From 2046b94e7c4fce92eb8165c2c36c6478f4927178 Mon Sep 17 00:00:00 2001 From: Fenghua Yu Date: Fri, 4 Apr 2008 11:05:59 -0700 Subject: [IA64] Multiple outstanding ptc.g instruction support According to SDM2.2, Itanium supports multiple outstanding ptc.g instructions. But current kernel function ia64_global_tlb_purge() uses a spinlock to serialize ptc.g instructions issued by multiple processors. This serialization might have scalability issue on a big SMP machine where many processors could purge TLB in parallel. The patch fixes this problem by issuing multiple ptc.g instructions in ia64_global_tlb_purge(). It also adds support for the "PALO" table to get a platform view of the max number of outstanding ptc.g instructions (which may be different from the processor view found from PAL_VM_SUMMARY). PALO specification can be found at: http://www.dig64.org/home/DIG64_PALO_R1_0.pdf spinaphore implementation by Matthew Wilcox. Signed-off-by: Fenghua Yu Signed-off-by: Tony Luck --- arch/ia64/kernel/efi.c | 46 +++++++++++++++++ arch/ia64/kernel/setup.c | 6 ++- arch/ia64/mm/tlb.c | 125 +++++++++++++++++++++++++++++++++++++++++------ 3 files changed, 160 insertions(+), 17 deletions(-) (limited to 'arch') diff --git a/arch/ia64/kernel/efi.c b/arch/ia64/kernel/efi.c index 728d7247a1a..003cd09b073 100644 --- a/arch/ia64/kernel/efi.c +++ b/arch/ia64/kernel/efi.c @@ -37,6 +37,7 @@ #include #include #include +#include #define EFI_DEBUG 0 @@ -403,6 +404,41 @@ efi_get_pal_addr (void) return NULL; } + +static u8 __init palo_checksum(u8 *buffer, u32 length) +{ + u8 sum = 0; + u8 *end = buffer + length; + + while (buffer < end) + sum = (u8) (sum + *(buffer++)); + + return sum; +} + +/* + * Parse and handle PALO table which is published at: + * http://www.dig64.org/home/DIG64_PALO_R1_0.pdf + */ +static void __init handle_palo(unsigned long palo_phys) +{ + struct palo_table *palo = __va(palo_phys); + u8 checksum; + + if (strncmp(palo->signature, PALO_SIG, sizeof(PALO_SIG) - 1)) { + printk(KERN_INFO "PALO signature incorrect.\n"); + return; + } + + checksum = palo_checksum((u8 *)palo, palo->length); + if (checksum) { + printk(KERN_INFO "PALO checksum incorrect.\n"); + return; + } + + setup_ptcg_sem(palo->max_tlb_purges, 1); +} + void efi_map_pal_code (void) { @@ -432,6 +468,7 @@ efi_init (void) u64 efi_desc_size; char *cp, vendor[100] = "unknown"; int i; + unsigned long palo_phys; /* * It's too early to be able to use the standard kernel command line @@ -496,6 +533,8 @@ efi_init (void) efi.hcdp = EFI_INVALID_TABLE_ADDR; efi.uga = EFI_INVALID_TABLE_ADDR; + palo_phys = EFI_INVALID_TABLE_ADDR; + for (i = 0; i < (int) efi.systab->nr_tables; i++) { if (efi_guidcmp(config_tables[i].guid, MPS_TABLE_GUID) == 0) { efi.mps = config_tables[i].table; @@ -515,10 +554,17 @@ efi_init (void) } else if (efi_guidcmp(config_tables[i].guid, HCDP_TABLE_GUID) == 0) { efi.hcdp = config_tables[i].table; printk(" HCDP=0x%lx", config_tables[i].table); + } else if (efi_guidcmp(config_tables[i].guid, + PROCESSOR_ABSTRACTION_LAYER_OVERWRITE_GUID) == 0) { + palo_phys = config_tables[i].table; + printk(" PALO=0x%lx", config_tables[i].table); } } printk("\n"); + if (palo_phys != EFI_INVALID_TABLE_ADDR) + handle_palo(palo_phys); + runtime = __va(efi.systab->runtime); efi.get_time = phys_get_time; efi.set_time = phys_set_time; diff --git a/arch/ia64/kernel/setup.c b/arch/ia64/kernel/setup.c index 4aa9eaea76c..1cbd26340d8 100644 --- a/arch/ia64/kernel/setup.c +++ b/arch/ia64/kernel/setup.c @@ -59,6 +59,7 @@ #include #include #include +#include #include #include @@ -946,9 +947,10 @@ cpu_init (void) #endif /* set ia64_ctx.max_rid to the maximum RID that is supported by all CPUs: */ - if (ia64_pal_vm_summary(NULL, &vmi) == 0) + if (ia64_pal_vm_summary(NULL, &vmi) == 0) { max_ctx = (1U << (vmi.pal_vm_info_2_s.rid_size - 3)) - 1; - else { + setup_ptcg_sem(vmi.pal_vm_info_2_s.max_purges, 0); + } else { printk(KERN_WARNING "cpu_init: PAL VM summary failed, assuming 18 RID bits\n"); max_ctx = (1U << 15) - 1; /* use architected minimum */ } diff --git a/arch/ia64/mm/tlb.c b/arch/ia64/mm/tlb.c index 655da240d13..d41d6076ed0 100644 --- a/arch/ia64/mm/tlb.c +++ b/arch/ia64/mm/tlb.c @@ -11,6 +11,9 @@ * Rohit Seth * Ken Chen * Christophe de Dinechin : Avoid ptc.e on memory allocation + * Copyright (C) 2007 Intel Corp + * Fenghua Yu + * Add multiple ptc.g/ptc.ga instruction support in global tlb purge. */ #include #include @@ -26,6 +29,7 @@ #include #include #include +#include static struct { unsigned long mask; /* mask of supported purge page-sizes */ @@ -84,14 +88,104 @@ wrap_mmu_context (struct mm_struct *mm) local_flush_tlb_all(); } +/* + * Implement "spinaphores" ... like counting semaphores, but they + * spin instead of sleeping. If there are ever any other users for + * this primitive it can be moved up to a spinaphore.h header. + */ +struct spinaphore { + atomic_t cur; +}; + +static inline void spinaphore_init(struct spinaphore *ss, int val) +{ + atomic_set(&ss->cur, val); +} + +static inline void down_spin(struct spinaphore *ss) +{ + while (unlikely(!atomic_add_unless(&ss->cur, -1, 0))) + while (atomic_read(&ss->cur) == 0) + cpu_relax(); +} + +static inline void up_spin(struct spinaphore *ss) +{ + atomic_add(1, &ss->cur); +} + +static struct spinaphore ptcg_sem; +static u16 nptcg = 1; +static int need_ptcg_sem = 1; +static int toolatetochangeptcgsem = 0; + +/* + * Maximum number of simultaneous ptc.g purges in the system can + * be defined by PAL_VM_SUMMARY (in which case we should take + * the smallest value for any cpu in the system) or by the PAL + * override table (in which case we should ignore the value from + * PAL_VM_SUMMARY). + * + * Complicating the logic here is the fact that num_possible_cpus() + * isn't fully setup until we start bringing cpus online. + */ +void +setup_ptcg_sem(int max_purges, int from_palo) +{ + static int have_palo; + static int firstcpu = 1; + + if (toolatetochangeptcgsem) { + BUG_ON(max_purges < nptcg); + return; + } + + if (from_palo) { + have_palo = 1; + + /* In PALO max_purges == 0 really means it! */ + if (max_purges == 0) + panic("Whoa! Platform does not support global TLB purges.\n"); + nptcg = max_purges; + if (nptcg == PALO_MAX_TLB_PURGES) { + need_ptcg_sem = 0; + return; + } + goto resetsema; + } + if (have_palo) { + if (nptcg != PALO_MAX_TLB_PURGES) + need_ptcg_sem = (num_possible_cpus() > nptcg); + return; + } + + /* In PAL_VM_SUMMARY max_purges == 0 actually means 1 */ + if (max_purges == 0) max_purges = 1; + + if (firstcpu) { + nptcg = max_purges; + firstcpu = 0; + } + if (max_purges < nptcg) + nptcg = max_purges; + if (nptcg == PAL_MAX_PURGES) { + need_ptcg_sem = 0; + return; + } else + need_ptcg_sem = (num_possible_cpus() > nptcg); + +resetsema: + spinaphore_init(&ptcg_sem, max_purges); +} + void ia64_global_tlb_purge (struct mm_struct *mm, unsigned long start, unsigned long end, unsigned long nbits) { - static DEFINE_SPINLOCK(ptcg_lock); - struct mm_struct *active_mm = current->active_mm; + toolatetochangeptcgsem = 1; + if (mm != active_mm) { /* Restore region IDs for mm */ if (mm && active_mm) { @@ -102,19 +196,20 @@ ia64_global_tlb_purge (struct mm_struct *mm, unsigned long start, } } - /* HW requires global serialization of ptc.ga. */ - spin_lock(&ptcg_lock); - { - do { - /* - * Flush ALAT entries also. - */ - ia64_ptcga(start, (nbits<<2)); - ia64_srlz_i(); - start += (1UL << nbits); - } while (start < end); - } - spin_unlock(&ptcg_lock); + if (need_ptcg_sem) + down_spin(&ptcg_sem); + + do { + /* + * Flush ALAT entries also. + */ + ia64_ptcga(start, (nbits << 2)); + ia64_srlz_i(); + start += (1UL << nbits); + } while (start < end); + + if (need_ptcg_sem) + up_spin(&ptcg_sem); if (mm != active_mm) { activate_context(active_mm); -- cgit v1.2.3 From a6c75b86ce9f01db4ea9912877b526c2dc4d2f0a Mon Sep 17 00:00:00 2001 From: Fenghua Yu Date: Fri, 14 Mar 2008 13:57:08 -0700 Subject: [IA64] Kernel parameter for max number of concurrent global TLB purges The patch defines kernel parameter "nptcg=". The parameter overrides max number of concurrent global TLB purges which is reported from either PAL_VM_SUMMARY or SAL PALO. Signed-off-by: Fenghua Yu Signed-off-by: Tony Luck --- arch/ia64/kernel/efi.c | 2 +- arch/ia64/kernel/setup.c | 2 +- arch/ia64/mm/tlb.c | 46 +++++++++++++++++++++++++++++++++++++++++----- 3 files changed, 43 insertions(+), 7 deletions(-) (limited to 'arch') diff --git a/arch/ia64/kernel/efi.c b/arch/ia64/kernel/efi.c index 003cd09b073..d45f215bc8f 100644 --- a/arch/ia64/kernel/efi.c +++ b/arch/ia64/kernel/efi.c @@ -436,7 +436,7 @@ static void __init handle_palo(unsigned long palo_phys) return; } - setup_ptcg_sem(palo->max_tlb_purges, 1); + setup_ptcg_sem(palo->max_tlb_purges, NPTCG_FROM_PALO); } void diff --git a/arch/ia64/kernel/setup.c b/arch/ia64/kernel/setup.c index 1cbd26340d8..f798c0769d3 100644 --- a/arch/ia64/kernel/setup.c +++ b/arch/ia64/kernel/setup.c @@ -949,7 +949,7 @@ cpu_init (void) /* set ia64_ctx.max_rid to the maximum RID that is supported by all CPUs: */ if (ia64_pal_vm_summary(NULL, &vmi) == 0) { max_ctx = (1U << (vmi.pal_vm_info_2_s.rid_size - 3)) - 1; - setup_ptcg_sem(vmi.pal_vm_info_2_s.max_purges, 0); + setup_ptcg_sem(vmi.pal_vm_info_2_s.max_purges, NPTCG_FROM_PAL); } else { printk(KERN_WARNING "cpu_init: PAL VM summary failed, assuming 18 RID bits\n"); max_ctx = (1U << 15) - 1; /* use architected minimum */ diff --git a/arch/ia64/mm/tlb.c b/arch/ia64/mm/tlb.c index d41d6076ed0..1a8948fd002 100644 --- a/arch/ia64/mm/tlb.c +++ b/arch/ia64/mm/tlb.c @@ -119,6 +119,27 @@ static u16 nptcg = 1; static int need_ptcg_sem = 1; static int toolatetochangeptcgsem = 0; +/* + * Kernel parameter "nptcg=" overrides max number of concurrent global TLB + * purges which is reported from either PAL or SAL PALO. + * + * We don't have sanity checking for nptcg value. It's the user's responsibility + * for valid nptcg value on the platform. Otherwise, kernel may hang in some + * cases. + */ +static int __init +set_nptcg(char *str) +{ + int value = 0; + + get_option(&str, &value); + setup_ptcg_sem(value, NPTCG_FROM_KERNEL_PARAMETER); + + return 1; +} + +__setup("nptcg=", set_nptcg); + /* * Maximum number of simultaneous ptc.g purges in the system can * be defined by PAL_VM_SUMMARY (in which case we should take @@ -126,13 +147,18 @@ static int toolatetochangeptcgsem = 0; * override table (in which case we should ignore the value from * PAL_VM_SUMMARY). * + * Kernel parameter "nptcg=" overrides maximum number of simultanesous ptc.g + * purges defined in either PAL_VM_SUMMARY or PAL override table. In this case, + * we should ignore the value from either PAL_VM_SUMMARY or PAL override table. + * * Complicating the logic here is the fact that num_possible_cpus() * isn't fully setup until we start bringing cpus online. */ void -setup_ptcg_sem(int max_purges, int from_palo) +setup_ptcg_sem(int max_purges, int nptcg_from) { - static int have_palo; + static int kp_override; + static int palo_override; static int firstcpu = 1; if (toolatetochangeptcgsem) { @@ -140,8 +166,18 @@ setup_ptcg_sem(int max_purges, int from_palo) return; } - if (from_palo) { - have_palo = 1; + if (nptcg_from == NPTCG_FROM_KERNEL_PARAMETER) { + kp_override = 1; + nptcg = max_purges; + goto resetsema; + } + if (kp_override) { + need_ptcg_sem = num_possible_cpus() > nptcg; + return; + } + + if (nptcg_from == NPTCG_FROM_PALO) { + palo_override = 1; /* In PALO max_purges == 0 really means it! */ if (max_purges == 0) @@ -153,7 +189,7 @@ setup_ptcg_sem(int max_purges, int from_palo) } goto resetsema; } - if (have_palo) { + if (palo_override) { if (nptcg != PALO_MAX_TLB_PURGES) need_ptcg_sem = (num_possible_cpus() > nptcg); return; -- cgit v1.2.3