Skip to contentSkip to conversation

NOPAGE_RETRY and 2.6.19

DISCUSSION Benjamin Herrenschmidt 2 participants 5 messages Oct 5, 2006 – Oct 6, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi Andrew !

Any chance that can be merged in 2.6.19 ?

The problem I have with signal and spufs is an actual bug affecting apps and I don't see other ways of fixing it.

In addition, we are having issues with infiniband and 64k pages (related to the way the hypervisor deals with some HV cards) that will require us to muck around with the MMU from within the IB driver's no_page() (it's a pSeries specific driver) and return to the caller the same way using NOPAGE_RETRY.

And to add to this, the graphics folks have been following a new approach of memory management that involves transparently swapping objects between video ram and main meory. To do that, they need installing PTEs from a no_page() handler as well and that also requires returning with NOPAGE_RETRY.

(For the later, they are currently using io_remap_pfn_range to install one PTE from no_page() which is a bit racy, we need to add a check for the PTE having already been installed afer taking the lock, but that's ok, they are only at the proof-of-concept stage. I'll send a patch adding a "clean" function to do that, we can use that from spufs too and get rid of the sparsemem hacks we do to create struct page for SPEs. Basically, that provides a generic solution for being able to have no_page() map hardware devices, which is something that I think sound driver folks have been asking for some time too).

All of these things depend on having the NOPAGE_RETRY exit path from no_page() handlers. Thus I was wondering if it was possible to have it in 2.6.19 since just adding the interface shouldn't affect anything badly.

Cheers,
Ben.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#1 lore raw
Quoted 1 line · On Fri, 06 Oct 2006 08:40:50 +1000 Benjamin Herrenschmidt…

Any chance that can be merged in 2.6.19 ?

Not if you don't show it to anyone ;)

I renamed your NOPAGE_RETRY to NOPAGE_REFAULT. Because I think that better communicates the fact that it returns all the way to userspace to redo the fault. Later, a NOPAGE_RETRY would be a lower-level thing which doesn't redo the fault.

From: Benjamin Herrenschmidt <[email protected]>

Add a way for a no_page() handler to request a retry of the faulting instruction. It goes back to userland on page faults and just tries again in get_user_pages(). I added a cond_resched() in the loop in that later case.

The problem I have with signal and spufs is an actual bug affecting apps and I don't see other ways of fixing it.

In addition, we are having issues with infiniband and 64k pages (related to the way the hypervisor deals with some HV cards) that will require us to muck around with the MMU from within the IB driver's no_page() (it's a pSeries specific driver) and return to the caller the same way using NOPAGE_REFAULT.

And to add to this, the graphics folks have been following a new approach of memory management that involves transparently swapping objects between video ram and main meory. To do that, they need installing PTEs from a no_page() handler as well and that also requires returning with NOPAGE_REFAULT.

(For the later, they are currently using io_remap_pfn_range to install one PTE from no_page() which is a bit racy, we need to add a check for the PTE having already been installed afer taking the lock, but that's ok, they are only at the proof-of-concept stage. I'll send a patch adding a "clean" function to do that, we can use that from spufs too and get rid of the sparsemem hacks we do to create struct page for SPEs. Basically, that provides a generic solution for being able to have no_page() map hardware devices, which is something that I think sound driver folks have been asking for some time too).

All of these things depend on having the NOPAGE_REFAULT exit path from no_page() handlers.

---

 include/linux/mm.h |    1 +
 mm/memory.c        |    9 ++++++---
 2 files changed, 7 insertions(+), 3 deletions(-)
include/linux/mm.h renamed from include/linux/mm.h~page-fault-retry-with-nopage_retry +1 -0
diff -puN include/linux/mm.h~page-fault-retry-with-nopage_retry include/linux/mm.h
@@ -593,6 +593,7 @@ static inline int page_mapped(struct pag
  */
 #define NOPAGE_SIGBUS	(NULL)
 #define NOPAGE_OOM	((struct page *) (-1))
+#define NOPAGE_REFAULT	((struct page *) (-2))	/* Return to userspace, rerun */
 
 /*
  * Error return values for the *_nopfn functions
mm/memory.c renamed from mm/memory.c~page-fault-retry-with-nopage_retry +6 -3
diff -puN mm/memory.c~page-fault-retry-with-nopage_retry mm/memory.c
@@ -1086,6 +1086,7 @@ int get_user_pages(struct task_struct *t
 				default:
 					BUG();
 				}
+				cond_resched();
 			}
 			if (pages) {
 				pages[i] = page;
@@ -2169,11 +2170,13 @@ retry:
 	 * after the next truncate_count read.
 	 */
 
-	/* no page was available -- either SIGBUS or OOM */
-	if (new_page == NOPAGE_SIGBUS)
+	/* no page was available -- either SIGBUS, OOM or REFAULT */
+	if (unlikely(new_page == NOPAGE_SIGBUS))
 		return VM_FAULT_SIGBUS;
-	if (new_page == NOPAGE_OOM)
+	else if (unlikely(new_page == NOPAGE_OOM))
 		return VM_FAULT_OOM;
+	else if (unlikely(new_page == NOPAGE_REFAULT))
+		return VM_FAULT_MINOR;
 
 	/*
 	 * Should we do an early C-O-W break?

_

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#2 lore raw
Quoted 6 lines · Andrew Morton
Quoted 1 line · On Fri, 06 Oct 2006 08:40:50 +1000 Benjamin Herrenschmidt…

Any chance that can be merged in 2.6.19 ?

Not if you don't show it to anyone ;)

Didn' I send it to you last week ?

Quoted 4 lines

I renamed your NOPAGE_RETRY to NOPAGE_REFAULT. Because I think that better communicates the fact that it returns all the way to userspace to redo the fault. Later, a NOPAGE_RETRY would be a lower-level thing which doesn't redo the fault.

Ok.

Thanks,
Ben.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#3 lore raw
Quoted 9 lines · On Fri, 06 Oct 2006 09:38:19 +1000 Benjamin Herrenschmidt…
Quoted 6 lines · Andrew Morton
Quoted 1 line · On Fri, 06 Oct 2006 08:40:50 +1000 Benjamin Herrenschmidt…

Any chance that can be merged in 2.6.19 ?

Not if you don't show it to anyone ;)

Didn' I send it to you last week ?

I saw a little dribble of diff at the end of some email thread, but unless it fixes some bug which I want to fix I'll generally ignore dribbly diffs and wait for the real patch.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#4 lore raw
Quoted 9 lines · Andrew Morton
Quoted 6 lines · On Fri, 06 Oct 2006 09:38:19 +1000 Benjamin Herrenschmidt…
Quoted 5 lines · Andrew Morton
Quoted 1 line · On Fri, 06 Oct 2006 08:40:50 +1000 Benjamin Herrenschmidt…

Any chance that can be merged in 2.6.19 ?

Quoted 7 lines
Quoted 3 lines
Quoted 1 line

Not if you don't show it to anyone ;)

Didn' I send it to you last week ?

I saw a little dribble of diff at the end of some email thread, but unless it fixes some bug which I want to fix I'll generally ignore dribbly diffs and wait for the real patch.

Well, it had a comment and a signed-off-by and is identical to the patch I just sent you :) Anyway, that's ok. I'll look into doing the matching fix to spufs so we can at least fix the signal bug and will send that to you today.

Ben.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>