[PATCH] RDMA/ucontext: fix building with CONFIG_MMU=n

[PATCH] RDMA/ucontext: fix building with CONFIG_MMU=n

[Arnd Bergmann]

The zap_vma_ptes() is declared but not defined on NOMMU kernels,
causing a link error for the newly added uverbs code:

drivers/infiniband/core/uverbs_main.o: In function `uverbs_user_mmap_disassociate':
uverbs_main.c:(.text+0x114c): undefined reference to `zap_vma_ptes'
drivers/infiniband/core/uverbs_main.o: In function `rdma_umap_open':
uverbs_main.c:(.text+0x53c): undefined reference to `zap_vma_ptes'

To fix this, we can either make uverbs depend on CONFIG_MMU, or try
to build it anyway. Since this is the only compile-time dependency,
I decided to allow building it with an extra compile-time check for
CONFIG_MMU before calling the one function.

Fixes: 5f9794dc94f5 ("RDMA/ucontext: Add a core API for mmaping driver IO memory")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 drivers/infiniband/core/uverbs_main.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 40dcf3d02a4b..5c1202af0748 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -884,7 +884,8 @@ static void rdma_umap_open(struct vm_area_struct *vma)
 	 * point, so zap it.
 	 */
 	vma->vm_private_data = NULL;
-	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
+	if (IS_ENABLED(CONFIG_MMU))
+		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
 }
 
 static void rdma_umap_close(struct vm_area_struct *vma)
@@ -1023,8 +1024,9 @@ void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
 				continue;
 			list_del_init(&priv->list);
 
-			zap_vma_ptes(vma, vma->vm_start,
-				     vma->vm_end - vma->vm_start);
+			if (IS_ENABLED(CONFIG_MMU))
+				zap_vma_ptes(vma, vma->vm_start,
+					     vma->vm_end - vma->vm_start);
 			vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
 		}
 		mutex_unlock(&ufile->umap_lock);
-- 
2.18.0

Re: [PATCH] RDMA/ucontext: fix building with CONFIG_MMU=n

[Jason Gunthorpe]

On Thu, Sep 27, 2018 at 12:10:03PM +0200, Arnd Bergmann wrote:
> The zap_vma_ptes() is declared but not defined on NOMMU kernels,
> causing a link error for the newly added uverbs code:
> 
> drivers/infiniband/core/uverbs_main.o: In function `uverbs_user_mmap_disassociate':
> uverbs_main.c:(.text+0x114c): undefined reference to `zap_vma_ptes'
> drivers/infiniband/core/uverbs_main.o: In function `rdma_umap_open':
> uverbs_main.c:(.text+0x53c): undefined reference to `zap_vma_ptes'

I'm a bit surprised by this - the code is not new, it just got moved
out of the drivers, and I don't see any protections there against
CONFIG_MMU ??
 
> To fix this, we can either make uverbs depend on CONFIG_MMU, or try
> to build it anyway. Since this is the only compile-time dependency,
> I decided to allow building it with an extra compile-time check for
> CONFIG_MMU before calling the one function.

Hrm. So this code doesn't work at all on NOMMU, it relies on
remap_pfn_range, which always fails on those kernels, 

Due to this of this the entirety of INFINIBAND_USER_ACCESS is broken.

I think a kconfig change is the better way to go?

Jason

Re: [PATCH] RDMA/ucontext: fix building with CONFIG_MMU=n

[Arnd Bergmann]

On Thu, Sep 27, 2018 at 6:26 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Thu, Sep 27, 2018 at 12:10:03PM +0200, Arnd Bergmann wrote:
> > The zap_vma_ptes() is declared but not defined on NOMMU kernels,
> > causing a link error for the newly added uverbs code:
> >
> > drivers/infiniband/core/uverbs_main.o: In function `uverbs_user_mmap_disassociate':
> > uverbs_main.c:(.text+0x114c): undefined reference to `zap_vma_ptes'
> > drivers/infiniband/core/uverbs_main.o: In function `rdma_umap_open':
> > uverbs_main.c:(.text+0x53c): undefined reference to `zap_vma_ptes'
>
> I'm a bit surprised by this - the code is not new, it just got moved
> out of the drivers, and I don't see any protections there against
> CONFIG_MMU ??

I think I just never hit those in my randconfig builds. Out of the three
drivers, one is 64-bit only, and the others depend on PCI. The nommu
configurations I'm testing in randconfig are all 32-bit without PCI.

> > To fix this, we can either make uverbs depend on CONFIG_MMU, or try
> > to build it anyway. Since this is the only compile-time dependency,
> > I decided to allow building it with an extra compile-time check for
> > CONFIG_MMU before calling the one function.
>
> Hrm. So this code doesn't work at all on NOMMU, it relies on
> remap_pfn_range, which always fails on those kernels,
>
> Due to this of this the entirety of INFINIBAND_USER_ACCESS is broken.
>
> I think a kconfig change is the better way to go?

Works for me. Can you just commit that yourself with 'Reported-by:
Arnd Bergmann <arnd@arndb.de>' then?

      Arnd