct scatterlist *sgs[],
1733 unsigned int out_sgs,
1734 unsigned int in_sgs,
1738 unsigned int i, total_sg = 0;
1740 /* Count them first. */
1741 for (i = 0; i < out_sgs + in_sgs; i++) {
1742 struct scatterlist *sg;
1744 for (sg = sgs[i]; sg; sg = sg_next(sg))
1747 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs,
1750 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
1753 * virtqueue_add_outbuf - expose output buffers to other end
1754 * @vq: the struct virtqueue we're talking about.
1755 * @sg: scatterlist (must be well-formed and terminated!)
1756 * @num: the number of entries in @sg readable by other side
1757 * @data: the token identifying the buffer.
1758 * @gfp: how to do memory allocations (if necessary).
1760 * Caller must ensure we don't call this with other virtqueue operations
1761 * at the same time (except where noted).
1763 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1765 int virtqueue_add_outbuf(struct virtqueue *vq,
1766 struct scatterlist *sg, unsigned int num,
1770 return virtqueue_add(vq, &sg, num, 1, 0, data, NULL, gfp);
1772 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
1775 * virtqueue_add_inbuf - expose input buffers to other end
1776 * @vq: the struct virtqueue we're talking about.
1777 * @sg: scatterlist (must be well-formed and terminated!)
1778 * @num: the number of entries in @sg writable by other side
1779 * @data: the token identifying the buffer.
1780 * @gfp: how to do memory allocations (if necessary).
1782 * Caller must ensure we don't call this with other virtqueue operations
1783 * at the same time (except where noted).
1785 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1787 int virtqueue_add_inbuf(struct virtqueue *vq,
1788 struct scatterlist *sg, unsigned int num,
1792 return virtqueue_add(vq, &sg, num, 0, 1, data, NULL, gfp);
1794 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
1797 * virtqueue_add_inbuf_ctx - expose input buffers to other end
1798 * @vq: the struct virtqueue we're talking about.
1799 * @sg: scatterlist (must be well-formed and terminated!)
1800 * @num: the number of entries in @sg writable by other side
1801 * @data: the token identifying the buffer.
1802 * @ctx: extra context for the token
1803 * @gfp: how to do memory allocations (if necessary).
1805 * Caller must ensure we don't call this with other virtqueue operations
1806 * at the same time (except where noted).
1808 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1810 int virtqueue_add_inbuf_ctx(struct virtqueue *vq,
1811 struct scatterlist *sg, unsigned int num,
1816 return virtqueue_add(vq, &sg, num, 0, 1, data, ctx, gfp);
1818 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf_ctx);
1821 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
1822 * @vq: the struct virtqueue
1824 * Instead of virtqueue_kick(), you can do:
1825 * if (virtqueue_kick_prepare(vq))
1826 * virtqueue_notify(vq);
1828 * This is sometimes useful because the virtqueue_kick_prepare() needs
1829 * to be serialized, but the actual virtqueue_notify() call does not.
1831 bool virtqueue_kick_prepare(struct virtqueue *_vq)
1833 struct vring_virtqueue *vq = to_vvq(_vq);
1835 return vq->packed_ring ? virtqueue_kick_prepare_packed(_vq) :
1836 virtqueue_kick_prepare_split(_vq);
1838 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
1841 * virtqueue_notify - second half of split virtqueue_kick call.
1842 * @vq: the struct virtqueue
1844 * This does not need to be serialized.
1846 * Returns false if host notify failed or queue is broken, otherwise true.
1848 bool virtqueue_notify(struct virtqueue *_vq)
1850 struct vring_virtqueue *vq = to_vvq(_vq);
1852 if (unlikely(vq->broken))
1855 /* Prod other side to tell it about changes. */
1856 if (!vq->notify(_vq)) {
1862 EXPORT_SYMBOL_GPL(virtqueue_notify);
1865 * virtqueue_kick - update after add_buf
1866 * @vq: the struct virtqueue
1868 * After one or more virtqueue_add_* calls, invoke this to kick
1871 * Caller must ensure we don't call this with other virtqueue
1872 * operations at the same time (except where noted).
1874 * Returns false if kick failed, otherwise true.
1876 bool virtqueue_kick(struct virtqueue *vq)
1878 if (virtqueue_kick_prepare(vq))
1879 return virtqueue_notify(vq);
1882 EXPORT_SYMBOL_GPL(virtqueue_kick);
1885 * virtqueue_get_buf - get the next used buffer
1886 * @vq: the struct virtqueue we're talking about.
1887 * @len: the length written into the buffer
1889 * If the device wrote data into the buffer, @len will be set to the
1890 * amount written. This means you don't need to clear the buffer
1891 * beforehand to ensure there's no data leakage in the case of short
1894 * Caller must ensure we don't call this with other virtqueue
1895 * operations at the same time (except where noted).
1897 * Returns NULL if there are no used buffers, or the "data" token
1898 * handed to virtqueue_add_*().
1900 void *virtqueue_get_buf_ctx(struct virtqueue *_vq, unsigned int *len,
1903 struct vring_virtqueue *vq = to_vvq(_vq);
1905 return vq->packed_ring ? virtqueue_get_buf_ctx_packed(_vq, len, ctx) :
1906 virtqueue_get_buf_ctx_split(_vq, len, ctx);
1908 EXPORT_SYMBOL_GPL(virtqueue_get_buf_ctx);
1910 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
1912 return virtqueue_get_buf_ctx(_vq, len, NULL);
1914 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
1916 * virtqueue_disable_cb - disable callbacks
1917 * @vq: the struct virtqueue we're talking about.
1919 * Note that this is not necessarily synchronous, hence unreliable and only
1920 * useful as an optimization.
1922 * Unlike other operations, this need not be serialized.
1924 void virtqueue_disable_cb(struct virtqueue *_vq)
1926 struct vring_virtqueue *vq = to_vvq(_vq);
1928 if (vq->packed_ring)
1929 virtqueue_disable_cb_packed(_vq);
1931 virtqueue_disable_cb_split(_vq);
1933 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
1936 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
1937 * @vq: the struct virtqueue we're talking about.
1939 * This re-enables callbacks; it returns current queue state
1940 * in an opaque unsigned value. This value should be later tested by
1941 * virtqueue_poll, to detect a possible race between the driver checking for
1942 * more work, and enabling callbacks.
1944 * Caller must ensure we don't call this with other virtqueue
1945 * operations at the same time (except where noted).
1947 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
1949 struct vring_virtqueue *vq = to_vvq(_vq);
1951 return vq->packed_ring ? virtqueue_enable_cb_prepare_packed(_vq) :
1952 virtqueue_enable_cb_prepare_split(_vq);
1954 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
1957 * virtqueue_poll - query pending used buffers
1958 * @vq: the struct virtqueue we're talking about.
1959 * @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
1961 * Returns "true" if there are pending used buffers in the queue.
1963 * This does not need to be serialized.
1965 bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
1967 struct vring_virtqueue *vq = to_vvq(_vq);
1969 virtio_mb(vq->weak_barriers);
1970 return vq->packed_ring ? virtqueue_poll_packed(_vq, last_used_idx) :
1971 virtqueue_poll_split(_vq, last_used_idx);
1973 EXPORT_SYMBOL_GPL(virtqueue_poll);
1976 * virtqueue_enable_cb - restart callbacks after disable_cb.
1977 * @vq: the struct virtqueue we're talking about.
1979 * This re-enables callbacks; it returns "false" if there are pending
1980 * buffers in the queue, to detect a possible race between the driver
1981 * checking for more work, and enabling callbacks.
1983 * Caller must ensure we don't call this with other virtqueue
1984 * operations at the same time (except where noted).
1986 bool virtqueue_enable_cb(struct virtqueue *_vq)
1988 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
1990 return !virtqueue_poll(_vq, last_used_idx);
1992 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
1995 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
1996 * @vq: the struct virtqueue we're talking about.
1998 * This re-enables callbacks but hints to the other side to delay
1999 * interrupts until most of the available buffers have been processed;
2000 * it returns "false" if there are many pending buffers in the queue,
2001 * to detect a possible race between the driver checking for more work,
2002 * and enabling callbacks.
2004 * Caller must ensure we don't call this with other virtqueue
2005 * operations at the same time (except where noted).
2007 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
2009 struct vring_virtqueue *vq = to_vvq(_vq);
2011 return vq->packed_ring ? virtqueue_enable_cb_delayed_packed(_vq) :
2012 virtqueue_enable_cb_delayed_split(_vq);
2014 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
2017 * virtqueue_detach_unused_buf - detach first unused buffer
2018 * @vq: the struct virtqueue we're talking about.
2020 * Returns NULL or the "data" token handed to virtqueue_add_*().
2021 * This is not valid on an active queue; it is useful only for device
2024 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
2026 struct vring_virtqueue *vq = to_vvq(_vq);
2028 return vq->packed_ring ? virtqueue_detach_unused_buf_packed(_vq) :
2029 virtqueue_detach_unused_buf_split(_vq);
2031 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
2033 static inline bool more_used(const struct vring_virtqueue *vq)
2035 return vq->packed_ring ? more_used_packed(vq) : more_used_split(vq);
2038 irqreturn_t vring_interrupt(int irq, void *_vq)
2040 struct vring_virtqueue *vq = to_vvq(_vq);
2042 if (!more_used(vq)) {
2043 pr_debug("virtqueue interrupt with no work for %p\n", vq);
2047 if (unlikely(vq->broken))
2050 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
2051 if (vq->vq.callback)
2052 vq->vq.callback(&vq->vq);
2056 EXPORT_SYMBOL_GPL(vring_interrupt);
2058 /* Only available for split ring */
2059 struct virtqueue *__vring_new_virtqueue(unsigned int index,
2061 struct virtio_device *vdev,
2064 bool (*notify)(struct virtqueue *),
2065 void (*callback)(struct virtqueue *),
2069 struct vring_virtqueue *vq;
2071 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED))
2074 vq = kmalloc(sizeof(*vq), GFP_KERNEL);
2078 vq->packed_ring = false;
2079 vq->vq.callback = callback;
2082 vq->vq.num_free = vring.num;
2083 vq->vq.index = index;
2084 vq->we_own_ring = false;
2085 vq->notify = notify;
2086 vq->weak_barriers = weak_barriers;
2088 vq->last_used_idx = 0;
2090 vq->use_dma_api = vring_use_dma_api(vdev);
2091 list_add_tail(&vq->vq.list, &vdev->vqs);
2094 vq->last_add_time_valid = false;
2097 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
2099 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
2101 if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
2102 vq->weak_barriers = false;
2104 vq->split.queue_dma_addr = 0;
2105 vq->split.queue_size_in_bytes = 0;
2107 vq->split.vring = vring;
2108 vq->split.avail_flags_shadow = 0;
2109 vq->split.avail_idx_shadow = 0;
2111 /* No callback? Tell other side not to bother us. */
2113 vq->split.avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
2115 vq->split.vring.avail->flags = cpu_to_virtio16(vdev,
2116 vq->split.avail_flags_shadow);
2119 vq->split.desc_state = kmalloc_array(vring.num,
2120 sizeof(struct vring_desc_state_split), GFP_KERNEL);
2121 if (!vq->split.desc_state) {
2126 /* Put everything in free lists. */
2128 for (i = 0; i < vring.num-1; i++)
2129 vq->split.vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
2130 memset(vq->split.desc_state, 0, vring.num *
2131 sizeof(struct vring_desc_state_split));
2135 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
2137 struct virtqueue *vring_create_virtqueue(
2140 unsigned int vring_align,
2141 struct virtio_device *vdev,
2143 bool may_reduce_num,
2145 bool (*notify)(struct virtqueue *),
2146 void (*callback)(struct virtqueue *),
2150 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED))
2151 return vring_create_virtqueue_packed(index, num, vring_align,
2152 vdev, weak_barriers, may_reduce_num,
2153 context, notify, callback, name);
2155 return vring_create_virtqueue_split(index, num, vring_align,
2156 vdev, weak_barriers, may_reduce_num,
2157 context, notify, callback, name);
2159 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
2161 /* Only available for split ring */
2162 struct virtqueue *vring_new_virtqueue(unsigned int index,
2164 unsigned int vring_align,
2165 struct virtio_device *vdev,
2169 bool (*notify)(struct virtqueue *vq),
2170 void (*callback)(struct virtqueue *vq),
2175 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED))
2178 vring_init(&vring, num, pages, vring_align);
2179 return __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
2180 notify, callback, name);
2182 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
2184 void vring_del_virtqueue(struct virtqueue *_vq)
2186 struct vring_virtqueue *vq = to_vvq(_vq);
2188 if (vq->we_own_ring) {
2189 if (vq->packed_ring) {
2190 vring_free_queue(vq->vq.vdev,
2191 vq->packed.ring_size_in_bytes,
2192 vq->packed.vring.desc,
2193 vq->packed.ring_dma_addr);
2195 vring_free_queue(vq->vq.vdev,
2196 vq->packed.event_size_in_bytes,
2197 vq->packed.vring.driver,
2198 vq->packed.driver_event_dma_addr);
2200 vring_free_queue(vq->vq.vdev,
2201 vq->packed.event_size_in_bytes,
2202 vq->packed.vring.device,
2203 vq->packed.device_event_dma_addr);
2205 kfree(vq->packed.desc_state);
2206 kfree(vq->packed.desc_extra);
2208 vring_free_queue(vq->vq.vdev,
2209 vq->split.queue_size_in_bytes,
2210 vq->split.vring.desc,
2211 vq->split.queue_dma_addr);
2213 kfree(vq->split.desc_state);
2216 list_del(&_vq->list);
2219 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
2221 /* Manipulates transport-specific feature bits. */
2222 void vring_transport_features(struct virtio_device *vdev)
2226 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
2228 case VIRTIO_RING_F_INDIRECT_DESC:
2230 case VIRTIO_RING_F_EVENT_IDX:
2232 case VIRTIO_F_VERSION_1:
2234 case VIRTIO_F_IOMMU_PLATFORM:
2236 case VIRTIO_F_RING_PACKED:
2238 case VIRTIO_F_ORDER_PLATFORM:
2241 /* We don't understand this bit. */
2242 __virtio_clear_bit(vdev, i);
2246 EXPORT_SYMBOL_GPL(vring_transport_features);
2249 * virtqueue_get_vring_size - return the size of the virtqueue's vring
2250 * @vq: the struct virtqueue containing the vring of interest.
2252 * Returns the size of the vring. This is mainly used for boasting to
2253 * userspace. Unlike other operations, this need not be serialized.
2255 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
2258 struct vring_virtqueue *vq = to_vvq(_vq);
2260 return vq->packed_ring ? vq->packed.vring.num : vq->split.vring.num;
2262 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
2264 bool virtqueue_is_broken(struct virtqueue *_vq)
2266 struct vring_virtqueue *vq = to_vvq(_vq);
2270 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
2273 * This should prevent the device from being used, allowing drivers to
2274 * recover. You may need to grab appropriate locks to flush.
2276 void virtio_break_device(struct virtio_device *dev)
2278 struct virtqueue *_vq;
2280 list_for_each_entry(_vq, &dev->vqs, list) {
2281 struct vring_virtqueue *vq = to_vvq(_vq);
2285 EXPORT_SYMBOL_GPL(virtio_break_device);
2287 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
2289 struct vring_virtqueue *vq = to_vvq(_vq);
2291 BUG_ON(!vq->we_own_ring);
2293 if (vq->packed_ring)
2294 return vq->packed.ring_dma_addr;
2296 return vq->split.queue_dma_addr;
2298 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
2300 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
2302 struct vring_virtqueue *vq = to_vvq(_vq);
2304 BUG_ON(!vq->we_own_ring);
2306 if (vq->packed_ring)
2307 return vq->packed.driver_event_dma_addr;
2309 return vq->split.queue_dma_addr +
2310 ((char *)vq->split.vring.avail - (char *)vq->split.vring.desc);
2312 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
2314 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
2316 struct vring_virtqueue *vq = to_vvq(_vq);
2318 BUG_ON(!vq->we_own_ring);
2320 if (vq->packed_ring)
2321 return vq->packed.device_event_dma_addr;
2323 return vq->split.queue_dma_addr +
2324 ((char *)vq->split.vring.used - (char *)vq->split.vring.desc);
2326 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
2328 /* Only available for split ring */
2329 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
2331 return &to_vvq(vq)->split.vring;
2333 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
2335 MODULE_LICENSE("GPL");