On Wed, Jan 15, 2025 at 8:39 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Wed, Jan 15, 2025 at 2:51 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Jan 10, 2025 at 08:25:59PM -0800, Suren Baghdasaryan wrote:
> > > Move several vma_area_struct members which are rarely or never used
> > > during page fault handling into the last cacheline to better pack
> > > vm_area_struct. As a result vm_area_struct will fit into 3 as opposed
> > > to 4 cachelines. New typical vm_area_struct layout:
> > >
> > > struct vm_area_struct {
> > > union {
> > > struct {
> > > long unsigned int vm_start; /* 0 8 */
> > > long unsigned int vm_end; /* 8 8 */
> > > }; /* 0 16 */
> > > freeptr_t vm_freeptr; /* 0 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 */
> > > unsigned int vm_lock_seq; /* 40 4 */
> >
> > Does it not make sense to move this seq field near the refcnt?
>
> In an earlier version, when vm_lock was not a refcount yet, I tried
> that and moving vm_lock_seq introduced regression in the pft test. We
> have that early vm_lock_seq check in the beginning of vma_start_read()
> and if it fails we bail out early without locking. I think that might
> be the reason why keeping vm_lock_seq in the first cacheling is
> beneficial. But I'll try moving it again now that we have vm_refcnt
> instead of the lock and see if pft still shows any regression.
I confirmed that moving vm_lock_seq next to vm_refcnt regresses
pagefault performance:
Hmean faults/cpu-1 508634.6876 ( 0.00%) 508548.5498 * -0.02%*
Hmean faults/cpu-4 474767.2684 ( 0.00%) 475620.7653 * 0.18%*
Hmean faults/cpu-7 451356.6844 ( 0.00%) 446738.2381 * -1.02%*
Hmean faults/cpu-12 360114.9092 ( 0.00%) 337121.8189 * -6.38%*
Hmean faults/cpu-21 227567.8237 ( 0.00%) 205277.2029 * -9.80%*
Hmean faults/cpu-30 163383.6765 ( 0.00%) 152765.1451 * -6.50%*
Hmean faults/cpu-48 118048.2568 ( 0.00%) 109959.2027 * -6.85%*
Hmean faults/cpu-56 103189.6761 ( 0.00%) 92989.3749 * -9.89%*
Hmean faults/sec-1 508228.4512 ( 0.00%) 508129.1963 * -0.02%*
Hmean faults/sec-4 1854868.9033 ( 0.00%) 1862443.6146 * 0.41%*
Hmean faults/sec-7 3088881.6158 ( 0.00%) 3050403.1664 * -1.25%*
Hmean faults/sec-12 4222540.9948 ( 0.00%) 3951163.9557 * -6.43%*
Hmean faults/sec-21 4555777.5386 ( 0.00%) 4130470.6021 * -9.34%*
Hmean faults/sec-30 4336721.3467 ( 0.00%) 4150477.5095 * -4.29%*
Hmean faults/sec-48 5163921.7465 ( 0.00%) 4857286.2166 * -5.94%*
Hmean faults/sec-56 5413622.8890 ( 0.00%) 4936484.0021 * -8.81%*
So, I kept it unchanged in v10
(https://lore.kernel.org/all/20250213224655.1680278-14-surenb@xxxxxxxxxx/)
>
> >
> > > /* XXX 4 bytes hole, try to pack */
> > >
> > > 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 */
> > > struct vma_numab_state * numab_state; /* 120 8 */
> > > /* --- cacheline 2 boundary (128 bytes) --- */
> > > refcount_t vm_refcnt (__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 anon_vma_name * anon_name; /* 168 8 */
> > > struct vm_userfaultfd_ctx vm_userfaultfd_ctx; /* 176 8 */
> > >
> > > /* size: 192, cachelines: 3, members: 18 */
> > > /* sum members: 176, holes: 2, sum holes: 8 */
> > > /* padding: 8 */
> > > /* forced alignments: 2, forced holes: 1, sum forced holes: 4 */
> > > } __attribute__((__aligned__(64)));
> >
> >
