[PATCH] nvme: fix cntlid uniqueness check for different hosts

[PATCH] nvme: fix cntlid uniqueness check for different hosts

[Sagi Grimberg]

A single physical host may use a separate {hostnqn, hostid} identifiers
when connecting to a controller. It is possible that the a distributed
nvmf controller may generate unique cntlid per hostnqn and not globally
unique. Hence, validate the cntlid is unique per nvmf host.

Fixes: 1b1031ca63b2 ("nvme: validate cntlid during controller initialisation")
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
---
 drivers/nvme/host/core.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index 11c24f89f4fd..86c59144da74 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -3265,6 +3265,10 @@ static bool nvme_validate_cntlid(struct nvme_subsystem *subsys,
 		if (nvme_state_terminal(tmp))
 			continue;
 
+		if ((ctrl->ops->flags & NVME_F_FABRICS) &&
+		    !nvmf_ctlr_matches_baseopts(tmp, ctrl->opts))
+			continue;
+
 		if (tmp->cntlid == ctrl->cntlid) {
 			dev_err(ctrl->device,
 				"Duplicate cntlid %u with %s, subsys %s, rejecting\n",
-- 
2.43.0

Re: [PATCH] nvme: fix cntlid uniqueness check for different hosts

[Sagi Grimberg]

On 23/08/2026 3:15, Sagi Grimberg wrote:
> A single physical host may use a separate {hostnqn, hostid} identifiers
> when connecting to a controller. It is possible that the a distributed
> nvmf controller may generate unique cntlid per hostnqn and not globally
> unique. Hence, validate the cntlid is unique per nvmf host.
>
> Fixes: 1b1031ca63b2 ("nvme: validate cntlid during controller initialisation")
> Signed-off-by: Sagi Grimberg <sagi@grimberg.me>

Any thoughts on this one?

Re: [PATCH] nvme: fix cntlid uniqueness check for different hosts

[Christoph Hellwig]

On Sun, Aug 23, 2026 at 03:15:09AM +0300, Sagi Grimberg wrote:
> A single physical host may use a separate {hostnqn, hostid} identifiers
> when connecting to a controller.

Not really.  As the hostid defines a host.  Linux just might be able to
implement multiple hosts when it say hosts multiple containers.

> It is possible that the a distributed
> nvmf controller may generate unique cntlid per hostnqn and not globally
> unique. Hence, validate the cntlid is unique per nvmf host.

If it does so it is not spec compliant, and we should not support it.

Re: [PATCH] nvme: fix cntlid uniqueness check for different hosts

[Sagi Grimberg]

On 02/09/2026 16:47, Christoph Hellwig wrote:
> On Sun, Aug 23, 2026 at 03:15:09AM +0300, Sagi Grimberg wrote:
>> A single physical host may use a separate {hostnqn, hostid} identifiers
>> when connecting to a controller.
> Not really.  As the hostid defines a host.  Linux just might be able to
> implement multiple hosts when it say hosts multiple containers.

Well, from the subsystem perspective, it is a separate host.

>
>> It is possible that the a distributed
>> nvmf controller may generate unique cntlid per hostnqn and not globally
>> unique. Hence, validate the cntlid is unique per nvmf host.
> If it does so it is not spec compliant, and we should not support it.

Hmm, interesting. It didn't occur to me that the cross-hosts controller IDs
are meaningful in any way.. The issue is that if a host connects to multiple
subsystem ports (say 16 for multipath), and also may use multiple 
hostnqn/hostid
pairs, the available number of cntlids starts dropping very quickly 
given that
there are only 16 bits in cntlid.

I am wandering if this issue was observed by others.