Enrich rdmatool with an option to enable or disable privileged QKEY.
When enabled, non-privileged users will be allowed to specify a
controlled QKEY.
By default this parameter is disabled in order to comply with IB spec.
According to the IB specification rel-1.6, section 3.5.3:
"QKEYs with the most significant bit set are considered controlled
QKEYs, and a HCA does not allow a consumer to arbitrarily specify a
controlled QKEY."
This allows old applications which existed before the kernel commit:
0cadb4db79e1 ("RDMA/uverbs: Restrict usage of privileged QKEYs")
they can use privileged QKEYs without being a privileged user to now
be able to work again without being privileged granted they turn on this
parameter.
rdma tool command examples and output.
$ rdma system show
netns shared privileged-qkey off copy-on-fork on
$ rdma system set privileged-qkey on
$ rdma system show
netns shared privileged-qkey on copy-on-fork on
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
---
rdma/sys.c | 46 ++++++++++++++++++++++++++++++++++++++++++++--
rdma/utils.c | 1 +
2 files changed, 45 insertions(+), 2 deletions(-)
diff --git a/rdma/sys.c b/rdma/sys.c
index fd785b25..db34cb41 100644
--- a/rdma/sys.c
+++ b/rdma/sys.c
@@ -40,6 +40,17 @@ static int sys_show_parse_cb(const struct nlmsghdr *nlh, void *data)
mode_str);
}
+ if (tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]) {
+ uint8_t pqkey_mode;
+
+ pqkey_mode =
+ mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]);
+
+ print_color_on_off(PRINT_ANY, COLOR_NONE, "privileged-qkey",
+ "privileged-qkey %s ", pqkey_mode);
+
+ }
+
if (tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK])
cof = mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK]);
@@ -67,8 +78,9 @@ static int sys_show_no_args(struct rd *rd)
static int sys_show(struct rd *rd)
{
const struct rd_cmd cmds[] = {
- { NULL, sys_show_no_args},
- { "netns", sys_show_no_args},
+ { NULL, sys_show_no_args},
+ { "netns", sys_show_no_args},
+ { "privileged-qkey", sys_show_no_args},
{ 0 }
};
@@ -86,6 +98,17 @@ static int sys_set_netns_cmd(struct rd *rd, bool enable)
return rd_sendrecv_msg(rd, seq);
}
+static int sys_set_privileged_qkey_cmd(struct rd *rd, bool enable)
+{
+ uint32_t seq;
+
+ rd_prepare_msg(rd, RDMA_NLDEV_CMD_SYS_SET,
+ &seq, (NLM_F_REQUEST | NLM_F_ACK));
+ mnl_attr_put_u8(rd->nlh, RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE, enable);
+
+ return rd_sendrecv_msg(rd, seq);
+}
+
static bool sys_valid_netns_cmd(const char *cmd)
{
int i;
@@ -111,10 +134,28 @@ static int sys_set_netns_args(struct rd *rd)
return sys_set_netns_cmd(rd, cmd);
}
+static int sys_set_privileged_qkey_args(struct rd *rd)
+{
+ bool cmd;
+ int ret;
+
+ if (rd_no_arg(rd)) {
+ pr_err("valid options are: { on | off }\n");
+ return -EINVAL;
+ }
+
+ cmd = parse_on_off("privileged-qkey", rd_argv(rd), &ret);
+ if (ret)
+ return -EINVAL;
+
+ return sys_set_privileged_qkey_cmd(rd, cmd);
+}
+
static int sys_set_help(struct rd *rd)
{
pr_out("Usage: %s system set [PARAM] value\n", rd->filename);
pr_out(" system set netns { shared | exclusive }\n");
+ pr_out(" system set privileged-qkey { on | off }\n");
return 0;
}
@@ -124,6 +165,7 @@ static int sys_set(struct rd *rd)
{ NULL, sys_set_help },
{ "help", sys_set_help },
{ "netns", sys_set_netns_args},
+ { "privileged-qkey", sys_set_privileged_qkey_args},
{ 0 }
};
diff --git a/rdma/utils.c b/rdma/utils.c
index 8a091c05..09985069 100644
--- a/rdma/utils.c
+++ b/rdma/utils.c
@@ -473,6 +473,7 @@ static const enum mnl_attr_data_type nldev_policy[RDMA_NLDEV_ATTR_MAX] = {
[RDMA_NLDEV_ATTR_STAT_AUTO_MODE_MASK] = MNL_TYPE_U32,
[RDMA_NLDEV_ATTR_DEV_DIM] = MNL_TYPE_U8,
[RDMA_NLDEV_ATTR_RES_RAW] = MNL_TYPE_BINARY,
+ [RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE] = MNL_TYPE_U8,
};
static int rd_attr_check(const struct nlattr *attr, int *typep)
--
2.18.1On 10/23/23 5:22 AM, Patrisious Haddad wrote:
> diff --git a/rdma/sys.c b/rdma/sys.c
> index fd785b25..db34cb41 100644
> --- a/rdma/sys.c
> +++ b/rdma/sys.c
> @@ -40,6 +40,17 @@ static int sys_show_parse_cb(const struct nlmsghdr *nlh, void *data)
> mode_str);
> }
>
> + if (tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]) {
> + uint8_t pqkey_mode;
> +
> + pqkey_mode =
> + mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]);
just make it mode so it fits on one line.
40 characters for an attribute name .... I will never understand this
fascination with writing a sentence for an attribute name.
> +
> + print_color_on_off(PRINT_ANY, COLOR_NONE, "privileged-qkey",
> + "privileged-qkey %s ", pqkey_mode);
> +
> + }
> +
> if (tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK])
> cof = mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK]);
>
keep Petr's reviewed-by tag on the respin.
David Ahern <dsahern@gmail.com> writes:
> On 10/23/23 5:22 AM, Patrisious Haddad wrote:
>> diff --git a/rdma/sys.c b/rdma/sys.c
>> index fd785b25..db34cb41 100644
>> --- a/rdma/sys.c
>> +++ b/rdma/sys.c
>> @@ -40,6 +40,17 @@ static int sys_show_parse_cb(const struct nlmsghdr *nlh, void *data)
>> mode_str);
>> }
>>
>> + if (tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]) {
>> + uint8_t pqkey_mode;
>> +
>> + pqkey_mode =
>> + mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]);
>
> just make it mode so it fits on one line.
>
> 40 characters for an attribute name .... I will never understand this
> fascination with writing a sentence for an attribute name.
Yeah, I noticed this after the review, and figured oh whatever. But this
would be better:
int at = RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE;
uint8_t pqkey_mode;
pqkey_mode = mnl_attr_get_u8(tb[at]);
print_color_on_off(PRINT_ANY, COLOR_NONE, "privileged-qkey",
"privileged-qkey %s ", pqkey_mode);
>> +
>> + print_color_on_off(PRINT_ANY, COLOR_NONE, "privileged-qkey",
>> + "privileged-qkey %s ", pqkey_mode);
>> +
>> + }
>> +
>> if (tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK])
>> cof = mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK]);
>>
>
> keep Petr's reviewed-by tag on the respin.
On 10/24/2023 8:02 PM, David Ahern wrote:
> External email: Use caution opening links or attachments
>
>
> On 10/23/23 5:22 AM, Patrisious Haddad wrote:
>> diff --git a/rdma/sys.c b/rdma/sys.c
>> index fd785b25..db34cb41 100644
>> --- a/rdma/sys.c
>> +++ b/rdma/sys.c
>> @@ -40,6 +40,17 @@ static int sys_show_parse_cb(const struct nlmsghdr *nlh, void *data)
>> mode_str);
>> }
>>
>> + if (tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]) {
>> + uint8_t pqkey_mode;
>> +
>> + pqkey_mode =
>> + mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE]);
> just make it mode so it fits on one line.
will do.
>
> 40 characters for an attribute name .... I will never understand this
> fascination with writing a sentence for an attribute name.
me neither, just following the naming convention in
rdma/include/uapi/rdma/rdma_netlink.h sadly ...
>
>> +
>> + print_color_on_off(PRINT_ANY, COLOR_NONE, "privileged-qkey",
>> + "privileged-qkey %s ", pqkey_mode);
>> +
>> + }
>> +
>> if (tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK])
>> cof = mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK]);
>>
> keep Petr's reviewed-by tag on the respin.
will do, just making sure I understand "the respin" , you mean when I
re-send it ?
Patrisious Haddad <phaddad@nvidia.com> writes: > On 10/24/2023 8:02 PM, David Ahern wrote: > >>> + >>> + print_color_on_off(PRINT_ANY, COLOR_NONE, "privileged-qkey", >>> + "privileged-qkey %s ", pqkey_mode); >>> + >>> + } >>> + >>> if (tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK]) >>> cof = mnl_attr_get_u8(tb[RDMA_NLDEV_SYS_ATTR_COPY_ON_FORK]); >>> >> keep Petr's reviewed-by tag on the respin. > > will do, just making sure I understand "the respin" , you mean when I > re-send it ? Yes, to send a new version with fixes.
Patrisious Haddad <phaddad@nvidia.com> writes:
> Enrich rdmatool with an option to enable or disable privileged QKEY.
> When enabled, non-privileged users will be allowed to specify a
> controlled QKEY.
>
> By default this parameter is disabled in order to comply with IB spec.
> According to the IB specification rel-1.6, section 3.5.3:
> "QKEYs with the most significant bit set are considered controlled
> QKEYs, and a HCA does not allow a consumer to arbitrarily specify a
> controlled QKEY."
>
> This allows old applications which existed before the kernel commit:
> 0cadb4db79e1 ("RDMA/uverbs: Restrict usage of privileged QKEYs")
> they can use privileged QKEYs without being a privileged user to now
> be able to work again without being privileged granted they turn on this
> parameter.
>
> rdma tool command examples and output.
>
> $ rdma system show
> netns shared privileged-qkey off copy-on-fork on
>
> $ rdma system set privileged-qkey on
>
> $ rdma system show
> netns shared privileged-qkey on copy-on-fork on
>
> Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
> Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
Again, I'm not familiar with the subject matter, but mechanically this
looks OK to me.
Reviewed-by: Petr Machata <petrm@nvidia.com>
© 2016 - 2026 Red Hat, Inc.