[PATCH 0/3] some small GART cleanups

[PATCH 0/3] some small GART cleanups

[Joerg Roedel]

Hi,

this small patch series contains some cleanups for the K8 GART driver in
the x86 architecture.

diffstat:

 arch/x86/kernel/pci-gart_64.c |   38 ++++++++++++++++++--------------------
 include/asm-x86/gart.h        |    2 ++
 2 files changed, 20 insertions(+), 20 deletions(-)

[PATCH 1/3] x86/iommu: make GART driver checkpatch clean

[Joerg Roedel]

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/pci-gart_64.c |   17 +++++++++--------
 include/asm-x86/gart.h        |    2 ++
 2 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/arch/x86/kernel/pci-gart_64.c b/arch/x86/kernel/pci-gart_64.c
index b85a2f9..aa569db 100644
--- a/arch/x86/kernel/pci-gart_64.c
+++ b/arch/x86/kernel/pci-gart_64.c
@@ -27,8 +27,8 @@
 #include <linux/scatterlist.h>
 #include <linux/iommu-helper.h>
 #include <linux/sysdev.h>
+#include <linux/io.h>
 #include <asm/atomic.h>
-#include <asm/io.h>
 #include <asm/mtrr.h>
 #include <asm/pgtable.h>
 #include <asm/proto.h>
@@ -175,7 +175,8 @@ static void dump_leak(void)
 	       iommu_leak_pages);
 	for (i = 0; i < iommu_leak_pages; i += 2) {
 		printk(KERN_DEBUG "%lu: ", iommu_pages-i);
-		printk_address((unsigned long) iommu_leak_tab[iommu_pages-i], 0);
+		printk_address((unsigned long) iommu_leak_tab[iommu_pages-i],
+				0);
 		printk(KERN_CONT "%c", (i+1)%2 == 0 ? '\n' : ' ');
 	}
 	printk(KERN_DEBUG "\n");
@@ -688,7 +689,8 @@ static __init int init_k8_gatt(struct agp_kern_info *info)
 	if (!error)
 		error = sysdev_register(&device_gart);
 	if (error)
-		panic("Could not register gart_sysdev -- would corrupt data on next suspend");
+		panic("Could not register gart_sysdev -- "
+		      "would corrupt data on next suspend");
 
 	flush_gart();
 
@@ -710,8 +712,6 @@ static __init int init_k8_gatt(struct agp_kern_info *info)
 	return -1;
 }
 
-extern int agp_amd64_init(void);
-
 static struct dma_mapping_ops gart_dma_ops = {
 	.map_single			= gart_map_single,
 	.unmap_single			= gart_unmap_single,
@@ -777,8 +777,8 @@ void __init gart_iommu_init(void)
 	    (no_agp && init_k8_gatt(&info) < 0)) {
 		if (max_pfn > MAX_DMA32_PFN) {
 			printk(KERN_WARNING "More than 4GB of memory "
-			       	          "but GART IOMMU not available.\n"
-			       KERN_WARNING "falling back to iommu=soft.\n");
+			       "but GART IOMMU not available.\n");
+			printk(KERN_WARNING "falling back to iommu=soft.\n");
 		}
 		return;
 	}
@@ -868,7 +868,8 @@ void __init gart_parse_options(char *p)
 	if (!strncmp(p, "leak", 4)) {
 		leak_trace = 1;
 		p += 4;
-		if (*p == '=') ++p;
+		if (*p == '=')
+			++p;
 		if (isdigit(*p) && get_option(&p, &arg))
 			iommu_leak_pages = arg;
 	}
diff --git a/include/asm-x86/gart.h b/include/asm-x86/gart.h
index 3f62a83..cdf4f78 100644
--- a/include/asm-x86/gart.h
+++ b/include/asm-x86/gart.h
@@ -29,6 +29,8 @@ extern int fix_aperture;
 #define AMD64_GARTCACHECTL	0x9c
 #define AMD64_GARTEN		(1<<0)
 
+extern int agp_amd64_init(void);
+
 static inline void enable_gart_translation(struct pci_dev *dev, u64 addr)
 {
 	u32 tmp, ctl;
-- 
1.5.6.4

[PATCH 2/3] x86/iommu: convert GART need_flush to bool

[Joerg Roedel]

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/pci-gart_64.c |   10 +++++-----
 1 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/arch/x86/kernel/pci-gart_64.c b/arch/x86/kernel/pci-gart_64.c
index aa569db..aecea06 100644
--- a/arch/x86/kernel/pci-gart_64.c
+++ b/arch/x86/kernel/pci-gart_64.c
@@ -80,7 +80,7 @@ AGPEXTERN int agp_memory_reserved;
 AGPEXTERN __u32 *agp_gatt_table;
 
 static unsigned long next_bit;  /* protected by iommu_bitmap_lock */
-static int need_flush;		/* global flush state. set for each gart wrap */
+static bool need_flush;		/* global flush state. set for each gart wrap */
 
 static unsigned long alloc_iommu(struct device *dev, int size,
 				 unsigned long align_mask)
@@ -98,7 +98,7 @@ static unsigned long alloc_iommu(struct device *dev, int size,
 	offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
 				  size, base_index, boundary_size, align_mask);
 	if (offset == -1) {
-		need_flush = 1;
+		need_flush = true;
 		offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
 					  size, base_index, boundary_size,
 					  align_mask);
@@ -107,11 +107,11 @@ static unsigned long alloc_iommu(struct device *dev, int size,
 		next_bit = offset+size;
 		if (next_bit >= iommu_pages) {
 			next_bit = 0;
-			need_flush = 1;
+			need_flush = true;
 		}
 	}
 	if (iommu_fullflush)
-		need_flush = 1;
+		need_flush = true;
 	spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
 
 	return offset;
@@ -136,7 +136,7 @@ static void flush_gart(void)
 	spin_lock_irqsave(&iommu_bitmap_lock, flags);
 	if (need_flush) {
 		k8_flush_garts();
-		need_flush = 0;
+		need_flush = false;
 	}
 	spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
 }
-- 
1.5.6.4

[PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Joerg Roedel]

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/pci-gart_64.c |   11 ++++-------
 1 files changed, 4 insertions(+), 7 deletions(-)

diff --git a/arch/x86/kernel/pci-gart_64.c b/arch/x86/kernel/pci-gart_64.c
index aecea06..84d541f 100644
--- a/arch/x86/kernel/pci-gart_64.c
+++ b/arch/x86/kernel/pci-gart_64.c
@@ -674,13 +674,13 @@ static __init int init_k8_gatt(struct agp_kern_info *info)
 	info->aper_size = aper_size >> 20;
 
 	gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
-	gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
+	gatt = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
+					get_order(gatt_size));
 	if (!gatt)
 		panic("Cannot allocate GATT table");
 	if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
 		panic("Could not set GART PTEs to uncacheable pages");
 
-	memset(gatt, 0, gatt_size);
 	agp_gatt_table = gatt;
 
 	enable_gart_translations();
@@ -788,18 +788,15 @@ void __init gart_iommu_init(void)
 	iommu_size = check_iommu_size(info.aper_base, aper_size);
 	iommu_pages = iommu_size >> PAGE_SHIFT;
 
-	iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
+	iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
 						      get_order(iommu_pages/8));
 	if (!iommu_gart_bitmap)
 		panic("Cannot allocate iommu bitmap\n");
-	memset(iommu_gart_bitmap, 0, iommu_pages/8);
 
 #ifdef CONFIG_IOMMU_LEAK
 	if (leak_trace) {
-		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
+		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
 				  get_order(iommu_pages*sizeof(void *)));
-		if (iommu_leak_tab)
-			memset(iommu_leak_tab, 0, iommu_pages * 8);
 		else
 			printk(KERN_DEBUG
 			       "PCI-DMA: Cannot allocate leak trace area\n");
-- 
1.5.6.4

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Ingo Molnar]

* Joerg Roedel <joerg.roedel@amd.com> wrote:

>  #ifdef CONFIG_IOMMU_LEAK
>  	if (leak_trace) {
> -		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
> +		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
>  				  get_order(iommu_pages*sizeof(void *)));
> -		if (iommu_leak_tab)
> -			memset(iommu_leak_tab, 0, iommu_pages * 8);
>  		else
>  			printk(KERN_DEBUG
>  			       "PCI-DMA: Cannot allocate leak trace area\n");

hm, that looks broken.

	Ingo

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Joerg Roedel]

On Thu, Sep 25, 2008 at 12:20:12PM +0200, Ingo Molnar wrote:
> 
> * Joerg Roedel <joerg.roedel@amd.com> wrote:
> 
> >  #ifdef CONFIG_IOMMU_LEAK
> >  	if (leak_trace) {
> > -		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
> > +		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
> >  				  get_order(iommu_pages*sizeof(void *)));
> > -		if (iommu_leak_tab)
> > -			memset(iommu_leak_tab, 0, iommu_pages * 8);
> >  		else
> >  			printk(KERN_DEBUG
> >  			       "PCI-DMA: Cannot allocate leak trace area\n");
> 
> hm, that looks broken.

Argh, yes. Please try this one:

=
>From c402fd7f5fe930d45a02ef863bf3fb61851e6ad0 Mon Sep 17 00:00:00 2001
From: Joerg Roedel <joerg.roedel@amd.com>
Date: Thu, 25 Sep 2008 12:39:06 +0200
Subject: [PATCH] x86/iommu: use __GFP_ZERO instead of memset for GART

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/pci-gart_64.c |   13 +++++--------
 1 files changed, 5 insertions(+), 8 deletions(-)

diff --git a/arch/x86/kernel/pci-gart_64.c b/arch/x86/kernel/pci-gart_64.c
index aecea06..d077116 100644
--- a/arch/x86/kernel/pci-gart_64.c
+++ b/arch/x86/kernel/pci-gart_64.c
@@ -674,13 +674,13 @@ static __init int init_k8_gatt(struct agp_kern_info *info)
 	info->aper_size = aper_size >> 20;
 
 	gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
-	gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
+	gatt = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
+					get_order(gatt_size));
 	if (!gatt)
 		panic("Cannot allocate GATT table");
 	if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
 		panic("Could not set GART PTEs to uncacheable pages");
 
-	memset(gatt, 0, gatt_size);
 	agp_gatt_table = gatt;
 
 	enable_gart_translations();
@@ -788,19 +788,16 @@ void __init gart_iommu_init(void)
 	iommu_size = check_iommu_size(info.aper_base, aper_size);
 	iommu_pages = iommu_size >> PAGE_SHIFT;
 
-	iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
+	iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
 						      get_order(iommu_pages/8));
 	if (!iommu_gart_bitmap)
 		panic("Cannot allocate iommu bitmap\n");
-	memset(iommu_gart_bitmap, 0, iommu_pages/8);
 
 #ifdef CONFIG_IOMMU_LEAK
 	if (leak_trace) {
-		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
+		iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
 				  get_order(iommu_pages*sizeof(void *)));
-		if (iommu_leak_tab)
-			memset(iommu_leak_tab, 0, iommu_pages * 8);
-		else
+		if (!iommu_leak_tab)
 			printk(KERN_DEBUG
 			       "PCI-DMA: Cannot allocate leak trace area\n");
 	}
-- 
1.5.6.4


-- 
           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Ingo Molnar]

* Joerg Roedel <joerg.roedel@amd.com> wrote:

> > > -		if (iommu_leak_tab)
> > > -			memset(iommu_leak_tab, 0, iommu_pages * 8);
> > >  		else
> > >  			printk(KERN_DEBUG
> > >  			       "PCI-DMA: Cannot allocate leak trace area\n");
> > 
> > hm, that looks broken.
> 
> Argh, yes. Please try this one:

applied these patches to tip/x86/iommu:

 0114267: x86/iommu: use __GFP_ZERO instead of memset for GART
 3610f21: x86/iommu: convert GART need_flush to bool
 237a622: x86/iommu: make GART driver checkpatch clean

thanks Joerg!

Fujita, do they look fine to you too?

	Ingo

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Ingo Molnar]

another thing:

How hard would it be to add an CONFIG_IOMMU_DEBUG option that forces as 
many DMA requests to go via the IOMMU as possible?

This slows things down of course so it's only for debugging - but it 
also makes sure that we utilize the IOMMU code to the maximum - which is 
not normally the case.

Would be nice to have it .config driven (default-disabled), so that 
-tip's randconfig testing can stumble upon it every now and then. I've 
got GART test-systems - this way we could find certain types of IOMMU 
breakages sooner.

	Ingo

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[FUJITA Tomonori]

On Sat, 27 Sep 2008 20:14:38 +0200
Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Joerg Roedel <joerg.roedel@amd.com> wrote:
> 
> > > > -		if (iommu_leak_tab)
> > > > -			memset(iommu_leak_tab, 0, iommu_pages * 8);
> > > >  		else
> > > >  			printk(KERN_DEBUG
> > > >  			       "PCI-DMA: Cannot allocate leak trace area\n");
> > > 
> > > hm, that looks broken.
> > 
> > Argh, yes. Please try this one:
> 
> applied these patches to tip/x86/iommu:
> 
>  0114267: x86/iommu: use __GFP_ZERO instead of memset for GART
>  3610f21: x86/iommu: convert GART need_flush to bool
>  237a622: x86/iommu: make GART driver checkpatch clean
> 
> thanks Joerg!
> 
> Fujita, do they look fine to you too?

Yeah, all the patches look trivial and fine.

Thanks,

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[FUJITA Tomonori]

On Sat, 27 Sep 2008 20:18:55 +0200
Ingo Molnar <mingo@elte.hu> wrote:

> another thing:
> 
> How hard would it be to add an CONFIG_IOMMU_DEBUG option that forces as 
> many DMA requests to go via the IOMMU as possible?
> 
> This slows things down of course so it's only for debugging - but it 
> also makes sure that we utilize the IOMMU code to the maximum - which is 
> not normally the case.

If you use iommu=force boot option, GART always tries to use the
IOMMU.

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Joerg Roedel]

On Sat, Sep 27, 2008 at 08:18:55PM +0200, Ingo Molnar wrote:
> 
> another thing:
> 
> How hard would it be to add an CONFIG_IOMMU_DEBUG option that forces as 
> many DMA requests to go via the IOMMU as possible?
> 
> This slows things down of course so it's only for debugging - but it 
> also makes sure that we utilize the IOMMU code to the maximum - which is 
> not normally the case.
> 
> Would be nice to have it .config driven (default-disabled), so that 
> -tip's randconfig testing can stumble upon it every now and then. I've 
> got GART test-systems - this way we could find certain types of IOMMU 
> breakages sooner.

For AMD IOMMU I disabled the round-robin allocator to stress-test the
code. This means that the address allocation bitmap is always traversed
from the first bit. In consequence the TLB flushing is stressed a lot
(both in hardware and software) because the same DMA addresses are used
again and again. For testing I hardcoded it into the driver but I can
also make it depend on CONFIG_IOMMU_DEBUG.

Joerg

-- 
           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Ingo Molnar]

* Joerg Roedel <joerg.roedel@amd.com> wrote:

> On Sat, Sep 27, 2008 at 08:18:55PM +0200, Ingo Molnar wrote:
> > 
> > another thing:
> > 
> > How hard would it be to add an CONFIG_IOMMU_DEBUG option that forces as 
> > many DMA requests to go via the IOMMU as possible?
> > 
> > This slows things down of course so it's only for debugging - but it 
> > also makes sure that we utilize the IOMMU code to the maximum - which is 
> > not normally the case.
> > 
> > Would be nice to have it .config driven (default-disabled), so that 
> > -tip's randconfig testing can stumble upon it every now and then. I've 
> > got GART test-systems - this way we could find certain types of IOMMU 
> > breakages sooner.
> 
> For AMD IOMMU I disabled the round-robin allocator to stress-test the 
> code. This means that the address allocation bitmap is always 
> traversed from the first bit. In consequence the TLB flushing is 
> stressed a lot (both in hardware and software) because the same DMA 
> addresses are used again and again. For testing I hardcoded it into 
> the driver but I can also make it depend on CONFIG_IOMMU_DEBUG.

yes - could you please make a new option for it, 
CONFIG_IOMMU_DEBUG_FORCE=y or so - and cover all iommus that support it?

	Ingo

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[FUJITA Tomonori]

On Mon, 29 Sep 2008 10:56:47 +0200
Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Joerg Roedel <joerg.roedel@amd.com> wrote:
> 
> > On Sat, Sep 27, 2008 at 08:18:55PM +0200, Ingo Molnar wrote:
> > > 
> > > another thing:
> > > 
> > > How hard would it be to add an CONFIG_IOMMU_DEBUG option that forces as 
> > > many DMA requests to go via the IOMMU as possible?
> > > 
> > > This slows things down of course so it's only for debugging - but it 
> > > also makes sure that we utilize the IOMMU code to the maximum - which is 
> > > not normally the case.
> > > 
> > > Would be nice to have it .config driven (default-disabled), so that 
> > > -tip's randconfig testing can stumble upon it every now and then. I've 
> > > got GART test-systems - this way we could find certain types of IOMMU 
> > > breakages sooner.
> > 
> > For AMD IOMMU I disabled the round-robin allocator to stress-test the 
> > code. This means that the address allocation bitmap is always 
> > traversed from the first bit. In consequence the TLB flushing is 
> > stressed a lot (both in hardware and software) because the same DMA 
> > addresses are used again and again. For testing I hardcoded it into 
> > the driver but I can also make it depend on CONFIG_IOMMU_DEBUG.
> 
> yes - could you please make a new option for it, 
> CONFIG_IOMMU_DEBUG_FORCE=y or so - and cover all iommus that support it?

I'm not sure we are talking about the same thing. Surely, I and Joerg
are talking different things.

GART driver doesn't need to use the IOMMU hardware at all times. GART
does a virtual mapping only when necessary (a device needs to handle
an address that it can't access to). But as I wrote, if you use
iommu=force, GART driver always use the IOMMU hardware.


Other x86 IOMMU drivers always use the IOMMU hardware. Except for
Intel VT-D, they manage free virtual I/O space in the round-robin
manner with the bitmap algorithm to avoid frequent IOTLB flush. Joerg
said he tested AMD IOMMU driver with the round-robin manner disabled
so AMD IOMMU driver uses the same virtual I/O space with lots of IOTLB
flush.

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[Ingo Molnar]

* FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp> wrote:

> > yes - could you please make a new option for it, 
> > CONFIG_IOMMU_DEBUG_FORCE=y or so - and cover all iommus that support 
> > it?
> 
> I'm not sure we are talking about the same thing. Surely, I and Joerg 
> are talking different things.
> 
> GART driver doesn't need to use the IOMMU hardware at all times. GART 
> does a virtual mapping only when necessary (a device needs to handle 
> an address that it can't access to). But as I wrote, if you use 
> iommu=force, GART driver always use the IOMMU hardware.

yes - but iommu=force is not randconfig covered, hence it never gets 
tested by -tip testing. So my suggestion was a really simple patch: a 
new Kconfig entry that makes iommu=force default. 

CONFIG_BOOTPARAM_IOMMU_FORCE=y would be the right name for it. (disabled 
by default, of course)

> Other x86 IOMMU drivers always use the IOMMU hardware. Except for 
> Intel VT-D, they manage free virtual I/O space in the round-robin 
> manner with the bitmap algorithm to avoid frequent IOTLB flush. Joerg 
> said he tested AMD IOMMU driver with the round-robin manner disabled 
> so AMD IOMMU driver uses the same virtual I/O space with lots of IOTLB 
> flush.

all i'm suggesting is to please expose existing debug capabilities in 
the .config space, so that it can be tested in automated setups.

	Ingo

Re: [PATCH 3/3] x86/iommu: use __GFP_ZERO instead of memset for GART

[H. Peter Anvin]

Ingo Molnar wrote:
> 
> all i'm suggesting is to please expose existing debug capabilities in 
> the .config space, so that it can be tested in automated setups.
> 

Boot configurations, too, can be tested in automated setups, and there 
is benefit to being able to do that without compilation.  However, what 
we'd need is a machine-readable description that one can select a 
configuration from.

	-hpa