Use unlocked page lookup for inmem() to avoid object lock contention

[mlaier]

Summary

Sponsored by: Dell EMC Isilon

Test Plan

Significant reduction in vm_object lock contention shown.

Seems to fit with: https://reviews.freebsd.org/D23847

[mlaier]

Sorry - bad WS

[markj]

Could you please upload with extra context? Add -U99999 if you are uploading a raw diff, per https://wiki.freebsd.org/Phabricator .

[mlaier]

More context - sorry, rusty fingers :)

[markj]

sys/kern/vfs_bio.c

3613

There is a vm_page_grab_pages_unlocked() that handles runs of multiple pages, did you consider using it?

3617

I think the gotos should be replaced with a plain return statement.

3621

Note that the shared busy state on its own does not block a concurrent vm_page_invalid() call on the page. I'm not sure if this matters much since in the previous implementation results were stale as soon as the object lock is dropped.

[mlaier]

sys/kern/vfs_bio.c

3613

I have decided against that, because it will block other users for longer - iff somebody wants an exclusive busy on the first page, while we are looking through the results. This way also preserved more of the original code flow.

3617

I was trying to preserve as much of the original code as possible, but I can rework it.

3621

Yeah - this interface is completely advisory. My goal was to make it as unimpactful as possible - hence the single page busy above as well...

There is an argument for changing this to to a VM_ALLOC_WIRED above and return the pages back to PQ_ACTIVE (because why would you check inmem if you're not planning to use it) ... but I didn't want to add side-effects.

[kib]

This is fine.

But since completely advisory (it is used to avoid unneeded read-ahead), you can avoid sbusy as well, I believe. False-negatives should be safe.
You would remove RLOCK as in patch, and use vm_radix_lookup_unlocked. Then after the check for validity you can re-check with vm_radix_lookup_unlocked() that the page is still within object.

In fact I think that it would not give false negatives because the vnode lock is owned.

[markj]

sys/kern/vfs_bio.c

3617

The diff is already quite small, I don't think it hurts anything to simplify the function a bit.

3621

You don't need to wire the page if it is busied. It would be reasonable to call vm_page_reference() before unbusying the page. If the page is already in PQ_ACTIVE, then the extra reference will be counted next time the page is visited. If the reference bit is observed during a PQ_INACTIVE scan, the page will either be requeued or moved to PQ_ACTIVE depending on whether it might be mapped.

[mlaier]

What @kib suggested(I think) - also trying to do the cleanup @markj asked for.

[markj]

sys/kern/vfs_bio.c

3620

I think this can move to the end of the loop body.

[kib]

sys/kern/vfs_bio.c

3589

Change return type to bool, since you are fixing style ?

3630

Does it make sense to check for m != n ? You can return true to stop read-ahead as well. We cannot handle ABA there anyway.

[mlaier]

sys/kern/vfs_bio.c

3589

will do

3620

will do

3630

This also catches the n == NULL case (e.g. we just lost the page) and we certainly want to do read-ahead then.

This is probably overly paranoid, and I could be convinced to just drop the second lookup altogether.

[bdrewery]

I had committed this prematurely and now reverted.

Building /usr/obj/usr/src/i386.i386/sys/SARAH/vfs_bio.o
--- vfs_bio.o ---
/usr/src/sys/kern/vfs_bio.c:3614:7: error: implicit declaration of function 'vm_page_lookup_unlocked' is invalid in C99 [-Werror,-Wimplicit-function-declaration]
                m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
                    ^
/usr/src/sys/kern/vfs_bio.c:3614:7: note: did you mean 'vm_map_lookup_locked'?
/usr/src/sys/vm/vm_map.h:480:5: note: 'vm_map_lookup_locked' declared here
int vm_map_lookup_locked(vm_map_t *, vm_offset_t, vm_prot_t, vm_map_entry_t *, vm_object_t *,
    ^
/usr/src/sys/kern/vfs_bio.c:3614:5: error: incompatible integer to pointer conversion assigning to 'vm_page_t' (aka 'struct vm_page *') from 'int' [-Werror,-Wint-conversion]
                m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
                  ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/src/sys/kern/vfs_bio.c:3624:5: error: incompatible integer to pointer conversion assigning to 'vm_page_t' (aka 'struct vm_page *') from 'int' [-Werror,-Wint-conversion]
                n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
                  ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
3 errors generated.
*** [vfs_bio.o] Error code 1

sys/kern/vfs_bio.c

3601

Ravi Pokala pointed out that this changed to bool but the return values did change to true/false.

[bdrewery]

sys/kern/vfs_bio.c

3601

did NOT

[markj]

sys/kern/vfs_bio.c

3614

(To be clear, the problem is that there is no vm_page_lookup_unlocked() yet. It should be straightforward to add. I would prefer that to using a naked vm_radix_lookup_unlocked() call here.)

[mlaier]

https://reviews.freebsd.org/D26653