Skip to contentSkip to conversation

New: [regression] performance drop for glx

Bug 10328 Applied Andrew Morton 5 participants 9 messages Mar 25, 2008 – Apr 28, 2008 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore
#1 lore raw
In reply to a message not in the archive

(switched to email. Please respond via emailed reply-to-all, not via the bugzilla web interface).

Quoted 17 lines · On Tue, 25 Mar 2008 15:11:15 -0700 (PDT) [email protected] wrote:

http://bugzilla.kernel.org/show_bug.cgi?id=10328

Summary: [regression] performance drop for glx
Product: Memory Management
Version: 2.5
KernelVersion: 2.6.25-rc6
Platform: All
OS/Version: Linux
Tree: Mainline
Status: NEW
Severity: normal
Priority: P1
Component: Other
AssignedTo: [email protected]
ReportedBy: [email protected]

Quoted 6 lines

after commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 i expirience some grafik related perfomance issues.

I used glxgears for test.
before this patch: 1281.005 FPS
and after: 765.000 FPS

It nearly halved.

Quoted 5 lines

latest tested commit a4083c9271e0a697278e089f2c0b9a95363ada0a still hase bad performance.

I use Pentium D with 2GB RAM, Grafick: i945G, ICH7

That's

: commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 : Author: Siddha, Suresh B <[email protected]>
: Date: Wed Jan 30 13:33:43 2008 +0100
:
: x86: set strong uncacheable where UC is really desired
:

:     Also use _PAGE_PWT for all the mappings which need uncache mapping.
:     Instead of existing PAT2 which is UC- (and can be overwritten by MTRRs),
:     we now use PAT3 which is strong uncacheable.
:     
:     This makes it consistent with pgprot_noncached()
:     
:     Signed-off-by: Suresh Siddha <[email protected]>
:     Signed-off-by: Ingo Molnar <[email protected]>
:     Signed-off-by: Thomas Gleixner <[email protected]>

:

#2 lore raw
Quoted 33 lines · Andrew Morton

(switched to email. Please respond via emailed reply-to-all, not via the bugzilla web interface).

Quoted 10 lines · On Tue, 25 Mar 2008 15:11:15 -0700 (PDT) [email protected] wrote:

http://bugzilla.kernel.org/show_bug.cgi?id=10328

Summary: [regression] performance drop for glx

after commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 i expirience some grafik related perfomance issues.

I used glxgears for test.
before this patch: 1281.005 FPS
and after: 765.000 FPS

It nearly halved.

Quoted 5 lines

latest tested commit a4083c9271e0a697278e089f2c0b9a95363ada0a still hase bad performance.

I use Pentium D with 2GB RAM, Grafick: i945G, ICH7

That's

: commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 : Author: Siddha, Suresh B <[email protected]>
: Date: Wed Jan 30 13:33:43 2008 +0100
:
: x86: set strong uncacheable where UC is really desired
:

Quoted 5 lines
:     Also use _PAGE_PWT for all the mappings which need uncache mapping.
:     Instead of existing PAT2 which is UC- (and can be overwritten by MTRRs),
:     we now use PAT3 which is strong uncacheable.

:
: This makes it consistent with pgprot_noncached()

Alexey, Can you please try the appended patch?

Andrew, can you please push the appended patch for 2.6.25? Thanks.
---

fb drivers are using ioremap()/ioremap_nocache(), followed by mtrr_add with WC attribute. Recent changes in page attribute code made both ioremap()/ioremap_nocache() mappings as UC (instead of previous UC-). This breaks the graphics performance, as the effective memory type is UC instead of expected WC.

The correct way to fix this is to add ioremap_wc() (which uses UC- in the absence of PAT kernel support and WC with PAT) and change all the fb drivers to use this new ioremap_wc() API.

We can take this correct and longer route for post 2.6.25. For now, revert back to the UC- behavior for ioremap/ioremap_nocache.

---

arch/x86/mm/ioremap.c +5 -1
index 4afaba0..794895c 100644
@@ -137,7 +137,11 @@ static void __iomem *__ioremap(resource_size_t phys_addr, unsigned long size,
 	switch (mode) {
 	case IOR_MODE_UNCACHED:
 	default:
-		prot = PAGE_KERNEL_NOCACHE;
+		/*
+		 * FIXME: we will use UC MINUS for now, as video fb drivers
+		 * depend on it. Upcoming ioremap_wc() will fix this behavior.
+		 */
+		prot = PAGE_KERNEL_UC_MINUS;
 		break;
 	case IOR_MODE_CACHED:
 		prot = PAGE_KERNEL;
include/asm-x86/pgtable.h +2 -0
index 174b877..9cf472a 100644
@@ -85,6 +85,7 @@ extern pteval_t __PAGE_KERNEL, __PAGE_KERNEL_EXEC;
 #define __PAGE_KERNEL_RX		(__PAGE_KERNEL_EXEC & ~_PAGE_RW)
 #define __PAGE_KERNEL_EXEC_NOCACHE	(__PAGE_KERNEL_EXEC | _PAGE_PCD | _PAGE_PWT)
 #define __PAGE_KERNEL_NOCACHE		(__PAGE_KERNEL | _PAGE_PCD | _PAGE_PWT)
+#define __PAGE_KERNEL_UC_MINUS		(__PAGE_KERNEL | _PAGE_PCD)
 #define __PAGE_KERNEL_VSYSCALL		(__PAGE_KERNEL_RX | _PAGE_USER)
 #define __PAGE_KERNEL_VSYSCALL_NOCACHE	(__PAGE_KERNEL_VSYSCALL | _PAGE_PCD | _PAGE_PWT)
 #define __PAGE_KERNEL_LARGE		(__PAGE_KERNEL | _PAGE_PSE)
@@ -101,6 +102,7 @@ extern pteval_t __PAGE_KERNEL, __PAGE_KERNEL_EXEC;
 #define PAGE_KERNEL_EXEC		MAKE_GLOBAL(__PAGE_KERNEL_EXEC)
 #define PAGE_KERNEL_RX			MAKE_GLOBAL(__PAGE_KERNEL_RX)
 #define PAGE_KERNEL_NOCACHE		MAKE_GLOBAL(__PAGE_KERNEL_NOCACHE)
+#define PAGE_KERNEL_UC_MINUS		MAKE_GLOBAL(__PAGE_KERNEL_UC_MINUS)
 #define PAGE_KERNEL_EXEC_NOCACHE	MAKE_GLOBAL(__PAGE_KERNEL_EXEC_NOCACHE)
 #define PAGE_KERNEL_LARGE		MAKE_GLOBAL(__PAGE_KERNEL_LARGE)
 #define PAGE_KERNEL_LARGE_EXEC		MAKE_GLOBAL(__PAGE_KERNEL_LARGE_EXEC)
#3 lore raw
Quoted 46 lines · Suresh Siddha
Quoted 25 lines · Andrew Morton

(switched to email. Please respond via emailed reply-to-all, not via the bugzilla web interface).

Quoted 10 lines · On Tue, 25 Mar 2008 15:11:15 -0700 (PDT) [email protected] wrote:

http://bugzilla.kernel.org/show_bug.cgi?id=10328

Summary: [regression] performance drop for glx

after commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 i expirience some grafik related perfomance issues.

I used glxgears for test.
before this patch: 1281.005 FPS
and after: 765.000 FPS

It nearly halved.

Quoted 5 lines

latest tested commit a4083c9271e0a697278e089f2c0b9a95363ada0a still hase bad performance.

I use Pentium D with 2GB RAM, Grafick: i945G, ICH7

That's

Quoted 8 lines

: commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 : Author: Siddha, Suresh B <[email protected]>
: Date: Wed Jan 30 13:33:43 2008 +0100
: : x86: set strong uncacheable where UC is really desired

:     :     Also use _PAGE_PWT for all the mappings which need uncache mapping.
:     Instead of existing PAT2 which is UC- (and can be overwritten by MTRRs),
:     we now use PAT3 which is strong uncacheable.
:     :     This makes it consistent with pgprot_noncached()

Alexey, Can you please try the appended patch?

Andrew, can you please push the appended patch for 2.6.25? Thanks.
---

fb drivers are using ioremap()/ioremap_nocache(), followed by mtrr_add with WC attribute. Recent changes in page attribute code made both ioremap()/ioremap_nocache() mappings as UC (instead of previous UC-). This breaks the graphics performance, as the effective memory type is UC instead of expected WC.

Quoted 6 lines

The correct way to fix this is to add ioremap_wc() (which uses UC- in the absence of PAT kernel support and WC with PAT) and change all the fb drivers to use this new ioremap_wc() API.

We can take this correct and longer route for post 2.6.25. For now, revert back to the UC- behavior for ioremap/ioremap_nocache.

I would still like to add an ioremap_wc() even in 2.6.25; even if it's for now identical to ioremap_nocache(). Better get the right API in place as soon as possible

#8 lore raw
In reply to Suresh Siddha
Quoted 6 lines · Suresh Siddha

fb drivers are using ioremap()/ioremap_nocache(), followed by mtrr_add with WC attribute. Recent changes in page attribute code made both ioremap()/ioremap_nocache() mappings as UC (instead of previous UC-). This breaks the graphics performance, as the effective memory type is UC instead of expected WC.

Quoted 6 lines

The correct way to fix this is to add ioremap_wc() (which uses UC- in the absence of PAT kernel support and WC with PAT) and change all the fb drivers to use this new ioremap_wc() API.

We can take this correct and longer route for post 2.6.25. For now, revert back to the UC- behavior for ioremap/ioremap_nocache.

thanks Suresh, applied.

Ingo

#9 lore raw
Quoted 9 lines · Ingo Molnar
Quoted 6 lines · Suresh Siddha

fb drivers are using ioremap()/ioremap_nocache(), followed by mtrr_add with WC attribute. Recent changes in page attribute code made both ioremap()/ioremap_nocache() mappings as UC (instead of previous UC-). This breaks the graphics performance, as the effective memory type is UC instead of expected WC.

Quoted 8 lines
Quoted 6 lines

The correct way to fix this is to add ioremap_wc() (which uses UC- in the absence of PAT kernel support and WC with PAT) and change all the fb drivers to use this new ioremap_wc() API.

We can take this correct and longer route for post 2.6.25. For now, revert back to the UC- behavior for ioremap/ioremap_nocache.

thanks Suresh, applied.

Well, we need to take care of set_memory_uc() aswell, as the previous version didn't fix Alexey's issue.

Alexey, can you please test and ack this, before Andrew/Ingo can push relevant bits to their trees. Thanks.
---

fb drivers are using ioremap()/ioremap_nocache(), followed by mtrr_add with WC attribute. Recent changes in page attribute code made both ioremap()/ioremap_nocache() mappings as UC (instead of previous UC-). This breaks the graphics performance, as the effective memory type is UC instead of expected WC.

The correct way to fix this is to add ioremap_wc() (which uses UC- in the absence of PAT kernel support and WC with PAT) and change all the drivers to use this new ioremap_wc() API.

We can take this correct and longer route for post 2.6.25. For now, revert back to the UC- behavior for ioremap/ioremap_nocache.

Cc Arjan van de Ven <[email protected]>
Cc Alexey Fisher <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Thomas Gleixner <[email protected]>

---

arch/x86/mm/ioremap.c +5 -1
index 4afaba0..794895c 100644
@@ -137,7 +137,11 @@ static void __iomem *__ioremap(resource_size_t phys_addr, unsigned long size,
 	switch (mode) {
 	case IOR_MODE_UNCACHED:
 	default:
-		prot = PAGE_KERNEL_NOCACHE;
+		/*
+		 * FIXME: we will use UC MINUS for now, as video fb drivers
+		 * depend on it. Upcoming ioremap_wc() will fix this behavior.
+		 */
+		prot = PAGE_KERNEL_UC_MINUS;
 		break;
 	case IOR_MODE_CACHED:
 		prot = PAGE_KERNEL;
arch/x86/mm/pageattr.c +1 -1
index 14e48b5..7b79f6b 100644
@@ -771,7 +771,7 @@ static inline int change_page_attr_clear(unsigned long addr, int numpages,
 int set_memory_uc(unsigned long addr, int numpages)
 {
 	return change_page_attr_set(addr, numpages,
-				    __pgprot(_PAGE_PCD | _PAGE_PWT));
+				    __pgprot(_PAGE_PCD));
 }
 EXPORT_SYMBOL(set_memory_uc);
 
include/asm-x86/pgtable.h +2 -0
index 174b877..9cf472a 100644
@@ -85,6 +85,7 @@ extern pteval_t __PAGE_KERNEL, __PAGE_KERNEL_EXEC;
 #define __PAGE_KERNEL_RX		(__PAGE_KERNEL_EXEC & ~_PAGE_RW)
 #define __PAGE_KERNEL_EXEC_NOCACHE	(__PAGE_KERNEL_EXEC | _PAGE_PCD | _PAGE_PWT)
 #define __PAGE_KERNEL_NOCACHE		(__PAGE_KERNEL | _PAGE_PCD | _PAGE_PWT)
+#define __PAGE_KERNEL_UC_MINUS		(__PAGE_KERNEL | _PAGE_PCD)
 #define __PAGE_KERNEL_VSYSCALL		(__PAGE_KERNEL_RX | _PAGE_USER)
 #define __PAGE_KERNEL_VSYSCALL_NOCACHE	(__PAGE_KERNEL_VSYSCALL | _PAGE_PCD | _PAGE_PWT)
 #define __PAGE_KERNEL_LARGE		(__PAGE_KERNEL | _PAGE_PSE)
@@ -101,6 +102,7 @@ extern pteval_t __PAGE_KERNEL, __PAGE_KERNEL_EXEC;
 #define PAGE_KERNEL_EXEC		MAKE_GLOBAL(__PAGE_KERNEL_EXEC)
 #define PAGE_KERNEL_RX			MAKE_GLOBAL(__PAGE_KERNEL_RX)
 #define PAGE_KERNEL_NOCACHE		MAKE_GLOBAL(__PAGE_KERNEL_NOCACHE)
+#define PAGE_KERNEL_UC_MINUS		MAKE_GLOBAL(__PAGE_KERNEL_UC_MINUS)
 #define PAGE_KERNEL_EXEC_NOCACHE	MAKE_GLOBAL(__PAGE_KERNEL_EXEC_NOCACHE)
 #define PAGE_KERNEL_LARGE		MAKE_GLOBAL(__PAGE_KERNEL_LARGE)
 #define PAGE_KERNEL_LARGE_EXEC		MAKE_GLOBAL(__PAGE_KERNEL_LARGE_EXEC)
#4 lore raw
In reply to Arjan van de Ven

Am Dienstag, den 25.03.2008, 21:42 -0700 schrieb Arjan van de Ven:

Quoted 39 lines
Quoted 38 lines · Suresh Siddha
Quoted 36 lines · Andrew Morton

(switched to email. Please respond via emailed reply-to-all, not via the bugzilla web interface).

Quoted 10 lines · On Tue, 25 Mar 2008 15:11:15 -0700 (PDT) [email protected] wrote:

http://bugzilla.kernel.org/show_bug.cgi?id=10328

Summary: [regression] performance drop for glx

after commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 i expirience some grafik related perfomance issues.

I used glxgears for test.
before this patch: 1281.005 FPS
and after: 765.000 FPS

It nearly halved.

Quoted 5 lines

latest tested commit a4083c9271e0a697278e089f2c0b9a95363ada0a still hase bad performance.

I use Pentium D with 2GB RAM, Grafick: i945G, ICH7

That's

: commit 4138cc3418f5eaa7524ff8e927102863f1ba0ea5 : Author: Siddha, Suresh B <[email protected]>
: Date: Wed Jan 30 13:33:43 2008 +0100
:
: x86: set strong uncacheable where UC is really desired
:

:     Also use _PAGE_PWT for all the mappings which need uncache mapping.
:     Instead of existing PAT2 which is UC- (and can be overwritten by MTRRs),
:     we now use PAT3 which is strong uncacheable.

:
: This makes it consistent with pgprot_noncached()

Quoted 5 lines
Quoted 5 lines

Alexey, Can you please try the appended patch?

Andrew, can you please push the appended patch for 2.6.25? Thanks.
---

This issue come again with commit:

commit 2e5d9c857d4e6c9e7b7d8c8c86a68a7842d213d6

Author: [email protected] <[email protected]>
Date:   Tue Mar 18 17:00:14 2008 -0700

x86: PAT infrastructure patch

Sets up pat_init() infrastructure.

PAT MSR has following setting.
PAT
|PCD
||PWT
|||
000 WB _PAGE_CACHE_WB
001 WC _PAGE_CACHE_WC
010 UC- _PAGE_CACHE_UC_MINUS
011 UC _PAGE_CACHE_UC

We are effectively changing WT from boot time setting to WC. UC_MINUS is used to provide backward compatibility to existing /dev/mem users(X).

reserve_memtype and free_memtype are new interfaces for maintaining
alias-free
mapping. It is currently implemented in a simple way with a linked list and not optimized. reserve and free tracks the effective memory type, as a
result
of PAT and MTRR setting rather than what is actually requested in PAT.

pat_init piggy backs on mtrr_init as the rules for setting both pat and
mtrr
are same.

#5 lore raw
Quoted 16 lines · Alexey Fisher
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Tue, 25 Mar 2008 15:11:15 -0700 (PDT) [email protected] wrote:

http://bugzilla.kernel.org/show_bug.cgi?id=10328

Quoted 7 lines
Quoted 7 lines
Quoted 1 line
Quoted 1 line

Summary: [regression] performance drop for glx

Alexey, Can you please try the appended patch?

Andrew, can you please push the appended patch for 2.6.25? Thanks.
---

This issue come again with commit:

Quoted 8 lines

commit 2e5d9c857d4e6c9e7b7d8c8c86a68a7842d213d6

Author: [email protected] <[email protected]>
Date:   Tue Mar 18 17:00:14 2008 -0700

x86: PAT infrastructure patch

Sets up pat_init() infrastructure.

Alexey, Sorry. I missed posting a fix in the post PAT series for this issue. Can you please try the appended patch and Ack?

copying X folks get their attention and their Ack on the patch and next steps outlined.

Thanks.
---

[patch] use UC minus for ioremap(), ioremap_nocache() and pci_mmap_page_range()

Use UC_MINUS for ioremap(), ioremap_nocache() instead of strong UC. Once all the X drivers move to ioremap_wc(), we can go back to strong UC semantics for ioremap() and ioremap_nocache().

To avoid attribute aliasing issues, pci_mmap_page_range() will also use UC_MINUS for default non write-combining mapping request.

Next steps:
a) change all the video drivers using ioremap() or ioremap_nocache() and adding WC MTTR using mttr_add() to ioremap_wc()

b) for strict usage, we can go back to strong uc semantics for ioremap() and ioremap_nocache() after some grace period for completing step-a.

c) user level X server needs to use the appropriate method for setting up WC mapping (like using resourceX_wc sysfs file instead of adding MTRR for WC and using /dev/mem or resourceX under /sys)

---

arch/x86/mm/ioremap.c +16 -4
index 3a4baf9..c053977 100644
@@ -176,11 +176,11 @@ static void __iomem *__ioremap(resource_size_t phys_addr, unsigned long size,
 		/*
 		 * Do not fallback to certain memory types with certain
 		 * requested type:
-		 * - request is uncached, return cannot be write-back
-		 * - request is uncached, return cannot be write-combine
+		 * - request is uc-, return cannot be write-back
+		 * - request is uc-, return cannot be write-combine
 		 * - request is write-combine, return cannot be write-back
 		 */
-		if ((prot_val == _PAGE_CACHE_UC &&
+		if ((prot_val == _PAGE_CACHE_UC_MINUS &&
 		     (new_prot_val == _PAGE_CACHE_WB ||
 		      new_prot_val == _PAGE_CACHE_WC)) ||
 		    (prot_val == _PAGE_CACHE_WC &&
@@ -201,6 +201,9 @@ static void __iomem *__ioremap(resource_size_t phys_addr, unsigned long size,
 	default:
 		prot = PAGE_KERNEL_NOCACHE;
 		break;
+	case _PAGE_CACHE_UC_MINUS:
+		prot = PAGE_KERNEL_UC_MINUS;
+		break;
 	case _PAGE_CACHE_WC:
 		prot = PAGE_KERNEL_WC;
 		break;
@@ -255,7 +258,16 @@ static void __iomem *__ioremap(resource_size_t phys_addr, unsigned long size,
  */
 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
 {
-	return __ioremap(phys_addr, size, _PAGE_CACHE_UC);
+	/*
+	 * Ideally, this should be:
+	 * 	pat_wc_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
+	 *
+	 * Till we fix all X drivers to use ioremap_wc(), we will use
+	 * UC MINUS.
+	 */
+	unsigned long val = _PAGE_CACHE_UC_MINUS;
+
+	return __ioremap(phys_addr, size, val);
 }
 EXPORT_SYMBOL(ioremap_nocache);
 
arch/x86/mm/pageattr.c +8 -2
index c29ebd0..ad3202c 100644
@@ -779,14 +779,20 @@ static inline int change_page_attr_clear(unsigned long addr, int numpages,
 
 int _set_memory_uc(unsigned long addr, int numpages)
 {
+	/*
+	 * for now UC MINUS. see comments in ioremap_nocache()
+	 */
 	return change_page_attr_set(addr, numpages,
-				    __pgprot(_PAGE_CACHE_UC));
+				    __pgprot(_PAGE_CACHE_UC_MINUS));
 }
 
 int set_memory_uc(unsigned long addr, int numpages)
 {
+	/*
+	 * for now UC MINUS. see comments in ioremap_nocache()
+	 */
 	if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
-	                    _PAGE_CACHE_UC, NULL))
+	                    _PAGE_CACHE_UC_MINUS, NULL))
 		return -EINVAL;
 
 	return _set_memory_uc(addr, numpages);
arch/x86/pci/i386.c +9 -3
index 94f6c73..8af0f0b 100644
@@ -301,6 +301,13 @@ int pci_mmap_page_range(struct pci_dev *dev, struct vm_area_struct *vma,
 	prot = pgprot_val(vma->vm_page_prot);
 	if (pat_wc_enabled && write_combine)
 		prot |= _PAGE_CACHE_WC;
+	else if (pat_wc_enabled)
+		/*
+		 * ioremap() and ioremap_nocache() defaults to UC MINUS for now.
+		 * To avoid attribute conflicts, request UC MINUS here
+		 * aswell.
+		 */
+		prot |= _PAGE_CACHE_UC_MINUS;
 	else if (boot_cpu_data.x86 > 3)
 		prot |= _PAGE_CACHE_UC;
 
@@ -319,9 +326,8 @@ int pci_mmap_page_range(struct pci_dev *dev, struct vm_area_struct *vma,
 		 * - request is uncached, return cannot be write-combine
 		 * - request is write-combine, return cannot be write-back
 		 */
-		if ((flags == _PAGE_CACHE_UC &&
-		     (new_flags == _PAGE_CACHE_WB ||
-		      new_flags == _PAGE_CACHE_WC)) ||
+		if ((flags == _PAGE_CACHE_UC_MINUS &&
+		     (new_flags == _PAGE_CACHE_WB)) ||
 		    (flags == _PAGE_CACHE_WC &&
 		     new_flags == _PAGE_CACHE_WB)) {
 			free_memtype(addr, addr+len);
#6 lore raw

Am Freitag, den 25.04.2008, 17:07 -0700 schrieb Suresh Siddha:

Quoted 31 lines · Alexey Fisher
Quoted 16 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines · On Tue, 25 Mar 2008 15:11:15 -0700 (PDT) [email protected] wrote:

http://bugzilla.kernel.org/show_bug.cgi?id=10328

Quoted 7 lines
Quoted 7 lines
Quoted 1 line
Quoted 1 line

Summary: [regression] performance drop for glx

Alexey, Can you please try the appended patch?

Andrew, can you please push the appended patch for 2.6.25? Thanks.
---

This issue come again with commit:

Quoted 8 lines

commit 2e5d9c857d4e6c9e7b7d8c8c86a68a7842d213d6

Author: [email protected] <[email protected]>
Date:   Tue Mar 18 17:00:14 2008 -0700

x86: PAT infrastructure patch

Sets up pat_init() infrastructure.

Alexey, Sorry. I missed posting a fix in the post PAT series for this issue. Can you please try the appended patch and Ack?

copying X folks get their attention and their Ack on the patch and next steps outlined.

Quoted 9 lines

Thanks.
---

[patch] use UC minus for ioremap(), ioremap_nocache() and pci_mmap_page_range()

Use UC_MINUS for ioremap(), ioremap_nocache() instead of strong UC. Once all the X drivers move to ioremap_wc(), we can go back to strong UC semantics for ioremap() and ioremap_nocache().

Thank you. This patch working.

Ack ?

Applied
#7 lore raw
In reply to Suresh Siddha
Quoted 5 lines · Suresh Siddha

Alexey, Sorry. I missed posting a fix in the post PAT series for this issue. Can you please try the appended patch and Ack?

copying X folks get their attention and their Ack on the patch and next steps outlined.

i queued up the patch below - hand-merged to x86.git.

Ingo

----------->
Subject: x86 PAT: fix performance drop for glx, use UC minus for ioremap(), ioremap_nocache() and pci_mmap_page_range()
From: Suresh Siddha <[email protected]>
Date: Fri, 25 Apr 2008 17:07:22 -0700

Use UC_MINUS for ioremap(), ioremap_nocache() instead of strong UC. Once all the X drivers move to ioremap_wc(), we can go back to strong UC semantics for ioremap() and ioremap_nocache().

To avoid attribute aliasing issues, pci_mmap_page_range() will also use UC_MINUS for default non write-combining mapping request.

Next steps:
a) change all the video drivers using ioremap() or ioremap_nocache() and adding WC MTTR using mttr_add() to ioremap_wc()

b) for strict usage, we can go back to strong uc semantics for ioremap() and ioremap_nocache() after some grace period for completing step-a.

c) user level X server needs to use the appropriate method for setting up WC mapping (like using resourceX_wc sysfs file instead of adding MTRR for WC and using /dev/mem or resourceX under /sys)

---
 arch/x86/mm/ioremap.c  |   20 ++++++++++++++++----
 arch/x86/mm/pageattr.c |   10 ++++++++--
 arch/x86/pci/i386.c    |   12 +++++++++---
 3 files changed, 33 insertions(+), 9 deletions(-)
arch/x86/mm/ioremap.c +16 -4
Index: linux-x86.q/arch/x86/mm/ioremap.c
===================================================================
@@ -176,11 +176,11 @@ static void __iomem *__ioremap(resource_
 		/*
 		 * Do not fallback to certain memory types with certain
 		 * requested type:
-		 * - request is uncached, return cannot be write-back
-		 * - request is uncached, return cannot be write-combine
+		 * - request is uc-, return cannot be write-back
+		 * - request is uc-, return cannot be write-combine
 		 * - request is write-combine, return cannot be write-back
 		 */
-		if ((prot_val == _PAGE_CACHE_UC &&
+		if ((prot_val == _PAGE_CACHE_UC_MINUS &&
 		     (new_prot_val == _PAGE_CACHE_WB ||
 		      new_prot_val == _PAGE_CACHE_WC)) ||
 		    (prot_val == _PAGE_CACHE_WC &&
@@ -201,6 +201,9 @@ static void __iomem *__ioremap(resource_
 	default:
 		prot = PAGE_KERNEL_NOCACHE;
 		break;
+	case _PAGE_CACHE_UC_MINUS:
+		prot = PAGE_KERNEL_UC_MINUS;
+		break;
 	case _PAGE_CACHE_WC:
 		prot = PAGE_KERNEL_WC;
 		break;
@@ -255,7 +258,16 @@ static void __iomem *__ioremap(resource_
  */
 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
 {
-	return __ioremap(phys_addr, size, _PAGE_CACHE_UC);
+	/*
+	 * Ideally, this should be:
+	 * 	pat_wc_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
+	 *
+	 * Till we fix all X drivers to use ioremap_wc(), we will use
+	 * UC MINUS.
+	 */
+	unsigned long val = _PAGE_CACHE_UC_MINUS;
+
+	return __ioremap(phys_addr, size, val);
 }
 EXPORT_SYMBOL(ioremap_nocache);
 
arch/x86/mm/pageattr.c +8 -2
Index: linux-x86.q/arch/x86/mm/pageattr.c
===================================================================
@@ -829,14 +829,20 @@ static inline int change_page_attr_clear
 
 int _set_memory_uc(unsigned long addr, int numpages)
 {
+	/*
+	 * for now UC MINUS. see comments in ioremap_nocache()
+	 */
 	return change_page_attr_set(&addr, numpages,
-				    __pgprot(_PAGE_CACHE_UC), 0);
+				    __pgprot(_PAGE_CACHE_UC_MINUS), 0);
 }
 
 int set_memory_uc(unsigned long addr, int numpages)
 {
+	/*
+	 * for now UC MINUS. see comments in ioremap_nocache()
+	 */
 	if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
-	                    _PAGE_CACHE_UC, NULL))
+			    _PAGE_CACHE_UC_MINUS, NULL))
 		return -EINVAL;
 
 	return _set_memory_uc(addr, numpages);
arch/x86/pci/i386.c +9 -3
Index: linux-x86.q/arch/x86/pci/i386.c
===================================================================
@@ -301,6 +301,13 @@ int pci_mmap_page_range(struct pci_dev *
 	prot = pgprot_val(vma->vm_page_prot);
 	if (pat_wc_enabled && write_combine)
 		prot |= _PAGE_CACHE_WC;
+	else if (pat_wc_enabled)
+		/*
+		 * ioremap() and ioremap_nocache() defaults to UC MINUS for now.
+		 * To avoid attribute conflicts, request UC MINUS here
+		 * aswell.
+		 */
+		prot |= _PAGE_CACHE_UC_MINUS;
 	else if (boot_cpu_data.x86 > 3)
 		prot |= _PAGE_CACHE_UC;
 
@@ -319,9 +326,8 @@ int pci_mmap_page_range(struct pci_dev *
 		 * - request is uncached, return cannot be write-combine
 		 * - request is write-combine, return cannot be write-back
 		 */
-		if ((flags == _PAGE_CACHE_UC &&
-		     (new_flags == _PAGE_CACHE_WB ||
-		      new_flags == _PAGE_CACHE_WC)) ||
+		if ((flags == _PAGE_CACHE_UC_MINUS &&
+		     (new_flags == _PAGE_CACHE_WB)) ||
 		    (flags == _PAGE_CACHE_WC &&
 		     new_flags == _PAGE_CACHE_WB)) {
 			free_memtype(addr, addr+len);