Re: git: 66b5296f1b29 - main - ctld: Add support for NVMe over Fabrics

From: John Baldwin <jhb_at_FreeBSD.org>
Date: Tue, 14 Apr 2026 10:31:41 UTC
On 4/13/26 13:36, Alan Somers wrote:
> On Mon, Apr 13, 2026 at 10:56 AM John Baldwin <jhb@freebsd.org> wrote:
>>
>> On 4/13/26 11:51, Alan Somers wrote:
>>> On Wed, Aug 6, 2025 at 2:10 PM John Baldwin <jhb@freebsd.org> wrote:
>>>>
>>>> The branch main has been updated by jhb:
>>>>
>>>> URL: https://cgit.FreeBSD.org/src/commit/?id=66b5296f1b29083634e2875ff08c32e7b6b866a8
>>>>
>>>> commit 66b5296f1b29083634e2875ff08c32e7b6b866a8
>>>> Author:     John Baldwin <jhb@FreeBSD.org>
>>>> AuthorDate: 2025-08-06 19:57:50 +0000
>>>> Commit:     John Baldwin <jhb@FreeBSD.org>
>>>> CommitDate: 2025-08-06 19:59:13 +0000
>>>>
>>>>       ctld: Add support for NVMe over Fabrics
>>>>
>>>>       While the overall structure is similar for NVMeoF controllers and
>>>>       iSCSI targets, there are sufficient differences that NVMe support uses
>>>>       an alternate configuration syntax.
>>>>
>>>>       - In authentication groups, permitted NVMeoF hosts can be allowed by
>>>>         names (NQNs) via "host-nqn" values (similar to "initiator-name" for
>>>>         iSCSI).  Similarly, "host-address" accepts permitted host addresses
>>>>         similar to "initiator-portal" for iSCSI.
>>>>
>>>>       - A new "transport-group" context enumerates transports that can be
>>>>         used by a group of NVMeoF controllers similar to the "portal-group"
>>>>         context for iSCSI.  In this section, the "listen" keyword accepts a
>>>>         transport as well as an address to permit other types of transports
>>>>         besides TCP in the future.  The "foreign", "offload", and "redirect"
>>>>         keywords are also not meaningful and thus not supported.
>>>>
>>>>       - A new "controller" context describes an NVMeoF I/O controller
>>>>         similar to the "target" context for iSCSI.  One key difference here
>>>>         is that "lun" objects are replaced by "namespace" objects.  However,
>>>>         a "namespace" can reference a named global lun permitting LUNs to be
>>>>         shared between iSCSI targets and NVMeoF controllers.
>>>>
>>>>       NB: Authentication via CHAP is not implemented for NVMeoF.
>>>>
>>>>       Reviewed by:    imp
>>>>       Sponsored by:   Chelsio Communications
>>>>       Differential Revision:  https://reviews.freebsd.org/D48773
>>> ...
>>>> +struct target *
>>>> +conf::add_controller(const char *name)
>>>> +{
>>>> +       if (!nvmf_nqn_valid_strict(name)) {
>>>> +               log_warnx("controller name \"%s\" is invalid for NVMe", name);
>>>> +               return nullptr;
>>>> +       }
>>>> +
>>>> +       /*
>>>> +        * Normalize the name to lowercase to match iSCSI.
>>>> +        */
>>>> +       std::string t_name(name);
>>>> +       for (char &c : t_name)
>>>> +               c = tolower(c);
>>> ...
>>>
>>> This makes it impossible to define an uppercase or mixed case target
>>> name in ctld.  I guess the intent was to comply with rfc3722[^1]?
>>> Even so, it's surprising, because such target names used to work.
>>> It's also inconsistent, because it's still possible to create an
>>> uppercase target name using ctladm directly, like this:
>>>
>>> ctladm port -c -d iscsi -O cfiscsi_portal_group_tag=257 -O
>>> cfiscsi_target=iqn.2018-10.myhost:TESTVOL1
>>>
>>> Should we warn the user if they specify an uppercase target name, or
>>> even fail to create it?
>>>
>>> [^1]: https://datatracker.ietf.org/doc/html/rfc3722
>>
>> Note that this function is for NVMe, not iSCSI.  iSCSI targets are created in
>> conf::add_target which has similar code:
>>
>> struct target *
>> conf::add_target(const char *name)
>> {
>>          if (!valid_iscsi_name(name, log_warnx))
>>                  return (nullptr);
>>
>>          /*
>>           * RFC 3722 requires us to normalize the name to lowercase.
>>           */
>>          std::string t_name(name);
>>          for (char &c : t_name)
>>                  c = tolower(c);
>>
>> Prior to the C++ commit, this change was already in place:
>>
>> struct target *
>> target_new(struct conf *conf, const char *name)
>> {
>>          struct target *targ;
>>          int i, len;
>>
>>          targ = target_find(conf, name);
>>          if (targ != NULL) {
>>                  log_warnx("duplicated target \"%s\"", name);
>>                  return (NULL);
>>          }
>>          if (valid_iscsi_name(name, log_warnx) == false) {
>>                  return (NULL);
>>          }
>>          targ = new target();
>>          targ->t_name = checked_strdup(name);
>>
>>          /*
>>           * RFC 3722 requires us to normalize the name to lowercase.
>>           */
>>          len = strlen(name);
>>          for (i = 0; i < len; i++)
>>                  targ->t_name[i] = tolower(targ->t_name[i]);
>>
>>          targ->t_conf = conf;
>>          TAILQ_INSERT_TAIL(&conf->conf_targets, targ, t_next);
>>
>>          return (targ);
>> }
>>
>> This was present in commit 009ea47eb2d21856af4529aaaca32cd67748daea
>> which brought in the iSCSI target, so it has always been present
>> in ctld.
>>
>> Also, AFAICT, the names are still accepted, they are just normalized.
>>
>> I guess one difference is that before, target_new() called target_find()
>> with the non-normalized name to check for duplicates, and now we check
>> for duplicates after normalizing the name.  I'm not sure how that worked
>> in the past in practice as you would have had two targets with the same
>> name (e.g. I wonder what the ctladm portlist output looked like for this
>> case and if it would have listed two ports with the same name)?  I suspect
>> that was more by accident and probably didn't work properly in practice
>> (e.g. the kernel handoff ioctl used the normalized name when invoking
>> CTL_ISCSI, so connections to both "names" probably were always mapped to
>> only one of the connections, and finding a port during login processing
>> probably only found the first target, and only if the initiator gave the
>> all-lowercase name).
>>
>> That is to say, you didn't get an error before, but it didn't work, and
>> now it tells you that it doesn't work AFAICT.
> 
> Excuse me, I spoke a little too soon.  You are correct that ctld has
> been converting target names to lower case before registering them in
> the kernel for a long time.  The change is that previously, if an
> initiator attempted to connect to an uppercase target name, ctld would
> accept it.  That's because port_find_in_pg used strcasecmp in
> stable/14.  But change 4b1aac931465f39c5c26bfa1d5539a428d340f20
> removed strcasecmp, replacing it by the C++ STL's find method on
> std::unordered_map.
> 
> So we used to accept connections case-insensitively, and now we accept
> them case-sensitively.  To restore the previous behavior, should we
> add tolower() on the target_name in iscsi_connection::login() ?

Yes, we should normalize there, and that indeed is my fault and warrants
a Fixes tag.

-- 
John Baldwin