[RFC PATCH v4 2/3] mm/filemap: allow filemap faults to retry under the VMA lock

[Hongru Zhang]

From: Hongru Zhang <zhanghongru@xxxxxxxxxx>

filemap_fault() currently returns VM_FAULT_RETRY_MMAP_LOCK on every
retry, which forces faults taken under the per-VMA lock to fall back
to mmap_lock even when the retry could complete under that lock alone.

Add vmf_retry_with_fault_lock() and use it in filemap_fault() so that
faults taken under the per-VMA lock return VM_FAULT_RETRY_VMA_LOCK.
This lets the bounded VMA-lock retry handle the second attempt before
falling back to mmap_lock.

All retry paths in filemap_fault() already use a common return path.
Use vmf_retry_with_fault_lock() there instead of hard-coding
VM_FAULT_RETRY_MMAP_LOCK.

try_handle_fault_under_vma_lock() bounds a VMA-lock retry by setting
FAULT_FLAG_TRIED before retrying under the per-VMA lock, so a retried
VMA-lock filemap fault may reach synchronous I/O while still holding
that lock. In stress testing with the workload [1], there were over
140 million filemap_fault() calls, but the synchronous I/O path under
retried VMA-lock faults was reached only 115 times, less than 0.0001%
of all filemap_fault() calls; all but one completed within 1 ms, and
none exceeded 4 ms. This did not show a problematic latency tail from
the bounded VMA-lock retry.

Based on the stress model from Kunwu Chan and Wang Lian in v2, we
adapted a benchmark [2] to a 20-core Intel i7-12700 desktop by reducing
the thread count and adjusting the memcg limits. The benchmark uses
concurrent page faults under memcg pressure with parallel munmap to
amplify mmap_lock read-write contention.

Throughput (higher is better):
+---------+------------+------------+-------------+
| Threads |  Vanilla   |  Patched   | Improvement |
+---------+------------+------------+-------------+
|   40    | 1069.34 /s | 1404.47 /s |   +31.3%    |
+---------+------------+------------+-------------+
|   60    | 1038.12 /s | 1682.88 /s |   +62.1%    |
+---------+------------+------------+-------------+
|   80    | 1042.62 /s | 1766.72 /s |   +69.5%    |
+---------+------------+------------+-------------+

mmap_lock contention count (lower is better):
+---------+-----------+---------+-----------+
| Threads |  Vanilla  | Patched | Reduction |
+---------+-----------+---------+-----------+
|   40    | 3,187,336 |  68,490 |  -97.9%   |
+---------+-----------+---------+-----------+
|   60    | 4,385,154 |  96,234 |  -97.8%   |
+---------+-----------+---------+-----------+
|   80    | 5,337,890 | 116,184 |  -97.8%   |
+---------+-----------+---------+-----------+

Since filemap_fault() can now return either VM_FAULT_RETRY_* value,
update the generic fault handling and fault-signal/accounting checks to
recognize VM_FAULT_RETRY_MASK as incomplete faults.

Update the filemap_fault() and helper comments for the new fault-lock
retry path.

[1] https://gist.github.com/zhr250/4d53483a6e91aa2f7ded091f6019e491
[2] https://gist.github.com/zhr250/c36c2c54d9351df37e12fd072d4926ef

Signed-off-by: Hongru Zhang <zhanghongru@xxxxxxxxxx>
Suggested-by: Barry Song <baohua@xxxxxxxxxx>
Suggested-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
---
 mm/filemap.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 6afec636881f..a86355c5f1c7 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -3709,7 +3709,7 @@ vm_fault_t filemap_fault(struct vm_fault *vmf)
 		filemap_invalidate_unlock_shared(mapping);
 	if (fpin)
 		fput(fpin);
-	return ret | VM_FAULT_RETRY;
+	return ret | VM_FAULT_RETRY | VM_FAULT_MAY_USE_VMA_LOCK;
 }
 EXPORT_SYMBOL(filemap_fault);
 
-- 
2.43.0