page_referenced() and VM_LOCKED

Ethan Solomita

author

2007-11-16 04:25 UTC #0 lore raw

page_referenced_file() checks for the vma to be VM_LOCKED|VM_MAYSHARE and adds returns 1. We don't do the same in page_referenced_anon(). I would've thought the point was to treat locked pages as active, never pushing them into the inactive list, but since that's not quite what's happening I was hoping someone could give me a clue.

Thanks,
-- Ethan

--
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>

KAMEZAWA Hiroyuki

2007-11-16 05:46 UTC #1 lore raw

Quoted 8 lines · On Thu, 15 Nov 2007 20:25:45 -0800 Ethan Solomita <[email protected]> wrote:

page_referenced_file() checks for the vma to be VM_LOCKED|VM_MAYSHARE and adds returns 1. We don't do the same in page_referenced_anon(). I would've thought the point was to treat locked pages as active, never pushing them into the inactive list, but since that's not quite what's happening I was hoping someone could give me a clue.
Thanks,
-- Ethan

Hmm,

== vmscan.c::shrink_page_list()

page_referenced() if returns 1 -> link to active list

add to swap # only works if anon

try_to_unmap() if VM_LOCKED -> SWAP_FAIL -> link to active list

==

Then, "VM_LOCKED & not referenced" anon page is added to swap cache (before pushed back to active list)

Seems intended ?

Thanks,
- Kame

--
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>

Hugh Dickins

2007-11-16 18:00 UTC #2 lore raw

Quoted 5 lines · KAMEZAWA Hiroyuki

Quoted 2 lines · On Thu, 15 Nov 2007 20:25:45 -0800 Ethan Solomita <[email protected]> wrote:
page_referenced_file() checks for the vma to be VM_LOCKED|VM_MAYSHARE and adds returns 1.

That's a case where it can deduce that the page is present and should be treated as referenced, without even examining the page tables.

Quoted 1 line

Quoted 1 line
We don't do the same in page_referenced_anon().

It cannot make that same deduction in the page_referenced_anon() case (different vmas may well contain different COWs of some original page).

Perhaps you're suggesting that page_referenced_one() ought to cover this case. Yes. I think we were coming at it from the 2.6.4 rmap.c in which VM_LOCKED was checked only in the try_to_unmap() case; but I was worried about the length of the vma lists and tried to short- circuit the full search in the one case I could; without thinking about doing the equivalent elsewhere for the other cases.

Quoted 3 lines

Quoted 3 lines
I would've thought the point was to treat locked pages as active, never pushing them into the inactive list, but since that's not quite what's happening I was hoping someone could give me a clue.

Rik and Lee and others have proposed that we keep VM_LOCKED pages off both active and inactive lists: that seems a better way forward.

Quoted 11 lines

Quoted 3 lines
Thanks,
-- Ethan
Hmm,
== vmscan.c::shrink_page_list()
page_referenced() if returns 1 -> link to active list
add to swap # only works if anon

Quoted 8 lines

   try_to_unmap()    if VM_LOCKED -> SWAP_FAIL -> link to active list
==
Then, "VM_LOCKED & not referenced" anon page is added to swap cache (before pushed back to active list)
Seems intended ?

Not intended, no. Rather a waste of swap. How about this patch?

mm/rmap.c +4 -1

@@ -283,7 +283,10 @@ static int page_referenced_one(struct pa
 	if (!pte)
 		goto out;
 
-	if (ptep_clear_flush_young(vma, address, pte))
+	if (vma->vm_flags & VM_LOCKED) {
+		referenced++;
+		*mapcount = 1;
+	} else if (ptep_clear_flush_young(vma, address, pte))
 		referenced++;
 
 	/* Pretend the page is referenced if the task has the

--
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>

[email protected]

2007-11-17 00:32 UTC #3 lore raw

Quoted 7 lines

Quoted 3 lines
Quoted 3 lines
I would've thought the point was to treat locked pages as active, never pushing them into the inactive list, but since that's not quite what's happening I was hoping someone could give me a clue.
Rik and Lee and others have proposed that we keep VM_LOCKED pages off both active and inactive lists: that seems a better way forward.

agreed.

Quoted 7 lines

Quoted 4 lines
Then, "VM_LOCKED & not referenced" anon page is added to swap cache (before pushed back to active list)
Seems intended ?
Not intended, no. Rather a waste of swap. How about this patch?

seems nice. I'd like to do some test in the next week,

Thanks,
-Kame

--
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>

KAMEZAWA Hiroyuki

Tested-by

2007-11-19 09:39 UTC #4 lore raw

Quoted 9 lines · On Sat, 17 Nov 2007 09:32:36 +0900 (JST) [email protected] wrote:

Quoted 7 lines
Quoted 3 lines
Quoted 3 lines
I would've thought the point was to treat locked pages as active, never pushing them into the inactive list, but since that's not quite what's happening I was hoping someone could give me a clue.
Rik and Lee and others have proposed that we keep VM_LOCKED pages off both active and inactive lists: that seems a better way forward.
agreed.

Quoted 9 lines

Quoted 7 lines
Quoted 4 lines
Then, "VM_LOCKED & not referenced" anon page is added to swap cache (before pushed back to active list)
Seems intended ?
Not intended, no. Rather a waste of swap. How about this patch?
seems nice. I'd like to do some test in the next week,

your patch helps the kernel to avoid a waste of Swap.

Tested-by KAMEZAWA Hiroyuki <[email protected]>

==
I tested your patch on x86_64/6GiB memory, + 2.6.24-rc3. mlock 5GiB and create 4GiB file by"dd".

[before patch]

MemTotal:      6072620 kB
MemFree:         50540 kB
Buffers:          4508 kB
Cached:         724828 kB
SwapCached:    5146960 kB
Active:        2683964 kB
Inactive:      3198752 kB

[after patch]

MemTotal:      6072620 kB
MemFree:         17112 kB
Buffers:          6816 kB
Cached:         744880 kB
SwapCached:      21724 kB
Active:        5175828 kB
Inactive:       744956 kB

Thanks,
-Kame

--
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>

Hugh Dickins

2007-11-19 19:09 UTC #5 lore raw

Quoted 4 lines · KAMEZAWA Hiroyuki

Quoted 1 line
your patch helps the kernel to avoid a waste of Swap.
Tested-by KAMEZAWA Hiroyuki <[email protected]>

Many thanks, Kame. I'll send it in with my next little lot. Hugh

Quoted 13 lines

==
I tested your patch on x86_64/6GiB memory, + 2.6.24-rc3. mlock 5GiB and create 4GiB file by"dd".
[before patch]
MemTotal:      6072620 kB
MemFree:         50540 kB
Buffers:          4508 kB
Cached:         724828 kB
SwapCached:    5146960 kB
Active:        2683964 kB
Inactive:      3198752 kB

Quoted 8 lines

[after patch]

MemTotal:      6072620 kB
MemFree:         17112 kB
Buffers:          6816 kB
Cached:         744880 kB
SwapCached:      21724 kB
Active:        5175828 kB
Inactive:       744956 kB

--
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>

Ethan Solomita

author

2007-12-05 00:26 UTC #6 lore raw

In reply to Hugh Dickins

Quoted 7 lines · Hugh Dickins

Quoted 5 lines · KAMEZAWA Hiroyuki
Quoted 2 lines · On Thu, 15 Nov 2007 20:25:45 -0800 Ethan Solomita <[email protected]> wrote:
page_referenced_file() checks for the vma to be VM_LOCKED|VM_MAYSHARE and adds returns 1.

Quoted 7 lines

That's a case where it can deduce that the page is present and should be treated as referenced, without even examining the page tables.
Quoted 1 line
Quoted 1 line
We don't do the same in page_referenced_anon().
It cannot make that same deduction in the page_referenced_anon() case (different vmas may well contain different COWs of some original page).

Sorry to come back in with this so late -- if the vma is VM_MAYSHARE,
would there be COWs of the original page?
-- Ethan

--
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>

Hugh Dickins

2007-12-05 15:55 UTC #7 lore raw

Quoted 11 lines · Ethan Solomita

Quoted 10 lines · Hugh Dickins
Quoted 5 lines · KAMEZAWA Hiroyuki
Quoted 2 lines · On Thu, 15 Nov 2007 20:25:45 -0800 Ethan Solomita <[email protected]> wrote:
page_referenced_file() checks for the vma to be VM_LOCKED|VM_MAYSHARE and adds returns 1.
That's a case where it can deduce that the page is present and should be treated as referenced, without even examining the page tables.

Quoted 7 lines

Quoted 4 lines
Quoted 1 line
Quoted 1 line
We don't do the same in page_referenced_anon().
It cannot make that same deduction in the page_referenced_anon() case (different vmas may well contain different COWs of some original page).
Sorry to come back in with this so late --
if the vma is VM_MAYSHARE, would there be COWs of the original page?

99.999% correct answer: No, that's why we're testing VM_MAYSHARE, because we know it has to be that original page which is present and locked there (it might be readwrite, or it might be readonly, but it won't be a COWed copy in a VM_MAYSHARE vma; the same would be true if we tested VM_SHARED, but that would miss readonly cases).

00.001% adjustment: Actually, there's an aberrant case in which do_wp_page() can put a COW into a VM_MAYSHARE area, supposedly to suit ptrace(). I overlooked that case when I put the VM_MAYSHARE test into page_referenced_file(); later when I learnt about it (and found Linus unwilling to change it), I did an audit of such oversights, but concluded this test wasn't worth changing.

Hugh

--
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>