[PATCH mm-unstable v2 16/16] mm: Respect mmap hint before THP alignment if allocation is possible
From: Kalesh Singh <hidden>
Date: 2024-12-11 23:28:28
Also in:
linux-alpha, linux-arm-kernel, linux-mips, linux-mm, linux-s390, linux-sh, lkml, loongarch, sparclinux
Subsystem:
memory management, memory management - thp (transparent huge page), memory mapping, the rest · Maintainers:
Andrew Morton, David Hildenbrand, Lorenzo Stoakes, Liam R. Howlett, Linus Torvalds
Commit 249608ee4713 ("mm: respect mmap hint address when aligning for THP")
fallsback to PAGE_SIZE alignment instead of THP alignment
for anonymous mapping as long as a hint address is provided by the user
-- even if we weren't able to allocate the unmapped area at the hint
address in the end.
This was done to address the immediate regression in anonymous mappings
where the hint address were being ignored in some cases; due to commit
efa7df3e3bb5 ("mm: align larger anonymous mappings on THP boundaries").
It was later pointed out that this issue also existed for file-backed
mappings from file systems that use thp_get_unmapped_area() for their
.get_unmapped_area() file operation.
The same fix was not applied for file-backed mappings since it would
mean any mmap requests that provide a hint address would be only
PAGE_SIZE-aligned regardless of whether allocation was successful at
the hint address or not.
Instead, use arch_mmap_hint() to first attempt allocation at the hint
address and fallback to THP alignment if there isn't sufficient VA space
to satisfy the allocation at the hint address.
Signed-off-by: Kalesh Singh <redacted>
---
mm/huge_memory.c | 17 ++++++++++-------
mm/mmap.c | 1 -
2 files changed, 10 insertions(+), 8 deletions(-)
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 2da5520bfe24..426761a30aff 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c@@ -1097,6 +1097,16 @@ static unsigned long __thp_get_unmapped_area(struct file *filp, loff_t off_align = round_up(off, size); unsigned long len_pad, ret, off_sub; + /* + * If allocation at the address hint succeeds; respect the hint and + * don't try to align to THP boundary; + * + * Or if an the requested extent is invalid return the error immediately. + */ + addr = arch_mmap_hint(filp, addr, len, off, flags); + if (addr) + return addr; + if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall()) return 0;
@@ -1117,13 +1127,6 @@ static unsigned long __thp_get_unmapped_area(struct file *filp, if (IS_ERR_VALUE(ret)) return 0; - /* - * Do not try to align to THP boundary if allocation at the address - * hint succeeds. - */ - if (ret == addr) - return addr; - off_sub = (off - ret) & (size - 1); if (test_bit(MMF_TOPDOWN, ¤t->mm->flags) && !off_sub)
diff --git a/mm/mmap.c b/mm/mmap.c
index 76dd6acdf051..3286fdff26f2 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c@@ -814,7 +814,6 @@ __get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, if (get_area) { addr = get_area(file, addr, len, pgoff, flags); } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && !file - && !addr /* no hint */ && IS_ALIGNED(len, PMD_SIZE)) { /* Ensures that larger anonymous mappings are THP aligned. */ addr = thp_get_unmapped_area_vmflags(file, addr, len,
--
2.47.0.338.g60cca15819-goog