1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* Virtio ring implementation.
4 * Copyright 2007 Rusty Russell IBM Corporation
6 #include <linux/virtio.h>
7 #include <linux/virtio_ring.h>
8 #include <linux/virtio_config.h>
9 #include <linux/device.h>
10 #include <linux/slab.h>
11 #include <linux/module.h>
12 #include <linux/hrtimer.h>
13 #include <linux/dma-mapping.h>
17 /* For development, we want to crash whenever the ring is screwed. */
18 #define BAD_RING(_vq, fmt, args...) \
20 dev_err(&(_vq)->vq.vdev->dev, \
21 "%s:"fmt, (_vq)->vq.name, ##args); \
24 /* Caller is supposed to guarantee no reentry. */
25 #define START_USE(_vq) \
28 panic("%s:in_use = %i\n", \
29 (_vq)->vq.name, (_vq)->in_use); \
30 (_vq)->in_use = __LINE__; \
32 #define END_USE(_vq) \
33 do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
34 #define LAST_ADD_TIME_UPDATE(_vq) \
36 ktime_t now = ktime_get(); \
38 /* No kick or get, with .1 second between? Warn. */ \
39 if ((_vq)->last_add_time_valid) \
40 WARN_ON(ktime_to_ms(ktime_sub(now, \
41 (_vq)->last_add_time)) > 100); \
42 (_vq)->last_add_time = now; \
43 (_vq)->last_add_time_valid = true; \
45 #define LAST_ADD_TIME_CHECK(_vq) \
47 if ((_vq)->last_add_time_valid) { \
48 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(), \
49 (_vq)->last_add_time)) > 100); \
52 #define LAST_ADD_TIME_INVALID(_vq) \
53 ((_vq)->last_add_time_valid = false)
55 #define BAD_RING(_vq, fmt, args...) \
57 dev_err(&_vq->vq.vdev->dev, \
58 "%s:"fmt, (_vq)->vq.name, ##args); \
59 (_vq)->broken = true; \
63 #define LAST_ADD_TIME_UPDATE(vq)
64 #define LAST_ADD_TIME_CHECK(vq)
65 #define LAST_ADD_TIME_INVALID(vq)
68 struct vring_desc_state_split {
69 void *data; /* Data for callback. */
70 struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
73 struct vring_desc_state_packed {
74 void *data; /* Data for callback. */
75 struct vring_packed_desc *indir_desc; /* Indirect descriptor, if any. */
76 u16 num; /* Descriptor list length. */
77 u16 next; /* The next desc state in a list. */
78 u16 last; /* The last desc state in a list. */
81 struct vring_desc_extra_packed {
82 dma_addr_t addr; /* Buffer DMA addr. */
83 u32 len; /* Buffer length. */
84 u16 flags; /* Descriptor flags. */
87 struct vring_virtqueue {
90 /* Is this a packed ring? */
93 /* Is DMA API used? */
96 /* Can we use weak barriers? */
99 /* Other side has made a mess, don't try any more. */
102 /* Host supports indirect buffers */
105 /* Host publishes avail event idx */
108 /* Head of free buffer list. */
109 unsigned int free_head;
110 /* Number we've added since last sync. */
111 unsigned int num_added;
113 /* Last used index we've seen. */
117 /* Available for split ring */
119 /* Actual memory layout for this queue. */
122 /* Last written value to avail->flags */
123 u16 avail_flags_shadow;
126 * Last written value to avail->idx in
129 u16 avail_idx_shadow;
131 /* Per-descriptor state. */
132 struct vring_desc_state_split *desc_state;
134 /* DMA address and size information */
135 dma_addr_t queue_dma_addr;
136 size_t queue_size_in_bytes;
139 /* Available for packed ring */
141 /* Actual memory layout for this queue. */
144 struct vring_packed_desc *desc;
145 struct vring_packed_desc_event *driver;
146 struct vring_packed_desc_event *device;
149 /* Driver ring wrap counter. */
150 bool avail_wrap_counter;
152 /* Device ring wrap counter. */
153 bool used_wrap_counter;
155 /* Avail used flags. */
156 u16 avail_used_flags;
158 /* Index of the next avail descriptor. */
162 * Last written value to driver->flags in
165 u16 event_flags_shadow;
167 /* Per-descriptor state. */
168 struct vring_desc_state_packed *desc_state;
169 struct vring_desc_extra_packed *desc_extra;
171 /* DMA address and size information */
172 dma_addr_t ring_dma_addr;
173 dma_addr_t driver_event_dma_addr;
174 dma_addr_t device_event_dma_addr;
175 size_t ring_size_in_bytes;
176 size_t event_size_in_bytes;
180 /* How to notify other side. FIXME: commonalize hcalls! */
181 bool (*notify)(struct virtqueue *vq);
183 /* DMA, allocation, and size information */
187 /* They're supposed to lock for us. */
190 /* Figure out if their kicks are too delayed. */
191 bool last_add_time_valid;
192 ktime_t last_add_time;
201 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
203 static inline bool virtqueue_use_indirect(struct virtqueue *_vq,
204 unsigned int total_sg)
206 struct vring_virtqueue *vq = to_vvq(_vq);
209 * If the host supports indirect descriptor tables, and we have multiple
210 * buffers, then go indirect. FIXME: tune this threshold
212 return (vq->indirect && total_sg > 1 && vq->vq.num_free);
216 * Modern virtio devices have feature bits to specify whether they need a
217 * quirk and bypass the IOMMU. If not there, just use the DMA API.
219 * If there, the interaction between virtio and DMA API is messy.
221 * On most systems with virtio, physical addresses match bus addresses,
222 * and it doesn't particularly matter whether we use the DMA API.
224 * On some systems, including Xen and any system with a physical device
225 * that speaks virtio behind a physical IOMMU, we must use the DMA API
226 * for virtio DMA to work at all.
228 * On other systems, including SPARC and PPC64, virtio-pci devices are
229 * enumerated as though they are behind an IOMMU, but the virtio host
230 * ignores the IOMMU, so we must either pretend that the IOMMU isn't
231 * there or somehow map everything as the identity.
233 * For the time being, we preserve historic behavior and bypass the DMA
236 * TODO: install a per-device DMA ops structure that does the right thing
237 * taking into account all the above quirks, and use the DMA API
238 * unconditionally on data path.
241 static bool vring_use_dma_api(struct virtio_device *vdev)
243 if (!virtio_has_iommu_quirk(vdev))
246 /* Otherwise, we are left to guess. */
248 * In theory, it's possible to have a buggy QEMU-supposed
249 * emulated Q35 IOMMU and Xen enabled at the same time. On
250 * such a configuration, virtio has never worked and will
251 * not work without an even larger kludge. Instead, enable
252 * the DMA API if we're a Xen guest, which at least allows
253 * all of the sensible Xen configurations to work correctly.
261 size_t virtio_max_dma_size(struct virtio_device *vdev)
263 size_t max_segment_size = SIZE_MAX;
265 if (vring_use_dma_api(vdev))
266 max_segment_size = dma_max_mapping_size(&vdev->dev);
268 return max_segment_size;
270 EXPORT_SYMBOL_GPL(virtio_max_dma_size);
272 static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
273 dma_addr_t *dma_handle, gfp_t flag)
275 if (vring_use_dma_api(vdev)) {
276 return dma_alloc_coherent(vdev->dev.parent, size,
279 void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
282 phys_addr_t phys_addr = virt_to_phys(queue);
283 *dma_handle = (dma_addr_t)phys_addr;
286 * Sanity check: make sure we dind't truncate
287 * the address. The only arches I can find that
288 * have 64-bit phys_addr_t but 32-bit dma_addr_t
289 * are certain non-highmem MIPS and x86
290 * configurations, but these configurations
291 * should never allocate physical pages above 32
292 * bits, so this is fine. Just in case, throw a
293 * warning and abort if we end up with an
294 * unrepresentable address.
296 if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
297 free_pages_exact(queue, PAGE_ALIGN(size));
305 static void vring_free_queue(struct virtio_device *vdev, size_t size,
306 void *queue, dma_addr_t dma_handle)
308 if (vring_use_dma_api(vdev))
309 dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
311 free_pages_exact(queue, PAGE_ALIGN(size));
315 * The DMA ops on various arches are rather gnarly right now, and
316 * making all of the arch DMA ops work on the vring device itself
317 * is a mess. For now, we use the parent device for DMA ops.
319 static inline struct device *vring_dma_dev(const struct vring_virtqueue *vq)
321 return vq->vq.vdev->dev.parent;
324 /* Map one sg entry. */
325 static dma_addr_t vring_map_one_sg(const struct vring_virtqueue *vq,
326 struct scatterlist *sg,
327 enum dma_data_direction direction)
329 if (!vq->use_dma_api)
330 return (dma_addr_t)sg_phys(sg);
333 * We can't use dma_map_sg, because we don't use scatterlists in
334 * the way it expects (we don't guarantee that the scatterlist
335 * will exist for the lifetime of the mapping).
337 return dma_map_page(vring_dma_dev(vq),
338 sg_page(sg), sg->offset, sg->length,
342 static dma_addr_t vring_map_single(const struct vring_virtqueue *vq,
343 void *cpu_addr, size_t size,
344 enum dma_data_direction direction)
346 if (!vq->use_dma_api)
347 return (dma_addr_t)virt_to_phys(cpu_addr);
349 return dma_map_single(vring_dma_dev(vq),
350 cpu_addr, size, direction);
353 static int vring_mapping_error(const struct vring_virtqueue *vq,
356 if (!vq->use_dma_api)
359 return dma_mapping_error(vring_dma_dev(vq), addr);
364 * Split ring specific functions - *_split().
367 static void vring_unmap_one_split(const struct vring_virtqueue *vq,
368 struct vring_desc *desc)
372 if (!vq->use_dma_api)
375 flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
377 if (flags & VRING_DESC_F_INDIRECT) {
378 dma_unmap_single(vring_dma_dev(vq),
379 virtio64_to_cpu(vq->vq.vdev, desc->addr),
380 virtio32_to_cpu(vq->vq.vdev, desc->len),
381 (flags & VRING_DESC_F_WRITE) ?
382 DMA_FROM_DEVICE : DMA_TO_DEVICE);
384 dma_unmap_page(vring_dma_dev(vq),
385 virtio64_to_cpu(vq->vq.vdev, desc->addr),
386 virtio32_to_cpu(vq->vq.vdev, desc->len),
387 (flags & VRING_DESC_F_WRITE) ?
388 DMA_FROM_DEVICE : DMA_TO_DEVICE);
392 static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
393 unsigned int total_sg,
396 struct vring_desc *desc;
400 * We require lowmem mappings for the descriptors because
401 * otherwise virt_to_phys will give us bogus addresses in the
404 gfp &= ~__GFP_HIGHMEM;
406 desc = kmalloc_array(total_sg, sizeof(struct vring_desc), gfp);
410 for (i = 0; i < total_sg; i++)
411 desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
415 static inline int virtqueue_add_split(struct virtqueue *_vq,
416 struct scatterlist *sgs[],
417 unsigned int total_sg,
418 unsigned int out_sgs,
424 struct vring_virtqueue *vq = to_vvq(_vq);
425 struct scatterlist *sg;
426 struct vring_desc *desc;
427 unsigned int i, n, avail, descs_used, uninitialized_var(prev), err_idx;
433 BUG_ON(data == NULL);
434 BUG_ON(ctx && vq->indirect);
436 if (unlikely(vq->broken)) {
441 LAST_ADD_TIME_UPDATE(vq);
443 BUG_ON(total_sg == 0);
445 head = vq->free_head;
447 if (virtqueue_use_indirect(_vq, total_sg))
448 desc = alloc_indirect_split(_vq, total_sg, gfp);
451 WARN_ON_ONCE(total_sg > vq->split.vring.num && !vq->indirect);
455 /* Use a single buffer which doesn't continue */
457 /* Set up rest to use this indirect table. */
462 desc = vq->split.vring.desc;
464 descs_used = total_sg;
467 if (vq->vq.num_free < descs_used) {
468 pr_debug("Can't add buf len %i - avail = %i\n",
469 descs_used, vq->vq.num_free);
470 /* FIXME: for historical reasons, we force a notify here if
471 * there are outgoing parts to the buffer. Presumably the
472 * host should service the ring ASAP. */
481 for (n = 0; n < out_sgs; n++) {
482 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
483 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
484 if (vring_mapping_error(vq, addr))
487 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
488 desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
489 desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
491 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
494 for (; n < (out_sgs + in_sgs); n++) {
495 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
496 dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
497 if (vring_mapping_error(vq, addr))
500 desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
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 /* Last one doesn't continue. */
508 desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
511 /* Now that the indirect table is filled in, map it. */
512 dma_addr_t addr = vring_map_single(
513 vq, desc, total_sg * sizeof(struct vring_desc),
515 if (vring_mapping_error(vq, addr))
518 vq->split.vring.desc[head].flags = cpu_to_virtio16(_vq->vdev,
519 VRING_DESC_F_INDIRECT);
520 vq->split.vring.desc[head].addr = cpu_to_virtio64(_vq->vdev,
523 vq->split.vring.desc[head].len = cpu_to_virtio32(_vq->vdev,
524 total_sg * sizeof(struct vring_desc));
527 /* We're using some buffers from the free list. */
528 vq->vq.num_free -= descs_used;
530 /* Update free pointer */
532 vq->free_head = virtio16_to_cpu(_vq->vdev,
533 vq->split.vring.desc[head].next);
537 /* Store token and indirect buffer state. */
538 vq->split.desc_state[head].data = data;
540 vq->split.desc_state[head].indir_desc = desc;
542 vq->split.desc_state[head].indir_desc = ctx;
544 /* Put entry in available array (but don't update avail->idx until they
546 avail = vq->split.avail_idx_shadow & (vq->split.vring.num - 1);
547 vq->split.vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
549 /* Descriptors and available array need to be set before we expose the
550 * new available array entries. */
551 virtio_wmb(vq->weak_barriers);
552 vq->split.avail_idx_shadow++;
553 vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
554 vq->split.avail_idx_shadow);
557 pr_debug("Added buffer head %i to %p\n", head, vq);
560 /* This is very unlikely, but theoretically possible. Kick
562 if (unlikely(vq->num_added == (1 << 16) - 1))
575 for (n = 0; n < total_sg; n++) {
578 vring_unmap_one_split(vq, &desc[i]);
579 i = virtio16_to_cpu(_vq->vdev, desc[i].next);
589 static bool virtqueue_kick_prepare_split(struct virtqueue *_vq)
591 struct vring_virtqueue *vq = to_vvq(_vq);
596 /* We need to expose available array entries before checking avail
598 virtio_mb(vq->weak_barriers);
600 old = vq->split.avail_idx_shadow - vq->num_added;
601 new = vq->split.avail_idx_shadow;
604 LAST_ADD_TIME_CHECK(vq);
605 LAST_ADD_TIME_INVALID(vq);
608 needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev,
609 vring_avail_event(&vq->split.vring)),
612 needs_kick = !(vq->split.vring.used->flags &
613 cpu_to_virtio16(_vq->vdev,
614 VRING_USED_F_NO_NOTIFY));
620 static void detach_buf_split(struct vring_virtqueue *vq, unsigned int head,
624 __virtio16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
626 /* Clear data ptr. */
627 vq->split.desc_state[head].data = NULL;
629 /* Put back on free list: unmap first-level descriptors and find end */
632 while (vq->split.vring.desc[i].flags & nextflag) {
633 vring_unmap_one_split(vq, &vq->split.vring.desc[i]);
634 i = virtio16_to_cpu(vq->vq.vdev, vq->split.vring.desc[i].next);
638 vring_unmap_one_split(vq, &vq->split.vring.desc[i]);
639 vq->split.vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev,
641 vq->free_head = head;
643 /* Plus final descriptor */
647 struct vring_desc *indir_desc =
648 vq->split.desc_state[head].indir_desc;
651 /* Free the indirect table, if any, now that it's unmapped. */
655 len = virtio32_to_cpu(vq->vq.vdev,
656 vq->split.vring.desc[head].len);
658 BUG_ON(!(vq->split.vring.desc[head].flags &
659 cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
660 BUG_ON(len == 0 || len % sizeof(struct vring_desc));
662 for (j = 0; j < len / sizeof(struct vring_desc); j++)
663 vring_unmap_one_split(vq, &indir_desc[j]);
666 vq->split.desc_state[head].indir_desc = NULL;
668 *ctx = vq->split.desc_state[head].indir_desc;
672 static inline bool more_used_split(const struct vring_virtqueue *vq)
674 return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev,
675 vq->split.vring.used->idx);
678 static void *virtqueue_get_buf_ctx_split(struct virtqueue *_vq,
682 struct vring_virtqueue *vq = to_vvq(_vq);
689 if (unlikely(vq->broken)) {
694 if (!more_used_split(vq)) {
695 pr_debug("No more buffers in queue\n");
700 /* Only get used array entries after they have been exposed by host. */
701 virtio_rmb(vq->weak_barriers);
703 last_used = (vq->last_used_idx & (vq->split.vring.num - 1));
704 i = virtio32_to_cpu(_vq->vdev,
705 vq->split.vring.used->ring[last_used].id);
706 *len = virtio32_to_cpu(_vq->vdev,
707 vq->split.vring.used->ring[last_used].len);
709 if (unlikely(i >= vq->split.vring.num)) {
710 BAD_RING(vq, "id %u out of range\n", i);
713 if (unlikely(!vq->split.desc_state[i].data)) {
714 BAD_RING(vq, "id %u is not a head!\n", i);
718 /* detach_buf_split clears data, so grab it now. */
719 ret = vq->split.desc_state[i].data;
720 detach_buf_split(vq, i, ctx);
722 /* If we expect an interrupt for the next entry, tell host
723 * by writing event index and flush out the write before
724 * the read in the next get_buf call. */
725 if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
726 virtio_store_mb(vq->weak_barriers,
727 &vring_used_event(&vq->split.vring),
728 cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
730 LAST_ADD_TIME_INVALID(vq);
736 static void virtqueue_disable_cb_split(struct virtqueue *_vq)
738 struct vring_virtqueue *vq = to_vvq(_vq);
740 if (!(vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
741 vq->split.avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
743 vq->split.vring.avail->flags =
744 cpu_to_virtio16(_vq->vdev,
745 vq->split.avail_flags_shadow);
749 static unsigned virtqueue_enable_cb_prepare_split(struct virtqueue *_vq)
751 struct vring_virtqueue *vq = to_vvq(_vq);
756 /* We optimistically turn back on interrupts, then check if there was
758 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
759 * either clear the flags bit or point the event index at the next
760 * entry. Always do both to keep code simple. */
761 if (vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
762 vq->split.avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
764 vq->split.vring.avail->flags =
765 cpu_to_virtio16(_vq->vdev,
766 vq->split.avail_flags_shadow);
768 vring_used_event(&vq->split.vring) = cpu_to_virtio16(_vq->vdev,
769 last_used_idx = vq->last_used_idx);
771 return last_used_idx;
774 static bool virtqueue_poll_split(struct virtqueue *_vq, unsigned last_used_idx)
776 struct vring_virtqueue *vq = to_vvq(_vq);
778 return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev,
779 vq->split.vring.used->idx);
782 static bool virtqueue_enable_cb_delayed_split(struct virtqueue *_vq)
784 struct vring_virtqueue *vq = to_vvq(_vq);
789 /* We optimistically turn back on interrupts, then check if there was
791 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
792 * either clear the flags bit or point the event index at the next
793 * entry. Always update the event index to keep code simple. */
794 if (vq->split.avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
795 vq->split.avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
797 vq->split.vring.avail->flags =
798 cpu_to_virtio16(_vq->vdev,
799 vq->split.avail_flags_shadow);
801 /* TODO: tune this threshold */
802 bufs = (u16)(vq->split.avail_idx_shadow - vq->last_used_idx) * 3 / 4;
804 virtio_store_mb(vq->weak_barriers,
805 &vring_used_event(&vq->split.vring),
806 cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
808 if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->split.vring.used->idx)
809 - vq->last_used_idx) > bufs)) {
818 static void *virtqueue_detach_unused_buf_split(struct virtqueue *_vq)
820 struct vring_virtqueue *vq = to_vvq(_vq);
826 for (i = 0; i < vq->split.vring.num; i++) {
827 if (!vq->split.desc_state[i].data)
829 /* detach_buf_split clears data, so grab it now. */
830 buf = vq->split.desc_state[i].data;
831 detach_buf_split(vq, i, NULL);
832 vq->split.avail_idx_shadow--;
833 vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
834 vq->split.avail_idx_shadow);
838 /* That should have freed everything. */
839 BUG_ON(vq->vq.num_free != vq->split.vring.num);
845 static struct virtqueue *vring_create_virtqueue_split(
848 unsigned int vring_align,
849 struct virtio_device *vdev,
853 bool (*notify)(struct virtqueue *),
854 void (*callback)(struct virtqueue *),
857 struct virtqueue *vq;
860 size_t queue_size_in_bytes;
863 /* We assume num is a power of 2. */
864 if (num & (num - 1)) {
865 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
869 /* TODO: allocate each queue chunk individually */
870 for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
871 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
873 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
884 /* Try to get a single page. You are my only hope! */
885 queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
886 &dma_addr, GFP_KERNEL|__GFP_ZERO);
891 queue_size_in_bytes = vring_size(num, vring_align);
892 vring_init(&vring, num, queue, vring_align);
894 vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
895 notify, callback, name);
897 vring_free_queue(vdev, queue_size_in_bytes, queue,
902 to_vvq(vq)->split.queue_dma_addr = dma_addr;
903 to_vvq(vq)->split.queue_size_in_bytes = queue_size_in_bytes;
904 to_vvq(vq)->we_own_ring = true;
911 * Packed ring specific functions - *_packed().
914 static void vring_unmap_state_packed(const struct vring_virtqueue *vq,
915 struct vring_desc_extra_packed *state)
919 if (!vq->use_dma_api)
922 flags = state->flags;
924 if (flags & VRING_DESC_F_INDIRECT) {
925 dma_unmap_single(vring_dma_dev(vq),
926 state->addr, state->len,
927 (flags & VRING_DESC_F_WRITE) ?
928 DMA_FROM_DEVICE : DMA_TO_DEVICE);
930 dma_unmap_page(vring_dma_dev(vq),
931 state->addr, state->len,
932 (flags & VRING_DESC_F_WRITE) ?
933 DMA_FROM_DEVICE : DMA_TO_DEVICE);
937 static void vring_unmap_desc_packed(const struct vring_virtqueue *vq,
938 struct vring_packed_desc *desc)
942 if (!vq->use_dma_api)
945 flags = le16_to_cpu(desc->flags);
947 if (flags & VRING_DESC_F_INDIRECT) {
948 dma_unmap_single(vring_dma_dev(vq),
949 le64_to_cpu(desc->addr),
950 le32_to_cpu(desc->len),
951 (flags & VRING_DESC_F_WRITE) ?
952 DMA_FROM_DEVICE : DMA_TO_DEVICE);
954 dma_unmap_page(vring_dma_dev(vq),
955 le64_to_cpu(desc->addr),
956 le32_to_cpu(desc->len),
957 (flags & VRING_DESC_F_WRITE) ?
958 DMA_FROM_DEVICE : DMA_TO_DEVICE);
962 static struct vring_packed_desc *alloc_indirect_packed(unsigned int total_sg,
965 struct vring_packed_desc *desc;
968 * We require lowmem mappings for the descriptors because
969 * otherwise virt_to_phys will give us bogus addresses in the
972 gfp &= ~__GFP_HIGHMEM;
974 desc = kmalloc_array(total_sg, sizeof(struct vring_packed_desc), gfp);
979 static int virtqueue_add_indirect_packed(struct vring_virtqueue *vq,
980 struct scatterlist *sgs[],
981 unsigned int total_sg,
982 unsigned int out_sgs,
987 struct vring_packed_desc *desc;
988 struct scatterlist *sg;
989 unsigned int i, n, err_idx;
993 head = vq->packed.next_avail_idx;
994 desc = alloc_indirect_packed(total_sg, gfp);
996 if (unlikely(vq->vq.num_free < 1)) {
997 pr_debug("Can't add buf len 1 - avail = 0\n");
1005 BUG_ON(id == vq->packed.vring.num);
1007 for (n = 0; n < out_sgs + in_sgs; n++) {
1008 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
1009 addr = vring_map_one_sg(vq, sg, n < out_sgs ?
1010 DMA_TO_DEVICE : DMA_FROM_DEVICE);
1011 if (vring_mapping_error(vq, addr))
1014 desc[i].flags = cpu_to_le16(n < out_sgs ?
1015 0 : VRING_DESC_F_WRITE);
1016 desc[i].addr = cpu_to_le64(addr);
1017 desc[i].len = cpu_to_le32(sg->length);
1022 /* Now that the indirect table is filled in, map it. */
1023 addr = vring_map_single(vq, desc,
1024 total_sg * sizeof(struct vring_packed_desc),
1026 if (vring_mapping_error(vq, addr))
1029 vq->packed.vring.desc[head].addr = cpu_to_le64(addr);
1030 vq->packed.vring.desc[head].len = cpu_to_le32(total_sg *
1031 sizeof(struct vring_packed_desc));
1032 vq->packed.vring.desc[head].id = cpu_to_le16(id);
1034 if (vq->use_dma_api) {
1035 vq->packed.desc_extra[id].addr = addr;
1036 vq->packed.desc_extra[id].len = total_sg *
1037 sizeof(struct vring_packed_desc);
1038 vq->packed.desc_extra[id].flags = VRING_DESC_F_INDIRECT |
1039 vq->packed.avail_used_flags;
1043 * A driver MUST NOT make the first descriptor in the list
1044 * available before all subsequent descriptors comprising
1045 * the list are made available.
1047 virtio_wmb(vq->weak_barriers);
1048 vq->packed.vring.desc[head].flags = cpu_to_le16(VRING_DESC_F_INDIRECT |
1049 vq->packed.avail_used_flags);
1051 /* We're using some buffers from the free list. */
1052 vq->vq.num_free -= 1;
1054 /* Update free pointer */
1056 if (n >= vq->packed.vring.num) {
1058 vq->packed.avail_wrap_counter ^= 1;
1059 vq->packed.avail_used_flags ^=
1060 1 << VRING_PACKED_DESC_F_AVAIL |
1061 1 << VRING_PACKED_DESC_F_USED;
1063 vq->packed.next_avail_idx = n;
1064 vq->free_head = vq->packed.desc_state[id].next;
1066 /* Store token and indirect buffer state. */
1067 vq->packed.desc_state[id].num = 1;
1068 vq->packed.desc_state[id].data = data;
1069 vq->packed.desc_state[id].indir_desc = desc;
1070 vq->packed.desc_state[id].last = id;
1074 pr_debug("Added buffer head %i to %p\n", head, vq);
1082 for (i = 0; i < err_idx; i++)
1083 vring_unmap_desc_packed(vq, &desc[i]);
1091 static inline int virtqueue_add_packed(struct virtqueue *_vq,
1092 struct scatterlist *sgs[],
1093 unsigned int total_sg,
1094 unsigned int out_sgs,
1095 unsigned int in_sgs,
1100 struct vring_virtqueue *vq = to_vvq(_vq);
1101 struct vring_packed_desc *desc;
1102 struct scatterlist *sg;
1103 unsigned int i, n, c, descs_used, err_idx;
1104 __le16 uninitialized_var(head_flags), flags;
1105 u16 head, id, uninitialized_var(prev), curr, avail_used_flags;
1109 BUG_ON(data == NULL);
1110 BUG_ON(ctx && vq->indirect);
1112 if (unlikely(vq->broken)) {
1117 LAST_ADD_TIME_UPDATE(vq);
1119 BUG_ON(total_sg == 0);
1121 if (virtqueue_use_indirect(_vq, total_sg))
1122 return virtqueue_add_indirect_packed(vq, sgs, total_sg,
1123 out_sgs, in_sgs, data, gfp);
1125 head = vq->packed.next_avail_idx;
1126 avail_used_flags = vq->packed.avail_used_flags;
1128 WARN_ON_ONCE(total_sg > vq->packed.vring.num && !vq->indirect);
1130 desc = vq->packed.vring.desc;
1132 descs_used = total_sg;
1134 if (unlikely(vq->vq.num_free < descs_used)) {
1135 pr_debug("Can't add buf len %i - avail = %i\n",
1136 descs_used, vq->vq.num_free);
1142 BUG_ON(id == vq->packed.vring.num);
1146 for (n = 0; n < out_sgs + in_sgs; n++) {
1147 for (sg = sgs[n]; sg; sg = sg_next(sg)) {
1148 dma_addr_t addr = vring_map_one_sg(vq, sg, n < out_sgs ?
1149 DMA_TO_DEVICE : DMA_FROM_DEVICE);
1150 if (vring_mapping_error(vq, addr))
1153 flags = cpu_to_le16(vq->packed.avail_used_flags |
1154 (++c == total_sg ? 0 : VRING_DESC_F_NEXT) |
1155 (n < out_sgs ? 0 : VRING_DESC_F_WRITE));
1159 desc[i].flags = flags;
1161 desc[i].addr = cpu_to_le64(addr);
1162 desc[i].len = cpu_to_le32(sg->length);
1163 desc[i].id = cpu_to_le16(id);
1165 if (unlikely(vq->use_dma_api)) {
1166 vq->packed.desc_extra[curr].addr = addr;
1167 vq->packed.desc_extra[curr].len = sg->length;
1168 vq->packed.desc_extra[curr].flags =
1172 curr = vq->packed.desc_state[curr].next;
1174 if ((unlikely(++i >= vq->packed.vring.num))) {
1176 vq->packed.avail_used_flags ^=
1177 1 << VRING_PACKED_DESC_F_AVAIL |
1178 1 << VRING_PACKED_DESC_F_USED;
1184 vq->packed.avail_wrap_counter ^= 1;
1186 /* We're using some buffers from the free list. */
1187 vq->vq.num_free -= descs_used;
1189 /* Update free pointer */
1190 vq->packed.next_avail_idx = i;
1191 vq->free_head = curr;
1194 vq->packed.desc_state[id].num = descs_used;
1195 vq->packed.desc_state[id].data = data;
1196 vq->packed.desc_state[id].indir_desc = ctx;
1197 vq->packed.desc_state[id].last = prev;
1200 * A driver MUST NOT make the first descriptor in the list
1201 * available before all subsequent descriptors comprising
1202 * the list are made available.
1204 virtio_wmb(vq->weak_barriers);
1205 vq->packed.vring.desc[head].flags = head_flags;
1206 vq->num_added += descs_used;
1208 pr_debug("Added buffer head %i to %p\n", head, vq);
1217 vq->packed.avail_used_flags = avail_used_flags;
1219 for (n = 0; n < total_sg; n++) {
1222 vring_unmap_desc_packed(vq, &desc[i]);
1224 if (i >= vq->packed.vring.num)
1232 static bool virtqueue_kick_prepare_packed(struct virtqueue *_vq)
1234 struct vring_virtqueue *vq = to_vvq(_vq);
1235 u16 new, old, off_wrap, flags, wrap_counter, event_idx;
1248 * We need to expose the new flags value before checking notification
1251 virtio_mb(vq->weak_barriers);
1253 old = vq->packed.next_avail_idx - vq->num_added;
1254 new = vq->packed.next_avail_idx;
1257 snapshot.u32 = *(u32 *)vq->packed.vring.device;
1258 flags = le16_to_cpu(snapshot.flags);
1260 LAST_ADD_TIME_CHECK(vq);
1261 LAST_ADD_TIME_INVALID(vq);
1263 if (flags != VRING_PACKED_EVENT_FLAG_DESC) {
1264 needs_kick = (flags != VRING_PACKED_EVENT_FLAG_DISABLE);
1268 off_wrap = le16_to_cpu(snapshot.off_wrap);
1270 wrap_counter = off_wrap >> VRING_PACKED_EVENT_F_WRAP_CTR;
1271 event_idx = off_wrap & ~(1 << VRING_PACKED_EVENT_F_WRAP_CTR);
1272 if (wrap_counter != vq->packed.avail_wrap_counter)
1273 event_idx -= vq->packed.vring.num;
1275 needs_kick = vring_need_event(event_idx, new, old);
1281 static void detach_buf_packed(struct vring_virtqueue *vq,
1282 unsigned int id, void **ctx)
1284 struct vring_desc_state_packed *state = NULL;
1285 struct vring_packed_desc *desc;
1286 unsigned int i, curr;
1288 state = &vq->packed.desc_state[id];
1290 /* Clear data ptr. */
1293 vq->packed.desc_state[state->last].next = vq->free_head;
1295 vq->vq.num_free += state->num;
1297 if (unlikely(vq->use_dma_api)) {
1299 for (i = 0; i < state->num; i++) {
1300 vring_unmap_state_packed(vq,
1301 &vq->packed.desc_extra[curr]);
1302 curr = vq->packed.desc_state[curr].next;
1309 /* Free the indirect table, if any, now that it's unmapped. */
1310 desc = state->indir_desc;
1314 if (vq->use_dma_api) {
1315 len = vq->packed.desc_extra[id].len;
1316 for (i = 0; i < len / sizeof(struct vring_packed_desc);
1318 vring_unmap_desc_packed(vq, &desc[i]);
1321 state->indir_desc = NULL;
1323 *ctx = state->indir_desc;
1327 static inline bool is_used_desc_packed(const struct vring_virtqueue *vq,
1328 u16 idx, bool used_wrap_counter)
1333 flags = le16_to_cpu(vq->packed.vring.desc[idx].flags);
1334 avail = !!(flags & (1 << VRING_PACKED_DESC_F_AVAIL));
1335 used = !!(flags & (1 << VRING_PACKED_DESC_F_USED));
1337 return avail == used && used == used_wrap_counter;
1340 static inline bool more_used_packed(const struct vring_virtqueue *vq)
1342 return is_used_desc_packed(vq, vq->last_used_idx,
1343 vq->packed.used_wrap_counter);
1346 static void *virtqueue_get_buf_ctx_packed(struct virtqueue *_vq,
1350 struct vring_virtqueue *vq = to_vvq(_vq);
1356 if (unlikely(vq->broken)) {
1361 if (!more_used_packed(vq)) {
1362 pr_debug("No more buffers in queue\n");
1367 /* Only get used elements after they have been exposed by host. */
1368 virtio_rmb(vq->weak_barriers);
1370 last_used = vq->last_used_idx;
1371 id = le16_to_cpu(vq->packed.vring.desc[last_used].id);
1372 *len = le32_to_cpu(vq->packed.vring.desc[last_used].len);
1374 if (unlikely(id >= vq->packed.vring.num)) {
1375 BAD_RING(vq, "id %u out of range\n", id);
1378 if (unlikely(!vq->packed.desc_state[id].data)) {
1379 BAD_RING(vq, "id %u is not a head!\n", id);
1383 /* detach_buf_packed clears data, so grab it now. */
1384 ret = vq->packed.desc_state[id].data;
1385 detach_buf_packed(vq, id, ctx);
1387 vq->last_used_idx += vq->packed.desc_state[id].num;
1388 if (unlikely(vq->last_used_idx >= vq->packed.vring.num)) {
1389 vq->last_used_idx -= vq->packed.vring.num;
1390 vq->packed.used_wrap_counter ^= 1;
1394 * If we expect an interrupt for the next entry, tell host
1395 * by writing event index and flush out the write before
1396 * the read in the next get_buf call.
1398 if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DESC)
1399 virtio_store_mb(vq->weak_barriers,
1400 &vq->packed.vring.driver->off_wrap,
1401 cpu_to_le16(vq->last_used_idx |
1402 (vq->packed.used_wrap_counter <<
1403 VRING_PACKED_EVENT_F_WRAP_CTR)));
1405 LAST_ADD_TIME_INVALID(vq);
1411 static void virtqueue_disable_cb_packed(struct virtqueue *_vq)
1413 struct vring_virtqueue *vq = to_vvq(_vq);
1415 if (vq->packed.event_flags_shadow != VRING_PACKED_EVENT_FLAG_DISABLE) {
1416 vq->packed.event_flags_shadow = VRING_PACKED_EVENT_FLAG_DISABLE;
1417 vq->packed.vring.driver->flags =
1418 cpu_to_le16(vq->packed.event_flags_shadow);
1422 static unsigned virtqueue_enable_cb_prepare_packed(struct virtqueue *_vq)
1424 struct vring_virtqueue *vq = to_vvq(_vq);
1429 * We optimistically turn back on interrupts, then check if there was
1434 vq->packed.vring.driver->off_wrap =
1435 cpu_to_le16(vq->last_used_idx |
1436 (vq->packed.used_wrap_counter <<
1437 VRING_PACKED_EVENT_F_WRAP_CTR));
1439 * We need to update event offset and event wrap
1440 * counter first before updating event flags.
1442 virtio_wmb(vq->weak_barriers);
1445 if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DISABLE) {
1446 vq->packed.event_flags_shadow = vq->event ?
1447 VRING_PACKED_EVENT_FLAG_DESC :
1448 VRING_PACKED_EVENT_FLAG_ENABLE;
1449 vq->packed.vring.driver->flags =
1450 cpu_to_le16(vq->packed.event_flags_shadow);
1454 return vq->last_used_idx | ((u16)vq->packed.used_wrap_counter <<
1455 VRING_PACKED_EVENT_F_WRAP_CTR);
1458 static bool virtqueue_poll_packed(struct virtqueue *_vq, u16 off_wrap)
1460 struct vring_virtqueue *vq = to_vvq(_vq);
1464 wrap_counter = off_wrap >> VRING_PACKED_EVENT_F_WRAP_CTR;
1465 used_idx = off_wrap & ~(1 << VRING_PACKED_EVENT_F_WRAP_CTR);
1467 return is_used_desc_packed(vq, used_idx, wrap_counter);
1470 static bool virtqueue_enable_cb_delayed_packed(struct virtqueue *_vq)
1472 struct vring_virtqueue *vq = to_vvq(_vq);
1473 u16 used_idx, wrap_counter;
1479 * We optimistically turn back on interrupts, then check if there was
1484 /* TODO: tune this threshold */
1485 bufs = (vq->packed.vring.num - vq->vq.num_free) * 3 / 4;
1486 wrap_counter = vq->packed.used_wrap_counter;
1488 used_idx = vq->last_used_idx + bufs;
1489 if (used_idx >= vq->packed.vring.num) {
1490 used_idx -= vq->packed.vring.num;
1494 vq->packed.vring.driver->off_wrap = cpu_to_le16(used_idx |
1495 (wrap_counter << VRING_PACKED_EVENT_F_WRAP_CTR));
1498 * We need to update event offset and event wrap
1499 * counter first before updating event flags.
1501 virtio_wmb(vq->weak_barriers);
1503 used_idx = vq->last_used_idx;
1504 wrap_counter = vq->packed.used_wrap_counter;
1507 if (vq->packed.event_flags_shadow == VRING_PACKED_EVENT_FLAG_DISABLE) {
1508 vq->packed.event_flags_shadow = vq->event ?
1509 VRING_PACKED_EVENT_FLAG_DESC :
1510 VRING_PACKED_EVENT_FLAG_ENABLE;
1511 vq->packed.vring.driver->flags =
1512 cpu_to_le16(vq->packed.event_flags_shadow);
1516 * We need to update event suppression structure first
1517 * before re-checking for more used buffers.
1519 virtio_mb(vq->weak_barriers);
1521 if (is_used_desc_packed(vq, used_idx, wrap_counter)) {
1530 static void *virtqueue_detach_unused_buf_packed(struct virtqueue *_vq)
1532 struct vring_virtqueue *vq = to_vvq(_vq);
1538 for (i = 0; i < vq->packed.vring.num; i++) {
1539 if (!vq->packed.desc_state[i].data)
1541 /* detach_buf clears data, so grab it now. */
1542 buf = vq->packed.desc_state[i].data;
1543 detach_buf_packed(vq, i, NULL);
1547 /* That should have freed everything. */
1548 BUG_ON(vq->vq.num_free != vq->packed.vring.num);
1554 static struct virtqueue *vring_create_virtqueue_packed(
1557 unsigned int vring_align,
1558 struct virtio_device *vdev,
1560 bool may_reduce_num,
1562 bool (*notify)(struct virtqueue *),
1563 void (*callback)(struct virtqueue *),
1566 struct vring_virtqueue *vq;
1567 struct vring_packed_desc *ring;
1568 struct vring_packed_desc_event *driver, *device;
1569 dma_addr_t ring_dma_addr, driver_event_dma_addr, device_event_dma_addr;
1570 size_t ring_size_in_bytes, event_size_in_bytes;
1573 ring_size_in_bytes = num * sizeof(struct vring_packed_desc);
1575 ring = vring_alloc_queue(vdev, ring_size_in_bytes,
1577 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1581 event_size_in_bytes = sizeof(struct vring_packed_desc_event);
1583 driver = vring_alloc_queue(vdev, event_size_in_bytes,
1584 &driver_event_dma_addr,
1585 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1589 device = vring_alloc_queue(vdev, event_size_in_bytes,
1590 &device_event_dma_addr,
1591 GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
1595 vq = kmalloc(sizeof(*vq), GFP_KERNEL);
1599 vq->vq.callback = callback;
1602 vq->vq.num_free = num;
1603 vq->vq.index = index;
1604 vq->we_own_ring = true;
1605 vq->notify = notify;
1606 vq->weak_barriers = weak_barriers;
1608 vq->last_used_idx = 0;
1610 vq->packed_ring = true;
1611 vq->use_dma_api = vring_use_dma_api(vdev);
1612 list_add_tail(&vq->vq.list, &vdev->vqs);
1615 vq->last_add_time_valid = false;
1618 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
1620 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
1622 if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
1623 vq->weak_barriers = false;
1625 vq->packed.ring_dma_addr = ring_dma_addr;
1626 vq->packed.driver_event_dma_addr = driver_event_dma_addr;
1627 vq->packed.device_event_dma_addr = device_event_dma_addr;
1629 vq->packed.ring_size_in_bytes = ring_size_in_bytes;
1630 vq->packed.event_size_in_bytes = event_size_in_bytes;
1632 vq->packed.vring.num = num;
1633 vq->packed.vring.desc = ring;
1634 vq->packed.vring.driver = driver;
1635 vq->packed.vring.device = device;
1637 vq->packed.next_avail_idx = 0;
1638 vq->packed.avail_wrap_counter = 1;
1639 vq->packed.used_wrap_counter = 1;
1640 vq->packed.event_flags_shadow = 0;
1641 vq->packed.avail_used_flags = 1 << VRING_PACKED_DESC_F_AVAIL;
1643 vq->packed.desc_state = kmalloc_array(num,
1644 sizeof(struct vring_desc_state_packed),
1646 if (!vq->packed.desc_state)
1647 goto err_desc_state;
1649 memset(vq->packed.desc_state, 0,
1650 num * sizeof(struct vring_desc_state_packed));
1652 /* Put everything in free lists. */
1654 for (i = 0; i < num-1; i++)
1655 vq->packed.desc_state[i].next = i + 1;
1657 vq->packed.desc_extra = kmalloc_array(num,
1658 sizeof(struct vring_desc_extra_packed),
1660 if (!vq->packed.desc_extra)
1661 goto err_desc_extra;
1663 memset(vq->packed.desc_extra, 0,
1664 num * sizeof(struct vring_desc_extra_packed));
1666 /* No callback? Tell other side not to bother us. */
1668 vq->packed.event_flags_shadow = VRING_PACKED_EVENT_FLAG_DISABLE;
1669 vq->packed.vring.driver->flags =
1670 cpu_to_le16(vq->packed.event_flags_shadow);
1676 kfree(vq->packed.desc_state);
1680 vring_free_queue(vdev, event_size_in_bytes, device, ring_dma_addr);
1682 vring_free_queue(vdev, event_size_in_bytes, driver, ring_dma_addr);
1684 vring_free_queue(vdev, ring_size_in_bytes, ring, ring_dma_addr);
1691 * Generic functions and exported symbols.
1694 static inline int virtqueue_add(struct virtqueue *_vq,
1695 struct scatterlist *sgs[],
1696 unsigned int total_sg,
1697 unsigned int out_sgs,
1698 unsigned int in_sgs,
1703 struct vring_virtqueue *vq = to_vvq(_vq);
1705 return vq->packed_ring ? virtqueue_add_packed(_vq, sgs, total_sg,
1706 out_sgs, in_sgs, data, ctx, gfp) :
1707 virtqueue_add_split(_vq, sgs, total_sg,
1708 out_sgs, in_sgs, data, ctx, gfp);
1712 * virtqueue_add_sgs - expose buffers to other end
1713 * @_vq: the struct virtqueue we're talking about.
1714 * @sgs: array of terminated scatterlists.
1715 * @out_sgs: the number of scatterlists readable by other side
1716 * @in_sgs: the number of scatterlists which are writable (after readable ones)
1717 * @data: the token identifying the buffer.
1718 * @gfp: how to do memory allocations (if necessary).
1720 * Caller must ensure we don't call this with other virtqueue operations
1721 * at the same time (except where noted).
1723 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1725 int virtqueue_add_sgs(struct virtqueue *_vq,
1726 struct scatterlist *sgs[],
1727 unsigned int out_sgs,
1728 unsigned int in_sgs,
1732 unsigned int i, total_sg = 0;
1734 /* Count them first. */
1735 for (i = 0; i < out_sgs + in_sgs; i++) {
1736 struct scatterlist *sg;
1738 for (sg = sgs[i]; sg; sg = sg_next(sg))
1741 return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs,
1744 EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
1747 * virtqueue_add_outbuf - expose output buffers to other end
1748 * @vq: the struct virtqueue we're talking about.
1749 * @sg: scatterlist (must be well-formed and terminated!)
1750 * @num: the number of entries in @sg readable by other side
1751 * @data: the token identifying the buffer.
1752 * @gfp: how to do memory allocations (if necessary).
1754 * Caller must ensure we don't call this with other virtqueue operations
1755 * at the same time (except where noted).
1757 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1759 int virtqueue_add_outbuf(struct virtqueue *vq,
1760 struct scatterlist *sg, unsigned int num,
1764 return virtqueue_add(vq, &sg, num, 1, 0, data, NULL, gfp);
1766 EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
1769 * virtqueue_add_inbuf - expose input buffers to other end
1770 * @vq: the struct virtqueue we're talking about.
1771 * @sg: scatterlist (must be well-formed and terminated!)
1772 * @num: the number of entries in @sg writable by other side
1773 * @data: the token identifying the buffer.
1774 * @gfp: how to do memory allocations (if necessary).
1776 * Caller must ensure we don't call this with other virtqueue operations
1777 * at the same time (except where noted).
1779 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1781 int virtqueue_add_inbuf(struct virtqueue *vq,
1782 struct scatterlist *sg, unsigned int num,
1786 return virtqueue_add(vq, &sg, num, 0, 1, data, NULL, gfp);
1788 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
1791 * virtqueue_add_inbuf_ctx - expose input buffers to other end
1792 * @vq: the struct virtqueue we're talking about.
1793 * @sg: scatterlist (must be well-formed and terminated!)
1794 * @num: the number of entries in @sg writable by other side
1795 * @data: the token identifying the buffer.
1796 * @ctx: extra context for the token
1797 * @gfp: how to do memory allocations (if necessary).
1799 * Caller must ensure we don't call this with other virtqueue operations
1800 * at the same time (except where noted).
1802 * Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
1804 int virtqueue_add_inbuf_ctx(struct virtqueue *vq,
1805 struct scatterlist *sg, unsigned int num,
1810 return virtqueue_add(vq, &sg, num, 0, 1, data, ctx, gfp);
1812 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf_ctx);
1815 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
1816 * @_vq: the struct virtqueue
1818 * Instead of virtqueue_kick(), you can do:
1819 * if (virtqueue_kick_prepare(vq))
1820 * virtqueue_notify(vq);
1822 * This is sometimes useful because the virtqueue_kick_prepare() needs
1823 * to be serialized, but the actual virtqueue_notify() call does not.
1825 bool virtqueue_kick_prepare(struct virtqueue *_vq)
1827 struct vring_virtqueue *vq = to_vvq(_vq);
1829 return vq->packed_ring ? virtqueue_kick_prepare_packed(_vq) :
1830 virtqueue_kick_prepare_split(_vq);
1832 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
1835 * virtqueue_notify - second half of split virtqueue_kick call.
1836 * @_vq: the struct virtqueue
1838 * This does not need to be serialized.
1840 * Returns false if host notify failed or queue is broken, otherwise true.
1842 bool virtqueue_notify(struct virtqueue *_vq)
1844 struct vring_virtqueue *vq = to_vvq(_vq);
1846 if (unlikely(vq->broken))
1849 /* Prod other side to tell it about changes. */
1850 if (!vq->notify(_vq)) {
1856 EXPORT_SYMBOL_GPL(virtqueue_notify);
1859 * virtqueue_kick - update after add_buf
1860 * @vq: the struct virtqueue
1862 * After one or more virtqueue_add_* calls, invoke this to kick
1865 * Caller must ensure we don't call this with other virtqueue
1866 * operations at the same time (except where noted).
1868 * Returns false if kick failed, otherwise true.
1870 bool virtqueue_kick(struct virtqueue *vq)
1872 if (virtqueue_kick_prepare(vq))
1873 return virtqueue_notify(vq);
1876 EXPORT_SYMBOL_GPL(virtqueue_kick);
1879 * virtqueue_get_buf - get the next used buffer
1880 * @_vq: the struct virtqueue we're talking about.
1881 * @len: the length written into the buffer
1882 * @ctx: extra context for the token
1884 * If the device wrote data into the buffer, @len will be set to the
1885 * amount written. This means you don't need to clear the buffer
1886 * beforehand to ensure there's no data leakage in the case of short
1889 * Caller must ensure we don't call this with other virtqueue
1890 * operations at the same time (except where noted).
1892 * Returns NULL if there are no used buffers, or the "data" token
1893 * handed to virtqueue_add_*().
1895 void *virtqueue_get_buf_ctx(struct virtqueue *_vq, unsigned int *len,
1898 struct vring_virtqueue *vq = to_vvq(_vq);
1900 return vq->packed_ring ? virtqueue_get_buf_ctx_packed(_vq, len, ctx) :
1901 virtqueue_get_buf_ctx_split(_vq, len, ctx);
1903 EXPORT_SYMBOL_GPL(virtqueue_get_buf_ctx);
1905 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
1907 return virtqueue_get_buf_ctx(_vq, len, NULL);
1909 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
1911 * virtqueue_disable_cb - disable callbacks
1912 * @_vq: the struct virtqueue we're talking about.
1914 * Note that this is not necessarily synchronous, hence unreliable and only
1915 * useful as an optimization.
1917 * Unlike other operations, this need not be serialized.
1919 void virtqueue_disable_cb(struct virtqueue *_vq)
1921 struct vring_virtqueue *vq = to_vvq(_vq);
1923 if (vq->packed_ring)
1924 virtqueue_disable_cb_packed(_vq);
1926 virtqueue_disable_cb_split(_vq);
1928 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
1931 * virtqueue_enable_cb_prepare - restart callbacks after disable_cb
1932 * @_vq: the struct virtqueue we're talking about.
1934 * This re-enables callbacks; it returns current queue state
1935 * in an opaque unsigned value. This value should be later tested by
1936 * virtqueue_poll, to detect a possible race between the driver checking for
1937 * more work, and enabling callbacks.
1939 * Caller must ensure we don't call this with other virtqueue
1940 * operations at the same time (except where noted).
1942 unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
1944 struct vring_virtqueue *vq = to_vvq(_vq);
1946 return vq->packed_ring ? virtqueue_enable_cb_prepare_packed(_vq) :
1947 virtqueue_enable_cb_prepare_split(_vq);
1949 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
1952 * virtqueue_poll - query pending used buffers
1953 * @_vq: the struct virtqueue we're talking about.
1954 * @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
1956 * Returns "true" if there are pending used buffers in the queue.
1958 * This does not need to be serialized.
1960 bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
1962 struct vring_virtqueue *vq = to_vvq(_vq);
1964 virtio_mb(vq->weak_barriers);
1965 return vq->packed_ring ? virtqueue_poll_packed(_vq, last_used_idx) :
1966 virtqueue_poll_split(_vq, last_used_idx);
1968 EXPORT_SYMBOL_GPL(virtqueue_poll);
1971 * virtqueue_enable_cb - restart callbacks after disable_cb.
1972 * @_vq: the struct virtqueue we're talking about.
1974 * This re-enables callbacks; it returns "false" if there are pending
1975 * buffers in the queue, to detect a possible race between the driver
1976 * checking for more work, and enabling callbacks.
1978 * Caller must ensure we don't call this with other virtqueue
1979 * operations at the same time (except where noted).
1981 bool virtqueue_enable_cb(struct virtqueue *_vq)
1983 unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
1985 return !virtqueue_poll(_vq, last_used_idx);
1987 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
1990 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
1991 * @_vq: the struct virtqueue we're talking about.
1993 * This re-enables callbacks but hints to the other side to delay
1994 * interrupts until most of the available buffers have been processed;
1995 * it returns "false" if there are many pending buffers in the queue,
1996 * to detect a possible race between the driver checking for more work,
1997 * and enabling callbacks.
1999 * Caller must ensure we don't call this with other virtqueue
2000 * operations at the same time (except where noted).
2002 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
2004 struct vring_virtqueue *vq = to_vvq(_vq);
2006 return vq->packed_ring ? virtqueue_enable_cb_delayed_packed(_vq) :
2007 virtqueue_enable_cb_delayed_split(_vq);
2009 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
2012 * virtqueue_detach_unused_buf - detach first unused buffer
2013 * @_vq: the struct virtqueue we're talking about.
2015 * Returns NULL or the "data" token handed to virtqueue_add_*().
2016 * This is not valid on an active queue; it is useful only for device
2019 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
2021 struct vring_virtqueue *vq = to_vvq(_vq);
2023 return vq->packed_ring ? virtqueue_detach_unused_buf_packed(_vq) :
2024 virtqueue_detach_unused_buf_split(_vq);
2026 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
2028 static inline bool more_used(const struct vring_virtqueue *vq)
2030 return vq->packed_ring ? more_used_packed(vq) : more_used_split(vq);
2033 irqreturn_t vring_interrupt(int irq, void *_vq)
2035 struct vring_virtqueue *vq = to_vvq(_vq);
2037 if (!more_used(vq)) {
2038 pr_debug("virtqueue interrupt with no work for %p\n", vq);
2042 if (unlikely(vq->broken))
2045 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
2046 if (vq->vq.callback)
2047 vq->vq.callback(&vq->vq);
2051 EXPORT_SYMBOL_GPL(vring_interrupt);
2053 /* Only available for split ring */
2054 struct virtqueue *__vring_new_virtqueue(unsigned int index,
2056 struct virtio_device *vdev,
2059 bool (*notify)(struct virtqueue *),
2060 void (*callback)(struct virtqueue *),
2064 struct vring_virtqueue *vq;
2066 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED))
2069 vq = kmalloc(sizeof(*vq), GFP_KERNEL);
2073 vq->packed_ring = false;
2074 vq->vq.callback = callback;
2077 vq->vq.num_free = vring.num;
2078 vq->vq.index = index;
2079 vq->we_own_ring = false;
2080 vq->notify = notify;
2081 vq->weak_barriers = weak_barriers;
2083 vq->last_used_idx = 0;
2085 vq->use_dma_api = vring_use_dma_api(vdev);
2086 list_add_tail(&vq->vq.list, &vdev->vqs);
2089 vq->last_add_time_valid = false;
2092 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC) &&
2094 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
2096 if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
2097 vq->weak_barriers = false;
2099 vq->split.queue_dma_addr = 0;
2100 vq->split.queue_size_in_bytes = 0;
2102 vq->split.vring = vring;
2103 vq->split.avail_flags_shadow = 0;
2104 vq->split.avail_idx_shadow = 0;
2106 /* No callback? Tell other side not to bother us. */
2108 vq->split.avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
2110 vq->split.vring.avail->flags = cpu_to_virtio16(vdev,
2111 vq->split.avail_flags_shadow);
2114 vq->split.desc_state = kmalloc_array(vring.num,
2115 sizeof(struct vring_desc_state_split), GFP_KERNEL);
2116 if (!vq->split.desc_state) {
2121 /* Put everything in free lists. */
2123 for (i = 0; i < vring.num-1; i++)
2124 vq->split.vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
2125 memset(vq->split.desc_state, 0, vring.num *
2126 sizeof(struct vring_desc_state_split));
2130 EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
2132 struct virtqueue *vring_create_virtqueue(
2135 unsigned int vring_align,
2136 struct virtio_device *vdev,
2138 bool may_reduce_num,
2140 bool (*notify)(struct virtqueue *),
2141 void (*callback)(struct virtqueue *),
2145 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED))
2146 return vring_create_virtqueue_packed(index, num, vring_align,
2147 vdev, weak_barriers, may_reduce_num,
2148 context, notify, callback, name);
2150 return vring_create_virtqueue_split(index, num, vring_align,
2151 vdev, weak_barriers, may_reduce_num,
2152 context, notify, callback, name);
2154 EXPORT_SYMBOL_GPL(vring_create_virtqueue);
2156 /* Only available for split ring */
2157 struct virtqueue *vring_new_virtqueue(unsigned int index,
2159 unsigned int vring_align,
2160 struct virtio_device *vdev,
2164 bool (*notify)(struct virtqueue *vq),
2165 void (*callback)(struct virtqueue *vq),
2170 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED))
2173 vring_init(&vring, num, pages, vring_align);
2174 return __vring_new_virtqueue(index, vring, vdev, weak_barriers, context,
2175 notify, callback, name);
2177 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
2179 void vring_del_virtqueue(struct virtqueue *_vq)
2181 struct vring_virtqueue *vq = to_vvq(_vq);
2183 if (vq->we_own_ring) {
2184 if (vq->packed_ring) {
2185 vring_free_queue(vq->vq.vdev,
2186 vq->packed.ring_size_in_bytes,
2187 vq->packed.vring.desc,
2188 vq->packed.ring_dma_addr);
2190 vring_free_queue(vq->vq.vdev,
2191 vq->packed.event_size_in_bytes,
2192 vq->packed.vring.driver,
2193 vq->packed.driver_event_dma_addr);
2195 vring_free_queue(vq->vq.vdev,
2196 vq->packed.event_size_in_bytes,
2197 vq->packed.vring.device,
2198 vq->packed.device_event_dma_addr);
2200 kfree(vq->packed.desc_state);
2201 kfree(vq->packed.desc_extra);
2203 vring_free_queue(vq->vq.vdev,
2204 vq->split.queue_size_in_bytes,
2205 vq->split.vring.desc,
2206 vq->split.queue_dma_addr);
2208 kfree(vq->split.desc_state);
2211 list_del(&_vq->list);
2214 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
2216 /* Manipulates transport-specific feature bits. */
2217 void vring_transport_features(struct virtio_device *vdev)
2221 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
2223 case VIRTIO_RING_F_INDIRECT_DESC:
2225 case VIRTIO_RING_F_EVENT_IDX:
2227 case VIRTIO_F_VERSION_1:
2229 case VIRTIO_F_IOMMU_PLATFORM:
2231 case VIRTIO_F_RING_PACKED:
2233 case VIRTIO_F_ORDER_PLATFORM:
2236 /* We don't understand this bit. */
2237 __virtio_clear_bit(vdev, i);
2241 EXPORT_SYMBOL_GPL(vring_transport_features);
2244 * virtqueue_get_vring_size - return the size of the virtqueue's vring
2245 * @_vq: the struct virtqueue containing the vring of interest.
2247 * Returns the size of the vring. This is mainly used for boasting to
2248 * userspace. Unlike other operations, this need not be serialized.
2250 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
2253 struct vring_virtqueue *vq = to_vvq(_vq);
2255 return vq->packed_ring ? vq->packed.vring.num : vq->split.vring.num;
2257 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
2259 bool virtqueue_is_broken(struct virtqueue *_vq)
2261 struct vring_virtqueue *vq = to_vvq(_vq);
2265 EXPORT_SYMBOL_GPL(virtqueue_is_broken);
2268 * This should prevent the device from being used, allowing drivers to
2269 * recover. You may need to grab appropriate locks to flush.
2271 void virtio_break_device(struct virtio_device *dev)
2273 struct virtqueue *_vq;
2275 list_for_each_entry(_vq, &dev->vqs, list) {
2276 struct vring_virtqueue *vq = to_vvq(_vq);
2280 EXPORT_SYMBOL_GPL(virtio_break_device);
2282 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
2284 struct vring_virtqueue *vq = to_vvq(_vq);
2286 BUG_ON(!vq->we_own_ring);
2288 if (vq->packed_ring)
2289 return vq->packed.ring_dma_addr;
2291 return vq->split.queue_dma_addr;
2293 EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
2295 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
2297 struct vring_virtqueue *vq = to_vvq(_vq);
2299 BUG_ON(!vq->we_own_ring);
2301 if (vq->packed_ring)
2302 return vq->packed.driver_event_dma_addr;
2304 return vq->split.queue_dma_addr +
2305 ((char *)vq->split.vring.avail - (char *)vq->split.vring.desc);
2307 EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
2309 dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
2311 struct vring_virtqueue *vq = to_vvq(_vq);
2313 BUG_ON(!vq->we_own_ring);
2315 if (vq->packed_ring)
2316 return vq->packed.device_event_dma_addr;
2318 return vq->split.queue_dma_addr +
2319 ((char *)vq->split.vring.used - (char *)vq->split.vring.desc);
2321 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
2323 /* Only available for split ring */
2324 const struct vring *virtqueue_get_vring(struct virtqueue *vq)
2326 return &to_vvq(vq)->split.vring;
2328 EXPORT_SYMBOL_GPL(virtqueue_get_vring);
2330 MODULE_LICENSE("GPL");