linux/drivers/virtio/virtio_ring.c

1208 lines
33 KiB
C
Raw Normal View History

/* Virtio ring implementation.
*
* Copyright 2007 Rusty Russell IBM Corporation
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <linux/virtio.h>
#include <linux/virtio_ring.h>
#include <linux/virtio_config.h>
#include <linux/device.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
#include <linux/slab.h>
#include <linux/module.h>
#include <linux/hrtimer.h>
#include <linux/kmemleak.h>
#include <linux/dma-mapping.h>
#include <xen/xen.h>
#ifdef DEBUG
/* For development, we want to crash whenever the ring is screwed. */
#define BAD_RING(_vq, fmt, args...) \
do { \
dev_err(&(_vq)->vq.vdev->dev, \
"%s:"fmt, (_vq)->vq.name, ##args); \
BUG(); \
} while (0)
/* Caller is supposed to guarantee no reentry. */
#define START_USE(_vq) \
do { \
if ((_vq)->in_use) \
panic("%s:in_use = %i\n", \
(_vq)->vq.name, (_vq)->in_use); \
(_vq)->in_use = __LINE__; \
} while (0)
#define END_USE(_vq) \
do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
#else
#define BAD_RING(_vq, fmt, args...) \
do { \
dev_err(&_vq->vq.vdev->dev, \
"%s:"fmt, (_vq)->vq.name, ##args); \
(_vq)->broken = true; \
} while (0)
#define START_USE(vq)
#define END_USE(vq)
#endif
struct vring_desc_state {
void *data; /* Data for callback. */
struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
};
struct vring_virtqueue {
struct virtqueue vq;
/* Actual memory layout for this queue */
struct vring vring;
/* Can we use weak barriers? */
bool weak_barriers;
/* Other side has made a mess, don't try any more. */
bool broken;
/* Host supports indirect buffers */
bool indirect;
/* Host publishes avail event idx */
bool event;
/* Head of free buffer list. */
unsigned int free_head;
/* Number we've added since last sync. */
unsigned int num_added;
/* Last used index we've seen. */
u16 last_used_idx;
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
/* Last written value to avail->flags */
u16 avail_flags_shadow;
/* Last written value to avail->idx in guest byte order */
u16 avail_idx_shadow;
/* How to notify other side. FIXME: commonalize hcalls! */
bool (*notify)(struct virtqueue *vq);
/* DMA, allocation, and size information */
bool we_own_ring;
size_t queue_size_in_bytes;
dma_addr_t queue_dma_addr;
#ifdef DEBUG
/* They're supposed to lock for us. */
unsigned int in_use;
/* Figure out if their kicks are too delayed. */
bool last_add_time_valid;
ktime_t last_add_time;
#endif
/* Per-descriptor state. */
struct vring_desc_state desc_state[];
};
#define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
/*
virtio: new feature to detect IOMMU device quirk The interaction between virtio and IOMMUs is messy. On most systems with virtio, physical addresses match bus addresses, and it doesn't particularly matter which one we use to program the device. On some systems, including Xen and any system with a physical device that speaks virtio behind a physical IOMMU, we must program the IOMMU for virtio DMA to work at all. On other systems, including SPARC and PPC64, virtio-pci devices are enumerated as though they are behind an IOMMU, but the virtio host ignores the IOMMU, so we must either pretend that the IOMMU isn't there or somehow map everything as the identity. Add a feature bit to detect that quirk: VIRTIO_F_IOMMU_PLATFORM. Any device with this feature bit set to 0 needs a quirk and has to be passed physical addresses (as opposed to bus addresses) even though the device is behind an IOMMU. Note: it has to be a per-device quirk because for example, there could be a mix of passed-through and virtual virtio devices. As another example, some devices could be implemented by an out of process hypervisor backend (in case of qemu vhost, or vhost-user) and so support for an IOMMU needs to be coded up separately. It would be cleanest to handle this in IOMMU core code, but that needs per-device DMA ops. While we are waiting for that to be implemented, use a work-around in virtio core. Note: a "noiommu" feature is a quirk - add a wrapper to make that clear. Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2016-04-18 09:58:14 +00:00
* Modern virtio devices have feature bits to specify whether they need a
* quirk and bypass the IOMMU. If not there, just use the DMA API.
*
* If there, the interaction between virtio and DMA API is messy.
*
* On most systems with virtio, physical addresses match bus addresses,
* and it doesn't particularly matter whether we use the DMA API.
*
* On some systems, including Xen and any system with a physical device
* that speaks virtio behind a physical IOMMU, we must use the DMA API
* for virtio DMA to work at all.
*
* On other systems, including SPARC and PPC64, virtio-pci devices are
* enumerated as though they are behind an IOMMU, but the virtio host
* ignores the IOMMU, so we must either pretend that the IOMMU isn't
* there or somehow map everything as the identity.
*
* For the time being, we preserve historic behavior and bypass the DMA
* API.
virtio: new feature to detect IOMMU device quirk The interaction between virtio and IOMMUs is messy. On most systems with virtio, physical addresses match bus addresses, and it doesn't particularly matter which one we use to program the device. On some systems, including Xen and any system with a physical device that speaks virtio behind a physical IOMMU, we must program the IOMMU for virtio DMA to work at all. On other systems, including SPARC and PPC64, virtio-pci devices are enumerated as though they are behind an IOMMU, but the virtio host ignores the IOMMU, so we must either pretend that the IOMMU isn't there or somehow map everything as the identity. Add a feature bit to detect that quirk: VIRTIO_F_IOMMU_PLATFORM. Any device with this feature bit set to 0 needs a quirk and has to be passed physical addresses (as opposed to bus addresses) even though the device is behind an IOMMU. Note: it has to be a per-device quirk because for example, there could be a mix of passed-through and virtual virtio devices. As another example, some devices could be implemented by an out of process hypervisor backend (in case of qemu vhost, or vhost-user) and so support for an IOMMU needs to be coded up separately. It would be cleanest to handle this in IOMMU core code, but that needs per-device DMA ops. While we are waiting for that to be implemented, use a work-around in virtio core. Note: a "noiommu" feature is a quirk - add a wrapper to make that clear. Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2016-04-18 09:58:14 +00:00
*
* TODO: install a per-device DMA ops structure that does the right thing
* taking into account all the above quirks, and use the DMA API
* unconditionally on data path.
*/
static bool vring_use_dma_api(struct virtio_device *vdev)
{
virtio: new feature to detect IOMMU device quirk The interaction between virtio and IOMMUs is messy. On most systems with virtio, physical addresses match bus addresses, and it doesn't particularly matter which one we use to program the device. On some systems, including Xen and any system with a physical device that speaks virtio behind a physical IOMMU, we must program the IOMMU for virtio DMA to work at all. On other systems, including SPARC and PPC64, virtio-pci devices are enumerated as though they are behind an IOMMU, but the virtio host ignores the IOMMU, so we must either pretend that the IOMMU isn't there or somehow map everything as the identity. Add a feature bit to detect that quirk: VIRTIO_F_IOMMU_PLATFORM. Any device with this feature bit set to 0 needs a quirk and has to be passed physical addresses (as opposed to bus addresses) even though the device is behind an IOMMU. Note: it has to be a per-device quirk because for example, there could be a mix of passed-through and virtual virtio devices. As another example, some devices could be implemented by an out of process hypervisor backend (in case of qemu vhost, or vhost-user) and so support for an IOMMU needs to be coded up separately. It would be cleanest to handle this in IOMMU core code, but that needs per-device DMA ops. While we are waiting for that to be implemented, use a work-around in virtio core. Note: a "noiommu" feature is a quirk - add a wrapper to make that clear. Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2016-04-18 09:58:14 +00:00
if (!virtio_has_iommu_quirk(vdev))
return true;
/* Otherwise, we are left to guess. */
/*
* In theory, it's possible to have a buggy QEMU-supposed
* emulated Q35 IOMMU and Xen enabled at the same time. On
* such a configuration, virtio has never worked and will
* not work without an even larger kludge. Instead, enable
* the DMA API if we're a Xen guest, which at least allows
* all of the sensible Xen configurations to work correctly.
*/
if (xen_domain())
return true;
return false;
}
/*
* The DMA ops on various arches are rather gnarly right now, and
* making all of the arch DMA ops work on the vring device itself
* is a mess. For now, we use the parent device for DMA ops.
*/
static inline struct device *vring_dma_dev(const struct vring_virtqueue *vq)
{
return vq->vq.vdev->dev.parent;
}
/* Map one sg entry. */
static dma_addr_t vring_map_one_sg(const struct vring_virtqueue *vq,
struct scatterlist *sg,
enum dma_data_direction direction)
{
if (!vring_use_dma_api(vq->vq.vdev))
return (dma_addr_t)sg_phys(sg);
/*
* We can't use dma_map_sg, because we don't use scatterlists in
* the way it expects (we don't guarantee that the scatterlist
* will exist for the lifetime of the mapping).
*/
return dma_map_page(vring_dma_dev(vq),
sg_page(sg), sg->offset, sg->length,
direction);
}
static dma_addr_t vring_map_single(const struct vring_virtqueue *vq,
void *cpu_addr, size_t size,
enum dma_data_direction direction)
{
if (!vring_use_dma_api(vq->vq.vdev))
return (dma_addr_t)virt_to_phys(cpu_addr);
return dma_map_single(vring_dma_dev(vq),
cpu_addr, size, direction);
}
static void vring_unmap_one(const struct vring_virtqueue *vq,
struct vring_desc *desc)
{
u16 flags;
if (!vring_use_dma_api(vq->vq.vdev))
return;
flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
if (flags & VRING_DESC_F_INDIRECT) {
dma_unmap_single(vring_dma_dev(vq),
virtio64_to_cpu(vq->vq.vdev, desc->addr),
virtio32_to_cpu(vq->vq.vdev, desc->len),
(flags & VRING_DESC_F_WRITE) ?
DMA_FROM_DEVICE : DMA_TO_DEVICE);
} else {
dma_unmap_page(vring_dma_dev(vq),
virtio64_to_cpu(vq->vq.vdev, desc->addr),
virtio32_to_cpu(vq->vq.vdev, desc->len),
(flags & VRING_DESC_F_WRITE) ?
DMA_FROM_DEVICE : DMA_TO_DEVICE);
}
}
static int vring_mapping_error(const struct vring_virtqueue *vq,
dma_addr_t addr)
{
if (!vring_use_dma_api(vq->vq.vdev))
return 0;
return dma_mapping_error(vring_dma_dev(vq), addr);
}
static struct vring_desc *alloc_indirect(struct virtqueue *_vq,
unsigned int total_sg, gfp_t gfp)
{
struct vring_desc *desc;
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
unsigned int i;
/*
* We require lowmem mappings for the descriptors because
* otherwise virt_to_phys will give us bogus addresses in the
* virtqueue.
*/
gfp &= ~__GFP_HIGHMEM;
desc = kmalloc(total_sg * sizeof(struct vring_desc), gfp);
if (!desc)
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
return NULL;
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
for (i = 0; i < total_sg; i++)
desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
return desc;
}
static inline int virtqueue_add(struct virtqueue *_vq,
struct scatterlist *sgs[],
unsigned int total_sg,
unsigned int out_sgs,
unsigned int in_sgs,
void *data,
gfp_t gfp)
{
struct vring_virtqueue *vq = to_vvq(_vq);
struct scatterlist *sg;
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
struct vring_desc *desc;
unsigned int i, n, avail, descs_used, uninitialized_var(prev), err_idx;
int head;
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
bool indirect;
START_USE(vq);
BUG_ON(data == NULL);
if (unlikely(vq->broken)) {
END_USE(vq);
return -EIO;
}
#ifdef DEBUG
{
ktime_t now = ktime_get();
/* No kick or get, with .1 second between? Warn. */
if (vq->last_add_time_valid)
WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
> 100);
vq->last_add_time = now;
vq->last_add_time_valid = true;
}
#endif
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
BUG_ON(total_sg > vq->vring.num);
BUG_ON(total_sg == 0);
head = vq->free_head;
/* If the host supports indirect descriptor tables, and we have multiple
* buffers, then go indirect. FIXME: tune this threshold */
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
if (vq->indirect && total_sg > 1 && vq->vq.num_free)
desc = alloc_indirect(_vq, total_sg, gfp);
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
else
desc = NULL;
if (desc) {
/* Use a single buffer which doesn't continue */
indirect = true;
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
/* Set up rest to use this indirect table. */
i = 0;
descs_used = 1;
} else {
indirect = false;
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
desc = vq->vring.desc;
i = head;
descs_used = total_sg;
}
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
if (vq->vq.num_free < descs_used) {
pr_debug("Can't add buf len %i - avail = %i\n",
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
descs_used, vq->vq.num_free);
/* FIXME: for historical reasons, we force a notify here if
* there are outgoing parts to the buffer. Presumably the
* host should service the ring ASAP. */
if (out_sgs)
vq->notify(&vq->vq);
if (indirect)
kfree(desc);
END_USE(vq);
return -ENOSPC;
}
for (n = 0; n < out_sgs; n++) {
for (sg = sgs[n]; sg; sg = sg_next(sg)) {
dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_TO_DEVICE);
if (vring_mapping_error(vq, addr))
goto unmap_release;
desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT);
desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
prev = i;
i = virtio16_to_cpu(_vq->vdev, desc[i].next);
}
}
for (; n < (out_sgs + in_sgs); n++) {
for (sg = sgs[n]; sg; sg = sg_next(sg)) {
dma_addr_t addr = vring_map_one_sg(vq, sg, DMA_FROM_DEVICE);
if (vring_mapping_error(vq, addr))
goto unmap_release;
desc[i].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_NEXT | VRING_DESC_F_WRITE);
desc[i].addr = cpu_to_virtio64(_vq->vdev, addr);
desc[i].len = cpu_to_virtio32(_vq->vdev, sg->length);
prev = i;
i = virtio16_to_cpu(_vq->vdev, desc[i].next);
}
}
/* Last one doesn't continue. */
desc[prev].flags &= cpu_to_virtio16(_vq->vdev, ~VRING_DESC_F_NEXT);
if (indirect) {
/* Now that the indirect table is filled in, map it. */
dma_addr_t addr = vring_map_single(
vq, desc, total_sg * sizeof(struct vring_desc),
DMA_TO_DEVICE);
if (vring_mapping_error(vq, addr))
goto unmap_release;
vq->vring.desc[head].flags = cpu_to_virtio16(_vq->vdev, VRING_DESC_F_INDIRECT);
vq->vring.desc[head].addr = cpu_to_virtio64(_vq->vdev, addr);
vq->vring.desc[head].len = cpu_to_virtio32(_vq->vdev, total_sg * sizeof(struct vring_desc));
}
/* We're using some buffers from the free list. */
vq->vq.num_free -= descs_used;
/* Update free pointer */
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
if (indirect)
vq->free_head = virtio16_to_cpu(_vq->vdev, vq->vring.desc[head].next);
virtio_ring: unify direct/indirect code paths. virtqueue_add() populates the virtqueue descriptor table from the sgs given. If it uses an indirect descriptor table, then it puts a single descriptor in the descriptor table pointing to the kmalloc'ed indirect table where the sg is populated. Previously vring_add_indirect() did the allocation and the simple linear layout. We replace that with alloc_indirect() which allocates the indirect table then chains it like the normal descriptor table so we can reuse the core logic. This slows down pktgen by less than 1/2 a percent (which uses direct descriptors), as well as vring_bench, but it's far neater. vring_bench before: 1061485790-1104800648(1.08254e+09+/-6.6e+06)ns vring_bench after: 1125610268-1183528965(1.14172e+09+/-8e+06)ns pktgen before: 787781-796334(793165+/-2.4e+03)pps 365-369(367.5+/-1.2)Mb/sec (365530384-369498976(3.68028e+08+/-1.1e+06)bps) errors: 0 pktgen after: 779988-790404(786391+/-2.5e+03)pps 361-366(364.35+/-1.3)Mb/sec (361914432-366747456(3.64885e+08+/-1.2e+06)bps) errors: 0 Now, if we make force indirect descriptors by turning off any_header_sg in virtio_net.c: pktgen before: 713773-721062(718374+/-2.1e+03)pps 331-334(332.95+/-0.92)Mb/sec (331190672-334572768(3.33325e+08+/-9.6e+05)bps) errors: 0 pktgen after: 710542-719195(714898+/-2.4e+03)pps 329-333(331.15+/-1.1)Mb/sec (329691488-333706480(3.31713e+08+/-1.1e+06)bps) errors: 0 Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-11 00:47:38 +00:00
else
vq->free_head = i;
/* Store token and indirect buffer state. */
vq->desc_state[head].data = data;
if (indirect)
vq->desc_state[head].indir_desc = desc;
/* Put entry in available array (but don't update avail->idx until they
* do sync). */
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
avail = vq->avail_idx_shadow & (vq->vring.num - 1);
vq->vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
/* Descriptors and available array need to be set before we expose the
* new available array entries. */
virtio_wmb(vq->weak_barriers);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
vq->avail_idx_shadow++;
vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
vq->num_added++;
pr_debug("Added buffer head %i to %p\n", head, vq);
END_USE(vq);
/* This is very unlikely, but theoretically possible. Kick
* just in case. */
if (unlikely(vq->num_added == (1 << 16) - 1))
virtqueue_kick(_vq);
return 0;
unmap_release:
err_idx = i;
i = head;
for (n = 0; n < total_sg; n++) {
if (i == err_idx)
break;
vring_unmap_one(vq, &desc[i]);
i = vq->vring.desc[i].next;
}
vq->vq.num_free += total_sg;
if (indirect)
kfree(desc);
END_USE(vq);
return -EIO;
}
/**
* virtqueue_add_sgs - expose buffers to other end
* @vq: the struct virtqueue we're talking about.
* @sgs: array of terminated scatterlists.
* @out_num: the number of scatterlists readable by other side
* @in_num: the number of scatterlists which are writable (after readable ones)
* @data: the token identifying the buffer.
* @gfp: how to do memory allocations (if necessary).
*
* Caller must ensure we don't call this with other virtqueue operations
* at the same time (except where noted).
*
* Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
*/
int virtqueue_add_sgs(struct virtqueue *_vq,
struct scatterlist *sgs[],
unsigned int out_sgs,
unsigned int in_sgs,
void *data,
gfp_t gfp)
{
unsigned int i, total_sg = 0;
/* Count them first. */
for (i = 0; i < out_sgs + in_sgs; i++) {
struct scatterlist *sg;
for (sg = sgs[i]; sg; sg = sg_next(sg))
total_sg++;
}
return virtqueue_add(_vq, sgs, total_sg, out_sgs, in_sgs, data, gfp);
}
EXPORT_SYMBOL_GPL(virtqueue_add_sgs);
/**
* virtqueue_add_outbuf - expose output buffers to other end
* @vq: the struct virtqueue we're talking about.
* @sg: scatterlist (must be well-formed and terminated!)
* @num: the number of entries in @sg readable by other side
* @data: the token identifying the buffer.
* @gfp: how to do memory allocations (if necessary).
*
* Caller must ensure we don't call this with other virtqueue operations
* at the same time (except where noted).
*
* Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
*/
int virtqueue_add_outbuf(struct virtqueue *vq,
struct scatterlist *sg, unsigned int num,
void *data,
gfp_t gfp)
{
return virtqueue_add(vq, &sg, num, 1, 0, data, gfp);
}
EXPORT_SYMBOL_GPL(virtqueue_add_outbuf);
/**
* virtqueue_add_inbuf - expose input buffers to other end
* @vq: the struct virtqueue we're talking about.
* @sg: scatterlist (must be well-formed and terminated!)
* @num: the number of entries in @sg writable by other side
* @data: the token identifying the buffer.
* @gfp: how to do memory allocations (if necessary).
*
* Caller must ensure we don't call this with other virtqueue operations
* at the same time (except where noted).
*
* Returns zero or a negative error (ie. ENOSPC, ENOMEM, EIO).
*/
int virtqueue_add_inbuf(struct virtqueue *vq,
struct scatterlist *sg, unsigned int num,
void *data,
gfp_t gfp)
{
return virtqueue_add(vq, &sg, num, 0, 1, data, gfp);
}
EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
/**
* virtqueue_kick_prepare - first half of split virtqueue_kick call.
* @vq: the struct virtqueue
*
* Instead of virtqueue_kick(), you can do:
* if (virtqueue_kick_prepare(vq))
* virtqueue_notify(vq);
*
* This is sometimes useful because the virtqueue_kick_prepare() needs
* to be serialized, but the actual virtqueue_notify() call does not.
*/
bool virtqueue_kick_prepare(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
u16 new, old;
bool needs_kick;
START_USE(vq);
/* We need to expose available array entries before checking avail
* event. */
virtio_mb(vq->weak_barriers);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
old = vq->avail_idx_shadow - vq->num_added;
new = vq->avail_idx_shadow;
vq->num_added = 0;
#ifdef DEBUG
if (vq->last_add_time_valid) {
WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
vq->last_add_time)) > 100);
}
vq->last_add_time_valid = false;
#endif
if (vq->event) {
needs_kick = vring_need_event(virtio16_to_cpu(_vq->vdev, vring_avail_event(&vq->vring)),
new, old);
} else {
needs_kick = !(vq->vring.used->flags & cpu_to_virtio16(_vq->vdev, VRING_USED_F_NO_NOTIFY));
}
END_USE(vq);
return needs_kick;
}
EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
/**
* virtqueue_notify - second half of split virtqueue_kick call.
* @vq: the struct virtqueue
*
* This does not need to be serialized.
*
* Returns false if host notify failed or queue is broken, otherwise true.
*/
bool virtqueue_notify(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
if (unlikely(vq->broken))
return false;
/* Prod other side to tell it about changes. */
if (!vq->notify(_vq)) {
vq->broken = true;
return false;
}
return true;
}
EXPORT_SYMBOL_GPL(virtqueue_notify);
/**
* virtqueue_kick - update after add_buf
* @vq: the struct virtqueue
*
* After one or more virtqueue_add_* calls, invoke this to kick
* the other side.
*
* Caller must ensure we don't call this with other virtqueue
* operations at the same time (except where noted).
*
* Returns false if kick failed, otherwise true.
*/
bool virtqueue_kick(struct virtqueue *vq)
{
if (virtqueue_kick_prepare(vq))
return virtqueue_notify(vq);
return true;
}
EXPORT_SYMBOL_GPL(virtqueue_kick);
static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
{
unsigned int i, j;
u16 nextflag = cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_NEXT);
/* Clear data ptr. */
vq->desc_state[head].data = NULL;
/* Put back on free list: unmap first-level descriptors and find end */
i = head;
while (vq->vring.desc[i].flags & nextflag) {
vring_unmap_one(vq, &vq->vring.desc[i]);
i = virtio16_to_cpu(vq->vq.vdev, vq->vring.desc[i].next);
vq->vq.num_free++;
}
vring_unmap_one(vq, &vq->vring.desc[i]);
vq->vring.desc[i].next = cpu_to_virtio16(vq->vq.vdev, vq->free_head);
vq->free_head = head;
/* Plus final descriptor */
vq->vq.num_free++;
/* Free the indirect table, if any, now that it's unmapped. */
if (vq->desc_state[head].indir_desc) {
struct vring_desc *indir_desc = vq->desc_state[head].indir_desc;
u32 len = virtio32_to_cpu(vq->vq.vdev, vq->vring.desc[head].len);
BUG_ON(!(vq->vring.desc[head].flags &
cpu_to_virtio16(vq->vq.vdev, VRING_DESC_F_INDIRECT)));
BUG_ON(len == 0 || len % sizeof(struct vring_desc));
for (j = 0; j < len / sizeof(struct vring_desc); j++)
vring_unmap_one(vq, &indir_desc[j]);
kfree(vq->desc_state[head].indir_desc);
vq->desc_state[head].indir_desc = NULL;
}
}
static inline bool more_used(const struct vring_virtqueue *vq)
{
return vq->last_used_idx != virtio16_to_cpu(vq->vq.vdev, vq->vring.used->idx);
}
/**
* virtqueue_get_buf - get the next used buffer
* @vq: the struct virtqueue we're talking about.
* @len: the length written into the buffer
*
* If the driver wrote data into the buffer, @len will be set to the
* amount written. This means you don't need to clear the buffer
* beforehand to ensure there's no data leakage in the case of short
* writes.
*
* Caller must ensure we don't call this with other virtqueue
* operations at the same time (except where noted).
*
* Returns NULL if there are no used buffers, or the "data" token
* handed to virtqueue_add_*().
*/
void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
{
struct vring_virtqueue *vq = to_vvq(_vq);
void *ret;
unsigned int i;
u16 last_used;
START_USE(vq);
if (unlikely(vq->broken)) {
END_USE(vq);
return NULL;
}
if (!more_used(vq)) {
pr_debug("No more buffers in queue\n");
END_USE(vq);
return NULL;
}
/* Only get used array entries after they have been exposed by host. */
virtio_rmb(vq->weak_barriers);
last_used = (vq->last_used_idx & (vq->vring.num - 1));
i = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].id);
*len = virtio32_to_cpu(_vq->vdev, vq->vring.used->ring[last_used].len);
if (unlikely(i >= vq->vring.num)) {
BAD_RING(vq, "id %u out of range\n", i);
return NULL;
}
if (unlikely(!vq->desc_state[i].data)) {
BAD_RING(vq, "id %u is not a head!\n", i);
return NULL;
}
/* detach_buf clears data, so grab it now. */
ret = vq->desc_state[i].data;
detach_buf(vq, i);
vq->last_used_idx++;
/* If we expect an interrupt for the next entry, tell host
* by writing event index and flush out the write before
* the read in the next get_buf call. */
if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT))
virtio_store_mb(vq->weak_barriers,
&vring_used_event(&vq->vring),
cpu_to_virtio16(_vq->vdev, vq->last_used_idx));
#ifdef DEBUG
vq->last_add_time_valid = false;
#endif
END_USE(vq);
return ret;
}
EXPORT_SYMBOL_GPL(virtqueue_get_buf);
/**
* virtqueue_disable_cb - disable callbacks
* @vq: the struct virtqueue we're talking about.
*
* Note that this is not necessarily synchronous, hence unreliable and only
* useful as an optimization.
*
* Unlike other operations, this need not be serialized.
*/
void virtqueue_disable_cb(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
if (!(vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT)) {
vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
if (!vq->event)
vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
}
}
EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
/**
* virtqueue_enable_cb_prepare - restart callbacks after disable_cb
* @vq: the struct virtqueue we're talking about.
*
* This re-enables callbacks; it returns current queue state
* in an opaque unsigned value. This value should be later tested by
* virtqueue_poll, to detect a possible race between the driver checking for
* more work, and enabling callbacks.
*
* Caller must ensure we don't call this with other virtqueue
* operations at the same time (except where noted).
*/
unsigned virtqueue_enable_cb_prepare(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
u16 last_used_idx;
START_USE(vq);
/* We optimistically turn back on interrupts, then check if there was
* more to do. */
/* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
* either clear the flags bit or point the event index at the next
* entry. Always do both to keep code simple. */
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
if (!vq->event)
vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
}
vring_used_event(&vq->vring) = cpu_to_virtio16(_vq->vdev, last_used_idx = vq->last_used_idx);
END_USE(vq);
return last_used_idx;
}
EXPORT_SYMBOL_GPL(virtqueue_enable_cb_prepare);
/**
* virtqueue_poll - query pending used buffers
* @vq: the struct virtqueue we're talking about.
* @last_used_idx: virtqueue state (from call to virtqueue_enable_cb_prepare).
*
* Returns "true" if there are pending used buffers in the queue.
*
* This does not need to be serialized.
*/
bool virtqueue_poll(struct virtqueue *_vq, unsigned last_used_idx)
{
struct vring_virtqueue *vq = to_vvq(_vq);
virtio_mb(vq->weak_barriers);
return (u16)last_used_idx != virtio16_to_cpu(_vq->vdev, vq->vring.used->idx);
}
EXPORT_SYMBOL_GPL(virtqueue_poll);
/**
* virtqueue_enable_cb - restart callbacks after disable_cb.
* @vq: the struct virtqueue we're talking about.
*
* This re-enables callbacks; it returns "false" if there are pending
* buffers in the queue, to detect a possible race between the driver
* checking for more work, and enabling callbacks.
*
* Caller must ensure we don't call this with other virtqueue
* operations at the same time (except where noted).
*/
bool virtqueue_enable_cb(struct virtqueue *_vq)
{
unsigned last_used_idx = virtqueue_enable_cb_prepare(_vq);
return !virtqueue_poll(_vq, last_used_idx);
}
EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
/**
* virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
* @vq: the struct virtqueue we're talking about.
*
* This re-enables callbacks but hints to the other side to delay
* interrupts until most of the available buffers have been processed;
* it returns "false" if there are many pending buffers in the queue,
* to detect a possible race between the driver checking for more work,
* and enabling callbacks.
*
* Caller must ensure we don't call this with other virtqueue
* operations at the same time (except where noted).
*/
bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
u16 bufs;
START_USE(vq);
/* We optimistically turn back on interrupts, then check if there was
* more to do. */
/* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
* either clear the flags bit or point the event index at the next
* entry. Always update the event index to keep code simple. */
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
if (vq->avail_flags_shadow & VRING_AVAIL_F_NO_INTERRUPT) {
vq->avail_flags_shadow &= ~VRING_AVAIL_F_NO_INTERRUPT;
if (!vq->event)
vq->vring.avail->flags = cpu_to_virtio16(_vq->vdev, vq->avail_flags_shadow);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
}
/* TODO: tune this threshold */
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
bufs = (u16)(vq->avail_idx_shadow - vq->last_used_idx) * 3 / 4;
virtio_store_mb(vq->weak_barriers,
&vring_used_event(&vq->vring),
cpu_to_virtio16(_vq->vdev, vq->last_used_idx + bufs));
if (unlikely((u16)(virtio16_to_cpu(_vq->vdev, vq->vring.used->idx) - vq->last_used_idx) > bufs)) {
END_USE(vq);
return false;
}
END_USE(vq);
return true;
}
EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
/**
* virtqueue_detach_unused_buf - detach first unused buffer
* @vq: the struct virtqueue we're talking about.
*
* Returns NULL or the "data" token handed to virtqueue_add_*().
* This is not valid on an active queue; it is useful only for device
* shutdown.
*/
void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
unsigned int i;
void *buf;
START_USE(vq);
for (i = 0; i < vq->vring.num; i++) {
if (!vq->desc_state[i].data)
continue;
/* detach_buf clears data, so grab it now. */
buf = vq->desc_state[i].data;
detach_buf(vq, i);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
vq->avail_idx_shadow--;
vq->vring.avail->idx = cpu_to_virtio16(_vq->vdev, vq->avail_idx_shadow);
END_USE(vq);
return buf;
}
/* That should have freed everything. */
BUG_ON(vq->vq.num_free != vq->vring.num);
END_USE(vq);
return NULL;
}
EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
irqreturn_t vring_interrupt(int irq, void *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
if (!more_used(vq)) {
pr_debug("virtqueue interrupt with no work for %p\n", vq);
return IRQ_NONE;
}
if (unlikely(vq->broken))
return IRQ_HANDLED;
pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
if (vq->vq.callback)
vq->vq.callback(&vq->vq);
return IRQ_HANDLED;
}
EXPORT_SYMBOL_GPL(vring_interrupt);
struct virtqueue *__vring_new_virtqueue(unsigned int index,
struct vring vring,
struct virtio_device *vdev,
bool weak_barriers,
bool (*notify)(struct virtqueue *),
void (*callback)(struct virtqueue *),
const char *name)
{
unsigned int i;
struct vring_virtqueue *vq;
vq = kmalloc(sizeof(*vq) + vring.num * sizeof(struct vring_desc_state),
GFP_KERNEL);
if (!vq)
return NULL;
vq->vring = vring;
vq->vq.callback = callback;
vq->vq.vdev = vdev;
vq->vq.name = name;
vq->vq.num_free = vring.num;
vq->vq.index = index;
vq->we_own_ring = false;
vq->queue_dma_addr = 0;
vq->queue_size_in_bytes = 0;
vq->notify = notify;
vq->weak_barriers = weak_barriers;
vq->broken = false;
vq->last_used_idx = 0;
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
vq->avail_flags_shadow = 0;
vq->avail_idx_shadow = 0;
vq->num_added = 0;
list_add_tail(&vq->vq.list, &vdev->vqs);
#ifdef DEBUG
vq->in_use = false;
vq->last_add_time_valid = false;
#endif
vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
/* No callback? Tell other side not to bother us. */
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
if (!callback) {
vq->avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;
if (!vq->event)
vq->vring.avail->flags = cpu_to_virtio16(vdev, vq->avail_flags_shadow);
virtio_ring: shadow available ring flags & index Improves cacheline transfer flow of available ring header. Virtqueues are implemented as a pair of rings, one producer->consumer avail ring and one consumer->producer used ring; preceding the avail ring in memory are two contiguous u16 fields -- avail->flags and avail->idx. A producer posts work by writing to avail->idx and a consumer reads avail->idx. The flags and idx fields only need to be written by a producer CPU and only read by a consumer CPU; when the producer and consumer are running on different CPUs and the virtio_ring code is structured to only have source writes/sink reads, we can continuously transfer the avail header cacheline between 'M' states between cores. This flow optimizes core -> core bandwidth on certain CPUs. (see: "Software Optimization Guide for AMD Family 15h Processors", Section 11.6; similar language appears in the 10h guide and should apply to CPUs w/ exclusive caches, using LLC as a transfer cache) Unfortunately the existing virtio_ring code issued reads to the avail->idx and read-modify-writes to avail->flags on the producer. This change shadows the flags and index fields in producer memory; the vring code now reads from the shadows and only ever writes to avail->flags and avail->idx, allowing the cacheline to transfer core -> core optimally. In a concurrent version of vring_bench, the time required for 10,000,000 buffer checkout/returns was reduced by ~2% (average across many runs) on an AMD Piledriver (15h) CPU: (w/o shadowing): Performance counter stats for './vring_bench': 5,451,082,016 L1-dcache-loads ... 2.221477739 seconds time elapsed (w/ shadowing): Performance counter stats for './vring_bench': 5,405,701,361 L1-dcache-loads ... 2.168405376 seconds time elapsed The further away (in a NUMA sense) virtio producers and consumers are from each other, the more we expect to benefit. Physical implementations of virtio devices and implementations of virtio where the consumer polls vring avail indexes (vhost) should also benefit. Signed-off-by: Venkatesh Srinivas <venkateshs@google.com> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2015-11-11 00:21:07 +00:00
}
/* Put everything in free lists. */
vq->free_head = 0;
for (i = 0; i < vring.num-1; i++)
vq->vring.desc[i].next = cpu_to_virtio16(vdev, i + 1);
memset(vq->desc_state, 0, vring.num * sizeof(struct vring_desc_state));
return &vq->vq;
}
EXPORT_SYMBOL_GPL(__vring_new_virtqueue);
static void *vring_alloc_queue(struct virtio_device *vdev, size_t size,
dma_addr_t *dma_handle, gfp_t flag)
{
if (vring_use_dma_api(vdev)) {
return dma_alloc_coherent(vdev->dev.parent, size,
dma_handle, flag);
} else {
void *queue = alloc_pages_exact(PAGE_ALIGN(size), flag);
if (queue) {
phys_addr_t phys_addr = virt_to_phys(queue);
*dma_handle = (dma_addr_t)phys_addr;
/*
* Sanity check: make sure we dind't truncate
* the address. The only arches I can find that
* have 64-bit phys_addr_t but 32-bit dma_addr_t
* are certain non-highmem MIPS and x86
* configurations, but these configurations
* should never allocate physical pages above 32
* bits, so this is fine. Just in case, throw a
* warning and abort if we end up with an
* unrepresentable address.
*/
if (WARN_ON_ONCE(*dma_handle != phys_addr)) {
free_pages_exact(queue, PAGE_ALIGN(size));
return NULL;
}
}
return queue;
}
}
static void vring_free_queue(struct virtio_device *vdev, size_t size,
void *queue, dma_addr_t dma_handle)
{
if (vring_use_dma_api(vdev)) {
dma_free_coherent(vdev->dev.parent, size, queue, dma_handle);
} else {
free_pages_exact(queue, PAGE_ALIGN(size));
}
}
struct virtqueue *vring_create_virtqueue(
unsigned int index,
unsigned int num,
unsigned int vring_align,
struct virtio_device *vdev,
bool weak_barriers,
bool may_reduce_num,
bool (*notify)(struct virtqueue *),
void (*callback)(struct virtqueue *),
const char *name)
{
struct virtqueue *vq;
void *queue = NULL;
dma_addr_t dma_addr;
size_t queue_size_in_bytes;
struct vring vring;
/* We assume num is a power of 2. */
if (num & (num - 1)) {
dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
return NULL;
}
/* TODO: allocate each queue chunk individually */
for (; num && vring_size(num, vring_align) > PAGE_SIZE; num /= 2) {
queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
&dma_addr,
GFP_KERNEL|__GFP_NOWARN|__GFP_ZERO);
if (queue)
break;
}
if (!num)
return NULL;
if (!queue) {
/* Try to get a single page. You are my only hope! */
queue = vring_alloc_queue(vdev, vring_size(num, vring_align),
&dma_addr, GFP_KERNEL|__GFP_ZERO);
}
if (!queue)
return NULL;
queue_size_in_bytes = vring_size(num, vring_align);
vring_init(&vring, num, queue, vring_align);
vq = __vring_new_virtqueue(index, vring, vdev, weak_barriers,
notify, callback, name);
if (!vq) {
vring_free_queue(vdev, queue_size_in_bytes, queue,
dma_addr);
return NULL;
}
to_vvq(vq)->queue_dma_addr = dma_addr;
to_vvq(vq)->queue_size_in_bytes = queue_size_in_bytes;
to_vvq(vq)->we_own_ring = true;
return vq;
}
EXPORT_SYMBOL_GPL(vring_create_virtqueue);
struct virtqueue *vring_new_virtqueue(unsigned int index,
unsigned int num,
unsigned int vring_align,
struct virtio_device *vdev,
bool weak_barriers,
void *pages,
bool (*notify)(struct virtqueue *vq),
void (*callback)(struct virtqueue *vq),
const char *name)
{
struct vring vring;
vring_init(&vring, num, pages, vring_align);
return __vring_new_virtqueue(index, vring, vdev, weak_barriers,
notify, callback, name);
}
EXPORT_SYMBOL_GPL(vring_new_virtqueue);
void vring_del_virtqueue(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
if (vq->we_own_ring) {
vring_free_queue(vq->vq.vdev, vq->queue_size_in_bytes,
vq->vring.desc, vq->queue_dma_addr);
}
list_del(&_vq->list);
kfree(vq);
}
EXPORT_SYMBOL_GPL(vring_del_virtqueue);
/* Manipulates transport-specific feature bits. */
void vring_transport_features(struct virtio_device *vdev)
{
unsigned int i;
for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
switch (i) {
case VIRTIO_RING_F_INDIRECT_DESC:
break;
case VIRTIO_RING_F_EVENT_IDX:
break;
case VIRTIO_F_VERSION_1:
break;
virtio: new feature to detect IOMMU device quirk The interaction between virtio and IOMMUs is messy. On most systems with virtio, physical addresses match bus addresses, and it doesn't particularly matter which one we use to program the device. On some systems, including Xen and any system with a physical device that speaks virtio behind a physical IOMMU, we must program the IOMMU for virtio DMA to work at all. On other systems, including SPARC and PPC64, virtio-pci devices are enumerated as though they are behind an IOMMU, but the virtio host ignores the IOMMU, so we must either pretend that the IOMMU isn't there or somehow map everything as the identity. Add a feature bit to detect that quirk: VIRTIO_F_IOMMU_PLATFORM. Any device with this feature bit set to 0 needs a quirk and has to be passed physical addresses (as opposed to bus addresses) even though the device is behind an IOMMU. Note: it has to be a per-device quirk because for example, there could be a mix of passed-through and virtual virtio devices. As another example, some devices could be implemented by an out of process hypervisor backend (in case of qemu vhost, or vhost-user) and so support for an IOMMU needs to be coded up separately. It would be cleanest to handle this in IOMMU core code, but that needs per-device DMA ops. While we are waiting for that to be implemented, use a work-around in virtio core. Note: a "noiommu" feature is a quirk - add a wrapper to make that clear. Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2016-04-18 09:58:14 +00:00
case VIRTIO_F_IOMMU_PLATFORM:
break;
default:
/* We don't understand this bit. */
__virtio_clear_bit(vdev, i);
}
}
}
EXPORT_SYMBOL_GPL(vring_transport_features);
/**
* virtqueue_get_vring_size - return the size of the virtqueue's vring
* @vq: the struct virtqueue containing the vring of interest.
*
* Returns the size of the vring. This is mainly used for boasting to
* userspace. Unlike other operations, this need not be serialized.
*/
unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
return vq->vring.num;
}
EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
bool virtqueue_is_broken(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
return vq->broken;
}
EXPORT_SYMBOL_GPL(virtqueue_is_broken);
/*
* This should prevent the device from being used, allowing drivers to
* recover. You may need to grab appropriate locks to flush.
*/
void virtio_break_device(struct virtio_device *dev)
{
struct virtqueue *_vq;
list_for_each_entry(_vq, &dev->vqs, list) {
struct vring_virtqueue *vq = to_vvq(_vq);
vq->broken = true;
}
}
EXPORT_SYMBOL_GPL(virtio_break_device);
dma_addr_t virtqueue_get_desc_addr(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
BUG_ON(!vq->we_own_ring);
return vq->queue_dma_addr;
}
EXPORT_SYMBOL_GPL(virtqueue_get_desc_addr);
dma_addr_t virtqueue_get_avail_addr(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
BUG_ON(!vq->we_own_ring);
return vq->queue_dma_addr +
((char *)vq->vring.avail - (char *)vq->vring.desc);
}
EXPORT_SYMBOL_GPL(virtqueue_get_avail_addr);
dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
{
struct vring_virtqueue *vq = to_vvq(_vq);
BUG_ON(!vq->we_own_ring);
return vq->queue_dma_addr +
((char *)vq->vring.used - (char *)vq->vring.desc);
}
EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
const struct vring *virtqueue_get_vring(struct virtqueue *vq)
{
return &to_vvq(vq)->vring;
}
EXPORT_SYMBOL_GPL(virtqueue_get_vring);
MODULE_LICENSE("GPL");