Re: [PATCH?][arch/parisc/kernel/pci-dma.c] pcxl_dma_ops.alloc_noncoherent = pa11_dma_alloc_consistent?

Previous thread: Re: 2.6.24-git2: Oracle 11g VKTM process enters R state on startup and is unkillable [still broken in 2.6.25-rc1] by Alessandro Suardi on Monday, February 11, 2008 - 9:20 am. (13 messages)

Next thread: [PATCH] drivers/base: export gpl (un)register_memory_notifier by Jan-Bernd Themann on Monday, February 11, 2008 - 9:24 am. (14 messages)
From: Roel Kluin
Date: Monday, February 11, 2008 - 9:23 am

duplicate pa11_dma_alloc_consistent; more appropriate appears
pa11_dma_alloc_noncoherent here. 

Not tested, please confirm that this fix is correct
---
fix noncoherent allocation

Signed-off-by: Roel Kluin <12o3l@tiscali.nl>
---
diff --git a/arch/parisc/kernel/pci-dma.c b/arch/parisc/kernel/pci-dma.c
index 9448d4e..0e8b71f 100644
--- a/arch/parisc/kernel/pci-dma.c
+++ b/arch/parisc/kernel/pci-dma.c
@@ -549,7 +549,7 @@ static void pa11_dma_sync_sg_for_device(struct device *dev, struct scatterlist *
 struct hppa_dma_ops pcxl_dma_ops = {
 	.dma_supported =	pa11_dma_supported,
 	.alloc_consistent =	pa11_dma_alloc_consistent,
-	.alloc_noncoherent =	pa11_dma_alloc_consistent,
+	.alloc_noncoherent =	pa11_dma_alloc_noncoherent,
 	.free_consistent =	pa11_dma_free_consistent,
 	.map_single =		pa11_dma_map_single,
 	.unmap_single =		pa11_dma_unmap_single,
--

From: Matthew Wilcox
Date: Monday, February 11, 2008 - 9:42 am

I don't think it is.  The memories are fading, so I don't recall why it
is we do it this way, but I'm pretty sure it's correct the way it is.

-- 
Intel are signing my paycheques ... these opinions are still mine
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours.  We can't possibly take such
a retrograde step."
--

From: Roel Kluin
Date: Monday, February 11, 2008 - 9:45 am

Maybe this helps a bit: later I found that something similar occurs in 
drivers/parisc/{ccio-dma.c, sba_iommu.c}:

1010: .alloc_consistent =     ccio_alloc_consistent,
      .alloc_noncoherent =    ccio_alloc_consistent,

1036: .alloc_consistent =     sba_alloc_consistent,
      .alloc_noncoherent =    sba_alloc_consistent,

However in these files the functions {ccio_alloc, sba_alloc}_noncoherent
do not exist.
--

From: James Bottomley
Date: Monday, February 11, 2008 - 9:55 am

dma_alloc_noncoherent is a fallback for boxes with coherency problems
which cannot allocate coherent memory.  The API asks for coherent memory
but the driver promises to behave correctly even if an incoherent area
is returned.  dma_alloc_coherent() requires only coherent memory.  The
pcx boxes (PA7200 and below) can't do the uncached trick for coherent
memory, so the API is designed for them.  The PCXL can, so we allocate
coherent memory even for dma_alloc_noncoherent() in that case (and nop
out the coherency handlers the driver uses).

James


--

From: James Bottomley
Date: Monday, February 11, 2008 - 9:45 am

No, it looks completely incorrect to me.  What makes you think a pcxl
box has a problem with coherency?

James


--

From: Roel Kluin
Date: Monday, February 11, 2008 - 9:57 am

Ok, please ignore the patch then. It just appeared suspicious to me
that the function did exist, but the names assigned were different.
--

From: John David Anglin
Date: Monday, February 11, 2008 - 10:08 am

How about a comment?

Dave
-- 
J. David Anglin                                  dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)
--

From: Roel Kluin
Date: Monday, February 11, 2008 - 11:56 am

Based on James Bottomley's explanation maybe a comment like this?
---
Explain why dma_alloc_noncoherent is only used for boxes PA7200 and below

Signed-off-by: Roel Kluin <12o3l@tiscali.nl>
---
diff --git a/arch/parisc/kernel/pci-dma.c b/arch/parisc/kernel/pci-dma.c
index 9448d4e..fc3325a 100644
--- a/arch/parisc/kernel/pci-dma.c
+++ b/arch/parisc/kernel/pci-dma.c
@@ -567,6 +567,10 @@ static void *fail_alloc_consistent(struct device *dev, size_t size,
 	return NULL;
 }
 
+/*
+ * dma_alloc_noncoherent is a fallback for boxes PA7200 and below which
+ * cannot allocate coherent memory.
+ */
 static void *pa11_dma_alloc_noncoherent(struct device *dev, size_t size,
 					  dma_addr_t *dma_handle, gfp_t flag)
 {
@@ -586,6 +590,10 @@ static void pa11_dma_free_noncoherent(struct device *dev, size_t size,
 	return;
 }
 
+/*
+ * PCXL allocates coherent memory even for dma_alloc_noncoherent() due to the
+ * uncached trick for coherent memory.
+ */
 struct hppa_dma_ops pcx_dma_ops = {
 	.dma_supported =	pa11_dma_supported,
 	.alloc_consistent =	fail_alloc_consistent,

--

From: Kyle McMartin
Date: Monday, February 11, 2008 - 11:58 am

Why? This is located below pcxl_dma_ops so its obviously only relevant

--

From: Kyle McMartin
Date: Monday, February 11, 2008 - 12:10 pm

How about just doing something like this:

diff --git a/arch/parisc/kernel/pci-dma.c b/arch/parisc/kernel/pci-dma.c
index 9448d4e..63f9b7f 100644
--- a/arch/parisc/kernel/pci-dma.c
+++ b/arch/parisc/kernel/pci-dma.c
@@ -409,7 +409,7 @@ pcxl_dma_init(void)
 
 __initcall(pcxl_dma_init);
 
-static void * pa11_dma_alloc_consistent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag)
+static void * pcxl_dma_alloc_consistent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag)
 {
 	unsigned long vaddr;
 	unsigned long paddr;
@@ -435,7 +435,7 @@ static void * pa11_dma_alloc_consistent (struct device *dev, size_t size, dma_ad
 	return (void *)vaddr;
 }
 
-static void pa11_dma_free_consistent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle)
+static void pcxl_dma_free_consistent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle)
 {
 	int order;
 
@@ -548,9 +548,9 @@ static void pa11_dma_sync_sg_for_device(struct device *dev, struct scatterlist *
 
 struct hppa_dma_ops pcxl_dma_ops = {
 	.dma_supported =	pa11_dma_supported,
-	.alloc_consistent =	pa11_dma_alloc_consistent,
-	.alloc_noncoherent =	pa11_dma_alloc_consistent,
-	.free_consistent =	pa11_dma_free_consistent,
+	.alloc_consistent =	pcxl_dma_alloc_consistent,
+	.alloc_noncoherent =	pcxl_dma_alloc_consistent,
+	.free_consistent =	pcxl_dma_free_consistent,
 	.map_single =		pa11_dma_map_single,
 	.unmap_single =		pa11_dma_unmap_single,
 	.map_sg =		pa11_dma_map_sg,
@@ -567,7 +567,7 @@ static void *fail_alloc_consistent(struct device *dev, size_t size,
 	return NULL;
 }
 
-static void *pa11_dma_alloc_noncoherent(struct device *dev, size_t size,
+static void *pcx_dma_alloc_noncoherent(struct device *dev, size_t size,
 					  dma_addr_t *dma_handle, gfp_t flag)
 {
 	void *addr;
@@ -579,7 +579,7 @@ static void *pa11_dma_alloc_noncoherent(struct device *dev, size_t size,
 	return addr;
 }
 
-static void pa11_dma_free_noncoherent(struct device *dev, ...
Previous thread: Re: 2.6.24-git2: Oracle 11g VKTM process enters R state on startup and is unkillable [still broken in 2.6.25-rc1] by Alessandro Suardi on Monday, February 11, 2008 - 9:20 am. (13 messages)

Next thread: [PATCH] drivers/base: export gpl (un)register_memory_notifier by Jan-Bernd Themann on Monday, February 11, 2008 - 9:24 am. (14 messages)