On Wed 17-07-24 14:55:08, Kirill A. Shutemov wrote:
On Wed, Jul 17, 2024 at 09:19:12AM +0200, Michal Hocko wrote:
On Tue 16-07-24 16:00:13, Kirill A. Shutemov wrote:
Unaccepted memory is considered unusable free memory, which is not counted as free on the zone watermark check. This causes get_page_from_freelist() to accept more memory to hit the high watermark, but it creates problems in the reclaim path.
The reclaim path encounters a failed zone watermark check and attempts to reclaim memory. This is usually successful, but if there is little or no reclaimable memory, it can result in endless reclaim with little to no progress. This can occur early in the boot process, just after start of the init process when the only reclaimable memory is the page cache of the init executable and its libraries.
How does this happen when try_to_accept_memory is the first thing to do when wmark check fails in the allocation path?
Good question.
I've lost access to the test setup and cannot check it directly right now.
Reading the code Looks like __alloc_pages_bulk() bypasses get_page_from_freelist() where we usually accept more pages and goes directly to __rmqueue_pcplist() -> rmqueue_bulk() -> __rmqueue().
Will look more into it when I have access to the test setup.
Could you describe what was the initial configuration of the system? How much of the unaccepted memory was there to trigger this?
This is large TDX guest VM: 176 vCPUs and ~800GiB of memory.
One thing that I noticed that the problem is only triggered when LRU_GEN enabled. But I failed to identify why.
The system hang (or have very little progress) shortly after systemd starts.
Please try to investigate this further. The patch as is looks rather questionable to me TBH. Spilling unaccepted memory into the reclaim seems like something we should avoid if possible as this is something page allocator should care about IMHO.
To address this issue, teach shrink_node() and shrink_zones() to accept memory before attempting to reclaim.
Signed-off-by: Kirill A. Shutemov kirill.shutemov@linux.intel.com Reported-by: Jianxiong Gao jxgao@google.com Fixes: dcdfdd40fa82 ("mm: Add support for unaccepted memory") Cc: stable@vger.kernel.org # v6.5+
[...]
static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) { unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed; struct lruvec *target_lruvec; bool reclaimable = false;
- /* Try to accept memory before going for reclaim */
- if (node_try_to_accept_memory(pgdat, sc)) {
if (!should_continue_reclaim(pgdat, 0, sc))return;- }
This would need an exemption from the memcg reclaim.
Hm. Could you elaborate why?
Because memcg reclaim doesn't look for memory but rather frees charges to reclaim for the new use so unaccepted memory is not really relevant as it couldn't have been charged.