On Mon, Nov 11, 2024 at 6:48 PM 'Liam R. Howlett' via kernel-team
<kernel-team@xxxxxxxxxxx> wrote:
>
> * Suren Baghdasaryan <surenb@xxxxxxxxxx> [241111 16:41]:
> > On Mon, Nov 11, 2024 at 12:55 PM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > >
> > > Back when per-vma locks were introduces, vm_lock was moved out of
> > > vm_area_struct in [1] because of the performance regression caused by
> > > false cacheline sharing. Recent investigation [2] revealed that the
> > > regressions is limited to a rather old Broadwell microarchitecture and
> > > even there it can be mitigated by disabling adjacent cacheline
> > > prefetching, see [3].
> > > This patchset moves vm_lock back into vm_area_struct, aligning it at the
> > > cacheline boundary and changing the cache to be cache-aligned as well.
> > > This causes VMA memory consumption to grow from 160 (vm_area_struct) + 40
> > > (vm_lock) bytes to 256 bytes:
> > >
> > > slabinfo before:
> > > <name> ... <objsize> <objperslab> <pagesperslab> : ...
> > > vma_lock ... 40 102 1 : ...
> > > vm_area_struct ... 160 51 2 : ...
> > >
> > > slabinfo after moving vm_lock:
> > > <name> ... <objsize> <objperslab> <pagesperslab> : ...
> > > vm_area_struct ... 256 32 2 : ...
> > >
> > > Aggregate VMA memory consumption per 1000 VMAs grows from 50 to 64 pages,
> > > which is 5.5MB per 100000 VMAs.
> > > To minimize memory overhead, vm_lock implementation is changed from
> > > using rw_semaphore (40 bytes) to an atomic (8 bytes) and several
> > > vm_area_struct members are moved into the last cacheline, resulting
> > > in a less fragmented structure:
>
> Wait a second, this is taking 40B down to 8B, but the alignment of the
> vma will surely absorb that 32B difference? The struct sum is 153B
> according to what you have below so we won't go over 192B. What am I
> missing?
Take a look at the last patch in the series called "[PATCH 4/4] mm:
move lesser used vma_area_struct members into the last cacheline". I
move some struct members from the earlier cachelines into cacheline #4
where the vm_lock is staying.
>
> > >
> > > struct vm_area_struct {
> > > union {
> > > struct {
> > > long unsigned int vm_start; /* 0 8 */
> > > long unsigned int vm_end; /* 8 8 */
> > > }; /* 0 16 */
> > > struct callback_head vm_rcu ; /* 0 16 */
> > > } __attribute__((__aligned__(8))); /* 0 16 */
> > > struct mm_struct * vm_mm; /* 16 8 */
> > > pgprot_t vm_page_prot; /* 24 8 */
> > > union {
> > > const vm_flags_t vm_flags; /* 32 8 */
> > > vm_flags_t __vm_flags; /* 32 8 */
> > > }; /* 32 8 */
> > > bool detached; /* 40 1 */
> > >
> > > /* XXX 3 bytes hole, try to pack */
> > >
> > > unsigned int vm_lock_seq; /* 44 4 */
> > > struct list_head anon_vma_chain; /* 48 16 */
> > > /* --- cacheline 1 boundary (64 bytes) --- */
> > > struct anon_vma * anon_vma; /* 64 8 */
> > > const struct vm_operations_struct * vm_ops; /* 72 8 */
> > > long unsigned int vm_pgoff; /* 80 8 */
> > > struct file * vm_file; /* 88 8 */
> > > void * vm_private_data; /* 96 8 */
> > > atomic_long_t swap_readahead_info; /* 104 8 */
> > > struct mempolicy * vm_policy; /* 112 8 */
> > >
> > > /* XXX 8 bytes hole, try to pack */
> > >
> > > /* --- cacheline 2 boundary (128 bytes) --- */
> > > struct vma_lock vm_lock (__aligned__(64)); /* 128 4 */
> > >
> > > /* XXX 4 bytes hole, try to pack */
> > >
> > > struct {
> > > struct rb_node rb (__aligned__(8)); /* 136 24 */
> > > long unsigned int rb_subtree_last; /* 160 8 */
> > > } __attribute__((__aligned__(8))) shared; /* 136 32 */
> > > struct vm_userfaultfd_ctx vm_userfaultfd_ctx; /* 168 0 */
> > >
> > > /* size: 192, cachelines: 3, members: 17 */
> > > /* sum members: 153, holes: 3, sum holes: 15 */
> > > /* padding: 24 */
> > > /* forced alignments: 3, forced holes: 2, sum forced holes: 12 */
> > > } __attribute__((__aligned__(64)));
> > >
> > > Memory consumption per 1000 VMAs becomes 48 pages, saving 2 pages compared
> > > to the 50 pages in the baseline:
> > >
> > > slabinfo after vm_area_struct changes:
> > > <name> ... <objsize> <objperslab> <pagesperslab> : ...
> > > vm_area_struct ... 192 42 2 : ...
> > >
> > > Performance measurements using pft test on x86 do not show considerable
> > > difference, on Pixel 6 running Android it results in 3-5% improvement in
> > > faults per second.
> > >
> > > [1] https://lore.kernel.org/all/20230227173632.3292573-34-surenb@xxxxxxxxxx/
> > > [2] https://lore.kernel.org/all/ZsQyI%2F087V34JoIt@xsang-OptiPlex-9020/
> > > [3] https://lore.kernel.org/all/CAJuCfpEisU8Lfe96AYJDZ+OM4NoPmnw9bP53cT_kbfP_pR+-2g@xxxxxxxxxxxxxx/
> >
> > And of course I forgot to update Lorenzo's new locking documentation :/
> > Will add that in the next version.
> >
> > >
> > > Suren Baghdasaryan (4):
> > > mm: introduce vma_start_read_locked{_nested} helpers
> > > mm: move per-vma lock into vm_area_struct
> > > mm: replace rw_semaphore with atomic_t in vma_lock
> > > mm: move lesser used vma_area_struct members into the last cacheline
> > >
> > > include/linux/mm.h | 163 +++++++++++++++++++++++++++++++++++---
> > > include/linux/mm_types.h | 59 +++++++++-----
> > > include/linux/mmap_lock.h | 3 +
> > > kernel/fork.c | 50 ++----------
> > > mm/init-mm.c | 2 +
> > > mm/userfaultfd.c | 14 ++--
> > > 6 files changed, 205 insertions(+), 86 deletions(-)
> > >
> > >
> > > base-commit: 931086f2a88086319afb57cd3925607e8cda0a9f
> > > --
> > > 2.47.0.277.g8800431eea-goog
> > >
>
> To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@xxxxxxxxxxx.
>
