On Mon, 2025-03-31 at 17:50 +0300, Mike Rapoport wrote:

On Mon, Mar 31, 2025 at 01:50:33PM +0100, David Woodhouse wrote:

quoted

On Tue, 2021-05-11 at 13:05 +0300, Mike Rapoport wrote:

quoted

 
+static void __init memmap_init_reserved_pages(void)
+{
+	struct memblock_region *region;
+	phys_addr_t start, end;
+	u64 i;
+
+	/* initialize struct pages for the reserved regions */
+	for_each_reserved_mem_range(i, &start, &end)
+		reserve_bootmem_region(start, end);
+
+	/* and also treat struct pages for the NOMAP regions as PageReserved */
+	for_each_mem_region(region) {
+		if (memblock_is_nomap(region)) {
+			start = region->base;
+			end = start + region->size;
+			reserve_bootmem_region(start, end);
+		}
+	}
+}
+

In some cases, that whole call to reserve_bootmem_region() may be a no-
op because pfn_valid() is not true for *any* address in that range.

But reserve_bootmem_region() spends a long time iterating of them all,
and eventually doing nothing:

void __meminit reserve_bootmem_region(phys_addr_t start,
                                      phys_addr_t end, int nid)
{
        unsigned long start_pfn = PFN_DOWN(start);
        unsigned long end_pfn = PFN_UP(end);

        for (; start_pfn < end_pfn; start_pfn++) {
                if (pfn_valid(start_pfn)) {
                        struct page *page = pfn_to_page(start_pfn);

                        init_reserved_page(start_pfn, nid);

                        /*
                         * no need for atomic set_bit because the struct
                         * page is not visible yet so nobody should
                         * access it yet.
                         */
                        __SetPageReserved(page);
                }
        }
}

On platforms with large NOMAP regions (e.g. which are actually reserved
for guest memory to keep it out of the Linux address map and allow for
kexec-based live update of the hypervisor), this pointless loop ends up
taking a significant amount of time which is visible as guest steal
time during the live update.

Can reserve_bootmem_region() skip the loop *completely* if no PFN in
the range from start to end is valid? Or tweak the loop itself to have
an 'else' case which skips to the next valid PFN? Something like

 for(...) {
    if (pfn_valid(start_pfn)) {
       ...
    } else {
       start_pfn = next_valid_pfn(start_pfn);
    }
 }

My understanding is that you have large reserved NOMAP ranges that don't
appear as memory at all, so no memory map for them is created and so
pfn_valid() is false for pfns in those ranges.

If this is the case one way indeed would be to make
reserve_bootmem_region() skip ranges with no valid pfns.

Another way could be to memblock_reserved_mark_noinit() such ranges and
then reserve_bootmem_region() won't even get called, but that would require
firmware to pass that information somehow.

I was thinking along these lines (not even build tested)...

I don't much like the (unsigned long)-1 part. I might make the helper
'static inline bool first_valid_pfn (unsigned long *pfn)' and return
success or failure. But that's an implementation detail.

index 6d1fb6162ac1..edd27ba3e908 100644

--- a/include/asm-generic/memory_model.h
+++ b/include/asm-generic/memory_model.h

@@ -29,8 +29,43 @@ static inline int pfn_valid(unsigned long pfn)
        return pfn >= pfn_offset && (pfn - pfn_offset) < max_mapnr;
 }
 #define pfn_valid pfn_valid
+
+static inline unsigned long first_valid_pfn(unsigned long pfn)
+{
+       /* avoid <linux/mm.h> include hell */
+       extern unsigned long max_mapnr;
+       unsigned long pfn_offset = ARCH_PFN_OFFSET;
+
+       if (pfn < pfn_offset)
+               return pfn_offset;
+
+       if ((pfn - pfn_offset) < max_mapnr)
+               return pfn;
+
+       return (unsigned long)(-1);
+}
+
+#ifndef for_each_valid_pfn
+#define for_each_valid_pfn(pfn, start_pfn, end_pfn)            \
+       /* Sanity check on the end condition */                 \
+       BUG_ON(end_pfn == (unsigned long)-1);                   \
+       for (pfn = first_valid_pfn(pfn); pfn < end_pfn;         \
+            pfn = first_valid_pfn(pfn + 1))
+#endif
+#endif
+
+/*
+ * If the architecture provides its own pfn_valid(), it can either
+ * provide a matching for_each_valid_pfn() or use the fallback which
+ * just iterates over them *all*, calling pfn_valid() for each.
+ */
+#ifndef for_each_valid_pfn
+#define for_each_valid_pfn(pfn, start_pfn, end_pfn)            \
+       for (pfn = start_pfn; pfn < end_pfn, pfn++) {           \
+               if (pfn_valid(pfn))
 #endif
 
+
 #elif defined(CONFIG_SPARSEMEM_VMEMMAP)
 
 /* memmap is virtually contiguous.  */