ovl: implement ->get_unmapped_area()
Subsystems (2)
PATCH 1/1 ovl: implement ->get_unmapped_area()
ovl_mmap() passes realfile to the underlying filesystem, so the VMA it returns is backed by realfile, not by the overlayfs file.
Address selection is not delegated. __get_unmapped_area() looks at the overlayfs file's f_op, ovl_file_operations does not implement ->get_unmapped_area(), and the mapping lands wherever mm_get_unmapped_area() puts it. ext4, xfs and btrfs all point ->get_unmapped_area() at thp_get_unmapped_area(); none of that is reached.
vaddr - file_offset is then not a multiple of PMD_SIZE, thp_vma_suitable_order() rejects the VMA, and the PMD-sized folios the underlying filesystem already has in the page cache are never mapped by a PMD.
Delegate to the file ovl_mmap() will use. Calling thp_get_unmapped_area() here instead would force alignment on filesystems that deliberately do not implement ->get_unmapped_area(), which commit 34d7cf637c43 ("mm:
don't try THP alignment for FS without get_unmapped_area") avoided. fs/proc delegates the same way in pde_get_unmapped_area().
Reading a 1025M ext4 file through mmap(NULL, ...), 97% of its page cache in PMD-sized folios, 15 runs:
PMD congruent minor faults page tables
ext4, directly 15/15 764 68K
overlayfs, before 0/5 16186 2056K
overlayfs, after 15/15 764 68K
Fault counts had 0 stddev. Median read time 17.4ms -> 4.80ms, against 4.92ms for the same file on ext4.
--- fs/overlayfs/file.c | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)
fs/overlayfs/file.c +26 -0
@@ -13,6 +13,7 @@#include <linux/security.h> #include <linux/fs.h> #include <linux/backing-file.h> +#include <linux/sched/mm.h> #include "overlayfs.h" static char ovl_whatisit (struct inode *inode,struct inode *realinode)
@@ -465,6 +466,30 @@ static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync)return vfs_fsync_range (upperfile, start, end, datasync); } +static unsigned long ovl_get_unmapped_area (struct file *file, +unsigned long addr,unsigned long len, +unsigned long pgoff,unsigned long flags) +{ +struct ovl_file *of = file->private_data; +struct file *realfile = of->realfile; + +/* +* ovl_mmap() hands realfile to the underlying filesystem, so the vma +* ends up backed by realfile. Let that filesystem pick the address +* too, or one that needs a specific alignment - to allow PMD mappings, +* for example - never gets asked for one. +*/ +if (realfile->f_op->get_unmapped_area) +return realfile->f_op->get_unmapped_area (realfile, addr, len, + pgoff, flags); + +#ifdef CONFIG_MMU +return mm_get_unmapped_area (file, addr, len, pgoff, flags); +#endif + +return addr; +} +static int ovl_mmap (struct file *file,struct vm_area_struct *vma) {struct ovl_file *of = file->private_data;
@@ -654,6 +679,7 @@ const struct file_operations ovl_file_operations = { .write_iter = ovl_write_iter, .fsync = ovl_fsync, .mmap = ovl_mmap, + .get_unmapped_area = ovl_get_unmapped_area, .fallocate = ovl_fallocate, .fadvise = ovl_fadvise, .flush = ovl_flush,
base-commit: 26260251022fbc2f248a3d747a9b2b961b18d2d8
Amir Goldstein
maintainerQuoted 65 lines · Yuan-Hao Hsu
ovl_mmap()passes realfile to the underlying filesystem, so the VMA it returns is backed by realfile, not by the overlayfs file.Address selection is not delegated.
__get_unmapped_area()looks at the overlayfs file'sf_op,ovl_file_operationsdoes not implement ->get_unmapped_area(), and the mapping lands wherevermm_get_unmapped_area()puts it. ext4, xfs and btrfs all point ->get_unmapped_area() atthp_get_unmapped_area(); none of that is reached.vaddr -
file_offsetis then not a multiple ofPMD_SIZE,thp_vma_suitable_order()rejects the VMA, and the PMD-sized folios the underlying filesystem already has in the page cache are never mapped by a PMD.Delegate to the file
ovl_mmap()will use. Callingthp_get_unmapped_area()here instead would force alignment on filesystems that deliberately do not implement ->get_unmapped_area(), which commit34d7cf637c43("mm:
don't try THP alignment for FS withoutget_unmapped_area") avoided. fs/proc delegates the same way inpde_get_unmapped_area().Reading a 1025M ext4 file through
mmap(NULL, ...), 97% of its page cache in PMD-sized folios, 15 runs:PMD congruent minor faults page tables
ext4, directly 15/15 764 68K
overlayfs, before 0/5 16186 2056K
overlayfs, after 15/15 764 68KFault counts had 0 stddev. Median read time 17.4ms -> 4.80ms, against 4.92ms for the same file on ext4.
Assisted-by Claude:claude-opus-5Signed-off-by Yuan-Hao Hsu <[email protected]>--- fs/overlayfs/file.c | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)fs/overlayfs/file.c +11
-0@@ -13,6 +13,7 @@#include <linux/security.h> #include <linux/fs.h> #include <linux/backing-file.h> +#include <linux/sched/mm.h> #include "overlayfs.h" static char ovl_whatisit (struct inode *inode,struct inode *realinode)@@ -465,6 +466,30 @@ static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync)return vfs_fsync_range (upperfile, start, end, datasync); } +static unsigned long ovl_get_unmapped_area (struct file *file, +unsigned long addr,unsigned long len, +unsigned long pgoff,unsigned long flags) +{ +struct ovl_file *of = file->private_data; +struct file *realfile = of->realfile; + +/* +* ovl_mmap() hands realfile to the underlying filesystem, so the vma +* ends up backed by realfile. Let that filesystem pick the address
fs/overlayfs/file.c static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync) · line 488
Quoted 10 lines
fs/overlayfs/file.c +10
-0+* too, or one that needs a specific alignment - to allow PMD mappings, +* for example - never gets asked for one. +*/ +if (realfile->f_op->get_unmapped_area) +return realfile->f_op->get_unmapped_area (realfile, addr, len, + pgoff, flags); + +#ifdef CONFIG_MMU +return mm_get_unmapped_area (file, addr, len, pgoff, flags); +#endif
That looks quite ugly.
Can't we have a noop inline helper for mm_get_unmapped_area() fo nommu?
Quoted 16 lines
fs/overlayfs/file.c +5
-0+ +return addr; +} +static int ovl_mmap (struct file *file,struct vm_area_struct *vma) {struct ovl_file *of = file->private_data;@@ -654,6 +679,7 @@ const struct file_operations ovl_file_operations = { .write_iter = ovl_write_iter, .fsync = ovl_fsync, .mmap = ovl_mmap, + .get_unmapped_area = ovl_get_unmapped_area, .fallocate = ovl_fallocate, .fadvise = ovl_fadvise, .flush = ovl_flush,
Quoted 1 line
base-commit:
26260251022fbc2f248a3d747a9b2b961b18d2d8
Since mmap is going through backing_file_mmap() probably a good idea to pass this method through a backing_file helper as well.
But more importantly, adding MM people to this patch review, mainly the ones that participated in the discussion over patch to fix similar issues in 2024 [1].
Most of the concerns in that patch seem to have been addressed (?)
Thanks,
Amir.
[1] https://lore.kernel.org/linux-unionfs/[email protected]/
Quoted 9 lines · Amir Goldstein
Since mmap is going through
backing_file_mmap()probably a good idea to pass this method through abacking_filehelper as well.But more importantly, adding MM people to this patch review, mainly the ones that participated in the discussion over patch to fix similar issues in 2024 [1].
Most of the concerns in that patch seem to have been addressed (?)[1] https://lore.kernel.org/linux-unionfs/[email protected]/
No, not at all. All this patch needs to be is:
+++ b/fs/overlayfs/file.c
@@ -657,6 +657,7 @@ const struct file_operations ovl_file_operations = {
.fallocate = ovl_fallocate,
.fadvise = ovl_fadvise,
.flush = ovl_flush,+ .get_unmapped_area = thp_get_unmapped_area,
.splice_read = ovl_splice_read,
.splice_write = ovl_splice_write,
All this "oh we have to call the underlying filesystem's get_unmapepd_area" betrays a lack of understanding of the problem.
And Lorenzo suggested this in the linked thread:
Miklos Szeredi
maintainerQuoted 12 lines · Matthew Wilcox
No, not at all. All this patch needs to be is:
+++
b/fs/overlayfs/file.c(quoted diff) +1
-0@@ -657,6 +657,7 @@ const struct file_operations ovl_file_operations = { .fallocate = ovl_fallocate, .fadvise = ovl_fadvise, .flush = ovl_flush, + .get_unmapped_area = thp_get_unmapped_area, .splice_read = ovl_splice_read, .splice_write = ovl_splice_write,
Quoted 6 lines
All this "oh we have to call the underlying filesystem's
get_unmapepd_area" betrays a lack of understanding of the problem.And Lorenzo suggested this in the linked thread:
There seems to be a logical fallacy in saying thp_get_unmapepd_area is the most suitable value, yet not making it the default for all filesystems.
Saying this without any understanding of the subtle difference between
setting .get_unmapepd_area to NULL and to thp_get_unmapepd_area. All
I can see that the difference is indeed subtle.
Can you please explain?
Thanks,
Miklos
Quoted 9 lines · Miklos Szeredi
There seems to be a logical fallacy in saying
thp_get_unmapepd_areais the most suitable value, yet not making it the default for all filesystems.Saying this without any understanding of the subtle difference between
setting .get_unmapepd_area to NULL and tothp_get_unmapepd_area. All
I can see that the difference is indeed subtle.Can you please explain?
I had a look at this. thp_get_unmapped_area() aligns anything with room for a whole PMD in it; nothing asks whether the file can use it. NULL means no alignment at all, and 34d7cf637c43 kept files there to avoid "unnecessary VMA fragmentation and potentially worse performance on filesystems that do not actually support THPs". The case where that matters is a lower without large folios. 32 x 3M files, mmap(NULL, ...):
ovl over ext4 ovl over squashfs
NULL (today) 16/32 aligned, no gaps 16/32 aligned, no gaps delegate (v1) 32/32, 31744K of gaps 16/32, no gaps thp_get_unmapped_area 32/32, 31744K of gaps 32/32, 31744K of gaps
squashfs has no large folios, so the alignment buys nothing there (fault counts were identical) and costs 1M of address space per mapping. v1 only avoids that by inheriting the lower's choice. fuse passthrough already pays the same, as Jann noted in the 2024 thread.
Hope this helps.
Thanks,
Yuan-Hao Hsu