Re: crypto/async_tx/* doesn't build on s390

Re: crypto/async_tx/* doesn't build on s390

[Cornelia Huck]

On Thu, 31 Jan 2008 12:49:00 -0700,
"Williams, Dan J" <dan.j.williams@intel.com> wrote:

> I am mistaken, the 'depends on ARCH...' precludes HAS_DMA.  Perhaps the compiler is emitting a call to async_tx_find_channel when it needs to be inline?  On x86 do_async_xor is successfully compiled away when CONFIG_DMA_ENGINE=n.

Just checked why it compiled for me on one box but not on the other and 
found that deactivating CONFIG_SECTION_MISMATCH makes it go away. Hmm...

Re: crypto/async_tx/* doesn't build on s390

[Dan Williams]

On Feb 1, 2008 4:37 AM, Cornelia Huck <cornelia.huck@de.ibm.com> wrote:
> On Thu, 31 Jan 2008 12:49:00 -0700,
> "Williams, Dan J" <dan.j.williams@intel.com> wrote:
>
> > I am mistaken, the 'depends on ARCH...' precludes HAS_DMA.  Perhaps the compiler is emitting a call to async_tx_find_channel when it needs to be inline?  On x86 do_async_xor is successfully compiled away when CONFIG_DMA_ENGINE=n.
>
> Just checked why it compiled for me on one box but not on the other and
> found that deactivating CONFIG_SECTION_MISMATCH makes it go away. Hmm...
>

Here is what I have come up with as a fix.

--
Dan

---
async_tx: prevent do_async_xor from compiling on !HAS_DMA archs

With the addition of -fno-inline in CONFIG_DEBUG_SECTION_MISMATCH
do_async_xor is no longer compiled away on !HAS_DMA archs like s390.
Other async_tx calls to the dma-api are already open coded inline.

Signed-off-by: Dan Williams <dan.j.williams@intel.com>
---

 crypto/async_tx/async_xor.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)


diff --git a/crypto/async_tx/async_xor.c b/crypto/async_tx/async_xor.c
index 2575f67..393f07d 100644
--- a/crypto/async_tx/async_xor.c
+++ b/crypto/async_tx/async_xor.c
@@ -41,6 +41,14 @@ do_async_xor(struct dma_async_tx_descriptor *tx,
struct dma_device *device,
        enum dma_data_direction dir;
        int i;

+       /* if this function is not inlined we need to prevent
+        * the rest of the routine from compiling on !HAS_DMA
+        * archs
+        */
+       #ifndef CONFIG_DMA_ENGINE
+       return;
+       #endif
+
        pr_debug("%s: len: %zu\n", __FUNCTION__, len);

        dir = (flags & ASYNC_TX_ASSUME_COHERENT) ?

Re: crypto/async_tx/* doesn't build on s390

[Heiko Carstens]

On Fri, Feb 01, 2008 at 02:31:28PM -0700, Dan Williams wrote:
> On Feb 1, 2008 4:37 AM, Cornelia Huck <cornelia.huck@de.ibm.com> wrote:
> > On Thu, 31 Jan 2008 12:49:00 -0700,
> > "Williams, Dan J" <dan.j.williams@intel.com> wrote:
> >
> > > I am mistaken, the 'depends on ARCH...' precludes HAS_DMA.  Perhaps the compiler is emitting a call to async_tx_find_channel when it needs to be inline?  On x86 do_async_xor is successfully compiled away when CONFIG_DMA_ENGINE=n.
> >
> > Just checked why it compiled for me on one box but not on the other and
> > found that deactivating CONFIG_SECTION_MISMATCH makes it go away. Hmm...
> >
> 
> Here is what I have come up with as a fix.

The fix works for me. Thanks! However your mailer replaced tabs with spaces
and added an extra line break.

> ---
> async_tx: prevent do_async_xor from compiling on !HAS_DMA archs
> 
> With the addition of -fno-inline in CONFIG_DEBUG_SECTION_MISMATCH
> do_async_xor is no longer compiled away on !HAS_DMA archs like s390.
> Other async_tx calls to the dma-api are already open coded inline.
> 
> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
> ---
> 
>  crypto/async_tx/async_xor.c |    8 ++++++++
>  1 files changed, 8 insertions(+), 0 deletions(-)
> 
> 
> diff --git a/crypto/async_tx/async_xor.c b/crypto/async_tx/async_xor.c
> index 2575f67..393f07d 100644
> --- a/crypto/async_tx/async_xor.c
> +++ b/crypto/async_tx/async_xor.c
> @@ -41,6 +41,14 @@ do_async_xor(struct dma_async_tx_descriptor *tx,
> struct dma_device *device,
>         enum dma_data_direction dir;
>         int i;
> 
> +       /* if this function is not inlined we need to prevent
> +        * the rest of the routine from compiling on !HAS_DMA
> +        * archs
> +        */
> +       #ifndef CONFIG_DMA_ENGINE
> +       return;
> +       #endif
> +
>         pr_debug("%s: len: %zu\n", __FUNCTION__, len);
> 
>         dir = (flags & ASYNC_TX_ASSUME_COHERENT) ?

Re: crypto/async_tx/* doesn't build on s390

[Dan Williams]

[-- Attachment #1: Type: text/plain, Size: 528 bytes --]

On Feb 3, 2008 4:40 AM, Heiko Carstens <heiko.carstens@de.ibm.com> wrote:
> On Fri, Feb 01, 2008 at 02:31:28PM -0700, Dan Williams wrote:
[..]
> The fix works for me. Thanks! However your mailer replaced tabs with spaces
> and added an extra line break.
>

The attached patch is slightly cleaner, and still works.  It marks
do_async_xor as '__always_inline'.

I have pushed this out to a git repository [1] and will be requesting
a pull later today.

--
Dan

[1] git://lost.foo-projects.org/~dwillia2/git/iop async-tx-for-linus

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: async-tx-fix-xor-compile.patch --]
[-- Type: text/x-patch; name=async-tx-fix-xor-compile.patch, Size: 995 bytes --]

async_tx: fix compile breakage, mark do_async_xor __always_inline

From: Dan Williams <dan.j.williams@intel.com>

do_async_xor must be compiled away on !HAS_DMA archs.

Signed-off-by: Dan Williams <dan.j.williams@intel.com>
---

 crypto/async_tx/async_xor.c |    6 +++++-
 1 files changed, 5 insertions(+), 1 deletions(-)


diff --git a/crypto/async_tx/async_xor.c b/crypto/async_tx/async_xor.c
index 2575f67..716885a 100644
--- a/crypto/async_tx/async_xor.c
+++ b/crypto/async_tx/async_xor.c
@@ -30,7 +30,11 @@
 #include <linux/raid/xor.h>
 #include <linux/async_tx.h>
 
-static void
+/* do_async_xor - dma map the pages and perform the xor with an engine.
+ * 	This routine is marked __always_inline so it can be compiled away
+ * 	when CONFIG_DMA_ENGINE=n
+ */
+static __always_inline void
 do_async_xor(struct dma_async_tx_descriptor *tx, struct dma_device *device,
 	struct dma_chan *chan, struct page *dest, struct page **src_list,
 	unsigned int offset, unsigned int src_cnt, size_t len,

Re: crypto/async_tx/* doesn't build on s390

[Cornelia Huck]

On Sun, 3 Feb 2008 07:10:50 -0700,
"Dan Williams" <dan.j.williams@intel.com> wrote:

> On Feb 3, 2008 4:40 AM, Heiko Carstens <heiko.carstens@de.ibm.com> wrote:
> > On Fri, Feb 01, 2008 at 02:31:28PM -0700, Dan Williams wrote:
> [..]
> > The fix works for me. Thanks! However your mailer replaced tabs with spaces
> > and added an extra line break.
> >
> 
> The attached patch is slightly cleaner, and still works.  It marks
> do_async_xor as '__always_inline'.
> 
I can confirm that the patch works for me. Thanks!

crypto/async_tx/* doesn't build on s390

[Heiko Carstens]

I get the following:

crypto/built-in.o: In function `do_async_xor':
async_xor.c:49: undefined reference to `dma_map_page'
async_xor.c:56: undefined reference to `dma_map_page'

This is mainly because s390 doesn't support DMA at all,
but these files get selected via MD_RAID456 anyway.

Any idea how to fix this?

Re: crypto/async_tx/* doesn't build on s390

[Cornelia Huck]

On Thu, 31 Jan 2008 15:50:58 +0100,
Heiko Carstens <heiko.carstens@de.ibm.com> wrote:

> I get the following:
> 
> crypto/built-in.o: In function `do_async_xor':
> async_xor.c:49: undefined reference to `dma_map_page'
> async_xor.c:56: undefined reference to `dma_map_page'
> 
> This is mainly because s390 doesn't support DMA at all,
> but these files get selected via MD_RAID456 anyway.

deja-vu: http://marc.info/?l=linux-kernel&m=118056140212764&w=2

and indeed, async_memcpy.c doesn't seem to suffer from this affliction.
async_xor.c needs a similar treatment; maybe async_memset.c as well
(didn't check).

(No, I didn't actually look at the code :))