On Thu, Aug 29, 2019 at 12:39:11PM -0700, Sukadev Bhattiprolu wrote:

Bharata B Rao [bharata@linux.ibm.com] wrote:

quoted

On Wed, Aug 28, 2019 at 08:02:19PM -0700, Sukadev Bhattiprolu wrote:

Where do we serialize two threads attempting to H_SVM_PAGE_IN the same gfn
at the same time? Or one thread issuing a H_SVM_PAGE_IN and another a
H_SVM_PAGE_OUT for the same page?

I am not not serializing page-in/out calls on same gfn, I thought you take
care of that in UV, guess UV doesn't yet.

I can probably use rmap_lock() and serialize such calls in HV if UV can't
prevent such calls easily.

quoted

quoted

quoted

+
+	if (!trylock_page(dpage))
+		goto out_clear;
+
+	*rmap = devm_pfn | KVMPPC_RMAP_DEVM_PFN;
+	pvt = kzalloc(sizeof(*pvt), GFP_ATOMIC);
+	if (!pvt)
+		goto out_unlock;

If we fail to alloc, we don't clear the KVMPPC_RMAP_DEVM_PFN?

Right, I will move the assignment to *rmap to after kzalloc.

Also, when/where do we clear this flag on an uv-page-out?
kvmppc_devm_drop_pages() drops the flag on a local variable but not
in the rmap? If we don't clear the flag on page-out, would the
subsequent H_SVM_PAGE_IN of this page fail?

It gets cleared in kvmppc_devm_page_free().

Ok. Nit. thought we can drop the "_fault" in the function name but would
collide the other "alloc_and_copy" function used during H_SVM_PAGE_IN.
If the two alloc_and_copy functions are symmetric, maybe they could
have "page_in" and "page_out" in the (already long) names.

Christoph also suggested to reorganize these two calls. Will take care.

Regards,
Bharata.