rdma: move the ib_wr_opcode enum to include/uapi

By making this enum accessible to the userspace API, we can avoid conflicts between the kernel-space ib_wr_opcode and the user-space ibv_wr_opcode enums.

When using the rxe software driver interface between kernel and user space, the enum value IBV_WR_SEND_WITH_INV was being improperly translated to IB_WR_READ_WITH_INV causing the userspace application to fail.

Signed-off-by Seth Howell <[email protected]>

 include/rdma/ib_verbs.h           | 31 -------------------------------
 include/uapi/rdma/ib_user_verbs.h | 31 +++++++++++++++++++++++++++++++
 2 files changed, 31 insertions(+), 31 deletions(-)

From e3c1b1b81373601fa6cbad2ba5fadd2c8cfdfaed Mon Sep 17 00:00:00 2001

By making this enum accessible to the userspace API, we can avoid conflicts between the kernel-space ib_wr_opcode and the user-space ibv_wr_opcode enums.

When using the rxe software driver interface between kernel and user space, the enum value IBV_WR_SEND_WITH_INV was being improperly translated to IB_WR_READ_WITH_INV causing the userspace application to fail.

Signed-off-by Seth Howell <[email protected]>

 include/rdma/ib_verbs.h           | 31 -------------------------------
 include/uapi/rdma/ib_user_verbs.h | 31 
+++++++++++++++++++++++++++++++
 2 files changed, 31 insertions(+), 31 deletions(-)

From e3c1b1b81373601fa6cbad2ba5fadd2c8cfdfaed Mon Sep 17 00:00:00 2001

Thank you for taking the time to review this issue. The patch you submitted will definitely fix the problem I was facing.

Thank you for taking the time to review this issue. The patch you submitted will definitely fix the problem I was facing.

Hi Jason,

I apologize for the few days of radio silence on this one. I was able to apply your patch on my local configuration and can confirm that it fixes the issue of send with invalidate being improperly mapped between a user-space process and the kernel rxe driver.

Thanks again for your time and help.

Seth Howell

-----Original Message-----
From: [email protected] [mailto:
[email protected]] On Behalf Of Jason Gunthorpe
Sent: Thursday, August 16, 2018 11:01 AM
To: Howell, Seth <[email protected]>
Cc: Walker, Benjamin <[email protected]>; Doug Ledford < [email protected]>; [email protected]; [email protected] ; Moni Shoua <[email protected]>
Subject: Re: [PATCH] rdma: move the ib_wr_opcode enum to include/uapi

On Wed, Aug 15, 2018 at 11:27:36PM +0000, Howell, Seth wrote:

Quoted 2 lines

Thank you for taking the time to review this issue. The patch you submitted will definitely fix the problem I was facing.

Can you test it to be certain?

Thanks,
Jason

Quoted 25 lines · Seth

Hi Jason,

I apologize for the few days of radio silence on this one. I was able to apply your patch on my local configuration and can confirm that it fixes the issue of send with invalidate being improperly mapped between a user-space process and the kernel rxe driver.

Thanks again for your time and help.

Seth Howell

From: [email protected] [mailto:
[email protected]] On Behalf Of Jason Gunthorpe
Sent: Thursday, August 16, 2018 11:01 AM
To: Howell, Seth <[email protected]>
Cc: Walker, Benjamin <[email protected]>; Doug Ledford < [email protected]>; [email protected]; [email protected] ; Moni Shoua <[email protected]>
Subject: Re: [PATCH] rdma: move the ib_wr_opcode enum to include/uapi

Quoted 2 lines · Seth

Thank you for taking the time to review this issue. The patch you submitted will definitely fix the problem I was facing.

Quoted 4 lines

Can you test it to be certain?

Thanks,
Jason

We've recently run into this same issue on i40iw, which appears to make the same mistake of using the kernel version of the enum instead of the userspace version.

What's the current status here? Can it be merged? I just checked and do not see it merged to Linux master.

Running a user-space NVMe-oF target with RDMA and a recent Linux kernel initiator is not currently possible on rxe or i40iw because it requires send with invalidate support.

Quoted 4 lines · Benjamin

We've recently run into this same issue on i40iw, which appears to make the same
mistake of using the kernel version of the enum instead of the userspace version.

Confused by this?? i40iw_upost_send does not handle the kernel numbers at all, as far as I can see? How does it develop a kernel dependency??

Quoted 2 lines

What's the current status here? Can it be merged? I just checked and do not see it merged to Linux master.

Oh! This apparently got lost, thanks for bringing it up again.

Quoted 3 lines

Running a user-space NVMe-oF target with RDMA and a recent Linux kernel initiator is not currently possible on rxe or i40iw because it requires send with invalidate support.

Okay, but i40iw doesn't seem to support send with invalidate at all in userspace?

i40iw_upost_send() swithces on opcode, doesn't handle SEND_INV and then blows up in the default clause - how does this patch make any difference???

Jason

Quoted 21 lines · Jason Gunthorpe

Quoted 4 lines · Benjamin

We've recently run into this same issue on i40iw, which appears to make the same
mistake of using the kernel version of the enum instead of the userspace version.

Confused by this?? i40iw_upost_send does not handle the kernel numbers at all, as far as I can see? How does it develop a kernel dependency??

Quoted 2 lines

What's the current status here? Can it be merged? I just checked and do not see it merged to Linux master.

Oh! This apparently got lost, thanks for bringing it up again.

Quoted 3 lines

Running a user-space NVMe-oF target with RDMA and a recent Linux kernel initiator is not currently possible on rxe or i40iw because it requires send with invalidate support.

Okay, but i40iw doesn't seem to support send with invalidate at all in userspace?

Quoted 3 lines

i40iw_upost_send() swithces on opcode, doesn't handle SEND_INV and then blows up in the default clause - how does this patch make any difference???

It appears I've read the error message incorrectly and was looking at the kernel version (i40iw_post_send) as opposed to the user version (i40iw_upost_send). Indeed, the NIC does not support SEND_WTIH_INVAL at all in that function. The NIC does support SEND_WITH_INVAL in the kernel i40iw_post_send.

What is the correct way for a user space application to check whether a NIC supports SEND_WITH_INVAL? We are currently examining the device_cap_flags in the structure returned by ibv_query_device. Specifically, we're looking at IBV_DEVICE_MEM_MGT_EXTENSIONS. However, for i40iw, that flag is set. I'm concerned that the feature support flags are common between user space and the kernel, but the actual support differs in this case.

Quoted 4 lines

What is the correct way for a user space application to check whether a NIC supports SEND_WITH_INVAL? We are currently examining the device_cap_flags in the structure returned by ibv_query_device. Specifically, we're looking at IBV_DEVICE_MEM_MGT_EXTENSIONS.

However,

Quoted 2 lines

for i40iw, that flag is set. I'm concerned that the feature support flags are common between user space and the kernel, but the actual

support differs in this case.

That is the correct thing to do. Sadly the i40iw driver is broken, it should be fixed to either mask those flags or implement the functionality.

Jason

Quoted 1 line

From e3c1b1b81373601fa6cbad2ba5fadd2c8cfdfaed Mon Sep 17 00:00:00 2001

From: Seth Howell <[email protected]>
Date: Tue, 14 Aug 2018 15:33:02 -0700
Subject: [PATCH] IB/rxe: Revise the ib_wr_opcode enum

This enum has become part of the uABI, as both RXE and the ib_uverbs_post_send() command expect userspace to supply values from this enum. So it should be properly placed in include/uapi/rdma.

In userspace this enum is called 'enum ibv_wr_opcode' as part of libibverbs.h. That enum defines different values for IB_WR_LOCAL_INV, IB_WR_SEND_WITH_INV, and IB_WR_LSO. These were introduced (incorrectly, it turns out) into libiberbs in 2015.

The kernel has changed its mind on the numbering for several of the IB_WC values over the years, but has remained stable on IB_WR_LOCAL_INV and below.

Based on this we can conclude that there is no real user space user of the values beyond IB_WR_ATOMIC_FETCH_AND_ADD, as they have never worked via rdma-core. This is confirmed by inspection, only rxe uses the kernel enum and implements the latter operations. rxe has clearly never worked with these attributes from userspace. Other drivers that support these opcodes implement the functionality without calling out to the kernel.

To make IB_WR_SEND_WITH_INV and related work for RXE in userspace we choose to renumber the IB_WR enum in the kernel to match the uABI that userspace has bee using since before Soft RoCE was merged. This is an overall simpler configuration for the whole software stack, and obviously can't break anything existing.

Reported-by Seth Howell <[email protected]>

Fixes 8700e3e7c485 ("Soft RoCE driver")

Cc <[email protected]>

Signed-off-by Jason Gunthorpe <[email protected]>

 include/rdma/ib_verbs.h           | 34 ++++++++++++++++++-------------
 include/uapi/rdma/ib_user_verbs.h | 20 +++++++++++++++++-
 2 files changed, 39 insertions(+), 15 deletions(-)

Quoted 2 lines

From e3c1b1b81373601fa6cbad2ba5fadd2c8cfdfaed Mon Sep 17 00:00:00 2001

From: Seth Howell <[email protected]>
Date: Tue, 14 Aug 2018 15:33:02 -0700
Subject: [PATCH] IB/rxe: Revise the ib_wr_opcode enum

This enum has become part of the uABI, as both RXE and the ib_uverbs_post_send() command expect userspace to supply values from this enum. So it should be properly placed in include/uapi/rdma.

In userspace this enum is called 'enum ibv_wr_opcode' as part of libibverbs.h. That enum defines different values for IB_WR_LOCAL_INV, IB_WR_SEND_WITH_INV, and IB_WR_LSO. These were introduced (incorrectly, it turns out) into libiberbs in 2015.

The kernel has changed its mind on the numbering for several of the IB_WC values over the years, but has remained stable on IB_WR_LOCAL_INV and below.

Based on this we can conclude that there is no real user space user of the values beyond IB_WR_ATOMIC_FETCH_AND_ADD, as they have never worked via rdma-core. This is confirmed by inspection, only rxe uses the kernel enum and implements the latter operations. rxe has clearly never worked with these attributes from userspace. Other drivers that support these opcodes implement the functionality without calling out to the kernel.

To make IB_WR_SEND_WITH_INV and related work for RXE in userspace we choose to renumber the IB_WR enum in the kernel to match the uABI that userspace has bee using since before Soft RoCE was merged. This is an overall simpler configuration for the whole software stack, and obviously can't break anything existing.

Reported-by Seth Howell <[email protected]>

Fixes 8700e3e7c485 ("Soft RoCE driver")

Cc <[email protected]>

Signed-off-by Jason Gunthorpe <[email protected]>

 include/rdma/ib_verbs.h           | 34 ++++++++++++++++++-------------
 include/uapi/rdma/ib_user_verbs.h | 20 +++++++++++++++++-
 2 files changed, 39 insertions(+), 15 deletions(-)