| Message ID | 20230829081142.3619-10-urezki@gmail.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Mitigate a vmap lock contention v2 | expand |
On 08/29/23 at 10:11am, Uladzislau Rezki (Sony) wrote: ...... > real 1m28.382s > user 0m0.014s > sys 0m0.026s > urezki@pc638:~$ > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> > --- > mm/vmalloc.c | 26 ++++++++++++++++++++++++++ > 1 file changed, 26 insertions(+) LGTM, Reviewed-by: Baoquan He <bhe@redhat.com> > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c > index 9cce012aecdb..08990f630c21 100644 > --- a/mm/vmalloc.c > +++ b/mm/vmalloc.c > @@ -796,6 +796,9 @@ struct vmap_node { > atomic_t fill_in_progress; > }; > > +#define MAX_NODES U8_MAX > +#define MAX_NODE_SIZE SZ_4M > + > static struct vmap_node *nodes, snode; > static __read_mostly unsigned int nr_nodes = 1; > static __read_mostly unsigned int node_size = 1; > @@ -4803,11 +4806,24 @@ static void vmap_init_free_space(void) > } > } > > +static unsigned int calculate_nr_nodes(void) > +{ > + unsigned int nr_cpus; > + > + nr_cpus = num_present_cpus(); > + if (nr_cpus <= 1) > + nr_cpus = num_possible_cpus(); > + > + /* Density factor. Two users per a node. */ > + return clamp_t(unsigned int, nr_cpus >> 1, 1, MAX_NODES); > +} > + > static void vmap_init_nodes(void) > { > struct vmap_node *vn; > int i; > > + nr_nodes = calculate_nr_nodes(); > nodes = &snode; > > if (nr_nodes > 1) { > @@ -4830,6 +4846,16 @@ static void vmap_init_nodes(void) > INIT_LIST_HEAD(&vn->free.head); > spin_lock_init(&vn->free.lock); > } > + > + /* > + * Scale a node size to number of CPUs. Each power of two > + * value doubles a node size. A high-threshold limit is set > + * to 4M. > + */ > +#if BITS_PER_LONG == 64 > + if (nr_nodes > 1) > + node_size = min(SZ_64K << fls(num_possible_cpus()), SZ_4M); > +#endif > } > > void __init vmalloc_init(void) > -- > 2.30.2 >
On Fri, Sep 15, 2023 at 09:03:29PM +0800, Baoquan He wrote: > On 08/29/23 at 10:11am, Uladzislau Rezki (Sony) wrote: > ...... > > real 1m28.382s > > user 0m0.014s > > sys 0m0.026s > > urezki@pc638:~$ > > > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> > > --- > > mm/vmalloc.c | 26 ++++++++++++++++++++++++++ > > 1 file changed, 26 insertions(+) > > LGTM, > > Reviewed-by: Baoquan He <bhe@redhat.com> > Applied. Thank you! -- Uladzislau Rezki
diff --git a/mm/vmalloc.c b/mm/vmalloc.c index 9cce012aecdb..08990f630c21 100644 --- a/mm/vmalloc.c +++ b/mm/vmalloc.c @@ -796,6 +796,9 @@ struct vmap_node { atomic_t fill_in_progress; }; +#define MAX_NODES U8_MAX +#define MAX_NODE_SIZE SZ_4M + static struct vmap_node *nodes, snode; static __read_mostly unsigned int nr_nodes = 1; static __read_mostly unsigned int node_size = 1; @@ -4803,11 +4806,24 @@ static void vmap_init_free_space(void) } } +static unsigned int calculate_nr_nodes(void) +{ + unsigned int nr_cpus; + + nr_cpus = num_present_cpus(); + if (nr_cpus <= 1) + nr_cpus = num_possible_cpus(); + + /* Density factor. Two users per a node. */ + return clamp_t(unsigned int, nr_cpus >> 1, 1, MAX_NODES); +} + static void vmap_init_nodes(void) { struct vmap_node *vn; int i; + nr_nodes = calculate_nr_nodes(); nodes = &snode; if (nr_nodes > 1) { @@ -4830,6 +4846,16 @@ static void vmap_init_nodes(void) INIT_LIST_HEAD(&vn->free.head); spin_lock_init(&vn->free.lock); } + + /* + * Scale a node size to number of CPUs. Each power of two + * value doubles a node size. A high-threshold limit is set + * to 4M. + */ +#if BITS_PER_LONG == 64 + if (nr_nodes > 1) + node_size = min(SZ_64K << fls(num_possible_cpus()), SZ_4M); +#endif } void __init vmalloc_init(void)
The density ratio is set to 2, i.e. two users per one node. For example if there are 6 cores in a system the "nr_nodes" is 3. The "node_size" also depends on number of physical cores. A high-threshold limit is hard-coded and set to SZ_4M. For 32-bit, single/dual core systems an access to a global vmap heap is not balanced. Such small systems do not suffer from lock contentions due to limitation of CPU-cores. Test on AMD Ryzen Threadripper 3970X 32-Core Processor: sudo ./test_vmalloc.sh run_test_mask=127 nr_threads=64 <default perf> 94.17% 0.90% [kernel] [k] _raw_spin_lock 93.27% 93.05% [kernel] [k] native_queued_spin_lock_slowpath 74.69% 0.25% [kernel] [k] __vmalloc_node_range 72.64% 0.01% [kernel] [k] __get_vm_area_node 72.04% 0.89% [kernel] [k] alloc_vmap_area 42.17% 0.00% [kernel] [k] vmalloc 32.53% 0.00% [kernel] [k] __vmalloc_node 24.91% 0.25% [kernel] [k] vfree 24.32% 0.01% [kernel] [k] remove_vm_area 22.63% 0.21% [kernel] [k] find_unlink_vmap_area 15.51% 0.00% [unknown] [k] 0xffffffffc09a74ac 14.35% 0.00% [kernel] [k] ret_from_fork_asm 14.35% 0.00% [kernel] [k] ret_from_fork 14.35% 0.00% [kernel] [k] kthread <default perf> vs <patch-series perf> 74.32% 2.42% [kernel] [k] __vmalloc_node_range 69.58% 0.01% [kernel] [k] vmalloc 54.21% 1.17% [kernel] [k] __alloc_pages_bulk 48.13% 47.91% [kernel] [k] clear_page_orig 43.60% 0.01% [unknown] [k] 0xffffffffc082f16f 32.06% 0.00% [kernel] [k] ret_from_fork_asm 32.06% 0.00% [kernel] [k] ret_from_fork 32.06% 0.00% [kernel] [k] kthread 31.30% 0.00% [unknown] [k] 0xffffffffc082f889 22.98% 4.16% [kernel] [k] vfree 14.36% 0.28% [kernel] [k] __get_vm_area_node 13.43% 3.35% [kernel] [k] alloc_vmap_area 10.86% 0.04% [kernel] [k] remove_vm_area 8.89% 2.75% [kernel] [k] _raw_spin_lock 7.19% 0.00% [unknown] [k] 0xffffffffc082fba3 6.65% 1.37% [kernel] [k] free_unref_page 6.13% 6.11% [kernel] [k] native_queued_spin_lock_slowpath <patch-series perf> confirms that a native_queued_spin_lock_slowpath bottle-neck can be considered as negligible for the patch-series version. The throughput is ~15x higher: urezki@pc638:~$ time sudo ./test_vmalloc.sh run_test_mask=127 nr_threads=64 Run the test with following parameters: run_test_mask=127 nr_threads=64 Done. Check the kernel ring buffer to see the summary. real 24m3.305s user 0m0.361s sys 0m0.013s urezki@pc638:~$ urezki@pc638:~$ time sudo ./test_vmalloc.sh run_test_mask=127 nr_threads=64 Run the test with following parameters: run_test_mask=127 nr_threads=64 Done. Check the kernel ring buffer to see the summary. real 1m28.382s user 0m0.014s sys 0m0.026s urezki@pc638:~$ Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> --- mm/vmalloc.c | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)