Skip to contentSkip to conversation

Improve x86 smpboot integration

PATCH In review Glauber Costa 4 participants 38 messages Jun 9, 2008 – Jun 11, 2008

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Improve x86 smpboot integration

Ingo,

First of all, sorry for being away for so long ;-)

Here it goes a series of improvements for the x86 smpboot integration. The final goal is the same: To reduce the difference between architectures, just this time I do a little bit more ;-)

Basically, this one reduces greatly the ifdef count:

[root@t60 linux-2.6-x86]# git-show 674ce:arch/x86/kernel/smpboot.c \
			  | grep -e "^#ifdef \+CONFIG_X86_\(32\|64\)" | wc -l
18
[root@t60 linux-2.6-x86]# git-show HEAD:arch/x86/kernel/smpboot.c \
			  | grep -e "^#ifdef \+CONFIG_X86_\(32\|64\)" | wc -l

8

The remaining ones are _mainly_ (not all) due to the fact that x86_64 uses the pda, while i386 goes with normal per-cpu data. They are things like:

per_cpu(current_task, cpu) = c_idle.idle; vs cpu_pda(cpu)->pcurrent = c_idle.idle;

Also, there is the low mappings piece of code Hugh detected. To that, I intend to also find a common base between them. Just I'm not doing it in this series, because it deserves special attention.

As usual, this series was compiled tested in a whole bunch of different configs ( ~ 10 for each architecture), including all i386 variants. Boot tested in all my hardware.

The final diffstat is:

 arch/x86/kernel/head_32.S             |    2 
 arch/x86/kernel/head_64.S             |   48 +---------------
 arch/x86/kernel/io_apic_32.c          |    2 
 arch/x86/kernel/setup64.c             |    1 
 arch/x86/kernel/smpboot.c             |   97 ++++------------------------------
 b/arch/x86/kernel/acpi/sleep.c        |    2 
 b/arch/x86/kernel/apic_32.c           |    6 --
 b/arch/x86/kernel/apic_64.c           |   10 +++
 b/arch/x86/kernel/head_32.S           |    6 +-
 b/arch/x86/kernel/head_64.S           |    5 +
 b/arch/x86/kernel/io_apic_32.c        |    5 +
 b/arch/x86/kernel/io_apic_64.c        |    9 ++-
 b/arch/x86/kernel/process_32.c        |   16 +++++
 b/arch/x86/kernel/setup64.c           |    5 -
 b/arch/x86/kernel/setup_64.c          |   27 +++++++++
 b/arch/x86/kernel/smpboot.c           |    8 --
 b/arch/x86/kernel/traps_32.c          |    3 -
 b/arch/x86/kernel/x8664_ksyms_64.c    |    5 -
 b/arch/x86/mach-voyager/voyager_smp.c |    2 
 b/include/asm-x86/desc.h              |   24 +++-----
 b/include/asm-x86/hw_irq.h            |    3 -
 b/include/asm-x86/numa_32.h           |    1 
 b/include/asm-x86/segment.h           |   24 ++++----
 b/include/asm-x86/smp.h               |    2 
 include/asm-x86/hw_irq.h              |    4 -
 25 files changed, 122 insertions(+), 195 deletions(-)
#1 lore raw

PATCH 1/15 x86: use stack_start in x86_64

call x86_64's init_rsp stack_start, just as i386 does. Put a zeroed stack segment for consistency. With this, we can eliminate one ugly ifdef in smpboot.c.

---
 arch/x86/kernel/acpi/sleep.c |    2 +-
 arch/x86/kernel/head_64.S    |    5 +++--
 arch/x86/kernel/smpboot.c    |    7 +------
 3 files changed, 5 insertions(+), 9 deletions(-)
arch/x86/kernel/acpi/sleep.c +1 -1
index afc25ee..0103af2 100644
@@ -72,7 +72,7 @@ int acpi_save_state_mem(void)
 	saved_magic = 0x12345678;
 #else /* CONFIG_64BIT */
 	header->trampoline_segment = setup_trampoline() >> 4;
-	init_rsp = (unsigned long)temp_stack + 4096;
+	stack_start.sp = temp_stack + 4096;
 	initial_code = (unsigned long)wakeup_long64;
 	saved_magic = 0x123456789abcdef0;
 #endif /* CONFIG_64BIT */
arch/x86/kernel/head_64.S +3 -2
index d8ed325..4949c32 100644
@@ -191,7 +191,7 @@ ENTRY(secondary_startup_64)
 	movq	%rax, %cr0
 
 	/* Setup a boot time stack */
-	movq init_rsp(%rip),%rsp
+	movq stack_start(%rip),%rsp
 
 	/* zero EFLAGS after setting rsp */
 	pushq $0
@@ -252,8 +252,9 @@ ENTRY(secondary_startup_64)
 	.quad	x86_64_start_kernel
 	__FINITDATA
 
-	ENTRY(init_rsp)
+	ENTRY(stack_start)
 	.quad  init_thread_union+THREAD_SIZE-8
+	.word  0
 
 bad_address:
 	jmp bad_address
arch/x86/kernel/smpboot.c +1 -6
index 722b24b..c77a7d2 100644
@@ -714,11 +714,7 @@ wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
 	 * target processor state.
 	 */
 	startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
-#ifdef CONFIG_X86_64
-			 (unsigned long)init_rsp);
-#else
 			 (unsigned long)stack_start.sp);
-#endif
 
 	/*
 	 * Run STARTUP IPI loop.
@@ -905,15 +901,14 @@ do_rest:
 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
 	c_idle.idle->thread.ip = (unsigned long) start_secondary;
 	/* Stack for startup_32 can be just as for start_secondary onwards */
-	stack_start.sp = (void *) c_idle.idle->thread.sp;
 	irq_ctx_init(cpu);
 #else
 	cpu_pda(cpu)->pcurrent = c_idle.idle;
-	init_rsp = c_idle.idle->thread.sp;
 	load_sp0(&per_cpu(init_tss, cpu), &c_idle.idle->thread);
 	initial_code = (unsigned long)start_secondary;
 	clear_tsk_thread_flag(c_idle.idle, TIF_FORK);
 #endif
+	stack_start.sp = (void *) c_idle.idle->thread.sp;
 
 	/* start_ip had better be page-aligned! */
 	start_ip = setup_trampoline();
Signature · 1 line
1.5.4.5
#2 lore raw

PATCH 2/15 x86: don't use gdt_page openly.

There's a macro available for that.

---
 arch/x86/kernel/traps_32.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/traps_32.c +1 -1
index a92b0c7..9ad896d 100644
@@ -1135,7 +1135,7 @@ void do_spurious_interrupt_bug(struct pt_regs *regs, long error_code)
 
 unsigned long patch_espfix_desc(unsigned long uesp, unsigned long kesp)
 {
-	struct desc_struct *gdt = __get_cpu_var(gdt_page).gdt;
+	struct desc_struct *gdt = get_cpu_gdt_table(smp_processor_id());
 	unsigned long base = (kesp - uesp) & -THREAD_SIZE;
 	unsigned long new_kesp = kesp - base;
 	unsigned long lim_pages = (new_kesp | (THREAD_SIZE - 1)) >> PAGE_SHIFT;
Signature · 1 line
1.5.4.5
#3 lore raw

PATCH 3/15 x86: remove early_gdt_descr reference

since we use switch_to_new_gdt, there is no point in assigning early_gdt_descr except for the first assignment, which is done manually.

CC James Bottomley <[email protected]>
---
 arch/x86/kernel/smpboot.c           |    1 -
 arch/x86/mach-voyager/voyager_smp.c |    1 -
 2 files changed, 0 insertions(+), 2 deletions(-)
arch/x86/kernel/smpboot.c +0 -1
index c77a7d2..181b109 100644
@@ -898,7 +898,6 @@ do_rest:
 #ifdef CONFIG_X86_32
 	per_cpu(current_task, cpu) = c_idle.idle;
 	init_gdt(cpu);
-	early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
 	c_idle.idle->thread.ip = (unsigned long) start_secondary;
 	/* Stack for startup_32 can be just as for start_secondary onwards */
 	irq_ctx_init(cpu);
arch/x86/mach-voyager/voyager_smp.c +0 -1
index 8dedd01..6b97d9b 100644
@@ -529,7 +529,6 @@ static void __init do_boot_cpu(__u8 cpu)
 
 	init_gdt(cpu);
 	per_cpu(current_task, cpu) = idle;
-	early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
 	irq_ctx_init(cpu);
 
 	/* Note: Don't modify initial ss override */
Signature · 1 line
1.5.4.5
#4 lore raw

PATCH 4/15 x86: move x86_64 gdt closer to i386

i386 and x86_64 used two different schemes for maintaining the gdt. With this patch, x86_64 initial gdt table is defined in a .c file, same way as i386 is now. Also, we call it "gdt_page", and the descriptor, "early_gdt_descr". This way we achieve common naming, which can allow for more code integration.

---
 arch/x86/kernel/head_64.S        |   48 ++++----------------------------------
 arch/x86/kernel/setup64.c        |    5 +---
 arch/x86/kernel/setup_64.c       |   27 +++++++++++++++++++++
 arch/x86/kernel/smpboot.c        |   10 +------
 arch/x86/kernel/x8664_ksyms_64.c |    5 ----
 include/asm-x86/desc.h           |   24 ++++++++-----------
 include/asm-x86/segment.h        |   23 +++++++++--------
 7 files changed, 57 insertions(+), 85 deletions(-)
arch/x86/kernel/head_64.S +5 -43
index 4949c32..febeabb 100644
@@ -203,7 +203,7 @@ ENTRY(secondary_startup_64)
 	 * addresses where we're currently running on. We have to do that here
 	 * because in 32bit we couldn't load a 64bit linear address.
 	 */
-	lgdt	cpu_gdt_descr(%rip)
+	lgdt	early_gdt_descr(%rip)
 
 	/* set up data segments. actually 0 would do too */
 	movl $__KERNEL_DS,%eax
@@ -391,54 +391,16 @@ NEXT_PAGE(level2_spare_pgt)
 
 	.data
 	.align 16
-	.globl cpu_gdt_descr
-cpu_gdt_descr:
-	.word	gdt_end-cpu_gdt_table-1
-gdt:
-	.quad	cpu_gdt_table
-#ifdef CONFIG_SMP
-	.rept	NR_CPUS-1
-	.word	0
-	.quad	0
-	.endr
-#endif
+	.globl early_gdt_descr
+early_gdt_descr:
+	.word	GDT_ENTRIES*8-1
+	.quad   per_cpu__gdt_page
 
 ENTRY(phys_base)
 	/* This must match the first entry in level2_kernel_pgt */
 	.quad   0x0000000000000000
 
-/* We need valid kernel segments for data and code in long mode too
- * IRET will check the segment types  kkeil 2000/10/28
- * Also sysret mandates a special GDT layout 
- */
-		 		
-	.section .data.page_aligned, "aw"
-	.align PAGE_SIZE
-
-/* The TLS descriptors are currently at a different place compared to i386.
-   Hopefully nobody expects them at a fixed place (Wine?) */
 	
-ENTRY(cpu_gdt_table)
-	.quad	0x0000000000000000	/* NULL descriptor */
-	.quad	0x00cf9b000000ffff	/* __KERNEL32_CS */
-	.quad	0x00af9b000000ffff	/* __KERNEL_CS */
-	.quad	0x00cf93000000ffff	/* __KERNEL_DS */
-	.quad	0x00cffb000000ffff	/* __USER32_CS */
-	.quad	0x00cff3000000ffff	/* __USER_DS, __USER32_DS  */
-	.quad	0x00affb000000ffff	/* __USER_CS */
-	.quad	0x0			/* unused */
-	.quad	0,0			/* TSS */
-	.quad	0,0			/* LDT */
-	.quad   0,0,0			/* three TLS descriptors */ 
-	.quad	0x0000f40000000000	/* node/CPU stored in limit */
-gdt_end:	
-	/* asm/segment.h:GDT_ENTRIES must match this */	
-	/* This should be a multiple of the cache line size */
-	/* GDTs of other CPUs are now dynamically allocated */
-
-	/* zero the remaining page */
-	.fill PAGE_SIZE / 8 - GDT_ENTRIES,8,0
-
 	.section .bss, "aw", @nobits
 	.align L1_CACHE_BYTES
 ENTRY(idt_table)
arch/x86/kernel/setup64.c +1 -4
index fc1a56d..70ff071 100644
@@ -202,11 +202,8 @@ void __cpuinit cpu_init (void)
 	 * Initialize the per-CPU GDT with the boot GDT,
 	 * and set up the GDT descriptor:
 	 */
-	if (cpu)
-		memcpy(get_cpu_gdt_table(cpu), cpu_gdt_table, GDT_SIZE);
 
-	cpu_gdt_descr[cpu].size = GDT_SIZE;
-	load_gdt((const struct desc_ptr *)&cpu_gdt_descr[cpu]);
+	switch_to_new_gdt();
 	load_idt((const struct desc_ptr *)&idt_descr);
 
 	memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
arch/x86/kernel/setup_64.c +27 -0
index 2aee441..c57223f 100644
@@ -83,6 +83,22 @@
 #define ARCH_SETUP
 #endif
 
+/* We need valid kernel segments for data and code in long mode too
+ * IRET will check the segment types  kkeil 2000/10/28
+ * Also sysret mandates a special GDT layout
+ */
+/* The TLS descriptors are currently at a different place compared to i386.
+   Hopefully nobody expects them at a fixed place (Wine?) */
+DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
+	[GDT_ENTRY_KERNEL32_CS] = { { { 0x0000ffff, 0x00cf9b00 } } },
+	[GDT_ENTRY_KERNEL_CS] = { { { 0x0000ffff, 0x00af9b00 } } },
+	[GDT_ENTRY_KERNEL_DS] = { { { 0x0000ffff, 0x00cf9300 } } },
+	[GDT_ENTRY_DEFAULT_USER32_CS] = { { { 0x0000ffff, 0x00cffb00 } } },
+	[GDT_ENTRY_DEFAULT_USER_DS] = { { { 0x0000ffff, 0x00cff300 } } },
+	[GDT_ENTRY_DEFAULT_USER_CS] = { { { 0x0000ffff, 0x00affb00 } } },
+} };
+EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
+
 /*
  * Machine setup..
  */
@@ -273,6 +289,17 @@ void __attribute__((weak)) __init memory_setup(void)
        machine_specific_memory_setup();
 }
 
+/* Current gdt points %fs at the "master" per-cpu area: after this,
+ * it's on the real one. */
+void switch_to_new_gdt(void)
+{
+	struct desc_ptr gdt_descr;
+
+	gdt_descr.address = (long)get_cpu_gdt_table(smp_processor_id());
+	gdt_descr.size = GDT_SIZE - 1;
+	load_gdt(&gdt_descr);
+}
+
 /*
  * setup_arch - architecture-specific boot-time initializations
  *
arch/x86/kernel/smpboot.c +2 -8
index 181b109..69639fd 100644
@@ -849,14 +849,8 @@ static int __cpuinit do_boot_cpu(int apicid, int cpu)
 		.done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
 	};
 	INIT_WORK(&c_idle.work, do_fork_idle);
-#ifdef CONFIG_X86_64
-	/* allocate memory for gdts of secondary cpus. Hotplug is considered */
-	if (!cpu_gdt_descr[cpu].address &&
-		!(cpu_gdt_descr[cpu].address = get_zeroed_page(GFP_KERNEL))) {
-		printk(KERN_ERR "Failed to allocate GDT for CPU %d\n", cpu);
-		return -1;
-	}
 
+#ifdef CONFIG_X86_64
 	/* Allocate node local memory for AP pdas */
 	if (cpu > 0) {
 		boot_error = get_local_pda(cpu);
@@ -1253,8 +1247,8 @@ void __init native_smp_prepare_boot_cpu(void)
 	int me = smp_processor_id();
 #ifdef CONFIG_X86_32
 	init_gdt(me);
-	switch_to_new_gdt();
 #endif
+	switch_to_new_gdt();
 	/* already set me in cpu_online_map in boot_cpu_init() */
 	cpu_set(me, cpu_callout_map);
 	per_cpu(cpu_state, me) = CPU_ONLINE;
arch/x86/kernel/x8664_ksyms_64.c +0 -5
index 122885b..16a6751 100644
@@ -60,8 +60,3 @@ EXPORT_SYMBOL(init_level4_pgt);
 EXPORT_SYMBOL(load_gs_index);
 
 EXPORT_SYMBOL(_proxy_pda);
-
-#ifdef CONFIG_PARAVIRT
-/* Virtualized guests may want to use it */
-EXPORT_SYMBOL_GPL(cpu_gdt_descr);
-#endif
include/asm-x86/desc.h +10 -14
index b3875d4..07f9f2b 100644
@@ -29,11 +29,17 @@ static inline void fill_ldt(struct desc_struct *desc,
 extern struct desc_ptr idt_descr;
 extern gate_desc idt_table[];
 
+struct gdt_page {
+	struct desc_struct gdt[GDT_ENTRIES];
+} __attribute__((aligned(PAGE_SIZE)));
+DECLARE_PER_CPU(struct gdt_page, gdt_page);
+
+static inline struct desc_struct *get_cpu_gdt_table(unsigned int cpu)
+{
+	return per_cpu(gdt_page, cpu).gdt;
+}
+
 #ifdef CONFIG_X86_64
-extern struct desc_struct cpu_gdt_table[GDT_ENTRIES];
-extern struct desc_ptr cpu_gdt_descr[];
-/* the cpu gdt accessor */
-#define get_cpu_gdt_table(x) ((struct desc_struct *)cpu_gdt_descr[x].address)
 
 static inline void pack_gate(gate_desc *gate, unsigned type, unsigned long func,
 			     unsigned dpl, unsigned ist, unsigned seg)
@@ -51,16 +57,6 @@ static inline void pack_gate(gate_desc *gate, unsigned type, unsigned long func,
 }
 
 #else
-struct gdt_page {
-	struct desc_struct gdt[GDT_ENTRIES];
-} __attribute__((aligned(PAGE_SIZE)));
-DECLARE_PER_CPU(struct gdt_page, gdt_page);
-
-static inline struct desc_struct *get_cpu_gdt_table(unsigned int cpu)
-{
-	return per_cpu(gdt_page, cpu).gdt;
-}
-
 static inline void pack_gate(gate_desc *gate, unsigned char type,
 			     unsigned long base, unsigned dpl, unsigned flags,
 			     unsigned short seg)
include/asm-x86/segment.h +12 -11
index ed5131d..dfc8601 100644
@@ -61,18 +61,14 @@
 #define GDT_ENTRY_TLS_MAX 	(GDT_ENTRY_TLS_MIN + GDT_ENTRY_TLS_ENTRIES - 1)
 
 #define GDT_ENTRY_DEFAULT_USER_CS	14
-#define __USER_CS (GDT_ENTRY_DEFAULT_USER_CS * 8 + 3)
 
 #define GDT_ENTRY_DEFAULT_USER_DS	15
-#define __USER_DS (GDT_ENTRY_DEFAULT_USER_DS * 8 + 3)
 
 #define GDT_ENTRY_KERNEL_BASE	12
 
 #define GDT_ENTRY_KERNEL_CS		(GDT_ENTRY_KERNEL_BASE + 0)
-#define __KERNEL_CS (GDT_ENTRY_KERNEL_CS * 8)
 
 #define GDT_ENTRY_KERNEL_DS		(GDT_ENTRY_KERNEL_BASE + 1)
-#define __KERNEL_DS (GDT_ENTRY_KERNEL_DS * 8)
 
 #define GDT_ENTRY_TSS			(GDT_ENTRY_KERNEL_BASE + 4)
 #define GDT_ENTRY_LDT			(GDT_ENTRY_KERNEL_BASE + 5)
@@ -139,10 +135,11 @@
 #else
 #include <asm/cache.h>
 
-#define __KERNEL_CS	0x10
-#define __KERNEL_DS	0x18
+#define GDT_ENTRY_KERNEL32_CS 1
+#define GDT_ENTRY_KERNEL_CS 2
+#define GDT_ENTRY_KERNEL_DS 3
 
-#define __KERNEL32_CS   0x08
+#define __KERNEL32_CS   (GDT_ENTRY_KERNEL32_CS * 8)
 
 /*
  * we cannot use the same code segment descriptor for user and kernel
@@ -150,10 +147,10 @@
  * The segment offset needs to contain a RPL. Grr. -AK
  * GDT layout to get 64bit syscall right (sysret hardcodes gdt offsets)
  */
-
-#define __USER32_CS   0x23   /* 4*8+3 */
-#define __USER_DS     0x2b   /* 5*8+3 */
-#define __USER_CS     0x33   /* 6*8+3 */
+#define GDT_ENTRY_DEFAULT_USER32_CS 4
+#define GDT_ENTRY_DEFAULT_USER_DS 5
+#define GDT_ENTRY_DEFAULT_USER_CS 6
+#define __USER32_CS   (GDT_ENTRY_DEFAULT_USER32_CS * 8 + 3)
 #define __USER32_DS	__USER_DS
 
 #define GDT_ENTRY_TSS 8	/* needs two entries */
@@ -175,6 +172,10 @@
 
 #endif
 
+#define __KERNEL_CS	(GDT_ENTRY_KERNEL_CS * 8)
+#define __KERNEL_DS	(GDT_ENTRY_KERNEL_DS * 8)
+#define __USER_DS     (GDT_ENTRY_DEFAULT_USER_DS* 8 + 3)
+#define __USER_CS     (GDT_ENTRY_DEFAULT_USER_CS* 8 + 3)
 #ifndef CONFIG_PARAVIRT
 #define get_kernel_rpl()  0
 #endif
Signature · 1 line
1.5.4.5
#5 lore raw

PATCH 5/15 x86: use initial_code for i386

x86_64 jumps to whatever is written in "initial_code" symbol, instead of a fixed address. Do it for i386 too. It will allow us to integrate more of the smp boot code.

---
 arch/x86/kernel/head_32.S |    5 ++++-
 1 files changed, 4 insertions(+), 1 deletions(-)
arch/x86/kernel/head_32.S +4 -1
index b98b338..ffb73a5 100644
@@ -455,7 +455,10 @@ is386:	movl $2,%ecx		# set MP
 	jmp initialize_secondary # all other CPUs call initialize_secondary
 1:
 #endif /* CONFIG_SMP */
-	jmp i386_start_kernel
+	jmp *(initial_code)
+.align 4
+ENTRY(initial_code)
+	.long i386_start_kernel
 
 /*
  * We depend on ET to be correct. This checks for 287/387.
Signature · 1 line
1.5.4.5
#6 lore raw

PATCH 6/15 x86: boot secondary cpus through initial_code

remove "initialize_secondary". Boot both architectures via initial_code.

---
 arch/x86/kernel/head_32.S |    2 +-
 arch/x86/kernel/smpboot.c |   25 +------------------------
 2 files changed, 2 insertions(+), 25 deletions(-)
arch/x86/kernel/head_32.S +1 -1
index ffb73a5..f67e934 100644
@@ -452,7 +452,7 @@ is386:	movl $2,%ecx		# set MP
 	je   1f
 	movl $(__KERNEL_PERCPU), %eax
 	movl %eax,%fs		# set this cpu's percpu
-	jmp initialize_secondary # all other CPUs call initialize_secondary
+	movl (stack_start), %esp
 1:
 #endif /* CONFIG_SMP */
 	jmp *(initial_code)
arch/x86/kernel/smpboot.c +1 -24
index 69639fd..0e4a44a 100644
@@ -349,28 +349,6 @@ static void __cpuinit start_secondary(void *unused)
 	cpu_idle();
 }
 
-#ifdef CONFIG_X86_32
-/*
- * Everything has been set up for the secondary
- * CPUs - they just need to reload everything
- * from the task structure
- * This function must not return.
- */
-void __devinit initialize_secondary(void)
-{
-	/*
-	 * We don't actually need to load the full TSS,
-	 * basically just the stack pointer and the ip.
-	 */
-
-	asm volatile(
-		"movl %0,%%esp\n\t"
-		"jmp *%1"
-		:
-		:"m" (current->thread.sp), "m" (current->thread.ip));
-}
-#endif
-
 static void __cpuinit smp_apply_quirks(struct cpuinfo_x86 *c)
 {
 #ifdef CONFIG_X86_32
@@ -892,15 +870,14 @@ do_rest:
 #ifdef CONFIG_X86_32
 	per_cpu(current_task, cpu) = c_idle.idle;
 	init_gdt(cpu);
-	c_idle.idle->thread.ip = (unsigned long) start_secondary;
 	/* Stack for startup_32 can be just as for start_secondary onwards */
 	irq_ctx_init(cpu);
 #else
 	cpu_pda(cpu)->pcurrent = c_idle.idle;
 	load_sp0(&per_cpu(init_tss, cpu), &c_idle.idle->thread);
-	initial_code = (unsigned long)start_secondary;
 	clear_tsk_thread_flag(c_idle.idle, TIF_FORK);
 #endif
+	initial_code = (unsigned long)start_secondary;
 	stack_start.sp = (void *) c_idle.idle->thread.sp;
 
 	/* start_ip had better be page-aligned! */
Signature · 1 line
1.5.4.5
#7 lore raw

PATCH 7/15 x86: clearing io_apic harmless for x86_64

so remove ifdef.

---
 arch/x86/kernel/smpboot.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)
arch/x86/kernel/smpboot.c +1 -2
index 0e4a44a..5a055a1 100644
@@ -1052,9 +1052,8 @@ static __init void disable_smp(void)
 {
 	cpu_present_map = cpumask_of_cpu(0);
 	cpu_possible_map = cpumask_of_cpu(0);
-#ifdef CONFIG_X86_32
 	smpboot_clear_io_apic_irqs();
-#endif
+
 	if (smp_found_config)
 		phys_cpu_present_map =
 				physid_mask_of_physid(boot_cpu_physical_apicid);
Signature · 1 line
1.5.4.5
#8 lore raw

PATCH 8/15 x86: remove ifdef from stepping

The stepping won't affect x86_64, since there are not x86_64 k7's or pentiums. So, although it adds to the binary size, remove the ifdef for smoother integration

---
 arch/x86/kernel/smpboot.c |    2 --
 1 files changed, 0 insertions(+), 2 deletions(-)
arch/x86/kernel/smpboot.c +0 -2
index 5a055a1..50d307e 100644
@@ -351,7 +351,6 @@ static void __cpuinit start_secondary(void *unused)
 
 static void __cpuinit smp_apply_quirks(struct cpuinfo_x86 *c)
 {
-#ifdef CONFIG_X86_32
 	/*
 	 * Mask B, Pentium, but not Pentium MMX
 	 */
@@ -401,7 +400,6 @@ static void __cpuinit smp_apply_quirks(struct cpuinfo_x86 *c)
 
 valid_k7:
 	;
-#endif
 }
 
 static void __cpuinit smp_checks(void)
Signature · 1 line
1.5.4.5
#9 lore raw

PATCH 9/15 x86: change __setup_vector_irq with setup_vector_irq

We create a version of it for i386, and then take the CONFIG_X86_64 ifdef out of the game. We could create a __setup_vector_irq for i386, but it would incur in an unnecessary lock taking. Moreover, it is better practice to only export setup_vector_irq anyway.

---
 arch/x86/kernel/io_apic_32.c |    5 +++++
 arch/x86/kernel/io_apic_64.c |    9 ++++++++-
 arch/x86/kernel/smpboot.c    |   11 ++---------
 include/asm-x86/hw_irq.h     |    2 +-
 4 files changed, 16 insertions(+), 11 deletions(-)
arch/x86/kernel/io_apic_32.c +5 -0
index 9dbe7e9..7fc071f 100644
@@ -1209,6 +1209,11 @@ static int assign_irq_vector(int irq)
 
 	return vector;
 }
+
+void setup_vector_irq(int cpu)
+{
+}
+
 static struct irq_chip ioapic_chip;
 
 #define IOAPIC_AUTO	-1
arch/x86/kernel/io_apic_64.c +8 -1
index 5ea376d..97d67c9 100644
@@ -780,7 +780,7 @@ static void __clear_irq_vector(int irq)
 	cpus_clear(cfg->domain);
 }
 
-void __setup_vector_irq(int cpu)
+static void __setup_vector_irq(int cpu)
 {
 	/* Initialize vector_irq on a new cpu */
 	/* This function must be called with vector_lock held */
@@ -803,6 +803,13 @@ void __setup_vector_irq(int cpu)
 	}
 }
 
+void setup_vector_irq(int cpu)
+{
+	spin_lock(&vector_lock);
+	__setup_vector_irq(smp_processor_id());
+	spin_unlock(&vector_lock);
+}
+
 
 static struct irq_chip ioapic_chip;
 
arch/x86/kernel/smpboot.c +2 -9
index 50d307e..9d8dfec 100644
@@ -329,15 +329,8 @@ static void __cpuinit start_secondary(void *unused)
 	 * smp_call_function().
 	 */
 	lock_ipi_call_lock();
-#ifdef CONFIG_X86_64
-	spin_lock(&vector_lock);
-
-	/* Setup the per cpu irq handling data structures */
-	__setup_vector_irq(smp_processor_id());
-	/*
-	 * Allow the master to continue.
-	 */
-	spin_unlock(&vector_lock);
+#ifdef CONFIG_X86_IO_APIC
+	setup_vector_irq(smp_processor_id());
 #endif
 	cpu_set(smp_processor_id(), cpu_online_map);
 	unlock_ipi_call_lock();
include/asm-x86/hw_irq.h +1 -1
index 1428b41..18f067c 100644
@@ -97,9 +97,9 @@ extern void (*const interrupt[NR_IRQS])(void);
 #else
 typedef int vector_irq_t[NR_VECTORS];
 DECLARE_PER_CPU(vector_irq_t, vector_irq);
-extern void __setup_vector_irq(int cpu);
 extern spinlock_t vector_lock;
 #endif
+extern void setup_vector_irq(int cpu);
 
 #endif /* !ASSEMBLY_ */
 
Signature · 1 line
1.5.4.5
#10 lore raw

PATCH 10/15 x86: provide connect_bsp_APIC for x86_64

Although it is not really needed, we provide it to get closer to i386. ifdefs around it are removed in smpboot.c

---
 arch/x86/kernel/apic_64.c |   10 ++++++++++
 arch/x86/kernel/smpboot.c |    5 +----
 2 files changed, 11 insertions(+), 4 deletions(-)
arch/x86/kernel/apic_64.c +10 -0
index 310bec1..8d62387 100644
@@ -942,6 +942,8 @@ int __init APIC_init_uniprocessor(void)
 
 	verify_local_APIC();
 
+	connect_bsp_APIC();
+
 	phys_cpu_present_map = physid_mask_of_physid(boot_cpu_physical_apicid);
 	apic_write(APIC_ID, SET_APIC_ID(boot_cpu_physical_apicid));
 
@@ -1023,6 +1025,14 @@ asmlinkage void smp_error_interrupt(void)
 	irq_exit();
 }
 
+/**
+ *  * connect_bsp_APIC - attach the APIC to the interrupt system
+ *   */
+void __init connect_bsp_APIC(void)
+{
+	enable_apic_mode();
+}
+
 void disconnect_bsp_APIC(int virt_wire_setup)
 {
 	/* Go back to Virtual Wire compatibility mode */
arch/x86/kernel/smpboot.c +1 -4
index 9d8dfec..c4bb997 100644
@@ -1117,9 +1117,7 @@ static int __init smp_sanity_check(unsigned max_cpus)
 
 		localise_nmi_watchdog();
 
-#ifdef CONFIG_X86_32
 		connect_bsp_APIC();
-#endif
 		setup_local_APIC();
 		end_local_APIC_setup();
 		return -1;
@@ -1174,9 +1172,8 @@ void __init native_smp_prepare_cpus(unsigned int max_cpus)
 	}
 	preempt_enable();
 
-#ifdef CONFIG_X86_32
 	connect_bsp_APIC();
-#endif
+
 	/*
 	 * Switch from PIC to APIC mode.
 	 */
Signature · 1 line
1.5.4.5
#11 lore raw

PATCH 11/15 x86: move enabling of io_apic to prepare_cpus

Do it, instead of keeping in io_apic_32.c. This is the way x86_64 already does.

---
 arch/x86/kernel/apic_32.c    |    6 +-----
 arch/x86/kernel/io_apic_32.c |    2 +-
 arch/x86/kernel/smpboot.c    |    3 ++-
 include/asm-x86/hw_irq.h     |    3 ---
 4 files changed, 4 insertions(+), 10 deletions(-)
arch/x86/kernel/apic_32.c +1 -5
index fa8cf79..a06442a 100644
@@ -1277,11 +1277,7 @@ int __init APIC_init_uniprocessor(void)
 #endif
 		localise_nmi_watchdog();
 	end_local_APIC_setup();
-#ifdef CONFIG_X86_IO_APIC
-	if (smp_found_config)
-		if (!skip_ioapic_setup && nr_ioapics)
-			setup_IO_APIC();
-#endif
+
 	setup_boot_clock();
 
 	return 0;
arch/x86/kernel/io_apic_32.c +1 -1
index 7fc071f..cb79ce3 100644
@@ -1606,7 +1606,7 @@ void /*__init*/ print_PIC(void)
 
 #endif  /*  0  */
 
-static void __init enable_IO_APIC(void)
+void __init enable_IO_APIC(void)
 {
 	union IO_APIC_reg_01 reg_01;
 	int i8259_apic, i8259_pin;
arch/x86/kernel/smpboot.c +2 -1
index c4bb997..49d2900 100644
@@ -1179,13 +1179,14 @@ void __init native_smp_prepare_cpus(unsigned int max_cpus)
 	 */
 	setup_local_APIC();
 
-#ifdef CONFIG_X86_64
+#ifdef CONFIG_X86_IO_APIC
 	/*
 	 * Enable IO APIC before setting up error vector
 	 */
 	if (!skip_ioapic_setup && nr_ioapics)
 		enable_IO_APIC();
 #endif
+
 	end_local_APIC_setup();
 
 	map_cpu_to_logical_apicid();
include/asm-x86/hw_irq.h +0 -3
index 18f067c..31b56cd 100644
@@ -66,10 +66,7 @@ extern void disable_IO_APIC(void);
 extern void print_IO_APIC(void);
 extern int IO_APIC_get_PCI_irq_vector(int bus, int slot, int fn);
 extern void setup_ioapic_dest(void);
-
-#ifdef CONFIG_X86_64
 extern void enable_IO_APIC(void);
-#endif
 
 /* IPI functions */
 extern void send_IPI_self(int vector);
Signature · 1 line
1.5.4.5
#12 lore raw

PATCH 12/15 x86: change naming to match x86_64

Change unmap_cpu_to_logical_apicid to numa_remove_cpu. Besides being shorter, it is the same name x86_64 uses. We can save an ifdef in the code this way.

---
 arch/x86/kernel/smpboot.c |    8 ++------
 1 files changed, 2 insertions(+), 6 deletions(-)
arch/x86/kernel/smpboot.c +2 -6
index 49d2900..6316d30 100644
@@ -181,13 +181,12 @@ static void map_cpu_to_logical_apicid(void)
 	map_cpu_to_node(cpu, node);
 }
 
-static void unmap_cpu_to_logical_apicid(int cpu)
+static void numa_remove_cpu(int cpu)
 {
 	cpu_2_logical_apicid[cpu] = BAD_APICID;
 	unmap_cpu_to_node(cpu);
 }
 #else
-#define unmap_cpu_to_logical_apicid(cpu) do {} while (0)
 #define map_cpu_to_logical_apicid()  do {} while (0)
 #endif
 
@@ -945,10 +944,7 @@ restore_state:
 
 	if (boot_error) {
 		/* Try to put things back the way they were before ... */
-		unmap_cpu_to_logical_apicid(cpu);
-#ifdef CONFIG_X86_64
 		numa_remove_cpu(cpu); /* was set by numa_add_cpu */
-#endif
 		cpu_clear(cpu, cpu_callout_map); /* was set by do_boot_cpu() */
 		cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
 		cpu_clear(cpu, cpu_possible_map);
@@ -1246,7 +1242,7 @@ void cpu_exit_clear(void)
 	cpu_clear(cpu, cpu_callout_map);
 	cpu_clear(cpu, cpu_callin_map);
 
-	unmap_cpu_to_logical_apicid(cpu);
+	numa_remove_cpu(cpu);
 }
 #  endif /* CONFIG_X86_32 */
 
Signature · 1 line
1.5.4.5
#13 lore raw

PATCH 13/15 x86: remove cpu from maps

during cpu disable, take cpus out of all maps in i386, instead of just the online map.

---
 arch/x86/kernel/smpboot.c |    2 --
 1 files changed, 0 insertions(+), 2 deletions(-)
arch/x86/kernel/smpboot.c +0 -2
index 6316d30..f6b99b6 100644
@@ -1321,13 +1321,11 @@ __init void prefill_possible_map(void)
 static void __ref remove_cpu_from_maps(int cpu)
 {
 	cpu_clear(cpu, cpu_online_map);
-#ifdef CONFIG_X86_64
 	cpu_clear(cpu, cpu_callout_map);
 	cpu_clear(cpu, cpu_callin_map);
 	/* was set by cpu_init() */
 	clear_bit(cpu, (unsigned long *)&cpu_initialized);
 	numa_remove_cpu(cpu);
-#endif
 }
 
 int __cpu_disable(void)
Signature · 1 line
1.5.4.5
#14 lore raw

PATCH 14/15 x86: move cpu_exit_clear to process_32.c

Take it out of smpboot.c, and move it to process_32.c, closer to its only user.

---
 arch/x86/kernel/process_32.c |   16 ++++++++++++++++
 arch/x86/kernel/smpboot.c    |   19 +------------------
 include/asm-x86/numa_32.h    |    1 +
 include/asm-x86/smp.h        |    1 -
 4 files changed, 18 insertions(+), 19 deletions(-)
arch/x86/kernel/process_32.c +16 -0
index 1aaca76..99f8e7d 100644
@@ -128,6 +128,22 @@ EXPORT_SYMBOL(default_idle);
 
 #ifdef CONFIG_HOTPLUG_CPU
 #include <asm/nmi.h>
+
+static void cpu_exit_clear(void)
+{
+	int cpu = raw_smp_processor_id();
+
+	idle_task_exit();
+
+	cpu_uninit();
+	irq_ctx_exit(cpu);
+
+	cpu_clear(cpu, cpu_callout_map);
+	cpu_clear(cpu, cpu_callin_map);
+
+	numa_remove_cpu(cpu);
+}
+
 /* We don't actually take CPU down, just spin without interrupts. */
 static inline void play_dead(void)
 {
arch/x86/kernel/smpboot.c +1 -18
index f6b99b6..7952c39 100644
@@ -181,7 +181,7 @@ static void map_cpu_to_logical_apicid(void)
 	map_cpu_to_node(cpu, node);
 }
 
-static void numa_remove_cpu(int cpu)
+void numa_remove_cpu(int cpu)
 {
 	cpu_2_logical_apicid[cpu] = BAD_APICID;
 	unmap_cpu_to_node(cpu);
@@ -1229,23 +1229,6 @@ void __init native_smp_cpus_done(unsigned int max_cpus)
 
 #ifdef CONFIG_HOTPLUG_CPU
 
-#  ifdef CONFIG_X86_32
-void cpu_exit_clear(void)
-{
-	int cpu = raw_smp_processor_id();
-
-	idle_task_exit();
-
-	cpu_uninit();
-	irq_ctx_exit(cpu);
-
-	cpu_clear(cpu, cpu_callout_map);
-	cpu_clear(cpu, cpu_callin_map);
-
-	numa_remove_cpu(cpu);
-}
-#  endif /* CONFIG_X86_32 */
-
 static void remove_siblinginfo(int cpu)
 {
 	int sibling;
include/asm-x86/numa_32.h +1 -0
index 03d0f7a..205015f 100644
@@ -2,6 +2,7 @@
 #define _ASM_X86_32_NUMA_H 1
 
 extern int pxm_to_nid(int pxm);
+extern void numa_remove_cpu(int cpu);
 
 #ifdef CONFIG_NUMA
 extern void __init remap_numa_kva(void);
include/asm-x86/smp.h +0 -1
index fc10073..fad45f6 100644
@@ -188,7 +188,6 @@ static inline int hard_smp_processor_id(void)
 #endif /* CONFIG_X86_LOCAL_APIC */
 
 #ifdef CONFIG_HOTPLUG_CPU
-extern void cpu_exit_clear(void);
 extern void cpu_uninit(void);
 #endif
 
Signature · 1 line
1.5.4.5
#15 lore raw

PATCH 15/15 x86: take load_sp0 out of smpboot.c

there's no particular reason to do load_sp0 in different places for i386 and x86_64. They should all be in cpu_init. Right now, cpu_init itself is not integrated, but with this patch, the code becomes closer to each other, making in easier to integrate when the time comes.

Furthermore, although doing it in do_boot_cpu for x86_64 is fine, since it's only a copy, load_sp0 should be executed in the cpu it refers to anyway.

---
 arch/x86/kernel/setup64.c |    1 +
 arch/x86/kernel/smpboot.c |    1 -
 2 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/setup64.c +1 -0
index 70ff071..151d315 100644
@@ -247,6 +247,7 @@ void __cpuinit cpu_init (void)
 		BUG();
 	enter_lazy_tlb(&init_mm, me);
 
+	load_sp0(t, &current->thread);
 	set_tss_desc(cpu, t);
 	load_TR_desc();
 	load_LDT(&init_mm.context);
arch/x86/kernel/smpboot.c +0 -1
index 7952c39..371113a 100644
@@ -864,7 +864,6 @@ do_rest:
 	irq_ctx_init(cpu);
 #else
 	cpu_pda(cpu)->pcurrent = c_idle.idle;
-	load_sp0(&per_cpu(init_tss, cpu), &c_idle.idle->thread);
 	clear_tsk_thread_flag(c_idle.idle, TIF_FORK);
 #endif
 	initial_code = (unsigned long)start_secondary;
Signature · 1 line
1.5.4.5
#16 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

In reply to Glauber Costa
Quoted 2 lines · Glauber Costa

Do it, instead of keeping in io_apic_32.c. This is the way x86_64 already does.

This change looks wrong -- is native_smp_prepare_cpus() at all run on
!CONFIG_SMP? I don't think so. You have to do this differently.

Maciej

#17 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 7 lines · Maciej W. Rozycki
Quoted 2 lines · Glauber Costa

Do it, instead of keeping in io_apic_32.c. This is the way x86_64 already does.

This change looks wrong -- is native_smp_prepare_cpus() at all run on
!CONFIG_SMP? I don't think so. You have to do this differently.

You're right.

I'll update this one.

#18 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

In reply to Maciej W. Rozycki
Quoted 9 lines · Maciej W. Rozycki
Quoted 2 lines · Glauber Costa

Do it, instead of keeping in io_apic_32.c. This is the way x86_64 already does.

This change looks wrong -- is native_smp_prepare_cpus() at all run on
!CONFIG_SMP? I don't think so. You have to do this differently.

Maciej

How about this one instead?

This one does enable_IO_APIC in the init_uniprocessor too, and should account for the !smp case.

Quoted 1 line

From 87f8366012e1dc3cf7212e241b0aa552d7825338 Mon Sep 17 00:00:00 2001

From: Glauber Costa <[email protected]>
Date: Wed, 28 May 2008 17:04:12 -0300
Subject: [PATCH] x86: move enabling of io_apic to prepare_cpus

Do it, instead of keeping in io_apic_32.c. This is the way x86_64 already does.

---
 arch/x86/kernel/apic_32.c    |    3 +++
 arch/x86/kernel/io_apic_32.c |    4 ++--
 arch/x86/kernel/smpboot.c    |    2 +-
 include/asm-x86/hw_irq.h     |    2 --
 4 files changed, 6 insertions(+), 5 deletions(-)
arch/x86/kernel/apic_32.c +3 -0
index fa8cf79..8385dc6 100644
@@ -1273,6 +1273,9 @@ int __init APIC_init_uniprocessor(void)
 	setup_local_APIC();
 
 #ifdef CONFIG_X86_IO_APIC
+	if (!skip_ioapic_setup && nr_ioapics)
+		enable_IO_APIC();
+
 	if (!smp_found_config || skip_ioapic_setup || !nr_ioapics)
 #endif
 		localise_nmi_watchdog();
arch/x86/kernel/io_apic_32.c +2 -2
index 7fc071f..fa4f0a0 100644
@@ -1606,7 +1606,7 @@ void /*__init*/ print_PIC(void)
 
 #endif  /*  0  */
 
-static void __init enable_IO_APIC(void)
+void __init enable_IO_APIC(void)
 {
 	union IO_APIC_reg_01 reg_01;
 	int i8259_apic, i8259_pin;
@@ -2295,7 +2295,7 @@ void __init setup_IO_APIC(void)
 	for (i = first_system_vector; i < NR_VECTORS; i++)
 		set_bit(i, used_vectors);
 
-	enable_IO_APIC();
+	/* calling enable_IO_APIC() is moved to setup_local_APIC for BP */
 
 	if (acpi_ioapic)
 		io_apic_irqs = ~0;	/* all IRQs go through IOAPIC */
arch/x86/kernel/smpboot.c +1 -1
index f5ac5b2..e537c7b 100644
@@ -1175,7 +1175,7 @@ void __init native_smp_prepare_cpus(unsigned int max_cpus)
 	 */
 	setup_local_APIC();
 
-#ifdef CONFIG_X86_64
+#ifdef CONFIG_X86_IO_APIC
 	/*
 	 * Enable IO APIC before setting up error vector
 	 */
include/asm-x86/hw_irq.h +0 -2
index 18f067c..4c81d1b 100644
@@ -67,9 +67,7 @@ extern void print_IO_APIC(void);
 extern int IO_APIC_get_PCI_irq_vector(int bus, int slot, int fn);
 extern void setup_ioapic_dest(void);
 
-#ifdef CONFIG_X86_64
 extern void enable_IO_APIC(void);
-#endif
 
 /* IPI functions */
 extern void send_IPI_self(int vector);
Signature · 1 line
1.5.4.5
#19 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 2 lines · Glauber Costa

This one does enable_IO_APIC in the init_uniprocessor too, and should account for the !smp case.

Hmm, it looks a little bit better, but why do you want to call
enable_IO_APIC() separately in the first place? There is a comment
stating: "Enable IO APIC before setting up error vector," but why is it needed on 64-bit systems? Especially as the very same system may run a 32-bit kernel and then it suddenly would not have to do this anymore? Strange...

Also since you are cleaning up this code -- why don't you actually take
the opportunity and get rid of the horrible #ifdefs interspersed throughout?

Maciej

#20 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 11 lines · Maciej W. Rozycki
Quoted 2 lines · Glauber Costa

This one does enable_IO_APIC in the init_uniprocessor too, and should account for the !smp case.

Hmm, it looks a little bit better, but why do you want to call
enable_IO_APIC() separately in the first place? There is a comment
stating: "Enable IO APIC before setting up error vector," but why is it needed on 64-bit systems? Especially as the very same system may run a 32-bit kernel and then it suddenly would not have to do this anymore? Strange...

that enable_IO_APIC() only call record old ioapic/pin for timer and clear_IO_APIC.

we need clear_IO_APIC before enable setup error vector, in case there is wrong setting in ioapic by BIOS...

YH

#21 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 2 lines · Yinghai Lu

we need clear_IO_APIC before enable setup error vector, in case there is wrong setting in ioapic by BIOS...

Fair enough, although it is still interesting why it would only trigger
in the 64-bit mode and why shouldn't the BIOS be fixed instead.

Maciej

#22 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 5 lines · Maciej W. Rozycki
Quoted 2 lines

we need clear_IO_APIC before enable setup error vector, in case there is wrong setting in ioapic by BIOS...

Fair enough, although it is still interesting why it would only trigger
in the 64-bit mode and why shouldn't the BIOS be fixed instead.

Then again -- what if X86_LOCAL_APIC is set, but X86_IO_APIC is not?

Maciej

#23 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 9 lines · Maciej W. Rozycki
Quoted 5 lines · Maciej W. Rozycki
Quoted 2 lines

we need clear_IO_APIC before enable setup error vector, in case there is wrong setting in ioapic by BIOS...

Fair enough, although it is still interesting why it would only trigger
in the 64-bit mode and why shouldn't the BIOS be fixed instead.

Then again -- what if X86_LOCAL_APIC is set, but X86_IO_APIC is not?

config X86_LOCAL_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_APIC || ((X86_VISWS ||

SMP) && !X86_VOYAGER) || X86_GENERICARCH))

config X86_IO_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_IOAPIC || (SMP &&

!(X86_VISWS || X86_VOYAGER)) || X86_GENERICARCH))

for 64bit, those are all set.

for 32bit, may need to null stub if X86_IO_APIC is not set

YH

#24 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 17 lines · Yinghai Lu
Quoted 7 lines · Maciej W. Rozycki
Quoted 4 lines · Maciej W. Rozycki
Quoted 2 lines

we need clear_IO_APIC before enable setup error vector, in case there is wrong setting in ioapic by BIOS...

Fair enough, although it is still interesting why it would only trigger
in the 64-bit mode and why shouldn't the BIOS be fixed instead.

Then again -- what if X86_LOCAL_APIC is set, but X86_IO_APIC is not?

config X86_LOCAL_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_APIC || ((X86_VISWS ||

SMP) && !X86_VOYAGER) || X86_GENERICARCH))

config X86_IO_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_IOAPIC || (SMP &&
Quoted 7 lines

!(X86_VISWS || X86_VOYAGER)) || X86_GENERICARCH))

for 64bit, those are all set.

for 32bit, may need to null stub if X86_IO_APIC is not set

YH

Fair Enough.

#25 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 12 lines · Glauber Costa
Quoted 12 lines
Quoted 1 line

Then again -- what if X86_LOCAL_APIC is set, but X86_IO_APIC is not?

config X86_LOCAL_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_APIC || ((X86_VISWS ||

SMP) && !X86_VOYAGER) || X86_GENERICARCH))

config X86_IO_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_IOAPIC || (SMP &&

!(X86_VISWS || X86_VOYAGER)) || X86_GENERICARCH))

Quoted 6 lines
Quoted 5 lines

for 64bit, those are all set.

for 32bit, may need to null stub if X86_IO_APIC is not set

YH

Fair Enough.

That does not answer the question what to do with a 32-bit kernel run on
a system that requires the fixup in the 64-bit mode. Papering over such firmware bugs in the kernel encourages hardware vendors to maintain the breakage.

Note that the I/O APIC comes out of reset with all the redirection
entries masked out, so if any error conditions are triggered before Linux configures the chip, they are a result of a deliberate misprogramming done to the chip by the firmware.

Does anyone have a reference to the original problem report which lead to
this workaround having been put in place?

Maciej

#26 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 18 lines · Maciej W. Rozycki
Quoted 16 lines · Glauber Costa
Quoted 16 lines
Quoted 1 line

Then again -- what if X86_LOCAL_APIC is set, but X86_IO_APIC is not?

config X86_LOCAL_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_APIC || ((X86_VISWS ||

SMP) && !X86_VOYAGER) || X86_GENERICARCH))

config X86_IO_APIC

        def_bool y
        depends on X86_64 || (X86_32 && (X86_UP_IOAPIC || (SMP &&

!(X86_VISWS || X86_VOYAGER)) || X86_GENERICARCH))

for 64bit, those are all set.

for 32bit, may need to null stub if X86_IO_APIC is not set

Quoted 7 lines
Quoted 2 lines
Quoted 1 line

YH

Fair Enough.

That does not answer the question what to do with a 32-bit kernel run on
a system that requires the fixup in the 64-bit mode. Papering over such firmware bugs in the kernel encourages hardware vendors to maintain the breakage.

kernel should not assume io apic register is set right by firmware. and kernel actually doesn't trust them, and clear the io apic registers. that patch just move that early before enable error vector.

YH

#27 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 1 line · Yinghai Lu

kernel should not assume io apic register is set right by firmware.

What is the basis of this assumption? Linux generally assumes the
chipset components have been placed by the firmware into a consistent state. That does not necessarily mean suitable for Linux, hence the need to reconfigure a bit here or there, but there should be no need to touch components that are not going to be used by Linux directly. The I/O APIC is no different.

Quoted 1 line

and kernel actually doesn't trust them, and clear the io apic registers.

The I/O APIC registers are cleared, because this is the only way you can
assure inconsistent configuration does not happen during reconfiguration. For example two inputs using the same vector or set up into the ExtINTA mode. The original intent of the code was not to paper over breakage in the firmware. You are trying to change it and it can be done, but it has to be justified well.

Quoted 1 line

that patch just move that early before enable error vector.

What I am saying repeatedly is clearing of the I/O APIC is not guaranteed
to happen for all the possible cases of Linux configuration. Which means this is a partial solution only -- please try to propose a better one or provide the original problem report so that someone else can have a look at it. What's triggering the error interrupt for example? Is it recoverable?

Maciej

#28 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 20 lines · Maciej W. Rozycki
Quoted 1 line · Yinghai Lu

kernel should not assume io apic register is set right by firmware.

What is the basis of this assumption? Linux generally assumes the
chipset components have been placed by the firmware into a consistent state. That does not necessarily mean suitable for Linux, hence the need to reconfigure a bit here or there, but there should be no need to touch components that are not going to be used by Linux directly. The I/O APIC is no different.

Quoted 1 line

and kernel actually doesn't trust them, and clear the io apic registers.

The I/O APIC registers are cleared, because this is the only way you can
assure inconsistent configuration does not happen during reconfiguration. For example two inputs using the same vector or set up into the ExtINTA mode. The original intent of the code was not to paper over breakage in the firmware. You are trying to change it and it can be done, but it has to be justified well.

Quoted 8 lines
Quoted 1 line

that patch just move that early before enable error vector.

What I am saying repeatedly is clearing of the I/O APIC is not guaranteed
to happen for all the possible cases of Linux configuration. Which means this is a partial solution only -- please try to propose a better one or provide the original problem report so that someone else can have a look at it. What's triggering the error interrupt for example? Is it recoverable?

ExtINT is routed to ioapic pin0. but the dst is set to 0.
and the systems has multi sockets with quadcore cpu, so the apic id of boot cpu is set to 4 instead of 0

YH

#29 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 3 lines · Yinghai Lu

ExtINT is routed to ioapic pin0. but the dst is set to 0.
and the systems has multi sockets with quadcore cpu, so the apic id of boot cpu is set to 4 instead of 0

Thanks for the info. Let me understand the situation better: local APIC
IDs are preassigned by the firmware based on their "socket address" and the socket where the lowest numbered quad would be is empty. Nevertheless the firmware sets the destination ID of the ExtINTA interrupt in the I/O APIC to 0 rather than the ID of the bootstrap CPU. Is that correct?

But it would mean the Virtual Wire interrupt delivery would not work, or
is the I/O APIC setup redundant and the local APIC of the bootstrap CPU is set up for ExtINTA delivery as well?

Maciej

#30 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 6 lines · Maciej W. Rozycki
Quoted 3 lines · Yinghai Lu

ExtINT is routed to ioapic pin0. but the dst is set to 0.
and the systems has multi sockets with quadcore cpu, so the apic id of boot cpu is set to 4 instead of 0

Quoted 6 lines

Thanks for the info. Let me understand the situation better: local APIC
IDs are preassigned by the firmware based on their "socket address" and the socket where the lowest numbered quad would be is empty. Nevertheless the firmware sets the destination ID of the ExtINTA interrupt in the I/O APIC to 0 rather than the ID of the bootstrap CPU. Is that correct?

Yes

after I asked bios engineer to change the dest apic id to 4, the error is gone.

before clear_IO_APIC()
number of MP IRQ sources: 15.
number of IO-APIC #0 registers: 24.
number of IO-APIC #1 registers: 7.
number of IO-APIC #2 registers: 7.
number of IO-APIC #3 registers: 24.
testing the IO APIC.......................

IO APIC #0......
.... register #00: 00000000
....... : physical APIC id: 00
.... register #01: 00170011

.......     : max redirection entries: 0017
.......     : PRQ implemented: 0
.......     : IO APIC version: 0011

.... register #02: 00000000
....... : arbitration: 00
.... IRQ redirection table:
NR Dst Mask Trig IRR Pol Stat Dmod Deli Vect:

 00 004 0    0    0   0   0    0    7    00
 01 000 1    0    0   0   0    0    0    00
 02 000 1    0    0   0   0    0    0    00
 03 000 1    0    0   0   0    0    0    00
 04 000 1    0    0   0   0    0    0    00
 05 000 1    0    0   0   0    0    0    00
 06 000 1    0    0   0   0    0    0    00
 07 000 1    0    0   0   0    0    0    00
 08 000 1    0    0   0   0    0    0    00
 09 000 1    0    0   0   0    0    0    00
 0a 000 1    0    0   0   0    0    0    00
 0b 000 1    0    0   0   0    0    0    00
 0c 000 1    0    0   0   0    0    0    00
 0d 000 1    0    0   0   0    0    0    00
 0e 000 1    0    0   0   0    0    0    00
 0f 000 1    0    0   0   0    0    0    00
 10 000 1    0    0   0   0    0    0    00
 11 000 1    0    0   0   0    0    0    00
 12 000 1    0    0   0   0    0    0    00
 13 000 1    0    0   0   0    0    0    00
 14 000 1    0    0   0   0    0    0    00
 15 000 1    0    0   0   0    0    0    00
 16 000 1    0    0   0   0    0    0    00
 17 000 1    0    0   0   0    0    0    00
Quoted 4 lines

But it would mean the Virtual Wire interrupt delivery would not work, or
is the I/O APIC setup redundant and the local APIC of the bootstrap CPU is set up for ExtINTA delivery as well?

it doesn't need to virtual wire for timer (ExtInt), timer is hpet and routed to ioapic pin2.

Just know at first BIOS engineer refused to change that to 4, because other os is not happy.

YH

#31 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

Quoted 11 lines · Yinghai Lu
Quoted 6 lines

Thanks for the info. Let me understand the situation better: local APIC
IDs are preassigned by the firmware based on their "socket address" and the socket where the lowest numbered quad would be is empty. Nevertheless the firmware sets the destination ID of the ExtINTA interrupt in the I/O APIC to 0 rather than the ID of the bootstrap CPU. Is that correct?

Yes

after I asked bios engineer to change the dest apic id to 4, the error is gone.

Thanks for the clarification.

Quoted 6 lines
Quoted 3 lines

But it would mean the Virtual Wire interrupt delivery would not work, or
is the I/O APIC setup redundant and the local APIC of the bootstrap CPU is set up for ExtINTA delivery as well?

it doesn't need to virtual wire for timer (ExtInt), timer is hpet and routed to ioapic pin2.

That's not what I asked about -- the timer does not matter here. The
Virtual Wire mode is a configuration, where one input of one APIC in the system is set up for the ExtINTA mode and acts transparently with the system software having no need to know about it. Instead a pair of legacy 8259A chips is used to deliver interrupts, including claiming the INTA cycles, providing vectors and prioritising sources, as defined in the PC/AT architecture.

Many pieces of software rely on the 8259A PICs, either because they
predate the APIC or because they have no means to make use of multiprocessor features anyway. They include various versions of DOS together with software run in that environment (as DOS programs quite frequently drive hardware at the low level), many versions of the Microsoft Windows system as well as other software. For these a legacy mode, either the Virtual Wire mode, or a mode where 8259A interrupts are delivered directly to one processor's INT line has to be implemented as mandated both by the Multiprocessor Specification and the Advanced Configuration and Power Interface Specification.

Coming back to my question -- how is such a mode implemented in the
affected system? Clearly the route through the I/O APIC is not going to work and moreover, the chip clutters the system with broken interrupt messages each time the 8259A signals an interrupt.

Please note Linux can use the Virtual Wire mode in the APIC/SMP mode too,
if requested by specifying the "noapic" command-line option -- have you verified the option works correctly with the affected system?

Quoted 2 lines

Just know at first BIOS engineer refused to change that to 4, because other os is not happy.

Well, this is just a confirmation my attitude is correct -- such problems
should not be papered over, because vendors will deny their existence then. At least a complaint message should be printed so that users have an opportunity to see it and ask their hardware supplier for an explanation.

In this case, a workaround for the 64-bit mode happens to be quite cheap,
but it should be extended to cover the 32-bit mode as well. The only solution for the 32-bit mode I have in mind would lead to a waste of resources for many users of good hardware. And this because of somebody's sloppiness -- as I have written -- this better be well justified.

Unless you have precise means to identify this system -- in that case I
think reconfiguring the bootstrap processor's local APIC ID to 0 would be the right approach. Have you tried it?

Maciej

#32 lore raw

Re: [PATCH 11/15] x86: move enabling of io_apic to prepare_cpus

In reply to Maciej W. Rozycki
Quoted 9 lines · Maciej W. Rozycki
Quoted 1 line · Glauber Costa

This one does enable_IO_APIC in the init_uniprocessor too, and should account for the !smp case.

Hmm, it looks a little bit better, but why do you want to call
enable_IO_APIC() separately in the first place? There is a comment
stating: "Enable IO APIC before setting up error vector," but why is it needed on 64-bit systems? Especially as the very same system may run a 32-bit kernel and then it suddenly would not have to do this anymore? Strange...

This was reported by Yinghai, but I think he already answered to that.

Quoted 3 lines

Also since you are cleaning up this code -- why don't you actually take
the opportunity and get rid of the horrible #ifdefs interspersed throughout?

throughout where?
They're all over the place ;)

My next target would be per-cpu data. But that's because there's _a lot_ of code in the tree that got ifdefs between 32 and 64-bit because of differences in that, specially irq statistics. A macro would do, but if we're gonna do it, let's do it right.

#33 lore raw

Re: [PATCH 03/15] x86: remove early_gdt_descr reference

In reply to Glauber Costa
Quoted 3 lines · Glauber Costa

since we use switch_to_new_gdt, there is no point
in assigning early_gdt_descr except for the first
assignment, which is done manually.

What makes you think you can do this? If you don't update the early boot gdt, they all end up using the Boot CPU one. The problem with this is that there's a time from start_secondary to switch_to_new_gdt where the per cpu selector (%fs) and the pda selector (%gs) are those of the
boot CPU. The former isn't a problem but the CPU number is in the
latter, and it's used in that path before we get to the initialisation.

James

#34 lore raw

Re: [PATCH 03/15] x86: remove early_gdt_descr reference

Quoted 11 lines · James Bottomley
Quoted 3 lines · Glauber Costa

since we use switch_to_new_gdt, there is no point
in assigning early_gdt_descr except for the first
assignment, which is done manually.

What makes you think you can do this? If you don't update the early boot gdt, they all end up using the Boot CPU one. The problem with this is that there's a time from start_secondary to switch_to_new_gdt where the per cpu selector (%fs) and the pda selector (%gs) are those of the
boot CPU. The former isn't a problem but the CPU number is in the
latter, and it's used in that path before we get to the initialisation.

You are right, I missed it.

However, it only seem to be used in cpu_init, and very early. Sure there are some users _before_ we load the new gdt, but nothing prevents them to be moved after it. (Of course, this patch is wrong anyway).

And if we do that, we can even take the %fs loading out of head_32.S Of course, it's only valid if those are indeed the only early users of it.

Is there any other use I'm missing?

#35 lore raw

Re: [PATCH 03/15] x86: remove early_gdt_descr reference

Quoted 15 lines · Glauber Costa
Quoted 11 lines · James Bottomley
Quoted 3 lines · Glauber Costa

since we use switch_to_new_gdt, there is no point
in assigning early_gdt_descr except for the first
assignment, which is done manually.

What makes you think you can do this? If you don't update the early boot gdt, they all end up using the Boot CPU one. The problem with this is that there's a time from start_secondary to switch_to_new_gdt where the per cpu selector (%fs) and the pda selector (%gs) are those of the
boot CPU. The former isn't a problem but the CPU number is in the
latter, and it's used in that path before we get to the initialisation.

You are right, I missed it.

Quoted 8 lines

However, it only seem to be used in cpu_init, and very early. Sure there are some users _before_ we load the new gdt, but nothing prevents them to be moved after it. (Of course, this patch is wrong anyway).

And if we do that, we can even take the %fs loading out of head_32.S Of course, it's only valid if those are indeed the only early users of it.

Is there any other use I'm missing?

Well, %fs loading there is done for the boot CPU. To eliminate that you have to not only verify that start_secondary doesn't use anything in per_cpu areas, but also verify that nothing in start_kernel() up until boot_cpu_init() does ... That's a lot of smp_processor_id() references to convert.

James

#36 lore raw

Re: [PATCH 03/15] x86: remove early_gdt_descr reference

Quoted 19 lines · James Bottomley
Quoted 18 lines · Glauber Costa
Quoted 10 lines · James Bottomley
Quoted 3 lines · Glauber Costa

since we use switch_to_new_gdt, there is no point
in assigning early_gdt_descr except for the first
assignment, which is done manually.

What makes you think you can do this? If you don't update the early boot gdt, they all end up using the Boot CPU one. The problem with this is that there's a time from start_secondary to switch_to_new_gdt where the per cpu selector (%fs) and the pda selector (%gs) are those of the
boot CPU. The former isn't a problem but the CPU number is in the
latter, and it's used in that path before we get to the initialisation.

You are right, I missed it.

However, it only seem to be used in cpu_init, and very early. Sure there are some users _before_ we load the new gdt, but nothing prevents them to be moved after it. (Of course, this patch is wrong anyway).

And if we do that, we can even take the %fs loading out of head_32.S Of course, it's only valid if those are indeed the only early users of it.

Quoted 7 lines
Quoted 1 line

Is there any other use I'm missing?

Well, %fs loading there is done for the boot CPU. To eliminate that you have to not only verify that start_secondary doesn't use anything in per_cpu areas, but also verify that nothing in start_kernel() up until boot_cpu_init() does ... That's a lot of smp_processor_id() references to convert.

Yes, after a second look, it would be tricky indeed. But only for cpu0.
For all the others, I still think we could get rid of the problem by switching to the new gdt earlier in cpu_init.

What do you think?

#37 lore raw

Re: [PATCH 03/15] x86: remove early_gdt_descr reference

Quoted 29 lines · Glauber Costa
Quoted 21 lines · James Bottomley
Quoted 20 lines · Glauber Costa
Quoted 10 lines · James Bottomley
Quoted 3 lines · Glauber Costa

since we use switch_to_new_gdt, there is no point
in assigning early_gdt_descr except for the first
assignment, which is done manually.

What makes you think you can do this? If you don't update the early boot gdt, they all end up using the Boot CPU one. The problem with this is that there's a time from start_secondary to switch_to_new_gdt where the per cpu selector (%fs) and the pda selector (%gs) are those of the
boot CPU. The former isn't a problem but the CPU number is in the
latter, and it's used in that path before we get to the initialisation.

You are right, I missed it.

However, it only seem to be used in cpu_init, and very early. Sure there are some users _before_ we load the new gdt, but nothing prevents them to be moved after it. (Of course, this patch is wrong anyway).

And if we do that, we can even take the %fs loading out of head_32.S Of course, it's only valid if those are indeed the only early users of it.

Quoted 7 lines
Quoted 1 line

Is there any other use I'm missing?

Well, %fs loading there is done for the boot CPU. To eliminate that you have to not only verify that start_secondary doesn't use anything in per_cpu areas, but also verify that nothing in start_kernel() up until boot_cpu_init() does ... That's a lot of smp_processor_id() references to convert.

Quoted 5 lines

Yes, after a second look, it would be tricky indeed. But only for cpu0. For all the others, I still think we could get rid of the problem by switching to the new gdt earlier in cpu_init.

What do you think?

Operating a CPU with a bogus GDT is very fragile. You can fix all the current issues with the secondary CPUs, but it gives a critical section within which none of the per_cpu operations will work. It only takes one patch violating this rule and we have a very subtle bug introduced.

It looks to me like the better fix might be just to initialise the gdt completely and properly in do_boot_cpu and just have the single switch in head_32.S be the correct one. That way there's no problem critical region.

James