mm: batch unlink_file_vma calls in free_pgd_range
Subsystems (3)
PATCH 1/1 mm: batch unlink_file_vma calls in free_pgd_range
Execs of dynamically linked binaries at 20-ish cores are bottlenecked on the i_mmap_rwsem semaphore, while the biggest singular contributor is free_pgd_range inducing the lock acquire back-to-back for all consecutive mappings of a given file.
Tracing the count of said acquires while building the kernel shows:
[1, 2) 799579 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| [2, 3) 0 | | [3, 4) 3009 | | [4, 5) 3009 | | [5, 6) 326442 |@@@@@@@@@@@@@@@@@@@@@ |
So in particular there were 326442 opportunities to coalesce 5 acquires into 1.
Doing so increases execs per second by 4% (~50k to ~52k) when running the benchmark linked below.
The lock remains the main bottleneck, I have not looked at other spots yet.
Bench can be found here:
http://apollo.backplane.com/DFlyMisc/doexec.c
$ cc -O2 -o shared-doexec doexec.c $ ./shared-doexec $(nproc)
Note this particular test makes sure binaries are separate, but the loader is shared.
Stats collected on the patched kernel (+ "noinline") with:
bpftrace -e 'kprobe:unlink_file_vma_batch_process
{ @ = lhist(((struct unlink_vma_file_batch *)arg0)->count, 0, 8, 1); }'
--- include/linux/mm.h | 8 ++++++++ mm/memory.c | 10 ++++++++-- mm/mmap.c | 41 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 57 insertions(+), 2 deletions(-)
include/linux/mm.h +8 -0
@@ -3272,6 +3272,11 @@ void anon_vma_interval_tree_verify(struct anon_vma_chain *node); avc; avc = anon_vma_interval_tree_iter_next(avc, start, last))/* mmap.c */ +struct unlink_vma_file_batch { +int count; +struct vm_area_struct *vmas[8 ]; +}; +extern int __vm_enough_memory (struct mm_struct *mm,long pages,int cap_sys_admin); extern int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma, unsigned long start, unsigned long end, pgoff_t pgoff,
@@ -3281,6 +3286,9 @@ extern int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,extern struct anon_vma *find_mergeable_anon_vma (struct vm_area_struct *);extern int insert_vm_struct (struct mm_struct *,struct vm_area_struct *);extern void unlink_file_vma (struct vm_area_struct *); +void unlink_file_vma_batch_init (struct unlink_vma_file_batch *); +void unlink_file_vma_batch_add (struct unlink_vma_file_batch *,struct vm_area_struct *); +void unlink_file_vma_batch_final (struct unlink_vma_file_batch *);extern struct vm_area_struct *copy_vma (struct vm_area_struct **,unsigned long addr,unsigned long len,pgoff_t pgoff,bool *need_rmap_locks);
mm/memory.c +8 -2
@@ -363,6 +363,8 @@ void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas, struct vm_area_struct *vma, unsigned long floor, unsigned long ceiling, bool mm_wr_locked) { + struct unlink_vma_file_batch vb; + do { unsigned long addr = vma->vm_start; struct vm_area_struct *next;
@@ -382,12 +384,15 @@ void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,if (mm_wr_locked)vma_start_write (vma);unlink_anon_vmas (vma); -unlink_file_vma (vma);if (is_vm_hugetlb_page (vma)) { +unlink_file_vma (vma);hugetlb_free_pgd_range (tlb, addr, vma->vm_end, floor, next ? next->vm_start : ceiling); }else { +unlink_file_vma_batch_init (&vb); +unlink_file_vma_batch_add (&vb, vma); +/* * Optimization: gather nearby vmas into one call down */
@@ -400,8 +405,9 @@ void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,if (mm_wr_locked)vma_start_write (vma);unlink_anon_vmas (vma); -unlink_file_vma (vma); +unlink_file_vma_batch_add (&vb, vma); } +unlink_file_vma_batch_final (&vb);free_pgd_range (tlb, addr, vma->vm_end, floor, next ? next->vm_start : ceiling); }
mm/mmap.c +41 -0
47 lines not on this page · raw message
Signature · 1 line
2.39.2
Quoted 11 lines · Mateusz Guzik
Execs of dynamically linked binaries at 20-ish cores are bottlenecked on the
i_mmap_rwsemsemaphore, while the biggest singular contributor isfree_pgd_rangeinducing the lock acquire back-to-back for all consecutive mappings of a given file.Tracing the count of said acquires while building the kernel shows:
[1, 2) 799579 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| [2, 3) 0 | | [3, 4) 3009 | | [4, 5) 3009 | | [5, 6) 326442 |@@@@@@@@@@@@@@@@@@@@@ |
This makes sense. A snippet of /proc/self/maps:
7f0a44725000-7f0a4474b000 r--p 00000000 fe:01 100663437 /usr/lib/x86_64-linux-gnu/libc.so.6 7f0a4474b000-7f0a448a0000 r-xp 00026000 fe:01 100663437 /usr/lib/x86_64-linux-gnu/libc.so.6 7f0a448a0000-7f0a448f4000 r--p 0017b000 fe:01 100663437 /usr/lib/x86_64-linux-gnu/libc.so.6 7f0a448f4000-7f0a448f8000 r--p 001cf000 fe:01 100663437 /usr/lib/x86_64-linux-gnu/libc.so.6 7f0a448f8000-7f0a448fa000 rw-p 001d3000 fe:01 100663437 /usr/lib/x86_64-linux-gnu/libc.so.6
so we frequently have the same file mmaped five times in a row.
Quoted 2 lines
The lock remains the main bottleneck, I have not looked at other spots yet.
You're not the first to report high contention on this lock.
https://lore.kernel.org/all/[email protected]/
for example.
include/linux/mm.h
Quoted 4 lines
include/linux/mm.h +0
-0
I do object to this going into mm.h. mm/internal.h would be better.
I haven't reviewed the patch in depth, but I don't have a problem with the idea. I think it's only a stopgap and we really do need a better data structure than this.
Quoted 6 lines · Matthew Wilcox
Quoted 5 lines · Mateusz Guzik
Execs of dynamically linked binaries at 20-ish cores are bottlenecked on the
i_mmap_rwsemsemaphore, while the biggest singular contributor isfree_pgd_rangeinducing the lock acquire back-to-back for all consecutive mappings of a given file.
Quoted 7 lines
include/linux/mm.h
Quoted 4 lines
include/linux/mm.h +0
-0I do object to this going into
mm.h.mm/internal.hwould be better.
Noted.
Quoted 4 lines
I haven't reviewed the patch in depth, but I don't have a problem with the idea. I think it's only a stopgap and we really do need a better data structure than this.
I'll send a v2 after some more reviews pour in.
The above indeed is just a low hanging fruit fixup in an unpleasant situation.
I think the real fix in the long run would provide the loader with means to be more efficient about it.
strace /bin/echo shows:
[snip]openat(AT_FDCWD, "/lib/x86_64-linux-gnu/libc.so.6", O_RDONLY|O_CLOEXEC) = 3read(3, "\177ELF\2\1\1\3\0\0\0\0\0\0\0\0\3\0>\0\1\0\0\0\220\243\2\0\0\0\0\0"..., 832) = 832 pread64(3, "\6\0\0\0\4\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0"..., 784, 64) = 784 fstat(3, {st_mode=S_IFREG|0755, st_size=2125328, ...}) = 0pread64(3, "\6\0\0\0\4\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0@\0\0\0\0\0\0\0"..., 784, 64) = 784 mmap(NULL, 2170256, PROT_READ, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x7dbda8a00000 mmap(0x7dbda8a28000, 1605632, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x28000) = 0x7dbda8a28000
mmap(0x7dbda8bb0000, 323584, PROT_READ, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x1b0000) = 0x7dbda8bb0000 mmap(0x7dbda8bff000, 24576, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x1fe000) = 0x7dbda8bff000 mmap(0x7dbda8c05000, 52624, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7dbda8c05000
[/snip]
Hence the 5 mappings.
Should there be a mechanism to issue all these mmaps at the same time there would definitely be savings in total work done, not only in terms of one i_mmap_rwsem lock trip.
The mechanism should be versatile enough to replace other back-to-back mmap uses. It would be great if on top of it it did not require the size argument, instead it could return a pair address + size. Then the typical combo of open + fstat + mmap could be shortened.
As in that was just a quick note, I have no intention of pursuing anything of the sort. I'll probably submit some other patches to damage-control the state without altering any design choices.