New DISCONTIG patch
Subsystems (7)
Attached is a new discontig patch. I would appreciate feedback on these changes & whether you see problems. Assuming there are no major issues, I'll move the patch to 2.5.
The patch is based on 2.4.20 and builds on the discontig-2.4.20-030304.patch from kimi. Apply the patch to a vanilla 2.4.20+mosberger tree.
The code boots/runs on the following hardware configurations:
- SGI SN2 platform
- DIG SP kernel
- DIG MP kernel using phys memmap
- DIG MP kernel with a hack to force it to use virt memmap
- DIG MP kernel with CONFIG_NUMA & CONFIG_DISCONTIGMEM enabled:
- boots using phy memmap
- boots with hack to force it to use virt memmap
The patch contains a number of improvements to the previous patches. Primarily, it shrinks the size & complexity of the discontig patch by using the VIRTUAL_MEMMAP code that is in 2.4.20.
Note that discontig_paging_init() is gone & that a lot of DISCONTIG code has been removed from other places except the for code that initializes the bootmem allocator.
The files that are the most interesting are:
arch/ia64/mm/discontig.c - paging_init is gone arch/ia64/mm/init.c - DISCONTIG uses VIRT_MEMMAP
- no #ifdef CONFIG_DISCONTIG
include/asm-ia64/pgtable.h - no #ifdef CONFIG_DISCONTIG include/asm-ia64/mmzone.h - add SGI & GENERIC platforms & remove
most of the ugly macros
I have run a few AIM7 benchmarks to compare the performance of the old patch that uses region 7 for page-structs vs this new patch that uses the virtual memmap in region 5. So far, I have not seen any significant differences. AIM7 shows a max of 0.5% degradation in thruput. Cpu time increases 0-2% depending on the number of users.
Next, I am working on making the patch work on GENERIC kernels but have not yet completed the work. I think this patch is close....
Notes:
- discontig_paging_init is completely gone. I use the paging_init() code that is in arch/ia64/mm/init.c There are a few changes in paging_init() but I believe it will work on all platforms (needs to be verified though).
- there were a few features in the original discontig patch that were missing in the latest discontig patch. A few of these have been added again. They are important on the SGI platform & I suspect they are important on other platforms too. I think that some of these features were implemented differently in 2.5 & were just dropped to simplify the backport of discontig to 2.4.20.
- node local cpu_data structures & mmu_gathers. Possible performance improvement but done primarily to reduce the amount of static data. This could be split into a separate patch.
---------------------------------------
discontig_2.4.20-20030502
arch/ia64/config.in +4 -0
@@ -66,6 +66,10 @@ fi if [ "$CONFIG_IA64_GENERIC" = "y" -o "$CONFIG_IA64_DIG" = "y" -o "$CONFIG_IA64_HP_ZX1" = "y" ]; then + bool ' Enable NUMA support' CONFIG_NUMA + if [ "$CONFIG_NUMA" = "y" ]; then + define_bool CONFIG_DISCONTIGMEM y + fi bool ' Enable IA-64 Machine Check Abort' CONFIG_IA64_MCA define_bool CONFIG_PM y fi
arch/ia64/kernel/acpi.c +211 -7
259 lines not on this page · raw message
arch/ia64/kernel/setup.c +98 -31
221 lines not on this page · raw message
arch/ia64/kernel/smpboot.c +64 -0
73 lines not on this page · raw message
arch/ia64/mm/Makefile +2 -0
@@ -12,5 +12,7 @@ export-objs := init.o obj-y := init.o fault.o tlb.o extable.o +obj-$(CONFIG_NUMA) += numa.o +obj-$(CONFIG_DISCONTIGMEM) += discontig.o include $(TOPDIR)/Rules.make
arch/ia64/mm/discontig.c +280 -0
280 lines not on this page · raw message
arch/ia64/mm/init.c +127 -114
382 lines not on this page · raw message
arch/ia64/mm/numa.c +46 -0
46 lines not on this page · raw message
drivers/acpi/Config.in +6 -0
@@ -36,6 +36,9 @@ tristate ' Fan' CONFIG_ACPI_FAN tristate ' Processor' CONFIG_ACPI_PROCESSOR dep_tristate ' Thermal Zone' CONFIG_ACPI_THERMAL $CONFIG_ACPI_PROCESSOR + if [ "$CONFIG_NUMA" = "y" ]; then + bool ' NUMA support' CONFIG_ACPI_NUMA + fi bool ' Debug Statements' CONFIG_ACPI_DEBUG fi
@@ -123,4 +126,7 @@ endmenu fi + if [ "$CONFIG_NUMA" = "y" ]; then + bool ' NUMA support' CONFIG_ACPI_NUMA + fi fi
drivers/acpi/Makefile +1 -0
@@ -51,5 +51,6 @@ obj-$(CONFIG_ACPI_THERMAL) += thermal.o obj-$(CONFIG_ACPI_SYSTEM) += system.o endif +obj-$(CONFIG_ACPI_NUMA) += numa.o include $(TOPDIR)/Rules.make
drivers/acpi/numa.c +188 -0
188 lines not on this page · raw message
drivers/acpi/tables.c +23 -9
@@ -224,11 +224,13 @@int __init -acpi_table_parse_madt ( +acpi_table_parse_madt_family (enum acpi_table_id id, +unsigned long madt_size, +int entry_id,acpi_madt_entry_handler handler) { -struct acpi_table_madt *madt =NULL ; +void *madt =NULL ;acpi_table_entry_header *entry =NULL ;unsigned long count =0 ;unsigned long madt_end =0 ;
@@ -240,19 +242,21 @@/* Locate the MADT (if exists). There should only be one. */ for (i =0 ; i < sdt.count; i++) { -if (sdt.entry[i].id !=ACPI_APIC ) +if (sdt.entry[i].id != id)continue ; - madt = (struct acpi_table_madt *) + madt = (void *)__acpi_map_table (sdt.entry[i].pa, sdt.entry[i].size);if (!madt) { - printk(KERN_WARNING PREFIX "Unable to map MADT\n"); + printk(KERN_WARNING PREFIX "Unable to map %s\n", + acpi_table_signatures[id]);return -ENODEV ; }break ; }if (!madt) { - printk(KERN_WARNING PREFIX "MADT not present\n"); + printk(KERN_WARNING PREFIX "%s not present\n", + acpi_table_signatures[id]);return -ENODEV ; }
@@ -261,21 +265,31 @@/* Parse all entries looking for a match. */ entry = (acpi_table_entry_header *) - ((unsigned long ) madt +sizeof (struct acpi_table_madt )); + ((unsigned long ) madt + madt_size);while (((unsigned long ) entry) < madt_end) { -if (entry->type = id) { +if (entry->type = entry_id) { count++;handler (entry); } entry = (acpi_table_entry_header *) - ((unsigned long ) entry += entry->length); + ((unsigned long ) entry + entry->length); }return count; } +int __init +acpi_table_parse_madt ( +enum acpi_madt_entry_id id, +acpi_madt_entry_handler handler) +{ +return acpi_table_parse_madt_family (ACPI_APIC ,sizeof (struct acpi_table_madt ), + id, handler); +} + + int __init acpi_table_parse ( enum acpi_table_id id,
include/asm-generic/tlb.h +6 -1
@@ -31,15 +31,20 @@pte_t ptes[FREE_PTE_NR ]; } mmu_gather_t; +#if defined(CONFIG_NUMA ) && defined(CONFIG_IA64 )/* Users of the generic TLB shootdown code must declare this storage space. */ +#define local_mmu_gathers local_cpu_data->mmu_gathers +#else extern mmu_gather_t mmu_gathers[NR_CPUS ]; +#define local_mmu_gathers &mmu_gathers[smp_processor_id()] +#endif /* tlb_gather_mmu * Return a pointer to an initialized mmu_gather_t. */ static inline mmu_gather_t *tlb_gather_mmu (struct mm_struct *mm) { -mmu_gather_t *tlb = &mmu_gathers[smp_processor_id ()]; +mmu_gather_t *tlb = local_mmu_gathers; tlb->mm = mm;/* Use fast mode if there is only one user of this mm (this process) */
include/asm-ia64/acpi.h +7 -6
@@ -97,17 +97,18 @@ } while (0) const char *acpi_get_sysname (void); -int acpi_boot_init (char *cdline);int acpi_request_vector (u32 int_type);int acpi_get_prt (struct pci_vector_struct **vectors,int *count);int acpi_get_interrupt_model (int *type);int acpi_irq_to_vector (u32 irq); -#ifdef CONFIG_DISCONTIGMEM -#define NODE_ARRAY_INDEX (x) ((x) / 8)/* 8 bits/char */ -#define NODE_ARRAY_OFFSET (x) ((x) % 8)/* 8 bits/char */ -#define MAX_PXM_DOMAINS (256) -#endif /* CONFIG_DISCONTIGMEM */ +#ifdef CONFIG_ACPI_NUMA +#include <asm/numa.h> +/* Proximity bitmap length; _PXM is at most 255 (8 bit)*/ +#define MAX_PXM_DOMAINS (256) +extern int __initdata pxm_to_nid_map[MAX_PXM_DOMAINS ]; +extern int __initdata nid_to_pxm_map[NR_NODES ]; +#endif #endif /*__KERNEL__*/
include/asm-ia64/mmzone.h +65 -0
65 lines not on this page · raw message
include/asm-ia64/nodedata.h +67 -0
67 lines not on this page · raw message
include/asm-ia64/numa.h +84 -0
84 lines not on this page · raw message
include/asm-ia64/numnodes.h +7 -0
@@ -0,0 +1,7 @@ +#ifndef _ASM_MAX_NUMNODES_H +#define _ASM_MAX_NUMNODES_H + +#include <asm/mmzone.h> +#define MAX_NUMNODES NR_NODES + +#endif /* _ASM_MAX_NUMNODES_H */
include/asm-ia64/page.h +0 -11
@@ -52,19 +52,8 @@*/ #define MAP_NR_DENSE (addr) (((unsigned long) (addr) - PAGE_OFFSET) >> PAGE_SHIFT) -#ifdef CONFIG_IA64_GENERIC -# include <asm/machvec.h> -# define virt_to_page (kaddr) (mem_map + platform_map_nr(kaddr)) -# define page_to_phys (page) ((page - mem_map) << PAGE_SHIFT) -#elif defined (CONFIG_IA64_SGI_SN1 ) -# ifndef CONFIG_DISCONTIGMEM -# define virt_to_page (kaddr) (mem_map + MAP_NR_DENSE(kaddr)) -# define page_to_phys (page) XXX fix me -# endif -#else # define virt_to_page (kaddr) (mem_map + MAP_NR_DENSE(kaddr))# define page_to_phys (page) ((page - mem_map) << PAGE_SHIFT) -#endif struct page ;extern int ia64_page_valid (struct page *);
include/asm-ia64/pgtable.h +0 -1
@@ -163,7 +163,6 @@return (addr & (local_cpu_data->unimpl_pa_mask)) =0 ; } -#ifndef CONFIG_DISCONTIGMEM /* * kern_addr_valid(ADDR) tests if ADDR is pointing to valid kernel * memory. For the return value to be meaningful, ADDR must be >@@ -179,7 +178,6 @@
*/
#define kern_addr_valid(addr) (1)
-#endif
/*
* Now come the defines and routines to manage and access the three-level
@@ -227,10 +225,8 @@
#define pte_none(pte) (!pte_val(pte))
#define pte_present(pte) (pte_val(pte) & (_PAGE_P | _PAGE_PROTNONE))
#define pte_clear(pte) (pte_val(*(pte)) = 0UL)
-#ifndef CONFIG_DISCONTIGMEM
/* pte_page() returns the "struct page *" corresponding to the PTE: */
#define pte_page(pte) (mem_map + (unsigned long) ((pte_val(pte) & _PFN_MASK) >> PAGE_SHIFT))
-#endif
#define pmd_none(pmd) (!pmd_val(pmd))
#define pmd_bad(pmd) (!ia64_phys_addr_valid(pmd_val(pmd)))
@@ -430,7 +426,8 @@
* for zero-mapped memory areas etc..
*/
extern unsigned long empty_zero_page[PAGE_SIZE/sizeof(unsigned long)];
-#define ZERO_PAGE(vaddr) (virt_to_page(empty_zero_page)) +extern struct page *zero_page_memmap_ptr;
+#define ZERO_PAGE(vaddr) (zero_page_memmap_ptr)
/* We provide our own get_unmapped_area to cope with VA holes for userland */ #define HAVE_ARCH_UNMAPPED_AREA
include/asm-ia64/processor.h +10 -4
@@ -87,6 +87,9 @@#include <asm/rse.h> #include <asm/unwind.h> #include <asm/atomic.h> +#ifdef CONFIG_NUMA +#include <asm/nodedata.h> +#endif /* like above but expressed as bitfields for more efficient access: */ struct ia64_psr {
@@ -187,9 +190,12 @@ } ipi;#endif #ifdef CONFIG_NUMA -void *node_directory; -int numa_node_id; +void * mmu_gathers; +struct ia64_node_data *node_data; +int nodeid;struct cpuinfo_ia64 *cpu_data[NR_CPUS ]; +/* mod this number by numnodes to get the node to use for this pagecache allocation */ +unsigned int page_cache_nodeval;#endif /* Platform specific word. MUST BE LAST IN STRUCT */ __u64 platform_specific;
@@ -213,9 +219,9 @@*/ #ifdef CONFIG_NUMA # define cpu_data (cpu) local_cpu_data->cpu_data[cpu] -# define numa_node_id () (local_cpu_data->numa_node_id) +# define numa_node_id () (local_cpu_data->nodeid)#else -extern struct cpuinfo_ia64 _cpu_data[NR_CPUS ]; +extern struct cpuinfo_ia64 _cpu_data[NR_CPUS ];# define cpu_data (cpu) (&_cpu_data[cpu])#endif
include/asm-ia64/smp.h +1 -0
@@ -122,6 +122,7 @@extern int smp_call_function_single (int cpuid,void (*func) (void *info),void *info,int retry,int wait); +extern void smp_build_cpu_map (void );#endif /* CONFIG_SMP */ #endif /* _ASM_IA64_SMP_H */
include/asm-ia64/topology.h +63 -0
63 lines not on this page · raw message
include/linux/acpi.h +8 -0
@@ -344,6 +344,14 @@void acpi_table_print (struct acpi_table_header *,unsigned long );void acpi_table_print_madt_entry (acpi_table_entry_header *); +#ifdef CONFIG_ACPI_NUMA +int __init acpi_numa_init (void ); +void __init acpi_numa_slit_init (struct acpi_table_slit *); +void __init acpi_numa_processor_affinity_init (struct acpi_table_processor_affinity *); +void __init acpi_numa_memory_affinity_init (struct acpi_table_memory_affinity *); +void __init acpi_numa_arch_fixup (void ); +#endif +#endif /*CONFIG_ACPI_BOOT*/
include/linux/mmzone.h +21 -3
@@ -8,6 +8,12 @@#include <linux/spinlock.h> #include <linux/list.h> #include <linux/wait.h> +#ifdef CONFIG_DISCONTIGMEM +#include <asm/numnodes.h> +#endif +#ifndef MAX_NUMNODES +#define MAX_NUMNODES 1 +#endif /* * Free memory management - zoned buddy allocator.
@@ -110,7 +116,7 @@* footprint of this construct is very small. */ typedef struct zonelist_struct { -zone_t * zones [MAX_NR_ZONES +1 ];// NULL delimited +zone_t * zones [MAX_NUMNODES *MAX_NR_ZONES +1 ];// NULL delimited }zonelist_t ;#define GFP_ZONEMASK 0x0f
@@ -144,8 +150,8 @@extern int numnodes;extern pg_data_t *pgdat_list; -#define memclass (pgzone, classzone) (((pgzone)->zone_pgdat = (classzone)->zone_pgdat) \ - && ((pgzone) <= (classzone))) +#define memclass (pgzone, classzone) (((pgzone) - (pgzone)->zone_pgdat->node_zones) <= \ +((classzone) - (classzone)->zone_pgdat->node_zones))/* * The following two are not meant for general usage. They are here as
@@ -212,6 +218,18 @@#define for_each_zone (zone) \ for(zone = pgdat_list->node_zones; zone; zone = next_zone(zone)) +#ifdef CONFIG_NUMA +#define MAX_NR_MEMBLKS BITS_PER_LONG/* Max number of Memory Blocks */ +#include <asm/topology.h> +#else /* !CONFIG_NUMA */ +#define MAX_NR_MEMBLKS 1 +#endif /* CONFIG_NUMA */ + +/* Returns the number of the current Node. */ + +#ifndef CONFIG_NUMA +#define numa_node_id () (__cpu_to_node(smp_processor_id())) +#endif #ifndef CONFIG_DISCONTIGMEM
init/main.c +2 -0
@@ -290,6 +290,7 @@extern void setup_arch (char **); +extern void __init build_all_zonelists (void );extern void cpu_idle (void );unsigned long wait_init_idle;
@@ -360,6 +361,7 @@lock_kernel ();printk (linux_banner);setup_arch (&command_line); +build_all_zonelists ();printk ("Kernel command line: %s\n" , saved_command_line);parse_options (command_line);trap_init ();
mm/page_alloc.c +56 -29
120 lines not on this page · raw message
Signature · 3 lines
Thanks Jack Steiner (651-683-5302) (vnet 233-5302) [email protected]