Re: [PATCH v1 4/4] powernv/memtrace: don't abuse memory hot(un)plug infrastructure for memory allocations

On 03.11.20 10:23, Michal Hocko wrote:

On Thu 29-10-20 17:27:18, David Hildenbrand wrote:

quoted

Let's use alloc_contig_pages() for allocating memory and remove the
linear mapping manually via arch_remove_linear_mapping(). Mark all pages
PG_offline, such that they will definitely not get touched - e.g.,
when hibernating. When freeing memory, try to revert what we did.

The original idea was discussed in:
  https://lkml.kernel.org/r/48340e96-7e6b-736f-9e23-d3111b915b6e@redhat.com

This is similar to CONFIG_DEBUG_PAGEALLOC handling on other
architectures, whereby only single pages are unmapped from the linear
mapping. Let's mimic what memory hot(un)plug would do with the linear
mapping.

We now need MEMORY_HOTPLUG and CONTIG_ALLOC as dependencies.

Simple test under QEMU TCG (10GB RAM, single NUMA node):

sh-5.0# mount -t debugfs none /sys/kernel/debug/
sh-5.0# cat /sys/devices/system/memory/block_size_bytes
40000000
sh-5.0# echo 0x40000000 > /sys/kernel/debug/powerpc/memtrace/enable
[   71.052836][  T356] memtrace: Allocated trace memory on node 0 at 0x0000000080000000
sh-5.0# echo 0x80000000 > /sys/kernel/debug/powerpc/memtrace/enable
[   75.424302][  T356] radix-mmu: Mapped 0x0000000080000000-0x00000000c0000000 with 64.0 KiB pages
[   75.430549][  T356] memtrace: Freed trace memory back on node 0
[   75.604520][  T356] memtrace: Allocated trace memory on node 0 at 0x0000000080000000
sh-5.0# echo 0x100000000 > /sys/kernel/debug/powerpc/memtrace/enable
[   80.418835][  T356] radix-mmu: Mapped 0x0000000080000000-0x0000000100000000 with 64.0 KiB pages
[   80.430493][  T356] memtrace: Freed trace memory back on node 0
[   80.433882][  T356] memtrace: Failed to allocate trace memory on node 0
sh-5.0# echo 0x40000000 > /sys/kernel/debug/powerpc/memtrace/enable
[   91.920158][  T356] memtrace: Allocated trace memory on node 0 at 0x0000000080000000

Note 1: We currently won't be allocating from ZONE_MOVABLE - because our
	pages are not movable. However, as we don't run with any memory
	hot(un)plug mechanism around, we could make an exception to
	increase the chance of allocations succeeding.

Note 2: PG_reserved isn't sufficient. E.g., kernel_page_present() used
	along PG_reserved in hibernation code will always return "true"
	on powerpc, resulting in the pages getting touched. It's too
	generic - e.g., indicates boot allocations.

Note 3: For now, we keep using memory_block_size_bytes() as minimum
	granularity. I'm not able to come up with a better guess (most
	probably, doing it on a section basis could be possible).

Suggested-by: Michal Hocko <mhocko@kernel.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <redacted>
Cc: Rashmica Gupta <redacted>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Wei Yang <redacted>
Signed-off-by: David Hildenbrand <redacted>

Thanks! This looks like a move into the right direction. I cannot really
judge implementation details because I am not familiar with the code.
I have only one tiny concern:
[...]

quoted

-/* called with device_hotplug_lock held */
-static bool memtrace_offline_pages(u32 nid, u64 start_pfn, u64 nr_pages)
+static u64 memtrace_alloc_node(u32 nid, u64 size)
  {
-	const unsigned long start = PFN_PHYS(start_pfn);
-	const unsigned long size = PFN_PHYS(nr_pages);
+	const unsigned long nr_pages = PHYS_PFN(size);
+	unsigned long pfn, start_pfn;
+	struct page *page;
  
-	if (walk_memory_blocks(start, size, NULL, check_memblock_online))
-		return false;
-
-	walk_memory_blocks(start, size, (void *)MEM_GOING_OFFLINE,
-			   change_memblock_state);
-
-	if (offline_pages(start_pfn, nr_pages)) {
-		walk_memory_blocks(start, size, (void *)MEM_ONLINE,
-				   change_memblock_state);
-		return false;
-	}
+	/*
+	 * Trace memory needs to be aligned to the size, which is guaranteed
+	 * by alloc_contig_pages().
+	 */
+	page = alloc_contig_pages(nr_pages, __GFP_THISNODE | __GFP_NOWARN,
+				  nid, NULL);

__GFP_THISNODE without other modifiers looks suspicious. I suspect you
want to enfore node locality and exclude movable zones by this. While
this works it is an antipattern. I would rather use GFP_KERNEL |
__GFP_THISNODE | __GFP_NOWARN to be more in line with other gfp usage.

Agreed GFP_KERNEL should be the right thing to do here.

If for no other reasons we want to be able to work inside a normal
compaction context (comparing to effectively GFP_NOIO which the above
implies). Also this looks like a sleepable context.

Yes it is. Thanks!

-- 
Thanks,

David / dhildenb