RDMA (smbdirect) testing

RDMA (smbdirect) testing

[Steve French]

Namjae and Hyeoncheol,
Have you had any luck setting up virtual RDMA devices for testing ksmbd RDMA?

(similar to "/usr/sbin/rdma link add siw0 type siw netdev etho" then
mount with "rdma" mount option)



-- 
Thanks,

Steve

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-20 5:41 GMT+09:00, Steve French <smfrench@gmail.com>:
Hi Steve,
> Namjae and Hyeoncheol,
> Have you had any luck setting up virtual RDMA devices for testing ksmbd
> RDMA?
You seem to be asking about soft-ROCE(or soft-iWARP). Hyunchul had
been testing RDMA
 of ksmbd with it before.
Hyunchul, please explain how to set-up it.

Thanks!
>
> (similar to "/usr/sbin/rdma link add siw0 type siw netdev etho" then
> mount with "rdma" mount option)
>
>
>
> --
> Thanks,
>
> Steve
>

Re: RDMA (smbdirect) testing

[Hyunchul Lee]

Hello Steve,

Please refer to the page below to configure soft-ROCE:
https://support.mellanox.com/s/article/howto-configure-soft-roce

These kernel configs have to be turned on:
CONFIG_CIFS_SMB_DIRECT
CONFIG_SMB_SERVER
CONFIG_SMB_SERVER_SMBDIRECT

And you can mount cifs with SMB-direct:
mount -t cifs -o rdma,... //<IP address of ethernet interface coupled
with soft-ROCE>/<share> ...


2022년 5월 20일 (금) 오전 8:06, Namjae Jeon <linkinjeon@kernel.org>님이 작성:
>
> 2022-05-20 5:41 GMT+09:00, Steve French <smfrench@gmail.com>:
> Hi Steve,
> > Namjae and Hyeoncheol,
> > Have you had any luck setting up virtual RDMA devices for testing ksmbd
> > RDMA?
> You seem to be asking about soft-ROCE(or soft-iWARP). Hyunchul had
> been testing RDMA
>  of ksmbd with it before.
> Hyunchul, please explain how to set-up it.
>
> Thanks!
> >
> > (similar to "/usr/sbin/rdma link add siw0 type siw netdev etho" then
> > mount with "rdma" mount option)
> >
> >
> >
> > --
> > Thanks,
> >
> > Steve
> >



-- 
Thanks,
Hyunchul

Re: RDMA (smbdirect) testing

[David Howells]

Namjae Jeon <linkinjeon@kernel.org> wrote:

> You seem to be asking about soft-ROCE(or soft-iWARP). Hyunchul had been
> testing RDMA of ksmbd with it before.

Yep.  I don't have any RDMA-capable cards.  I have managed to use soft-IWarp
with NFS.

Also, if you know how to make Samba do RDMA, that would be useful.

Thanks,
David

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-20 15:20 GMT+09:00, David Howells <dhowells@redhat.com>:
> Namjae Jeon <linkinjeon@kernel.org> wrote:
>
>> You seem to be asking about soft-ROCE(or soft-iWARP). Hyunchul had been
>> testing RDMA of ksmbd with it before.
>
> Yep.  I don't have any RDMA-capable cards.  I have managed to use
> soft-IWarp
> with NFS.
Ah, I have the real hw, so I haven't tried it. I'll do it at this time.
>
> Also, if you know how to make Samba do RDMA, that would be useful.
As I know, samba doesn't support it yet. It is known as in-progress.

Thanks!
>
> Thanks,
> David
>
>

Re: RDMA (smbdirect) testing

[Tom Talpey]

SoftROCE is a bit of a hot mess in upstream right now. It's
getting a lot of attention, but it's still pretty shaky.
If you're testing, I'd STRONGLY recommend SoftiWARP.

Tom.

On 5/20/2022 2:01 AM, Hyunchul Lee wrote:
> Hello Steve,
> 
> Please refer to the page below to configure soft-ROCE:
> https://support.mellanox.com/s/article/howto-configure-soft-roce
> 
> These kernel configs have to be turned on:
> CONFIG_CIFS_SMB_DIRECT
> CONFIG_SMB_SERVER
> CONFIG_SMB_SERVER_SMBDIRECT
> 
> And you can mount cifs with SMB-direct:
> mount -t cifs -o rdma,... //<IP address of ethernet interface coupled
> with soft-ROCE>/<share> ...
> 
> 
> 2022년 5월 20일 (금) 오전 8:06, Namjae Jeon <linkinjeon@kernel.org>님이 작성:
>>
>> 2022-05-20 5:41 GMT+09:00, Steve French <smfrench@gmail.com>:
>> Hi Steve,
>>> Namjae and Hyeoncheol,
>>> Have you had any luck setting up virtual RDMA devices for testing ksmbd
>>> RDMA?
>> You seem to be asking about soft-ROCE(or soft-iWARP). Hyunchul had
>> been testing RDMA
>>   of ksmbd with it before.
>> Hyunchul, please explain how to set-up it.
>>
>> Thanks!
>>>
>>> (similar to "/usr/sbin/rdma link add siw0 type siw netdev etho" then
>>> mount with "rdma" mount option)
>>>
>>>
>>>
>>> --
>>> Thanks,
>>>
>>> Steve
>>>
> 
> 
> 

Re: RDMA (smbdirect) testing

[David Howells]

Tom Talpey <tom@talpey.com> wrote:

> SoftROCE is a bit of a hot mess in upstream right now. It's
> getting a lot of attention, but it's still pretty shaky.
> If you're testing, I'd STRONGLY recommend SoftiWARP.

I'm having problems getting that working.  I'm setting the client up with:

rdma link add siw0 type siw netdev enp6s0
mount //192.168.6.1/scratch /xfstest.scratch -o rdma,user=shares,pass=...

and then see:

CIFS: Attempting to mount \\192.168.6.1\scratch
CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too small
CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too small
CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too small
CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too small
CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
CIFS: VFS: cifs_mount failed w/return code = -2

in dmesg.

Problem is, I don't know what to do about it:-/

David

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 5/20/2022 2:12 PM, David Howells wrote:
> Tom Talpey <tom@talpey.com> wrote:
> 
>> SoftROCE is a bit of a hot mess in upstream right now. It's
>> getting a lot of attention, but it's still pretty shaky.
>> If you're testing, I'd STRONGLY recommend SoftiWARP.
> 
> I'm having problems getting that working.  I'm setting the client up with:
> 
> rdma link add siw0 type siw netdev enp6s0
> mount //192.168.6.1/scratch /xfstest.scratch -o rdma,user=shares,pass=...
> 
> and then see:
> 
> CIFS: Attempting to mount \\192.168.6.1\scratch
> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too small
> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too small
> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too small
> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too small
> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
> CIFS: VFS: cifs_mount failed w/return code = -2
> 
> in dmesg.
> 
> Problem is, I don't know what to do about it:-/

It looks like the client is hardcoding 16 sge's, and has no option to
configure a smaller value, or reduce its requested number. That's bad,
because providers all have their own limits - and SIW_MAX_SGE is 6. I
thought I'd seen this working (metze?), but either the code changed or
someone built a custom version.

Namjae/Long, have you used siw successfully? Why does the code require
16 sge's, regardless of other size limits? Normally, if the lower layer
supports fewer, the upper layer will simply reduce its operation sizes.

Tom.

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>
> On 5/20/2022 2:12 PM, David Howells wrote:
>> Tom Talpey <tom@talpey.com> wrote:
>>
>>> SoftROCE is a bit of a hot mess in upstream right now. It's
>>> getting a lot of attention, but it's still pretty shaky.
>>> If you're testing, I'd STRONGLY recommend SoftiWARP.
>>
>> I'm having problems getting that working.  I'm setting the client up
>> with:
>>
>> rdma link add siw0 type siw netdev enp6s0
>> mount //192.168.6.1/scratch /xfstest.scratch -o rdma,user=shares,pass=...
>>
>> and then see:
>>
>> CIFS: Attempting to mount \\192.168.6.1\scratch
>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too
>> small
>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too
>> small
>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too
>> small
>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too
>> small
>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>> CIFS: VFS: cifs_mount failed w/return code = -2
>>
>> in dmesg.
>>
>> Problem is, I don't know what to do about it:-/
>
> It looks like the client is hardcoding 16 sge's, and has no option to
> configure a smaller value, or reduce its requested number. That's bad,
> because providers all have their own limits - and SIW_MAX_SGE is 6. I
> thought I'd seen this working (metze?), but either the code changed or
> someone built a custom version.
I also fully agree that we should provide users with the path to
configure this value.
>
> Namjae/Long, have you used siw successfully?
No. I was able to reproduce the same problem that David reported. I
and Hyunchul will take a look. I also confirmed that RDMA work well
without any problems with soft-ROCE. Until this problem is fixed, I'd
like to say David to use soft-ROCE.

> Why does the code require
> 16 sge's, regardless of other size limits? Normally, if the lower layer
> supports fewer, the upper layer will simply reduce its operation sizes.
This should be answered by Long Li. It seems that he set the optimized
value for the NICs he used to implement RDMA in cifs.
>
> Tom.
>

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 5/23/2022 9:07 AM, David Howells wrote:
> Namjae Jeon <linkinjeon@kernel.org> wrote:
> 
>> No. I was able to reproduce the same problem that David reported. I
>> and Hyunchul will take a look. I also confirmed that RDMA work well
>> without any problems with soft-ROCE. Until this problem is fixed, I'd
>> like to say David to use soft-ROCE.
> 
> I managed to set up soft-RoCE and get cifs working over that.
> 
> One thing I'm not sure about: on my server, I see the roce device with
> ibv_devices, but on my test client, I don't.  I'm wondering if there's
> something I need to configure to get that.

Probably not, but I'd recommend to avoid wading into that if your
testing is going ok. Send a report to linux-rdma and let them know,
if so?

We are told that RedHat is removing SoftRoCE from RHEL9, this too might
affect your future testing plans...

Tom.

> I've attached the config for reference.  Note that I build a monolithic kernel
> so that I can PXE boot it directly out of the build dir rather than building
> and installing a fresh rpm each time.
> 
> David

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 5/22/2022 7:06 PM, Namjae Jeon wrote:
> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>>
>> On 5/20/2022 2:12 PM, David Howells wrote:
>>> Tom Talpey <tom@talpey.com> wrote:
>>>
>>>> SoftROCE is a bit of a hot mess in upstream right now. It's
>>>> getting a lot of attention, but it's still pretty shaky.
>>>> If you're testing, I'd STRONGLY recommend SoftiWARP.
>>>
>>> I'm having problems getting that working.  I'm setting the client up
>>> with:
>>>
>>> rdma link add siw0 type siw netdev enp6s0
>>> mount //192.168.6.1/scratch /xfstest.scratch -o rdma,user=shares,pass=...
>>>
>>> and then see:
>>>
>>> CIFS: Attempting to mount \\192.168.6.1\scratch
>>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too
>>> small
>>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too
>>> small
>>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too
>>> small
>>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too
>>> small
>>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>>> CIFS: VFS: cifs_mount failed w/return code = -2
>>>
>>> in dmesg.
>>>
>>> Problem is, I don't know what to do about it:-/
>>
>> It looks like the client is hardcoding 16 sge's, and has no option to
>> configure a smaller value, or reduce its requested number. That's bad,
>> because providers all have their own limits - and SIW_MAX_SGE is 6. I
>> thought I'd seen this working (metze?), but either the code changed or
>> someone built a custom version.
> I also fully agree that we should provide users with the path to
> configure this value.
>>
>> Namjae/Long, have you used siw successfully?
> No. I was able to reproduce the same problem that David reported. I
> and Hyunchul will take a look. I also confirmed that RDMA work well
> without any problems with soft-ROCE. Until this problem is fixed, I'd
> like to say David to use soft-ROCE.
> 
>> Why does the code require
>> 16 sge's, regardless of other size limits? Normally, if the lower layer
>> supports fewer, the upper layer will simply reduce its operation sizes.
> This should be answered by Long Li. It seems that he set the optimized
> value for the NICs he used to implement RDMA in cifs.

"Optimized" is a funny choice of words. If the provider doesn't support
the value, it's not much of an optimization to insist on 16. :)

Personally, I'd try building a kernel with smbdirect.h changed to have
SMBDIRECT_MAX_SGE set to 6, and see what happens. You might have to
reduce the r/w sizes in mount, depending on any other issues this may
reveal.

Tom.

Re: RDMA (smbdirect) testing

[Stefan Metzmacher]

Am 21.05.22 um 13:54 schrieb Tom Talpey:
> 
> On 5/20/2022 2:12 PM, David Howells wrote:
>> Tom Talpey <tom@talpey.com> wrote:
>>
>>> SoftROCE is a bit of a hot mess in upstream right now. It's
>>> getting a lot of attention, but it's still pretty shaky.
>>> If you're testing, I'd STRONGLY recommend SoftiWARP.
>>
>> I'm having problems getting that working.  I'm setting the client up with:
>>
>> rdma link add siw0 type siw netdev enp6s0
>> mount //192.168.6.1/scratch /xfstest.scratch -o rdma,user=shares,pass=...
>>
>> and then see:
>>
>> CIFS: Attempting to mount \\192.168.6.1\scratch
>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too small
>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too small
>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6 too small
>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6 too small
>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>> CIFS: VFS: cifs_mount failed w/return code = -2
>>
>> in dmesg.
>>
>> Problem is, I don't know what to do about it:-/
> 
> It looks like the client is hardcoding 16 sge's, and has no option to
> configure a smaller value, or reduce its requested number. That's bad,
> because providers all have their own limits - and SIW_MAX_SGE is 6. I
> thought I'd seen this working (metze?), but either the code changed or
> someone built a custom version.

No, I only tested with my smbdirect driver and there I don't have
such a problem.

I never god cifs.ko to work with smbdirect.

metze

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>>>
>>> On 5/20/2022 2:12 PM, David Howells wrote:
>>>> Tom Talpey <tom@talpey.com> wrote:
>>>>
>>>>> SoftROCE is a bit of a hot mess in upstream right now. It's
>>>>> getting a lot of attention, but it's still pretty shaky.
>>>>> If you're testing, I'd STRONGLY recommend SoftiWARP.
>>>>
>>>> I'm having problems getting that working.  I'm setting the client up
>>>> with:
>>>>
>>>> rdma link add siw0 type siw netdev enp6s0
>>>> mount //192.168.6.1/scratch /xfstest.scratch -o
>>>> rdma,user=shares,pass=...
>>>>
>>>> and then see:
>>>>
>>>> CIFS: Attempting to mount \\192.168.6.1\scratch
>>>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6
>>>> too
>>>> small
>>>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>>>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6
>>>> too
>>>> small
>>>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>>>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>>>> CIFS: VFS: _smbd_get_connection:1513 warning: device max_send_sge = 6
>>>> too
>>>> small
>>>> CIFS: VFS: _smbd_get_connection:1516 Queue Pair creation may fail
>>>> CIFS: VFS: _smbd_get_connection:1519 warning: device max_recv_sge = 6
>>>> too
>>>> small
>>>> CIFS: VFS: _smbd_get_connection:1522 Queue Pair creation may fail
>>>> CIFS: VFS: _smbd_get_connection:1559 rdma_create_qp failed -22
>>>> CIFS: VFS: cifs_mount failed w/return code = -2
>>>>
>>>> in dmesg.
>>>>
>>>> Problem is, I don't know what to do about it:-/
>>>
>>> It looks like the client is hardcoding 16 sge's, and has no option to
>>> configure a smaller value, or reduce its requested number. That's bad,
>>> because providers all have their own limits - and SIW_MAX_SGE is 6. I
>>> thought I'd seen this working (metze?), but either the code changed or
>>> someone built a custom version.
>> I also fully agree that we should provide users with the path to
>> configure this value.
>>>
>>> Namjae/Long, have you used siw successfully?
>> No. I was able to reproduce the same problem that David reported. I
>> and Hyunchul will take a look. I also confirmed that RDMA work well
>> without any problems with soft-ROCE. Until this problem is fixed, I'd
>> like to say David to use soft-ROCE.
>>
>>> Why does the code require
>>> 16 sge's, regardless of other size limits? Normally, if the lower layer
>>> supports fewer, the upper layer will simply reduce its operation sizes.
>> This should be answered by Long Li. It seems that he set the optimized
>> value for the NICs he used to implement RDMA in cifs.
>
> "Optimized" is a funny choice of words. If the provider doesn't support
> the value, it's not much of an optimization to insist on 16. :)
Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
the same with ksmbd...
>
> Personally, I'd try building a kernel with smbdirect.h changed to have
> SMBDIRECT_MAX_SGE set to 6, and see what happens. You might have to
> reduce the r/w sizes in mount, depending on any other issues this may
> reveal.
Agreed, and ksmbd should also be changed as well as cifs for test. We
are preparing the patches to improve this in ksmbd, rather than
changing/building this hardcoding every time.

diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h
index a87fca82a796..7003722ab004 100644
--- a/fs/cifs/smbdirect.h
+++ b/fs/cifs/smbdirect.h
@@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
 } __packed;

 /* Default maximum number of SGEs in a RDMA send/recv */
-#define SMBDIRECT_MAX_SGE      16
+#define SMBDIRECT_MAX_SGE      6
 /* The context for a SMBD request */
 struct smbd_request {
        struct smbd_connection *info;
diff --git a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c
index e646d79554b8..70662b3bd590 100644
--- a/fs/ksmbd/transport_rdma.c
+++ b/fs/ksmbd/transport_rdma.c
@@ -42,7 +42,7 @@
 /* SMB_DIRECT negotiation timeout in seconds */
 #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120

-#define SMB_DIRECT_MAX_SEND_SGES               8
+#define SMB_DIRECT_MAX_SEND_SGES               6
 #define SMB_DIRECT_MAX_RECV_SGES               1

 /*

Thanks!
>
> Tom.
>

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 5/23/2022 11:05 AM, Namjae Jeon wrote:
> 2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
>> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
>>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>>>> ...
>>>> Why does the code require
>>>> 16 sge's, regardless of other size limits? Normally, if the lower layer
>>>> supports fewer, the upper layer will simply reduce its operation sizes.
>>> This should be answered by Long Li. It seems that he set the optimized
>>> value for the NICs he used to implement RDMA in cifs.
>>
>> "Optimized" is a funny choice of words. If the provider doesn't support
>> the value, it's not much of an optimization to insist on 16. :)
> Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
> the same with ksmbd...
>>
>> Personally, I'd try building a kernel with smbdirect.h changed to have
>> SMBDIRECT_MAX_SGE set to 6, and see what happens. You might have to
>> reduce the r/w sizes in mount, depending on any other issues this may
>> reveal.
> Agreed, and ksmbd should also be changed as well as cifs for test. We
> are preparing the patches to improve this in ksmbd, rather than
> changing/building this hardcoding every time.

So, the patch is just for this test, right? Because I don't think any
kernel-based storage upper layer should ever need more than 2 or 3.
How many memory regions are you doing per operation? I would
expect one for the SMB3 headers, and another, if needed, for data.
These would all be lmr-type and would not require actual new memreg's.

And for bulk data, I would hope you're using fast-register, which
takes a different path and doesn't use the same sge's.

Getting this right, and keeping things efficient both in SGE bookkeeping
as well as memory registration efficiency, is the rocket science behind
RDMA performance and correctness. Slapping "16" or "6" or whatever isn't
the long-term fix.

Tom.

> diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h
> index a87fca82a796..7003722ab004 100644
> --- a/fs/cifs/smbdirect.h
> +++ b/fs/cifs/smbdirect.h
> @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>   } __packed;
> 
>   /* Default maximum number of SGEs in a RDMA send/recv */
> -#define SMBDIRECT_MAX_SGE      16
> +#define SMBDIRECT_MAX_SGE      6
>   /* The context for a SMBD request */
>   struct smbd_request {
>          struct smbd_connection *info;
> diff --git a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c
> index e646d79554b8..70662b3bd590 100644
> --- a/fs/ksmbd/transport_rdma.c
> +++ b/fs/ksmbd/transport_rdma.c
> @@ -42,7 +42,7 @@
>   /* SMB_DIRECT negotiation timeout in seconds */
>   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
> 
> -#define SMB_DIRECT_MAX_SEND_SGES               8
> +#define SMB_DIRECT_MAX_SEND_SGES               6
>   #define SMB_DIRECT_MAX_RECV_SGES               1
> 
>   /*
> 
> Thanks!
>>
>> Tom.
>>
> 

RE: RDMA (smbdirect) testing

[Long Li]

> Subject: Re: RDMA (smbdirect) testing
> 
> On 5/23/2022 11:05 AM, Namjae Jeon wrote:
> > 2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
> >> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
> >>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
> >>>> ...
> >>>> Why does the code require
> >>>> 16 sge's, regardless of other size limits? Normally, if the lower
> >>>> layer supports fewer, the upper layer will simply reduce its operation sizes.
> >>> This should be answered by Long Li. It seems that he set the
> >>> optimized value for the NICs he used to implement RDMA in cifs.
> >>
> >> "Optimized" is a funny choice of words. If the provider doesn't
> >> support the value, it's not much of an optimization to insist on 16.
> >> :)
> > Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
> > the same with ksmbd...
> >>
> >> Personally, I'd try building a kernel with smbdirect.h changed to
> >> have SMBDIRECT_MAX_SGE set to 6, and see what happens. You might have
> >> to reduce the r/w sizes in mount, depending on any other issues this
> >> may reveal.
> > Agreed, and ksmbd should also be changed as well as cifs for test. We
> > are preparing the patches to improve this in ksmbd, rather than
> > changing/building this hardcoding every time.
> 
> So, the patch is just for this test, right? Because I don't think any kernel-based
> storage upper layer should ever need more than 2 or 3.
> How many memory regions are you doing per operation? I would expect one for
> the SMB3 headers, and another, if needed, for data.
> These would all be lmr-type and would not require actual new memreg's.
> 
> And for bulk data, I would hope you're using fast-register, which takes a
> different path and doesn't use the same sge's.
> 
> Getting this right, and keeping things efficient both in SGE bookkeeping as well
> as memory registration efficiency, is the rocket science behind RDMA
> performance and correctness. Slapping "16" or "6" or whatever isn't the long-
> term fix.

I found max_sge is extremely large on Mellanox hardware, but smaller on other iWARP hardware.

Hardcoding it to 16 is certainly not a good choice. I think we should set it to the smaller value of 1) a predefined value (e.g. 8), and the 2) the max_sge the hardware reports.

If the CIFS upper layer ever sends data with larger number of SGEs, the send will fail.

Long

> 
> Tom.
> 
> > diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
> > a87fca82a796..7003722ab004 100644
> > --- a/fs/cifs/smbdirect.h
> > +++ b/fs/cifs/smbdirect.h
> > @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
> >   } __packed;
> >
> >   /* Default maximum number of SGEs in a RDMA send/recv */
> > -#define SMBDIRECT_MAX_SGE      16
> > +#define SMBDIRECT_MAX_SGE      6
> >   /* The context for a SMBD request */
> >   struct smbd_request {
> >          struct smbd_connection *info; diff --git
> > a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
> > e646d79554b8..70662b3bd590 100644
> > --- a/fs/ksmbd/transport_rdma.c
> > +++ b/fs/ksmbd/transport_rdma.c
> > @@ -42,7 +42,7 @@
> >   /* SMB_DIRECT negotiation timeout in seconds */
> >   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
> >
> > -#define SMB_DIRECT_MAX_SEND_SGES               8
> > +#define SMB_DIRECT_MAX_SEND_SGES               6
> >   #define SMB_DIRECT_MAX_RECV_SGES               1
> >
> >   /*
> >
> > Thanks!
> >>
> >> Tom.
> >>
> >

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-24 1:05 GMT+09:00, Tom Talpey <tom@talpey.com>:
> On 5/23/2022 11:05 AM, Namjae Jeon wrote:
>> 2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
>>> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
>>>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>>>>> ...
>>>>> Why does the code require
>>>>> 16 sge's, regardless of other size limits? Normally, if the lower
>>>>> layer
>>>>> supports fewer, the upper layer will simply reduce its operation
>>>>> sizes.
>>>> This should be answered by Long Li. It seems that he set the optimized
>>>> value for the NICs he used to implement RDMA in cifs.
>>>
>>> "Optimized" is a funny choice of words. If the provider doesn't support
>>> the value, it's not much of an optimization to insist on 16. :)
>> Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
>> the same with ksmbd...
>>>
>>> Personally, I'd try building a kernel with smbdirect.h changed to have
>>> SMBDIRECT_MAX_SGE set to 6, and see what happens. You might have to
>>> reduce the r/w sizes in mount, depending on any other issues this may
>>> reveal.
>> Agreed, and ksmbd should also be changed as well as cifs for test. We
>> are preparing the patches to improve this in ksmbd, rather than
>> changing/building this hardcoding every time.
>
> So, the patch is just for this test, right?
Yes.
> Because I don't think any
> kernel-based storage upper layer should ever need more than 2 or 3.
> How many memory regions are you doing per operation? I would
> expect one for the SMB3 headers, and another, if needed, for data.
> These would all be lmr-type and would not require actual new memreg's.
Maximum 4. (smb transform header , smb3 header+ response, data,
smb-direct header)
>
> And for bulk data, I would hope you're using fast-register, which
> takes a different path and doesn't use the same sge's.
For bulk data, ksmbd already using it.
>
> Getting this right, and keeping things efficient both in SGE bookkeeping
> as well as memory registration efficiency, is the rocket science behind
> RDMA performance and correctness. Slapping "16" or "6" or whatever isn't
> the long-term fix.
Okay.
>
> Tom.
>
>> diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h
>> index a87fca82a796..7003722ab004 100644
>> --- a/fs/cifs/smbdirect.h
>> +++ b/fs/cifs/smbdirect.h
>> @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>>   } __packed;
>>
>>   /* Default maximum number of SGEs in a RDMA send/recv */
>> -#define SMBDIRECT_MAX_SGE      16
>> +#define SMBDIRECT_MAX_SGE      6
>>   /* The context for a SMBD request */
>>   struct smbd_request {
>>          struct smbd_connection *info;
>> diff --git a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c
>> index e646d79554b8..70662b3bd590 100644
>> --- a/fs/ksmbd/transport_rdma.c
>> +++ b/fs/ksmbd/transport_rdma.c
>> @@ -42,7 +42,7 @@
>>   /* SMB_DIRECT negotiation timeout in seconds */
>>   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
>>
>> -#define SMB_DIRECT_MAX_SEND_SGES               8
>> +#define SMB_DIRECT_MAX_SEND_SGES               6
>>   #define SMB_DIRECT_MAX_RECV_SGES               1
>>
>>   /*
>>
>> Thanks!
>>>
>>> Tom.
>>>
>>
>

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-24 4:17 GMT+09:00, Long Li <longli@microsoft.com>:
>> Subject: Re: RDMA (smbdirect) testing
>>
>> On 5/23/2022 11:05 AM, Namjae Jeon wrote:
>> > 2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
>> >> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
>> >>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>> >>>> ...
>> >>>> Why does the code require
>> >>>> 16 sge's, regardless of other size limits? Normally, if the lower
>> >>>> layer supports fewer, the upper layer will simply reduce its
>> >>>> operation sizes.
>> >>> This should be answered by Long Li. It seems that he set the
>> >>> optimized value for the NICs he used to implement RDMA in cifs.
>> >>
>> >> "Optimized" is a funny choice of words. If the provider doesn't
>> >> support the value, it's not much of an optimization to insist on 16.
>> >> :)
>> > Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
>> > the same with ksmbd...
>> >>
>> >> Personally, I'd try building a kernel with smbdirect.h changed to
>> >> have SMBDIRECT_MAX_SGE set to 6, and see what happens. You might have
>> >> to reduce the r/w sizes in mount, depending on any other issues this
>> >> may reveal.
>> > Agreed, and ksmbd should also be changed as well as cifs for test. We
>> > are preparing the patches to improve this in ksmbd, rather than
>> > changing/building this hardcoding every time.
>>
>> So, the patch is just for this test, right? Because I don't think any
>> kernel-based
>> storage upper layer should ever need more than 2 or 3.
>> How many memory regions are you doing per operation? I would expect one
>> for
>> the SMB3 headers, and another, if needed, for data.
>> These would all be lmr-type and would not require actual new memreg's.
>>
>> And for bulk data, I would hope you're using fast-register, which takes a
>> different path and doesn't use the same sge's.
>>
>> Getting this right, and keeping things efficient both in SGE bookkeeping
>> as well
>> as memory registration efficiency, is the rocket science behind RDMA
>> performance and correctness. Slapping "16" or "6" or whatever isn't the
>> long-
>> term fix.
>
Hi Long,
> I found max_sge is extremely large on Mellanox hardware, but smaller on
> other iWARP hardware.
>
> Hardcoding it to 16 is certainly not a good choice. I think we should set it
> to the smaller value of 1) a predefined value (e.g. 8), and the 2) the
> max_sge the hardware reports.
Okay, Could you please send the patch for cifs.ko ?

Thanks.
>
> If the CIFS upper layer ever sends data with larger number of SGEs, the send
> will fail.
>
> Long
>
>>
>> Tom.
>>
>> > diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
>> > a87fca82a796..7003722ab004 100644
>> > --- a/fs/cifs/smbdirect.h
>> > +++ b/fs/cifs/smbdirect.h
>> > @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>> >   } __packed;
>> >
>> >   /* Default maximum number of SGEs in a RDMA send/recv */
>> > -#define SMBDIRECT_MAX_SGE      16
>> > +#define SMBDIRECT_MAX_SGE      6
>> >   /* The context for a SMBD request */
>> >   struct smbd_request {
>> >          struct smbd_connection *info; diff --git
>> > a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
>> > e646d79554b8..70662b3bd590 100644
>> > --- a/fs/ksmbd/transport_rdma.c
>> > +++ b/fs/ksmbd/transport_rdma.c
>> > @@ -42,7 +42,7 @@
>> >   /* SMB_DIRECT negotiation timeout in seconds */
>> >   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
>> >
>> > -#define SMB_DIRECT_MAX_SEND_SGES               8
>> > +#define SMB_DIRECT_MAX_SEND_SGES               6
>> >   #define SMB_DIRECT_MAX_RECV_SGES               1
>> >
>> >   /*
>> >
>> > Thanks!
>> >>
>> >> Tom.
>> >>
>> >
>

Re: RDMA (smbdirect) testing

[David Howells]

Is there some way for cifs to ask the RDMA layer what it supports?

David

Re: RDMA (smbdirect) testing

[Steve French]

Or alternatively - the "query network interfaces" ioctl - wonder if
additional flags that could be added (beyond what is defined in
MS-SMB2 2.2.32.5

On Tue, May 24, 2022 at 4:16 AM David Howells <dhowells@redhat.com> wrote:
>
> Is there some way for cifs to ask the RDMA layer what it supports?
>
> David
>


-- 
Thanks,

Steve

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 5/24/2022 1:49 PM, Steve French wrote:
> Or alternatively - the "query network interfaces" ioctl - wonder if
> additional flags that could be added (beyond what is defined in
> MS-SMB2 2.2.32.5

Huh? I see no reason to expose this over the wire, it's entirely a local
device API question. Zero impact on the far side of the connection.

> On Tue, May 24, 2022 at 4:16 AM David Howells <dhowells@redhat.com> wrote:
>>
>> Is there some way for cifs to ask the RDMA layer what it supports?

Yes, and in fact the client and server already fetch it. It's a
standard verbs device attibute.

client fs/cifs/smbdirect.c:

> 	if (info->id->device->attrs.max_send_sge < SMBDIRECT_MAX_SGE) {
> 		log_rdma_event(ERR,
> 			"warning: device max_send_sge = %d too small\n",
> 			info->id->device->attrs.max_send_sge);
> 		log_rdma_event(ERR, "Queue Pair creation may fail\n");
> 	}
> 	if (info->id->device->attrs.max_recv_sge < SMBDIRECT_MAX_SGE) {
> 		log_rdma_event(ERR,
> 			"warning: device max_recv_sge = %d too small\n",
> 			info->id->device->attrs.max_recv_sge);
> 		log_rdma_event(ERR, "Queue Pair creation may fail\n");
> 	}

server fs/ksmbd/transport_rdma.c:

> 
> 	if (device->attrs.max_send_sge < SMB_DIRECT_MAX_SEND_SGES) {
> 		pr_err("warning: device max_send_sge = %d too small\n",
> 		       device->attrs.max_send_sge);
> 		return -EINVAL;
> 	}
> 	if (device->attrs.max_recv_sge < SMB_DIRECT_MAX_RECV_SGES) {
> 		pr_err("warning: device max_recv_sge = %d too small\n",
> 		       device->attrs.max_recv_sge);
> 		return -EINVAL;
> 	}

Tom.

Re: RDMA (smbdirect) testing

[David Howells]

Okay - I got it working; somewhat at least.  Now to take out the mass of print
statements.

David

RE: RDMA (smbdirect) testing

[Long Li]

> Subject: Re: RDMA (smbdirect) testing
> 
> 2022-05-24 4:17 GMT+09:00, Long Li <longli@microsoft.com>:
> >> Subject: Re: RDMA (smbdirect) testing
> >>
> >> On 5/23/2022 11:05 AM, Namjae Jeon wrote:
> >> > 2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
> >> >> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
> >> >>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
> >> >>>> ...
> >> >>>> Why does the code require
> >> >>>> 16 sge's, regardless of other size limits? Normally, if the
> >> >>>> lower layer supports fewer, the upper layer will simply reduce
> >> >>>> its operation sizes.
> >> >>> This should be answered by Long Li. It seems that he set the
> >> >>> optimized value for the NICs he used to implement RDMA in cifs.
> >> >>
> >> >> "Optimized" is a funny choice of words. If the provider doesn't
> >> >> support the value, it's not much of an optimization to insist on 16.
> >> >> :)
> >> > Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
> >> > the same with ksmbd...
> >> >>
> >> >> Personally, I'd try building a kernel with smbdirect.h changed to
> >> >> have SMBDIRECT_MAX_SGE set to 6, and see what happens. You
> might
> >> >> have to reduce the r/w sizes in mount, depending on any other
> >> >> issues this may reveal.
> >> > Agreed, and ksmbd should also be changed as well as cifs for test.
> >> > We are preparing the patches to improve this in ksmbd, rather than
> >> > changing/building this hardcoding every time.
> >>
> >> So, the patch is just for this test, right? Because I don't think any
> >> kernel-based storage upper layer should ever need more than 2 or 3.
> >> How many memory regions are you doing per operation? I would expect
> >> one for the SMB3 headers, and another, if needed, for data.
> >> These would all be lmr-type and would not require actual new memreg's.
> >>
> >> And for bulk data, I would hope you're using fast-register, which
> >> takes a different path and doesn't use the same sge's.
> >>
> >> Getting this right, and keeping things efficient both in SGE
> >> bookkeeping as well as memory registration efficiency, is the rocket
> >> science behind RDMA performance and correctness. Slapping "16" or "6"
> >> or whatever isn't the
> >> long-
> >> term fix.
> >
> Hi Long,
> > I found max_sge is extremely large on Mellanox hardware, but smaller
> > on other iWARP hardware.
> >
> > Hardcoding it to 16 is certainly not a good choice. I think we should
> > set it to the smaller value of 1) a predefined value (e.g. 8), and the
> > 2) the max_sge the hardware reports.
> Okay, Could you please send the patch for cifs.ko ?

Yes, will do.

Long

> 
> Thanks.
> >
> > If the CIFS upper layer ever sends data with larger number of SGEs,
> > the send will fail.
> >
> > Long
> >
> >>
> >> Tom.
> >>
> >> > diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
> >> > a87fca82a796..7003722ab004 100644
> >> > --- a/fs/cifs/smbdirect.h
> >> > +++ b/fs/cifs/smbdirect.h
> >> > @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
> >> >   } __packed;
> >> >
> >> >   /* Default maximum number of SGEs in a RDMA send/recv */
> >> > -#define SMBDIRECT_MAX_SGE      16
> >> > +#define SMBDIRECT_MAX_SGE      6
> >> >   /* The context for a SMBD request */
> >> >   struct smbd_request {
> >> >          struct smbd_connection *info; diff --git
> >> > a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
> >> > e646d79554b8..70662b3bd590 100644
> >> > --- a/fs/ksmbd/transport_rdma.c
> >> > +++ b/fs/ksmbd/transport_rdma.c
> >> > @@ -42,7 +42,7 @@
> >> >   /* SMB_DIRECT negotiation timeout in seconds */
> >> >   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
> >> >
> >> > -#define SMB_DIRECT_MAX_SEND_SGES               8
> >> > +#define SMB_DIRECT_MAX_SEND_SGES               6
> >> >   #define SMB_DIRECT_MAX_RECV_SGES               1
> >> >
> >> >   /*
> >> >
> >> > Thanks!
> >> >>
> >> >> Tom.
> >> >>
> >> >
> >

Re: RDMA (smbdirect) testing

[David Howells]

Namjae Jeon <linkinjeon@kernel.org> wrote:

> diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h
> index a87fca82a796..7003722ab004 100644
> --- a/fs/cifs/smbdirect.h
> +++ b/fs/cifs/smbdirect.h
> @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>  } __packed;
> 
>  /* Default maximum number of SGEs in a RDMA send/recv */
> -#define SMBDIRECT_MAX_SGE      16
> +#define SMBDIRECT_MAX_SGE      6
>  /* The context for a SMBD request */
>  struct smbd_request {
>         struct smbd_connection *info;
> diff --git a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c
> index e646d79554b8..70662b3bd590 100644
> --- a/fs/ksmbd/transport_rdma.c
> +++ b/fs/ksmbd/transport_rdma.c
> @@ -42,7 +42,7 @@
>  /* SMB_DIRECT negotiation timeout in seconds */
>  #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
> 
> -#define SMB_DIRECT_MAX_SEND_SGES               8
> +#define SMB_DIRECT_MAX_SEND_SGES               6
>  #define SMB_DIRECT_MAX_RECV_SGES               1
> 
>  /*

With that, iWarp works for me.  You can add:

Tested-by: David Howells <dhowells@redhat.com>

Re: RDMA (smbdirect) testing

[David Howells]

Stefan Metzmacher <metze@samba.org> wrote:

> I never god cifs.ko to work with smbdirect.

It works for me.  I think you need the following modules:

	modprobe rdma_cm
	modprobe ib_umad
	modprobe siw # softiwarp
	modprobe rdma_rxe # softroce

I do:

	rdma link add siw0 type siw netdev enp2s0

for softiwarp or:

	rdma link add rxe0 type rxe netdev enp6s0

for softroce on the client before doing:

	mount //192.168.6.1/test /xfstest.test -o rdma,user=shares,pass=...

David

Re: RDMA (smbdirect) testing

[David Howells]

> With that, iWarp works for me.  You can add:

Note also that whilst wireshark can decode iwarp traffic carrying NFS, it
doesn't recognise iwarp traffic carrying cifs.

David

Re: RDMA (smbdirect) testing

[Stefan Metzmacher]

Hi David,

> Note also that whilst wireshark can decode iwarp traffic carrying NFS, it
> doesn't recognise iwarp traffic carrying cifs.

It works fine for me.

Can you share the capture file?

What version of wireshark are you using?

metze

Re: RDMA (smbdirect) testing

[David Howells]

[-- Attachment #1: Type: text/plain, Size: 223 bytes --]

Stefan Metzmacher <metze@samba.org> wrote:

> Can you share the capture file?

See attached.

> What version of wireshark are you using?

wireshark-3.6.2-1.fc35.x86_64
Wireshark 3.7.0 (v3.7.0rc0-1686-g64dfed53330f)

David


[-- Attachment #2: pcap-rs.gz --]
[-- Type: application/gzip, Size: 8560 bytes --]

Re: RDMA (smbdirect) testing

[Stefan Metzmacher]

Am 25.05.22 um 12:20 schrieb David Howells:
> Stefan Metzmacher <metze@samba.org> wrote:
> 
>> Can you share the capture file?
> 
> See attached.
> 
>> What version of wireshark are you using?
> 
> wireshark-3.6.2-1.fc35.x86_64
> Wireshark 3.7.0 (v3.7.0rc0-1686-g64dfed53330f)

Works fine for me with 3.6.2-2 on ubuntu and also a recent wireshark master version.

I just fixed a minor problem with fragmented iwrap_ddp_rdma_send messages in
frames 91-96. Which seem to happen because ksmbd.ko negotiated a preferred send size of 8192,
which is accepted by the cifs.ko. Windows only uses 1364, which means each send operation fits into
a single ethernet frame...

See https://gitlab.com/wireshark/wireshark/-/merge_requests/7025

metze

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 5/26/2022 10:56 AM, Stefan Metzmacher wrote:
> 
> Am 25.05.22 um 12:20 schrieb David Howells:
>> Stefan Metzmacher <metze@samba.org> wrote:
>>
>>> Can you share the capture file?
>>
>> See attached.
>>
>>> What version of wireshark are you using?
>>
>> wireshark-3.6.2-1.fc35.x86_64
>> Wireshark 3.7.0 (v3.7.0rc0-1686-g64dfed53330f)
> 
> Works fine for me with 3.6.2-2 on ubuntu and also a recent wireshark 
> master version.
> 
> I just fixed a minor problem with fragmented iwrap_ddp_rdma_send 
> messages in
> frames 91-96. Which seem to happen because ksmbd.ko negotiated a 
> preferred send size of 8192,
> which is accepted by the cifs.ko. Windows only uses 1364, which means 
> each send operation fits into
> a single ethernet frame...

That was not an accident. :)

There's a talk somewhere in which I mentioned how we tried to optimize
the smbdirect fragment size and always landed back on 1364. However,
this is only an implementation choice, the protocol supports a wide
range. So it's appropriate for wireshark to be accommodating.

> See https://gitlab.com/wireshark/wireshark/-/merge_requests/7025

I get a blank frame when I view both changes on this link. Do I need
a gitlab account??

Tom.

> 
> metze
> 

Re: RDMA (smbdirect) testing

[Stefan Metzmacher]

Hi Tom,

>>>> Can you share the capture file?
>>>
>>> See attached.
>>>
>>>> What version of wireshark are you using?
>>>
>>> wireshark-3.6.2-1.fc35.x86_64
>>> Wireshark 3.7.0 (v3.7.0rc0-1686-g64dfed53330f)
>>
>> Works fine for me with 3.6.2-2 on ubuntu and also a recent wireshark master version.
>>
>> I just fixed a minor problem with fragmented iwrap_ddp_rdma_send messages in
>> frames 91-96. Which seem to happen because ksmbd.ko negotiated a preferred send size of 8192,
>> which is accepted by the cifs.ko. Windows only uses 1364, which means each send operation fits into
>> a single ethernet frame...
> 
> That was not an accident. :)
> 
> There's a talk somewhere in which I mentioned how we tried to optimize
> the smbdirect fragment size and always landed back on 1364.

Yes, I remember 4096/3 :-)

> However, this is only an implementation choice, the protocol supports a wide
> range. So it's appropriate for wireshark to be accommodating.

Yes, but the past 20 years showed more than once that, everything but matching
Windows just leads to bugs, because of untested code paths.

>> See https://gitlab.com/wireshark/wireshark/-/merge_requests/7025
> 
> I get a blank frame when I view both changes on this link. Do I need
> a gitlab account??

I don't think so, I'm seeing it even without being logged in.

Does https://gitlab.com/wireshark/wireshark/-/merge_requests/7025.patch
or https://gitlab.com/wireshark/wireshark/-/merge_requests/7025/diffs
work?

metze

UAF in smbd_reconnect() when using softIWarp

[David Howells]

Hi Steve,

I switch to using the softIWarp driver as there's a deadlock in the softRoCE
driver.  However, this comes up with a repeatable UAF detected by KASAN.

The RDMA link was brought up with:

	rdma link add siw0 type siw netdev enp6s0

and then I started running xfstests with -g quick.

MOUNT_OPTIONS -- -ordma,username=shares,password=...,mfsymlinks -o context=system_u:object_r:root_t:s0 //carina/scratch /xfstest.scratch

The kernel was v5.18 + iwarp SGE patch.

The KASAN splat is attached.  Some decoded bits:

 smbd_reconnect+0xba/0x1a6
 smbd_reconnect (fs/cifs/smbdirect.c:1427):
	if (server->smbd_conn->transport_status == SMBD_CONNECTED) {

 _smbd_get_connection+0xce/0x1367
 _smbd_get_connection (fs/cifs/smbdirect.c:1530):
	info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);

 smbd_destroy+0x852/0x899
 smbd_destroy (fs/cifs/smbdirect.c:1323):
	(probably the kfree at the end on line 1407)

 __cifs_reconnect+0x315/0x4b3
 __cifs_reconnect (fs/cifs/connect.c:311 fs/cifs/connect.c:358) 
		smbd_destroy(server);

David
---
run fstests generic/005 at 2022-05-27 11:18:41
run fstests generic/006 at 2022-05-27 11:18:51
CIFS: VFS: smbd_recv_buf:1889 disconnected
==================================================================
BUG: KASAN: use-after-free in smbd_reconnect+0xba/0x1a6
Read of size 4 at addr ffff88813029e000 by task cifsd/4509

CPU: 2 PID: 4509 Comm: cifsd Not tainted 5.18.0-build2+ #467
Hardware name: ASUS All Series/H97-PLUS, BIOS 2306 10/09/2014
Call Trace:
 <TASK>
 dump_stack_lvl+0x45/0x59
 print_address_description.constprop.0+0x1f/0x2ce
 ? smbd_reconnect+0xba/0x1a6
 print_report+0xf0/0x1d6
 ? smbd_reconnect+0xba/0x1a6
 ? do_raw_spin_lock+0x13a/0x17b
 ? smbd_reconnect+0xba/0x1a6
 kasan_report+0x81/0xa1
 ? smbd_reconnect+0xba/0x1a6
 smbd_reconnect+0xba/0x1a6
 __cifs_reconnect+0x351/0x4b3
 ? cifs_mark_tcp_ses_conns_for_reconnect+0x1b3/0x1b3
 ? __raw_spin_lock_init+0x85/0x85
 cifs_readv_from_socket+0x29a/0x2f4
 cifs_read_from_socket+0x95/0xc5
 ? cifs_readv_from_socket+0x2f4/0x2f4
 ? cifs_small_buf_get+0x50/0x5d
 ? allocate_buffers+0xfb/0x186
 cifs_demultiplex_thread+0x19b/0xb64
 ? cifs_handle_standard+0x27e/0x27e
 ? lock_downgrade+0xad/0xad
 ? rcu_read_lock_bh_held+0xab/0xab
 ? pci_mmcfg_check_reserved+0xbd/0xbd
 ? preempt_count_sub+0x18/0xba
 ? _raw_spin_unlock_irqrestore+0x39/0x4c
 ? cifs_handle_standard+0x27e/0x27e
 kthread+0x164/0x173
 ? kthread_complete_and_exit+0x20/0x20
 ret_from_fork+0x1f/0x30
 </TASK>

Allocated by task 4505:
 stack_trace_save+0x8f/0xbe
 kasan_save_stack+0x1e/0x39
 kasan_set_track+0x21/0x26
 ____kasan_kmalloc+0x68/0x72
 kmem_cache_alloc_trace+0x121/0x162
 _smbd_get_connection+0xce/0x1367
 smbd_get_connection+0x21/0x3e
 cifs_get_tcp_session.part.0+0x853/0xbda
 mount_get_conns+0x51/0x594
 cifs_mount+0x8d/0x279
 cifs_smb3_do_mount+0x186/0x471
 smb3_get_tree+0x58/0x91
 vfs_get_tree+0x46/0x150
 do_new_mount+0x19f/0x2c9
 path_mount+0x6a5/0x6e3
 do_mount+0x9e/0xe1
 __do_sys_mount+0x150/0x17c
 do_syscall_64+0x39/0x46
 entry_SYSCALL_64_after_hwframe+0x44/0xae

Freed by task 4509:
 stack_trace_save+0x8f/0xbe
 kasan_save_stack+0x1e/0x39
 kasan_set_track+0x21/0x26
 kasan_set_free_info+0x20/0x2f
 ____kasan_slab_free+0xad/0xc9
 kfree+0x125/0x14b
 smbd_destroy+0x852/0x899
 __cifs_reconnect+0x315/0x4b3
 cifs_readv_from_socket+0x29a/0x2f4
 cifs_read_from_socket+0x95/0xc5
 cifs_demultiplex_thread+0x19b/0xb64
 kthread+0x164/0x173
 ret_from_fork+0x1f/0x30

Last potentially related work creation:
 stack_trace_save+0x8f/0xbe
 kasan_save_stack+0x1e/0x39
 __kasan_record_aux_stack+0x62/0x68
 insert_work+0x30/0xaf
 __queue_work+0x4b9/0x4dc
 queue_work_on+0x4d/0x67
 __ib_process_cq+0x219/0x268
 ib_poll_handler+0x3f/0x14c
 irq_poll_softirq+0xd8/0x1ab
 __do_softirq+0x202/0x489

Second to last potentially related work creation:
 stack_trace_save+0x8f/0xbe
 kasan_save_stack+0x1e/0x39
 __kasan_record_aux_stack+0x62/0x68
 insert_work+0x30/0xaf
 __queue_work+0x4b9/0x4dc
 queue_work_on+0x4d/0x67
 recv_done+0x16f/0x727
 __ib_process_cq+0x219/0x268
 ib_poll_handler+0x3f/0x14c
 irq_poll_softirq+0xd8/0x1ab
 __do_softirq+0x202/0x489

The buggy address belongs to the object at ffff88813029e000
 which belongs to the cache kmalloc-4k of size 4096
The buggy address is located 0 bytes inside of
 4096-byte region [ffff88813029e000, ffff88813029f000)
The buggy address belongs to the physical page:
page:0000000001f91160 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x13029e
head:0000000001f91160 order:1 compound_mapcount:0 compound_pincount:0
flags: 0x200000000010200(slab|head|node=0|zone=2)
raw: 0200000000010200 ffffea0004c06d08 ffffea0004c0a288 ffff888100040900
raw: 0000000000000000 ffff88813029e000 0000000100000001 0000000000000000
page dumped because: kasan: bad access detected

Memory state around the buggy address:
 ffff88813029df00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
 ffff88813029df80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
>ffff88813029e000: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
                   ^
 ffff88813029e080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
 ffff88813029e100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
==================================================================
Disabling lock debugging due to kernel taint
CIFS: VFS: RDMA transport re-established
CIFS: VFS: smbd_recv_buf:1889 disconnected

Re: RDMA (smbdirect) testing

[David Howells]

Stefan Metzmacher <metze@samba.org> wrote:

> Does https://gitlab.com/wireshark/wireshark/-/merge_requests/7025.patch

I tried applying it (there appear to be two patches therein), but it doesn't
seem to have any effect when viewing the pcap file I posted.  It says it
should decode the port as Artemis.

David

Re: RDMA (smbdirect) testing

[Stefan Metzmacher]

Am 27.05.22 um 13:46 schrieb David Howells:
> Stefan Metzmacher <metze@samba.org> wrote:
> 
>> Does https://gitlab.com/wireshark/wireshark/-/merge_requests/7025.patch
> 
> I tried applying it (there appear to be two patches therein), but it doesn't
> seem to have any effect when viewing the pcap file I posted.  It says it
> should decode the port as Artemis.

I found that it's needed to set the
tcp.try_heuristic_first preference to TRUE

Check this:

grep '^tcp' preferences
tcp.try_heuristic_first: TRUE

For ~/.config/wireshark/ (or ~/.wireshark/' as fallback)

metze

Re: RDMA (smbdirect) testing

[David Howells]

Stefan Metzmacher <metze@samba.org> wrote:

> I found that it's needed to set the
> tcp.try_heuristic_first preference to TRUE

That works, thanks.

Tested-by: David Howells <dhowells@redhat.com>

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-05-25 6:08 GMT+09:00, Long Li <longli@microsoft.com>:
>> Subject: Re: RDMA (smbdirect) testing
>>
>> 2022-05-24 4:17 GMT+09:00, Long Li <longli@microsoft.com>:
>> >> Subject: Re: RDMA (smbdirect) testing
>> >>
>> >> On 5/23/2022 11:05 AM, Namjae Jeon wrote:
>> >> > 2022-05-23 22:45 GMT+09:00, Tom Talpey <tom@talpey.com>:
>> >> >> On 5/22/2022 7:06 PM, Namjae Jeon wrote:
>> >> >>> 2022-05-21 20:54 GMT+09:00, Tom Talpey <tom@talpey.com>:
>> >> >>>> ...
>> >> >>>> Why does the code require
>> >> >>>> 16 sge's, regardless of other size limits? Normally, if the
>> >> >>>> lower layer supports fewer, the upper layer will simply reduce
>> >> >>>> its operation sizes.
>> >> >>> This should be answered by Long Li. It seems that he set the
>> >> >>> optimized value for the NICs he used to implement RDMA in cifs.
>> >> >>
>> >> >> "Optimized" is a funny choice of words. If the provider doesn't
>> >> >> support the value, it's not much of an optimization to insist on
>> >> >> 16.
>> >> >> :)
>> >> > Ah, It's obvious that cifs haven't been tested with soft-iWARP. And
>> >> > the same with ksmbd...
>> >> >>
>> >> >> Personally, I'd try building a kernel with smbdirect.h changed to
>> >> >> have SMBDIRECT_MAX_SGE set to 6, and see what happens. You
>> might
>> >> >> have to reduce the r/w sizes in mount, depending on any other
>> >> >> issues this may reveal.
>> >> > Agreed, and ksmbd should also be changed as well as cifs for test.
>> >> > We are preparing the patches to improve this in ksmbd, rather than
>> >> > changing/building this hardcoding every time.
>> >>
>> >> So, the patch is just for this test, right? Because I don't think any
>> >> kernel-based storage upper layer should ever need more than 2 or 3.
>> >> How many memory regions are you doing per operation? I would expect
>> >> one for the SMB3 headers, and another, if needed, for data.
>> >> These would all be lmr-type and would not require actual new memreg's.
>> >>
>> >> And for bulk data, I would hope you're using fast-register, which
>> >> takes a different path and doesn't use the same sge's.
>> >>
>> >> Getting this right, and keeping things efficient both in SGE
>> >> bookkeeping as well as memory registration efficiency, is the rocket
>> >> science behind RDMA performance and correctness. Slapping "16" or "6"
>> >> or whatever isn't the
>> >> long-
>> >> term fix.
>> >
>> Hi Long,
>> > I found max_sge is extremely large on Mellanox hardware, but smaller
>> > on other iWARP hardware.
>> >
>> > Hardcoding it to 16 is certainly not a good choice. I think we should
>> > set it to the smaller value of 1) a predefined value (e.g. 8), and the
>> > 2) the max_sge the hardware reports.
>> Okay, Could you please send the patch for cifs.ko ?
Long, My Chelsio(iWARP) NIC reports this value as 4. When I set it
with hw report value in cifs.ko, There is kernel oops in cifs.ko. Have
you checked smb-direct of cifs.ko with Chelsio and any iWARP NICs
before ? or only Mellanox NICs ?

Thanks!
>
> Yes, will do.
>
> Long
>
>>
>> Thanks.
>> >
>> > If the CIFS upper layer ever sends data with larger number of SGEs,
>> > the send will fail.
>> >
>> > Long
>> >
>> >>
>> >> Tom.
>> >>
>> >> > diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
>> >> > a87fca82a796..7003722ab004 100644
>> >> > --- a/fs/cifs/smbdirect.h
>> >> > +++ b/fs/cifs/smbdirect.h
>> >> > @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>> >> >   } __packed;
>> >> >
>> >> >   /* Default maximum number of SGEs in a RDMA send/recv */
>> >> > -#define SMBDIRECT_MAX_SGE      16
>> >> > +#define SMBDIRECT_MAX_SGE      6
>> >> >   /* The context for a SMBD request */
>> >> >   struct smbd_request {
>> >> >          struct smbd_connection *info; diff --git
>> >> > a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
>> >> > e646d79554b8..70662b3bd590 100644
>> >> > --- a/fs/ksmbd/transport_rdma.c
>> >> > +++ b/fs/ksmbd/transport_rdma.c
>> >> > @@ -42,7 +42,7 @@
>> >> >   /* SMB_DIRECT negotiation timeout in seconds */
>> >> >   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
>> >> >
>> >> > -#define SMB_DIRECT_MAX_SEND_SGES               8
>> >> > +#define SMB_DIRECT_MAX_SEND_SGES               6
>> >> >   #define SMB_DIRECT_MAX_RECV_SGES               1
>> >> >
>> >> >   /*
>> >> >
>> >> > Thanks!
>> >> >>
>> >> >> Tom.
>> >> >>
>> >> >
>> >
>

RE: RDMA (smbdirect) testing

[Long Li]

> Long, My Chelsio(iWARP) NIC reports this value as 4. When I set it with hw
> report value in cifs.ko, There is kernel oops in cifs.ko. Have you checked smb-
> direct of cifs.ko with Chelsio and any iWARP NICs before ? or only Mellanox
> NICs ?
> 
> Thanks!

Yes, I have tested on Chelsio. I didn't see kernel panic. In fact, I can pass a larger value (8) and successfully create a QP on Chelsio.

Can you paste your kernel panic trace?

> >> > If the CIFS upper layer ever sends data with larger number of SGEs,
> >> > the send will fail.
> >> >
> >> > Long
> >> >
> >> >>
> >> >> Tom.
> >> >>
> >> >> > diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
> >> >> > a87fca82a796..7003722ab004 100644
> >> >> > --- a/fs/cifs/smbdirect.h
> >> >> > +++ b/fs/cifs/smbdirect.h
> >> >> > @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
> >> >> >   } __packed;
> >> >> >
> >> >> >   /* Default maximum number of SGEs in a RDMA send/recv */
> >> >> > -#define SMBDIRECT_MAX_SGE      16
> >> >> > +#define SMBDIRECT_MAX_SGE      6
> >> >> >   /* The context for a SMBD request */
> >> >> >   struct smbd_request {
> >> >> >          struct smbd_connection *info; diff --git
> >> >> > a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
> >> >> > e646d79554b8..70662b3bd590 100644
> >> >> > --- a/fs/ksmbd/transport_rdma.c
> >> >> > +++ b/fs/ksmbd/transport_rdma.c
> >> >> > @@ -42,7 +42,7 @@
> >> >> >   /* SMB_DIRECT negotiation timeout in seconds */
> >> >> >   #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
> >> >> >
> >> >> > -#define SMB_DIRECT_MAX_SEND_SGES               8
> >> >> > +#define SMB_DIRECT_MAX_SEND_SGES               6
> >> >> >   #define SMB_DIRECT_MAX_RECV_SGES               1
> >> >> >
> >> >> >   /*
> >> >> >
> >> >> > Thanks!
> >> >> >>
> >> >> >> Tom.
> >> >> >>
> >> >> >
> >> >
> >

Re: RDMA (smbdirect) testing

[Tom Talpey]

On 6/2/2022 8:07 PM, Long Li wrote:
>> Long, My Chelsio(iWARP) NIC reports this value as 4. When I set it with hw
>> report value in cifs.ko, There is kernel oops in cifs.ko. Have you checked smb-
>> direct of cifs.ko with Chelsio and any iWARP NICs before ? or only Mellanox
>> NICs ?
>>
>> Thanks!
> 
> Yes, I have tested on Chelsio. I didn't see kernel panic. In fact, I can pass a larger value (8) and successfully create a QP on Chelsio.

There are many generations of Chelsio RDMA adapters, and this number
is very likely to be different. You both should be sure you're testing
with multiple configurations (and don't forget all the other in-kernel
RDMA NICs). But a constant value of "8" is still arbitrary.

The kernel should definitely not throw an oops when the NIC doesn't
support some precompiled constant value. Namjae, what oops, exactly?

Tom.

> Can you paste your kernel panic trace?
> 
>>>>> If the CIFS upper layer ever sends data with larger number of SGEs,
>>>>> the send will fail.
>>>>>
>>>>> Long
>>>>>
>>>>>>
>>>>>> Tom.
>>>>>>
>>>>>>> diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
>>>>>>> a87fca82a796..7003722ab004 100644
>>>>>>> --- a/fs/cifs/smbdirect.h
>>>>>>> +++ b/fs/cifs/smbdirect.h
>>>>>>> @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>>>>>>>    } __packed;
>>>>>>>
>>>>>>>    /* Default maximum number of SGEs in a RDMA send/recv */
>>>>>>> -#define SMBDIRECT_MAX_SGE      16
>>>>>>> +#define SMBDIRECT_MAX_SGE      6
>>>>>>>    /* The context for a SMBD request */
>>>>>>>    struct smbd_request {
>>>>>>>           struct smbd_connection *info; diff --git
>>>>>>> a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
>>>>>>> e646d79554b8..70662b3bd590 100644
>>>>>>> --- a/fs/ksmbd/transport_rdma.c
>>>>>>> +++ b/fs/ksmbd/transport_rdma.c
>>>>>>> @@ -42,7 +42,7 @@
>>>>>>>    /* SMB_DIRECT negotiation timeout in seconds */
>>>>>>>    #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
>>>>>>>
>>>>>>> -#define SMB_DIRECT_MAX_SEND_SGES               8
>>>>>>> +#define SMB_DIRECT_MAX_SEND_SGES               6
>>>>>>>    #define SMB_DIRECT_MAX_RECV_SGES               1
>>>>>>>
>>>>>>>    /*
>>>>>>>
>>>>>>> Thanks!
>>>>>>>>
>>>>>>>> Tom.
>>>>>>>>
>>>>>>>
>>>>>
>>>

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-06-08 2:26 GMT+09:00, Tom Talpey <tom@talpey.com>:
> On 6/2/2022 8:07 PM, Long Li wrote:
>>> Long, My Chelsio(iWARP) NIC reports this value as 4. When I set it with
>>> hw
>>> report value in cifs.ko, There is kernel oops in cifs.ko. Have you
>>> checked smb-
>>> direct of cifs.ko with Chelsio and any iWARP NICs before ? or only
>>> Mellanox
>>> NICs ?
>>>
>>> Thanks!
>>
>> Yes, I have tested on Chelsio. I didn't see kernel panic. In fact, I can
>> pass a larger value (8) and successfully create a QP on Chelsio.
>
> There are many generations of Chelsio RDMA adapters, and this number
> is very likely to be different. You both should be sure you're testing
> with multiple configurations (and don't forget all the other in-kernel
> RDMA NICs). But a constant value of "8" is still arbitrary.
>
> The kernel should definitely not throw an oops when the NIC doesn't
> support some precompiled constant value. Namjae, what oops, exactly?
Sorry for noise, This is not cifs.ko issue(with max sge 8). There
seems to be some issue in chelsio driver(cxgb4) on the latest kernel.
I have reported this issue to chelsio maintainers.
They was already aware of this problem and are trying to fix it. When
I checked again on linux-5.15, there was no problem. I am going to
send a reply about this mail after checking it with cxgb4 patch on the
latest kernel, It has not been fixed yet, still waiting for the patch
from them.

Thanks!
>
> Tom.
>
>> Can you paste your kernel panic trace?
>>
>>>>>> If the CIFS upper layer ever sends data with larger number of SGEs,
>>>>>> the send will fail.
>>>>>>
>>>>>> Long
>>>>>>
>>>>>>>
>>>>>>> Tom.
>>>>>>>
>>>>>>>> diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h index
>>>>>>>> a87fca82a796..7003722ab004 100644
>>>>>>>> --- a/fs/cifs/smbdirect.h
>>>>>>>> +++ b/fs/cifs/smbdirect.h
>>>>>>>> @@ -226,7 +226,7 @@ struct smbd_buffer_descriptor_v1 {
>>>>>>>>    } __packed;
>>>>>>>>
>>>>>>>>    /* Default maximum number of SGEs in a RDMA send/recv */
>>>>>>>> -#define SMBDIRECT_MAX_SGE      16
>>>>>>>> +#define SMBDIRECT_MAX_SGE      6
>>>>>>>>    /* The context for a SMBD request */
>>>>>>>>    struct smbd_request {
>>>>>>>>           struct smbd_connection *info; diff --git
>>>>>>>> a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c index
>>>>>>>> e646d79554b8..70662b3bd590 100644
>>>>>>>> --- a/fs/ksmbd/transport_rdma.c
>>>>>>>> +++ b/fs/ksmbd/transport_rdma.c
>>>>>>>> @@ -42,7 +42,7 @@
>>>>>>>>    /* SMB_DIRECT negotiation timeout in seconds */
>>>>>>>>    #define SMB_DIRECT_NEGOTIATE_TIMEOUT           120
>>>>>>>>
>>>>>>>> -#define SMB_DIRECT_MAX_SEND_SGES               8
>>>>>>>> +#define SMB_DIRECT_MAX_SEND_SGES               6
>>>>>>>>    #define SMB_DIRECT_MAX_RECV_SGES               1
>>>>>>>>
>>>>>>>>    /*
>>>>>>>>
>>>>>>>> Thanks!
>>>>>>>>>
>>>>>>>>> Tom.
>>>>>>>>>
>>>>>>>>
>>>>>>
>>>>
>

Re: RDMA (smbdirect) testing

[David Howells]

David Howells <dhowells@redhat.com> wrote:

> With that, iWarp works for me.  You can add:
> 
> Tested-by: David Howells <dhowells@redhat.com>

Would it be possible to get this pushed upstream as a stopgap fix?  iWarp
still doesn't work in 5.19 with cifs.

David

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-08-03 0:10 GMT+09:00, David Howells <dhowells@redhat.com>:
> David Howells <dhowells@redhat.com> wrote:
>
>> With that, iWarp works for me.  You can add:
>>
>> Tested-by: David Howells <dhowells@redhat.com>
>
> Would it be possible to get this pushed upstream as a stopgap fix?  iWarp
> still doesn't work in 5.19 with cifs.
I will try this.

Thanks.
>
> David
>
>

Re: RDMA (smbdirect) testing

[Namjae Jeon]

2022-08-03 9:55 GMT+09:00, Namjae Jeon <linkinjeon@kernel.org>:
> 2022-08-03 0:10 GMT+09:00, David Howells <dhowells@redhat.com>:
>> David Howells <dhowells@redhat.com> wrote:
>>
>>> With that, iWarp works for me.  You can add:
>>>
>>> Tested-by: David Howells <dhowells@redhat.com>
>>
>> Would it be possible to get this pushed upstream as a stopgap fix?  iWarp
>> still doesn't work in 5.19 with cifs.
> I will try this.
I have updated the patch, so I don't add your tested-by tag for
previous test patch.
Long said to change that value to 8. But max_sge in cifs need to be
set to 6 for sw-iWARP . I wonder if there is a problem with values
lower than 8...

Thanks.
>
> Thanks.
>>
>> David
>>
>>
>

Re: RDMA (smbdirect) testing

[David Howells]

Namjae Jeon <linkinjeon@kernel.org> wrote:

> Long said to change that value to 8. But max_sge in cifs need to be
> set to 6 for sw-iWARP . I wonder if there is a problem with values
> lower than 8...

Looking at the code, I think 5 might suffice.

David