On 2016/3/30 0:28, Julien Grall wrote:

Hi Shannon,

On 24/03/16 14:44, Shannon Zhao wrote:

quoted

Make xen_xlate_map_ballooned_pages work with 64K pages. In that case
Kernel pages are 64K in size but Xen pages remain 4K in size. Xen pfns
refer to 4K pages.

Signed-off-by: Shannon Zhao <redacted>
Reviewed-by: Stefano Stabellini <redacted>
---
  drivers/xen/xlate_mmu.c | 26 ++++++++++++++++----------
  1 file changed, 16 insertions(+), 10 deletions(-)

diff --git a/drivers/xen/xlate_mmu.c b/drivers/xen/xlate_mmu.c
index 9692656..28f728b 100644
--- a/drivers/xen/xlate_mmu.c
+++ b/drivers/xen/xlate_mmu.c

@@ -207,9 +207,12 @@ int __init

xen_xlate_map_ballooned_pages(xen_pfn_t **gfns, void **virt,
      void *vaddr;
      int rc;
      unsigned int i;
+    unsigned long nr_pages;
+    xen_pfn_t xen_pfn = 0;

      BUG_ON(nr_grant_frames == 0);
-    pages = kcalloc(nr_grant_frames, sizeof(pages[0]), GFP_KERNEL);
+    nr_pages = DIV_ROUND_UP(nr_grant_frames, XEN_PFN_PER_PAGE);
+    pages = kcalloc(nr_pages, sizeof(pages[0]), GFP_KERNEL);
      if (!pages)
          return -ENOMEM;

@@ -218,22 +221,25 @@ int __init

xen_xlate_map_ballooned_pages(xen_pfn_t **gfns, void **virt,
          kfree(pages);
          return -ENOMEM;
      }
-    rc = alloc_xenballooned_pages(nr_grant_frames, pages);
+    rc = alloc_xenballooned_pages(nr_pages, pages);
      if (rc) {
-        pr_warn("%s Couldn't balloon alloc %ld pfns rc:%d\n", __func__,
-            nr_grant_frames, rc);
+        pr_warn("%s Couldn't balloon alloc %ld pages rc:%d\n", __func__,
+            nr_pages, rc);
          kfree(pages);
          kfree(pfns);
          return rc;
      }
-    for (i = 0; i < nr_grant_frames; i++)
-        pfns[i] = page_to_pfn(pages[i]);
+    for (i = 0; i < nr_grant_frames; i++) {
+        if ((i % XEN_PFN_PER_PAGE) == 0)
+            xen_pfn = page_to_xen_pfn(pages[i / XEN_PFN_PER_PAGE]);
+        pfns[i] = pfn_to_gfn(xen_pfn++);
+    }

Would it be possible to re-use xen_for_each_gfn? This will avoid
open-coding the loop to break down the Linux page.

I don't think so. Using xen_acpi_guest_init will require factoring
"pfns[i] = pfn_to_gfn(xen_pfn++)" to a function with parameter pfns[i].
How can we pass pfns[i]?

Thanks,
-- 
Shannon