1
The following changes since commit 43ab9a5376c95c61ae898a222c4d04bdf60e239b:
1
The following changes since commit 352998df1c53b366413690d95b35f76d0721ebed:
2
2
3
hw/i386/vmport: fix missing definitions with non-log trace backends (2017-12-21 22:52:28 +0000)
3
Merge tag 'i2c-20220314' of https://github.com/philmd/qemu into staging (2022-03-14 14:39:33 +0000)
4
4
5
are available in the git repository at:
5
are available in the git repository at:
6
6
7
https://github.com/jasowang/qemu.git tags/net-pull-request
7
https://github.com/jasowang/qemu.git tags/net-pull-request
8
8
9
for you to fetch changes up to 0065e915192cdf83c2700bb377e5323c2649476e:
9
for you to fetch changes up to 12a195fa343aae2ead1301ce04727bd0ae25eb15:
10
10
11
qemu-doc: Update the deprecation information of -tftp, -bootp, -redir and -smb (2017-12-22 10:06:05 +0800)
11
vdpa: Expose VHOST_F_LOG_ALL on SVQ (2022-03-15 13:57:44 +0800)
12
12
13
----------------------------------------------------------------
13
----------------------------------------------------------------
14
14
15
Changes since V2:
16
- fix 32bit build errros
17
15
----------------------------------------------------------------
18
----------------------------------------------------------------
16
Ed Swierk via Qemu-devel (2):
19
Eugenio Pérez (14):
17
e1000, e1000e: Move per-packet TX offload flags out of context state
20
vhost: Add VhostShadowVirtqueue
18
e1000: Separate TSO and non-TSO contexts, fixing UDP TX corruption
21
vhost: Add Shadow VirtQueue kick forwarding capabilities
22
vhost: Add Shadow VirtQueue call forwarding capabilities
23
vhost: Add vhost_svq_valid_features to shadow vq
24
virtio: Add vhost_svq_get_vring_addr
25
vdpa: adapt vhost_ops callbacks to svq
26
vhost: Shadow virtqueue buffers forwarding
27
util: Add iova_tree_alloc_map
28
util: add iova_tree_find_iova
29
vhost: Add VhostIOVATree
30
vdpa: Add custom IOTLB translations to SVQ
31
vdpa: Adapt vhost_vdpa_get_vring_base to SVQ
32
vdpa: Never set log_base addr if SVQ is enabled
33
vdpa: Expose VHOST_F_LOG_ALL on SVQ
19
34
20
Mark Cave-Ayland (13):
35
Jason Wang (1):
21
net: move CRC32 calculation from compute_mcast_idx() into its own net_crc32() function
36
virtio-net: fix map leaking on error during receive
22
net: introduce net_crc32_le() function
23
pcnet: switch pcnet over to use net_crc32_le()
24
eepro100: switch eepro100 e100_compute_mcast_idx() over to use net_crc32()
25
sunhme: switch sunhme over to use net_crc32_le()
26
sungem: fix multicast filter CRC calculation
27
eepro100: use inline net_crc32() and bitshift instead of compute_mcast_idx()
28
opencores_eth: use inline net_crc32() and bitshift instead of compute_mcast_idx()
29
lan9118: use inline net_crc32() and bitshift instead of compute_mcast_idx()
30
ftgmac100: use inline net_crc32() and bitshift instead of compute_mcast_idx()
31
ne2000: use inline net_crc32() and bitshift instead of compute_mcast_idx()
32
rtl8139: use inline net_crc32() and bitshift instead of compute_mcast_idx()
33
net: remove unused compute_mcast_idx() function
34
37
35
Thomas Huth (3):
38
hw/net/virtio-net.c | 1 +
36
net: Remove the legacy "-net channel" parameter
39
hw/virtio/meson.build | 2 +-
37
qemu-doc: The "-net nic" option can be used with "netdev=...", too
40
hw/virtio/vhost-iova-tree.c | 110 +++++++
38
qemu-doc: Update the deprecation information of -tftp, -bootp, -redir and -smb
41
hw/virtio/vhost-iova-tree.h | 27 ++
39
42
hw/virtio/vhost-shadow-virtqueue.c | 636 +++++++++++++++++++++++++++++++++++++
40
hw/net/e1000.c | 92 ++++++++++++++++++++++++++++----------------------
43
hw/virtio/vhost-shadow-virtqueue.h | 87 +++++
41
hw/net/e1000e.c | 4 +--
44
hw/virtio/vhost-vdpa.c | 522 +++++++++++++++++++++++++++++-
42
hw/net/e1000e_core.c | 16 ++++-----
45
include/hw/virtio/vhost-vdpa.h | 8 +
43
hw/net/e1000e_core.h | 2 ++
46
include/qemu/iova-tree.h | 38 ++-
44
hw/net/e1000x_common.h | 2 --
47
util/iova-tree.c | 170 ++++++++++
45
hw/net/eepro100.c | 32 +++---------------
48
10 files changed, 1584 insertions(+), 17 deletions(-)
46
hw/net/ftgmac100.c | 2 +-
49
create mode 100644 hw/virtio/vhost-iova-tree.c
47
hw/net/lan9118.c | 3 +-
50
create mode 100644 hw/virtio/vhost-iova-tree.h
48
hw/net/ne2000.c | 4 ++-
51
create mode 100644 hw/virtio/vhost-shadow-virtqueue.c
49
hw/net/opencores_eth.c | 3 +-
52
create mode 100644 hw/virtio/vhost-shadow-virtqueue.h
50
hw/net/pcnet.c | 22 ++----------
51
hw/net/rtl8139.c | 2 +-
52
hw/net/sungem.c | 5 ++-
53
hw/net/sunhme.c | 25 +-------------
54
include/net/net.h | 5 ++-
55
include/net/slirp.h | 2 --
56
net/net.c | 40 +++++++++++++++-------
57
net/slirp.c | 34 -------------------
58
qemu-doc.texi | 38 +++++++++++----------
59
qemu-options.hx | 14 ++++----
60
20 files changed, 144 insertions(+), 203 deletions(-)
61
53
62
54
55
diff view generated by jsdifflib
1
From: Thomas Huth <thuth@redhat.com>
1
Commit bedd7e93d0196 ("virtio-net: fix use after unmap/free for sg")
2
tries to fix the use after free of the sg by caching the virtqueue
3
elements in an array and unmap them at once after receiving the
4
packets, But it forgot to unmap the cached elements on error which
5
will lead to leaking of mapping and other unexpected results.
2
6
3
Looks like we missed to document that it is also possible to specify
7
Fixing this by detaching the cached elements on error. This addresses
4
a netdev with "-net nic" - which is very useful if you want to
8
CVE-2022-26353.
5
configure your on-board NIC to use a backend that has been specified
6
with "-netdev".
7
9
8
Signed-off-by: Thomas Huth <thuth@redhat.com>
10
Reported-by: Victor Tom <vv474172261@gmail.com>
11
Cc: qemu-stable@nongnu.org
12
Fixes: CVE-2022-26353
13
Fixes: bedd7e93d0196 ("virtio-net: fix use after unmap/free for sg")
14
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
9
Signed-off-by: Jason Wang <jasowang@redhat.com>
15
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
---
16
---
11
qemu-options.hx | 14 ++++++++------
17
hw/net/virtio-net.c | 1 +
12
1 file changed, 8 insertions(+), 6 deletions(-)
18
1 file changed, 1 insertion(+)
13
19
14
diff --git a/qemu-options.hx b/qemu-options.hx
20
diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
15
index XXXXXXX..XXXXXXX 100644
21
index XXXXXXX..XXXXXXX 100644
16
--- a/qemu-options.hx
22
--- a/hw/net/virtio-net.c
17
+++ b/qemu-options.hx
23
+++ b/hw/net/virtio-net.c
18
@@ -XXX,XX +XXX,XX @@ DEF("netdev", HAS_ARG, QEMU_OPTION_netdev,
24
@@ -XXX,XX +XXX,XX @@ static ssize_t virtio_net_receive_rcu(NetClientState *nc, const uint8_t *buf,
19
"-netdev hubport,id=str,hubid=n\n"
25
20
" configure a hub port on QEMU VLAN 'n'\n", QEMU_ARCH_ALL)
26
err:
21
DEF("net", HAS_ARG, QEMU_OPTION_net,
27
for (j = 0; j < i; j++) {
22
- "-net nic[,vlan=n][,macaddr=mac][,model=type][,name=str][,addr=str][,vectors=v]\n"
28
+ virtqueue_detach_element(q->rx_vq, elems[j], lens[j]);
23
- " old way to create a new NIC and connect it to VLAN 'n'\n"
29
g_free(elems[j]);
24
- " (use the '-device devtype,netdev=str' option if possible instead)\n"
30
}
25
+ "-net nic[,vlan=n][,netdev=nd][,macaddr=mac][,model=type][,name=str][,addr=str][,vectors=v]\n"
31
26
+ " configure or create an on-board (or machine default) NIC and\n"
27
+ " connect it either to VLAN 'n' or the netdev 'nd' (for pluggable\n"
28
+ " NICs please use '-device devtype,netdev=nd' instead)\n"
29
"-net dump[,vlan=n][,file=f][,len=n]\n"
30
" dump traffic on vlan 'n' to file 'f' (max n bytes per packet)\n"
31
"-net none use it alone to have zero network devices. If no -net option\n"
32
@@ -XXX,XX +XXX,XX @@ DEF("net", HAS_ARG, QEMU_OPTION_net,
33
" old way to initialize a host network interface\n"
34
" (use the -netdev option if possible instead)\n", QEMU_ARCH_ALL)
35
STEXI
36
-@item -net nic[,vlan=@var{n}][,macaddr=@var{mac}][,model=@var{type}] [,name=@var{name}][,addr=@var{addr}][,vectors=@var{v}]
37
+@item -net nic[,vlan=@var{n}][,netdev=@var{nd}][,macaddr=@var{mac}][,model=@var{type}] [,name=@var{name}][,addr=@var{addr}][,vectors=@var{v}]
38
@findex -net
39
-Create a new Network Interface Card and connect it to VLAN @var{n} (@var{n}
40
-= 0 is the default). The NIC is an e1000 by default on the PC
41
+Configure or create an on-board (or machine default) Network Interface Card
42
+(NIC) and connect it either to VLAN @var{n} (@var{n} = 0 is the default), or
43
+to the netdev @var{nd}. The NIC is an e1000 by default on the PC
44
target. Optionally, the MAC address can be changed to @var{mac}, the
45
device address set to @var{addr} (PCI cards only),
46
and a @var{name} can be assigned for use in monitor commands.
47
--
32
--
48
2.7.4
33
2.7.4
49
50
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
This makes it much easier to compare the multicast CRC calculation endian and
3
Vhost shadow virtqueue (SVQ) is an intermediate jump for virtqueue
4
bitshift against the Linux driver implementation.
4
notifications and buffers, allowing qemu to track them. While qemu is
5
forwarding the buffers and virtqueue changes, it is able to commit the
6
memory it's being dirtied, the same way regular qemu's VirtIO devices
7
do.
5
8
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
9
This commit only exposes basic SVQ allocation and free. Next patches of
10
the series add functionality like notifications and buffers forwarding.
11
12
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
13
Acked-by: Michael S. Tsirkin <mst@redhat.com>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
14
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
15
---
9
hw/net/rtl8139.c | 2 +-
16
hw/virtio/meson.build | 2 +-
10
1 file changed, 1 insertion(+), 1 deletion(-)
17
hw/virtio/vhost-shadow-virtqueue.c | 62 ++++++++++++++++++++++++++++++++++++++
18
hw/virtio/vhost-shadow-virtqueue.h | 28 +++++++++++++++++
19
3 files changed, 91 insertions(+), 1 deletion(-)
20
create mode 100644 hw/virtio/vhost-shadow-virtqueue.c
21
create mode 100644 hw/virtio/vhost-shadow-virtqueue.h
11
22
12
diff --git a/hw/net/rtl8139.c b/hw/net/rtl8139.c
23
diff --git a/hw/virtio/meson.build b/hw/virtio/meson.build
13
index XXXXXXX..XXXXXXX 100644
24
index XXXXXXX..XXXXXXX 100644
14
--- a/hw/net/rtl8139.c
25
--- a/hw/virtio/meson.build
15
+++ b/hw/net/rtl8139.c
26
+++ b/hw/virtio/meson.build
16
@@ -XXX,XX +XXX,XX @@ static ssize_t rtl8139_do_receive(NetClientState *nc, const uint8_t *buf, size_t
27
@@ -XXX,XX +XXX,XX @@ softmmu_ss.add(when: 'CONFIG_ALL', if_true: files('vhost-stub.c'))
17
return size;
28
18
}
29
virtio_ss = ss.source_set()
19
30
virtio_ss.add(files('virtio.c'))
20
- int mcast_idx = compute_mcast_idx(buf);
31
-virtio_ss.add(when: 'CONFIG_VHOST', if_true: files('vhost.c', 'vhost-backend.c'))
21
+ int mcast_idx = net_crc32(buf, ETH_ALEN) >> 26;
32
+virtio_ss.add(when: 'CONFIG_VHOST', if_true: files('vhost.c', 'vhost-backend.c', 'vhost-shadow-virtqueue.c'))
22
33
virtio_ss.add(when: 'CONFIG_VHOST_USER', if_true: files('vhost-user.c'))
23
if (!(s->mult[mcast_idx >> 3] & (1 << (mcast_idx & 7))))
34
virtio_ss.add(when: 'CONFIG_VHOST_VDPA', if_true: files('vhost-vdpa.c'))
24
{
35
virtio_ss.add(when: 'CONFIG_VIRTIO_BALLOON', if_true: files('virtio-balloon.c'))
36
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
37
new file mode 100644
38
index XXXXXXX..XXXXXXX
39
--- /dev/null
40
+++ b/hw/virtio/vhost-shadow-virtqueue.c
41
@@ -XXX,XX +XXX,XX @@
42
+/*
43
+ * vhost shadow virtqueue
44
+ *
45
+ * SPDX-FileCopyrightText: Red Hat, Inc. 2021
46
+ * SPDX-FileContributor: Author: Eugenio Pérez <eperezma@redhat.com>
47
+ *
48
+ * SPDX-License-Identifier: GPL-2.0-or-later
49
+ */
50
+
51
+#include "qemu/osdep.h"
52
+#include "hw/virtio/vhost-shadow-virtqueue.h"
53
+
54
+#include "qemu/error-report.h"
55
+
56
+/**
57
+ * Creates vhost shadow virtqueue, and instructs the vhost device to use the
58
+ * shadow methods and file descriptors.
59
+ *
60
+ * Returns the new virtqueue or NULL.
61
+ *
62
+ * In case of error, reason is reported through error_report.
63
+ */
64
+VhostShadowVirtqueue *vhost_svq_new(void)
65
+{
66
+ g_autofree VhostShadowVirtqueue *svq = g_new0(VhostShadowVirtqueue, 1);
67
+ int r;
68
+
69
+ r = event_notifier_init(&svq->hdev_kick, 0);
70
+ if (r != 0) {
71
+ error_report("Couldn't create kick event notifier: %s (%d)",
72
+ g_strerror(errno), errno);
73
+ goto err_init_hdev_kick;
74
+ }
75
+
76
+ r = event_notifier_init(&svq->hdev_call, 0);
77
+ if (r != 0) {
78
+ error_report("Couldn't create call event notifier: %s (%d)",
79
+ g_strerror(errno), errno);
80
+ goto err_init_hdev_call;
81
+ }
82
+
83
+ return g_steal_pointer(&svq);
84
+
85
+err_init_hdev_call:
86
+ event_notifier_cleanup(&svq->hdev_kick);
87
+
88
+err_init_hdev_kick:
89
+ return NULL;
90
+}
91
+
92
+/**
93
+ * Free the resources of the shadow virtqueue.
94
+ *
95
+ * @pvq: gpointer to SVQ so it can be used by autofree functions.
96
+ */
97
+void vhost_svq_free(gpointer pvq)
98
+{
99
+ VhostShadowVirtqueue *vq = pvq;
100
+ event_notifier_cleanup(&vq->hdev_kick);
101
+ event_notifier_cleanup(&vq->hdev_call);
102
+ g_free(vq);
103
+}
104
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
105
new file mode 100644
106
index XXXXXXX..XXXXXXX
107
--- /dev/null
108
+++ b/hw/virtio/vhost-shadow-virtqueue.h
109
@@ -XXX,XX +XXX,XX @@
110
+/*
111
+ * vhost shadow virtqueue
112
+ *
113
+ * SPDX-FileCopyrightText: Red Hat, Inc. 2021
114
+ * SPDX-FileContributor: Author: Eugenio Pérez <eperezma@redhat.com>
115
+ *
116
+ * SPDX-License-Identifier: GPL-2.0-or-later
117
+ */
118
+
119
+#ifndef VHOST_SHADOW_VIRTQUEUE_H
120
+#define VHOST_SHADOW_VIRTQUEUE_H
121
+
122
+#include "qemu/event_notifier.h"
123
+
124
+/* Shadow virtqueue to relay notifications */
125
+typedef struct VhostShadowVirtqueue {
126
+ /* Shadow kick notifier, sent to vhost */
127
+ EventNotifier hdev_kick;
128
+ /* Shadow call notifier, sent to vhost */
129
+ EventNotifier hdev_call;
130
+} VhostShadowVirtqueue;
131
+
132
+VhostShadowVirtqueue *vhost_svq_new(void);
133
+
134
+void vhost_svq_free(gpointer vq);
135
+G_DEFINE_AUTOPTR_CLEANUP_FUNC(VhostShadowVirtqueue, vhost_svq_free);
136
+
137
+#endif
25
--
138
--
26
2.7.4
139
2.7.4
27
140
28
141
diff view generated by jsdifflib
1
From: Thomas Huth <thuth@redhat.com>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
It has never been documented, so hardly anybody knows about this
3
At this mode no buffer forwarding will be performed in SVQ mode: Qemu
4
parameter, and it is marked as deprecated since QEMU v2.6.
4
will just forward the guest's kicks to the device.
5
Time to let it go now.
5
6
6
Host memory notifiers regions are left out for simplicity, and they will
7
Reviewed-by: Samuel Thibault <samuel.thibault@ens-lyon.org>
7
not be addressed in this series.
8
Signed-off-by: Thomas Huth <thuth@redhat.com>
8
9
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
10
Acked-by: Michael S. Tsirkin <mst@redhat.com>
9
Signed-off-by: Jason Wang <jasowang@redhat.com>
11
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
---
12
---
11
include/net/slirp.h | 2 --
13
hw/virtio/vhost-shadow-virtqueue.c | 55 ++++++++++++++
12
net/net.c | 7 -------
14
hw/virtio/vhost-shadow-virtqueue.h | 14 ++++
13
net/slirp.c | 34 ----------------------------------
15
hw/virtio/vhost-vdpa.c | 144 ++++++++++++++++++++++++++++++++++++-
14
qemu-doc.texi | 5 -----
16
include/hw/virtio/vhost-vdpa.h | 4 ++
15
4 files changed, 48 deletions(-)
17
4 files changed, 215 insertions(+), 2 deletions(-)
16
18
17
diff --git a/include/net/slirp.h b/include/net/slirp.h
19
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
18
index XXXXXXX..XXXXXXX 100644
20
index XXXXXXX..XXXXXXX 100644
19
--- a/include/net/slirp.h
21
--- a/hw/virtio/vhost-shadow-virtqueue.c
20
+++ b/include/net/slirp.h
22
+++ b/hw/virtio/vhost-shadow-virtqueue.c
21
@@ -XXX,XX +XXX,XX @@ void hmp_hostfwd_remove(Monitor *mon, const QDict *qdict);
23
@@ -XXX,XX +XXX,XX @@
22
24
#include "hw/virtio/vhost-shadow-virtqueue.h"
23
int net_slirp_redir(const char *redir_str);
25
24
26
#include "qemu/error-report.h"
25
-int net_slirp_parse_legacy(QemuOptsList *opts_list, const char *optarg, int *ret);
27
+#include "qemu/main-loop.h"
26
-
28
+#include "linux-headers/linux/vhost.h"
27
int net_slirp_smb(const char *exported_dir);
29
+
28
30
+/**
29
void hmp_info_usernet(Monitor *mon, const QDict *qdict);
31
+ * Forward guest notifications.
30
diff --git a/net/net.c b/net/net.c
32
+ *
33
+ * @n: guest kick event notifier, the one that guest set to notify svq.
34
+ */
35
+static void vhost_handle_guest_kick(EventNotifier *n)
36
+{
37
+ VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue, svq_kick);
38
+ event_notifier_test_and_clear(n);
39
+ event_notifier_set(&svq->hdev_kick);
40
+}
41
+
42
+/**
43
+ * Set a new file descriptor for the guest to kick the SVQ and notify for avail
44
+ *
45
+ * @svq: The svq
46
+ * @svq_kick_fd: The svq kick fd
47
+ *
48
+ * Note that the SVQ will never close the old file descriptor.
49
+ */
50
+void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd)
51
+{
52
+ EventNotifier *svq_kick = &svq->svq_kick;
53
+ bool poll_stop = VHOST_FILE_UNBIND != event_notifier_get_fd(svq_kick);
54
+ bool poll_start = svq_kick_fd != VHOST_FILE_UNBIND;
55
+
56
+ if (poll_stop) {
57
+ event_notifier_set_handler(svq_kick, NULL);
58
+ }
59
+
60
+ /*
61
+ * event_notifier_set_handler already checks for guest's notifications if
62
+ * they arrive at the new file descriptor in the switch, so there is no
63
+ * need to explicitly check for them.
64
+ */
65
+ if (poll_start) {
66
+ event_notifier_init_fd(svq_kick, svq_kick_fd);
67
+ event_notifier_set(svq_kick);
68
+ event_notifier_set_handler(svq_kick, vhost_handle_guest_kick);
69
+ }
70
+}
71
+
72
+/**
73
+ * Stop the shadow virtqueue operation.
74
+ * @svq: Shadow Virtqueue
75
+ */
76
+void vhost_svq_stop(VhostShadowVirtqueue *svq)
77
+{
78
+ event_notifier_set_handler(&svq->svq_kick, NULL);
79
+}
80
81
/**
82
* Creates vhost shadow virtqueue, and instructs the vhost device to use the
83
@@ -XXX,XX +XXX,XX @@ VhostShadowVirtqueue *vhost_svq_new(void)
84
goto err_init_hdev_call;
85
}
86
87
+ event_notifier_init_fd(&svq->svq_kick, VHOST_FILE_UNBIND);
88
return g_steal_pointer(&svq);
89
90
err_init_hdev_call:
91
@@ -XXX,XX +XXX,XX @@ err_init_hdev_kick:
92
void vhost_svq_free(gpointer pvq)
93
{
94
VhostShadowVirtqueue *vq = pvq;
95
+ vhost_svq_stop(vq);
96
event_notifier_cleanup(&vq->hdev_kick);
97
event_notifier_cleanup(&vq->hdev_call);
98
g_free(vq);
99
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
31
index XXXXXXX..XXXXXXX 100644
100
index XXXXXXX..XXXXXXX 100644
32
--- a/net/net.c
101
--- a/hw/virtio/vhost-shadow-virtqueue.h
33
+++ b/net/net.c
102
+++ b/hw/virtio/vhost-shadow-virtqueue.h
34
@@ -XXX,XX +XXX,XX @@ int net_init_clients(void)
103
@@ -XXX,XX +XXX,XX @@ typedef struct VhostShadowVirtqueue {
35
104
EventNotifier hdev_kick;
36
int net_client_parse(QemuOptsList *opts_list, const char *optarg)
105
/* Shadow call notifier, sent to vhost */
37
{
106
EventNotifier hdev_call;
38
-#if defined(CONFIG_SLIRP)
107
+
39
- int ret;
108
+ /*
40
- if (net_slirp_parse_legacy(opts_list, optarg, &ret)) {
109
+ * Borrowed virtqueue's guest to host notifier. To borrow it in this event
41
- return ret;
110
+ * notifier allows to recover the VhostShadowVirtqueue from the event loop
42
- }
111
+ * easily. If we use the VirtQueue's one, we don't have an easy way to
43
-#endif
112
+ * retrieve VhostShadowVirtqueue.
44
-
113
+ *
45
if (!qemu_opts_parse_noisily(opts_list, optarg, true)) {
114
+ * So shadow virtqueue must not clean it, or we would lose VirtQueue one.
46
return -1;
115
+ */
47
}
116
+ EventNotifier svq_kick;
48
diff --git a/net/slirp.c b/net/slirp.c
117
} VhostShadowVirtqueue;
118
119
+void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd);
120
+
121
+void vhost_svq_stop(VhostShadowVirtqueue *svq);
122
+
123
VhostShadowVirtqueue *vhost_svq_new(void);
124
125
void vhost_svq_free(gpointer vq);
126
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
49
index XXXXXXX..XXXXXXX 100644
127
index XXXXXXX..XXXXXXX 100644
50
--- a/net/slirp.c
128
--- a/hw/virtio/vhost-vdpa.c
51
+++ b/net/slirp.c
129
+++ b/hw/virtio/vhost-vdpa.c
52
@@ -XXX,XX +XXX,XX @@ int net_init_slirp(const Netdev *netdev, const char *name,
130
@@ -XXX,XX +XXX,XX @@
53
131
#include "hw/virtio/vhost.h"
132
#include "hw/virtio/vhost-backend.h"
133
#include "hw/virtio/virtio-net.h"
134
+#include "hw/virtio/vhost-shadow-virtqueue.h"
135
#include "hw/virtio/vhost-vdpa.h"
136
#include "exec/address-spaces.h"
137
#include "qemu/main-loop.h"
138
#include "cpu.h"
139
#include "trace.h"
140
#include "qemu-common.h"
141
+#include "qapi/error.h"
142
143
/*
144
* Return one past the end of the end of section. Be careful with uint64_t
145
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_one_time_request(struct vhost_dev *dev)
146
return v->index != 0;
147
}
148
149
+static int vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v,
150
+ Error **errp)
151
+{
152
+ g_autoptr(GPtrArray) shadow_vqs = NULL;
153
+
154
+ if (!v->shadow_vqs_enabled) {
155
+ return 0;
156
+ }
157
+
158
+ shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free);
159
+ for (unsigned n = 0; n < hdev->nvqs; ++n) {
160
+ g_autoptr(VhostShadowVirtqueue) svq = vhost_svq_new();
161
+
162
+ if (unlikely(!svq)) {
163
+ error_setg(errp, "Cannot create svq %u", n);
164
+ return -1;
165
+ }
166
+ g_ptr_array_add(shadow_vqs, g_steal_pointer(&svq));
167
+ }
168
+
169
+ v->shadow_vqs = g_steal_pointer(&shadow_vqs);
170
+ return 0;
171
+}
172
+
173
static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp)
174
{
175
struct vhost_vdpa *v;
176
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp)
177
dev->opaque = opaque ;
178
v->listener = vhost_vdpa_memory_listener;
179
v->msg_type = VHOST_IOTLB_MSG_V2;
180
+ ret = vhost_vdpa_init_svq(dev, v, errp);
181
+ if (ret) {
182
+ goto err;
183
+ }
184
185
vhost_vdpa_get_iova_range(v);
186
187
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp)
188
VIRTIO_CONFIG_S_DRIVER);
189
190
return 0;
191
+
192
+err:
193
+ ram_block_discard_disable(false);
194
+ return ret;
195
}
196
197
static void vhost_vdpa_host_notifier_uninit(struct vhost_dev *dev,
198
@@ -XXX,XX +XXX,XX @@ static void vhost_vdpa_host_notifiers_uninit(struct vhost_dev *dev, int n)
199
200
static void vhost_vdpa_host_notifiers_init(struct vhost_dev *dev)
201
{
202
+ struct vhost_vdpa *v = dev->opaque;
203
int i;
204
205
+ if (v->shadow_vqs_enabled) {
206
+ /* FIXME SVQ is not compatible with host notifiers mr */
207
+ return;
208
+ }
209
+
210
for (i = dev->vq_index; i < dev->vq_index + dev->nvqs; i++) {
211
if (vhost_vdpa_host_notifier_init(dev, i)) {
212
goto err;
213
@@ -XXX,XX +XXX,XX @@ err:
214
return;
215
}
216
217
+static void vhost_vdpa_svq_cleanup(struct vhost_dev *dev)
218
+{
219
+ struct vhost_vdpa *v = dev->opaque;
220
+ size_t idx;
221
+
222
+ if (!v->shadow_vqs) {
223
+ return;
224
+ }
225
+
226
+ for (idx = 0; idx < v->shadow_vqs->len; ++idx) {
227
+ vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, idx));
228
+ }
229
+ g_ptr_array_free(v->shadow_vqs, true);
230
+}
231
+
232
static int vhost_vdpa_cleanup(struct vhost_dev *dev)
233
{
234
struct vhost_vdpa *v;
235
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_cleanup(struct vhost_dev *dev)
236
trace_vhost_vdpa_cleanup(dev, v);
237
vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
238
memory_listener_unregister(&v->listener);
239
+ vhost_vdpa_svq_cleanup(dev);
240
241
dev->opaque = NULL;
242
ram_block_discard_disable(false);
243
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_get_device_id(struct vhost_dev *dev,
54
return ret;
244
return ret;
55
}
245
}
56
-
246
57
-int net_slirp_parse_legacy(QemuOptsList *opts_list, const char *optarg, int *ret)
247
+static void vhost_vdpa_reset_svq(struct vhost_vdpa *v)
58
-{
248
+{
59
- if (strcmp(opts_list->name, "net") != 0 ||
249
+ if (!v->shadow_vqs_enabled) {
60
- strncmp(optarg, "channel,", strlen("channel,")) != 0) {
250
+ return;
61
- return 0;
251
+ }
62
- }
252
+
63
-
253
+ for (unsigned i = 0; i < v->shadow_vqs->len; ++i) {
64
- error_report("The '-net channel' option is deprecated. "
254
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
65
- "Please use '-netdev user,guestfwd=...' instead.");
255
+ vhost_svq_stop(svq);
66
-
256
+ }
67
- /* handle legacy -net channel,port:chr */
257
+}
68
- optarg += strlen("channel,");
258
+
69
-
259
static int vhost_vdpa_reset_device(struct vhost_dev *dev)
70
- if (QTAILQ_EMPTY(&slirp_stacks)) {
260
{
71
- struct slirp_config_str *config;
261
+ struct vhost_vdpa *v = dev->opaque;
72
-
262
int ret;
73
- config = g_malloc(sizeof(*config));
263
uint8_t status = 0;
74
- pstrcpy(config->str, sizeof(config->str), optarg);
264
75
- config->flags = SLIRP_CFG_LEGACY;
265
+ vhost_vdpa_reset_svq(v);
76
- config->next = slirp_configs;
266
+
77
- slirp_configs = config;
267
ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &status);
78
- *ret = 0;
268
trace_vhost_vdpa_reset_device(dev, status);
79
- } else {
269
return ret;
80
- Error *err = NULL;
270
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_get_config(struct vhost_dev *dev, uint8_t *config,
81
- *ret = slirp_guestfwd(QTAILQ_FIRST(&slirp_stacks), optarg, 1, &err);
271
return ret;
82
- if (*ret < 0) {
272
}
83
- error_report_err(err);
273
84
- }
274
+static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev,
85
- }
275
+ struct vhost_vring_file *file)
86
-
276
+{
87
- return 1;
277
+ trace_vhost_vdpa_set_vring_kick(dev, file->index, file->fd);
88
-}
278
+ return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file);
89
-
279
+}
90
diff --git a/qemu-doc.texi b/qemu-doc.texi
280
+
281
+/**
282
+ * Set the shadow virtqueue descriptors to the device
283
+ *
284
+ * @dev: The vhost device model
285
+ * @svq: The shadow virtqueue
286
+ * @idx: The index of the virtqueue in the vhost device
287
+ * @errp: Error
288
+ */
289
+static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
290
+ VhostShadowVirtqueue *svq, unsigned idx,
291
+ Error **errp)
292
+{
293
+ struct vhost_vring_file file = {
294
+ .index = dev->vq_index + idx,
295
+ };
296
+ const EventNotifier *event_notifier = &svq->hdev_kick;
297
+ int r;
298
+
299
+ file.fd = event_notifier_get_fd(event_notifier);
300
+ r = vhost_vdpa_set_vring_dev_kick(dev, &file);
301
+ if (unlikely(r != 0)) {
302
+ error_setg_errno(errp, -r, "Can't set device kick fd");
303
+ }
304
+
305
+ return r == 0;
306
+}
307
+
308
+static bool vhost_vdpa_svqs_start(struct vhost_dev *dev)
309
+{
310
+ struct vhost_vdpa *v = dev->opaque;
311
+ Error *err = NULL;
312
+ unsigned i;
313
+
314
+ if (!v->shadow_vqs) {
315
+ return true;
316
+ }
317
+
318
+ for (i = 0; i < v->shadow_vqs->len; ++i) {
319
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
320
+ bool ok = vhost_vdpa_svq_setup(dev, svq, i, &err);
321
+ if (unlikely(!ok)) {
322
+ error_reportf_err(err, "Cannot setup SVQ %u: ", i);
323
+ return false;
324
+ }
325
+ }
326
+
327
+ return true;
328
+}
329
+
330
static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started)
331
{
332
struct vhost_vdpa *v = dev->opaque;
333
+ bool ok;
334
trace_vhost_vdpa_dev_start(dev, started);
335
336
if (started) {
337
vhost_vdpa_host_notifiers_init(dev);
338
+ ok = vhost_vdpa_svqs_start(dev);
339
+ if (unlikely(!ok)) {
340
+ return -1;
341
+ }
342
vhost_vdpa_set_vring_ready(dev);
343
} else {
344
vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
345
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_get_vring_base(struct vhost_dev *dev,
346
static int vhost_vdpa_set_vring_kick(struct vhost_dev *dev,
347
struct vhost_vring_file *file)
348
{
349
- trace_vhost_vdpa_set_vring_kick(dev, file->index, file->fd);
350
- return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file);
351
+ struct vhost_vdpa *v = dev->opaque;
352
+ int vdpa_idx = file->index - dev->vq_index;
353
+
354
+ if (v->shadow_vqs_enabled) {
355
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx);
356
+ vhost_svq_set_svq_kick_fd(svq, file->fd);
357
+ return 0;
358
+ } else {
359
+ return vhost_vdpa_set_vring_dev_kick(dev, file);
360
+ }
361
}
362
363
static int vhost_vdpa_set_vring_call(struct vhost_dev *dev,
364
diff --git a/include/hw/virtio/vhost-vdpa.h b/include/hw/virtio/vhost-vdpa.h
91
index XXXXXXX..XXXXXXX 100644
365
index XXXXXXX..XXXXXXX 100644
92
--- a/qemu-doc.texi
366
--- a/include/hw/virtio/vhost-vdpa.h
93
+++ b/qemu-doc.texi
367
+++ b/include/hw/virtio/vhost-vdpa.h
94
@@ -XXX,XX +XXX,XX @@ The ``-smb /some/dir'' argument is now a synonym for setting
368
@@ -XXX,XX +XXX,XX @@
95
the ``-netdev user,smb=/some/dir'' argument instead. The new
369
#ifndef HW_VIRTIO_VHOST_VDPA_H
96
syntax allows different settings to be provided per NIC.
370
#define HW_VIRTIO_VHOST_VDPA_H
97
371
98
-@subsection -net channel (since 2.6.0)
372
+#include <gmodule.h>
99
-
373
+
100
-The ``--net channel,ARGS'' argument is now a synonym for setting
374
#include "hw/virtio/virtio.h"
101
-the ``-netdev user,guestfwd=ARGS'' argument instead.
375
#include "standard-headers/linux/vhost_types.h"
102
-
376
103
@subsection -net vlan (since 2.9.0)
377
@@ -XXX,XX +XXX,XX @@ typedef struct vhost_vdpa {
104
378
bool iotlb_batch_begin_sent;
105
The ``-net vlan=NN'' argument is partially replaced with the
379
MemoryListener listener;
380
struct vhost_vdpa_iova_range iova_range;
381
+ bool shadow_vqs_enabled;
382
+ GPtrArray *shadow_vqs;
383
struct vhost_dev *dev;
384
VhostVDPAHostNotifier notifier[VIRTIO_QUEUE_MAX];
385
} VhostVDPA;
106
--
386
--
107
2.7.4
387
2.7.4
108
388
109
389
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
Now that all of the callers have been converted to compute the multicast index
3
This will make qemu aware of the device used buffers, allowing it to
4
inline using new net CRC functions, this function can now be dropped.
4
write the guest memory with its contents if needed.
5
5
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
6
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
7
Acked-by: Michael S. Tsirkin <mst@redhat.com>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
9
---
9
net/net.c | 5 -----
10
hw/virtio/vhost-shadow-virtqueue.c | 38 ++++++++++++++++++++++++++++++++++++++
10
1 file changed, 5 deletions(-)
11
hw/virtio/vhost-shadow-virtqueue.h | 4 ++++
12
hw/virtio/vhost-vdpa.c | 31 +++++++++++++++++++++++++++++--
13
3 files changed, 71 insertions(+), 2 deletions(-)
11
14
12
diff --git a/net/net.c b/net/net.c
15
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
13
index XXXXXXX..XXXXXXX 100644
16
index XXXXXXX..XXXXXXX 100644
14
--- a/net/net.c
17
--- a/hw/virtio/vhost-shadow-virtqueue.c
15
+++ b/net/net.c
18
+++ b/hw/virtio/vhost-shadow-virtqueue.c
16
@@ -XXX,XX +XXX,XX @@ uint32_t net_crc32_le(const uint8_t *p, int len)
19
@@ -XXX,XX +XXX,XX @@ static void vhost_handle_guest_kick(EventNotifier *n)
17
return crc;
18
}
20
}
19
21
20
-unsigned compute_mcast_idx(const uint8_t *ep)
22
/**
21
-{
23
+ * Forward vhost notifications
22
- return net_crc32(ep, ETH_ALEN) >> 26;
24
+ *
23
-}
25
+ * @n: hdev call event notifier, the one that device set to notify svq.
24
-
26
+ */
25
QemuOptsList qemu_netdev_opts = {
27
+static void vhost_svq_handle_call(EventNotifier *n)
26
.name = "netdev",
28
+{
27
.implied_opt_name = "type",
29
+ VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue,
30
+ hdev_call);
31
+ event_notifier_test_and_clear(n);
32
+ event_notifier_set(&svq->svq_call);
33
+}
34
+
35
+/**
36
+ * Set the call notifier for the SVQ to call the guest
37
+ *
38
+ * @svq: Shadow virtqueue
39
+ * @call_fd: call notifier
40
+ *
41
+ * Called on BQL context.
42
+ */
43
+void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd)
44
+{
45
+ if (call_fd == VHOST_FILE_UNBIND) {
46
+ /*
47
+ * Fail event_notifier_set if called handling device call.
48
+ *
49
+ * SVQ still needs device notifications, since it needs to keep
50
+ * forwarding used buffers even with the unbind.
51
+ */
52
+ memset(&svq->svq_call, 0, sizeof(svq->svq_call));
53
+ } else {
54
+ event_notifier_init_fd(&svq->svq_call, call_fd);
55
+ }
56
+}
57
+
58
+/**
59
* Set a new file descriptor for the guest to kick the SVQ and notify for avail
60
*
61
* @svq: The svq
62
@@ -XXX,XX +XXX,XX @@ VhostShadowVirtqueue *vhost_svq_new(void)
63
}
64
65
event_notifier_init_fd(&svq->svq_kick, VHOST_FILE_UNBIND);
66
+ event_notifier_set_handler(&svq->hdev_call, vhost_svq_handle_call);
67
return g_steal_pointer(&svq);
68
69
err_init_hdev_call:
70
@@ -XXX,XX +XXX,XX @@ void vhost_svq_free(gpointer pvq)
71
VhostShadowVirtqueue *vq = pvq;
72
vhost_svq_stop(vq);
73
event_notifier_cleanup(&vq->hdev_kick);
74
+ event_notifier_set_handler(&vq->hdev_call, NULL);
75
event_notifier_cleanup(&vq->hdev_call);
76
g_free(vq);
77
}
78
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
79
index XXXXXXX..XXXXXXX 100644
80
--- a/hw/virtio/vhost-shadow-virtqueue.h
81
+++ b/hw/virtio/vhost-shadow-virtqueue.h
82
@@ -XXX,XX +XXX,XX @@ typedef struct VhostShadowVirtqueue {
83
* So shadow virtqueue must not clean it, or we would lose VirtQueue one.
84
*/
85
EventNotifier svq_kick;
86
+
87
+ /* Guest's call notifier, where the SVQ calls guest. */
88
+ EventNotifier svq_call;
89
} VhostShadowVirtqueue;
90
91
void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd);
92
+void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd);
93
94
void vhost_svq_stop(VhostShadowVirtqueue *svq);
95
96
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
97
index XXXXXXX..XXXXXXX 100644
98
--- a/hw/virtio/vhost-vdpa.c
99
+++ b/hw/virtio/vhost-vdpa.c
100
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev,
101
return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file);
102
}
103
104
+static int vhost_vdpa_set_vring_dev_call(struct vhost_dev *dev,
105
+ struct vhost_vring_file *file)
106
+{
107
+ trace_vhost_vdpa_set_vring_call(dev, file->index, file->fd);
108
+ return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file);
109
+}
110
+
111
/**
112
* Set the shadow virtqueue descriptors to the device
113
*
114
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev,
115
* @svq: The shadow virtqueue
116
* @idx: The index of the virtqueue in the vhost device
117
* @errp: Error
118
+ *
119
+ * Note that this function does not rewind kick file descriptor if cannot set
120
+ * call one.
121
*/
122
static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
123
VhostShadowVirtqueue *svq, unsigned idx,
124
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
125
r = vhost_vdpa_set_vring_dev_kick(dev, &file);
126
if (unlikely(r != 0)) {
127
error_setg_errno(errp, -r, "Can't set device kick fd");
128
+ return false;
129
+ }
130
+
131
+ event_notifier = &svq->hdev_call;
132
+ file.fd = event_notifier_get_fd(event_notifier);
133
+ r = vhost_vdpa_set_vring_dev_call(dev, &file);
134
+ if (unlikely(r != 0)) {
135
+ error_setg_errno(errp, -r, "Can't set device call fd");
136
}
137
138
return r == 0;
139
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_kick(struct vhost_dev *dev,
140
static int vhost_vdpa_set_vring_call(struct vhost_dev *dev,
141
struct vhost_vring_file *file)
142
{
143
- trace_vhost_vdpa_set_vring_call(dev, file->index, file->fd);
144
- return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file);
145
+ struct vhost_vdpa *v = dev->opaque;
146
+
147
+ if (v->shadow_vqs_enabled) {
148
+ int vdpa_idx = file->index - dev->vq_index;
149
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx);
150
+
151
+ vhost_svq_set_svq_call_fd(svq, file->fd);
152
+ return 0;
153
+ } else {
154
+ return vhost_vdpa_set_vring_dev_call(dev, file);
155
+ }
156
}
157
158
static int vhost_vdpa_get_features(struct vhost_dev *dev,
28
--
159
--
29
2.7.4
160
2.7.4
30
161
31
162
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
This provides a standard ethernet CRC32 little-endian implementation.
3
This allows SVQ to negotiate features with the guest and the device. For
4
the device, SVQ is a driver. While this function bypasses all
5
non-transport features, it needs to disable the features that SVQ does
6
not support when forwarding buffers. This includes packed vq layout,
7
indirect descriptors or event idx.
4
8
5
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
9
Future changes can add support to offer more features to the guest,
6
Reviewed-by: Eric Blake <eblake@redhat.com>
10
since the use of VirtQueue gives this for free. This is left out at the
7
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
11
moment for simplicity.
12
13
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
14
Acked-by: Michael S. Tsirkin <mst@redhat.com>
8
Signed-off-by: Jason Wang <jasowang@redhat.com>
15
Signed-off-by: Jason Wang <jasowang@redhat.com>
9
---
16
---
10
include/net/net.h | 2 ++
17
hw/virtio/vhost-shadow-virtqueue.c | 44 ++++++++++++++++++++++++++++++++++++++
11
net/net.c | 22 ++++++++++++++++++++++
18
hw/virtio/vhost-shadow-virtqueue.h | 2 ++
12
2 files changed, 24 insertions(+)
19
hw/virtio/vhost-vdpa.c | 15 +++++++++++++
20
3 files changed, 61 insertions(+)
13
21
14
diff --git a/include/net/net.h b/include/net/net.h
22
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
15
index XXXXXXX..XXXXXXX 100644
23
index XXXXXXX..XXXXXXX 100644
16
--- a/include/net/net.h
24
--- a/hw/virtio/vhost-shadow-virtqueue.c
17
+++ b/include/net/net.h
25
+++ b/hw/virtio/vhost-shadow-virtqueue.c
18
@@ -XXX,XX +XXX,XX @@ NetClientState *net_hub_port_find(int hub_id);
26
@@ -XXX,XX +XXX,XX @@
19
void qdev_set_nic_properties(DeviceState *dev, NICInfo *nd);
27
#include "hw/virtio/vhost-shadow-virtqueue.h"
20
28
21
#define POLYNOMIAL_BE 0x04c11db6
29
#include "qemu/error-report.h"
22
+#define POLYNOMIAL_LE 0xedb88320
30
+#include "qapi/error.h"
23
uint32_t net_crc32(const uint8_t *p, int len);
31
#include "qemu/main-loop.h"
24
+uint32_t net_crc32_le(const uint8_t *p, int len);
32
#include "linux-headers/linux/vhost.h"
25
unsigned compute_mcast_idx(const uint8_t *ep);
33
26
34
/**
27
#define vmstate_offset_macaddr(_state, _field) \
35
+ * Validate the transport device features that both guests can use with the SVQ
28
diff --git a/net/net.c b/net/net.c
36
+ * and SVQs can use with the device.
29
index XXXXXXX..XXXXXXX 100644
37
+ *
30
--- a/net/net.c
38
+ * @dev_features: The features
31
+++ b/net/net.c
39
+ * @errp: Error pointer
32
@@ -XXX,XX +XXX,XX @@ uint32_t net_crc32(const uint8_t *p, int len)
40
+ */
33
return crc;
41
+bool vhost_svq_valid_features(uint64_t features, Error **errp)
34
}
35
36
+uint32_t net_crc32_le(const uint8_t *p, int len)
37
+{
42
+{
38
+ uint32_t crc;
43
+ bool ok = true;
39
+ int carry, i, j;
44
+ uint64_t svq_features = features;
40
+ uint8_t b;
41
+
45
+
42
+ crc = 0xffffffff;
46
+ for (uint64_t b = VIRTIO_TRANSPORT_F_START; b <= VIRTIO_TRANSPORT_F_END;
43
+ for (i = 0; i < len; i++) {
47
+ ++b) {
44
+ b = *p++;
48
+ switch (b) {
45
+ for (j = 0; j < 8; j++) {
49
+ case VIRTIO_F_ANY_LAYOUT:
46
+ carry = (crc & 0x1) ^ (b & 0x01);
50
+ continue;
47
+ crc >>= 1;
51
+
48
+ b >>= 1;
52
+ case VIRTIO_F_ACCESS_PLATFORM:
49
+ if (carry) {
53
+ /* SVQ trust in the host's IOMMU to translate addresses */
50
+ crc ^= POLYNOMIAL_LE;
54
+ case VIRTIO_F_VERSION_1:
55
+ /* SVQ trust that the guest vring is little endian */
56
+ if (!(svq_features & BIT_ULL(b))) {
57
+ svq_features |= BIT_ULL(b);
58
+ ok = false;
59
+ }
60
+ continue;
61
+
62
+ default:
63
+ if (svq_features & BIT_ULL(b)) {
64
+ svq_features &= ~BIT_ULL(b);
65
+ ok = false;
51
+ }
66
+ }
52
+ }
67
+ }
53
+ }
68
+ }
54
+
69
+
55
+ return crc;
70
+ if (!ok) {
71
+ error_setg(errp, "SVQ Invalid device feature flags, offer: 0x%"PRIx64
72
+ ", ok: 0x%"PRIx64, features, svq_features);
73
+ }
74
+ return ok;
56
+}
75
+}
57
+
76
+
58
unsigned compute_mcast_idx(const uint8_t *ep)
77
+/**
78
* Forward guest notifications.
79
*
80
* @n: guest kick event notifier, the one that guest set to notify svq.
81
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
82
index XXXXXXX..XXXXXXX 100644
83
--- a/hw/virtio/vhost-shadow-virtqueue.h
84
+++ b/hw/virtio/vhost-shadow-virtqueue.h
85
@@ -XXX,XX +XXX,XX @@ typedef struct VhostShadowVirtqueue {
86
EventNotifier svq_call;
87
} VhostShadowVirtqueue;
88
89
+bool vhost_svq_valid_features(uint64_t features, Error **errp);
90
+
91
void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd);
92
void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd);
93
94
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
95
index XXXXXXX..XXXXXXX 100644
96
--- a/hw/virtio/vhost-vdpa.c
97
+++ b/hw/virtio/vhost-vdpa.c
98
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v,
99
Error **errp)
59
{
100
{
60
return net_crc32(ep, ETH_ALEN) >> 26;
101
g_autoptr(GPtrArray) shadow_vqs = NULL;
102
+ uint64_t dev_features, svq_features;
103
+ int r;
104
+ bool ok;
105
106
if (!v->shadow_vqs_enabled) {
107
return 0;
108
}
109
110
+ r = hdev->vhost_ops->vhost_get_features(hdev, &dev_features);
111
+ if (r != 0) {
112
+ error_setg_errno(errp, -r, "Can't get vdpa device features");
113
+ return r;
114
+ }
115
+
116
+ svq_features = dev_features;
117
+ ok = vhost_svq_valid_features(svq_features, errp);
118
+ if (unlikely(!ok)) {
119
+ return -1;
120
+ }
121
+
122
shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free);
123
for (unsigned n = 0; n < hdev->nvqs; ++n) {
124
g_autoptr(VhostShadowVirtqueue) svq = vhost_svq_new();
61
--
125
--
62
2.7.4
126
2.7.4
63
127
64
128
diff view generated by jsdifflib
1
From: Thomas Huth <thuth@redhat.com>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
The information how to update the deprecated parameters was too scarce,
3
It reports the shadow virtqueue address from qemu virtual address space.
4
so that some people did not update to the new syntax yet. Provide some
5
more information to make sure that it is clear how to update from the
6
old syntax to the new one.
7
4
8
Signed-off-by: Thomas Huth <thuth@redhat.com>
5
Since this will be different from the guest's vaddr, but the device can
6
access it, SVQ takes special care about its alignment & lack of garbage
7
data. It assumes that IOMMU will work in host_page_size ranges for that.
8
9
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
10
Acked-by: Michael S. Tsirkin <mst@redhat.com>
9
Signed-off-by: Jason Wang <jasowang@redhat.com>
11
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
---
12
---
11
qemu-doc.texi | 33 +++++++++++++++++++++------------
13
hw/virtio/vhost-shadow-virtqueue.c | 29 +++++++++++++++++++++++++++++
12
1 file changed, 21 insertions(+), 12 deletions(-)
14
hw/virtio/vhost-shadow-virtqueue.h | 9 +++++++++
15
2 files changed, 38 insertions(+)
13
16
14
diff --git a/qemu-doc.texi b/qemu-doc.texi
17
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
15
index XXXXXXX..XXXXXXX 100644
18
index XXXXXXX..XXXXXXX 100644
16
--- a/qemu-doc.texi
19
--- a/hw/virtio/vhost-shadow-virtqueue.c
17
+++ b/qemu-doc.texi
20
+++ b/hw/virtio/vhost-shadow-virtqueue.c
18
@@ -XXX,XX +XXX,XX @@ combined with ``-vnc tls-creds=tls0'
21
@@ -XXX,XX +XXX,XX @@ void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd)
19
22
}
20
@subsection -tftp (since 2.6.0)
23
21
24
/**
22
-The ``-tftp /some/dir'' argument is now a synonym for setting
25
+ * Get the shadow vq vring address.
23
-the ``-netdev user,tftp=/some/dir' argument. The new syntax
26
+ * @svq: Shadow virtqueue
24
-allows different settings to be provided per NIC.
27
+ * @addr: Destination to store address
25
+The ``-tftp /some/dir'' argument is replaced by
28
+ */
26
+``-netdev user,id=x,tftp=/some/dir'', either accompanied with
29
+void vhost_svq_get_vring_addr(const VhostShadowVirtqueue *svq,
27
+``-device ...,netdev=x'' (for pluggable NICs) or ``-net nic,netdev=x''
30
+ struct vhost_vring_addr *addr)
28
+(for embedded NICs). The new syntax allows different settings to be
31
+{
29
+provided per NIC.
32
+ addr->desc_user_addr = (uint64_t)(intptr_t)svq->vring.desc;
30
33
+ addr->avail_user_addr = (uint64_t)(intptr_t)svq->vring.avail;
31
@subsection -bootp (since 2.6.0)
34
+ addr->used_user_addr = (uint64_t)(intptr_t)svq->vring.used;
32
35
+}
33
-The ``-bootp /some/file'' argument is now a synonym for setting
36
+
34
-the ``-netdev user,bootp=/some/file' argument. The new syntax
37
+size_t vhost_svq_driver_area_size(const VhostShadowVirtqueue *svq)
35
-allows different settings to be provided per NIC.
38
+{
36
+The ``-bootp /some/file'' argument is replaced by
39
+ size_t desc_size = sizeof(vring_desc_t) * svq->vring.num;
37
+``-netdev user,id=x,bootp=/some/file'', either accompanied with
40
+ size_t avail_size = offsetof(vring_avail_t, ring) +
38
+``-device ...,netdev=x'' (for pluggable NICs) or ``-net nic,netdev=x''
41
+ sizeof(uint16_t) * svq->vring.num;
39
+(for embedded NICs). The new syntax allows different settings to be
42
+
40
+provided per NIC.
43
+ return ROUND_UP(desc_size + avail_size, qemu_real_host_page_size);
41
44
+}
42
@subsection -redir (since 2.6.0)
45
+
43
46
+size_t vhost_svq_device_area_size(const VhostShadowVirtqueue *svq)
44
-The ``-redir ARGS'' argument is now a synonym for setting
47
+{
45
-the ``-netdev user,hostfwd=ARGS'' argument instead. The new
48
+ size_t used_size = offsetof(vring_used_t, ring) +
46
-syntax allows different settings to be provided per NIC.
49
+ sizeof(vring_used_elem_t) * svq->vring.num;
47
+The ``-redir [tcp|udp]:hostport:[guestaddr]:guestport'' argument is
50
+ return ROUND_UP(used_size, qemu_real_host_page_size);
48
+replaced by ``-netdev
51
+}
49
+user,id=x,hostfwd=[tcp|udp]:[hostaddr]:hostport-[guestaddr]:guestport'',
52
+
50
+either accompanied with ``-device ...,netdev=x'' (for pluggable NICs) or
53
+/**
51
+``-net nic,netdev=x'' (for embedded NICs). The new syntax allows different
54
* Set a new file descriptor for the guest to kick the SVQ and notify for avail
52
+settings to be provided per NIC.
55
*
53
56
* @svq: The svq
54
@subsection -smb (since 2.6.0)
57
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
55
58
index XXXXXXX..XXXXXXX 100644
56
-The ``-smb /some/dir'' argument is now a synonym for setting
59
--- a/hw/virtio/vhost-shadow-virtqueue.h
57
-the ``-netdev user,smb=/some/dir'' argument instead. The new
60
+++ b/hw/virtio/vhost-shadow-virtqueue.h
58
-syntax allows different settings to be provided per NIC.
61
@@ -XXX,XX +XXX,XX @@
59
+The ``-smb /some/dir'' argument is replaced by
62
#define VHOST_SHADOW_VIRTQUEUE_H
60
+``-netdev user,id=x,smb=/some/dir'', either accompanied with
63
61
+``-device ...,netdev=x'' (for pluggable NICs) or ``-net nic,netdev=x''
64
#include "qemu/event_notifier.h"
62
+(for embedded NICs). The new syntax allows different settings to be
65
+#include "hw/virtio/virtio.h"
63
+provided per NIC.
66
+#include "standard-headers/linux/vhost_types.h"
64
67
65
@subsection -net vlan (since 2.9.0)
68
/* Shadow virtqueue to relay notifications */
69
typedef struct VhostShadowVirtqueue {
70
+ /* Shadow vring */
71
+ struct vring vring;
72
+
73
/* Shadow kick notifier, sent to vhost */
74
EventNotifier hdev_kick;
75
/* Shadow call notifier, sent to vhost */
76
@@ -XXX,XX +XXX,XX @@ bool vhost_svq_valid_features(uint64_t features, Error **errp);
77
78
void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd);
79
void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd);
80
+void vhost_svq_get_vring_addr(const VhostShadowVirtqueue *svq,
81
+ struct vhost_vring_addr *addr);
82
+size_t vhost_svq_driver_area_size(const VhostShadowVirtqueue *svq);
83
+size_t vhost_svq_device_area_size(const VhostShadowVirtqueue *svq);
84
85
void vhost_svq_stop(VhostShadowVirtqueue *svq);
66
86
67
--
87
--
68
2.7.4
88
2.7.4
69
89
70
90
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
This makes it much easier to compare the multicast CRC calculation endian and
3
First half of the buffers forwarding part, preparing vhost-vdpa
4
bitshift against the Linux driver implementation.
4
callbacks to SVQ to offer it. QEMU cannot enable it at this moment, so
5
this is effectively dead code at the moment, but it helps to reduce
6
patch size.
5
7
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
8
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
9
Acked-by: Michael S. Tsirkin <mst@redhat.com>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
11
---
9
hw/net/ne2000.c | 4 +++-
12
hw/virtio/vhost-vdpa.c | 48 +++++++++++++++++++++++++++++++++++++++++-------
10
1 file changed, 3 insertions(+), 1 deletion(-)
13
1 file changed, 41 insertions(+), 7 deletions(-)
11
14
12
diff --git a/hw/net/ne2000.c b/hw/net/ne2000.c
15
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
13
index XXXXXXX..XXXXXXX 100644
16
index XXXXXXX..XXXXXXX 100644
14
--- a/hw/net/ne2000.c
17
--- a/hw/virtio/vhost-vdpa.c
15
+++ b/hw/net/ne2000.c
18
+++ b/hw/virtio/vhost-vdpa.c
16
@@ -XXX,XX +XXX,XX @@
19
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_get_config(struct vhost_dev *dev, uint8_t *config,
17
*/
20
return ret;
18
#include "qemu/osdep.h"
21
}
19
#include "hw/pci/pci.h"
22
20
+#include "net/net.h"
23
+static int vhost_vdpa_set_dev_vring_base(struct vhost_dev *dev,
21
+#include "net/eth.h"
24
+ struct vhost_vring_state *ring)
22
#include "ne2000.h"
25
+{
23
#include "hw/loader.h"
26
+ trace_vhost_vdpa_set_vring_base(dev, ring->index, ring->num);
24
#include "sysemu/sysemu.h"
27
+ return vhost_vdpa_call(dev, VHOST_SET_VRING_BASE, ring);
25
@@ -XXX,XX +XXX,XX @@ ssize_t ne2000_receive(NetClientState *nc, const uint8_t *buf, size_t size_)
28
+}
26
/* multicast */
29
+
27
if (!(s->rxcr & 0x08))
30
static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev,
28
return size;
31
struct vhost_vring_file *file)
29
- mcast_idx = compute_mcast_idx(buf);
32
{
30
+ mcast_idx = net_crc32(buf, ETH_ALEN) >> 26;
33
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_dev_call(struct vhost_dev *dev,
31
if (!(s->mult[mcast_idx >> 3] & (1 << (mcast_idx & 7))))
34
return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file);
32
return size;
35
}
33
} else if (s->mem[0] == buf[0] &&
36
37
+static int vhost_vdpa_set_vring_dev_addr(struct vhost_dev *dev,
38
+ struct vhost_vring_addr *addr)
39
+{
40
+ trace_vhost_vdpa_set_vring_addr(dev, addr->index, addr->flags,
41
+ addr->desc_user_addr, addr->used_user_addr,
42
+ addr->avail_user_addr,
43
+ addr->log_guest_addr);
44
+
45
+ return vhost_vdpa_call(dev, VHOST_SET_VRING_ADDR, addr);
46
+
47
+}
48
+
49
/**
50
* Set the shadow virtqueue descriptors to the device
51
*
52
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_log_base(struct vhost_dev *dev, uint64_t base,
53
static int vhost_vdpa_set_vring_addr(struct vhost_dev *dev,
54
struct vhost_vring_addr *addr)
55
{
56
- trace_vhost_vdpa_set_vring_addr(dev, addr->index, addr->flags,
57
- addr->desc_user_addr, addr->used_user_addr,
58
- addr->avail_user_addr,
59
- addr->log_guest_addr);
60
- return vhost_vdpa_call(dev, VHOST_SET_VRING_ADDR, addr);
61
+ struct vhost_vdpa *v = dev->opaque;
62
+
63
+ if (v->shadow_vqs_enabled) {
64
+ /*
65
+ * Device vring addr was set at device start. SVQ base is handled by
66
+ * VirtQueue code.
67
+ */
68
+ return 0;
69
+ }
70
+
71
+ return vhost_vdpa_set_vring_dev_addr(dev, addr);
72
}
73
74
static int vhost_vdpa_set_vring_num(struct vhost_dev *dev,
75
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_num(struct vhost_dev *dev,
76
static int vhost_vdpa_set_vring_base(struct vhost_dev *dev,
77
struct vhost_vring_state *ring)
78
{
79
- trace_vhost_vdpa_set_vring_base(dev, ring->index, ring->num);
80
- return vhost_vdpa_call(dev, VHOST_SET_VRING_BASE, ring);
81
+ struct vhost_vdpa *v = dev->opaque;
82
+
83
+ if (v->shadow_vqs_enabled) {
84
+ /*
85
+ * Device vring base was set at device start. SVQ base is handled by
86
+ * VirtQueue code.
87
+ */
88
+ return 0;
89
+ }
90
+
91
+ return vhost_vdpa_set_dev_vring_base(dev, ring);
92
}
93
94
static int vhost_vdpa_get_vring_base(struct vhost_dev *dev,
34
--
95
--
35
2.7.4
96
2.7.4
36
97
37
98
diff view generated by jsdifflib
1
From: Ed Swierk via Qemu-devel <qemu-devel@nongnu.org>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
sum_needed and cptse flags are received from the guest within each
3
Initial version of shadow virtqueue that actually forward buffers. There
4
transmit data descriptor. They are not part of the offload context;
4
is no iommu support at the moment, and that will be addressed in future
5
instead, they determine how to apply a previously received context to
5
patches of this series. Since all vhost-vdpa devices use forced IOMMU,
6
the packet being transmitted:
6
this means that SVQ is not usable at this point of the series on any
7
7
device.
8
- If cptse is set, perform both segmentation and checksum offload
8
9
using the parameters in the TSO context; otherwise just do checksum
9
For simplicity it only supports modern devices, that expects vring
10
offload. (Currently the e1000 device incorrectly stores only one
10
in little endian, with split ring and no event idx or indirect
11
context, which will be fixed in a subsequent patch.)
11
descriptors. Support for them will not be added in this series.
12
12
13
- Depending on the bits set in sum_needed, possibly perform L4
13
It reuses the VirtQueue code for the device part. The driver part is
14
checksum offload and/or IP checksum offload, using the parameters in
14
based on Linux's virtio_ring driver, but with stripped functionality
15
the appropriate context.
15
and optimizations so it's easier to review.
16
16
17
Move these flags out of struct e1000x_txd_props, which is otherwise
17
However, forwarding buffers have some particular pieces: One of the most
18
dedicated to storing values from a context descriptor, and into the
18
unexpected ones is that a guest's buffer can expand through more than
19
per-packet TX struct.
19
one descriptor in SVQ. While this is handled gracefully by qemu's
20
20
emulated virtio devices, it may cause unexpected SVQ queue full. This
21
Signed-off-by: Ed Swierk <eswierk@skyportsystems.com>
21
patch also solves it by checking for this condition at both guest's
22
kicks and device's calls. The code may be more elegant in the future if
23
SVQ code runs in its own iocontext.
24
25
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
26
Acked-by: Michael S. Tsirkin <mst@redhat.com>
22
Signed-off-by: Jason Wang <jasowang@redhat.com>
27
Signed-off-by: Jason Wang <jasowang@redhat.com>
23
---
28
---
24
hw/net/e1000.c | 30 ++++++++++++++++--------------
29
hw/virtio/vhost-shadow-virtqueue.c | 352 ++++++++++++++++++++++++++++++++++++-
25
hw/net/e1000e.c | 4 ++--
30
hw/virtio/vhost-shadow-virtqueue.h | 26 +++
26
hw/net/e1000e_core.c | 16 ++++++++--------
31
hw/virtio/vhost-vdpa.c | 155 +++++++++++++++-
27
hw/net/e1000e_core.h | 2 ++
32
3 files changed, 522 insertions(+), 11 deletions(-)
28
hw/net/e1000x_common.h | 2 --
33
29
5 files changed, 28 insertions(+), 26 deletions(-)
34
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
30
31
diff --git a/hw/net/e1000.c b/hw/net/e1000.c
32
index XXXXXXX..XXXXXXX 100644
35
index XXXXXXX..XXXXXXX 100644
33
--- a/hw/net/e1000.c
36
--- a/hw/virtio/vhost-shadow-virtqueue.c
34
+++ b/hw/net/e1000.c
37
+++ b/hw/virtio/vhost-shadow-virtqueue.c
35
@@ -XXX,XX +XXX,XX @@ typedef struct E1000State_st {
38
@@ -XXX,XX +XXX,XX @@
36
unsigned char data[0x10000];
39
#include "qemu/error-report.h"
37
uint16_t size;
40
#include "qapi/error.h"
38
unsigned char vlan_needed;
41
#include "qemu/main-loop.h"
39
+ unsigned char sum_needed;
42
+#include "qemu/log.h"
40
+ bool cptse;
43
+#include "qemu/memalign.h"
41
e1000x_txd_props props;
44
#include "linux-headers/linux/vhost.h"
42
uint16_t tso_frames;
45
43
} tx;
46
/**
44
@@ -XXX,XX +XXX,XX @@ xmit_seg(E1000State *s)
47
@@ -XXX,XX +XXX,XX @@ bool vhost_svq_valid_features(uint64_t features, Error **errp)
45
unsigned int frames = s->tx.tso_frames, css, sofar;
46
struct e1000_tx *tp = &s->tx;
47
48
- if (tp->props.tse && tp->props.cptse) {
49
+ if (tp->props.tse && tp->cptse) {
50
css = tp->props.ipcss;
51
DBGOUT(TXSUM, "frames %d size %d ipcss %d\n",
52
frames, tp->size, css);
53
@@ -XXX,XX +XXX,XX @@ xmit_seg(E1000State *s)
54
}
55
} else /* UDP */
56
stw_be_p(tp->data+css+4, len);
57
- if (tp->props.sum_needed & E1000_TXD_POPTS_TXSM) {
58
+ if (tp->sum_needed & E1000_TXD_POPTS_TXSM) {
59
unsigned int phsum;
60
// add pseudo-header length before checksum calculation
61
void *sp = tp->data + tp->props.tucso;
62
@@ -XXX,XX +XXX,XX @@ xmit_seg(E1000State *s)
63
tp->tso_frames++;
64
}
65
66
- if (tp->props.sum_needed & E1000_TXD_POPTS_TXSM) {
67
+ if (tp->sum_needed & E1000_TXD_POPTS_TXSM) {
68
putsum(tp->data, tp->size, tp->props.tucso,
69
tp->props.tucss, tp->props.tucse);
70
}
71
- if (tp->props.sum_needed & E1000_TXD_POPTS_IXSM) {
72
+ if (tp->sum_needed & E1000_TXD_POPTS_IXSM) {
73
putsum(tp->data, tp->size, tp->props.ipcso,
74
tp->props.ipcss, tp->props.ipcse);
75
}
76
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
77
} else if (dtype == (E1000_TXD_CMD_DEXT | E1000_TXD_DTYP_D)) {
78
// data descriptor
79
if (tp->size == 0) {
80
- tp->props.sum_needed = le32_to_cpu(dp->upper.data) >> 8;
81
+ tp->sum_needed = le32_to_cpu(dp->upper.data) >> 8;
82
}
83
- tp->props.cptse = (txd_lower & E1000_TXD_CMD_TSE) ? 1 : 0;
84
+ tp->cptse = (txd_lower & E1000_TXD_CMD_TSE) ? 1 : 0;
85
} else {
86
// legacy descriptor
87
- tp->props.cptse = 0;
88
+ tp->cptse = 0;
89
}
90
91
if (e1000x_vlan_enabled(s->mac_reg) &&
92
e1000x_is_vlan_txd(txd_lower) &&
93
- (tp->props.cptse || txd_lower & E1000_TXD_CMD_EOP)) {
94
+ (tp->cptse || txd_lower & E1000_TXD_CMD_EOP)) {
95
tp->vlan_needed = 1;
96
stw_be_p(tp->vlan_header,
97
le16_to_cpu(s->mac_reg[VET]));
98
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
99
}
100
101
addr = le64_to_cpu(dp->buffer_addr);
102
- if (tp->props.tse && tp->props.cptse) {
103
+ if (tp->props.tse && tp->cptse) {
104
msh = tp->props.hdr_len + tp->props.mss;
105
do {
106
bytes = split_size;
107
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
108
}
109
split_size -= bytes;
110
} while (bytes && split_size);
111
- } else if (!tp->props.tse && tp->props.cptse) {
112
+ } else if (!tp->props.tse && tp->cptse) {
113
// context descriptor TSE is not set, while data descriptor TSE is set
114
DBGOUT(TXERR, "TCP segmentation error\n");
115
} else {
116
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
117
118
if (!(txd_lower & E1000_TXD_CMD_EOP))
119
return;
120
- if (!(tp->props.tse && tp->props.cptse && tp->size < tp->props.hdr_len)) {
121
+ if (!(tp->props.tse && tp->cptse && tp->size < tp->props.hdr_len)) {
122
xmit_seg(s);
123
}
124
tp->tso_frames = 0;
125
- tp->props.sum_needed = 0;
126
+ tp->sum_needed = 0;
127
tp->vlan_needed = 0;
128
tp->size = 0;
129
- tp->props.cptse = 0;
130
+ tp->cptse = 0;
131
}
48
}
132
49
133
static uint32_t
50
/**
134
@@ -XXX,XX +XXX,XX @@ static const VMStateDescription vmstate_e1000 = {
51
- * Forward guest notifications.
135
VMSTATE_UINT16(tx.props.mss, E1000State),
52
+ * Number of descriptors that the SVQ can make available from the guest.
136
VMSTATE_UINT16(tx.size, E1000State),
53
+ *
137
VMSTATE_UINT16(tx.tso_frames, E1000State),
54
+ * @svq: The svq
138
- VMSTATE_UINT8(tx.props.sum_needed, E1000State),
55
+ */
139
+ VMSTATE_UINT8(tx.sum_needed, E1000State),
56
+static uint16_t vhost_svq_available_slots(const VhostShadowVirtqueue *svq)
140
VMSTATE_INT8(tx.props.ip, E1000State),
57
+{
141
VMSTATE_INT8(tx.props.tcp, E1000State),
58
+ return svq->vring.num - (svq->shadow_avail_idx - svq->shadow_used_idx);
142
VMSTATE_BUFFER(tx.header, E1000State),
59
+}
143
diff --git a/hw/net/e1000e.c b/hw/net/e1000e.c
60
+
144
index XXXXXXX..XXXXXXX 100644
61
+static void vhost_vring_write_descs(VhostShadowVirtqueue *svq,
145
--- a/hw/net/e1000e.c
62
+ const struct iovec *iovec, size_t num,
146
+++ b/hw/net/e1000e.c
63
+ bool more_descs, bool write)
147
@@ -XXX,XX +XXX,XX @@ static const VMStateDescription e1000e_vmstate_tx = {
64
+{
148
.version_id = 1,
65
+ uint16_t i = svq->free_head, last = svq->free_head;
149
.minimum_version_id = 1,
66
+ unsigned n;
150
.fields = (VMStateField[]) {
67
+ uint16_t flags = write ? cpu_to_le16(VRING_DESC_F_WRITE) : 0;
151
- VMSTATE_UINT8(props.sum_needed, struct e1000e_tx),
68
+ vring_desc_t *descs = svq->vring.desc;
152
+ VMSTATE_UINT8(sum_needed, struct e1000e_tx),
69
+
153
VMSTATE_UINT8(props.ipcss, struct e1000e_tx),
70
+ if (num == 0) {
154
VMSTATE_UINT8(props.ipcso, struct e1000e_tx),
71
+ return;
155
VMSTATE_UINT16(props.ipcse, struct e1000e_tx),
72
+ }
156
@@ -XXX,XX +XXX,XX @@ static const VMStateDescription e1000e_vmstate_tx = {
73
+
157
VMSTATE_INT8(props.ip, struct e1000e_tx),
74
+ for (n = 0; n < num; n++) {
158
VMSTATE_INT8(props.tcp, struct e1000e_tx),
75
+ if (more_descs || (n + 1 < num)) {
159
VMSTATE_BOOL(props.tse, struct e1000e_tx),
76
+ descs[i].flags = flags | cpu_to_le16(VRING_DESC_F_NEXT);
160
- VMSTATE_BOOL(props.cptse, struct e1000e_tx),
77
+ } else {
161
+ VMSTATE_BOOL(cptse, struct e1000e_tx),
78
+ descs[i].flags = flags;
162
VMSTATE_BOOL(skip_cp, struct e1000e_tx),
79
+ }
163
VMSTATE_END_OF_LIST()
80
+ descs[i].addr = cpu_to_le64((hwaddr)(intptr_t)iovec[n].iov_base);
164
}
81
+ descs[i].len = cpu_to_le32(iovec[n].iov_len);
165
diff --git a/hw/net/e1000e_core.c b/hw/net/e1000e_core.c
82
+
166
index XXXXXXX..XXXXXXX 100644
83
+ last = i;
167
--- a/hw/net/e1000e_core.c
84
+ i = cpu_to_le16(descs[i].next);
168
+++ b/hw/net/e1000e_core.c
85
+ }
169
@@ -XXX,XX +XXX,XX @@ e1000e_rss_parse_packet(E1000ECore *core,
86
+
170
static void
87
+ svq->free_head = le16_to_cpu(descs[last].next);
171
e1000e_setup_tx_offloads(E1000ECore *core, struct e1000e_tx *tx)
88
+}
89
+
90
+static bool vhost_svq_add_split(VhostShadowVirtqueue *svq,
91
+ VirtQueueElement *elem, unsigned *head)
92
+{
93
+ unsigned avail_idx;
94
+ vring_avail_t *avail = svq->vring.avail;
95
+
96
+ *head = svq->free_head;
97
+
98
+ /* We need some descriptors here */
99
+ if (unlikely(!elem->out_num && !elem->in_num)) {
100
+ qemu_log_mask(LOG_GUEST_ERROR,
101
+ "Guest provided element with no descriptors");
102
+ return false;
103
+ }
104
+
105
+ vhost_vring_write_descs(svq, elem->out_sg, elem->out_num, elem->in_num > 0,
106
+ false);
107
+ vhost_vring_write_descs(svq, elem->in_sg, elem->in_num, false, true);
108
+
109
+ /*
110
+ * Put the entry in the available array (but don't update avail->idx until
111
+ * they do sync).
112
+ */
113
+ avail_idx = svq->shadow_avail_idx & (svq->vring.num - 1);
114
+ avail->ring[avail_idx] = cpu_to_le16(*head);
115
+ svq->shadow_avail_idx++;
116
+
117
+ /* Update the avail index after write the descriptor */
118
+ smp_wmb();
119
+ avail->idx = cpu_to_le16(svq->shadow_avail_idx);
120
+
121
+ return true;
122
+}
123
+
124
+static bool vhost_svq_add(VhostShadowVirtqueue *svq, VirtQueueElement *elem)
125
+{
126
+ unsigned qemu_head;
127
+ bool ok = vhost_svq_add_split(svq, elem, &qemu_head);
128
+ if (unlikely(!ok)) {
129
+ return false;
130
+ }
131
+
132
+ svq->ring_id_maps[qemu_head] = elem;
133
+ return true;
134
+}
135
+
136
+static void vhost_svq_kick(VhostShadowVirtqueue *svq)
137
+{
138
+ /*
139
+ * We need to expose the available array entries before checking the used
140
+ * flags
141
+ */
142
+ smp_mb();
143
+ if (svq->vring.used->flags & VRING_USED_F_NO_NOTIFY) {
144
+ return;
145
+ }
146
+
147
+ event_notifier_set(&svq->hdev_kick);
148
+}
149
+
150
+/**
151
+ * Forward available buffers.
152
+ *
153
+ * @svq: Shadow VirtQueue
154
+ *
155
+ * Note that this function does not guarantee that all guest's available
156
+ * buffers are available to the device in SVQ avail ring. The guest may have
157
+ * exposed a GPA / GIOVA contiguous buffer, but it may not be contiguous in
158
+ * qemu vaddr.
159
+ *
160
+ * If that happens, guest's kick notifications will be disabled until the
161
+ * device uses some buffers.
162
+ */
163
+static void vhost_handle_guest_kick(VhostShadowVirtqueue *svq)
164
+{
165
+ /* Clear event notifier */
166
+ event_notifier_test_and_clear(&svq->svq_kick);
167
+
168
+ /* Forward to the device as many available buffers as possible */
169
+ do {
170
+ virtio_queue_set_notification(svq->vq, false);
171
+
172
+ while (true) {
173
+ VirtQueueElement *elem;
174
+ bool ok;
175
+
176
+ if (svq->next_guest_avail_elem) {
177
+ elem = g_steal_pointer(&svq->next_guest_avail_elem);
178
+ } else {
179
+ elem = virtqueue_pop(svq->vq, sizeof(*elem));
180
+ }
181
+
182
+ if (!elem) {
183
+ break;
184
+ }
185
+
186
+ if (elem->out_num + elem->in_num > vhost_svq_available_slots(svq)) {
187
+ /*
188
+ * This condition is possible since a contiguous buffer in GPA
189
+ * does not imply a contiguous buffer in qemu's VA
190
+ * scatter-gather segments. If that happens, the buffer exposed
191
+ * to the device needs to be a chain of descriptors at this
192
+ * moment.
193
+ *
194
+ * SVQ cannot hold more available buffers if we are here:
195
+ * queue the current guest descriptor and ignore further kicks
196
+ * until some elements are used.
197
+ */
198
+ svq->next_guest_avail_elem = elem;
199
+ return;
200
+ }
201
+
202
+ ok = vhost_svq_add(svq, elem);
203
+ if (unlikely(!ok)) {
204
+ /* VQ is broken, just return and ignore any other kicks */
205
+ return;
206
+ }
207
+ vhost_svq_kick(svq);
208
+ }
209
+
210
+ virtio_queue_set_notification(svq->vq, true);
211
+ } while (!virtio_queue_empty(svq->vq));
212
+}
213
+
214
+/**
215
+ * Handle guest's kick.
216
*
217
* @n: guest kick event notifier, the one that guest set to notify svq.
218
*/
219
-static void vhost_handle_guest_kick(EventNotifier *n)
220
+static void vhost_handle_guest_kick_notifier(EventNotifier *n)
172
{
221
{
173
- if (tx->props.tse && tx->props.cptse) {
222
VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue, svq_kick);
174
+ if (tx->props.tse && tx->cptse) {
223
event_notifier_test_and_clear(n);
175
net_tx_pkt_build_vheader(tx->tx_pkt, true, true, tx->props.mss);
224
- event_notifier_set(&svq->hdev_kick);
176
net_tx_pkt_update_ip_checksums(tx->tx_pkt);
225
+ vhost_handle_guest_kick(svq);
177
e1000x_inc_reg_if_not_full(core->mac, TSCTC);
226
+}
178
return;
227
+
179
}
228
+static bool vhost_svq_more_used(VhostShadowVirtqueue *svq)
180
229
+{
181
- if (tx->props.sum_needed & E1000_TXD_POPTS_TXSM) {
230
+ if (svq->last_used_idx != svq->shadow_used_idx) {
182
+ if (tx->sum_needed & E1000_TXD_POPTS_TXSM) {
231
+ return true;
183
net_tx_pkt_build_vheader(tx->tx_pkt, false, true, 0);
232
+ }
184
}
233
+
185
234
+ svq->shadow_used_idx = cpu_to_le16(svq->vring.used->idx);
186
- if (tx->props.sum_needed & E1000_TXD_POPTS_IXSM) {
235
+
187
+ if (tx->sum_needed & E1000_TXD_POPTS_IXSM) {
236
+ return svq->last_used_idx != svq->shadow_used_idx;
188
net_tx_pkt_update_ip_hdr_checksum(tx->tx_pkt);
237
}
238
239
/**
240
- * Forward vhost notifications
241
+ * Enable vhost device calls after disable them.
242
+ *
243
+ * @svq: The svq
244
+ *
245
+ * It returns false if there are pending used buffers from the vhost device,
246
+ * avoiding the possible races between SVQ checking for more work and enabling
247
+ * callbacks. True if SVQ used vring has no more pending buffers.
248
+ */
249
+static bool vhost_svq_enable_notification(VhostShadowVirtqueue *svq)
250
+{
251
+ svq->vring.avail->flags &= ~cpu_to_le16(VRING_AVAIL_F_NO_INTERRUPT);
252
+ /* Make sure the flag is written before the read of used_idx */
253
+ smp_mb();
254
+ return !vhost_svq_more_used(svq);
255
+}
256
+
257
+static void vhost_svq_disable_notification(VhostShadowVirtqueue *svq)
258
+{
259
+ svq->vring.avail->flags |= cpu_to_le16(VRING_AVAIL_F_NO_INTERRUPT);
260
+}
261
+
262
+static VirtQueueElement *vhost_svq_get_buf(VhostShadowVirtqueue *svq,
263
+ uint32_t *len)
264
+{
265
+ vring_desc_t *descs = svq->vring.desc;
266
+ const vring_used_t *used = svq->vring.used;
267
+ vring_used_elem_t used_elem;
268
+ uint16_t last_used;
269
+
270
+ if (!vhost_svq_more_used(svq)) {
271
+ return NULL;
272
+ }
273
+
274
+ /* Only get used array entries after they have been exposed by dev */
275
+ smp_rmb();
276
+ last_used = svq->last_used_idx & (svq->vring.num - 1);
277
+ used_elem.id = le32_to_cpu(used->ring[last_used].id);
278
+ used_elem.len = le32_to_cpu(used->ring[last_used].len);
279
+
280
+ svq->last_used_idx++;
281
+ if (unlikely(used_elem.id >= svq->vring.num)) {
282
+ qemu_log_mask(LOG_GUEST_ERROR, "Device %s says index %u is used",
283
+ svq->vdev->name, used_elem.id);
284
+ return NULL;
285
+ }
286
+
287
+ if (unlikely(!svq->ring_id_maps[used_elem.id])) {
288
+ qemu_log_mask(LOG_GUEST_ERROR,
289
+ "Device %s says index %u is used, but it was not available",
290
+ svq->vdev->name, used_elem.id);
291
+ return NULL;
292
+ }
293
+
294
+ descs[used_elem.id].next = svq->free_head;
295
+ svq->free_head = used_elem.id;
296
+
297
+ *len = used_elem.len;
298
+ return g_steal_pointer(&svq->ring_id_maps[used_elem.id]);
299
+}
300
+
301
+static void vhost_svq_flush(VhostShadowVirtqueue *svq,
302
+ bool check_for_avail_queue)
303
+{
304
+ VirtQueue *vq = svq->vq;
305
+
306
+ /* Forward as many used buffers as possible. */
307
+ do {
308
+ unsigned i = 0;
309
+
310
+ vhost_svq_disable_notification(svq);
311
+ while (true) {
312
+ uint32_t len;
313
+ g_autofree VirtQueueElement *elem = vhost_svq_get_buf(svq, &len);
314
+ if (!elem) {
315
+ break;
316
+ }
317
+
318
+ if (unlikely(i >= svq->vring.num)) {
319
+ qemu_log_mask(LOG_GUEST_ERROR,
320
+ "More than %u used buffers obtained in a %u size SVQ",
321
+ i, svq->vring.num);
322
+ virtqueue_fill(vq, elem, len, i);
323
+ virtqueue_flush(vq, i);
324
+ return;
325
+ }
326
+ virtqueue_fill(vq, elem, len, i++);
327
+ }
328
+
329
+ virtqueue_flush(vq, i);
330
+ event_notifier_set(&svq->svq_call);
331
+
332
+ if (check_for_avail_queue && svq->next_guest_avail_elem) {
333
+ /*
334
+ * Avail ring was full when vhost_svq_flush was called, so it's a
335
+ * good moment to make more descriptors available if possible.
336
+ */
337
+ vhost_handle_guest_kick(svq);
338
+ }
339
+ } while (!vhost_svq_enable_notification(svq));
340
+}
341
+
342
+/**
343
+ * Forward used buffers.
344
*
345
* @n: hdev call event notifier, the one that device set to notify svq.
346
+ *
347
+ * Note that we are not making any buffers available in the loop, there is no
348
+ * way that it runs more than virtqueue size times.
349
*/
350
static void vhost_svq_handle_call(EventNotifier *n)
351
{
352
VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue,
353
hdev_call);
354
event_notifier_test_and_clear(n);
355
- event_notifier_set(&svq->svq_call);
356
+ vhost_svq_flush(svq, true);
357
}
358
359
/**
360
@@ -XXX,XX +XXX,XX @@ void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd)
361
if (poll_start) {
362
event_notifier_init_fd(svq_kick, svq_kick_fd);
363
event_notifier_set(svq_kick);
364
- event_notifier_set_handler(svq_kick, vhost_handle_guest_kick);
365
+ event_notifier_set_handler(svq_kick, vhost_handle_guest_kick_notifier);
366
+ }
367
+}
368
+
369
+/**
370
+ * Start the shadow virtqueue operation.
371
+ *
372
+ * @svq: Shadow Virtqueue
373
+ * @vdev: VirtIO device
374
+ * @vq: Virtqueue to shadow
375
+ */
376
+void vhost_svq_start(VhostShadowVirtqueue *svq, VirtIODevice *vdev,
377
+ VirtQueue *vq)
378
+{
379
+ size_t desc_size, driver_size, device_size;
380
+
381
+ svq->next_guest_avail_elem = NULL;
382
+ svq->shadow_avail_idx = 0;
383
+ svq->shadow_used_idx = 0;
384
+ svq->last_used_idx = 0;
385
+ svq->vdev = vdev;
386
+ svq->vq = vq;
387
+
388
+ svq->vring.num = virtio_queue_get_num(vdev, virtio_get_queue_index(vq));
389
+ driver_size = vhost_svq_driver_area_size(svq);
390
+ device_size = vhost_svq_device_area_size(svq);
391
+ svq->vring.desc = qemu_memalign(qemu_real_host_page_size, driver_size);
392
+ desc_size = sizeof(vring_desc_t) * svq->vring.num;
393
+ svq->vring.avail = (void *)((char *)svq->vring.desc + desc_size);
394
+ memset(svq->vring.desc, 0, driver_size);
395
+ svq->vring.used = qemu_memalign(qemu_real_host_page_size, device_size);
396
+ memset(svq->vring.used, 0, device_size);
397
+ svq->ring_id_maps = g_new0(VirtQueueElement *, svq->vring.num);
398
+ for (unsigned i = 0; i < svq->vring.num - 1; i++) {
399
+ svq->vring.desc[i].next = cpu_to_le16(i + 1);
189
}
400
}
190
}
401
}
191
@@ -XXX,XX +XXX,XX @@ e1000e_process_tx_desc(E1000ECore *core,
402
192
return;
403
@@ -XXX,XX +XXX,XX @@ void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd)
193
} else if (dtype == (E1000_TXD_CMD_DEXT | E1000_TXD_DTYP_D)) {
404
void vhost_svq_stop(VhostShadowVirtqueue *svq)
194
/* data descriptor */
405
{
195
- tx->props.sum_needed = le32_to_cpu(dp->upper.data) >> 8;
406
event_notifier_set_handler(&svq->svq_kick, NULL);
196
- tx->props.cptse = (txd_lower & E1000_TXD_CMD_TSE) ? 1 : 0;
407
+ g_autofree VirtQueueElement *next_avail_elem = NULL;
197
+ tx->sum_needed = le32_to_cpu(dp->upper.data) >> 8;
408
+
198
+ tx->cptse = (txd_lower & E1000_TXD_CMD_TSE) ? 1 : 0;
409
+ if (!svq->vq) {
199
e1000e_process_ts_option(core, dp);
410
+ return;
411
+ }
412
+
413
+ /* Send all pending used descriptors to guest */
414
+ vhost_svq_flush(svq, false);
415
+
416
+ for (unsigned i = 0; i < svq->vring.num; ++i) {
417
+ g_autofree VirtQueueElement *elem = NULL;
418
+ elem = g_steal_pointer(&svq->ring_id_maps[i]);
419
+ if (elem) {
420
+ virtqueue_detach_element(svq->vq, elem, 0);
421
+ }
422
+ }
423
+
424
+ next_avail_elem = g_steal_pointer(&svq->next_guest_avail_elem);
425
+ if (next_avail_elem) {
426
+ virtqueue_detach_element(svq->vq, next_avail_elem, 0);
427
+ }
428
+ svq->vq = NULL;
429
+ g_free(svq->ring_id_maps);
430
+ qemu_vfree(svq->vring.desc);
431
+ qemu_vfree(svq->vring.used);
432
}
433
434
/**
435
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
436
index XXXXXXX..XXXXXXX 100644
437
--- a/hw/virtio/vhost-shadow-virtqueue.h
438
+++ b/hw/virtio/vhost-shadow-virtqueue.h
439
@@ -XXX,XX +XXX,XX @@ typedef struct VhostShadowVirtqueue {
440
441
/* Guest's call notifier, where the SVQ calls guest. */
442
EventNotifier svq_call;
443
+
444
+ /* Virtio queue shadowing */
445
+ VirtQueue *vq;
446
+
447
+ /* Virtio device */
448
+ VirtIODevice *vdev;
449
+
450
+ /* Map for use the guest's descriptors */
451
+ VirtQueueElement **ring_id_maps;
452
+
453
+ /* Next VirtQueue element that guest made available */
454
+ VirtQueueElement *next_guest_avail_elem;
455
+
456
+ /* Next head to expose to the device */
457
+ uint16_t shadow_avail_idx;
458
+
459
+ /* Next free descriptor */
460
+ uint16_t free_head;
461
+
462
+ /* Last seen used idx */
463
+ uint16_t shadow_used_idx;
464
+
465
+ /* Next head to consume from the device */
466
+ uint16_t last_used_idx;
467
} VhostShadowVirtqueue;
468
469
bool vhost_svq_valid_features(uint64_t features, Error **errp);
470
@@ -XXX,XX +XXX,XX @@ void vhost_svq_get_vring_addr(const VhostShadowVirtqueue *svq,
471
size_t vhost_svq_driver_area_size(const VhostShadowVirtqueue *svq);
472
size_t vhost_svq_device_area_size(const VhostShadowVirtqueue *svq);
473
474
+void vhost_svq_start(VhostShadowVirtqueue *svq, VirtIODevice *vdev,
475
+ VirtQueue *vq);
476
void vhost_svq_stop(VhostShadowVirtqueue *svq);
477
478
VhostShadowVirtqueue *vhost_svq_new(void);
479
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
480
index XXXXXXX..XXXXXXX 100644
481
--- a/hw/virtio/vhost-vdpa.c
482
+++ b/hw/virtio/vhost-vdpa.c
483
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_dev_addr(struct vhost_dev *dev,
484
* Note that this function does not rewind kick file descriptor if cannot set
485
* call one.
486
*/
487
-static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
488
- VhostShadowVirtqueue *svq, unsigned idx,
489
- Error **errp)
490
+static int vhost_vdpa_svq_set_fds(struct vhost_dev *dev,
491
+ VhostShadowVirtqueue *svq, unsigned idx,
492
+ Error **errp)
493
{
494
struct vhost_vring_file file = {
495
.index = dev->vq_index + idx,
496
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
497
r = vhost_vdpa_set_vring_dev_kick(dev, &file);
498
if (unlikely(r != 0)) {
499
error_setg_errno(errp, -r, "Can't set device kick fd");
500
- return false;
501
+ return r;
502
}
503
504
event_notifier = &svq->hdev_call;
505
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
506
error_setg_errno(errp, -r, "Can't set device call fd");
507
}
508
509
+ return r;
510
+}
511
+
512
+/**
513
+ * Unmap a SVQ area in the device
514
+ */
515
+static bool vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v, hwaddr iova,
516
+ hwaddr size)
517
+{
518
+ int r;
519
+
520
+ size = ROUND_UP(size, qemu_real_host_page_size);
521
+ r = vhost_vdpa_dma_unmap(v, iova, size);
522
+ return r == 0;
523
+}
524
+
525
+static bool vhost_vdpa_svq_unmap_rings(struct vhost_dev *dev,
526
+ const VhostShadowVirtqueue *svq)
527
+{
528
+ struct vhost_vdpa *v = dev->opaque;
529
+ struct vhost_vring_addr svq_addr;
530
+ size_t device_size = vhost_svq_device_area_size(svq);
531
+ size_t driver_size = vhost_svq_driver_area_size(svq);
532
+ bool ok;
533
+
534
+ vhost_svq_get_vring_addr(svq, &svq_addr);
535
+
536
+ ok = vhost_vdpa_svq_unmap_ring(v, svq_addr.desc_user_addr, driver_size);
537
+ if (unlikely(!ok)) {
538
+ return false;
539
+ }
540
+
541
+ return vhost_vdpa_svq_unmap_ring(v, svq_addr.used_user_addr, device_size);
542
+}
543
+
544
+/**
545
+ * Map the shadow virtqueue rings in the device
546
+ *
547
+ * @dev: The vhost device
548
+ * @svq: The shadow virtqueue
549
+ * @addr: Assigned IOVA addresses
550
+ * @errp: Error pointer
551
+ */
552
+static bool vhost_vdpa_svq_map_rings(struct vhost_dev *dev,
553
+ const VhostShadowVirtqueue *svq,
554
+ struct vhost_vring_addr *addr,
555
+ Error **errp)
556
+{
557
+ struct vhost_vdpa *v = dev->opaque;
558
+ size_t device_size = vhost_svq_device_area_size(svq);
559
+ size_t driver_size = vhost_svq_driver_area_size(svq);
560
+ int r;
561
+
562
+ ERRP_GUARD();
563
+ vhost_svq_get_vring_addr(svq, addr);
564
+
565
+ r = vhost_vdpa_dma_map(v, addr->desc_user_addr, driver_size,
566
+ (void *)(uintptr_t)addr->desc_user_addr, true);
567
+ if (unlikely(r != 0)) {
568
+ error_setg_errno(errp, -r, "Cannot create vq driver region: ");
569
+ return false;
570
+ }
571
+
572
+ r = vhost_vdpa_dma_map(v, addr->used_user_addr, device_size,
573
+ (void *)(intptr_t)addr->used_user_addr, false);
574
+ if (unlikely(r != 0)) {
575
+ error_setg_errno(errp, -r, "Cannot create vq device region: ");
576
+ }
577
+
578
+ return r == 0;
579
+}
580
+
581
+static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
582
+ VhostShadowVirtqueue *svq, unsigned idx,
583
+ Error **errp)
584
+{
585
+ uint16_t vq_index = dev->vq_index + idx;
586
+ struct vhost_vring_state s = {
587
+ .index = vq_index,
588
+ };
589
+ int r;
590
+
591
+ r = vhost_vdpa_set_dev_vring_base(dev, &s);
592
+ if (unlikely(r)) {
593
+ error_setg_errno(errp, -r, "Cannot set vring base");
594
+ return false;
595
+ }
596
+
597
+ r = vhost_vdpa_svq_set_fds(dev, svq, idx, errp);
598
return r == 0;
599
}
600
601
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_svqs_start(struct vhost_dev *dev)
602
}
603
604
for (i = 0; i < v->shadow_vqs->len; ++i) {
605
+ VirtQueue *vq = virtio_get_queue(dev->vdev, dev->vq_index + i);
606
VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
607
+ struct vhost_vring_addr addr = {
608
+ .index = i,
609
+ };
610
+ int r;
611
bool ok = vhost_vdpa_svq_setup(dev, svq, i, &err);
612
if (unlikely(!ok)) {
613
- error_reportf_err(err, "Cannot setup SVQ %u: ", i);
614
+ goto err;
615
+ }
616
+
617
+ vhost_svq_start(svq, dev->vdev, vq);
618
+ ok = vhost_vdpa_svq_map_rings(dev, svq, &addr, &err);
619
+ if (unlikely(!ok)) {
620
+ goto err_map;
621
+ }
622
+
623
+ /* Override vring GPA set by vhost subsystem */
624
+ r = vhost_vdpa_set_vring_dev_addr(dev, &addr);
625
+ if (unlikely(r != 0)) {
626
+ error_setg_errno(&err, -r, "Cannot set device address");
627
+ goto err_set_addr;
628
+ }
629
+ }
630
+
631
+ return true;
632
+
633
+err_set_addr:
634
+ vhost_vdpa_svq_unmap_rings(dev, g_ptr_array_index(v->shadow_vqs, i));
635
+
636
+err_map:
637
+ vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, i));
638
+
639
+err:
640
+ error_reportf_err(err, "Cannot setup SVQ %u: ", i);
641
+ for (unsigned j = 0; j < i; ++j) {
642
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, j);
643
+ vhost_vdpa_svq_unmap_rings(dev, svq);
644
+ vhost_svq_stop(svq);
645
+ }
646
+
647
+ return false;
648
+}
649
+
650
+static bool vhost_vdpa_svqs_stop(struct vhost_dev *dev)
651
+{
652
+ struct vhost_vdpa *v = dev->opaque;
653
+
654
+ if (!v->shadow_vqs) {
655
+ return true;
656
+ }
657
+
658
+ for (unsigned i = 0; i < v->shadow_vqs->len; ++i) {
659
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
660
+ bool ok = vhost_vdpa_svq_unmap_rings(dev, svq);
661
+ if (unlikely(!ok)) {
662
return false;
663
}
664
}
665
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started)
666
}
667
vhost_vdpa_set_vring_ready(dev);
200
} else {
668
} else {
201
/* legacy descriptor */
669
+ ok = vhost_vdpa_svqs_stop(dev);
202
e1000e_process_ts_option(core, dp);
670
+ if (unlikely(!ok)) {
203
- tx->props.cptse = 0;
671
+ return -1;
204
+ tx->cptse = 0;
672
+ }
673
vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
205
}
674
}
206
675
207
addr = le64_to_cpu(dp->buffer_addr);
208
@@ -XXX,XX +XXX,XX @@ e1000e_process_tx_desc(E1000ECore *core,
209
tx->skip_cp = false;
210
net_tx_pkt_reset(tx->tx_pkt);
211
212
- tx->props.sum_needed = 0;
213
- tx->props.cptse = 0;
214
+ tx->sum_needed = 0;
215
+ tx->cptse = 0;
216
}
217
}
218
219
diff --git a/hw/net/e1000e_core.h b/hw/net/e1000e_core.h
220
index XXXXXXX..XXXXXXX 100644
221
--- a/hw/net/e1000e_core.h
222
+++ b/hw/net/e1000e_core.h
223
@@ -XXX,XX +XXX,XX @@ struct E1000Core {
224
e1000x_txd_props props;
225
226
bool skip_cp;
227
+ unsigned char sum_needed;
228
+ bool cptse;
229
struct NetTxPkt *tx_pkt;
230
} tx[E1000E_NUM_QUEUES];
231
232
diff --git a/hw/net/e1000x_common.h b/hw/net/e1000x_common.h
233
index XXXXXXX..XXXXXXX 100644
234
--- a/hw/net/e1000x_common.h
235
+++ b/hw/net/e1000x_common.h
236
@@ -XXX,XX +XXX,XX @@ void e1000x_update_regs_on_autoneg_done(uint32_t *mac, uint16_t *phy);
237
void e1000x_increase_size_stats(uint32_t *mac, const int *size_regs, int size);
238
239
typedef struct e1000x_txd_props {
240
- unsigned char sum_needed;
241
uint8_t ipcss;
242
uint8_t ipcso;
243
uint16_t ipcse;
244
@@ -XXX,XX +XXX,XX @@ typedef struct e1000x_txd_props {
245
int8_t ip;
246
int8_t tcp;
247
bool tse;
248
- bool cptse;
249
} e1000x_txd_props;
250
251
void e1000x_read_tx_ctx_descr(struct e1000_context_desc *d,
252
--
676
--
253
2.7.4
677
2.7.4
254
678
255
679
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
Instead of sunhme_crc32_le() using its own implementation, we can simply call
3
This iova tree function allows it to look for a hole in allocated
4
net_crc32_le() directly and apply the bit shift inline.
4
regions and return a totally new translation for a given translated
5
5
address.
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
6
7
Reviewed-by: Eric Blake <eblake@redhat.com>
7
It's usage is mainly to allow devices to access qemu address space,
8
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
8
remapping guest's one into a new iova space where qemu can add chunks of
9
addresses.
10
11
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
12
Reviewed-by: Peter Xu <peterx@redhat.com>
13
Acked-by: Michael S. Tsirkin <mst@redhat.com>
9
Signed-off-by: Jason Wang <jasowang@redhat.com>
14
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
---
15
---
11
hw/net/sunhme.c | 25 +------------------------
16
include/qemu/iova-tree.h | 18 +++++++
12
1 file changed, 1 insertion(+), 24 deletions(-)
17
util/iova-tree.c | 136 +++++++++++++++++++++++++++++++++++++++++++++++
13
18
2 files changed, 154 insertions(+)
14
diff --git a/hw/net/sunhme.c b/hw/net/sunhme.c
19
20
diff --git a/include/qemu/iova-tree.h b/include/qemu/iova-tree.h
15
index XXXXXXX..XXXXXXX 100644
21
index XXXXXXX..XXXXXXX 100644
16
--- a/hw/net/sunhme.c
22
--- a/include/qemu/iova-tree.h
17
+++ b/hw/net/sunhme.c
23
+++ b/include/qemu/iova-tree.h
18
@@ -XXX,XX +XXX,XX @@ static inline void sunhme_set_rx_ring_nr(SunHMEState *s, int i)
24
@@ -XXX,XX +XXX,XX @@
19
s->erxregs[HME_ERXI_RING >> 2] = ring;
25
#define IOVA_OK (0)
26
#define IOVA_ERR_INVALID (-1) /* Invalid parameters */
27
#define IOVA_ERR_OVERLAP (-2) /* IOVA range overlapped */
28
+#define IOVA_ERR_NOMEM (-3) /* Cannot allocate */
29
30
typedef struct IOVATree IOVATree;
31
typedef struct DMAMap {
32
@@ -XXX,XX +XXX,XX @@ const DMAMap *iova_tree_find_address(const IOVATree *tree, hwaddr iova);
33
void iova_tree_foreach(IOVATree *tree, iova_tree_iterator iterator);
34
35
/**
36
+ * iova_tree_alloc_map:
37
+ *
38
+ * @tree: the iova tree to allocate from
39
+ * @map: the new map (as translated addr & size) to allocate in the iova region
40
+ * @iova_begin: the minimum address of the allocation
41
+ * @iova_end: the maximum addressable direction of the allocation
42
+ *
43
+ * Allocates a new region of a given size, between iova_min and iova_max.
44
+ *
45
+ * Return: Same as iova_tree_insert, but cannot overlap and can return error if
46
+ * iova tree is out of free contiguous range. The caller gets the assigned iova
47
+ * in map->iova.
48
+ */
49
+int iova_tree_alloc_map(IOVATree *tree, DMAMap *map, hwaddr iova_begin,
50
+ hwaddr iova_end);
51
+
52
+/**
53
* iova_tree_destroy:
54
*
55
* @tree: the iova tree to destroy
56
diff --git a/util/iova-tree.c b/util/iova-tree.c
57
index XXXXXXX..XXXXXXX 100644
58
--- a/util/iova-tree.c
59
+++ b/util/iova-tree.c
60
@@ -XXX,XX +XXX,XX @@ struct IOVATree {
61
GTree *tree;
62
};
63
64
+/* Args to pass to iova_tree_alloc foreach function. */
65
+struct IOVATreeAllocArgs {
66
+ /* Size of the desired allocation */
67
+ size_t new_size;
68
+
69
+ /* The minimum address allowed in the allocation */
70
+ hwaddr iova_begin;
71
+
72
+ /* Map at the left of the hole, can be NULL if "this" is first one */
73
+ const DMAMap *prev;
74
+
75
+ /* Map at the right of the hole, can be NULL if "prev" is the last one */
76
+ const DMAMap *this;
77
+
78
+ /* If found, we fill in the IOVA here */
79
+ hwaddr iova_result;
80
+
81
+ /* Whether have we found a valid IOVA */
82
+ bool iova_found;
83
+};
84
+
85
+/**
86
+ * Iterate args to the next hole
87
+ *
88
+ * @args: The alloc arguments
89
+ * @next: The next mapping in the tree. Can be NULL to signal the last one
90
+ */
91
+static void iova_tree_alloc_args_iterate(struct IOVATreeAllocArgs *args,
92
+ const DMAMap *next)
93
+{
94
+ args->prev = args->this;
95
+ args->this = next;
96
+}
97
+
98
static int iova_tree_compare(gconstpointer a, gconstpointer b, gpointer data)
99
{
100
const DMAMap *m1 = a, *m2 = b;
101
@@ -XXX,XX +XXX,XX @@ int iova_tree_remove(IOVATree *tree, const DMAMap *map)
102
return IOVA_OK;
20
}
103
}
21
104
22
-#define POLYNOMIAL_LE 0xedb88320
105
+/**
23
-static uint32_t sunhme_crc32_le(const uint8_t *p, int len)
106
+ * Try to find an unallocated IOVA range between prev and this elements.
24
-{
107
+ *
25
- uint32_t crc;
108
+ * @args: Arguments to allocation
26
- int carry, i, j;
109
+ *
27
- uint8_t b;
110
+ * Cases:
28
-
111
+ *
29
- crc = 0xffffffff;
112
+ * (1) !prev, !this: No entries allocated, always succeed
30
- for (i = 0; i < len; i++) {
113
+ *
31
- b = *p++;
114
+ * (2) !prev, this: We're iterating at the 1st element.
32
- for (j = 0; j < 8; j++) {
115
+ *
33
- carry = (crc & 0x1) ^ (b & 0x01);
116
+ * (3) prev, !this: We're iterating at the last element.
34
- crc >>= 1;
117
+ *
35
- b >>= 1;
118
+ * (4) prev, this: this is the most common case, we'll try to find a hole
36
- if (carry) {
119
+ * between "prev" and "this" mapping.
37
- crc = crc ^ POLYNOMIAL_LE;
120
+ *
38
- }
121
+ * Note that this function assumes the last valid iova is HWADDR_MAX, but it
39
- }
122
+ * searches linearly so it's easy to discard the result if it's not the case.
40
- }
123
+ */
41
-
124
+static void iova_tree_alloc_map_in_hole(struct IOVATreeAllocArgs *args)
42
- return crc;
125
+{
43
-}
126
+ const DMAMap *prev = args->prev, *this = args->this;
44
-
127
+ uint64_t hole_start, hole_last;
45
#define MIN_BUF_SIZE 60
128
+
46
129
+ if (this && this->iova + this->size < args->iova_begin) {
47
static ssize_t sunhme_receive(NetClientState *nc, const uint8_t *buf,
130
+ return;
48
@@ -XXX,XX +XXX,XX @@ static ssize_t sunhme_receive(NetClientState *nc, const uint8_t *buf,
131
+ }
49
trace_sunhme_rx_filter_bcast_match();
132
+
50
} else if (s->macregs[HME_MACI_RXCFG >> 2] & HME_MAC_RXCFG_HENABLE) {
133
+ hole_start = MAX(prev ? prev->iova + prev->size + 1 : 0, args->iova_begin);
51
/* Didn't match local address, check hash filter */
134
+ hole_last = this ? this->iova : HWADDR_MAX;
52
- int mcast_idx = sunhme_crc32_le(buf, 6) >> 26;
135
+
53
+ int mcast_idx = net_crc32_le(buf, ETH_ALEN) >> 26;
136
+ if (hole_last - hole_start > args->new_size) {
54
if (!(s->macregs[(HME_MACI_HASHTAB0 >> 2) - (mcast_idx >> 4)] &
137
+ args->iova_result = hole_start;
55
(1 << (mcast_idx & 0xf)))) {
138
+ args->iova_found = true;
56
/* Didn't match hash filter */
139
+ }
140
+}
141
+
142
+/**
143
+ * Foreach dma node in the tree, compare if there is a hole with its previous
144
+ * node (or minimum iova address allowed) and the node.
145
+ *
146
+ * @key: Node iterating
147
+ * @value: Node iterating
148
+ * @pargs: Struct to communicate with the outside world
149
+ *
150
+ * Return: false to keep iterating, true if needs break.
151
+ */
152
+static gboolean iova_tree_alloc_traverse(gpointer key, gpointer value,
153
+ gpointer pargs)
154
+{
155
+ struct IOVATreeAllocArgs *args = pargs;
156
+ DMAMap *node = value;
157
+
158
+ assert(key == value);
159
+
160
+ iova_tree_alloc_args_iterate(args, node);
161
+ iova_tree_alloc_map_in_hole(args);
162
+ return args->iova_found;
163
+}
164
+
165
+int iova_tree_alloc_map(IOVATree *tree, DMAMap *map, hwaddr iova_begin,
166
+ hwaddr iova_last)
167
+{
168
+ struct IOVATreeAllocArgs args = {
169
+ .new_size = map->size,
170
+ .iova_begin = iova_begin,
171
+ };
172
+
173
+ if (unlikely(iova_last < iova_begin)) {
174
+ return IOVA_ERR_INVALID;
175
+ }
176
+
177
+ /*
178
+ * Find a valid hole for the mapping
179
+ *
180
+ * Assuming low iova_begin, so no need to do a binary search to
181
+ * locate the first node.
182
+ *
183
+ * TODO: Replace all this with g_tree_node_first/next/last when available
184
+ * (from glib since 2.68). To do it with g_tree_foreach complicates the
185
+ * code a lot.
186
+ *
187
+ */
188
+ g_tree_foreach(tree->tree, iova_tree_alloc_traverse, &args);
189
+ if (!args.iova_found) {
190
+ /*
191
+ * Either tree is empty or the last hole is still not checked.
192
+ * g_tree_foreach does not compare (last, iova_last] range, so we check
193
+ * it here.
194
+ */
195
+ iova_tree_alloc_args_iterate(&args, NULL);
196
+ iova_tree_alloc_map_in_hole(&args);
197
+ }
198
+
199
+ if (!args.iova_found || args.iova_result + map->size > iova_last) {
200
+ return IOVA_ERR_NOMEM;
201
+ }
202
+
203
+ map->iova = args.iova_result;
204
+ return iova_tree_insert(tree, map);
205
+}
206
+
207
void iova_tree_destroy(IOVATree *tree)
208
{
209
g_tree_destroy(tree->tree);
57
--
210
--
58
2.7.4
211
2.7.4
59
212
60
213
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
Separate out the standard ethernet CRC32 calculation into a new net_crc32()
3
This function does the reverse operation of iova_tree_find: To look for
4
function, renaming the constant POLYNOMIAL to POLYNOMIAL_BE to make it clear
4
a mapping that match a translated address so we can do the reverse.
5
that this is a big-endian CRC32 calculation.
6
5
7
As part of the constant rename, remove the duplicate definition of POLYNOMIAL
6
This have linear complexity instead of logarithmic, but it supports
8
from eepro100.c and use the new POLYNOMIAL_BE constant instead.
7
overlapping HVA. Future developments could reduce it.
9
8
10
Once this is complete remove the existing CRC32 implementation from
9
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
11
compute_mcast_idx() and call the new net_crc32() function in its place.
10
Acked-by: Michael S. Tsirkin <mst@redhat.com>
12
13
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
14
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
15
Signed-off-by: Jason Wang <jasowang@redhat.com>
11
Signed-off-by: Jason Wang <jasowang@redhat.com>
16
---
12
---
17
hw/net/eepro100.c | 4 +---
13
include/qemu/iova-tree.h | 20 +++++++++++++++++++-
18
include/net/net.h | 3 ++-
14
util/iova-tree.c | 34 ++++++++++++++++++++++++++++++++++
19
net/net.c | 16 +++++++++++-----
15
2 files changed, 53 insertions(+), 1 deletion(-)
20
3 files changed, 14 insertions(+), 9 deletions(-)
21
16
22
diff --git a/hw/net/eepro100.c b/hw/net/eepro100.c
17
diff --git a/include/qemu/iova-tree.h b/include/qemu/iova-tree.h
23
index XXXXXXX..XXXXXXX 100644
18
index XXXXXXX..XXXXXXX 100644
24
--- a/hw/net/eepro100.c
19
--- a/include/qemu/iova-tree.h
25
+++ b/hw/net/eepro100.c
20
+++ b/include/qemu/iova-tree.h
26
@@ -XXX,XX +XXX,XX @@ static const uint16_t eepro100_mdi_mask[] = {
21
@@ -XXX,XX +XXX,XX @@ int iova_tree_remove(IOVATree *tree, const DMAMap *map);
27
0xffff, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
22
* @tree: the iova tree to search from
23
* @map: the mapping to search
24
*
25
- * Search for a mapping in the iova tree that overlaps with the
26
+ * Search for a mapping in the iova tree that iova overlaps with the
27
* mapping range specified. Only the first found mapping will be
28
* returned.
29
*
30
@@ -XXX,XX +XXX,XX @@ int iova_tree_remove(IOVATree *tree, const DMAMap *map);
31
const DMAMap *iova_tree_find(const IOVATree *tree, const DMAMap *map);
32
33
/**
34
+ * iova_tree_find_iova:
35
+ *
36
+ * @tree: the iova tree to search from
37
+ * @map: the mapping to search
38
+ *
39
+ * Search for a mapping in the iova tree that translated_addr overlaps with the
40
+ * mapping range specified. Only the first found mapping will be
41
+ * returned.
42
+ *
43
+ * Return: DMAMap pointer if found, or NULL if not found. Note that
44
+ * the returned DMAMap pointer is maintained internally. User should
45
+ * only read the content but never modify or free the content. Also,
46
+ * user is responsible to make sure the pointer is valid (say, no
47
+ * concurrent deletion in progress).
48
+ */
49
+const DMAMap *iova_tree_find_iova(const IOVATree *tree, const DMAMap *map);
50
+
51
+/**
52
* iova_tree_find_address:
53
*
54
* @tree: the iova tree to search from
55
diff --git a/util/iova-tree.c b/util/iova-tree.c
56
index XXXXXXX..XXXXXXX 100644
57
--- a/util/iova-tree.c
58
+++ b/util/iova-tree.c
59
@@ -XXX,XX +XXX,XX @@ struct IOVATreeAllocArgs {
60
bool iova_found;
28
};
61
};
29
62
30
-#define POLYNOMIAL 0x04c11db6
63
+typedef struct IOVATreeFindIOVAArgs {
31
-
64
+ const DMAMap *needle;
32
static E100PCIDeviceInfo *eepro100_get_class(EEPRO100State *s);
65
+ const DMAMap *result;
33
66
+} IOVATreeFindIOVAArgs;
34
/* From FreeBSD (locally modified). */
35
@@ -XXX,XX +XXX,XX @@ static unsigned e100_compute_mcast_idx(const uint8_t *ep)
36
crc <<= 1;
37
b >>= 1;
38
if (carry) {
39
- crc = ((crc ^ POLYNOMIAL) | carry);
40
+ crc = ((crc ^ POLYNOMIAL_BE) | carry);
41
}
42
}
43
}
44
diff --git a/include/net/net.h b/include/net/net.h
45
index XXXXXXX..XXXXXXX 100644
46
--- a/include/net/net.h
47
+++ b/include/net/net.h
48
@@ -XXX,XX +XXX,XX @@ NetClientState *net_hub_port_find(int hub_id);
49
50
void qdev_set_nic_properties(DeviceState *dev, NICInfo *nd);
51
52
-#define POLYNOMIAL 0x04c11db6
53
+#define POLYNOMIAL_BE 0x04c11db6
54
+uint32_t net_crc32(const uint8_t *p, int len);
55
unsigned compute_mcast_idx(const uint8_t *ep);
56
57
#define vmstate_offset_macaddr(_state, _field) \
58
diff --git a/net/net.c b/net/net.c
59
index XXXXXXX..XXXXXXX 100644
60
--- a/net/net.c
61
+++ b/net/net.c
62
@@ -XXX,XX +XXX,XX @@ int net_client_parse(QemuOptsList *opts_list, const char *optarg)
63
64
/* From FreeBSD */
65
/* XXX: optimize */
66
-unsigned compute_mcast_idx(const uint8_t *ep)
67
+uint32_t net_crc32(const uint8_t *p, int len)
68
{
69
uint32_t crc;
70
int carry, i, j;
71
uint8_t b;
72
73
crc = 0xffffffff;
74
- for (i = 0; i < 6; i++) {
75
- b = *ep++;
76
+ for (i = 0; i < len; i++) {
77
+ b = *p++;
78
for (j = 0; j < 8; j++) {
79
carry = ((crc & 0x80000000L) ? 1 : 0) ^ (b & 0x01);
80
crc <<= 1;
81
b >>= 1;
82
if (carry) {
83
- crc = ((crc ^ POLYNOMIAL) | carry);
84
+ crc = ((crc ^ POLYNOMIAL_BE) | carry);
85
}
86
}
87
}
88
- return crc >> 26;
89
+
67
+
90
+ return crc;
68
/**
69
* Iterate args to the next hole
70
*
71
@@ -XXX,XX +XXX,XX @@ const DMAMap *iova_tree_find(const IOVATree *tree, const DMAMap *map)
72
return g_tree_lookup(tree->tree, map);
73
}
74
75
+static gboolean iova_tree_find_address_iterator(gpointer key, gpointer value,
76
+ gpointer data)
77
+{
78
+ const DMAMap *map = key;
79
+ IOVATreeFindIOVAArgs *args = data;
80
+ const DMAMap *needle;
81
+
82
+ g_assert(key == value);
83
+
84
+ needle = args->needle;
85
+ if (map->translated_addr + map->size < needle->translated_addr ||
86
+ needle->translated_addr + needle->size < map->translated_addr) {
87
+ return false;
88
+ }
89
+
90
+ args->result = map;
91
+ return true;
91
+}
92
+}
92
+
93
+
93
+unsigned compute_mcast_idx(const uint8_t *ep)
94
+const DMAMap *iova_tree_find_iova(const IOVATree *tree, const DMAMap *map)
94
+{
95
+{
95
+ return net_crc32(ep, ETH_ALEN) >> 26;
96
+ IOVATreeFindIOVAArgs args = {
96
}
97
+ .needle = map,
97
98
+ };
98
QemuOptsList qemu_netdev_opts = {
99
+
100
+ g_tree_foreach(tree->tree, iova_tree_find_address_iterator, &args);
101
+ return args.result;
102
+}
103
+
104
const DMAMap *iova_tree_find_address(const IOVATree *tree, hwaddr iova)
105
{
106
const DMAMap map = { .iova = iova, .size = 0 };
99
--
107
--
100
2.7.4
108
2.7.4
101
109
102
110
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
This makes it much easier to compare the multicast CRC calculation endian and
3
This tree is able to look for a translated address from an IOVA address.
4
bitshift against the Linux driver implementation.
5
4
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
5
At first glance it is similar to util/iova-tree. However, SVQ working on
6
devices with limited IOVA space need more capabilities, like allocating
7
IOVA chunks or performing reverse translations (qemu addresses to iova).
8
9
The allocation capability, as "assign a free IOVA address to this chunk
10
of memory in qemu's address space" allows shadow virtqueue to create a
11
new address space that is not restricted by guest's addressable one, so
12
we can allocate shadow vqs vrings outside of it.
13
14
It duplicates the tree so it can search efficiently in both directions,
15
and it will signal overlap if iova or the translated address is present
16
in any tree.
17
18
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
19
Acked-by: Michael S. Tsirkin <mst@redhat.com>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
20
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
21
---
9
hw/net/ftgmac100.c | 2 +-
22
hw/virtio/meson.build | 2 +-
10
1 file changed, 1 insertion(+), 1 deletion(-)
23
hw/virtio/vhost-iova-tree.c | 110 ++++++++++++++++++++++++++++++++++++++++++++
24
hw/virtio/vhost-iova-tree.h | 27 +++++++++++
25
3 files changed, 138 insertions(+), 1 deletion(-)
26
create mode 100644 hw/virtio/vhost-iova-tree.c
27
create mode 100644 hw/virtio/vhost-iova-tree.h
11
28
12
diff --git a/hw/net/ftgmac100.c b/hw/net/ftgmac100.c
29
diff --git a/hw/virtio/meson.build b/hw/virtio/meson.build
13
index XXXXXXX..XXXXXXX 100644
30
index XXXXXXX..XXXXXXX 100644
14
--- a/hw/net/ftgmac100.c
31
--- a/hw/virtio/meson.build
15
+++ b/hw/net/ftgmac100.c
32
+++ b/hw/virtio/meson.build
16
@@ -XXX,XX +XXX,XX @@ static int ftgmac100_filter(FTGMAC100State *s, const uint8_t *buf, size_t len)
33
@@ -XXX,XX +XXX,XX @@ softmmu_ss.add(when: 'CONFIG_ALL', if_true: files('vhost-stub.c'))
17
}
34
18
35
virtio_ss = ss.source_set()
19
/* TODO: this does not seem to work for ftgmac100 */
36
virtio_ss.add(files('virtio.c'))
20
- mcast_idx = compute_mcast_idx(buf);
37
-virtio_ss.add(when: 'CONFIG_VHOST', if_true: files('vhost.c', 'vhost-backend.c', 'vhost-shadow-virtqueue.c'))
21
+ mcast_idx = net_crc32(buf, ETH_ALEN) >> 26;
38
+virtio_ss.add(when: 'CONFIG_VHOST', if_true: files('vhost.c', 'vhost-backend.c', 'vhost-shadow-virtqueue.c', 'vhost-iova-tree.c'))
22
if (!(s->math[mcast_idx / 32] & (1 << (mcast_idx % 32)))) {
39
virtio_ss.add(when: 'CONFIG_VHOST_USER', if_true: files('vhost-user.c'))
23
return 0;
40
virtio_ss.add(when: 'CONFIG_VHOST_VDPA', if_true: files('vhost-vdpa.c'))
24
}
41
virtio_ss.add(when: 'CONFIG_VIRTIO_BALLOON', if_true: files('virtio-balloon.c'))
42
diff --git a/hw/virtio/vhost-iova-tree.c b/hw/virtio/vhost-iova-tree.c
43
new file mode 100644
44
index XXXXXXX..XXXXXXX
45
--- /dev/null
46
+++ b/hw/virtio/vhost-iova-tree.c
47
@@ -XXX,XX +XXX,XX @@
48
+/*
49
+ * vhost software live migration iova tree
50
+ *
51
+ * SPDX-FileCopyrightText: Red Hat, Inc. 2021
52
+ * SPDX-FileContributor: Author: Eugenio Pérez <eperezma@redhat.com>
53
+ *
54
+ * SPDX-License-Identifier: GPL-2.0-or-later
55
+ */
56
+
57
+#include "qemu/osdep.h"
58
+#include "qemu/iova-tree.h"
59
+#include "vhost-iova-tree.h"
60
+
61
+#define iova_min_addr qemu_real_host_page_size
62
+
63
+/**
64
+ * VhostIOVATree, able to:
65
+ * - Translate iova address
66
+ * - Reverse translate iova address (from translated to iova)
67
+ * - Allocate IOVA regions for translated range (linear operation)
68
+ */
69
+struct VhostIOVATree {
70
+ /* First addressable iova address in the device */
71
+ uint64_t iova_first;
72
+
73
+ /* Last addressable iova address in the device */
74
+ uint64_t iova_last;
75
+
76
+ /* IOVA address to qemu memory maps. */
77
+ IOVATree *iova_taddr_map;
78
+};
79
+
80
+/**
81
+ * Create a new IOVA tree
82
+ *
83
+ * Returns the new IOVA tree
84
+ */
85
+VhostIOVATree *vhost_iova_tree_new(hwaddr iova_first, hwaddr iova_last)
86
+{
87
+ VhostIOVATree *tree = g_new(VhostIOVATree, 1);
88
+
89
+ /* Some devices do not like 0 addresses */
90
+ tree->iova_first = MAX(iova_first, iova_min_addr);
91
+ tree->iova_last = iova_last;
92
+
93
+ tree->iova_taddr_map = iova_tree_new();
94
+ return tree;
95
+}
96
+
97
+/**
98
+ * Delete an iova tree
99
+ */
100
+void vhost_iova_tree_delete(VhostIOVATree *iova_tree)
101
+{
102
+ iova_tree_destroy(iova_tree->iova_taddr_map);
103
+ g_free(iova_tree);
104
+}
105
+
106
+/**
107
+ * Find the IOVA address stored from a memory address
108
+ *
109
+ * @tree: The iova tree
110
+ * @map: The map with the memory address
111
+ *
112
+ * Return the stored mapping, or NULL if not found.
113
+ */
114
+const DMAMap *vhost_iova_tree_find_iova(const VhostIOVATree *tree,
115
+ const DMAMap *map)
116
+{
117
+ return iova_tree_find_iova(tree->iova_taddr_map, map);
118
+}
119
+
120
+/**
121
+ * Allocate a new mapping
122
+ *
123
+ * @tree: The iova tree
124
+ * @map: The iova map
125
+ *
126
+ * Returns:
127
+ * - IOVA_OK if the map fits in the container
128
+ * - IOVA_ERR_INVALID if the map does not make sense (like size overflow)
129
+ * - IOVA_ERR_NOMEM if tree cannot allocate more space.
130
+ *
131
+ * It returns assignated iova in map->iova if return value is VHOST_DMA_MAP_OK.
132
+ */
133
+int vhost_iova_tree_map_alloc(VhostIOVATree *tree, DMAMap *map)
134
+{
135
+ /* Some vhost devices do not like addr 0. Skip first page */
136
+ hwaddr iova_first = tree->iova_first ?: qemu_real_host_page_size;
137
+
138
+ if (map->translated_addr + map->size < map->translated_addr ||
139
+ map->perm == IOMMU_NONE) {
140
+ return IOVA_ERR_INVALID;
141
+ }
142
+
143
+ /* Allocate a node in IOVA address */
144
+ return iova_tree_alloc_map(tree->iova_taddr_map, map, iova_first,
145
+ tree->iova_last);
146
+}
147
+
148
+/**
149
+ * Remove existing mappings from iova tree
150
+ *
151
+ * @iova_tree: The vhost iova tree
152
+ * @map: The map to remove
153
+ */
154
+void vhost_iova_tree_remove(VhostIOVATree *iova_tree, const DMAMap *map)
155
+{
156
+ iova_tree_remove(iova_tree->iova_taddr_map, map);
157
+}
158
diff --git a/hw/virtio/vhost-iova-tree.h b/hw/virtio/vhost-iova-tree.h
159
new file mode 100644
160
index XXXXXXX..XXXXXXX
161
--- /dev/null
162
+++ b/hw/virtio/vhost-iova-tree.h
163
@@ -XXX,XX +XXX,XX @@
164
+/*
165
+ * vhost software live migration iova tree
166
+ *
167
+ * SPDX-FileCopyrightText: Red Hat, Inc. 2021
168
+ * SPDX-FileContributor: Author: Eugenio Pérez <eperezma@redhat.com>
169
+ *
170
+ * SPDX-License-Identifier: GPL-2.0-or-later
171
+ */
172
+
173
+#ifndef HW_VIRTIO_VHOST_IOVA_TREE_H
174
+#define HW_VIRTIO_VHOST_IOVA_TREE_H
175
+
176
+#include "qemu/iova-tree.h"
177
+#include "exec/memory.h"
178
+
179
+typedef struct VhostIOVATree VhostIOVATree;
180
+
181
+VhostIOVATree *vhost_iova_tree_new(uint64_t iova_first, uint64_t iova_last);
182
+void vhost_iova_tree_delete(VhostIOVATree *iova_tree);
183
+G_DEFINE_AUTOPTR_CLEANUP_FUNC(VhostIOVATree, vhost_iova_tree_delete);
184
+
185
+const DMAMap *vhost_iova_tree_find_iova(const VhostIOVATree *iova_tree,
186
+ const DMAMap *map);
187
+int vhost_iova_tree_map_alloc(VhostIOVATree *iova_tree, DMAMap *map);
188
+void vhost_iova_tree_remove(VhostIOVATree *iova_tree, const DMAMap *map);
189
+
190
+#endif
25
--
191
--
26
2.7.4
192
2.7.4
27
193
28
194
diff view generated by jsdifflib
1
From: Ed Swierk via Qemu-devel <qemu-devel@nongnu.org>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
The device is supposed to maintain two distinct contexts for transmit
3
Use translations added in VhostIOVATree in SVQ.
4
offloads: one has parameters for both segmentation and checksum
4
5
offload, the other only for checksum offload. The guest driver can
5
Only introduce usage here, not allocation and deallocation. As with
6
send two context descriptors, one for each context (the TSE flag
6
previous patches, we use the dead code paths of shadow_vqs_enabled to
7
specifies which). Then the guest can refer to one or the other context
7
avoid commiting too many changes at once. These are impossible to take
8
in subsequent transmit data descriptors, depending on what offloads it
8
at the moment.
9
wants applied to each packet.
9
10
10
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
11
Currently the e1000 device stores just one context, and misinterprets
11
Acked-by: Michael S. Tsirkin <mst@redhat.com>
12
the TSE flags in the context and data descriptors. This is often okay:
13
Linux happens to send a fresh context descriptor before every data
14
descriptor, so forgetting the other context doesn't matter. Windows
15
does rely on separate contexts for TSO vs. non-TSO packets, but for
16
mostly-TCP traffic the two contexts have identical TCP-specific
17
offload parameters so confusing them doesn't matter.
18
19
One case where this confusion matters is when a Windows guest sets up
20
a TSO context for TCP and a non-TSO context for UDP, and then
21
transmits both TCP and UDP traffic in parallel. The e1000 device
22
sometimes ends up using TCP-specific parameters while doing checksum
23
offload on a UDP datagram: it writes the checksum to offset 16 (the
24
correct location for a TCP checksum), stomping on two bytes of UDP
25
data, and leaving the wrong value in the actual UDP checksum field at
26
offset 6. (Even worse, the host network stack may then recompute the
27
UDP checksum, "correcting" it to match the corrupt data before sending
28
it out a physical interface.)
29
30
Correct this by tracking the TSO context independently of the non-TSO
31
context, and selecting the appropriate context based on the TSE flag
32
in each transmit data descriptor.
33
34
Signed-off-by: Ed Swierk <eswierk@skyportsystems.com>
35
Signed-off-by: Jason Wang <jasowang@redhat.com>
12
Signed-off-by: Jason Wang <jasowang@redhat.com>
36
---
13
---
37
hw/net/e1000.c | 70 +++++++++++++++++++++++++++++++++-------------------------
14
hw/virtio/vhost-shadow-virtqueue.c | 86 +++++++++++++++++++++++---
38
1 file changed, 40 insertions(+), 30 deletions(-)
15
hw/virtio/vhost-shadow-virtqueue.h | 6 +-
39
16
hw/virtio/vhost-vdpa.c | 122 +++++++++++++++++++++++++++++++------
40
diff --git a/hw/net/e1000.c b/hw/net/e1000.c
17
include/hw/virtio/vhost-vdpa.h | 3 +
18
4 files changed, 187 insertions(+), 30 deletions(-)
19
20
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
41
index XXXXXXX..XXXXXXX 100644
21
index XXXXXXX..XXXXXXX 100644
42
--- a/hw/net/e1000.c
22
--- a/hw/virtio/vhost-shadow-virtqueue.c
43
+++ b/hw/net/e1000.c
23
+++ b/hw/virtio/vhost-shadow-virtqueue.c
44
@@ -XXX,XX +XXX,XX @@ typedef struct E1000State_st {
24
@@ -XXX,XX +XXX,XX @@ static uint16_t vhost_svq_available_slots(const VhostShadowVirtqueue *svq)
45
unsigned char sum_needed;
25
return svq->vring.num - (svq->shadow_avail_idx - svq->shadow_used_idx);
46
bool cptse;
26
}
47
e1000x_txd_props props;
27
48
+ e1000x_txd_props tso_props;
28
-static void vhost_vring_write_descs(VhostShadowVirtqueue *svq,
49
uint16_t tso_frames;
29
+/**
50
} tx;
30
+ * Translate addresses between the qemu's virtual address and the SVQ IOVA
51
31
+ *
52
@@ -XXX,XX +XXX,XX @@ xmit_seg(E1000State *s)
32
+ * @svq: Shadow VirtQueue
53
uint16_t len;
33
+ * @vaddr: Translated IOVA addresses
54
unsigned int frames = s->tx.tso_frames, css, sofar;
34
+ * @iovec: Source qemu's VA addresses
55
struct e1000_tx *tp = &s->tx;
35
+ * @num: Length of iovec and minimum length of vaddr
56
+ struct e1000x_txd_props *props = tp->cptse ? &tp->tso_props : &tp->props;
36
+ */
57
37
+static bool vhost_svq_translate_addr(const VhostShadowVirtqueue *svq,
58
- if (tp->props.tse && tp->cptse) {
38
+ hwaddr *addrs, const struct iovec *iovec,
59
- css = tp->props.ipcss;
39
+ size_t num)
60
+ if (tp->cptse) {
40
+{
61
+ css = props->ipcss;
41
+ if (num == 0) {
62
DBGOUT(TXSUM, "frames %d size %d ipcss %d\n",
42
+ return true;
63
frames, tp->size, css);
43
+ }
64
- if (tp->props.ip) { /* IPv4 */
44
+
65
+ if (props->ip) { /* IPv4 */
45
+ for (size_t i = 0; i < num; ++i) {
66
stw_be_p(tp->data+css+2, tp->size - css);
46
+ DMAMap needle = {
67
stw_be_p(tp->data+css+4,
47
+ .translated_addr = (hwaddr)(uintptr_t)iovec[i].iov_base,
68
lduw_be_p(tp->data + css + 4) + frames);
48
+ .size = iovec[i].iov_len,
69
} else { /* IPv6 */
49
+ };
70
stw_be_p(tp->data+css+4, tp->size - css);
50
+ Int128 needle_last, map_last;
51
+ size_t off;
52
+
53
+ const DMAMap *map = vhost_iova_tree_find_iova(svq->iova_tree, &needle);
54
+ /*
55
+ * Map cannot be NULL since iova map contains all guest space and
56
+ * qemu already has a physical address mapped
57
+ */
58
+ if (unlikely(!map)) {
59
+ qemu_log_mask(LOG_GUEST_ERROR,
60
+ "Invalid address 0x%"HWADDR_PRIx" given by guest",
61
+ needle.translated_addr);
62
+ return false;
63
+ }
64
+
65
+ off = needle.translated_addr - map->translated_addr;
66
+ addrs[i] = map->iova + off;
67
+
68
+ needle_last = int128_add(int128_make64(needle.translated_addr),
69
+ int128_make64(iovec[i].iov_len));
70
+ map_last = int128_make64(map->translated_addr + map->size);
71
+ if (unlikely(int128_gt(needle_last, map_last))) {
72
+ qemu_log_mask(LOG_GUEST_ERROR,
73
+ "Guest buffer expands over iova range");
74
+ return false;
75
+ }
76
+ }
77
+
78
+ return true;
79
+}
80
+
81
+static void vhost_vring_write_descs(VhostShadowVirtqueue *svq, hwaddr *sg,
82
const struct iovec *iovec, size_t num,
83
bool more_descs, bool write)
84
{
85
@@ -XXX,XX +XXX,XX @@ static void vhost_vring_write_descs(VhostShadowVirtqueue *svq,
86
} else {
87
descs[i].flags = flags;
71
}
88
}
72
- css = tp->props.tucss;
89
- descs[i].addr = cpu_to_le64((hwaddr)(intptr_t)iovec[n].iov_base);
73
+ css = props->tucss;
90
+ descs[i].addr = cpu_to_le64(sg[n]);
74
len = tp->size - css;
91
descs[i].len = cpu_to_le32(iovec[n].iov_len);
75
- DBGOUT(TXSUM, "tcp %d tucss %d len %d\n", tp->props.tcp, css, len);
92
76
- if (tp->props.tcp) {
93
last = i;
77
- sofar = frames * tp->props.mss;
94
@@ -XXX,XX +XXX,XX @@ static bool vhost_svq_add_split(VhostShadowVirtqueue *svq,
78
+ DBGOUT(TXSUM, "tcp %d tucss %d len %d\n", props->tcp, css, len);
95
{
79
+ if (props->tcp) {
96
unsigned avail_idx;
80
+ sofar = frames * props->mss;
97
vring_avail_t *avail = svq->vring.avail;
81
stl_be_p(tp->data+css+4, ldl_be_p(tp->data+css+4)+sofar); /* seq */
98
+ bool ok;
82
- if (tp->props.paylen - sofar > tp->props.mss) {
99
+ g_autofree hwaddr *sgs = g_new(hwaddr, MAX(elem->out_num, elem->in_num));
83
+ if (props->paylen - sofar > props->mss) {
100
84
tp->data[css + 13] &= ~9; /* PSH, FIN */
101
*head = svq->free_head;
85
} else if (frames) {
102
86
e1000x_inc_reg_if_not_full(s->mac_reg, TSCTC);
103
@@ -XXX,XX +XXX,XX @@ static bool vhost_svq_add_split(VhostShadowVirtqueue *svq,
87
}
104
return false;
88
- } else /* UDP */
105
}
89
+ } else { /* UDP */
106
90
stw_be_p(tp->data+css+4, len);
107
- vhost_vring_write_descs(svq, elem->out_sg, elem->out_num, elem->in_num > 0,
108
- false);
109
- vhost_vring_write_descs(svq, elem->in_sg, elem->in_num, false, true);
110
+ ok = vhost_svq_translate_addr(svq, sgs, elem->out_sg, elem->out_num);
111
+ if (unlikely(!ok)) {
112
+ return false;
113
+ }
114
+ vhost_vring_write_descs(svq, sgs, elem->out_sg, elem->out_num,
115
+ elem->in_num > 0, false);
116
+
117
+
118
+ ok = vhost_svq_translate_addr(svq, sgs, elem->in_sg, elem->in_num);
119
+ if (unlikely(!ok)) {
120
+ return false;
121
+ }
122
+
123
+ vhost_vring_write_descs(svq, sgs, elem->in_sg, elem->in_num, false, true);
124
125
/*
126
* Put the entry in the available array (but don't update avail->idx until
127
@@ -XXX,XX +XXX,XX @@ void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd)
128
void vhost_svq_get_vring_addr(const VhostShadowVirtqueue *svq,
129
struct vhost_vring_addr *addr)
130
{
131
- addr->desc_user_addr = (uint64_t)(intptr_t)svq->vring.desc;
132
- addr->avail_user_addr = (uint64_t)(intptr_t)svq->vring.avail;
133
- addr->used_user_addr = (uint64_t)(intptr_t)svq->vring.used;
134
+ addr->desc_user_addr = (uint64_t)(uintptr_t)svq->vring.desc;
135
+ addr->avail_user_addr = (uint64_t)(uintptr_t)svq->vring.avail;
136
+ addr->used_user_addr = (uint64_t)(uintptr_t)svq->vring.used;
137
}
138
139
size_t vhost_svq_driver_area_size(const VhostShadowVirtqueue *svq)
140
@@ -XXX,XX +XXX,XX @@ void vhost_svq_stop(VhostShadowVirtqueue *svq)
141
* Creates vhost shadow virtqueue, and instructs the vhost device to use the
142
* shadow methods and file descriptors.
143
*
144
+ * @iova_tree: Tree to perform descriptors translations
145
+ *
146
* Returns the new virtqueue or NULL.
147
*
148
* In case of error, reason is reported through error_report.
149
*/
150
-VhostShadowVirtqueue *vhost_svq_new(void)
151
+VhostShadowVirtqueue *vhost_svq_new(VhostIOVATree *iova_tree)
152
{
153
g_autofree VhostShadowVirtqueue *svq = g_new0(VhostShadowVirtqueue, 1);
154
int r;
155
@@ -XXX,XX +XXX,XX @@ VhostShadowVirtqueue *vhost_svq_new(void)
156
157
event_notifier_init_fd(&svq->svq_kick, VHOST_FILE_UNBIND);
158
event_notifier_set_handler(&svq->hdev_call, vhost_svq_handle_call);
159
+ svq->iova_tree = iova_tree;
160
return g_steal_pointer(&svq);
161
162
err_init_hdev_call:
163
diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
164
index XXXXXXX..XXXXXXX 100644
165
--- a/hw/virtio/vhost-shadow-virtqueue.h
166
+++ b/hw/virtio/vhost-shadow-virtqueue.h
167
@@ -XXX,XX +XXX,XX @@
168
#include "qemu/event_notifier.h"
169
#include "hw/virtio/virtio.h"
170
#include "standard-headers/linux/vhost_types.h"
171
+#include "hw/virtio/vhost-iova-tree.h"
172
173
/* Shadow virtqueue to relay notifications */
174
typedef struct VhostShadowVirtqueue {
175
@@ -XXX,XX +XXX,XX @@ typedef struct VhostShadowVirtqueue {
176
/* Virtio device */
177
VirtIODevice *vdev;
178
179
+ /* IOVA mapping */
180
+ VhostIOVATree *iova_tree;
181
+
182
/* Map for use the guest's descriptors */
183
VirtQueueElement **ring_id_maps;
184
185
@@ -XXX,XX +XXX,XX @@ void vhost_svq_start(VhostShadowVirtqueue *svq, VirtIODevice *vdev,
186
VirtQueue *vq);
187
void vhost_svq_stop(VhostShadowVirtqueue *svq);
188
189
-VhostShadowVirtqueue *vhost_svq_new(void);
190
+VhostShadowVirtqueue *vhost_svq_new(VhostIOVATree *iova_tree);
191
192
void vhost_svq_free(gpointer vq);
193
G_DEFINE_AUTOPTR_CLEANUP_FUNC(VhostShadowVirtqueue, vhost_svq_free);
194
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
195
index XXXXXXX..XXXXXXX 100644
196
--- a/hw/virtio/vhost-vdpa.c
197
+++ b/hw/virtio/vhost-vdpa.c
198
@@ -XXX,XX +XXX,XX @@ static void vhost_vdpa_listener_region_add(MemoryListener *listener,
199
vaddr, section->readonly);
200
201
llsize = int128_sub(llend, int128_make64(iova));
202
+ if (v->shadow_vqs_enabled) {
203
+ DMAMap mem_region = {
204
+ .translated_addr = (hwaddr)(uintptr_t)vaddr,
205
+ .size = int128_get64(llsize) - 1,
206
+ .perm = IOMMU_ACCESS_FLAG(true, section->readonly),
207
+ };
208
+
209
+ int r = vhost_iova_tree_map_alloc(v->iova_tree, &mem_region);
210
+ if (unlikely(r != IOVA_OK)) {
211
+ error_report("Can't allocate a mapping (%d)", r);
212
+ goto fail;
91
+ }
213
+ }
92
if (tp->sum_needed & E1000_TXD_POPTS_TXSM) {
214
+
93
unsigned int phsum;
215
+ iova = mem_region.iova;
94
// add pseudo-header length before checksum calculation
216
+ }
95
- void *sp = tp->data + tp->props.tucso;
217
96
+ void *sp = tp->data + props->tucso;
218
vhost_vdpa_iotlb_batch_begin_once(v);
97
219
ret = vhost_vdpa_dma_map(v, iova, int128_get64(llsize),
98
phsum = lduw_be_p(sp) + len;
220
@@ -XXX,XX +XXX,XX @@ static void vhost_vdpa_listener_region_del(MemoryListener *listener,
99
phsum = (phsum >> 16) + (phsum & 0xffff);
221
100
@@ -XXX,XX +XXX,XX @@ xmit_seg(E1000State *s)
222
llsize = int128_sub(llend, int128_make64(iova));
223
224
+ if (v->shadow_vqs_enabled) {
225
+ const DMAMap *result;
226
+ const void *vaddr = memory_region_get_ram_ptr(section->mr) +
227
+ section->offset_within_region +
228
+ (iova - section->offset_within_address_space);
229
+ DMAMap mem_region = {
230
+ .translated_addr = (hwaddr)(uintptr_t)vaddr,
231
+ .size = int128_get64(llsize) - 1,
232
+ };
233
+
234
+ result = vhost_iova_tree_find_iova(v->iova_tree, &mem_region);
235
+ iova = result->iova;
236
+ vhost_iova_tree_remove(v->iova_tree, &mem_region);
237
+ }
238
vhost_vdpa_iotlb_batch_begin_once(v);
239
ret = vhost_vdpa_dma_unmap(v, iova, int128_get64(llsize));
240
if (ret) {
241
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v,
242
243
shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free);
244
for (unsigned n = 0; n < hdev->nvqs; ++n) {
245
- g_autoptr(VhostShadowVirtqueue) svq = vhost_svq_new();
246
+ g_autoptr(VhostShadowVirtqueue) svq = vhost_svq_new(v->iova_tree);
247
248
if (unlikely(!svq)) {
249
error_setg(errp, "Cannot create svq %u", n);
250
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_svq_set_fds(struct vhost_dev *dev,
251
/**
252
* Unmap a SVQ area in the device
253
*/
254
-static bool vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v, hwaddr iova,
255
- hwaddr size)
256
+static bool vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v,
257
+ const DMAMap *needle)
258
{
259
+ const DMAMap *result = vhost_iova_tree_find_iova(v->iova_tree, needle);
260
+ hwaddr size;
261
int r;
262
263
- size = ROUND_UP(size, qemu_real_host_page_size);
264
- r = vhost_vdpa_dma_unmap(v, iova, size);
265
+ if (unlikely(!result)) {
266
+ error_report("Unable to find SVQ address to unmap");
267
+ return false;
268
+ }
269
+
270
+ size = ROUND_UP(result->size, qemu_real_host_page_size);
271
+ r = vhost_vdpa_dma_unmap(v, result->iova, size);
272
return r == 0;
273
}
274
275
static bool vhost_vdpa_svq_unmap_rings(struct vhost_dev *dev,
276
const VhostShadowVirtqueue *svq)
277
{
278
+ DMAMap needle = {};
279
struct vhost_vdpa *v = dev->opaque;
280
struct vhost_vring_addr svq_addr;
281
- size_t device_size = vhost_svq_device_area_size(svq);
282
- size_t driver_size = vhost_svq_driver_area_size(svq);
283
bool ok;
284
285
vhost_svq_get_vring_addr(svq, &svq_addr);
286
287
- ok = vhost_vdpa_svq_unmap_ring(v, svq_addr.desc_user_addr, driver_size);
288
+ needle.translated_addr = svq_addr.desc_user_addr;
289
+ ok = vhost_vdpa_svq_unmap_ring(v, &needle);
290
if (unlikely(!ok)) {
291
return false;
101
}
292
}
102
293
103
if (tp->sum_needed & E1000_TXD_POPTS_TXSM) {
294
- return vhost_vdpa_svq_unmap_ring(v, svq_addr.used_user_addr, device_size);
104
- putsum(tp->data, tp->size, tp->props.tucso,
295
+ needle.translated_addr = svq_addr.used_user_addr;
105
- tp->props.tucss, tp->props.tucse);
296
+ return vhost_vdpa_svq_unmap_ring(v, &needle);
106
+ putsum(tp->data, tp->size, props->tucso, props->tucss, props->tucse);
297
+}
298
+
299
+/**
300
+ * Map the SVQ area in the device
301
+ *
302
+ * @v: Vhost-vdpa device
303
+ * @needle: The area to search iova
304
+ * @errorp: Error pointer
305
+ */
306
+static bool vhost_vdpa_svq_map_ring(struct vhost_vdpa *v, DMAMap *needle,
307
+ Error **errp)
308
+{
309
+ int r;
310
+
311
+ r = vhost_iova_tree_map_alloc(v->iova_tree, needle);
312
+ if (unlikely(r != IOVA_OK)) {
313
+ error_setg(errp, "Cannot allocate iova (%d)", r);
314
+ return false;
315
+ }
316
+
317
+ r = vhost_vdpa_dma_map(v, needle->iova, needle->size + 1,
318
+ (void *)(uintptr_t)needle->translated_addr,
319
+ needle->perm == IOMMU_RO);
320
+ if (unlikely(r != 0)) {
321
+ error_setg_errno(errp, -r, "Cannot map region to device");
322
+ vhost_iova_tree_remove(v->iova_tree, needle);
323
+ }
324
+
325
+ return r == 0;
326
}
327
328
/**
329
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_svq_map_rings(struct vhost_dev *dev,
330
struct vhost_vring_addr *addr,
331
Error **errp)
332
{
333
+ DMAMap device_region, driver_region;
334
+ struct vhost_vring_addr svq_addr;
335
struct vhost_vdpa *v = dev->opaque;
336
size_t device_size = vhost_svq_device_area_size(svq);
337
size_t driver_size = vhost_svq_driver_area_size(svq);
338
- int r;
339
+ size_t avail_offset;
340
+ bool ok;
341
342
ERRP_GUARD();
343
- vhost_svq_get_vring_addr(svq, addr);
344
+ vhost_svq_get_vring_addr(svq, &svq_addr);
345
346
- r = vhost_vdpa_dma_map(v, addr->desc_user_addr, driver_size,
347
- (void *)(uintptr_t)addr->desc_user_addr, true);
348
- if (unlikely(r != 0)) {
349
- error_setg_errno(errp, -r, "Cannot create vq driver region: ");
350
+ driver_region = (DMAMap) {
351
+ .translated_addr = svq_addr.desc_user_addr,
352
+ .size = driver_size - 1,
353
+ .perm = IOMMU_RO,
354
+ };
355
+ ok = vhost_vdpa_svq_map_ring(v, &driver_region, errp);
356
+ if (unlikely(!ok)) {
357
+ error_prepend(errp, "Cannot create vq driver region: ");
358
return false;
107
}
359
}
108
if (tp->sum_needed & E1000_TXD_POPTS_IXSM) {
360
+ addr->desc_user_addr = driver_region.iova;
109
- putsum(tp->data, tp->size, tp->props.ipcso,
361
+ avail_offset = svq_addr.avail_user_addr - svq_addr.desc_user_addr;
110
- tp->props.ipcss, tp->props.ipcse);
362
+ addr->avail_user_addr = driver_region.iova + avail_offset;
111
+ putsum(tp->data, tp->size, props->ipcso, props->ipcss, props->ipcse);
363
364
- r = vhost_vdpa_dma_map(v, addr->used_user_addr, device_size,
365
- (void *)(intptr_t)addr->used_user_addr, false);
366
- if (unlikely(r != 0)) {
367
- error_setg_errno(errp, -r, "Cannot create vq device region: ");
368
+ device_region = (DMAMap) {
369
+ .translated_addr = svq_addr.used_user_addr,
370
+ .size = device_size - 1,
371
+ .perm = IOMMU_RW,
372
+ };
373
+ ok = vhost_vdpa_svq_map_ring(v, &device_region, errp);
374
+ if (unlikely(!ok)) {
375
+ error_prepend(errp, "Cannot create vq device region: ");
376
+ vhost_vdpa_svq_unmap_ring(v, &driver_region);
112
}
377
}
113
if (tp->vlan_needed) {
378
+ addr->used_user_addr = device_region.iova;
114
memmove(tp->vlan, tp->data, 4);
379
115
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
380
- return r == 0;
116
381
+ return ok;
117
s->mit_ide |= (txd_lower & E1000_TXD_CMD_IDE);
382
}
118
if (dtype == E1000_TXD_CMD_DEXT) { /* context descriptor */
383
119
- e1000x_read_tx_ctx_descr(xp, &tp->props);
384
static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
120
- tp->tso_frames = 0;
385
diff --git a/include/hw/virtio/vhost-vdpa.h b/include/hw/virtio/vhost-vdpa.h
121
- if (tp->props.tucso == 0) { /* this is probably wrong */
386
index XXXXXXX..XXXXXXX 100644
122
- DBGOUT(TXSUM, "TCP/UDP: cso 0!\n");
387
--- a/include/hw/virtio/vhost-vdpa.h
123
- tp->props.tucso = tp->props.tucss + (tp->props.tcp ? 16 : 6);
388
+++ b/include/hw/virtio/vhost-vdpa.h
124
+ if (le32_to_cpu(xp->cmd_and_length) & E1000_TXD_CMD_TSE) {
389
@@ -XXX,XX +XXX,XX @@
125
+ e1000x_read_tx_ctx_descr(xp, &tp->tso_props);
390
126
+ tp->tso_frames = 0;
391
#include <gmodule.h>
127
+ } else {
392
128
+ e1000x_read_tx_ctx_descr(xp, &tp->props);
393
+#include "hw/virtio/vhost-iova-tree.h"
129
}
394
#include "hw/virtio/virtio.h"
130
return;
395
#include "standard-headers/linux/vhost_types.h"
131
} else if (dtype == (E1000_TXD_CMD_DEXT | E1000_TXD_DTYP_D)) {
396
132
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
397
@@ -XXX,XX +XXX,XX @@ typedef struct vhost_vdpa {
133
}
398
MemoryListener listener;
134
399
struct vhost_vdpa_iova_range iova_range;
135
addr = le64_to_cpu(dp->buffer_addr);
400
bool shadow_vqs_enabled;
136
- if (tp->props.tse && tp->cptse) {
401
+ /* IOVA mapping used by the Shadow Virtqueue */
137
- msh = tp->props.hdr_len + tp->props.mss;
402
+ VhostIOVATree *iova_tree;
138
+ if (tp->cptse) {
403
GPtrArray *shadow_vqs;
139
+ msh = tp->tso_props.hdr_len + tp->tso_props.mss;
404
struct vhost_dev *dev;
140
do {
405
VhostVDPAHostNotifier notifier[VIRTIO_QUEUE_MAX];
141
bytes = split_size;
142
if (tp->size + bytes > msh)
143
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
144
bytes = MIN(sizeof(tp->data) - tp->size, bytes);
145
pci_dma_read(d, addr, tp->data + tp->size, bytes);
146
sz = tp->size + bytes;
147
- if (sz >= tp->props.hdr_len && tp->size < tp->props.hdr_len) {
148
- memmove(tp->header, tp->data, tp->props.hdr_len);
149
+ if (sz >= tp->tso_props.hdr_len
150
+ && tp->size < tp->tso_props.hdr_len) {
151
+ memmove(tp->header, tp->data, tp->tso_props.hdr_len);
152
}
153
tp->size = sz;
154
addr += bytes;
155
if (sz == msh) {
156
xmit_seg(s);
157
- memmove(tp->data, tp->header, tp->props.hdr_len);
158
- tp->size = tp->props.hdr_len;
159
+ memmove(tp->data, tp->header, tp->tso_props.hdr_len);
160
+ tp->size = tp->tso_props.hdr_len;
161
}
162
split_size -= bytes;
163
} while (bytes && split_size);
164
- } else if (!tp->props.tse && tp->cptse) {
165
- // context descriptor TSE is not set, while data descriptor TSE is set
166
- DBGOUT(TXERR, "TCP segmentation error\n");
167
} else {
168
split_size = MIN(sizeof(tp->data) - tp->size, split_size);
169
pci_dma_read(d, addr, tp->data + tp->size, split_size);
170
@@ -XXX,XX +XXX,XX @@ process_tx_desc(E1000State *s, struct e1000_tx_desc *dp)
171
172
if (!(txd_lower & E1000_TXD_CMD_EOP))
173
return;
174
- if (!(tp->props.tse && tp->cptse && tp->size < tp->props.hdr_len)) {
175
+ if (!(tp->cptse && tp->size < tp->tso_props.hdr_len)) {
176
xmit_seg(s);
177
}
178
tp->tso_frames = 0;
179
@@ -XXX,XX +XXX,XX @@ static const VMStateDescription vmstate_e1000_full_mac_state = {
180
181
static const VMStateDescription vmstate_e1000 = {
182
.name = "e1000",
183
- .version_id = 2,
184
+ .version_id = 3,
185
.minimum_version_id = 1,
186
.pre_save = e1000_pre_save,
187
.post_load = e1000_post_load,
188
@@ -XXX,XX +XXX,XX @@ static const VMStateDescription vmstate_e1000 = {
189
VMSTATE_UINT32_SUB_ARRAY(mac_reg, E1000State, RA, 32),
190
VMSTATE_UINT32_SUB_ARRAY(mac_reg, E1000State, MTA, 128),
191
VMSTATE_UINT32_SUB_ARRAY(mac_reg, E1000State, VFTA, 128),
192
+ VMSTATE_UINT8_V(tx.tso_props.ipcss, E1000State, 3),
193
+ VMSTATE_UINT8_V(tx.tso_props.ipcso, E1000State, 3),
194
+ VMSTATE_UINT16_V(tx.tso_props.ipcse, E1000State, 3),
195
+ VMSTATE_UINT8_V(tx.tso_props.tucss, E1000State, 3),
196
+ VMSTATE_UINT8_V(tx.tso_props.tucso, E1000State, 3),
197
+ VMSTATE_UINT16_V(tx.tso_props.tucse, E1000State, 3),
198
+ VMSTATE_UINT32_V(tx.tso_props.paylen, E1000State, 3),
199
+ VMSTATE_UINT8_V(tx.tso_props.hdr_len, E1000State, 3),
200
+ VMSTATE_UINT16_V(tx.tso_props.mss, E1000State, 3),
201
+ VMSTATE_INT8_V(tx.tso_props.ip, E1000State, 3),
202
+ VMSTATE_INT8_V(tx.tso_props.tcp, E1000State, 3),
203
VMSTATE_END_OF_LIST()
204
},
205
.subsections = (const VMStateDescription*[]) {
206
--
406
--
207
2.7.4
407
2.7.4
208
408
209
409
diff view generated by jsdifflib
Deleted patch
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
2
1
3
Instead of lnc_mchash() using its own implementation, we can simply call
4
net_crc32_le() directly and apply the bit shift inline.
5
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
7
Reviewed-by: Eric Blake <eblake@redhat.com>
8
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
9
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
---
11
hw/net/pcnet.c | 22 ++--------------------
12
1 file changed, 2 insertions(+), 20 deletions(-)
13
14
diff --git a/hw/net/pcnet.c b/hw/net/pcnet.c
15
index XXXXXXX..XXXXXXX 100644
16
--- a/hw/net/pcnet.c
17
+++ b/hw/net/pcnet.c
18
@@ -XXX,XX +XXX,XX @@
19
#include "qemu/osdep.h"
20
#include "hw/qdev.h"
21
#include "net/net.h"
22
+#include "net/eth.h"
23
#include "qemu/timer.h"
24
#include "qemu/sockets.h"
25
#include "sysemu/sysemu.h"
26
@@ -XXX,XX +XXX,XX @@ static inline void pcnet_rmd_store(PCNetState *s, struct pcnet_RMD *rmd,
27
be16_to_cpu(hdr->ether_type)); \
28
} while (0)
29
30
-#define MULTICAST_FILTER_LEN 8
31
-
32
-static inline uint32_t lnc_mchash(const uint8_t *ether_addr)
33
-{
34
-#define LNC_POLYNOMIAL 0xEDB88320UL
35
- uint32_t crc = 0xFFFFFFFF;
36
- int idx, bit;
37
- uint8_t data;
38
-
39
- for (idx = 0; idx < 6; idx++) {
40
- for (data = *ether_addr++, bit = 0; bit < MULTICAST_FILTER_LEN; bit++) {
41
- crc = (crc >> 1) ^ (((crc ^ data) & 1) ? LNC_POLYNOMIAL : 0);
42
- data >>= 1;
43
- }
44
- }
45
- return crc;
46
-#undef LNC_POLYNOMIAL
47
-}
48
-
49
#define CRC(crc, ch)     (crc = (crc >> 8) ^ crctab[(crc ^ (ch)) & 0xff])
50
51
/* generated using the AUTODIN II polynomial
52
@@ -XXX,XX +XXX,XX @@ static inline int ladr_match(PCNetState *s, const uint8_t *buf, int size)
53
s->csr[10] & 0xff, s->csr[10] >> 8,
54
s->csr[11] & 0xff, s->csr[11] >> 8
55
};
56
- int index = lnc_mchash(hdr->ether_dhost) >> 26;
57
+ int index = net_crc32_le(hdr->ether_dhost, ETH_ALEN) >> 26;
58
return !!(ladr[index >> 3] & (1 << (index & 7)));
59
}
60
return 0;
61
--
62
2.7.4
63
64
diff view generated by jsdifflib
Deleted patch
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
2
1
3
Instead of e100_compute_mcast_idx() using its own implementation, we can
4
simply call net_crc32() directly and apply the bit shift inline.
5
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
7
Reviewed-by: Stefan Weil <sw@weilnetz.de>
8
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
9
Signed-off-by: Jason Wang <jasowang@redhat.com>
10
---
11
hw/net/eepro100.c | 28 ++++------------------------
12
1 file changed, 4 insertions(+), 24 deletions(-)
13
14
diff --git a/hw/net/eepro100.c b/hw/net/eepro100.c
15
index XXXXXXX..XXXXXXX 100644
16
--- a/hw/net/eepro100.c
17
+++ b/hw/net/eepro100.c
18
@@ -XXX,XX +XXX,XX @@
19
#include "hw/hw.h"
20
#include "hw/pci/pci.h"
21
#include "net/net.h"
22
+#include "net/eth.h"
23
#include "hw/nvram/eeprom93xx.h"
24
#include "sysemu/sysemu.h"
25
#include "sysemu/dma.h"
26
@@ -XXX,XX +XXX,XX @@ static const uint16_t eepro100_mdi_mask[] = {
27
28
static E100PCIDeviceInfo *eepro100_get_class(EEPRO100State *s);
29
30
-/* From FreeBSD (locally modified). */
31
-static unsigned e100_compute_mcast_idx(const uint8_t *ep)
32
-{
33
- uint32_t crc;
34
- int carry, i, j;
35
- uint8_t b;
36
-
37
- crc = 0xffffffff;
38
- for (i = 0; i < 6; i++) {
39
- b = *ep++;
40
- for (j = 0; j < 8; j++) {
41
- carry = ((crc & 0x80000000L) ? 1 : 0) ^ (b & 0x01);
42
- crc <<= 1;
43
- b >>= 1;
44
- if (carry) {
45
- crc = ((crc ^ POLYNOMIAL_BE) | carry);
46
- }
47
- }
48
- }
49
- return (crc & BITS(7, 2)) >> 2;
50
-}
51
-
52
/* Read a 16 bit control/status (CSR) register. */
53
static uint16_t e100_read_reg2(EEPRO100State *s, E100RegisterOffset addr)
54
{
55
@@ -XXX,XX +XXX,XX @@ static void set_multicast_list(EEPRO100State *s)
56
uint8_t multicast_addr[6];
57
pci_dma_read(&s->dev, s->cb_address + 10 + i, multicast_addr, 6);
58
TRACE(OTHER, logout("multicast entry %s\n", nic_dump(multicast_addr, 6)));
59
- unsigned mcast_idx = e100_compute_mcast_idx(multicast_addr);
60
+ unsigned mcast_idx = (net_crc32(multicast_addr, ETH_ALEN) &
61
+ BITS(7, 2)) >> 2;
62
assert(mcast_idx < 64);
63
s->mult[mcast_idx >> 3] |= (1 << (mcast_idx & 7));
64
}
65
@@ -XXX,XX +XXX,XX @@ static ssize_t nic_receive(NetClientState *nc, const uint8_t * buf, size_t size)
66
if (s->configuration[21] & BIT(3)) {
67
/* Multicast all bit is set, receive all multicast frames. */
68
} else {
69
- unsigned mcast_idx = e100_compute_mcast_idx(buf);
70
+ unsigned mcast_idx = (net_crc32(buf, ETH_ALEN) & BITS(7, 2)) >> 2;
71
assert(mcast_idx < 64);
72
if (s->mult[mcast_idx >> 3] & (1 << (mcast_idx & 7))) {
73
/* Multicast frame is allowed in hash table. */
74
--
75
2.7.4
76
77
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
This makes it much easier to compare the multicast CRC calculation endian and
3
This is needed to achieve migration, so the destination can restore its
4
bitshift against the Linux driver implementation.
4
index.
5
5
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
6
Setting base as last used idx, so destination will see as available all
7
the entries that the device did not use, including the in-flight
8
processing ones.
9
10
This is ok for networking, but other kinds of devices might have
11
problems with these retransmissions.
12
13
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
14
Acked-by: Michael S. Tsirkin <mst@redhat.com>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
15
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
16
---
9
hw/net/lan9118.c | 3 ++-
17
hw/virtio/vhost-vdpa.c | 17 +++++++++++++++++
10
1 file changed, 2 insertions(+), 1 deletion(-)
18
1 file changed, 17 insertions(+)
11
19
12
diff --git a/hw/net/lan9118.c b/hw/net/lan9118.c
20
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
13
index XXXXXXX..XXXXXXX 100644
21
index XXXXXXX..XXXXXXX 100644
14
--- a/hw/net/lan9118.c
22
--- a/hw/virtio/vhost-vdpa.c
15
+++ b/hw/net/lan9118.c
23
+++ b/hw/virtio/vhost-vdpa.c
16
@@ -XXX,XX +XXX,XX @@
24
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_base(struct vhost_dev *dev,
17
#include "qemu/osdep.h"
25
static int vhost_vdpa_get_vring_base(struct vhost_dev *dev,
18
#include "hw/sysbus.h"
26
struct vhost_vring_state *ring)
19
#include "net/net.h"
27
{
20
+#include "net/eth.h"
28
+ struct vhost_vdpa *v = dev->opaque;
21
#include "hw/devices.h"
29
int ret;
22
#include "sysemu/sysemu.h"
30
23
#include "hw/ptimer.h"
31
+ if (v->shadow_vqs_enabled) {
24
@@ -XXX,XX +XXX,XX @@ static int lan9118_filter(lan9118_state *s, const uint8_t *addr)
32
+ VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs,
25
}
33
+ ring->index);
26
} else {
34
+
27
/* Hash matching */
35
+ /*
28
- hash = compute_mcast_idx(addr);
36
+ * Setting base as last used idx, so destination will see as available
29
+ hash = net_crc32(addr, ETH_ALEN) >> 26;
37
+ * all the entries that the device did not use, including the in-flight
30
if (hash & 0x20) {
38
+ * processing ones.
31
return (s->mac_hashh >> (hash & 0x1f)) & 1;
39
+ *
32
} else {
40
+ * TODO: This is ok for networking, but other kinds of devices might
41
+ * have problems with these retransmissions.
42
+ */
43
+ ring->num = svq->last_used_idx;
44
+ return 0;
45
+ }
46
+
47
ret = vhost_vdpa_call(dev, VHOST_GET_VRING_BASE, ring);
48
trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num);
49
return ret;
33
--
50
--
34
2.7.4
51
2.7.4
35
52
36
53
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
This makes it much easier to compare the multicast CRC calculation endian and
3
Setting the log address would make the device start reporting invalid
4
bitshift against the Linux driver implementation.
4
dirty memory because the SVQ vrings are located in qemu's memory.
5
5
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
6
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
7
Acked-by: Michael S. Tsirkin <mst@redhat.com>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
9
---
9
hw/net/opencores_eth.c | 3 ++-
10
hw/virtio/vhost-vdpa.c | 3 ++-
10
1 file changed, 2 insertions(+), 1 deletion(-)
11
1 file changed, 2 insertions(+), 1 deletion(-)
11
12
12
diff --git a/hw/net/opencores_eth.c b/hw/net/opencores_eth.c
13
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
13
index XXXXXXX..XXXXXXX 100644
14
index XXXXXXX..XXXXXXX 100644
14
--- a/hw/net/opencores_eth.c
15
--- a/hw/virtio/vhost-vdpa.c
15
+++ b/hw/net/opencores_eth.c
16
+++ b/hw/virtio/vhost-vdpa.c
16
@@ -XXX,XX +XXX,XX @@
17
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started)
17
#include "hw/net/mii.h"
18
static int vhost_vdpa_set_log_base(struct vhost_dev *dev, uint64_t base,
18
#include "hw/sysbus.h"
19
struct vhost_log *log)
19
#include "net/net.h"
20
{
20
+#include "net/eth.h"
21
- if (vhost_vdpa_one_time_request(dev)) {
21
#include "sysemu/sysemu.h"
22
+ struct vhost_vdpa *v = dev->opaque;
22
#include "trace.h"
23
+ if (v->shadow_vqs_enabled || vhost_vdpa_one_time_request(dev)) {
23
24
return 0;
24
@@ -XXX,XX +XXX,XX @@ static ssize_t open_eth_receive(NetClientState *nc,
25
}
25
if (memcmp(buf, bcast_addr, sizeof(bcast_addr)) == 0) {
26
26
miss = GET_REGBIT(s, MODER, BRO);
27
} else if ((buf[0] & 0x1) || GET_REGBIT(s, MODER, IAM)) {
28
- unsigned mcast_idx = compute_mcast_idx(buf);
29
+ unsigned mcast_idx = net_crc32(buf, ETH_ALEN) >> 26;
30
miss = !(s->regs[HASH0 + mcast_idx / 32] &
31
(1 << (mcast_idx % 32)));
32
trace_open_eth_receive_mcast(
33
--
27
--
34
2.7.4
28
2.7.4
35
29
36
30
diff view generated by jsdifflib
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
1
From: Eugenio Pérez <eperezma@redhat.com>
2
2
3
From the Linux sungem driver, we know that the multicast filter CRC is
3
SVQ is able to log the dirty bits by itself, so let's use it to not
4
implemented using ether_crc_le() which isn't the same as calling zlib's
4
block migration.
5
crc32() function (the zlib implementation requires a complemented initial value
6
and also returns the complemented result).
7
5
8
Fix the multicast filter by simply using the new net_crc32_le() function.
6
Also, ignore set and clear of VHOST_F_LOG_ALL on set_features if SVQ is
7
enabled. Even if the device supports it, the reports would be nonsense
8
because SVQ memory is in the qemu region.
9
9
10
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
10
The log region is still allocated. Future changes might skip that, but
11
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
11
this series is already long enough.
12
13
Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
14
Acked-by: Michael S. Tsirkin <mst@redhat.com>
12
Signed-off-by: Jason Wang <jasowang@redhat.com>
15
Signed-off-by: Jason Wang <jasowang@redhat.com>
13
---
16
---
14
hw/net/sungem.c | 5 ++---
17
hw/virtio/vhost-vdpa.c | 39 +++++++++++++++++++++++++++++++++++----
15
1 file changed, 2 insertions(+), 3 deletions(-)
18
include/hw/virtio/vhost-vdpa.h | 1 +
19
2 files changed, 36 insertions(+), 4 deletions(-)
16
20
17
diff --git a/hw/net/sungem.c b/hw/net/sungem.c
21
diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
18
index XXXXXXX..XXXXXXX 100644
22
index XXXXXXX..XXXXXXX 100644
19
--- a/hw/net/sungem.c
23
--- a/hw/virtio/vhost-vdpa.c
20
+++ b/hw/net/sungem.c
24
+++ b/hw/virtio/vhost-vdpa.c
21
@@ -XXX,XX +XXX,XX @@
25
@@ -XXX,XX +XXX,XX @@ static bool vhost_vdpa_one_time_request(struct vhost_dev *dev)
22
#include "hw/pci/pci.h"
26
return v->index != 0;
23
#include "qemu/log.h"
27
}
24
#include "net/net.h"
28
25
+#include "net/eth.h"
29
+static int vhost_vdpa_get_dev_features(struct vhost_dev *dev,
26
#include "net/checksum.h"
30
+ uint64_t *features)
27
#include "hw/net/mii.h"
31
+{
28
#include "sysemu/sysemu.h"
32
+ int ret;
29
#include "trace.h"
33
+
30
-/* For crc32 */
34
+ ret = vhost_vdpa_call(dev, VHOST_GET_FEATURES, features);
31
-#include <zlib.h>
35
+ trace_vhost_vdpa_get_features(dev, *features);
32
36
+ return ret;
33
#define TYPE_SUNGEM "sungem"
37
+}
34
38
+
35
@@ -XXX,XX +XXX,XX @@ static ssize_t sungem_receive(NetClientState *nc, const uint8_t *buf,
39
static int vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v,
40
Error **errp)
41
{
42
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v,
43
return 0;
36
}
44
}
37
45
38
/* Get MAC crc */
46
- r = hdev->vhost_ops->vhost_get_features(hdev, &dev_features);
39
- mac_crc = crc32(~0, buf, 6);
47
+ r = vhost_vdpa_get_dev_features(hdev, &dev_features);
40
+ mac_crc = net_crc32_le(buf, ETH_ALEN);
48
if (r != 0) {
41
49
error_setg_errno(errp, -r, "Can't get vdpa device features");
42
/* Packet isn't for me ? */
50
return r;
43
rx_cond = sungem_check_rx_mac(s, buf, mac_crc);
51
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_mem_table(struct vhost_dev *dev,
52
static int vhost_vdpa_set_features(struct vhost_dev *dev,
53
uint64_t features)
54
{
55
+ struct vhost_vdpa *v = dev->opaque;
56
int ret;
57
58
if (vhost_vdpa_one_time_request(dev)) {
59
return 0;
60
}
61
62
+ if (v->shadow_vqs_enabled) {
63
+ if ((v->acked_features ^ features) == BIT_ULL(VHOST_F_LOG_ALL)) {
64
+ /*
65
+ * QEMU is just trying to enable or disable logging. SVQ handles
66
+ * this sepparately, so no need to forward this.
67
+ */
68
+ v->acked_features = features;
69
+ return 0;
70
+ }
71
+
72
+ v->acked_features = features;
73
+
74
+ /* We must not ack _F_LOG if SVQ is enabled */
75
+ features &= ~BIT_ULL(VHOST_F_LOG_ALL);
76
+ }
77
+
78
trace_vhost_vdpa_set_features(dev, features);
79
ret = vhost_vdpa_call(dev, VHOST_SET_FEATURES, &features);
80
if (ret) {
81
@@ -XXX,XX +XXX,XX @@ static int vhost_vdpa_set_vring_call(struct vhost_dev *dev,
82
static int vhost_vdpa_get_features(struct vhost_dev *dev,
83
uint64_t *features)
84
{
85
- int ret;
86
+ struct vhost_vdpa *v = dev->opaque;
87
+ int ret = vhost_vdpa_get_dev_features(dev, features);
88
+
89
+ if (ret == 0 && v->shadow_vqs_enabled) {
90
+ /* Add SVQ logging capabilities */
91
+ *features |= BIT_ULL(VHOST_F_LOG_ALL);
92
+ }
93
94
- ret = vhost_vdpa_call(dev, VHOST_GET_FEATURES, features);
95
- trace_vhost_vdpa_get_features(dev, *features);
96
return ret;
97
}
98
99
diff --git a/include/hw/virtio/vhost-vdpa.h b/include/hw/virtio/vhost-vdpa.h
100
index XXXXXXX..XXXXXXX 100644
101
--- a/include/hw/virtio/vhost-vdpa.h
102
+++ b/include/hw/virtio/vhost-vdpa.h
103
@@ -XXX,XX +XXX,XX @@ typedef struct vhost_vdpa {
104
bool iotlb_batch_begin_sent;
105
MemoryListener listener;
106
struct vhost_vdpa_iova_range iova_range;
107
+ uint64_t acked_features;
108
bool shadow_vqs_enabled;
109
/* IOVA mapping used by the Shadow Virtqueue */
110
VhostIOVATree *iova_tree;
44
--
111
--
45
2.7.4
112
2.7.4
46
113
47
114
diff view generated by jsdifflib
Deleted patch
1
From: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
2
1
3
This makes it much easier to compare the multicast CRC calculation endian and
4
bitshift against the Linux driver implementation.
5
6
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
7
Signed-off-by: Jason Wang <jasowang@redhat.com>
8
---
9
hw/net/eepro100.c | 2 +-
10
1 file changed, 1 insertion(+), 1 deletion(-)
11
12
diff --git a/hw/net/eepro100.c b/hw/net/eepro100.c
13
index XXXXXXX..XXXXXXX 100644
14
--- a/hw/net/eepro100.c
15
+++ b/hw/net/eepro100.c
16
@@ -XXX,XX +XXX,XX @@ static ssize_t nic_receive(NetClientState *nc, const uint8_t * buf, size_t size)
17
rfd_status |= 0x0004;
18
} else if (s->configuration[20] & BIT(6)) {
19
/* Multiple IA bit set. */
20
- unsigned mcast_idx = compute_mcast_idx(buf);
21
+ unsigned mcast_idx = net_crc32(buf, ETH_ALEN) >> 26;
22
assert(mcast_idx < 64);
23
if (s->mult[mcast_idx >> 3] & (1 << (mcast_idx & 7))) {
24
TRACE(RXTX, logout("%p accepted, multiple IA bit set\n", s));
25
--
26
2.7.4
27
28
diff view generated by jsdifflib