Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1 | /* Copyright (C) 2009 Red Hat, Inc. |
| 2 | * Author: Michael S. Tsirkin <mst@redhat.com> |
| 3 | * |
| 4 | * This work is licensed under the terms of the GNU GPL, version 2. |
| 5 | * |
| 6 | * virtio-net server in host kernel. |
| 7 | */ |
| 8 | |
| 9 | #include <linux/compat.h> |
| 10 | #include <linux/eventfd.h> |
| 11 | #include <linux/vhost.h> |
| 12 | #include <linux/virtio_net.h> |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 13 | #include <linux/miscdevice.h> |
| 14 | #include <linux/module.h> |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 15 | #include <linux/moduleparam.h> |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 16 | #include <linux/mutex.h> |
| 17 | #include <linux/workqueue.h> |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 18 | #include <linux/file.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 19 | #include <linux/slab.h> |
Michael S. Tsirkin | 23cc5a9 | 2013-01-23 21:46:47 +0100 | [diff] [blame] | 20 | #include <linux/vmalloc.h> |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 21 | |
| 22 | #include <linux/net.h> |
| 23 | #include <linux/if_packet.h> |
| 24 | #include <linux/if_arp.h> |
| 25 | #include <linux/if_tun.h> |
Arnd Bergmann | 501c774 | 2010-02-18 05:46:50 +0000 | [diff] [blame] | 26 | #include <linux/if_macvlan.h> |
Basil Gor | c53cff5e | 2012-05-03 22:55:23 +0000 | [diff] [blame] | 27 | #include <linux/if_vlan.h> |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 28 | |
| 29 | #include <net/sock.h> |
| 30 | |
| 31 | #include "vhost.h" |
| 32 | |
Michael S. Tsirkin | f9611c4 | 2012-12-06 14:56:00 +0200 | [diff] [blame] | 33 | static int experimental_zcopytx = 1; |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 34 | module_param(experimental_zcopytx, int, 0444); |
Michael S. Tsirkin | f9611c4 | 2012-12-06 14:56:00 +0200 | [diff] [blame] | 35 | MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;" |
| 36 | " 1 -Enable; 0 - Disable"); |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 37 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 38 | /* Max number of bytes transferred before requeueing the job. |
| 39 | * Using this limit prevents one virtqueue from starving others. */ |
| 40 | #define VHOST_NET_WEIGHT 0x80000 |
| 41 | |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 42 | /* MAX number of TX used buffers for outstanding zerocopy */ |
| 43 | #define VHOST_MAX_PEND 128 |
| 44 | #define VHOST_GOODCOPY_LEN 256 |
| 45 | |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 46 | /* |
| 47 | * For transmit, used buffer len is unused; we override it to track buffer |
| 48 | * status internally; used for zerocopy tx only. |
| 49 | */ |
| 50 | /* Lower device DMA failed */ |
Michael S. Tsirkin | bf99573 | 2014-10-24 11:49:27 +0300 | [diff] [blame] | 51 | #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3) |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 52 | /* Lower device DMA done */ |
Michael S. Tsirkin | bf99573 | 2014-10-24 11:49:27 +0300 | [diff] [blame] | 53 | #define VHOST_DMA_DONE_LEN ((__force __virtio32)2) |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 54 | /* Lower device DMA in progress */ |
Michael S. Tsirkin | bf99573 | 2014-10-24 11:49:27 +0300 | [diff] [blame] | 55 | #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1) |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 56 | /* Buffer unused */ |
Michael S. Tsirkin | bf99573 | 2014-10-24 11:49:27 +0300 | [diff] [blame] | 57 | #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0) |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 58 | |
Michael S. Tsirkin | bf99573 | 2014-10-24 11:49:27 +0300 | [diff] [blame] | 59 | #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN) |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 60 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 61 | enum { |
Asias He | 8570a6e | 2013-05-06 16:38:20 +0800 | [diff] [blame] | 62 | VHOST_NET_FEATURES = VHOST_FEATURES | |
| 63 | (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) | |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 64 | (1ULL << VIRTIO_NET_F_MRG_RXBUF) | |
| 65 | (1ULL << VIRTIO_F_IOMMU_PLATFORM) |
Asias He | 8570a6e | 2013-05-06 16:38:20 +0800 | [diff] [blame] | 66 | }; |
| 67 | |
| 68 | enum { |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 69 | VHOST_NET_VQ_RX = 0, |
| 70 | VHOST_NET_VQ_TX = 1, |
| 71 | VHOST_NET_VQ_MAX = 2, |
| 72 | }; |
| 73 | |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 74 | struct vhost_net_ubuf_ref { |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 75 | /* refcount follows semantics similar to kref: |
| 76 | * 0: object is released |
| 77 | * 1: no outstanding ubufs |
| 78 | * >1: outstanding ubufs |
| 79 | */ |
| 80 | atomic_t refcount; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 81 | wait_queue_head_t wait; |
| 82 | struct vhost_virtqueue *vq; |
| 83 | }; |
| 84 | |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 85 | struct vhost_net_virtqueue { |
| 86 | struct vhost_virtqueue vq; |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 87 | size_t vhost_hlen; |
| 88 | size_t sock_hlen; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 89 | /* vhost zerocopy support fields below: */ |
| 90 | /* last used idx for outstanding DMA zerocopy buffers */ |
| 91 | int upend_idx; |
| 92 | /* first used idx for DMA done zerocopy buffers */ |
| 93 | int done_idx; |
| 94 | /* an array of userspace buffers info */ |
| 95 | struct ubuf_info *ubuf_info; |
| 96 | /* Reference counting for outstanding ubufs. |
| 97 | * Protected by vq mutex. Writers must also take device mutex. */ |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 98 | struct vhost_net_ubuf_ref *ubufs; |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 99 | }; |
| 100 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 101 | struct vhost_net { |
| 102 | struct vhost_dev dev; |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 103 | struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX]; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 104 | struct vhost_poll poll[VHOST_NET_VQ_MAX]; |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 105 | /* Number of TX recently submitted. |
| 106 | * Protected by tx vq lock. */ |
| 107 | unsigned tx_packets; |
| 108 | /* Number of times zerocopy TX recently failed. |
| 109 | * Protected by tx vq lock. */ |
| 110 | unsigned tx_zcopy_err; |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 111 | /* Flush in progress. Protected by tx vq lock. */ |
| 112 | bool tx_flush; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 113 | }; |
| 114 | |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 115 | static unsigned vhost_net_zcopy_mask __read_mostly; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 116 | |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 117 | static void vhost_net_enable_zcopy(int vq) |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 118 | { |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 119 | vhost_net_zcopy_mask |= 0x1 << vq; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 120 | } |
| 121 | |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 122 | static struct vhost_net_ubuf_ref * |
| 123 | vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy) |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 124 | { |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 125 | struct vhost_net_ubuf_ref *ubufs; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 126 | /* No zero copy backend? Nothing to count. */ |
| 127 | if (!zcopy) |
| 128 | return NULL; |
| 129 | ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL); |
| 130 | if (!ubufs) |
| 131 | return ERR_PTR(-ENOMEM); |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 132 | atomic_set(&ubufs->refcount, 1); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 133 | init_waitqueue_head(&ubufs->wait); |
| 134 | ubufs->vq = vq; |
| 135 | return ubufs; |
| 136 | } |
| 137 | |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 138 | static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs) |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 139 | { |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 140 | int r = atomic_sub_return(1, &ubufs->refcount); |
| 141 | if (unlikely(!r)) |
| 142 | wake_up(&ubufs->wait); |
| 143 | return r; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 144 | } |
| 145 | |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 146 | static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs) |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 147 | { |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 148 | vhost_net_ubuf_put(ubufs); |
| 149 | wait_event(ubufs->wait, !atomic_read(&ubufs->refcount)); |
Michael S. Tsirkin | c38e39c | 2013-06-25 17:29:46 +0300 | [diff] [blame] | 150 | } |
| 151 | |
| 152 | static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs) |
| 153 | { |
| 154 | vhost_net_ubuf_put_and_wait(ubufs); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 155 | kfree(ubufs); |
| 156 | } |
| 157 | |
Asias He | b1ad849 | 2013-05-06 11:16:00 +0800 | [diff] [blame] | 158 | static void vhost_net_clear_ubuf_info(struct vhost_net *n) |
| 159 | { |
Asias He | b1ad849 | 2013-05-06 11:16:00 +0800 | [diff] [blame] | 160 | int i; |
| 161 | |
Michael S. Tsirkin | 288cfe7 | 2013-06-06 15:20:46 +0300 | [diff] [blame] | 162 | for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { |
| 163 | kfree(n->vqs[i].ubuf_info); |
| 164 | n->vqs[i].ubuf_info = NULL; |
Asias He | b1ad849 | 2013-05-06 11:16:00 +0800 | [diff] [blame] | 165 | } |
| 166 | } |
| 167 | |
Asias He | 0a1febf | 2013-06-05 21:17:38 +0800 | [diff] [blame] | 168 | static int vhost_net_set_ubuf_info(struct vhost_net *n) |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 169 | { |
| 170 | bool zcopy; |
| 171 | int i; |
| 172 | |
Michael S. Tsirkin | 288cfe7 | 2013-06-06 15:20:46 +0300 | [diff] [blame] | 173 | for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 174 | zcopy = vhost_net_zcopy_mask & (0x1 << i); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 175 | if (!zcopy) |
| 176 | continue; |
| 177 | n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) * |
| 178 | UIO_MAXIOV, GFP_KERNEL); |
| 179 | if (!n->vqs[i].ubuf_info) |
| 180 | goto err; |
| 181 | } |
| 182 | return 0; |
| 183 | |
| 184 | err: |
Michael S. Tsirkin | 288cfe7 | 2013-06-06 15:20:46 +0300 | [diff] [blame] | 185 | vhost_net_clear_ubuf_info(n); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 186 | return -ENOMEM; |
| 187 | } |
| 188 | |
Asias He | 0a1febf | 2013-06-05 21:17:38 +0800 | [diff] [blame] | 189 | static void vhost_net_vq_reset(struct vhost_net *n) |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 190 | { |
| 191 | int i; |
| 192 | |
Michael S. Tsirkin | 288cfe7 | 2013-06-06 15:20:46 +0300 | [diff] [blame] | 193 | vhost_net_clear_ubuf_info(n); |
| 194 | |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 195 | for (i = 0; i < VHOST_NET_VQ_MAX; i++) { |
| 196 | n->vqs[i].done_idx = 0; |
| 197 | n->vqs[i].upend_idx = 0; |
| 198 | n->vqs[i].ubufs = NULL; |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 199 | n->vqs[i].vhost_hlen = 0; |
| 200 | n->vqs[i].sock_hlen = 0; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 201 | } |
| 202 | |
| 203 | } |
| 204 | |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 205 | static void vhost_net_tx_packet(struct vhost_net *net) |
| 206 | { |
| 207 | ++net->tx_packets; |
| 208 | if (net->tx_packets < 1024) |
| 209 | return; |
| 210 | net->tx_packets = 0; |
| 211 | net->tx_zcopy_err = 0; |
| 212 | } |
| 213 | |
| 214 | static void vhost_net_tx_err(struct vhost_net *net) |
| 215 | { |
| 216 | ++net->tx_zcopy_err; |
| 217 | } |
| 218 | |
| 219 | static bool vhost_net_tx_select_zcopy(struct vhost_net *net) |
| 220 | { |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 221 | /* TX flush waits for outstanding DMAs to be done. |
| 222 | * Don't start new DMAs. |
| 223 | */ |
| 224 | return !net->tx_flush && |
| 225 | net->tx_packets / 64 >= net->tx_zcopy_err; |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 226 | } |
| 227 | |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 228 | static bool vhost_sock_zcopy(struct socket *sock) |
| 229 | { |
| 230 | return unlikely(experimental_zcopytx) && |
| 231 | sock_flag(sock->sk, SOCK_ZEROCOPY); |
| 232 | } |
| 233 | |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 234 | /* In case of DMA done not in order in lower device driver for some reason. |
| 235 | * upend_idx is used to track end of used idx, done_idx is used to track head |
| 236 | * of used idx. Once lower device DMA done contiguously, we will signal KVM |
| 237 | * guest used idx. |
| 238 | */ |
Jason Wang | 094afe7 | 2013-09-02 16:40:56 +0800 | [diff] [blame] | 239 | static void vhost_zerocopy_signal_used(struct vhost_net *net, |
| 240 | struct vhost_virtqueue *vq) |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 241 | { |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 242 | struct vhost_net_virtqueue *nvq = |
| 243 | container_of(vq, struct vhost_net_virtqueue, vq); |
Jason Wang | c92112a | 2013-09-02 16:40:57 +0800 | [diff] [blame] | 244 | int i, add; |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 245 | int j = 0; |
| 246 | |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 247 | for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) { |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 248 | if (vq->heads[i].len == VHOST_DMA_FAILED_LEN) |
| 249 | vhost_net_tx_err(net); |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 250 | if (VHOST_DMA_IS_DONE(vq->heads[i].len)) { |
| 251 | vq->heads[i].len = VHOST_DMA_CLEAR_LEN; |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 252 | ++j; |
| 253 | } else |
| 254 | break; |
| 255 | } |
Jason Wang | c92112a | 2013-09-02 16:40:57 +0800 | [diff] [blame] | 256 | while (j) { |
| 257 | add = min(UIO_MAXIOV - nvq->done_idx, j); |
| 258 | vhost_add_used_and_signal_n(vq->dev, vq, |
| 259 | &vq->heads[nvq->done_idx], add); |
| 260 | nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV; |
| 261 | j -= add; |
| 262 | } |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 263 | } |
| 264 | |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 265 | static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success) |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 266 | { |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 267 | struct vhost_net_ubuf_ref *ubufs = ubuf->ctx; |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 268 | struct vhost_virtqueue *vq = ubufs->vq; |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 269 | int cnt; |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 270 | |
Michael S. Tsirkin | b0c057c | 2014-02-13 11:45:11 +0200 | [diff] [blame] | 271 | rcu_read_lock_bh(); |
| 272 | |
Jason Wang | 19c73b3 | 2013-09-02 16:41:00 +0800 | [diff] [blame] | 273 | /* set len to mark this desc buffers done DMA */ |
| 274 | vq->heads[ubuf->desc].len = success ? |
| 275 | VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN; |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 276 | cnt = vhost_net_ubuf_put(ubufs); |
Jason Wang | 19c73b3 | 2013-09-02 16:41:00 +0800 | [diff] [blame] | 277 | |
Michael S. Tsirkin | 24eb21a | 2012-11-01 09:16:55 +0000 | [diff] [blame] | 278 | /* |
| 279 | * Trigger polling thread if guest stopped submitting new buffers: |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 280 | * in this case, the refcount after decrement will eventually reach 1. |
Michael S. Tsirkin | 24eb21a | 2012-11-01 09:16:55 +0000 | [diff] [blame] | 281 | * We also trigger polling periodically after each 16 packets |
| 282 | * (the value 16 here is more or less arbitrary, it's tuned to trigger |
| 283 | * less than 10% of times). |
| 284 | */ |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 285 | if (cnt <= 1 || !(cnt % 16)) |
Michael S. Tsirkin | 24eb21a | 2012-11-01 09:16:55 +0000 | [diff] [blame] | 286 | vhost_poll_queue(&vq->poll); |
Michael S. Tsirkin | b0c057c | 2014-02-13 11:45:11 +0200 | [diff] [blame] | 287 | |
| 288 | rcu_read_unlock_bh(); |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 289 | } |
| 290 | |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 291 | static inline unsigned long busy_clock(void) |
| 292 | { |
| 293 | return local_clock() >> 10; |
| 294 | } |
| 295 | |
| 296 | static bool vhost_can_busy_poll(struct vhost_dev *dev, |
| 297 | unsigned long endtime) |
| 298 | { |
| 299 | return likely(!need_resched()) && |
| 300 | likely(!time_after(busy_clock(), endtime)) && |
| 301 | likely(!signal_pending(current)) && |
| 302 | !vhost_has_work(dev); |
| 303 | } |
| 304 | |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 305 | static void vhost_net_disable_vq(struct vhost_net *n, |
| 306 | struct vhost_virtqueue *vq) |
| 307 | { |
| 308 | struct vhost_net_virtqueue *nvq = |
| 309 | container_of(vq, struct vhost_net_virtqueue, vq); |
| 310 | struct vhost_poll *poll = n->poll + (nvq - n->vqs); |
| 311 | if (!vq->private_data) |
| 312 | return; |
| 313 | vhost_poll_stop(poll); |
| 314 | } |
| 315 | |
| 316 | static int vhost_net_enable_vq(struct vhost_net *n, |
| 317 | struct vhost_virtqueue *vq) |
| 318 | { |
| 319 | struct vhost_net_virtqueue *nvq = |
| 320 | container_of(vq, struct vhost_net_virtqueue, vq); |
| 321 | struct vhost_poll *poll = n->poll + (nvq - n->vqs); |
| 322 | struct socket *sock; |
| 323 | |
| 324 | sock = vq->private_data; |
| 325 | if (!sock) |
| 326 | return 0; |
| 327 | |
| 328 | return vhost_poll_start(poll, sock->file); |
| 329 | } |
| 330 | |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 331 | static int vhost_net_tx_get_vq_desc(struct vhost_net *net, |
| 332 | struct vhost_virtqueue *vq, |
| 333 | struct iovec iov[], unsigned int iov_size, |
| 334 | unsigned int *out_num, unsigned int *in_num) |
| 335 | { |
| 336 | unsigned long uninitialized_var(endtime); |
| 337 | int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 338 | out_num, in_num, NULL, NULL); |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 339 | |
| 340 | if (r == vq->num && vq->busyloop_timeout) { |
| 341 | preempt_disable(); |
| 342 | endtime = busy_clock() + vq->busyloop_timeout; |
| 343 | while (vhost_can_busy_poll(vq->dev, endtime) && |
| 344 | vhost_vq_avail_empty(vq->dev, vq)) |
| 345 | cpu_relax_lowlatency(); |
| 346 | preempt_enable(); |
| 347 | r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 348 | out_num, in_num, NULL, NULL); |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 349 | } |
| 350 | |
| 351 | return r; |
| 352 | } |
| 353 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 354 | /* Expects to be always run from workqueue - which acts as |
| 355 | * read-size critical section for our kind of RCU. */ |
| 356 | static void handle_tx(struct vhost_net *net) |
| 357 | { |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 358 | struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 359 | struct vhost_virtqueue *vq = &nvq->vq; |
Al Viro | 98a527a | 2014-12-10 15:00:58 -0500 | [diff] [blame] | 360 | unsigned out, in; |
Michael S. Tsirkin | d5675bd | 2010-06-24 16:59:59 +0300 | [diff] [blame] | 361 | int head; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 362 | struct msghdr msg = { |
| 363 | .msg_name = NULL, |
| 364 | .msg_namelen = 0, |
| 365 | .msg_control = NULL, |
| 366 | .msg_controllen = 0, |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 367 | .msg_flags = MSG_DONTWAIT, |
| 368 | }; |
| 369 | size_t len, total_len = 0; |
Jason Wang | 70181d51 | 2013-04-10 20:50:48 +0000 | [diff] [blame] | 370 | int err; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 371 | size_t hdr_size; |
Arnd Bergmann | 28457ee | 2010-03-09 19:24:45 +0100 | [diff] [blame] | 372 | struct socket *sock; |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 373 | struct vhost_net_ubuf_ref *uninitialized_var(ubufs); |
Michael S. Tsirkin | cedb9bd | 2012-12-06 17:00:18 +0200 | [diff] [blame] | 374 | bool zcopy, zcopy_used; |
Arnd Bergmann | 28457ee | 2010-03-09 19:24:45 +0100 | [diff] [blame] | 375 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 376 | mutex_lock(&vq->mutex); |
Asias He | 2e26af7 | 2013-05-07 14:54:33 +0800 | [diff] [blame] | 377 | sock = vq->private_data; |
| 378 | if (!sock) |
| 379 | goto out; |
| 380 | |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 381 | if (!vq_iotlb_prefetch(vq)) |
| 382 | goto out; |
| 383 | |
Michael S. Tsirkin | 8ea8cf8 | 2011-05-20 02:10:54 +0300 | [diff] [blame] | 384 | vhost_disable_notify(&net->dev, vq); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 385 | |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 386 | hdr_size = nvq->vhost_hlen; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 387 | zcopy = nvq->ubufs; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 388 | |
| 389 | for (;;) { |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 390 | /* Release DMAs done buffers first */ |
| 391 | if (zcopy) |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 392 | vhost_zerocopy_signal_used(net, vq); |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 393 | |
Jason Wang | f7c6be4 | 2013-09-02 16:41:01 +0800 | [diff] [blame] | 394 | /* If more outstanding DMAs, queue the work. |
| 395 | * Handle upend_idx wrap around |
| 396 | */ |
| 397 | if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND) |
| 398 | % UIO_MAXIOV == nvq->done_idx)) |
| 399 | break; |
| 400 | |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 401 | head = vhost_net_tx_get_vq_desc(net, vq, vq->iov, |
| 402 | ARRAY_SIZE(vq->iov), |
| 403 | &out, &in); |
Michael S. Tsirkin | d5675bd | 2010-06-24 16:59:59 +0300 | [diff] [blame] | 404 | /* On error, stop handling until the next kick. */ |
Michael S. Tsirkin | 7b3384f | 2010-07-01 18:40:12 +0300 | [diff] [blame] | 405 | if (unlikely(head < 0)) |
Michael S. Tsirkin | d5675bd | 2010-06-24 16:59:59 +0300 | [diff] [blame] | 406 | break; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 407 | /* Nothing new? Wait for eventfd to tell us they refilled. */ |
| 408 | if (head == vq->num) { |
Michael S. Tsirkin | 8ea8cf8 | 2011-05-20 02:10:54 +0300 | [diff] [blame] | 409 | if (unlikely(vhost_enable_notify(&net->dev, vq))) { |
| 410 | vhost_disable_notify(&net->dev, vq); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 411 | continue; |
| 412 | } |
| 413 | break; |
| 414 | } |
| 415 | if (in) { |
| 416 | vq_err(vq, "Unexpected descriptor format for TX: " |
| 417 | "out %d, int %d\n", out, in); |
| 418 | break; |
| 419 | } |
| 420 | /* Skip header. TODO: support TSO. */ |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 421 | len = iov_length(vq->iov, out); |
Al Viro | c0371da | 2014-11-24 10:42:55 -0500 | [diff] [blame] | 422 | iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len); |
Al Viro | 98a527a | 2014-12-10 15:00:58 -0500 | [diff] [blame] | 423 | iov_iter_advance(&msg.msg_iter, hdr_size); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 424 | /* Sanity check */ |
Al Viro | 01e97e6 | 2014-12-15 21:39:31 -0500 | [diff] [blame] | 425 | if (!msg_data_left(&msg)) { |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 426 | vq_err(vq, "Unexpected header len for TX: " |
| 427 | "%zd expected %zd\n", |
Al Viro | 98a527a | 2014-12-10 15:00:58 -0500 | [diff] [blame] | 428 | len, hdr_size); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 429 | break; |
| 430 | } |
Al Viro | 01e97e6 | 2014-12-15 21:39:31 -0500 | [diff] [blame] | 431 | len = msg_data_left(&msg); |
Jason Wang | ce21a02 | 2013-09-02 16:40:59 +0800 | [diff] [blame] | 432 | |
| 433 | zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN |
| 434 | && (nvq->upend_idx + 1) % UIO_MAXIOV != |
| 435 | nvq->done_idx |
| 436 | && vhost_net_tx_select_zcopy(net); |
Michael S. Tsirkin | cedb9bd | 2012-12-06 17:00:18 +0200 | [diff] [blame] | 437 | |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 438 | /* use msg_control to pass vhost zerocopy ubuf info to skb */ |
Michael S. Tsirkin | cedb9bd | 2012-12-06 17:00:18 +0200 | [diff] [blame] | 439 | if (zcopy_used) { |
Jason Wang | ce21a02 | 2013-09-02 16:40:59 +0800 | [diff] [blame] | 440 | struct ubuf_info *ubuf; |
| 441 | ubuf = nvq->ubuf_info + nvq->upend_idx; |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 442 | |
Michael S. Tsirkin | 8b38694 | 2014-10-24 14:19:48 +0300 | [diff] [blame] | 443 | vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head); |
Jason Wang | ce21a02 | 2013-09-02 16:40:59 +0800 | [diff] [blame] | 444 | vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS; |
| 445 | ubuf->callback = vhost_zerocopy_callback; |
| 446 | ubuf->ctx = nvq->ubufs; |
| 447 | ubuf->desc = nvq->upend_idx; |
| 448 | msg.msg_control = ubuf; |
| 449 | msg.msg_controllen = sizeof(ubuf); |
| 450 | ubufs = nvq->ubufs; |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 451 | atomic_inc(&ubufs->refcount); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 452 | nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV; |
Jason Wang | ce21a02 | 2013-09-02 16:40:59 +0800 | [diff] [blame] | 453 | } else { |
Jason Wang | 4364d5f | 2013-06-05 15:40:46 +0800 | [diff] [blame] | 454 | msg.msg_control = NULL; |
Jason Wang | ce21a02 | 2013-09-02 16:40:59 +0800 | [diff] [blame] | 455 | ubufs = NULL; |
| 456 | } |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 457 | /* TODO: Check specific error and bomb out unless ENOBUFS? */ |
Ying Xue | 1b78414 | 2015-03-02 15:37:48 +0800 | [diff] [blame] | 458 | err = sock->ops->sendmsg(sock, &msg, len); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 459 | if (unlikely(err < 0)) { |
Michael S. Tsirkin | cedb9bd | 2012-12-06 17:00:18 +0200 | [diff] [blame] | 460 | if (zcopy_used) { |
Jason Wang | ce21a02 | 2013-09-02 16:40:59 +0800 | [diff] [blame] | 461 | vhost_net_ubuf_put(ubufs); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 462 | nvq->upend_idx = ((unsigned)nvq->upend_idx - 1) |
| 463 | % UIO_MAXIOV; |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 464 | } |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 465 | vhost_discard_vq_desc(vq, 1); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 466 | break; |
| 467 | } |
| 468 | if (err != len) |
Michael S. Tsirkin | 95c0ec6 | 2010-06-24 17:10:25 +0300 | [diff] [blame] | 469 | pr_debug("Truncated TX packet: " |
| 470 | " len %d != %zd\n", err, len); |
Michael S. Tsirkin | cedb9bd | 2012-12-06 17:00:18 +0200 | [diff] [blame] | 471 | if (!zcopy_used) |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 472 | vhost_add_used_and_signal(&net->dev, vq, head, 0); |
Jason Wang | c8fb217 | 2012-05-02 11:42:41 +0800 | [diff] [blame] | 473 | else |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 474 | vhost_zerocopy_signal_used(net, vq); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 475 | total_len += len; |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 476 | vhost_net_tx_packet(net); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 477 | if (unlikely(total_len >= VHOST_NET_WEIGHT)) { |
| 478 | vhost_poll_queue(&vq->poll); |
| 479 | break; |
| 480 | } |
| 481 | } |
Asias He | 2e26af7 | 2013-05-07 14:54:33 +0800 | [diff] [blame] | 482 | out: |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 483 | mutex_unlock(&vq->mutex); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 484 | } |
| 485 | |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 486 | static int peek_head_len(struct sock *sk) |
| 487 | { |
Jason Wang | 1576d98 | 2016-06-30 14:45:36 +0800 | [diff] [blame] | 488 | struct socket *sock = sk->sk_socket; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 489 | struct sk_buff *head; |
| 490 | int len = 0; |
Jason Wang | 783e398 | 2011-01-17 16:11:17 +0800 | [diff] [blame] | 491 | unsigned long flags; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 492 | |
Jason Wang | 1576d98 | 2016-06-30 14:45:36 +0800 | [diff] [blame] | 493 | if (sock->ops->peek_len) |
| 494 | return sock->ops->peek_len(sock); |
| 495 | |
Jason Wang | 783e398 | 2011-01-17 16:11:17 +0800 | [diff] [blame] | 496 | spin_lock_irqsave(&sk->sk_receive_queue.lock, flags); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 497 | head = skb_peek(&sk->sk_receive_queue); |
Basil Gor | c53cff5e | 2012-05-03 22:55:23 +0000 | [diff] [blame] | 498 | if (likely(head)) { |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 499 | len = head->len; |
Jiri Pirko | df8a39d | 2015-01-13 17:13:44 +0100 | [diff] [blame] | 500 | if (skb_vlan_tag_present(head)) |
Basil Gor | c53cff5e | 2012-05-03 22:55:23 +0000 | [diff] [blame] | 501 | len += VLAN_HLEN; |
| 502 | } |
| 503 | |
Jason Wang | 783e398 | 2011-01-17 16:11:17 +0800 | [diff] [blame] | 504 | spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 505 | return len; |
| 506 | } |
| 507 | |
Jason Wang | 1576d98 | 2016-06-30 14:45:36 +0800 | [diff] [blame] | 508 | static int sk_has_rx_data(struct sock *sk) |
| 509 | { |
| 510 | struct socket *sock = sk->sk_socket; |
| 511 | |
| 512 | if (sock->ops->peek_len) |
| 513 | return sock->ops->peek_len(sock); |
| 514 | |
| 515 | return skb_queue_empty(&sk->sk_receive_queue); |
| 516 | } |
| 517 | |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 518 | static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk) |
| 519 | { |
| 520 | struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; |
| 521 | struct vhost_virtqueue *vq = &nvq->vq; |
| 522 | unsigned long uninitialized_var(endtime); |
| 523 | int len = peek_head_len(sk); |
| 524 | |
| 525 | if (!len && vq->busyloop_timeout) { |
| 526 | /* Both tx vq and rx socket were polled here */ |
Jason Wang | 1cb8175 | 2018-03-26 16:10:23 +0800 | [diff] [blame] | 527 | mutex_lock_nested(&vq->mutex, 1); |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 528 | vhost_disable_notify(&net->dev, vq); |
| 529 | |
| 530 | preempt_disable(); |
| 531 | endtime = busy_clock() + vq->busyloop_timeout; |
| 532 | |
| 533 | while (vhost_can_busy_poll(&net->dev, endtime) && |
Jason Wang | 1576d98 | 2016-06-30 14:45:36 +0800 | [diff] [blame] | 534 | !sk_has_rx_data(sk) && |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 535 | vhost_vq_avail_empty(&net->dev, vq)) |
| 536 | cpu_relax_lowlatency(); |
| 537 | |
| 538 | preempt_enable(); |
| 539 | |
Jason Wang | f5755c0 | 2017-09-05 09:22:05 +0800 | [diff] [blame] | 540 | if (!vhost_vq_avail_empty(&net->dev, vq)) |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 541 | vhost_poll_queue(&vq->poll); |
Jason Wang | f5755c0 | 2017-09-05 09:22:05 +0800 | [diff] [blame] | 542 | else if (unlikely(vhost_enable_notify(&net->dev, vq))) { |
| 543 | vhost_disable_notify(&net->dev, vq); |
| 544 | vhost_poll_queue(&vq->poll); |
| 545 | } |
| 546 | |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 547 | mutex_unlock(&vq->mutex); |
| 548 | |
| 549 | len = peek_head_len(sk); |
| 550 | } |
| 551 | |
| 552 | return len; |
| 553 | } |
| 554 | |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 555 | /* This is a multi-buffer version of vhost_get_desc, that works if |
| 556 | * vq has read descriptors only. |
| 557 | * @vq - the relevant virtqueue |
| 558 | * @datalen - data length we'll be reading |
| 559 | * @iovcount - returned count of io vectors we fill |
| 560 | * @log - vhost log |
| 561 | * @log_num - log offset |
Jason Wang | 9424936 | 2011-01-17 16:11:08 +0800 | [diff] [blame] | 562 | * @quota - headcount quota, 1 for big buffer |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 563 | * returns number of buffer heads allocated, negative on error |
| 564 | */ |
| 565 | static int get_rx_bufs(struct vhost_virtqueue *vq, |
| 566 | struct vring_used_elem *heads, |
| 567 | int datalen, |
| 568 | unsigned *iovcount, |
| 569 | struct vhost_log *log, |
Jason Wang | 9424936 | 2011-01-17 16:11:08 +0800 | [diff] [blame] | 570 | unsigned *log_num, |
| 571 | unsigned int quota) |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 572 | { |
| 573 | unsigned int out, in; |
| 574 | int seg = 0; |
| 575 | int headcount = 0; |
| 576 | unsigned d; |
| 577 | int r, nlogs = 0; |
Michael S. Tsirkin | 8b38694 | 2014-10-24 14:19:48 +0300 | [diff] [blame] | 578 | /* len is always initialized before use since we are always called with |
| 579 | * datalen > 0. |
| 580 | */ |
| 581 | u32 uninitialized_var(len); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 582 | |
Jason Wang | 9424936 | 2011-01-17 16:11:08 +0800 | [diff] [blame] | 583 | while (datalen > 0 && headcount < quota) { |
Jason Wang | e0e9b40 | 2010-09-14 23:53:05 +0800 | [diff] [blame] | 584 | if (unlikely(seg >= UIO_MAXIOV)) { |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 585 | r = -ENOBUFS; |
| 586 | goto err; |
| 587 | } |
Michael S. Tsirkin | 47283be | 2014-06-05 15:20:27 +0300 | [diff] [blame] | 588 | r = vhost_get_vq_desc(vq, vq->iov + seg, |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 589 | ARRAY_SIZE(vq->iov) - seg, &out, |
| 590 | &in, log, log_num); |
Michael S. Tsirkin | a39ee44 | 2014-03-27 12:53:37 +0200 | [diff] [blame] | 591 | if (unlikely(r < 0)) |
| 592 | goto err; |
| 593 | |
| 594 | d = r; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 595 | if (d == vq->num) { |
| 596 | r = 0; |
| 597 | goto err; |
| 598 | } |
| 599 | if (unlikely(out || in <= 0)) { |
| 600 | vq_err(vq, "unexpected descriptor format for RX: " |
| 601 | "out %d, in %d\n", out, in); |
| 602 | r = -EINVAL; |
| 603 | goto err; |
| 604 | } |
| 605 | if (unlikely(log)) { |
| 606 | nlogs += *log_num; |
| 607 | log += *log_num; |
| 608 | } |
Michael S. Tsirkin | 8b38694 | 2014-10-24 14:19:48 +0300 | [diff] [blame] | 609 | heads[headcount].id = cpu_to_vhost32(vq, d); |
| 610 | len = iov_length(vq->iov + seg, in); |
| 611 | heads[headcount].len = cpu_to_vhost32(vq, len); |
| 612 | datalen -= len; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 613 | ++headcount; |
| 614 | seg += in; |
| 615 | } |
Michael S. Tsirkin | 99975cc | 2015-01-07 10:51:00 +0200 | [diff] [blame] | 616 | heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 617 | *iovcount = seg; |
| 618 | if (unlikely(log)) |
| 619 | *log_num = nlogs; |
Michael S. Tsirkin | d8316f3 | 2014-03-27 12:00:26 +0200 | [diff] [blame] | 620 | |
| 621 | /* Detect overrun */ |
| 622 | if (unlikely(datalen > 0)) { |
| 623 | r = UIO_MAXIOV + 1; |
| 624 | goto err; |
| 625 | } |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 626 | return headcount; |
| 627 | err: |
| 628 | vhost_discard_vq_desc(vq, headcount); |
| 629 | return r; |
| 630 | } |
| 631 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 632 | /* Expects to be always run from workqueue - which acts as |
| 633 | * read-size critical section for our kind of RCU. */ |
Jason Wang | 9424936 | 2011-01-17 16:11:08 +0800 | [diff] [blame] | 634 | static void handle_rx(struct vhost_net *net) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 635 | { |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 636 | struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX]; |
| 637 | struct vhost_virtqueue *vq = &nvq->vq; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 638 | unsigned uninitialized_var(in), log; |
| 639 | struct vhost_log *vq_log; |
| 640 | struct msghdr msg = { |
| 641 | .msg_name = NULL, |
| 642 | .msg_namelen = 0, |
| 643 | .msg_control = NULL, /* FIXME: get and handle RX aux data. */ |
| 644 | .msg_controllen = 0, |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 645 | .msg_flags = MSG_DONTWAIT, |
| 646 | }; |
Jason Wang | 0960b64 | 2015-02-15 16:35:17 +0800 | [diff] [blame] | 647 | struct virtio_net_hdr hdr = { |
| 648 | .flags = 0, |
| 649 | .gso_type = VIRTIO_NET_HDR_GSO_NONE |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 650 | }; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 651 | size_t total_len = 0; |
Michael S. Tsirkin | 910a578 | 2012-10-24 20:37:51 +0200 | [diff] [blame] | 652 | int err, mergeable; |
| 653 | s16 headcount; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 654 | size_t vhost_hlen, sock_hlen; |
| 655 | size_t vhost_len, sock_len; |
Asias He | 2e26af7 | 2013-05-07 14:54:33 +0800 | [diff] [blame] | 656 | struct socket *sock; |
Al Viro | ba7438a | 2014-12-10 15:51:28 -0500 | [diff] [blame] | 657 | struct iov_iter fixup; |
Jason Wang | 0960b64 | 2015-02-15 16:35:17 +0800 | [diff] [blame] | 658 | __virtio16 num_buffers; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 659 | |
Jason Wang | 1cb8175 | 2018-03-26 16:10:23 +0800 | [diff] [blame] | 660 | mutex_lock_nested(&vq->mutex, 0); |
Asias He | 2e26af7 | 2013-05-07 14:54:33 +0800 | [diff] [blame] | 661 | sock = vq->private_data; |
| 662 | if (!sock) |
| 663 | goto out; |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 664 | |
| 665 | if (!vq_iotlb_prefetch(vq)) |
| 666 | goto out; |
| 667 | |
Michael S. Tsirkin | 8ea8cf8 | 2011-05-20 02:10:54 +0300 | [diff] [blame] | 668 | vhost_disable_notify(&net->dev, vq); |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 669 | vhost_net_disable_vq(net, vq); |
Asias He | 2e26af7 | 2013-05-07 14:54:33 +0800 | [diff] [blame] | 670 | |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 671 | vhost_hlen = nvq->vhost_hlen; |
| 672 | sock_hlen = nvq->sock_hlen; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 673 | |
Michael S. Tsirkin | ea16c51 | 2014-06-05 15:20:23 +0300 | [diff] [blame] | 674 | vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ? |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 675 | vq->log : NULL; |
Michael S. Tsirkin | ea16c51 | 2014-06-05 15:20:23 +0300 | [diff] [blame] | 676 | mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 677 | |
Jason Wang | 0308813 | 2016-03-04 06:24:53 -0500 | [diff] [blame] | 678 | while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) { |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 679 | sock_len += sock_hlen; |
| 680 | vhost_len = sock_len + vhost_hlen; |
| 681 | headcount = get_rx_bufs(vq, vq->heads, vhost_len, |
Jason Wang | 9424936 | 2011-01-17 16:11:08 +0800 | [diff] [blame] | 682 | &in, vq_log, &log, |
| 683 | likely(mergeable) ? UIO_MAXIOV : 1); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 684 | /* On error, stop handling until the next kick. */ |
| 685 | if (unlikely(headcount < 0)) |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 686 | goto out; |
Michael S. Tsirkin | d8316f3 | 2014-03-27 12:00:26 +0200 | [diff] [blame] | 687 | /* On overrun, truncate and discard */ |
| 688 | if (unlikely(headcount > UIO_MAXIOV)) { |
Al Viro | c0371da | 2014-11-24 10:42:55 -0500 | [diff] [blame] | 689 | iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1); |
Ying Xue | 1b78414 | 2015-03-02 15:37:48 +0800 | [diff] [blame] | 690 | err = sock->ops->recvmsg(sock, &msg, |
Michael S. Tsirkin | d8316f3 | 2014-03-27 12:00:26 +0200 | [diff] [blame] | 691 | 1, MSG_DONTWAIT | MSG_TRUNC); |
| 692 | pr_debug("Discarded rx packet: len %zd\n", sock_len); |
| 693 | continue; |
| 694 | } |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 695 | /* OK, now we need to know about added descriptors. */ |
| 696 | if (!headcount) { |
Michael S. Tsirkin | 8ea8cf8 | 2011-05-20 02:10:54 +0300 | [diff] [blame] | 697 | if (unlikely(vhost_enable_notify(&net->dev, vq))) { |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 698 | /* They have slipped one in as we were |
| 699 | * doing that: check again. */ |
Michael S. Tsirkin | 8ea8cf8 | 2011-05-20 02:10:54 +0300 | [diff] [blame] | 700 | vhost_disable_notify(&net->dev, vq); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 701 | continue; |
| 702 | } |
| 703 | /* Nothing new? Wait for eventfd to tell us |
| 704 | * they refilled. */ |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 705 | goto out; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 706 | } |
| 707 | /* We don't need to be notified again. */ |
Al Viro | ba7438a | 2014-12-10 15:51:28 -0500 | [diff] [blame] | 708 | iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len); |
| 709 | fixup = msg.msg_iter; |
| 710 | if (unlikely((vhost_hlen))) { |
| 711 | /* We will supply the header ourselves |
| 712 | * TODO: support TSO. |
| 713 | */ |
| 714 | iov_iter_advance(&msg.msg_iter, vhost_hlen); |
Al Viro | ba7438a | 2014-12-10 15:51:28 -0500 | [diff] [blame] | 715 | } |
Ying Xue | 1b78414 | 2015-03-02 15:37:48 +0800 | [diff] [blame] | 716 | err = sock->ops->recvmsg(sock, &msg, |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 717 | sock_len, MSG_DONTWAIT | MSG_TRUNC); |
| 718 | /* Userspace might have consumed the packet meanwhile: |
| 719 | * it's not supposed to do this usually, but might be hard |
| 720 | * to prevent. Discard data we got (if any) and keep going. */ |
| 721 | if (unlikely(err != sock_len)) { |
| 722 | pr_debug("Discarded rx packet: " |
| 723 | " len %d, expected %zd\n", err, sock_len); |
| 724 | vhost_discard_vq_desc(vq, headcount); |
| 725 | continue; |
| 726 | } |
Al Viro | ba7438a | 2014-12-10 15:51:28 -0500 | [diff] [blame] | 727 | /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */ |
Michael S. Tsirkin | 4c5a844 | 2015-02-25 15:19:28 +0100 | [diff] [blame] | 728 | if (unlikely(vhost_hlen)) { |
| 729 | if (copy_to_iter(&hdr, sizeof(hdr), |
| 730 | &fixup) != sizeof(hdr)) { |
| 731 | vq_err(vq, "Unable to write vnet_hdr " |
| 732 | "at addr %p\n", vq->iov->iov_base); |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 733 | goto out; |
Michael S. Tsirkin | 4c5a844 | 2015-02-25 15:19:28 +0100 | [diff] [blame] | 734 | } |
| 735 | } else { |
| 736 | /* Header came from socket; we'll need to patch |
| 737 | * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF |
| 738 | */ |
| 739 | iov_iter_advance(&fixup, sizeof(hdr)); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 740 | } |
| 741 | /* TODO: Should check and handle checksum. */ |
Michael S. Tsirkin | 5201aa4 | 2015-02-03 11:07:06 +0200 | [diff] [blame] | 742 | |
Jason Wang | 0960b64 | 2015-02-15 16:35:17 +0800 | [diff] [blame] | 743 | num_buffers = cpu_to_vhost16(vq, headcount); |
Jason Wang | cfbdab9 | 2011-01-17 16:10:59 +0800 | [diff] [blame] | 744 | if (likely(mergeable) && |
Michael S. Tsirkin | 0d79a49 | 2015-02-25 15:20:01 +0100 | [diff] [blame] | 745 | copy_to_iter(&num_buffers, sizeof num_buffers, |
| 746 | &fixup) != sizeof num_buffers) { |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 747 | vq_err(vq, "Failed num_buffers write"); |
| 748 | vhost_discard_vq_desc(vq, headcount); |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 749 | goto out; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 750 | } |
| 751 | vhost_add_used_and_signal_n(&net->dev, vq, vq->heads, |
| 752 | headcount); |
| 753 | if (unlikely(vq_log)) |
Jason Wang | b35efeb | 2019-01-16 16:54:42 +0800 | [diff] [blame] | 754 | vhost_log_write(vq, vq_log, log, vhost_len, |
| 755 | vq->iov, in); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 756 | total_len += vhost_len; |
| 757 | if (unlikely(total_len >= VHOST_NET_WEIGHT)) { |
| 758 | vhost_poll_queue(&vq->poll); |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 759 | goto out; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 760 | } |
| 761 | } |
Jason Wang | 8241a1e | 2016-06-01 01:56:33 -0400 | [diff] [blame] | 762 | vhost_net_enable_vq(net, vq); |
Asias He | 2e26af7 | 2013-05-07 14:54:33 +0800 | [diff] [blame] | 763 | out: |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 764 | mutex_unlock(&vq->mutex); |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 765 | } |
| 766 | |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 767 | static void handle_tx_kick(struct vhost_work *work) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 768 | { |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 769 | struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, |
| 770 | poll.work); |
| 771 | struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); |
| 772 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 773 | handle_tx(net); |
| 774 | } |
| 775 | |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 776 | static void handle_rx_kick(struct vhost_work *work) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 777 | { |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 778 | struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, |
| 779 | poll.work); |
| 780 | struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); |
| 781 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 782 | handle_rx(net); |
| 783 | } |
| 784 | |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 785 | static void handle_tx_net(struct vhost_work *work) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 786 | { |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 787 | struct vhost_net *net = container_of(work, struct vhost_net, |
| 788 | poll[VHOST_NET_VQ_TX].work); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 789 | handle_tx(net); |
| 790 | } |
| 791 | |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 792 | static void handle_rx_net(struct vhost_work *work) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 793 | { |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 794 | struct vhost_net *net = container_of(work, struct vhost_net, |
| 795 | poll[VHOST_NET_VQ_RX].work); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 796 | handle_rx(net); |
| 797 | } |
| 798 | |
| 799 | static int vhost_net_open(struct inode *inode, struct file *f) |
| 800 | { |
Michael S. Tsirkin | 23cc5a9 | 2013-01-23 21:46:47 +0100 | [diff] [blame] | 801 | struct vhost_net *n; |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 802 | struct vhost_dev *dev; |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 803 | struct vhost_virtqueue **vqs; |
Zhi Yong Wu | 59566b6e | 2013-12-07 04:13:03 +0800 | [diff] [blame] | 804 | int i; |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 805 | |
Michael S. Tsirkin | 23cc5a9 | 2013-01-23 21:46:47 +0100 | [diff] [blame] | 806 | n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT); |
| 807 | if (!n) { |
| 808 | n = vmalloc(sizeof *n); |
| 809 | if (!n) |
| 810 | return -ENOMEM; |
| 811 | } |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 812 | vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL); |
| 813 | if (!vqs) { |
Romain Francoise | d04257b | 2014-06-12 10:42:34 +0200 | [diff] [blame] | 814 | kvfree(n); |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 815 | return -ENOMEM; |
| 816 | } |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 817 | |
| 818 | dev = &n->dev; |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 819 | vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq; |
| 820 | vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq; |
| 821 | n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick; |
| 822 | n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 823 | for (i = 0; i < VHOST_NET_VQ_MAX; i++) { |
| 824 | n->vqs[i].ubufs = NULL; |
| 825 | n->vqs[i].ubuf_info = NULL; |
| 826 | n->vqs[i].upend_idx = 0; |
| 827 | n->vqs[i].done_idx = 0; |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 828 | n->vqs[i].vhost_hlen = 0; |
| 829 | n->vqs[i].sock_hlen = 0; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 830 | } |
Zhi Yong Wu | 59566b6e | 2013-12-07 04:13:03 +0800 | [diff] [blame] | 831 | vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 832 | |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 833 | vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev); |
| 834 | vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 835 | |
| 836 | f->private_data = n; |
| 837 | |
| 838 | return 0; |
| 839 | } |
| 840 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 841 | static struct socket *vhost_net_stop_vq(struct vhost_net *n, |
| 842 | struct vhost_virtqueue *vq) |
| 843 | { |
| 844 | struct socket *sock; |
| 845 | |
| 846 | mutex_lock(&vq->mutex); |
Asias He | 22fa90c | 2013-05-07 14:54:36 +0800 | [diff] [blame] | 847 | sock = vq->private_data; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 848 | vhost_net_disable_vq(n, vq); |
Asias He | 22fa90c | 2013-05-07 14:54:36 +0800 | [diff] [blame] | 849 | vq->private_data = NULL; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 850 | mutex_unlock(&vq->mutex); |
| 851 | return sock; |
| 852 | } |
| 853 | |
| 854 | static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock, |
| 855 | struct socket **rx_sock) |
| 856 | { |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 857 | *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq); |
| 858 | *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 859 | } |
| 860 | |
| 861 | static void vhost_net_flush_vq(struct vhost_net *n, int index) |
| 862 | { |
| 863 | vhost_poll_flush(n->poll + index); |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 864 | vhost_poll_flush(&n->vqs[index].vq.poll); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 865 | } |
| 866 | |
| 867 | static void vhost_net_flush(struct vhost_net *n) |
| 868 | { |
| 869 | vhost_net_flush_vq(n, VHOST_NET_VQ_TX); |
| 870 | vhost_net_flush_vq(n, VHOST_NET_VQ_RX); |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 871 | if (n->vqs[VHOST_NET_VQ_TX].ubufs) { |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 872 | mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 873 | n->tx_flush = true; |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 874 | mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 875 | /* Wait for all lower device DMAs done. */ |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 876 | vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs); |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 877 | mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 878 | n->tx_flush = false; |
Michael S. Tsirkin | 0ad8b48 | 2014-02-13 11:42:05 +0200 | [diff] [blame] | 879 | atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1); |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 880 | mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 881 | } |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 882 | } |
| 883 | |
| 884 | static int vhost_net_release(struct inode *inode, struct file *f) |
| 885 | { |
| 886 | struct vhost_net *n = f->private_data; |
| 887 | struct socket *tx_sock; |
| 888 | struct socket *rx_sock; |
| 889 | |
| 890 | vhost_net_stop(n, &tx_sock, &rx_sock); |
| 891 | vhost_net_flush(n); |
Michael S. Tsirkin | b211616 | 2012-11-01 09:16:46 +0000 | [diff] [blame] | 892 | vhost_dev_stop(&n->dev); |
Michael S. Tsirkin | ea5d404 | 2011-11-27 19:05:58 +0200 | [diff] [blame] | 893 | vhost_dev_cleanup(&n->dev, false); |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 894 | vhost_net_vq_reset(n); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 895 | if (tx_sock) |
Al Viro | 09aaacf | 2014-03-05 20:39:00 -0500 | [diff] [blame] | 896 | sockfd_put(tx_sock); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 897 | if (rx_sock) |
Al Viro | 09aaacf | 2014-03-05 20:39:00 -0500 | [diff] [blame] | 898 | sockfd_put(rx_sock); |
Michael S. Tsirkin | b0c057c | 2014-02-13 11:45:11 +0200 | [diff] [blame] | 899 | /* Make sure no callbacks are outstanding */ |
| 900 | synchronize_rcu_bh(); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 901 | /* We do an extra flush before freeing memory, |
| 902 | * since jobs can re-queue themselves. */ |
| 903 | vhost_net_flush(n); |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 904 | kfree(n->dev.vqs); |
Romain Francoise | d04257b | 2014-06-12 10:42:34 +0200 | [diff] [blame] | 905 | kvfree(n); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 906 | return 0; |
| 907 | } |
| 908 | |
| 909 | static struct socket *get_raw_socket(int fd) |
| 910 | { |
| 911 | struct { |
| 912 | struct sockaddr_ll sa; |
| 913 | char buf[MAX_ADDR_LEN]; |
| 914 | } uaddr; |
| 915 | int uaddr_len = sizeof uaddr, r; |
| 916 | struct socket *sock = sockfd_lookup(fd, &r); |
Krishna Kumar | d47effe | 2011-03-01 17:06:37 +0530 | [diff] [blame] | 917 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 918 | if (!sock) |
| 919 | return ERR_PTR(-ENOTSOCK); |
| 920 | |
| 921 | /* Parameter checking */ |
| 922 | if (sock->sk->sk_type != SOCK_RAW) { |
| 923 | r = -ESOCKTNOSUPPORT; |
| 924 | goto err; |
| 925 | } |
| 926 | |
| 927 | r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, |
| 928 | &uaddr_len, 0); |
| 929 | if (r) |
| 930 | goto err; |
| 931 | |
| 932 | if (uaddr.sa.sll_family != AF_PACKET) { |
| 933 | r = -EPFNOSUPPORT; |
| 934 | goto err; |
| 935 | } |
| 936 | return sock; |
| 937 | err: |
Al Viro | 09aaacf | 2014-03-05 20:39:00 -0500 | [diff] [blame] | 938 | sockfd_put(sock); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 939 | return ERR_PTR(r); |
| 940 | } |
| 941 | |
Arnd Bergmann | 501c774 | 2010-02-18 05:46:50 +0000 | [diff] [blame] | 942 | static struct socket *get_tap_socket(int fd) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 943 | { |
| 944 | struct file *file = fget(fd); |
| 945 | struct socket *sock; |
Krishna Kumar | d47effe | 2011-03-01 17:06:37 +0530 | [diff] [blame] | 946 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 947 | if (!file) |
| 948 | return ERR_PTR(-EBADF); |
| 949 | sock = tun_get_socket(file); |
Arnd Bergmann | 501c774 | 2010-02-18 05:46:50 +0000 | [diff] [blame] | 950 | if (!IS_ERR(sock)) |
| 951 | return sock; |
| 952 | sock = macvtap_get_socket(file); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 953 | if (IS_ERR(sock)) |
| 954 | fput(file); |
| 955 | return sock; |
| 956 | } |
| 957 | |
| 958 | static struct socket *get_socket(int fd) |
| 959 | { |
| 960 | struct socket *sock; |
Krishna Kumar | d47effe | 2011-03-01 17:06:37 +0530 | [diff] [blame] | 961 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 962 | /* special case to disable backend */ |
| 963 | if (fd == -1) |
| 964 | return NULL; |
| 965 | sock = get_raw_socket(fd); |
| 966 | if (!IS_ERR(sock)) |
| 967 | return sock; |
Arnd Bergmann | 501c774 | 2010-02-18 05:46:50 +0000 | [diff] [blame] | 968 | sock = get_tap_socket(fd); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 969 | if (!IS_ERR(sock)) |
| 970 | return sock; |
| 971 | return ERR_PTR(-ENOTSOCK); |
| 972 | } |
| 973 | |
| 974 | static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd) |
| 975 | { |
| 976 | struct socket *sock, *oldsock; |
| 977 | struct vhost_virtqueue *vq; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 978 | struct vhost_net_virtqueue *nvq; |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 979 | struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 980 | int r; |
| 981 | |
| 982 | mutex_lock(&n->dev.mutex); |
| 983 | r = vhost_dev_check_owner(&n->dev); |
| 984 | if (r) |
| 985 | goto err; |
| 986 | |
| 987 | if (index >= VHOST_NET_VQ_MAX) { |
| 988 | r = -ENOBUFS; |
| 989 | goto err; |
| 990 | } |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 991 | vq = &n->vqs[index].vq; |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 992 | nvq = &n->vqs[index]; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 993 | mutex_lock(&vq->mutex); |
| 994 | |
| 995 | /* Verify that ring has been setup correctly. */ |
| 996 | if (!vhost_vq_access_ok(vq)) { |
| 997 | r = -EFAULT; |
Jeff Dike | 1dace8c | 2010-03-04 16:10:14 -0500 | [diff] [blame] | 998 | goto err_vq; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 999 | } |
| 1000 | sock = get_socket(fd); |
| 1001 | if (IS_ERR(sock)) { |
| 1002 | r = PTR_ERR(sock); |
Jeff Dike | 1dace8c | 2010-03-04 16:10:14 -0500 | [diff] [blame] | 1003 | goto err_vq; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1004 | } |
| 1005 | |
| 1006 | /* start polling new socket */ |
Asias He | 22fa90c | 2013-05-07 14:54:36 +0800 | [diff] [blame] | 1007 | oldsock = vq->private_data; |
David S. Miller | 11fe883 | 2010-07-20 18:25:24 -0700 | [diff] [blame] | 1008 | if (sock != oldsock) { |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 1009 | ubufs = vhost_net_ubuf_alloc(vq, |
| 1010 | sock && vhost_sock_zcopy(sock)); |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 1011 | if (IS_ERR(ubufs)) { |
| 1012 | r = PTR_ERR(ubufs); |
| 1013 | goto err_ubufs; |
| 1014 | } |
Jason Wang | 692a998 | 2013-01-28 01:05:17 +0000 | [diff] [blame] | 1015 | |
Krishna Kumar | d47effe | 2011-03-01 17:06:37 +0530 | [diff] [blame] | 1016 | vhost_net_disable_vq(n, vq); |
Asias He | 22fa90c | 2013-05-07 14:54:36 +0800 | [diff] [blame] | 1017 | vq->private_data = sock; |
Greg Kurz | 80f7d03 | 2016-02-16 15:59:44 +0100 | [diff] [blame] | 1018 | r = vhost_vq_init_access(vq); |
Jason Wang | f59281d | 2011-06-21 18:04:27 +0800 | [diff] [blame] | 1019 | if (r) |
Jason Wang | 692a998 | 2013-01-28 01:05:17 +0000 | [diff] [blame] | 1020 | goto err_used; |
Jason Wang | 2b8b328 | 2013-01-28 01:05:18 +0000 | [diff] [blame] | 1021 | r = vhost_net_enable_vq(n, vq); |
| 1022 | if (r) |
| 1023 | goto err_used; |
Jason Wang | 692a998 | 2013-01-28 01:05:17 +0000 | [diff] [blame] | 1024 | |
Asias He | 2839400 | 2013-04-27 15:07:46 +0800 | [diff] [blame] | 1025 | oldubufs = nvq->ubufs; |
| 1026 | nvq->ubufs = ubufs; |
Michael S. Tsirkin | 64e9a9b | 2012-12-03 07:31:51 +0000 | [diff] [blame] | 1027 | |
| 1028 | n->tx_packets = 0; |
| 1029 | n->tx_zcopy_err = 0; |
Michael S. Tsirkin | 1280c27 | 2012-12-04 00:17:14 +0200 | [diff] [blame] | 1030 | n->tx_flush = false; |
Jeff Dike | dd1f407 | 2010-03-04 16:10:14 -0500 | [diff] [blame] | 1031 | } |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1032 | |
Michael S. Tsirkin | 1680e90 | 2010-07-15 15:19:12 +0300 | [diff] [blame] | 1033 | mutex_unlock(&vq->mutex); |
| 1034 | |
Michael S. Tsirkin | c047e5f | 2011-07-20 13:41:31 +0300 | [diff] [blame] | 1035 | if (oldubufs) { |
Michael S. Tsirkin | c38e39c | 2013-06-25 17:29:46 +0300 | [diff] [blame] | 1036 | vhost_net_ubuf_put_wait_and_free(oldubufs); |
Michael S. Tsirkin | c047e5f | 2011-07-20 13:41:31 +0300 | [diff] [blame] | 1037 | mutex_lock(&vq->mutex); |
Michael S. Tsirkin | eaae813 | 2012-11-01 09:16:51 +0000 | [diff] [blame] | 1038 | vhost_zerocopy_signal_used(n, vq); |
Michael S. Tsirkin | c047e5f | 2011-07-20 13:41:31 +0300 | [diff] [blame] | 1039 | mutex_unlock(&vq->mutex); |
| 1040 | } |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 1041 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1042 | if (oldsock) { |
| 1043 | vhost_net_flush_vq(n, index); |
Al Viro | 09aaacf | 2014-03-05 20:39:00 -0500 | [diff] [blame] | 1044 | sockfd_put(oldsock); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1045 | } |
Jeff Dike | 1dace8c | 2010-03-04 16:10:14 -0500 | [diff] [blame] | 1046 | |
Michael S. Tsirkin | 1680e90 | 2010-07-15 15:19:12 +0300 | [diff] [blame] | 1047 | mutex_unlock(&n->dev.mutex); |
| 1048 | return 0; |
| 1049 | |
Jason Wang | 692a998 | 2013-01-28 01:05:17 +0000 | [diff] [blame] | 1050 | err_used: |
Asias He | 22fa90c | 2013-05-07 14:54:36 +0800 | [diff] [blame] | 1051 | vq->private_data = oldsock; |
Jason Wang | 692a998 | 2013-01-28 01:05:17 +0000 | [diff] [blame] | 1052 | vhost_net_enable_vq(n, vq); |
| 1053 | if (ubufs) |
Michael S. Tsirkin | c38e39c | 2013-06-25 17:29:46 +0300 | [diff] [blame] | 1054 | vhost_net_ubuf_put_wait_and_free(ubufs); |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 1055 | err_ubufs: |
Jason Wang | e11eb6a | 2018-06-21 13:11:31 +0800 | [diff] [blame] | 1056 | if (sock) |
| 1057 | sockfd_put(sock); |
Jeff Dike | 1dace8c | 2010-03-04 16:10:14 -0500 | [diff] [blame] | 1058 | err_vq: |
| 1059 | mutex_unlock(&vq->mutex); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1060 | err: |
| 1061 | mutex_unlock(&n->dev.mutex); |
| 1062 | return r; |
| 1063 | } |
| 1064 | |
| 1065 | static long vhost_net_reset_owner(struct vhost_net *n) |
| 1066 | { |
| 1067 | struct socket *tx_sock = NULL; |
| 1068 | struct socket *rx_sock = NULL; |
| 1069 | long err; |
Jason Wang | a9709d6 | 2016-06-23 02:04:31 -0400 | [diff] [blame] | 1070 | struct vhost_umem *umem; |
Krishna Kumar | d47effe | 2011-03-01 17:06:37 +0530 | [diff] [blame] | 1071 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1072 | mutex_lock(&n->dev.mutex); |
| 1073 | err = vhost_dev_check_owner(&n->dev); |
| 1074 | if (err) |
| 1075 | goto done; |
Jason Wang | a9709d6 | 2016-06-23 02:04:31 -0400 | [diff] [blame] | 1076 | umem = vhost_dev_reset_owner_prepare(); |
| 1077 | if (!umem) { |
Michael S. Tsirkin | 150b9e5 | 2013-04-28 17:12:08 +0300 | [diff] [blame] | 1078 | err = -ENOMEM; |
| 1079 | goto done; |
| 1080 | } |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1081 | vhost_net_stop(n, &tx_sock, &rx_sock); |
| 1082 | vhost_net_flush(n); |
Jason Wang | 73adb3b | 2018-01-25 22:03:52 +0800 | [diff] [blame] | 1083 | vhost_dev_stop(&n->dev); |
Jason Wang | a9709d6 | 2016-06-23 02:04:31 -0400 | [diff] [blame] | 1084 | vhost_dev_reset_owner(&n->dev, umem); |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 1085 | vhost_net_vq_reset(n); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1086 | done: |
| 1087 | mutex_unlock(&n->dev.mutex); |
| 1088 | if (tx_sock) |
Al Viro | 09aaacf | 2014-03-05 20:39:00 -0500 | [diff] [blame] | 1089 | sockfd_put(tx_sock); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1090 | if (rx_sock) |
Al Viro | 09aaacf | 2014-03-05 20:39:00 -0500 | [diff] [blame] | 1091 | sockfd_put(rx_sock); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1092 | return err; |
| 1093 | } |
| 1094 | |
| 1095 | static int vhost_net_set_features(struct vhost_net *n, u64 features) |
| 1096 | { |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 1097 | size_t vhost_hlen, sock_hlen, hdr_len; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1098 | int i; |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 1099 | |
Michael S. Tsirkin | e4fca7d | 2014-10-24 14:23:52 +0300 | [diff] [blame] | 1100 | hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | |
| 1101 | (1ULL << VIRTIO_F_VERSION_1))) ? |
David Stevens | 8dd014a | 2010-07-27 18:52:21 +0300 | [diff] [blame] | 1102 | sizeof(struct virtio_net_hdr_mrg_rxbuf) : |
| 1103 | sizeof(struct virtio_net_hdr); |
| 1104 | if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) { |
| 1105 | /* vhost provides vnet_hdr */ |
| 1106 | vhost_hlen = hdr_len; |
| 1107 | sock_hlen = 0; |
| 1108 | } else { |
| 1109 | /* socket provides vnet_hdr */ |
| 1110 | vhost_hlen = 0; |
| 1111 | sock_hlen = hdr_len; |
| 1112 | } |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1113 | mutex_lock(&n->dev.mutex); |
| 1114 | if ((features & (1 << VHOST_F_LOG_ALL)) && |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 1115 | !vhost_log_access_ok(&n->dev)) |
| 1116 | goto out_unlock; |
| 1117 | |
| 1118 | if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) { |
| 1119 | if (vhost_init_device_iotlb(&n->dev, true)) |
| 1120 | goto out_unlock; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1121 | } |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 1122 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1123 | for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 1124 | mutex_lock(&n->vqs[i].vq.mutex); |
Michael S. Tsirkin | ea16c51 | 2014-06-05 15:20:23 +0300 | [diff] [blame] | 1125 | n->vqs[i].vq.acked_features = features; |
Michael S. Tsirkin | 81f95a5 | 2013-04-28 15:51:40 +0300 | [diff] [blame] | 1126 | n->vqs[i].vhost_hlen = vhost_hlen; |
| 1127 | n->vqs[i].sock_hlen = sock_hlen; |
Asias He | 3ab2e42 | 2013-04-27 11:16:48 +0800 | [diff] [blame] | 1128 | mutex_unlock(&n->vqs[i].vq.mutex); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1129 | } |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1130 | mutex_unlock(&n->dev.mutex); |
| 1131 | return 0; |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 1132 | |
| 1133 | out_unlock: |
| 1134 | mutex_unlock(&n->dev.mutex); |
| 1135 | return -EFAULT; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1136 | } |
| 1137 | |
Asias He | b1ad849 | 2013-05-06 11:16:00 +0800 | [diff] [blame] | 1138 | static long vhost_net_set_owner(struct vhost_net *n) |
| 1139 | { |
| 1140 | int r; |
| 1141 | |
| 1142 | mutex_lock(&n->dev.mutex); |
Michael S. Tsirkin | 05c0535 | 2013-06-06 15:20:39 +0300 | [diff] [blame] | 1143 | if (vhost_dev_has_owner(&n->dev)) { |
| 1144 | r = -EBUSY; |
| 1145 | goto out; |
| 1146 | } |
Asias He | b1ad849 | 2013-05-06 11:16:00 +0800 | [diff] [blame] | 1147 | r = vhost_net_set_ubuf_info(n); |
| 1148 | if (r) |
| 1149 | goto out; |
| 1150 | r = vhost_dev_set_owner(&n->dev); |
| 1151 | if (r) |
| 1152 | vhost_net_clear_ubuf_info(n); |
| 1153 | vhost_net_flush(n); |
| 1154 | out: |
| 1155 | mutex_unlock(&n->dev.mutex); |
| 1156 | return r; |
| 1157 | } |
| 1158 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1159 | static long vhost_net_ioctl(struct file *f, unsigned int ioctl, |
| 1160 | unsigned long arg) |
| 1161 | { |
| 1162 | struct vhost_net *n = f->private_data; |
| 1163 | void __user *argp = (void __user *)arg; |
| 1164 | u64 __user *featurep = argp; |
| 1165 | struct vhost_vring_file backend; |
| 1166 | u64 features; |
| 1167 | int r; |
Krishna Kumar | d47effe | 2011-03-01 17:06:37 +0530 | [diff] [blame] | 1168 | |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1169 | switch (ioctl) { |
| 1170 | case VHOST_NET_SET_BACKEND: |
Takuya Yoshikawa | d3553a5 | 2010-05-27 19:01:58 +0900 | [diff] [blame] | 1171 | if (copy_from_user(&backend, argp, sizeof backend)) |
| 1172 | return -EFAULT; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1173 | return vhost_net_set_backend(n, backend.index, backend.fd); |
| 1174 | case VHOST_GET_FEATURES: |
Stefan Hajnoczi | 0dd05a3 | 2012-07-21 06:55:36 +0000 | [diff] [blame] | 1175 | features = VHOST_NET_FEATURES; |
Takuya Yoshikawa | d3553a5 | 2010-05-27 19:01:58 +0900 | [diff] [blame] | 1176 | if (copy_to_user(featurep, &features, sizeof features)) |
| 1177 | return -EFAULT; |
| 1178 | return 0; |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1179 | case VHOST_SET_FEATURES: |
Takuya Yoshikawa | d3553a5 | 2010-05-27 19:01:58 +0900 | [diff] [blame] | 1180 | if (copy_from_user(&features, featurep, sizeof features)) |
| 1181 | return -EFAULT; |
Stefan Hajnoczi | 0dd05a3 | 2012-07-21 06:55:36 +0000 | [diff] [blame] | 1182 | if (features & ~VHOST_NET_FEATURES) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1183 | return -EOPNOTSUPP; |
| 1184 | return vhost_net_set_features(n, features); |
| 1185 | case VHOST_RESET_OWNER: |
| 1186 | return vhost_net_reset_owner(n); |
Asias He | b1ad849 | 2013-05-06 11:16:00 +0800 | [diff] [blame] | 1187 | case VHOST_SET_OWNER: |
| 1188 | return vhost_net_set_owner(n); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1189 | default: |
| 1190 | mutex_lock(&n->dev.mutex); |
Michael S. Tsirkin | 935cdee | 2012-12-06 14:03:34 +0200 | [diff] [blame] | 1191 | r = vhost_dev_ioctl(&n->dev, ioctl, argp); |
| 1192 | if (r == -ENOIOCTLCMD) |
| 1193 | r = vhost_vring_ioctl(&n->dev, ioctl, argp); |
| 1194 | else |
| 1195 | vhost_net_flush(n); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1196 | mutex_unlock(&n->dev.mutex); |
| 1197 | return r; |
| 1198 | } |
| 1199 | } |
| 1200 | |
| 1201 | #ifdef CONFIG_COMPAT |
| 1202 | static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl, |
| 1203 | unsigned long arg) |
| 1204 | { |
| 1205 | return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg)); |
| 1206 | } |
| 1207 | #endif |
| 1208 | |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 1209 | static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to) |
| 1210 | { |
| 1211 | struct file *file = iocb->ki_filp; |
| 1212 | struct vhost_net *n = file->private_data; |
| 1213 | struct vhost_dev *dev = &n->dev; |
| 1214 | int noblock = file->f_flags & O_NONBLOCK; |
| 1215 | |
| 1216 | return vhost_chr_read_iter(dev, to, noblock); |
| 1217 | } |
| 1218 | |
| 1219 | static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb, |
| 1220 | struct iov_iter *from) |
| 1221 | { |
| 1222 | struct file *file = iocb->ki_filp; |
| 1223 | struct vhost_net *n = file->private_data; |
| 1224 | struct vhost_dev *dev = &n->dev; |
| 1225 | |
| 1226 | return vhost_chr_write_iter(dev, from); |
| 1227 | } |
| 1228 | |
| 1229 | static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait) |
| 1230 | { |
| 1231 | struct vhost_net *n = file->private_data; |
| 1232 | struct vhost_dev *dev = &n->dev; |
| 1233 | |
| 1234 | return vhost_chr_poll(file, dev, wait); |
| 1235 | } |
| 1236 | |
Tobias Klauser | 373a83a | 2010-05-17 15:12:49 +0200 | [diff] [blame] | 1237 | static const struct file_operations vhost_net_fops = { |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1238 | .owner = THIS_MODULE, |
| 1239 | .release = vhost_net_release, |
Jason Wang | 6b1e6cc | 2016-06-23 02:04:32 -0400 | [diff] [blame] | 1240 | .read_iter = vhost_net_chr_read_iter, |
| 1241 | .write_iter = vhost_net_chr_write_iter, |
| 1242 | .poll = vhost_net_chr_poll, |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1243 | .unlocked_ioctl = vhost_net_ioctl, |
| 1244 | #ifdef CONFIG_COMPAT |
| 1245 | .compat_ioctl = vhost_net_compat_ioctl, |
| 1246 | #endif |
| 1247 | .open = vhost_net_open, |
Arnd Bergmann | 6038f37 | 2010-08-15 18:52:59 +0200 | [diff] [blame] | 1248 | .llseek = noop_llseek, |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1249 | }; |
| 1250 | |
| 1251 | static struct miscdevice vhost_net_misc = { |
stephen hemminger | 7c7c7f0 | 2012-01-11 19:30:38 +0000 | [diff] [blame] | 1252 | .minor = VHOST_NET_MINOR, |
| 1253 | .name = "vhost-net", |
| 1254 | .fops = &vhost_net_fops, |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1255 | }; |
| 1256 | |
Christoph Hellwig | a8d3782 | 2010-04-13 14:11:25 -0400 | [diff] [blame] | 1257 | static int vhost_net_init(void) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1258 | { |
Michael S. Tsirkin | bab632d | 2011-07-18 03:48:46 +0000 | [diff] [blame] | 1259 | if (experimental_zcopytx) |
Asias He | fe729a5 | 2013-05-06 16:38:24 +0800 | [diff] [blame] | 1260 | vhost_net_enable_zcopy(VHOST_NET_VQ_TX); |
Tejun Heo | c23f3445 | 2010-06-02 20:40:00 +0200 | [diff] [blame] | 1261 | return misc_register(&vhost_net_misc); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1262 | } |
| 1263 | module_init(vhost_net_init); |
| 1264 | |
Christoph Hellwig | a8d3782 | 2010-04-13 14:11:25 -0400 | [diff] [blame] | 1265 | static void vhost_net_exit(void) |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1266 | { |
| 1267 | misc_deregister(&vhost_net_misc); |
Michael S. Tsirkin | 3a4d5c94e | 2010-01-14 06:17:27 +0000 | [diff] [blame] | 1268 | } |
| 1269 | module_exit(vhost_net_exit); |
| 1270 | |
| 1271 | MODULE_VERSION("0.0.1"); |
| 1272 | MODULE_LICENSE("GPL v2"); |
| 1273 | MODULE_AUTHOR("Michael S. Tsirkin"); |
| 1274 | MODULE_DESCRIPTION("Host kernel accelerator for virtio net"); |
stephen hemminger | 7c7c7f0 | 2012-01-11 19:30:38 +0000 | [diff] [blame] | 1275 | MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR); |
| 1276 | MODULE_ALIAS("devname:vhost-net"); |