Skip to contentSkip to conversation

agp: don't lock pages

PATCH In review Nick Piggin 4 participants 7 messages Jul 25, 2007 – Jul 26, 2007 Agpgart driver

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 agp: don't lock pages

Hi,

Does this patch solve the X problem? Does anyone see anything wrong with it or know why agp was locking the pages?

--
AGP should not need to lock pages. They are not protecting any race because there is no lock_page calls, only SetPageLocked.

This is causing hangs with d00806b183152af6d24f46f0c33f14162ca1262a.

drivers/char/agp/generic.c +0 -2
index d535c40..3db4f40 100644
@@ -1170,7 +1170,6 @@ void *agp_generic_alloc_page(struct agp_
 	map_page_into_agp(page);
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -1187,7 +1186,6 @@ void agp_generic_destroy_page(void *addr
 	page = virt_to_page(addr);
 	unmap_page_from_agp(page);
 	put_page(page);
-	unlock_page(page);
 	free_page((unsigned long)addr);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/intel-agp.c +0 -2
index a124060..2f319f4 100644
@@ -213,7 +213,6 @@ static void *i8xx_alloc_pages(void)
 	}
 	global_flush_tlb();
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -229,7 +228,6 @@ static void i8xx_destroy_pages(void *add
 	change_page_attr(page, 4, PAGE_KERNEL);
 	global_flush_tlb();
 	put_page(page);
-	unlock_page(page);
 	__free_pages(page, 2);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/sgi-agp.c +0 -1
index cda608c..98cf8ab 100644
@@ -51,7 +51,6 @@ static void *sgi_tioca_alloc_page(struct
 		return NULL;
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
#1 lore raw
Quoted 4 lines · Nick Piggin

Hi,

Does this patch solve the X problem? Does anyone see anything wrong with it or know why agp was locking the pages?

Yeap the mac mini is happy again
Thanks

Quoted 66 lines

--
AGP should not need to lock pages. They are not protecting any race because there is no lock_page calls, only SetPageLocked.

This is causing hangs with d00806b183152af6d24f46f0c33f14162ca1262a.

drivers/char/agp/generic.c +0 -2
index d535c40..3db4f40 100644
@@ -1170,7 +1170,6 @@ void *agp_generic_alloc_page(struct agp_
        map_page_into_agp(page);
 
        get_page(page);
-       SetPageLocked(page);
        atomic_inc(&agp_bridge->current_memory_agp);
        return page_address(page);
 }
@@ -1187,7 +1186,6 @@ void agp_generic_destroy_page(void *addr
        page = virt_to_page(addr);
        unmap_page_from_agp(page);
        put_page(page);
-       unlock_page(page);
        free_page((unsigned long)addr);
        atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/intel-agp.c +0 -2
index a124060..2f319f4 100644
@@ -213,7 +213,6 @@ static void *i8xx_alloc_pages(void)
        }
        global_flush_tlb();
        get_page(page);
-       SetPageLocked(page);
        atomic_inc(&agp_bridge->current_memory_agp);
        return page_address(page);
 }
@@ -229,7 +228,6 @@ static void i8xx_destroy_pages(void *add
        change_page_attr(page, 4, PAGE_KERNEL);
        global_flush_tlb();
        put_page(page);
-       unlock_page(page);
        __free_pages(page, 2);
        atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/sgi-agp.c +0 -2
index cda608c..98cf8ab 100644
@@ -51,7 +51,6 @@ static void *sgi_tioca_alloc_page(struct
                return NULL;
 
        get_page(page);
-       SetPageLocked(page);
        atomic_inc(&agp_bridge->current_memory_agp);
        return page_address(page);
 }
-

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#2 lore raw
In reply to Nick Piggin
Quoted 4 lines · Nick Piggin

Hi,

Does this patch solve the X problem? Does anyone see anything wrong with it or know why agp was locking the pages?

We need to do a little bit of auditing here, but I suspect it will turn out all right. I think the reason it locked them in the first place was to avoid AGP pages mapped into process space from being swapped out.

I think that should be taken care of by appropriate vma flags nowadays, but we need to double check. It also might have been a way around dodgy refcounting at one point but I think we got that right nowadays (I remember fixing issues in that area when we removed PageReserved from those pages back then).

Ben.

Quoted 60 lines

--
AGP should not need to lock pages. They are not protecting any race because there is no lock_page calls, only SetPageLocked.

This is causing hangs with d00806b183152af6d24f46f0c33f14162ca1262a.

drivers/char/agp/generic.c +0 -2
index d535c40..3db4f40 100644
@@ -1170,7 +1170,6 @@ void *agp_generic_alloc_page(struct agp_
 	map_page_into_agp(page);
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -1187,7 +1186,6 @@ void agp_generic_destroy_page(void *addr
 	page = virt_to_page(addr);
 	unmap_page_from_agp(page);
 	put_page(page);
-	unlock_page(page);
 	free_page((unsigned long)addr);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/intel-agp.c +0 -2
index a124060..2f319f4 100644
@@ -213,7 +213,6 @@ static void *i8xx_alloc_pages(void)
 	}
 	global_flush_tlb();
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -229,7 +228,6 @@ static void i8xx_destroy_pages(void *add
 	change_page_attr(page, 4, PAGE_KERNEL);
 	global_flush_tlb();
 	put_page(page);
-	unlock_page(page);
 	__free_pages(page, 2);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/sgi-agp.c +0 -1
index cda608c..98cf8ab 100644
@@ -51,7 +51,6 @@ static void *sgi_tioca_alloc_page(struct
 		return NULL;
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
#3 lore raw

[forgot to cc Dave Jones...]

Quoted 10 lines · Benjamin Herrenschmidt
Quoted 4 lines · Nick Piggin

Hi,

Does this patch solve the X problem? Does anyone see anything wrong with it or know why agp was locking the pages?

We need to do a little bit of auditing here, but I suspect it will turn out all right. I think the reason it locked them in the first place was to avoid AGP pages mapped into process space from being swapped out.

Quoted 5 lines

I think that should be taken care of by appropriate vma flags nowadays, but we need to double check. It also might have been a way around dodgy refcounting at one point but I think we got that right nowadays (I remember fixing issues in that area when we removed PageReserved from those pages back then).

Yeah I had a bit of a look around, and it seems OK (but would appreciate an ack from someone who knows the code).

These pages will never get seen by page reclaim, so we're OK there. There is a get_page before the SetPageLocked and a put_page right before the unlock_page, so refcounting should not be broken if it wasn't already: note that the lock_page doesn't pin a reference on a page in general -- we can use it as such for pagecache (although it isn't very clean), because the lock pins the page in pagecache and the pagecache holds a ref.

Anyway, if Dave or David can take a look, that would be appreciated. We'll need this for 2.6.23.

Nick

Quoted 63 lines

Ben.

Quoted 60 lines

--
AGP should not need to lock pages. They are not protecting any race because there is no lock_page calls, only SetPageLocked.

This is causing hangs with d00806b183152af6d24f46f0c33f14162ca1262a.

drivers/char/agp/generic.c +0 -2
index d535c40..3db4f40 100644
@@ -1170,7 +1170,6 @@ void *agp_generic_alloc_page(struct agp_
 	map_page_into_agp(page);
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -1187,7 +1186,6 @@ void agp_generic_destroy_page(void *addr
 	page = virt_to_page(addr);
 	unmap_page_from_agp(page);
 	put_page(page);
-	unlock_page(page);
 	free_page((unsigned long)addr);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/intel-agp.c +0 -2
index a124060..2f319f4 100644
@@ -213,7 +213,6 @@ static void *i8xx_alloc_pages(void)
 	}
 	global_flush_tlb();
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -229,7 +228,6 @@ static void i8xx_destroy_pages(void *add
 	change_page_attr(page, 4, PAGE_KERNEL);
 	global_flush_tlb();
 	put_page(page);
-	unlock_page(page);
 	__free_pages(page, 2);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/sgi-agp.c +0 -1
index cda608c..98cf8ab 100644
@@ -51,7 +51,6 @@ static void *sgi_tioca_alloc_page(struct
 		return NULL;
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
#4 lore raw

[one more try]

Quoted 99 lines · Nick Piggin

[forgot to cc Dave Jones...]

Quoted 10 lines · Benjamin Herrenschmidt
Quoted 4 lines · Nick Piggin

Hi,

Does this patch solve the X problem? Does anyone see anything wrong with it or know why agp was locking the pages?

We need to do a little bit of auditing here, but I suspect it will turn out all right. I think the reason it locked them in the first place was to avoid AGP pages mapped into process space from being swapped out.

Quoted 5 lines

I think that should be taken care of by appropriate vma flags nowadays, but we need to double check. It also might have been a way around dodgy refcounting at one point but I think we got that right nowadays (I remember fixing issues in that area when we removed PageReserved from those pages back then).

Yeah I had a bit of a look around, and it seems OK (but would appreciate an ack from someone who knows the code).

These pages will never get seen by page reclaim, so we're OK there. There is a get_page before the SetPageLocked and a put_page right before the unlock_page, so refcounting should not be broken if it wasn't already: note that the lock_page doesn't pin a reference on a page in general -- we can use it as such for pagecache (although it isn't very clean), because the lock pins the page in pagecache and the pagecache holds a ref.

Anyway, if Dave or David can take a look, that would be appreciated. We'll need this for 2.6.23.

Nick

Quoted 63 lines

Ben.

Quoted 60 lines

--
AGP should not need to lock pages. They are not protecting any race because there is no lock_page calls, only SetPageLocked.

This is causing hangs with d00806b183152af6d24f46f0c33f14162ca1262a.

drivers/char/agp/generic.c +0 -2
index d535c40..3db4f40 100644
@@ -1170,7 +1170,6 @@ void *agp_generic_alloc_page(struct agp_
 	map_page_into_agp(page);
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -1187,7 +1186,6 @@ void agp_generic_destroy_page(void *addr
 	page = virt_to_page(addr);
 	unmap_page_from_agp(page);
 	put_page(page);
-	unlock_page(page);
 	free_page((unsigned long)addr);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/intel-agp.c +0 -2
index a124060..2f319f4 100644
@@ -213,7 +213,6 @@ static void *i8xx_alloc_pages(void)
 	}
 	global_flush_tlb();
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
@@ -229,7 +228,6 @@ static void i8xx_destroy_pages(void *add
 	change_page_attr(page, 4, PAGE_KERNEL);
 	global_flush_tlb();
 	put_page(page);
-	unlock_page(page);
 	__free_pages(page, 2);
 	atomic_dec(&agp_bridge->current_memory_agp);
 }
drivers/char/agp/sgi-agp.c +0 -1
index cda608c..98cf8ab 100644
@@ -51,7 +51,6 @@ static void *sgi_tioca_alloc_page(struct
 		return NULL;
 
 	get_page(page);
-	SetPageLocked(page);
 	atomic_inc(&agp_bridge->current_memory_agp);
 	return page_address(page);
 }
#5 lore raw
In reply to Nick Piggin
Quoted 29 lines · Benjamin Herrenschmidt
Quoted 10 lines · Nick Piggin
Quoted 4 lines

Hi,

Does this patch solve the X problem? Does anyone see anything wrong with it or know why agp was locking the pages?

We need to do a little bit of auditing here, but I suspect it will turn out all right. I think the reason it locked them in the first place was to avoid AGP pages mapped into process space from being swapped out.

Quoted 5 lines

I think that should be taken care of by appropriate vma flags nowadays, but we need to double check. It also might have been a way around dodgy refcounting at one point but I think we got that right nowadays (I remember fixing issues in that area when we removed PageReserved from those pages back then).

Yeah I had a bit of a look around, and it seems OK (but would appreciate an ack from someone who knows the code).

These pages will never get seen by page reclaim, so we're OK there. There is a get_page before the SetPageLocked and a put_page right before the unlock_page, so refcounting should not be broken if it wasn't already: note that the lock_page doesn't pin a reference on a page in general -- we can use it as such for pagecache (although it isn't very clean), because the lock pins the page in pagecache and the pagecache holds a ref.

Quoted 2 lines

Anyway, if Dave or David can take a look, that would be appreciated. We'll need this for 2.6.23.

I talked with Ben earlier and I can't see anything inherently wrong with removing the lock_page, I assume it was put there to stop things getting swapped but if the get/put does that then I'd be happy to remove it.

I'm just a bit confused how this didn't get picked up in -mm at all.

Dave.

#6 lore raw
Quoted 12 lines · Dave Airlie
Quoted 12 lines

Yeah I had a bit of a look around, and it seems OK (but would appreciate an ack from someone who knows the code).

These pages will never get seen by page reclaim, so we're OK there. There is a get_page before the SetPageLocked and a put_page right before the unlock_page, so refcounting should not be broken if it wasn't already: note that the lock_page doesn't pin a reference on a page in general -- we can use it as such for pagecache (although it isn't very clean), because the lock pins the page in pagecache and the pagecache holds a ref.

Quoted 7 lines
Quoted 2 lines

Anyway, if Dave or David can take a look, that would be appreciated. We'll need this for 2.6.23.

I talked with Ben earlier and I can't see anything inherently wrong with removing the lock_page, I assume it was put there to stop things getting swapped but if the get/put does that then I'd be happy to remove it.

Well it is prevented from being swapped out because it never gets put on swapout lists, but the get/put certainly doesn't hurt :)

Quoted 1 line

I'm just a bit confused how this didn't get picked up in -mm at all.

Beats me. It was in there for nearly 5 months. Mustn't have been tested or reported.