| Message ID | 20240716111346.3676969-7-rppt@kernel.org |
|---|---|
| State | New |
| Headers | show |
| Series | mm: introduce numa_memblks | expand |
On Tue, 16 Jul 2024 14:13:35 +0300 Mike Rapoport <rppt@kernel.org> wrote: > From: "Mike Rapoport (Microsoft)" <rppt@kernel.org> > > Allocation of numa_distance uses memblock_phys_alloc_range() to limit > allocation to be below the last mapped page. > > But NUMA initializaition runs after the direct map is populated and initialization (one too many 'i's) > there is also code in setup_arch() that adjusts memblock limit to > reflect how much memory is already mapped in the direct map. > > Simplify the allocation of numa_distance and use plain memblock_alloc(). > This makes the code clearer and ensures that when numa_distance is not > allocated it is always NULL. Doesn't this break the comment in numa_set_distance() kernel-doc? " * If such table cannot be allocated, a warning is printed and further * calls are ignored until the distance table is reset with * numa_reset_distance(). " Superficially that looks to be to avoid repeatedly hitting the singleton bit at the top of numa_set_distance() as SRAT or similar parsing occurs. > > Signed-off-by: Mike Rapoport (Microsoft) <rppt@kernel.org> > --- > arch/x86/mm/numa.c | 12 +++--------- > 1 file changed, 3 insertions(+), 9 deletions(-) > > diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c > index 5e1dde26674b..ab2d4ecef786 100644 > --- a/arch/x86/mm/numa.c > +++ b/arch/x86/mm/numa.c > @@ -319,8 +319,7 @@ void __init numa_reset_distance(void) > { > size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); > > - /* numa_distance could be 1LU marking allocation failure, test cnt */ > - if (numa_distance_cnt) > + if (numa_distance) > memblock_free(numa_distance, size); > numa_distance_cnt = 0; > numa_distance = NULL; /* enable table creation */ > @@ -331,7 +330,6 @@ static int __init numa_alloc_distance(void) > nodemask_t nodes_parsed; > size_t size; > int i, j, cnt = 0; > - u64 phys; > > /* size the new table and allocate it */ > nodes_parsed = numa_nodes_parsed; > @@ -342,16 +340,12 @@ static int __init numa_alloc_distance(void) > cnt++; > size = cnt * cnt * sizeof(numa_distance[0]); > > - phys = memblock_phys_alloc_range(size, PAGE_SIZE, 0, > - PFN_PHYS(max_pfn_mapped)); > - if (!phys) { > + numa_distance = memblock_alloc(size, PAGE_SIZE); > + if (!numa_distance) { > pr_warn("Warning: can't allocate distance table!\n"); > - /* don't retry until explicitly reset */ > - numa_distance = (void *)1LU; > return -ENOMEM; > } > > - numa_distance = __va(phys); > numa_distance_cnt = cnt; > > /* fill with the default distances */
On Fri, Jul 19, 2024 at 05:28:49PM +0100, Jonathan Cameron wrote: > On Tue, 16 Jul 2024 14:13:35 +0300 > Mike Rapoport <rppt@kernel.org> wrote: > > > From: "Mike Rapoport (Microsoft)" <rppt@kernel.org> > > > > Allocation of numa_distance uses memblock_phys_alloc_range() to limit > > allocation to be below the last mapped page. > > > > But NUMA initializaition runs after the direct map is populated and > > initialization (one too many 'i's) Thanks. > > there is also code in setup_arch() that adjusts memblock limit to > > reflect how much memory is already mapped in the direct map. > > > > Simplify the allocation of numa_distance and use plain memblock_alloc(). > > This makes the code clearer and ensures that when numa_distance is not > > allocated it is always NULL. > Doesn't this break the comment in numa_set_distance() kernel-doc? > " > * If such table cannot be allocated, a warning is printed and further > * calls are ignored until the distance table is reset with > * numa_reset_distance(). > " > > Superficially that looks to be to avoid repeatedly hitting the > singleton bit at the top of numa_set_distance() as SRAT or similar > parsing occurs. I believe it's there to avoid allocation of numa_distance in the middle of distance parsing (SLIT or DT numa-distance-map). If the allocation fails for the first element in the table, the numa_distance and numa_distance_cnt remain zero and node_distance() falls back to return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; It's different from arch_numa that always tries to allocate MAX_NUMNODES * MAX_NUMNODES for numa_distance and treats the allocation failure as a failure to initialize NUMA. I like the general approach x86 uses more, i.e. in case distance parsing fails in some way NUMA is still initialized with probably suboptimal distances between nodes. I'm going to restore that "singleton" behavior for now and will look into making this all less cumbersome later. > > Signed-off-by: Mike Rapoport (Microsoft) <rppt@kernel.org> > > --- > > arch/x86/mm/numa.c | 12 +++--------- > > 1 file changed, 3 insertions(+), 9 deletions(-) > > > > diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c > > index 5e1dde26674b..ab2d4ecef786 100644 > > --- a/arch/x86/mm/numa.c > > +++ b/arch/x86/mm/numa.c > > @@ -319,8 +319,7 @@ void __init numa_reset_distance(void) > > { > > size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); > > > > - /* numa_distance could be 1LU marking allocation failure, test cnt */ > > - if (numa_distance_cnt) > > + if (numa_distance) > > memblock_free(numa_distance, size); > > numa_distance_cnt = 0; > > numa_distance = NULL; /* enable table creation */ > > @@ -331,7 +330,6 @@ static int __init numa_alloc_distance(void) > > nodemask_t nodes_parsed; > > size_t size; > > int i, j, cnt = 0; > > - u64 phys; > > > > /* size the new table and allocate it */ > > nodes_parsed = numa_nodes_parsed; > > @@ -342,16 +340,12 @@ static int __init numa_alloc_distance(void) > > cnt++; > > size = cnt * cnt * sizeof(numa_distance[0]); > > > > - phys = memblock_phys_alloc_range(size, PAGE_SIZE, 0, > > - PFN_PHYS(max_pfn_mapped)); > > - if (!phys) { > > + numa_distance = memblock_alloc(size, PAGE_SIZE); > > + if (!numa_distance) { > > pr_warn("Warning: can't allocate distance table!\n"); > > - /* don't retry until explicitly reset */ > > - numa_distance = (void *)1LU; > > return -ENOMEM; > > } > > > > - numa_distance = __va(phys); > > numa_distance_cnt = cnt; > > > > /* fill with the default distances */ > >
diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c index 5e1dde26674b..ab2d4ecef786 100644 --- a/arch/x86/mm/numa.c +++ b/arch/x86/mm/numa.c @@ -319,8 +319,7 @@ void __init numa_reset_distance(void) { size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); - /* numa_distance could be 1LU marking allocation failure, test cnt */ - if (numa_distance_cnt) + if (numa_distance) memblock_free(numa_distance, size); numa_distance_cnt = 0; numa_distance = NULL; /* enable table creation */ @@ -331,7 +330,6 @@ static int __init numa_alloc_distance(void) nodemask_t nodes_parsed; size_t size; int i, j, cnt = 0; - u64 phys; /* size the new table and allocate it */ nodes_parsed = numa_nodes_parsed; @@ -342,16 +340,12 @@ static int __init numa_alloc_distance(void) cnt++; size = cnt * cnt * sizeof(numa_distance[0]); - phys = memblock_phys_alloc_range(size, PAGE_SIZE, 0, - PFN_PHYS(max_pfn_mapped)); - if (!phys) { + numa_distance = memblock_alloc(size, PAGE_SIZE); + if (!numa_distance) { pr_warn("Warning: can't allocate distance table!\n"); - /* don't retry until explicitly reset */ - numa_distance = (void *)1LU; return -ENOMEM; } - numa_distance = __va(phys); numa_distance_cnt = cnt; /* fill with the default distances */