[PATCHv2 0/2] Use netlink to create veth device pair when netlink is supported
V1 here: https://www.redhat.com/archives/libvir-list/2020-November/msg00898.html Since V1: * Include <linux/veth.h> rather than introducing enum VETH_INFO_* * Have complete functions within an #ifdefs rather than putting #ifdefs in function Shi Lei (2): util:netlink: Enable virNetlinkNewLink to support veth util:veth: Create veth device pair by netlink src/util/virnetdevveth.c | 85 ++++++++++++++++++++++++++-------------- src/util/virnetlink.c | 14 +++++++ src/util/virnetlink.h | 1 + 3 files changed, 71 insertions(+), 29 deletions(-) -- 2.25.1
Signed-off-by: Shi Lei <shi_lei@massclouds.com> --- src/util/virnetlink.c | 14 ++++++++++++++ src/util/virnetlink.h | 1 + 2 files changed, 15 insertions(+) diff --git a/src/util/virnetlink.c b/src/util/virnetlink.c index fdd3a6a4..8625b896 100644 --- a/src/util/virnetlink.c +++ b/src/util/virnetlink.c @@ -24,6 +24,7 @@ #include <config.h> #include <unistd.h> +#include <linux/veth.h> #include "virnetlink.h" #include "virnetdev.h" @@ -535,6 +536,19 @@ virNetlinkNewLink(const char *ifname, NETLINK_MSG_NEST_END(nl_msg, infodata); } + if (STREQ(type, "veth") && extra_args && extra_args->veth_peer) { + struct nlattr *infoveth = NULL; + + NETLINK_MSG_NEST_START(nl_msg, infodata, IFLA_INFO_DATA); + NETLINK_MSG_NEST_START(nl_msg, infoveth, VETH_INFO_PEER); + nlmsg_reserve(nl_msg, sizeof(struct ifinfomsg), 0); + NETLINK_MSG_PUT(nl_msg, IFLA_IFNAME, + (strlen(extra_args->veth_peer) + 1), + extra_args->veth_peer); + NETLINK_MSG_NEST_END(nl_msg, infoveth); + NETLINK_MSG_NEST_END(nl_msg, infodata); + } + NETLINK_MSG_NEST_END(nl_msg, linkinfo); if (extra_args) { diff --git a/src/util/virnetlink.h b/src/util/virnetlink.h index 7121eac4..7c4ed202 100644 --- a/src/util/virnetlink.h +++ b/src/util/virnetlink.h @@ -84,6 +84,7 @@ struct _virNetlinkNewLinkData { const int *ifindex; /* The index for the 'link' device */ const virMacAddr *mac; /* The MAC address of the device */ const uint32_t *macvlan_mode; /* The mode of macvlan */ + const char *veth_peer; /* The peer name for veth */ }; int virNetlinkNewLink(const char *ifname, -- 2.25.1
When netlink is supported, use netlink to create veth device pair rather than 'ip link' command. Signed-off-by: Shi Lei <shi_lei@massclouds.com> --- src/util/virnetdevveth.c | 85 ++++++++++++++++++++++++++-------------- 1 file changed, 56 insertions(+), 29 deletions(-) diff --git a/src/util/virnetdevveth.c b/src/util/virnetdevveth.c index b3eee1af..b4074371 100644 --- a/src/util/virnetdevveth.c +++ b/src/util/virnetdevveth.c @@ -27,6 +27,7 @@ #include "virfile.h" #include "virstring.h" #include "virnetdev.h" +#include "virnetlink.h" #define VIR_FROM_THIS VIR_FROM_NONE @@ -75,6 +76,55 @@ static int virNetDevVethGetFreeNum(int startDev) return -1; } +#if defined(WITH_LIBNL) +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + virNetlinkNewLinkData data = { .veth_peer = veth2 }; + + return virNetlinkNewLink(veth1, "veth", &data, status); +} + +static int +virNetDevVethDeleteInternal(const char *veth) +{ + return virNetlinkDelLink(veth, NULL); +} +#else +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + g_autoptr(virCommand) cmd = virCommandNew("ip"); + virCommandAddArgList(cmd, "link", "add", veth1, "type", "veth", + "peer", "name", veth2, NULL); + + return virCommandRun(cmd, status); +} + +static int +virNetDevVethDeleteInternal(const char *veth) +{ + int status; + g_autoptr(virCommand) cmd = virCommandNewArgList("ip", "link", + "del", veth, NULL); + + if (virCommandRun(cmd, &status) < 0) + return -1; + + if (status != 0) { + if (!virNetDevExists(veth)) { + VIR_DEBUG("Device %s already deleted (by kernel namespace cleanup)", veth); + return 0; + } + virReportError(VIR_ERR_INTERNAL_ERROR, + _("Failed to delete veth device %s"), veth); + return -1; + } + + return 0; +} +#endif /* WITH_LIBNL */ + /** * virNetDevVethCreate: * @veth1: pointer to name for parent end of veth pair @@ -100,7 +150,7 @@ static int virNetDevVethGetFreeNum(int startDev) * * Returns 0 on success or -1 in case of error */ -int virNetDevVethCreate(char** veth1, char** veth2) +int virNetDevVethCreate(char **veth1, char **veth2) { int ret = -1; int vethNum = 0; @@ -116,9 +166,8 @@ int virNetDevVethCreate(char** veth1, char** veth2) for (i = 0; i < MAX_VETH_RETRIES; i++) { g_autofree char *veth1auto = NULL; g_autofree char *veth2auto = NULL; - g_autoptr(virCommand) cmd = NULL; - int status; + int status = 0; if (!*veth1) { int veth1num; if ((veth1num = virNetDevVethGetFreeNum(vethNum)) < 0) @@ -136,14 +185,9 @@ int virNetDevVethCreate(char** veth1, char** veth2) vethNum = veth2num + 1; } - cmd = virCommandNew("ip"); - virCommandAddArgList(cmd, "link", "add", - *veth1 ? *veth1 : veth1auto, - "type", "veth", "peer", "name", - *veth2 ? *veth2 : veth2auto, - NULL); - - if (virCommandRun(cmd, &status) < 0) + if (virNetDevVethCreateInternal(*veth1 ? *veth1 : veth1auto, + *veth2 ? *veth2 : veth2auto, + &status) < 0) goto cleanup; if (status == 0) { @@ -188,22 +232,5 @@ int virNetDevVethCreate(char** veth1, char** veth2) */ int virNetDevVethDelete(const char *veth) { - int status; - g_autoptr(virCommand) cmd = virCommandNewArgList("ip", "link", - "del", veth, NULL); - - if (virCommandRun(cmd, &status) < 0) - return -1; - - if (status != 0) { - if (!virNetDevExists(veth)) { - VIR_DEBUG("Device %s already deleted (by kernel namespace cleanup)", veth); - return 0; - } - virReportError(VIR_ERR_INTERNAL_ERROR, - _("Failed to delete veth device %s"), veth); - return -1; - } - - return 0; + return virNetDevVethDeleteInternal(veth); } -- 2.25.1
On 11/22/20 10:28 PM, Shi Lei wrote:
When netlink is supported, use netlink to create veth device pair rather than 'ip link' command.
Signed-off-by: Shi Lei <shi_lei@massclouds.com> --- src/util/virnetdevveth.c | 85 ++++++++++++++++++++++++++-------------- 1 file changed, 56 insertions(+), 29 deletions(-)
diff --git a/src/util/virnetdevveth.c b/src/util/virnetdevveth.c index b3eee1af..b4074371 100644 --- a/src/util/virnetdevveth.c +++ b/src/util/virnetdevveth.c @@ -27,6 +27,7 @@ #include "virfile.h" #include "virstring.h" #include "virnetdev.h" +#include "virnetlink.h"
#define VIR_FROM_THIS VIR_FROM_NONE
@@ -75,6 +76,55 @@ static int virNetDevVethGetFreeNum(int startDev) return -1; }
+#if defined(WITH_LIBNL) +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + virNetlinkNewLinkData data = { .veth_peer = veth2 }; + + return virNetlinkNewLink(veth1, "veth", &data, status); +}
The only thing that makes me uncomfortable in this patch is that the two versions of virNetDevVethCreateInternal() each return something different for "status". In this first case, it is returning 0 on success, and -errno on failure...
[...]
+#else +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + g_autoptr(virCommand) cmd = virCommandNew("ip"); + virCommandAddArgList(cmd, "link", "add", veth1, "type", "veth", + "peer", "name", veth2, NULL); + + return virCommandRun(cmd, status); +}
But in this case it is returning the exit code of the "ip link add" command. It turns out that the value of status is only checked for 0 / not-0, so in practice it doesn't matter, but it could lead to confusion in the future. If we want these patches to be applied before your "netdevveth: Simplify virNetDevVethCreate by using virNetDevGenerateName" patch (which I'll get to as soon as I send this mail), then we do still need a way to differentiate between "The requested device already exists" and "Permanent Failure", and in that case I would suggest that "status" be replaced with a variable called "retry" which would be set to true if retrying with a different device name might lead to success (it would be set based on an interpretation of status made by each of the vir*Internal() functions) However, if we decide to apply the *other* patchset first, then we will never retry anyway (because we've already checked if the device exists before we try to create it, and there is therefore no loop) and so we could just eliminate the final argument completely, and keep each vir*Internal() function's status internal to itself. (As a matter of fact, status in the virCommandRun() version could simply be replaced with NULL in the arglist to virCommandRun(), and not defined at all; status in the case of virNetlinkNewLink() would still need to be defined and passed into the function, but its value would just be ignored). I think it may be cleaner to do the latter, and it looks like your other series was written to be applied without this series anyway; let me look at those patches and I'll reply a 2nd time to this based on the results of reviewing the other series... BTW, I appreciate your patience - I had looked at these patches nearly two weeks ago (soon after you sent them), but then a holiday got in the way, and I forgot to post my reply after I returned to work :-/
[...] - - if (virCommandRun(cmd, &status) < 0) + if (virNetDevVethCreateInternal(*veth1 ? *veth1 : veth1auto, + *veth2 ? *veth2 : veth2auto, + &status) < 0) goto cleanup;
if (status == 0) {
This spot here ^ is the only place that status is examined, and I actually think it's not going to work as you'd like in the case of netlink - status will always be 0 if virNetDevVethCreateInternal() returns 0, and when the return is < 0, status may or may not be set to -errno of the attempted operation - if status is 0, that means there was a serious error before even trying to create the device (e.g. OOM), and if it is non-0, then it might be because a device with the desired name already exists, or it might be because we don't have enough privilege to create a new device. Anyway, let me look at the other patchset...
Okay. According to your comment, I will deal with another series first. Then I get rid of that 'status' in this patch and post V3 of it. Thanks! Shi Lei On 2020-12-08 at 06:54, Laine Stump wrote:
On 11/22/20 10:28 PM, Shi Lei wrote:
When netlink is supported, use netlink to create veth device pair rather than 'ip link' command.
Signed-off-by: Shi Lei <shi_lei@massclouds.com> --- src/util/virnetdevveth.c | 85 ++++++++++++++++++++++++++-------------- 1 file changed, 56 insertions(+), 29 deletions(-)
diff --git a/src/util/virnetdevveth.c b/src/util/virnetdevveth.c index b3eee1af..b4074371 100644 --- a/src/util/virnetdevveth.c +++ b/src/util/virnetdevveth.c @@ -27,6 +27,7 @@ #include "virfile.h" #include "virstring.h" #include "virnetdev.h" +#include "virnetlink.h" #define VIR_FROM_THIS VIR_FROM_NONE @@ -75,6 +76,55 @@ static int virNetDevVethGetFreeNum(int startDev) return -1; } +#if defined(WITH_LIBNL) +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + virNetlinkNewLinkData data = { .veth_peer = veth2 }; + + return virNetlinkNewLink(veth1, "veth", &data, status); +}
The only thing that makes me uncomfortable in this patch is that the two versions of virNetDevVethCreateInternal() each return something different for "status". In this first case, it is returning 0 on success, and -errno on failure...
[...]
+#else +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + g_autoptr(virCommand) cmd = virCommandNew("ip"); + virCommandAddArgList(cmd, "link", "add", veth1, "type", "veth", + "peer", "name", veth2, NULL); + + return virCommandRun(cmd, status); +}
But in this case it is returning the exit code of the "ip link add" command.
It turns out that the value of status is only checked for 0 / not-0, so in practice it doesn't matter, but it could lead to confusion in the future.
If we want these patches to be applied before your "netdevveth: Simplify virNetDevVethCreate by using virNetDevGenerateName" patch (which I'll get to as soon as I send this mail), then we do still need a way to differentiate between "The requested device already exists" and "Permanent Failure", and in that case I would suggest that "status" be replaced with a variable called "retry" which would be set to true if retrying with a different device name might lead to success (it would be set based on an interpretation of status made by each of the vir*Internal() functions)
However, if we decide to apply the *other* patchset first, then we will never retry anyway (because we've already checked if the device exists before we try to create it, and there is therefore no loop) and so we could just eliminate the final argument completely, and keep each vir*Internal() function's status internal to itself. (As a matter of fact, status in the virCommandRun() version could simply be replaced with NULL in the arglist to virCommandRun(), and not defined at all; status in the case of virNetlinkNewLink() would still need to be defined and passed into the function, but its value would just be ignored).
I think it may be cleaner to do the latter, and it looks like your other series was written to be applied without this series anyway; let me look at those patches and I'll reply a 2nd time to this based on the results of reviewing the other series...
BTW, I appreciate your patience - I had looked at these patches nearly two weeks ago (soon after you sent them), but then a holiday got in the way, and I forgot to post my reply after I returned to work :-/
[...] - - if (virCommandRun(cmd, &status) < 0) + if (virNetDevVethCreateInternal(*veth1 ? *veth1 : veth1auto, + *veth2 ? *veth2 : veth2auto, + &status) < 0) goto cleanup; if (status == 0) {
This spot here ^ is the only place that status is examined, and I actually think it's not going to work as you'd like in the case of netlink - status will always be 0 if virNetDevVethCreateInternal() returns 0, and when the return is < 0, status may or may not be set to -errno of the attempted operation - if status is 0, that means there was a serious error before even trying to create the device (e.g. OOM), and if it is non-0, then it might be because a device with the desired name already exists, or it might be because we don't have enough privilege to create a new device.
Anyway, let me look at the other patchset...
On Mon, Dec 07, 2020 at 05:54:51PM -0500, Laine Stump wrote:
On 11/22/20 10:28 PM, Shi Lei wrote:
When netlink is supported, use netlink to create veth device pair rather than 'ip link' command.
Signed-off-by: Shi Lei <shi_lei@massclouds.com> --- src/util/virnetdevveth.c | 85 ++++++++++++++++++++++++++-------------- 1 file changed, 56 insertions(+), 29 deletions(-)
diff --git a/src/util/virnetdevveth.c b/src/util/virnetdevveth.c index b3eee1af..b4074371 100644 --- a/src/util/virnetdevveth.c +++ b/src/util/virnetdevveth.c @@ -27,6 +27,7 @@ #include "virfile.h" #include "virstring.h" #include "virnetdev.h" +#include "virnetlink.h" #define VIR_FROM_THIS VIR_FROM_NONE @@ -75,6 +76,55 @@ static int virNetDevVethGetFreeNum(int startDev) return -1; } +#if defined(WITH_LIBNL) +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + virNetlinkNewLinkData data = { .veth_peer = veth2 }; + + return virNetlinkNewLink(veth1, "veth", &data, status); +}
The only thing that makes me uncomfortable in this patch is that the two versions of virNetDevVethCreateInternal() each return something different for "status". In this first case, it is returning 0 on success, and -errno on failure...
Just change this to return -1, as we very rarely want to return -errno in libvirt, as we have formal error reporting.
[...]
+#else +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + g_autoptr(virCommand) cmd = virCommandNew("ip"); + virCommandAddArgList(cmd, "link", "add", veth1, "type", "veth", + "peer", "name", veth2, NULL); + + return virCommandRun(cmd, status); +}
But in this case it is returning the exit code of the "ip link add" command.
It turns out that the value of status is only checked for 0 / not-0, so in practice it doesn't matter, but it could lead to confusion in the future.
If we want these patches to be applied before your "netdevveth: Simplify virNetDevVethCreate by using virNetDevGenerateName" patch (which I'll get to as soon as I send this mail), then we do still need a way to differentiate between "The requested device already exists" and "Permanent Failure", and in that case I would suggest that "status" be replaced with a variable called "retry" which would be set to true if retrying with a different device name might lead to success (it would be set based on an interpretation of status made by each of the vir*Internal() functions)
However, if we decide to apply the *other* patchset first, then we will never retry anyway (because we've already checked if the device exists before we try to create it, and there is therefore no loop) and so we could just eliminate the final argument completely, and keep each vir*Internal() function's status internal to itself. (As a matter of fact, status in the virCommandRun() version could simply be replaced with NULL in the arglist to virCommandRun(), and not defined at all; status in the case of virNetlinkNewLink() would still need to be defined and passed into the function, but its value would just be ignored).
I think it may be cleaner to do the latter, and it looks like your other series was written to be applied without this series anyway; let me look at those patches and I'll reply a 2nd time to this based on the results of reviewing the other series...
BTW, I appreciate your patience - I had looked at these patches nearly two weeks ago (soon after you sent them), but then a holiday got in the way, and I forgot to post my reply after I returned to work :-/
[...] - - if (virCommandRun(cmd, &status) < 0) + if (virNetDevVethCreateInternal(*veth1 ? *veth1 : veth1auto, + *veth2 ? *veth2 : veth2auto, + &status) < 0) goto cleanup; if (status == 0) {
This spot here ^ is the only place that status is examined, and I actually think it's not going to work as you'd like in the case of netlink - status will always be 0 if virNetDevVethCreateInternal() returns 0, and when the return is < 0, status may or may not be set to -errno of the attempted operation - if status is 0, that means there was a serious error before even trying to create the device (e.g. OOM), and if it is non-0, then it might be because a device with the desired name already exists, or it might be because we don't have enough privilege to create a new device.
Anyway, let me look at the other patchset...
Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
On 12/8/20 4:19 AM, Daniel P. Berrangé wrote:
On Mon, Dec 07, 2020 at 05:54:51PM -0500, Laine Stump wrote:
On 11/22/20 10:28 PM, Shi Lei wrote:
When netlink is supported, use netlink to create veth device pair rather than 'ip link' command.
Signed-off-by: Shi Lei <shi_lei@massclouds.com> --- src/util/virnetdevveth.c | 85 ++++++++++++++++++++++++++-------------- 1 file changed, 56 insertions(+), 29 deletions(-)
diff --git a/src/util/virnetdevveth.c b/src/util/virnetdevveth.c index b3eee1af..b4074371 100644 --- a/src/util/virnetdevveth.c +++ b/src/util/virnetdevveth.c @@ -27,6 +27,7 @@ #include "virfile.h" #include "virstring.h" #include "virnetdev.h" +#include "virnetlink.h" #define VIR_FROM_THIS VIR_FROM_NONE @@ -75,6 +76,55 @@ static int virNetDevVethGetFreeNum(int startDev) return -1; } +#if defined(WITH_LIBNL) +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + virNetlinkNewLinkData data = { .veth_peer = veth2 }; + + return virNetlinkNewLink(veth1, "veth", &data, status); +}
The only thing that makes me uncomfortable in this patch is that the two versions of virNetDevVethCreateInternal() each return something different for "status". In this first case, it is returning 0 on success, and -errno on failure...
Just change this to return -1, as we very rarely want to return -errno in libvirt, as we have formal error reporting.
I guess I should give a bit more detail here - the "-errno" returned by virNetlinkNewLink() isn't the value of errno in the libvirt process at the time it learns there was an error - it is the value of errno returned in the netlink response message, ie what errno was set to in the kernel context that is handling and responding to netlink messages. errno in the libvirt process will not be set to this value. And since the libvirt functions that are decoding the response messages don't themselves log any errors, they aren't reporting that exact error... ... Hmmm. this points out that we really need to be logging the exact error in virNetDevVethCreateInternal(), since once we return from that function we no longer have full info on the reason for the failure. This would have been a problem if we had any inclination to retry the operation (with a new device name, which the current code must do), but once the other series from Shi is applied (the one that changes veth name generation to mimic what is done for tap and macvtap), we won't need to worry about retrying, and so it will be acceptable to immediately log the error.
[...]
+#else +static int +virNetDevVethCreateInternal(const char *veth1, const char *veth2, int *status) +{ + g_autoptr(virCommand) cmd = virCommandNew("ip"); + virCommandAddArgList(cmd, "link", "add", veth1, "type", "veth", + "peer", "name", veth2, NULL); + + return virCommandRun(cmd, status); +}
But in this case it is returning the exit code of the "ip link add" command.
Also in this case if status != NULL then virCommandRun() will only log an error if it failed to run the "ip" command; if ip runs but fails to create the device and returns an error exit code, virCommandRun() will silently return the exit code. In this case, we can just call "virCommandRun(cmd, NULL)" - when status is NULL, virCommandRun() will log an error if the exec fails, and also if the exec is successful but the command returns an error. And since I'm already here typing - when I reviewed the other series (the ones that change the device name generation) I did arrive at the opinion that we should apply those patches first, and these patches on top of that - that way you won't have to worry about the possibility of needing to retry virNetDevVethCreateInternal with a new name, and it will be okay for its implementations to immediately log errors.
participants (3)
-
Daniel P. Berrangé -
Laine Stump -
Shi Lei