Page MenuHomeFreeBSD

swap_pager: deoptimize swapoff_object
ClosedPublic

Authored by dougm on Sep 22 2024, 6:03 PM.
Tags
None
Referenced Files
F102163098: D46753.diff
Fri, Nov 8, 9:42 AM
Unknown Object (File)
Tue, Nov 5, 5:48 AM
Unknown Object (File)
Mon, Nov 4, 3:37 PM
Unknown Object (File)
Fri, Nov 1, 2:13 PM
Unknown Object (File)
Tue, Oct 29, 6:30 AM
Unknown Object (File)
Thu, Oct 17, 1:02 PM
Unknown Object (File)
Sep 28 2024, 10:32 AM
Unknown Object (File)
Sep 27 2024, 5:52 AM
Subscribers

Details

Summary

Undo a change in swap_pager_swapoff_object that assumed a swap block would be valid after reacquiring a lock, when that validity cannot be assumed.

Diff Detail

Lint
Lint Skipped
Unit
Tests Skipped

Event Timeline

dougm requested review of this revision.Sep 22 2024, 6:03 PM
dougm created this revision.
In D46753#1065827, @alc wrote:

No, this is not what @markj meant.

Maybe not.

Here's what I see. In stable/14, there's this code (https://cgit.freebsd.org/src/plain/sys/vm/swap_pager.c?h=stable/14):

			vm_object_pip_add(object, 1);
			rahead = SWAP_META_PAGES;
			rv = swap_pager_getpages_locked(object, &m, 1, NULL,
			    &rahead);
			if (rv != VM_PAGER_OK)
				panic("%s: read from swap failed: %d",
				    __func__, rv);
			VM_OBJECT_WLOCK(object);
			vm_object_pip_wakeupn(object, 1);
			vm_page_xunbusy(m);

			/*
			 * The object lock was dropped so we must restart the
			 * scan of this swap block.  Pages paged in during this
			 * iteration will be marked dirty in a future iteration.
			 */
			break;
		}

The two lines @markj has highlighted aren't there. I put them in main. Assuming that they are a problem, and that stable/14 is correct, I seek to restore correctness by removing them. I thought that including them was an optimization. I assume that in stable/14 the next iteration will likely walk the same swapblk and find that page and process it properly, and that this change would have the same effect.

In D46753#1065827, @alc wrote:

No, this is not what @markj meant.

Maybe not.

It is. :)

I meant to point out that the update of &sb->d[i] following reacquisition of the VM object lock seems wrong.

P.S. I am traveling until Oct 3 and will be slower than usual to reply to reviews before then.

Here's what I see. In stable/14, there's this code (https://cgit.freebsd.org/src/plain/sys/vm/swap_pager.c?h=stable/14):

			vm_object_pip_add(object, 1);
			rahead = SWAP_META_PAGES;
			rv = swap_pager_getpages_locked(object, &m, 1, NULL,
			    &rahead);
			if (rv != VM_PAGER_OK)
				panic("%s: read from swap failed: %d",
				    __func__, rv);
			VM_OBJECT_WLOCK(object);
			vm_object_pip_wakeupn(object, 1);
			vm_page_xunbusy(m);

			/*
			 * The object lock was dropped so we must restart the
			 * scan of this swap block.  Pages paged in during this
			 * iteration will be marked dirty in a future iteration.
			 */
			break;
		}

The two lines @markj has highlighted aren't there. I put them in main. Assuming that they are a problem, and that stable/14 is correct, I seek to restore correctness by removing them. I thought that including them was an optimization. I assume that in stable/14 the next iteration will likely walk the same swapblk and find that page and process it properly, and that this change would have the same effect.

That is my understanding as well.

sys/vm/swap_pager.c
1955

I think the assertion can and should remain.

dougm marked an inline comment as done.

Let the assertion remain.

This revision is now accepted and ready to land.Sep 23 2024, 4:54 PM
This revision was automatically updated to reflect the committed changes.