On Mon, Dec 9, 2024 at 7:37 PM Yang Shi [off-list ref] wrote:

On Mon, Dec 9, 2024 at 6:45 PM Kalesh Singh [off-list ref] wrote:

quoted

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 that fails.

Thanks for taking time to try to fix this.

quoted

Signed-off-by: Kalesh Singh <redacted>
---
 mm/huge_memory.c | 15 ++++++++-------
 mm/mmap.c        |  1 -
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 137abeda8602..f070c89dafc9 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c

@@ -1097,6 +1097,14 @@ 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.
+        */
+       addr = arch_mmap_hint(filp, addr, len, off, flags);
+       if (addr)
+               return addr;
+

Hi Yang,

Thanks for the comments.

IIUC, arch_mmap_hint() will be called in arch_get_unmapped_area() and
arch_get_unmapped_area_topdown() again. So we will actually look up
maple tree twice. It sounds like the second hint address search is
pointless. You should be able to set addr to 0 before calling
mm_get_unmapped_area_vmflags() in order to skip the second hint
address search.

You are right that it would call into arch_mmap_hint() twice but it
only attempts the lookup once since on the second attempt addr == 0.

Thanks,
Kalesh

quoted

        if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall())
                return 0;

@@ -1117,13 +1125,6 @@ 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, &current->mm->flags) && !off_sub)

diff --git a/mm/mmap.c b/mm/mmap.c
index 59bf7d127aa1..6bfeec80152a 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c

@@ -807,7 +807,6 @@ __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