diff options
author | Tejun Heo <tj@kernel.org> | 2009-08-14 02:00:53 -0400 |
---|---|---|
committer | Tejun Heo <tj@kernel.org> | 2009-08-14 02:00:53 -0400 |
commit | bcb2107fdbecef3de55d597d23453747af81ba88 (patch) | |
tree | f5d7095b80a3483f65a58d8eecdb02b06475fa3d /arch/sparc/kernel/smp_64.c | |
parent | e933a73f48e3b2d40cfa56d81e2646f194b5a66a (diff) |
sparc64: use embedding percpu first chunk allocator
sparc64 currently allocates a large page for each cpu and partially
remap them into vmalloc area much like what lpage first chunk
allocator did. As a 4M page is used for each cpu, this results in
very large unit size and also adds TLB pressure due to the double
mapping of pages in the first chunk.
This patch converts sparc64 to use the embedding percpu first chunk
allocator which now knows how to handle NUMA configurations. This
simplifies the code a lot, doesn't incur any extra TLB pressure and
results in better utilization of address space.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'arch/sparc/kernel/smp_64.c')
-rw-r--r-- | arch/sparc/kernel/smp_64.c | 128 |
1 files changed, 18 insertions, 110 deletions
diff --git a/arch/sparc/kernel/smp_64.c b/arch/sparc/kernel/smp_64.c index b03fd362c629..ff68373ce6d6 100644 --- a/arch/sparc/kernel/smp_64.c +++ b/arch/sparc/kernel/smp_64.c | |||
@@ -1389,8 +1389,8 @@ void smp_send_stop(void) | |||
1389 | * RETURNS: | 1389 | * RETURNS: |
1390 | * Pointer to the allocated area on success, NULL on failure. | 1390 | * Pointer to the allocated area on success, NULL on failure. |
1391 | */ | 1391 | */ |
1392 | static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size, | 1392 | static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size, |
1393 | unsigned long align) | 1393 | size_t align) |
1394 | { | 1394 | { |
1395 | const unsigned long goal = __pa(MAX_DMA_ADDRESS); | 1395 | const unsigned long goal = __pa(MAX_DMA_ADDRESS); |
1396 | #ifdef CONFIG_NEED_MULTIPLE_NODES | 1396 | #ifdef CONFIG_NEED_MULTIPLE_NODES |
@@ -1415,123 +1415,31 @@ static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size, | |||
1415 | #endif | 1415 | #endif |
1416 | } | 1416 | } |
1417 | 1417 | ||
1418 | #define PCPU_CHUNK_SIZE (4UL * 1024UL * 1024UL) | 1418 | static void __init pcpu_free_bootmem(void *ptr, size_t size) |
1419 | |||
1420 | static void __init pcpu_map_range(unsigned long start, unsigned long end, | ||
1421 | struct page *page) | ||
1422 | { | 1419 | { |
1423 | unsigned long pfn = page_to_pfn(page); | 1420 | free_bootmem(__pa(ptr), size); |
1424 | unsigned long pte_base; | 1421 | } |
1425 | |||
1426 | BUG_ON((pfn<<PAGE_SHIFT)&(PCPU_CHUNK_SIZE - 1UL)); | ||
1427 | |||
1428 | pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U | | ||
1429 | _PAGE_CP_4U | _PAGE_CV_4U | | ||
1430 | _PAGE_P_4U | _PAGE_W_4U); | ||
1431 | if (tlb_type == hypervisor) | ||
1432 | pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V | | ||
1433 | _PAGE_CP_4V | _PAGE_CV_4V | | ||
1434 | _PAGE_P_4V | _PAGE_W_4V); | ||
1435 | |||
1436 | while (start < end) { | ||
1437 | pgd_t *pgd = pgd_offset_k(start); | ||
1438 | unsigned long this_end; | ||
1439 | pud_t *pud; | ||
1440 | pmd_t *pmd; | ||
1441 | pte_t *pte; | ||
1442 | |||
1443 | pud = pud_offset(pgd, start); | ||
1444 | if (pud_none(*pud)) { | ||
1445 | pmd_t *new; | ||
1446 | |||
1447 | new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE); | ||
1448 | pud_populate(&init_mm, pud, new); | ||
1449 | } | ||
1450 | |||
1451 | pmd = pmd_offset(pud, start); | ||
1452 | if (!pmd_present(*pmd)) { | ||
1453 | pte_t *new; | ||
1454 | |||
1455 | new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE); | ||
1456 | pmd_populate_kernel(&init_mm, pmd, new); | ||
1457 | } | ||
1458 | |||
1459 | pte = pte_offset_kernel(pmd, start); | ||
1460 | this_end = (start + PMD_SIZE) & PMD_MASK; | ||
1461 | if (this_end > end) | ||
1462 | this_end = end; | ||
1463 | |||
1464 | while (start < this_end) { | ||
1465 | unsigned long paddr = pfn << PAGE_SHIFT; | ||
1466 | |||
1467 | pte_val(*pte) = (paddr | pte_base); | ||
1468 | 1422 | ||
1469 | start += PAGE_SIZE; | 1423 | static int pcpu_cpu_distance(unsigned int from, unsigned int to) |
1470 | pte++; | 1424 | { |
1471 | pfn++; | 1425 | if (cpu_to_node(from) == cpu_to_node(to)) |
1472 | } | 1426 | return LOCAL_DISTANCE; |
1473 | } | 1427 | else |
1428 | return REMOTE_DISTANCE; | ||
1474 | } | 1429 | } |
1475 | 1430 | ||
1476 | void __init setup_per_cpu_areas(void) | 1431 | void __init setup_per_cpu_areas(void) |
1477 | { | 1432 | { |
1478 | static struct vm_struct vm; | 1433 | unsigned long delta; |
1479 | struct pcpu_alloc_info *ai; | 1434 | unsigned int cpu; |
1480 | unsigned long delta, cpu; | ||
1481 | size_t size_sum; | ||
1482 | size_t ptrs_size; | ||
1483 | void **ptrs; | ||
1484 | int rc; | 1435 | int rc; |
1485 | 1436 | ||
1486 | ai = pcpu_alloc_alloc_info(1, nr_cpu_ids); | 1437 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
1487 | 1438 | PERCPU_DYNAMIC_RESERVE, 4 << 20, | |
1488 | ai->static_size = __per_cpu_end - __per_cpu_start; | 1439 | pcpu_cpu_distance, pcpu_alloc_bootmem, |
1489 | ai->reserved_size = PERCPU_MODULE_RESERVE; | 1440 | pcpu_free_bootmem); |
1490 | |||
1491 | size_sum = PFN_ALIGN(ai->static_size + ai->reserved_size + | ||
1492 | PERCPU_DYNAMIC_RESERVE); | ||
1493 | |||
1494 | ai->dyn_size = size_sum - ai->static_size - ai->reserved_size; | ||
1495 | ai->unit_size = PCPU_CHUNK_SIZE; | ||
1496 | ai->atom_size = PCPU_CHUNK_SIZE; | ||
1497 | ai->alloc_size = PCPU_CHUNK_SIZE; | ||
1498 | ai->groups[0].nr_units = nr_cpu_ids; | ||
1499 | |||
1500 | for_each_possible_cpu(cpu) | ||
1501 | ai->groups[0].cpu_map[cpu] = cpu; | ||
1502 | |||
1503 | ptrs_size = PFN_ALIGN(nr_cpu_ids * sizeof(ptrs[0])); | ||
1504 | ptrs = alloc_bootmem(ptrs_size); | ||
1505 | |||
1506 | for_each_possible_cpu(cpu) { | ||
1507 | ptrs[cpu] = pcpu_alloc_bootmem(cpu, PCPU_CHUNK_SIZE, | ||
1508 | PCPU_CHUNK_SIZE); | ||
1509 | |||
1510 | free_bootmem(__pa(ptrs[cpu] + size_sum), | ||
1511 | PCPU_CHUNK_SIZE - size_sum); | ||
1512 | |||
1513 | memcpy(ptrs[cpu], __per_cpu_load, ai->static_size); | ||
1514 | } | ||
1515 | |||
1516 | /* allocate address and map */ | ||
1517 | vm.flags = VM_ALLOC; | ||
1518 | vm.size = nr_cpu_ids * PCPU_CHUNK_SIZE; | ||
1519 | vm_area_register_early(&vm, PCPU_CHUNK_SIZE); | ||
1520 | |||
1521 | for_each_possible_cpu(cpu) { | ||
1522 | unsigned long start = (unsigned long) vm.addr; | ||
1523 | unsigned long end; | ||
1524 | |||
1525 | start += cpu * PCPU_CHUNK_SIZE; | ||
1526 | end = start + PCPU_CHUNK_SIZE; | ||
1527 | pcpu_map_range(start, end, virt_to_page(ptrs[cpu])); | ||
1528 | } | ||
1529 | |||
1530 | rc = pcpu_setup_first_chunk(ai, vm.addr); | ||
1531 | if (rc) | 1441 | if (rc) |
1532 | panic("failed to setup percpu first chunk (%d)", rc); | 1442 | panic("failed to initialize first chunk (%d)", rc); |
1533 | |||
1534 | free_bootmem(__pa(ptrs), ptrs_size); | ||
1535 | 1443 | ||
1536 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | 1444 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
1537 | for_each_possible_cpu(cpu) | 1445 | for_each_possible_cpu(cpu) |