Re: [ANNOUNCE]: ConfigFS enabled Generic Target Mode and iSCSI TargetStack on v2.6.27-rc7

From: Vladislav Bolkhovitin
Date: Tue Oct 21 2008 - 14:21:23 EST

Next message: Alessandro Zummo: "Re: [rtc-linux] [PATCH] rtc-s3c: Fix section mismatch warnings"
Previous message: Phillip Lougher: "Re: Subject: [PATCH 12/16] Squashfs: header files"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Nicholas A. Bellinger wrote:

On Fri, 2008-10-10 at 21:48 +0400, Vladislav Bolkhovitin wrote:
Point taken however that $TARGET_MOD could, and probably should have
some manner of generic ACL infrastructure available through FABRIC <->
TARGET API. I will have a look at scst_register() and
scst_register_session() and see where it should be adapted to
target_core_mod.

Btw, saying that "management of all security stuff should be purely duty
of the mid-layer" is incorrect however. The generic target engine needs
to make it *EASIER* for $FABRIC to allow those initiator ports access to
Mapped LUNs through fabric *DEPENDENT* endpoints, but trying to put all
fabric depepdent ACL endpoint logic in target_core_mod is IMHO a bad
idea.

Since each SCSI fabric's method of attaching SCSI LUN to Initiator Port
Endpoints in $FABRIC_MOD to SCSI Device (I have been calling
this /sys/kernel/config/target/core/$STORAGE_OBJECT for target_core_mod)
to create the SCSI Target Port is different. The reference I use for
iscsi_target_mod (and hence wrt target_core_mod) is proper T10/SCSI
terminlogy AFAIK. Lets reference the objects in
http://www.haifa.il.ibm.com/satran/ips/EddyQuicksall-iSCSI-in-diagrams/portal_groups.pdf for the discussion so we can make sure we are on the same page..

For example, just because iSCSI uses TargetName + TargetPortalGroupTag
to attach target_core_mod's $STORAGE_OBJECTs at iSCSI Logical Units to,
does not mean that SAS, or another SCSI based target fabric know
anything about TargetName or TargetPortalGroupTag. In iSCSI, this is
defined in Section 2.1:

The I_T nexus can be identified by the conjunction of the SCSI port
names; that is, the I_T nexus identifier is the tuple (iSCSI
Initiator Name + ',i,'+ ISID, iSCSI Target Name + ',t,'+ Portal Group Tag).

Obviously the Initiator and Target Ports wrt iSCSI fabric are more
"symbolic" than devices attached to say a legacy Parallel SCSI bus
because of IP storage having multiple IP network portals across multiple
independent backbone providers and subnets (if you are using MC/S or
SCTP), etc, etc. This is this reason I think it does not make sense to
try to locate fabric dependent ACLs
under /sys/kernel/config/target/core/$STORAGE_OBJECT.

The type of things that need to be under $STORAGE_OBJECT, and that do
have a direct effect for $FABRIC mapped LUN endpoints are things like
device_type, max_sectors, sector_size, queue_depth and global READ-ONLY.
Of course, we want to be able to see *ALL* of
the /sys/kernel/config/target/$FABRIC dependent ACLs that have been
symlinked to said $STORAGE_OBJECT (this is one of the items on my list,
but not implemented in my current work).

Nicholas, you think too iSCSI centric. From access control POV only 2 thing matter:

Not true. Thre is *NOTHING* in target_core_mod's configfs layout that
is "iSCSI centric", or $FABRIC centric at all. We are talking about
configfs symbolic links with /bin/ls from target_core_mod storage
objects and $FABRIC_MOD portal group ports for $FABRIC LUNs. How
Initiators logging into those $FABRIC_MOD endpoints (Node ACLs) and
accessing those $FABRIC LUNs (LUN ACLs) is still $FABRIC dependent.

There is nothing iSCSI, SCSI, ATA or NBD centric about it, it is UNIX
centric and works generically across any fabric, that is the whole point
of having target_core_mod. Why would we want to limit the generic
target engine to having Parallel SCSI (see below) centric ACLs..?

1. Target name - to assign to it a default access control group (ACL, if you like that name), i.e. an ACL for initiators not listed in other ACLs

Wrong. For iSCSI, Section 2.1 of RFC-3720 defines it as TargetName +
TargetPortalGroupTag, and this is the method that all of my upstream
work and any proper implemention of target node endpoint and target
portal group assignement.

2. Initiator name - to assign it to the corresponding ACL.

It doesn't matter if those names are IQNs for iSCSI or WWNs for FC, or bus:id:lun for parallel SCSI.

For example, consider target "TTT", which has 2 ACLs: "Default" with Device1 as LUN 0 and "Group1" with "Device2" as LUN 0. "Group1" specified for initiator "III1". Then when initiator "III1" connected to target "TTT", it would be assigned to "Group1" and see "Device2". If then initiator "III2" connected, it would be assigned to "Default" ACL and see "Device1". "Default" group can be empty, if necessary. There's nothing transport specific in this approach at all.

Your example limits all iSCSI ACLs to TargetName, instead of TargetName
+TargetPortalGroupTag.

That is why everything related to iscsi_target_mod operation is
below /sys/kernel/config/target/iscsi/$IQN/$TPGT and
not /sys/kernel/config/target/iscsi/$IQN.

Obviously I am not going to limit my upstream iscsi_target_mod to an ACL
structure that does not take into account a complete RFC-3720
implementation, but I would be more than happy to see you update your
ACL code to reflect proper TargetName+TargetPortalGroupTag that RFC-3720
lays out for the iSCSI Target Port <-> SCSI Target Port mapping.

I strongly suggest you to look at SCST access control approach and make sure you understand it before reply. It would save us a lot of time and effort. Note, this approach isn't something theoretical. It's proved by 4 years of successful usage.

I don't really care about history, I care about code. Why don't you
start breaking out which code you want to go upstream so that it makes
my job easier or start integrating your own ACL control model into
drivers/lio-core/target_core_configfs.c and post a patch and then we can
discuss!

In all honesty however, the ACL code is a small nit-pick compared to how
we are going to merge your $FABRIC <-> $TARGET API with
drivers/lio-core. Why don't you start there first while I consider what
can be made generic for ACL code for the target_core_mod configfs
upstream work.

Also, it would be good, if you shift your terminology to be less iSCSI specific and use the corresponding terms from SAM, where possible. We are discussing a config interface for a generic target engine, aren't we? Otherwise sometimes it's quite hard for me to understand you and I have strong suspicions that other people are getting or already got lost in it.

Heh, why do you think I moved my upstream work to ConfigFS..? Being
able to use two 'mkdir -p' and two 'ln -s' to create two iSCSI Initiator
Node ACLs and four iSCSI Initiator Node ACLs is as easy as it gets!?

Being able to call a *SINGLE* mkdir -p to create Network Portal on an
iSCSI Target Portal Group, and from an unloaded iscsi_target_mod preform
four different iSCSI target mod ops is a simple as it gets

target_core_mod is a generic target engine that uses the most advanced
and complete iscsi_target_mod, so one must put effort into understanding
the drivers/lio-core/*configfs* to understand the simplicity of the
code.

Thus, I believe, all the ACL management should be done not in $FABRIC/, but in $TARGET/. It would remove all the corresponding configfs headaches from the target drivers writers.

But, in fact, I asked about completely different thing. SCSI target mid-layer in some cases needs to export in user space amount of data, which doesn't fit one page. /proc/scsi_tgt/sessions is one example. What should we do for it?

I did address point above in my work, and my commits
under /sys/kernel/config/target/iscsi implement how I get around the
PAGE_SIZE limitiations, which was something that I ran into (moving from
IOCTL and all, which requires overly complex kernel level information
code to get lots of output), to using ConfigFS, which has the same as
procfs and sysfs limits that you need to use seq_file() for > PAGE_SIZE.
Anyways, I did not end up using seq_file() for iscsi_target_mod current
configfs code, here is what I am using to address your above example wrt
getting all of session output:

Hmm, I looked at the code and in lio_target_initiator_nacl_info() saw something like:

rb += sprintf(page+rb, "LIO Session ID: %u "
"ISID: 0x%02x %02x %02x %02x %02x %02x "
"TSIH: %hu ", sess->sid,
sess->isid[0], sess->isid[1], sess->isid[2],
sess->isid[3], sess->isid[4], sess->isid[5],
sess->tsih);
rb += sprintf(page+rb, "SessionType: %s\n",
(SESS_OPS(sess)->SessionType) ?
"Discovery" : "Normal");
rb += sprintf(page+rb, "Cmds in Session Pool: %d ",
atomic_read(&sess->pool_count));
rb += sprintf(page+rb, "Session State: ");

It doesn't look for me like it addresses the PAGE_SIZE limitation issue.

You are still completely missing the point here.. Because I broke out
my projects *LEGACY* information code (just like every other upstream
project is required to do) I do not have gigantic nested loops in my
target_core_mod and iscsi_target_mod code that can only dump output
using seq_file() out of procfs or through god awful IOCTL code.

Every other upstream project that *HAS* broken out its legacy
informational code into sysfs (which again, has the same limitiation) or
another sane virtual FS control interface (like configfs) is working
just fine. Sysfs is used by people on many many millions of Linux
boxes, and all existing upstream projects that use sysfs have no problem
getting lots and lots and lots of info using /bin/cat even with the
PAGE_SIZE limitiation in place.

So this means you have two choices:

*) Fix your legacy code to use a sane informational output interface for
your upstream branch.

*) Produce a patch to solve the limitiation and produce an API and post
it to linuxfs-devel.

Again, for my upstream work with iscsi_target_mod, everyone will just be
using '/bin/cat' and wildcards (*) to grok the thousands
of /sys/kernel/config/target/iscsi/$IQNs configfs objects running on the
production systems. Because of this reason, I am not pained by this
limitiation (as some of your code appears to be) so please don't expect
me to produce this patch.

Sorry, Nicholas, but it's pretty hard to discuss something with you. Your complicated manner to express yourself (this isn't a critic, just statement of fact, I'm also pretty much not an ideal in this area) requires from your interlocutor a lot of effort to simply understand you, but I don't feel that you put comparable effort to understand what's written to you.

Let's restart our discussion and do it step by step. At first, some of terms you use are pretty confusing for me and, I suspect, many other people, as well as some terms I use seem confuse you. So, let's start from finding a common terminological ground. It will remove future misunderstandings and allow people to easier follow us. Below I'll propose some terms. I'll tried to make them as close to the regular Linux practice as possible, but if I'm not right somewhere everybody is welcome to correct me.

1. Let's use term "SCSI transport" instead of "fabric", which you use. This is well corresponding to the regular Linux practice as well as to SAM. Particularly, SAM doesn't have the word "fabric" anywhere.

2. Target name - an opaque string passed from target driver to SCSI target mid-layer. It contains whatever the target driver would like. For example, for iSCSI it can be Target Name, or Target Port Name + Target Portal Group Tag in string form. For Fibre Channel it can be WWN of the corresponding target port. For parallel SCSI it can be target's bus:id:lun numbers in string form. SCSI target mid-layer uses it to provide access control.

3. Initiator name - an opaque string passed from target driver to SCSI target mid-layer. It contains whatever the target driver would like. For example, for iSCSI it can be Initiator Name, or Initiator User Name @ Initiator Name in string form, like joe@xxxxxxxxxxxxxxxxxxx:01:1661f9ee7b5. For Fibre Channel it can be WWN of the corresponding initiator port. For parallel SCSI it can be initiator's bus:id:lun numbers in string form. SCSI target mid-layer uses it to provide access control.

Next, how access control works in SCST. A target driver registers using scst_register() a "target", which is an opaque object, used by target mid-layer to group sessions and some other related activities. It can be for target port for Fibre Channel or Target Portal, or Target Portal Group for iSCSI. During registration the target driver supplies target's name (see above its definition). Then, the target driver registers each new session using scst_register_session() binding it to the already registered target. During registration it provides initiator's name (see above it's definition).

The target mid-layer has predefined by administrator a set of ACLs. Each ACL contains a list of LUNs and a list of initiator names allowed to be bound to this group. There are also special "default" ACLs: one per each target (i.e. target name) and one global for targets without default ACL defined. In scst_register_session() the target mid-layer goes over all ACLs searching for one, containing the initiator name. If such ACL found, the session bound to it. If no such ACL found, the target mid-layer looks if the corresponding target has default ACL defined. If there is such ACL, the session bound to it. Otherwise, it's bound to the global default ACL.

Such approach has the following 2 advantages:

1. It's pretty simple to implement

2. It's transport independent. The only duty target drivers have with it is to do initiators authentication to be sure that the initiator name isn't a fake one. Most transports don't need such authentication. In fact, AFAIK, currently in Linux only iSCSI transport supports it. So, in this approach most target driver are *completely* free from caring about access control, *everything* is done by the mid-layer.

With your approach you push a lot of access control functionality from the target mid-layer to target drivers. Particularly:

1. Target driver need to care about user space configuration interface and each target driver will have a duplicated code. Is it good with tens of target drivers?

2. You need to define and then maintain the corresponding interface between target drivers and target mid-layer for access control helpers functions, provided by the target mid-layer to target drivers.

From other side, your approach don't have any advantages over one already used by SCST, which I described above.

Next, the PAGE_SIZE limit issue.

What you have implemented is "access allowed only for explicitly specified initiators and forbidden for all others". But there is also another approach: "access forbidden only for explicitly specified initiators and allowed for all others". How about it? In fact, it's a lot wider used in practice, than one you've implemented.

Then, if we add the "others are allowed" mode, we need a way to somehow show from the kernel to user space a list of such initiators. This list can be potentially huge with thousands of entries. Also the target mid-layer also needs a way to show all existing sessions with some transport independent parameters, like ACL, to which each session bound, count of outstanding commands, etc.

Thus, the need to show big lists is unavoidable. We can do it only in the following 3 ways. Correct me, if I wrong.

1. Add in configfs ability to display large files. This looks for me as a huge piece of work, which can easily take many months to be properly done, so I don't think this is an option for us.

2. Add a sysfs hierarchy in which we would be able to create for each list entry we want to show a dedicated subdirectory, in with we would show all the necessary attributes as one or more files. Like:

/sys/scsi_target/
`-- target
`-- sessions
`-- session1
| |-- initiator_name
| |-- target_name
| |-- acl_name
| `-- commands
|
`-- session2
| |-- initiator_name
| |-- target_name
| |-- acl_name
| `-- commands
|
.
.
.

In this case we would have 2 /sys hierarchies: one in sysfs and one in configfs. It's pretty much bad looking, isn't it?

3. Don't use configfs at all and do everything in sysfs. Actually, I don't see any real difference between:

# mkdir -p "$FABRIC/$DEF_IQN/tpgt_1/np/172.16.201.137:3260"

and

# echo "add_target $DEF_IQN 72.16.201.137:3260" >/sys/scsi_target/iscsi/control

as well as between:

# ln -s "$FABRIC/$DEF_IQN/tpgt_1/lun/lun_0" "$FABRIC/$DEF_IQN/tpgt_1/acls/$INITIATOR_DEBIAN/lun_0/."

and

# echo "add_lun $DEVICE_NAME" >/sys/scsi_target/acls/$DEF_IQN/control

Next, I haven't started breaking SCST code on one which should go to upstream and which shouldn't, because it's already "broken" a lot ago. All the /proc interface concentrated in scst_proc.c file and this file interacts with SCST core via well defined interface. All other SCST code in my opinion should go as is. I don't see what STGT or (sorry) LIO core can add to it. It's pretty much well polished.

Finally, I very much dislike your "my upstream iscsi_target_mod" and "my upstream work" kind of attitude. This is *our* work, right? Or should I stop wasting my time in a discussion with predefined result?

Vlad
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Next message: Alessandro Zummo: "Re: [rtc-linux] [PATCH] rtc-s3c: Fix section mismatch warnings"
Previous message: Phillip Lougher: "Re: Subject: [PATCH 12/16] Squashfs: header files"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]