Re: [PATCH] oom killer (Core)

[Andrea Arcangeli]

On Fri, Dec 10, 2004 at 05:52:33PM +0100, Thomas Gleixner wrote:
> +			if ((p->flags & PF_MEMDIE) ||
> +			    ((p->flags & PF_EXITING) && !(p->flags & PF_DEAD)))

this had to be:

			if (((p->flags & PF_MEMDIE) || (p->flags & PF_EXITING)) &&
			    !(p->flags & PF_DEAD))

I didn't take zombies into account. A task may be in memdie state and zombie at
the same time. We must not wait a PF_MEMDIE to go away completely, we must wait
only until PF_DEAD is not set. After PF_DEAD is set we know all ram we
were waiting for is already in the free list.

I noticed some deadlocks during an oom-torture-test before applying the stuff
into the suse kernel. The above change fixed all my deadlocks. Everything else
was working fine already.

Actually before finding the above bug I fixed PF_MEMDIE too and I converted it
to p->memdie, an unsigned char. All archs should support 1 byte granularity
for per-process atomic instructions, it's near used_math that I also converted
to a char to signal it cannot be a bitflag sharing the same char with globals.

Struct layout looks like this.

	/*
	 * All archs should support atomic ops with
	 * 1 byte granularity.
	 */
	unsigned char memdie;
	/*
	 * Must be changed atomically so it shouldn't be
	 * be a shareable bitflag.
	 */
	unsigned char used_math;
	/*
	 * OOM kill score adjustment (bit shift).
	 * Cannot live together with used_math since
	 * used_math and oomkilladj can be changed at the
	 * same time, so they would race if they're in the
	 * same atomic block.
	 */
	short oomkilladj;

As for the |= PF_MEMALLOC in oom_kill_task that was a gratuitous smp breakage,
I didn't need to do anything in PF_MEMALLOC since alloc_pages checks _both_
PF_MEMALLOC and p->memdie already.

I also added a !chosen in the below code to make that logic more self
contained and less dependent on the badness implementation (should never
be necessary though):

			if (points > maxpoints || !chosen) {
				chosen = p;
				maxpoints = points;
			}

I can port the final patch (including fixage of PF_MEMDIE races) to 2.6.10-rc
after I finished.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/