Per-cpu patches on top of PDA stuff...

Per-cpu patches on top of PDA stuff...

[Rusty Russell]

Hi Jeremy, all,

	Sorry this took so long, spent last week in Japan at OSDL conf then
netconf.  After several false starts, I ended up with a very simple
implementation, which clashes significantly with your work since then
8(.  I've pushed the patches anyway, but it's going to be significant
work for me to re-merge them, so I wanted your feedback first.

	The first patch simply changes the GDTs to be a straight per-cpu
variable.  I notice that you did a similar thing with your patches, but
this is simpler and avoids wasting space in the UP case.  It's a bit
tricky since we've never referred to per-cpu vars from asm before, but
since we're only referring to the pre-setup versions, it's ok.

	The second patch changes gs to be the per-cpu offset, and by
implication, avoids using it altogether on UP.  This avoids a special
"pda" structure, instead allowing all per-cpu variables to be accessed
this way.  It avoids __thread, which I gave up after creating a horribly
complicated patch which still didn't quite work, and was no more
efficient if we want the kernel to run under Xen anyway.

I really think this is the way to go, and I'll start work on merging
now.

Cheers!
Rusty.
-- 
Help! Save Australia from the worst of the DMCA: http://linux.org.au/law

Re: Per-cpu patches on top of PDA stuff...

[Rusty Russell]

On Tue, 2006-09-19 at 13:13 +1000, Rusty Russell wrote:
> Hi Jeremy, all,
> 
> 	Sorry this took so long, spent last week in Japan at OSDL conf then
> netconf.  After several false starts, I ended up with a very simple
> implementation, which clashes significantly with your work since then
> 8(.  I've pushed the patches anyway, but it's going to be significant
> work for me to re-merge them, so I wanted your feedback first.

OK, here's a patch against 2.6.18-rc6-mm2.  Tested on UP and SMP.
Crashes on hotplugging CPU, but crashes in same way as before the patch
8).

Replace PDA with per-cpu section, and put GDT in per-cpu section.

This patch uses the "gs" segment register which Jeremy Fitzhardinge
freed up for kernel use, for the per-cpu section.  This means that
instead of having a special per-cpu struct which we can access in a
single instruction, any per-cpu variable can be accessed in a single
instruction.  In addition, it avoids introducing the concept of a
"pda" into the kernel, in favour of the well-known "percpu" concept.

So, arch-specific code (eg. smp_processor_id()) can use
x86_write_percpu()/x86_read_percpu() directly.  Generic code expects
an lvalue from __get_cpu_var(), but it takes two instruction to get
the address of a per-cpu variable (still not bad).  Ideally, we could
use the __thread extension, and GCC would then generate optimal code
when an lvalue isn't needed, however, the linker wants to use a
negative offset within the gs register, which cannot be used with Xen
(or any similar hypervisor), because it requires a 4GB segment, which
would allow the OS to access the hypervisor memory.

As an additional simplification, the GDT is placed directly in a
per-cpu variable, rather than allocated dynamically.  This is optimal
for the UP case (previously, we made a copy even here), and
signficantly simplfies the code.  It's a little unusual to have asm
access a per-cpu var, but it is only done early at boot, where the
per-cpu GDT is sitting in the to-be-discarded section.

More cleanups/optimizations are possible:
1) Don't save/restore %gs on UP.  The cost is measurable, and we don't use it.
2) Remove early_smp_processor_id(), by setting up the per-cpu
   processor_id field correctly before starting a CPU.
3) Similarly, get rid of early_current().
4) Implement cpu_local_* in terms of x86_read_percpu etc.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/arch/i386/kernel/cpu/common.c working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/cpu/common.c
--- linux-2.6.18-rc6-mm2/arch/i386/kernel/cpu/common.c	2006-09-19 14:54:22.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/cpu/common.c	2006-09-19 15:27:29.000000000 +1000
@@ -19,18 +19,14 @@
 #include <asm/apic.h>
 #include <mach_apic.h>
 #endif
-#include <asm/pda.h>
 
 #include "cpu.h"
 
-DEFINE_PER_CPU(struct Xgt_desc_struct, cpu_gdt_descr);
-EXPORT_PER_CPU_SYMBOL(cpu_gdt_descr);
-
 DEFINE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
 EXPORT_PER_CPU_SYMBOL(cpu_16bit_stack);
 
-struct i386_pda *_cpu_pda[NR_CPUS] __read_mostly;
-EXPORT_SYMBOL(_cpu_pda);
+DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
+EXPORT_PER_CPU_SYMBOL(current_task);
 
 static int cachesize_override __cpuinitdata = -1;
 static int disable_x86_fxsr __cpuinitdata;
@@ -592,141 +587,10 @@ void __init early_cpu_init(void)
 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
 {
 	memset(regs, 0, sizeof(struct pt_regs));
-	regs->xgs = __KERNEL_PDA;
+	regs->xgs = __KERNEL_PERCPU;
 	return regs;
 }
 
-__cpuinit int alloc_gdt(int cpu)
-{
-	struct Xgt_desc_struct *cpu_gdt_descr = &per_cpu(cpu_gdt_descr, cpu);
-	struct desc_struct *gdt;
-	struct i386_pda *pda;
-
-	gdt = (struct desc_struct *)cpu_gdt_descr->address;
-	pda = cpu_pda(cpu);
-
-	/*
-	 * This is a horrible hack to allocate the GDT.  The problem
-	 * is that cpu_init() is called really early for the boot CPU
-	 * (and hence needs bootmem) but much later for the secondary
-	 * CPUs, when bootmem will have gone away
-	 */
-	if (NODE_DATA(0)->bdata->node_bootmem_map) {
-		BUG_ON(gdt != NULL || pda != NULL);
-
-		gdt = alloc_bootmem_pages(PAGE_SIZE);
-		pda = alloc_bootmem(sizeof(*pda));
-		/* alloc_bootmem(_pages) panics on failure, so no check */
-
-		memset(gdt, 0, PAGE_SIZE);
-		memset(pda, 0, sizeof(*pda));
-	} else {
-		/* GDT and PDA might already have been allocated if
-		   this is a CPU hotplug re-insertion. */
-		if (gdt == NULL)
-			gdt = (struct desc_struct *)get_zeroed_page(GFP_KERNEL);
-
-		if (pda == NULL)
-			pda = kmalloc_node(sizeof(*pda), GFP_KERNEL, cpu_to_node(cpu));
-
-		if (unlikely(!gdt || !pda)) {
-			free_pages((unsigned long)gdt, 0);
-			kfree(pda);
-			return 0;
-		}
-	}
-
- 	cpu_gdt_descr->address = (unsigned long)gdt;
-	cpu_pda(cpu) = pda;
-
-	return 1;
-}
-
-static __cpuinit void pda_init(int cpu, struct task_struct *curr)
-{
-	struct i386_pda *pda = cpu_pda(cpu);
-
-	memset(pda, 0, sizeof(*pda));
-
-	pda->cpu_number = cpu;
-	pda->pcurrent = curr;
-
-	printk("cpu %d current %p\n", cpu, curr);
-}
-
-static inline void set_kernel_gs(void)
-{
-	/* Set %gs for this CPU's PDA.  Memory clobber is to create a
-	   barrier with respect to any PDA operations, so the compiler
-	   doesn't move any before here. */
-	asm volatile ("mov %0, %%gs" : : "r" (__KERNEL_PDA) : "memory");
-}
-
-/* Initialize the CPU's GDT and PDA */
-static __cpuinit void init_gdt(void)
-{
-	int cpu = early_smp_processor_id();
-	struct task_struct *curr = early_current();
-	struct Xgt_desc_struct *cpu_gdt_descr = &per_cpu(cpu_gdt_descr, cpu);
-	__u32 stk16_off = (__u32)&per_cpu(cpu_16bit_stack, cpu);
-	struct desc_struct *gdt;
-	struct i386_pda *pda;
-
-	/* For non-boot CPUs, the GDT and PDA should already have been
-	   allocated. */
-	if (!alloc_gdt(cpu)) {
-		printk(KERN_CRIT "CPU%d failed to allocate GDT or PDA\n", cpu);
-		for (;;)
-			local_irq_enable();
-	}
-
-	gdt = (struct desc_struct *)cpu_gdt_descr->address;
-	pda = cpu_pda(cpu);
-
-	BUG_ON(gdt == NULL || pda == NULL);
-
-	/*
-	 * Initialize the per-CPU GDT with the boot GDT,
-	 * and set up the GDT descriptor:
-	 */
- 	memcpy(gdt, cpu_gdt_table, GDT_SIZE);
-	cpu_gdt_descr->size = GDT_SIZE - 1;
-
-	/* Set up GDT entry for 16bit stack */
- 	*(__u64 *)(&gdt[GDT_ENTRY_ESPFIX_SS]) |=
-		((((__u64)stk16_off) << 16) & 0x000000ffffff0000ULL) |
-		((((__u64)stk16_off) << 32) & 0xff00000000000000ULL) |
-		(CPU_16BIT_STACK_SIZE - 1);
-
-	pack_descriptor((u32 *)&gdt[GDT_ENTRY_PDA].a,
-			(u32 *)&gdt[GDT_ENTRY_PDA].b,
-			(unsigned long)pda, sizeof(*pda) - 1,
-			0x80 | DESCTYPE_S | 0x2, 0); /* present read-write data segment */
-
-	load_gdt(cpu_gdt_descr);
-	set_kernel_gs();
-
-	/* Do this once everything GDT-related has been set up. */
-	pda_init(cpu, curr);
-}
-
-/* Set up a very early PDA for the boot CPU so that smp_processor_id()
-   and current will work. */
-void __init smp_setup_processor_id(void)
-{
-	static __initdata struct i386_pda boot_pda;
-
-	pack_descriptor((u32 *)&cpu_gdt_table[GDT_ENTRY_PDA].a,
-			(u32 *)&cpu_gdt_table[GDT_ENTRY_PDA].b,
-			(unsigned long)&boot_pda, sizeof(struct i386_pda) - 1,
-			0x80 | DESCTYPE_S | 0x2, 0); /* present read-write data segment */
-
-	boot_pda.pcurrent = early_current();
-
-	/* Set %gs for this CPU's PDA */
-	set_kernel_gs();
-}
-
 /*
  * cpu_init() initializes state that is per-CPU. Some data is already
  * initialized (naturally) in the bootstrap process, such as the GDT
@@ -740,15 +604,27 @@ void __cpuinit cpu_init(void)
 
 	struct tss_struct * t = &per_cpu(init_tss, cpu);
 	struct thread_struct *thread = &curr->thread;
+	struct desc_struct *gdt;
+	u32 stk16_off;
 
 	if (cpu_test_and_set(cpu, cpu_initialized)) {
 		printk(KERN_WARNING "CPU#%d already initialized!\n", cpu);
 		for (;;) local_irq_enable();
 	}
 
-	/* Init the GDT and PDA early, before calling printk(),
-	   since it may end up using the PDA indirectly. */
-	init_gdt();
+	/* Complete percpu area setup early, before calling printk(),
+	   since it may end up using it indirectly. */
+	setup_percpu_for_this_cpu(cpu);
+	/* FIXME: Always the idle thread, can get rid of early_current. */
+	__get_cpu_var(current_task) = curr;
+
+	/* Set up GDT entry for 16bit stack */
+	stk16_off = (u32)&__get_cpu_var(cpu_16bit_stack);
+	gdt = __get_cpu_var(cpu_gdt_table);
+	*(__u64 *)(&gdt[GDT_ENTRY_ESPFIX_SS]) |=
+		((((__u64)stk16_off) << 16) & 0x000000ffffff0000ULL) |
+		((((__u64)stk16_off) << 32) & 0xff00000000000000ULL) |
+		(CPU_16BIT_STACK_SIZE - 1);
 
 	printk(KERN_INFO "Initializing CPU#%d\n", cpu);
 
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/arch/i386/kernel/entry.S working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/entry.S
--- linux-2.6.18-rc6-mm2/arch/i386/kernel/entry.S	2006-09-19 14:54:23.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/entry.S	2006-09-19 15:26:28.000000000 +1000
@@ -125,7 +125,7 @@ VM_MASK		= 0x00020000
 	movl $(__USER_DS), %edx; \
 	movl %edx, %ds; \
 	movl %edx, %es; \
-	movl $(__KERNEL_PDA), %edx; \
+	movl $(__KERNEL_PERCPU), %edx; \
 	movl %edx, %gs
 
 #define RESTORE_INT_REGS \
@@ -638,7 +638,7 @@ error_code:
 	movl $(__USER_DS), %ecx
 	movl %ecx, %ds
 	movl %ecx, %es
-	movl $(__KERNEL_PDA), %ecx
+	movl $(__KERNEL_PERCPU), %ecx
 	movl %ecx, %gs
 	movl %esp,%eax			# pt_regs pointer
 	call *%edi
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/arch/i386/kernel/head.S working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/head.S
--- linux-2.6.18-rc6-mm2/arch/i386/kernel/head.S	2006-09-19 14:54:23.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/head.S	2006-09-19 15:23:48.000000000 +1000
@@ -302,7 +302,7 @@ is386:	movl $2,%ecx		# set MP
 	movl %eax,%cr0
 
 	call check_x87
-	lgdt cpu_gdt_descr
+	lgdt per_cpu__cpu_gdt_descr
 	lidt idt_descr
 	ljmp $(__KERNEL_CS),$1f
 1:	movl $(__KERNEL_DS),%eax	# reload all the segment registers
@@ -523,12 +523,6 @@ idt_descr:
 	.word IDT_ENTRIES*8-1		# idt contains 256 entries
 	.long idt_table
 
-# boot GDT descriptor (later on used by CPU#0):
-	.word 0				# 32 bit align gdt_desc.address
-cpu_gdt_descr:
-	.word GDT_ENTRIES*8-1
-	.long cpu_gdt_table
-
 /*
  * The boot_gdt_table must mirror the equivalent in setup.S and is
  * used only for booting.
@@ -539,55 +533,3 @@ ENTRY(boot_gdt_table)
 	.quad 0x00cf9a000000ffff	/* kernel 4GB code at 0x00000000 */
 	.quad 0x00cf92000000ffff	/* kernel 4GB data at 0x00000000 */
 
-/*
- * The Global Descriptor Table contains 28 quadwords, per-CPU.
- */
-	.align L1_CACHE_BYTES
-ENTRY(cpu_gdt_table)
-	.quad 0x0000000000000000	/* NULL descriptor */
-	.quad 0x0000000000000000	/* 0x0b reserved */
-	.quad 0x0000000000000000	/* 0x13 reserved */
-	.quad 0x0000000000000000	/* 0x1b reserved */
-	.quad 0x0000000000000000	/* 0x20 unused */
-	.quad 0x0000000000000000	/* 0x28 unused */
-	.quad 0x0000000000000000	/* 0x33 TLS entry 1 */
-	.quad 0x0000000000000000	/* 0x3b TLS entry 2 */
-	.quad 0x0000000000000000	/* 0x43 TLS entry 3 */
-	.quad 0x0000000000000000	/* 0x4b reserved */
-	.quad 0x0000000000000000	/* 0x53 reserved */
-	.quad 0x0000000000000000	/* 0x5b reserved */
-
-	.quad 0x00cf9a000000ffff	/* 0x60 kernel 4GB code at 0x00000000 */
-	.quad 0x00cf92000000ffff	/* 0x68 kernel 4GB data at 0x00000000 */
-	.quad 0x00cffa000000ffff	/* 0x73 user 4GB code at 0x00000000 */
-	.quad 0x00cff2000000ffff	/* 0x7b user 4GB data at 0x00000000 */
-
-	.quad 0x0000000000000000	/* 0x80 TSS descriptor */
-	.quad 0x0000000000000000	/* 0x88 LDT descriptor */
-
-	/*
-	 * Segments used for calling PnP BIOS have byte granularity.
-	 * They code segments and data segments have fixed 64k limits,
-	 * the transfer segment sizes are set at run time.
-	 */
-	.quad 0x00409a000000ffff	/* 0x90 32-bit code */
-	.quad 0x00009a000000ffff	/* 0x98 16-bit code */
-	.quad 0x000092000000ffff	/* 0xa0 16-bit data */
-	.quad 0x0000920000000000	/* 0xa8 16-bit data */
-	.quad 0x0000920000000000	/* 0xb0 16-bit data */
-
-	/*
-	 * The APM segments have byte granularity and their bases
-	 * are set at run time.  All have 64k limits.
-	 */
-	.quad 0x00409a000000ffff	/* 0xb8 APM CS    code */
-	.quad 0x00009a000000ffff	/* 0xc0 APM CS 16 code (16 bit) */
-	.quad 0x004092000000ffff	/* 0xc8 APM DS    data */
-
-	.quad 0x0000920000000000	/* 0xd0 - ESPFIX 16-bit SS */
-	.quad 0x0000000000000000	/* 0xd8 - PDA */
-	.quad 0x0000000000000000	/* 0xe0 - unused */
-	.quad 0x0000000000000000	/* 0xe8 - unused */
-	.quad 0x0000000000000000	/* 0xf0 - unused */
-	.quad 0x0000000000000000	/* 0xf8 - GDT entry 31: double-fault TSS */
-
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/arch/i386/kernel/process.c working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/process.c
--- linux-2.6.18-rc6-mm2/arch/i386/kernel/process.c	2006-09-19 14:54:24.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/process.c	2006-09-19 15:26:28.000000000 +1000
@@ -38,6 +38,7 @@
 #include <linux/ptrace.h>
 #include <linux/random.h>
 #include <linux/personality.h>
+#include <linux/percpu.h>
 
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
@@ -56,7 +57,6 @@
 
 #include <asm/tlbflush.h>
 #include <asm/cpu.h>
-#include <asm/pda.h>
 
 asmlinkage void ret_from_fork(void) __asm__("ret_from_fork");
 
@@ -345,7 +345,7 @@ int kernel_thread(int (*fn)(void *), voi
 
 	regs.xds = __USER_DS;
 	regs.xes = __USER_DS;
-	regs.xgs = __KERNEL_PDA;
+	regs.xgs = __KERNEL_PERCPU;
 	regs.orig_eax = -1;
 	regs.eip = (unsigned long) kernel_thread_helper;
 	regs.xcs = __KERNEL_CS | get_kernel_rpl();
@@ -684,7 +684,7 @@ struct task_struct fastcall * __switch_t
 	if (unlikely(prev->fs | next->fs))
 		loadsegment(fs, next->fs);
 
-	write_pda(pcurrent, next_p);
+	x86_write_percpu(current_task, next_p);
 
 	/*
 	 * Restore IOPL if needed.
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/arch/i386/kernel/setup.c working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/setup.c
--- linux-2.6.18-rc6-mm2/arch/i386/kernel/setup.c	2006-09-19 14:54:24.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/setup.c	2006-09-19 15:26:28.000000000 +1000
@@ -1470,6 +1470,52 @@ void __init setup_arch(char **cmdline_p)
 	tsc_init();
 }
 
+/*
+ * The Global Descriptor Table contains 28 quadwords, per-CPU.
+ */
+__attribute__((aligned(L1_CACHE_BYTES)))
+DEFINE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]) =
+{
+	/* kernel 4GB code at 0x00000000 */
+	[GDT_ENTRY_KERNEL_CS] = { 0x0000ffff, 0x00cf9a00 },
+	/* kernel 4GB data at 0x00000000 */
+	[GDT_ENTRY_KERNEL_DS] = { 0x0000ffff, 0x00cf9200 },
+	/* user 4GB code at 0x00000000 */
+	[GDT_ENTRY_DEFAULT_USER_CS] = { 0x0000ffff, 0x00cffa00 },
+	/* user 4GB data at 0x00000000 */
+	[GDT_ENTRY_DEFAULT_USER_DS] = { 0x0000ffff, 0x00cff200 },
+	/*
+	 * Segments used for calling PnP BIOS have byte granularity.
+	 * They code segments and data segments have fixed 64k limits,
+	 * the transfer segment sizes are set at run time.
+	 */
+	[GDT_ENTRY_PNPBIOS_BASE] =
+	{ 0x0000ffff, 0x00409a00 }, /* 32-bit code */
+	{ 0x0000ffff, 0x00009a00 }, /* 16-bit code */
+	{ 0x0000ffff, 0x00009200 }, /* 16-bit data */
+	{ 0x00000000, 0x00009200 }, /* 16-bit data */
+	{ 0x00000000, 0x00009200 }, /* 16-bit data */
+
+	/*
+	 * The APM segments have byte granularity and their bases
+	 * are set at run time.  All have 64k limits.
+	 */
+	[GDT_ENTRY_APMBIOS_BASE] =
+	{ 0x0000ffff, 0x00409a00 }, /* APM CS    code */
+	{ 0x0000ffff, 0x00009a00 }, /* APM CS 16 code (16 bit) */
+	{ 0x0000ffff, 0x00409200 }, /* APM DS    data */
+
+	/* ESPFIX 16-bit SS */
+	[GDT_ENTRY_ESPFIX_SS] = { 0x00000000, 0x00009200 },
+	/* FIXME: We save/restore %gs even on UP: fix entry.S. */
+	[GDT_ENTRY_PERCPU] = { 0x0000ffff, 0x00cf9200 },
+};
+
+/* Early in boot we use the master per-cpu gdt_table directly. */
+DEFINE_PER_CPU(struct Xgt_desc_struct, cpu_gdt_descr) 
+= { .size = GDT_ENTRIES*8-1, .address = (long)&per_cpu__cpu_gdt_table };
+EXPORT_PER_CPU_SYMBOL(cpu_gdt_descr);
+
 static __init int add_pcspkr(void)
 {
 	struct platform_device *pd;
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/arch/i386/kernel/smpboot.c working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/smpboot.c
--- linux-2.6.18-rc6-mm2/arch/i386/kernel/smpboot.c	2006-09-19 14:54:24.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/arch/i386/kernel/smpboot.c	2006-09-19 15:26:28.000000000 +1000
@@ -60,6 +60,9 @@
 /* Set if we find a B stepping CPU */
 static int __devinitdata smp_b_stepping;
 
+DEFINE_PER_CPU(unsigned int, processor_id);
+EXPORT_PER_CPU_SYMBOL(processor_id);
+
 /* Number of siblings per CPU package */
 int smp_num_siblings = 1;
 #ifdef CONFIG_X86_HT
@@ -104,6 +107,9 @@ EXPORT_SYMBOL(x86_cpu_to_apicid);
 
 u8 apicid_2_node[MAX_APICID];
 
+DEFINE_PER_CPU(unsigned long, this_cpu_off);
+EXPORT_PER_CPU_SYMBOL(this_cpu_off);
+
 /*
  * Trampoline 80x86 program as an array.
  */
@@ -934,14 +940,6 @@ static int __devinit do_boot_cpu(int api
 	unsigned long start_eip;
 	unsigned short nmi_high = 0, nmi_low = 0;
 
-	/* Pre-allocate the CPU's GDT and PDA so it doesn't have to do
-	   any memory allocation during the delicate CPU-bringup
-	   phase. */
-	if (!alloc_gdt(cpu)) {
-		printk(KERN_INFO "Couldn't allocate GDT/PDA for CPU %d\n", cpu);
-		return -1;	/* ? */
-	}
-
 	++cpucount;
 	alternatives_smp_switch(1);
 
@@ -1072,7 +1070,6 @@ static int __cpuinit __smp_prepare_cpu(i
 	struct warm_boot_cpu_info info;
 	struct work_struct task;
 	int	apicid, ret;
-	struct Xgt_desc_struct *cpu_gdt_descr = &per_cpu(cpu_gdt_descr, cpu);
 
 	apicid = x86_cpu_to_apicid[cpu];
 	if (apicid == BAD_APICID) {
@@ -1080,18 +1077,6 @@ static int __cpuinit __smp_prepare_cpu(i
 		goto exit;
 	}
 
-	/*
-	 * the CPU isn't initialized at boot time, allocate gdt table here.
-	 * cpu_init will initialize it
-	 */
-	if (!cpu_gdt_descr->address) {
-		cpu_gdt_descr->address = get_zeroed_page(GFP_KERNEL);
-		if (!cpu_gdt_descr->address)
-			printk(KERN_CRIT "CPU%d failed to allocate GDT\n", cpu);
-			ret = -ENOMEM;
-			goto exit;
-	}
-
 	info.complete = &done;
 	info.apicid = apicid;
 	info.cpu = cpu;
@@ -1330,6 +1315,37 @@ static void __init smp_boot_cpus(unsigne
 		synchronize_tsc_bp();
 }
 
+static inline void set_kernel_gs(void)
+{
+	/* Set %gs for this CPU's per-cpu area.  Memory clobber is to create a
+	   barrier with respect to any per-cpu operations, so the compiler
+	   doesn't move any before here. */
+	asm volatile ("mov %0, %%gs" : : "r" (__KERNEL_PERCPU) : "memory");
+}
+
+static __cpuinit void setup_percpu_descriptor(struct desc_struct *gdt,
+					      unsigned long per_cpu_off)
+{
+	unsigned limit, flags;
+
+	limit = (1 << 20);
+	flags = 0x8;		/* 4k granularity */
+
+	/* present read-write data segment */
+	pack_descriptor((u32 *)&gdt->a, (u32 *)&gdt->b,
+			per_cpu_off, limit - 1,
+			0x80 | DESCTYPE_S | 0x2, flags);
+}
+
+/* Set up a very early per-cpu for the boot CPU so that smp_processor_id()
+   and current will work. */
+void __init smp_setup_processor_id(void)
+{
+	/* We use the per-cpu template area (__per_cpu_offset[0] == 0). */
+	__per_cpu_offset[0] = 0;
+	setup_percpu_for_this_cpu(0);
+}
+
 /* These are wrappers to interface to the new boot process.  Someone
    who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
 void __init smp_prepare_cpus(unsigned int max_cpus)
@@ -1340,8 +1356,26 @@ void __init smp_prepare_cpus(unsigned in
 	smp_boot_cpus(max_cpus);
 }
 
+/* Be careful not to use %gs references until this is setup: needs to
+ * be done on this CPU. */
+void __init setup_percpu_for_this_cpu(unsigned int cpu)
+{
+	struct desc_struct *gdt = per_cpu(cpu_gdt_table, cpu);
+	struct Xgt_desc_struct *cpu_gdt_descr = &per_cpu(cpu_gdt_descr, cpu);
+
+	per_cpu(processor_id, cpu) = cpu;
+	per_cpu(this_cpu_off, cpu) = __per_cpu_offset[cpu];
+	setup_percpu_descriptor(&gdt[GDT_ENTRY_PERCPU],	__per_cpu_offset[cpu]);
+	cpu_gdt_descr->address = (unsigned long)gdt;
+	cpu_gdt_descr->size = GDT_SIZE - 1;
+	load_gdt(cpu_gdt_descr);
+	set_kernel_gs();
+}
+
 void __devinit smp_prepare_boot_cpu(void)
 {
+	setup_percpu_for_this_cpu(0);
+
 	cpu_set(smp_processor_id(), cpu_online_map);
 	cpu_set(smp_processor_id(), cpu_callout_map);
 	cpu_set(smp_processor_id(), cpu_present_map);
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/current.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/current.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/current.h	2006-09-19 14:55:55.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/current.h	2006-09-19 15:26:28.000000000 +1000
@@ -2,7 +2,7 @@
 #define _I386_CURRENT_H
 
 #include <linux/thread_info.h>
-#include <asm/pda.h>
+#include <asm/percpu.h>
 
 struct task_struct;
 
@@ -11,11 +11,7 @@ static __always_inline struct task_struc
 	return current_thread_info()->task;
 }
 
-static __always_inline struct task_struct *get_current(void)
-{
-	return read_pda(pcurrent);
-}
-
-#define current get_current()
+DECLARE_PER_CPU(struct task_struct *, current_task);
+#define current x86_read_percpu(current_task)
 
 #endif /* !(_I386_CURRENT_H) */
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/desc.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/desc.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/desc.h	2006-09-19 14:55:55.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/desc.h	2006-09-19 15:23:48.000000000 +1000
@@ -14,8 +14,8 @@
 
 #include <asm/mmu.h>
 
-extern struct desc_struct cpu_gdt_table[GDT_ENTRIES];
-
+DECLARE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
+DECLARE_PER_CPU(struct Xgt_desc_struct, cpu_gdt_descr);
 DECLARE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
 
 struct Xgt_desc_struct {
@@ -25,8 +25,6 @@ struct Xgt_desc_struct {
 } __attribute__ ((packed));
 
 extern struct Xgt_desc_struct idt_descr;
-DECLARE_PER_CPU(struct Xgt_desc_struct, cpu_gdt_descr);
-
 
 static inline struct desc_struct *get_cpu_gdt_table(unsigned int cpu)
 {
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/pda.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/pda.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/pda.h	2006-09-19 14:55:56.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/pda.h	1970-01-01 10:00:00.000000000 +1000
@@ -1,68 +0,0 @@
-#ifndef _I386_PDA_H
-#define _I386_PDA_H
-
-struct i386_pda
-{
-	struct task_struct *pcurrent;	/* current process */
-	int cpu_number;
-};
-
-extern struct i386_pda *_cpu_pda[];
-
-#define cpu_pda(i)	(_cpu_pda[i])
-
-#define pda_offset(field) offsetof(struct i386_pda, field)
-
-extern void __bad_pda_field(void);
-
-extern struct i386_pda _proxy_pda;
-
-#define pda_to_op(op,field,val)						\
-	do {								\
-		typedef typeof(_proxy_pda.field) T__;			\
-		if (0) { T__ tmp__; tmp__ = (val); }			\
-		switch (sizeof(_proxy_pda.field)) {			\
-		case 2:							\
-			asm(op "w %1,%%gs:%c2"				\
-			    : "+m" (_proxy_pda.field)			\
-			    :"ri" ((T__)val),				\
-			     "i"(pda_offset(field)));			\
-			break;						\
-		case 4:							\
-			asm(op "l %1,%%gs:%c2"				\
-			    : "+m" (_proxy_pda.field)			\
-			    :"ri" ((T__)val),				\
-			     "i"(pda_offset(field)));			\
-			break;						\
-		default: __bad_pda_field();				\
-		}							\
-	} while (0)
-
-#define pda_from_op(op,field)						\
-	({								\
-		typeof(_proxy_pda.field) ret__;				\
-		switch (sizeof(_proxy_pda.field)) {			\
-		case 2:							\
-			asm(op "w %%gs:%c1,%0"				\
-			    : "=r" (ret__)				\
-			    : "i" (pda_offset(field)),			\
-			      "m" (_proxy_pda.field));			\
-			break;						\
-		case 4:							\
-			asm(op "l %%gs:%c1,%0"				\
-			    : "=r" (ret__)				\
-			    : "i" (pda_offset(field)),			\
-			      "m" (_proxy_pda.field));			\
-			break;						\
-		default: __bad_pda_field();				\
-		}							\
-		ret__; })
-
-
-#define read_pda(field) pda_from_op("mov",field)
-#define write_pda(field,val) pda_to_op("mov",field,val)
-#define add_pda(field,val) pda_to_op("add",field,val)
-#define sub_pda(field,val) pda_to_op("sub",field,val)
-#define or_pda(field,val) pda_to_op("or",field,val)
-
-#endif	/* _I386_PDA_H */
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/percpu.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/percpu.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/percpu.h	2004-02-04 14:44:44.000000000 +1100
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/percpu.h	2006-09-19 15:26:28.000000000 +1000
@@ -1,6 +1,107 @@
 #ifndef __ARCH_I386_PERCPU__
 #define __ARCH_I386_PERCPU__
 
+#ifdef CONFIG_SMP
+/* Same as generic implementation except for optimized local access. */
+#define __GENERIC_PER_CPU
+
+/* This is used for other cpus to find our section. */
+extern unsigned long __per_cpu_offset[NR_CPUS];
+
+/* Separate out the type, so (int[3], foo) works. */
+#define DECLARE_PER_CPU(type, name) extern __typeof__(type) per_cpu__##name
+#define DEFINE_PER_CPU(type, name) \
+    __attribute__((__section__(".data.percpu"))) __typeof__(type) per_cpu__##name
+
+/* We can use this directly for local CPU (faster). */
+DECLARE_PER_CPU(unsigned long, this_cpu_off);
+
+/* var is in discarded region: offset to particular copy we want */
+#define per_cpu(var, cpu) (*({				\
+	extern int simple_indentifier_##var(void);	\
+	RELOC_HIDE(&per_cpu__##var, __per_cpu_offset[cpu]); }))
+
+#define __raw_get_cpu_var(var) (*({					\
+	extern int simple_indentifier_##var(void);			\
+	RELOC_HIDE(&per_cpu__##var, x86_read_percpu(this_cpu_off));	\
+}))
+
+#define __get_cpu_var(var) __raw_get_cpu_var(var)
+
+/* A macro to avoid #include hell... */
+#define percpu_modcopy(pcpudst, src, size)			\
+do {								\
+	unsigned int __i;					\
+	for_each_possible_cpu(__i)				\
+		memcpy((pcpudst)+__per_cpu_offset[__i],		\
+		       (src), (size));				\
+} while (0)
+
+#define EXPORT_PER_CPU_SYMBOL(var) EXPORT_SYMBOL(per_cpu__##var)
+#define EXPORT_PER_CPU_SYMBOL_GPL(var) EXPORT_SYMBOL_GPL(per_cpu__##var)
+
+/* gs segment starts at (positive) offset == __per_cpu_offset[cpu] */
+#define __percpu_seg "%%gs:"
+#else  /* !SMP */
 #include <asm-generic/percpu.h>
+#define __percpu_seg ""
+#endif	/* SMP */
+
+/* For arch-specific code, we can use direct single-insn ops (they
+ * don't give an lvalue though). */
+extern void __bad_percpu_size(void);
+
+#define percpu_to_op(op,var,val)				\
+	do {							\
+		typedef typeof(var) T__;			\
+		if (0) { T__ tmp__; tmp__ = (val); }		\
+		switch (sizeof(var)) {				\
+		case 1:						\
+			asm(op "b %1,"__percpu_seg"%0"		\
+			    : "+m" (var)			\
+			    :"ri" ((T__)val));			\
+			break;					\
+		case 2:						\
+			asm(op "w %1,"__percpu_seg"%0"		\
+			    : "+m" (var)			\
+			    :"ri" ((T__)val));			\
+			break;					\
+		case 4:						\
+			asm(op "l %1,"__percpu_seg"%0"		\
+			    : "+m" (var)			\
+			    :"ri" ((T__)val));			\
+			break;					\
+		default: __bad_percpu_size();			\
+		}						\
+	} while (0)
+
+#define percpu_from_op(op,var)					\
+	({							\
+		typeof(var) ret__;				\
+		switch (sizeof(var)) {				\
+		case 1:						\
+			asm(op "b "__percpu_seg"%1,%0"		\
+			    : "=r" (ret__)			\
+			    : "m" (var));			\
+			break;					\
+		case 2:						\
+			asm(op "w "__percpu_seg"%1,%0"		\
+			    : "=r" (ret__)			\
+			    : "m" (var));			\
+			break;					\
+		case 4:						\
+			asm(op "l "__percpu_seg"%1,%0"		\
+			    : "=r" (ret__)			\
+			    : "m" (var));			\
+			break;					\
+		default: __bad_percpu_size();			\
+		}						\
+		ret__; })
+
+#define x86_read_percpu(var) percpu_from_op("mov", per_cpu__##var)
+#define x86_write_percpu(var,val) percpu_to_op("mov", per_cpu__##var, val)
+#define x86_add_percpu(var,val) percpu_to_op("add", per_cpu__##var, val)
+#define x86_sub_percpu(var,val) percpu_to_op("sub", per_cpu__##var, val)
+#define x86_or_percpu(var,val) percpu_to_op("or", per_cpu__##var, val)
 
 #endif /* __ARCH_I386_PERCPU__ */
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/processor.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/processor.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/processor.h	2006-09-19 14:55:56.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/processor.h	2006-09-19 15:26:28.000000000 +1000
@@ -473,7 +473,7 @@ struct thread_struct {
 	.vm86_info = NULL,						\
 	.sysenter_cs = __KERNEL_CS,					\
 	.io_bitmap_ptr = NULL,						\
-	.gs = __KERNEL_PDA,						\
+	.gs = __KERNEL_PERCPU,						\
 }
 
 /*
@@ -728,6 +728,5 @@ extern void select_idle_routine(const st
 extern unsigned long boot_option_idle_override;
 extern void enable_sep_cpu(void);
 extern int sysenter_setup(void);
-extern int alloc_gdt(int cpu);
 
 #endif /* __ASM_I386_PROCESSOR_H */
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/segment.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/segment.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/segment.h	2006-09-19 14:55:56.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/segment.h	2006-09-19 15:26:28.000000000 +1000
@@ -39,7 +39,7 @@
  *  25 - APM BIOS support 
  *
  *  26 - ESPFIX small SS
- *  27 - PDA				[ per-cpu private data area ]
+ *  27 - PERCPU				[ offset segment for per-cpu area ]
  *  28 - unused
  *  29 - unused
  *  30 - unused
@@ -74,8 +74,8 @@
 #define GDT_ENTRY_ESPFIX_SS		(GDT_ENTRY_KERNEL_BASE + 14)
 #define __ESPFIX_SS (GDT_ENTRY_ESPFIX_SS * 8)
 
-#define GDT_ENTRY_PDA			(GDT_ENTRY_KERNEL_BASE + 15)
-#define __KERNEL_PDA (GDT_ENTRY_PDA * 8)
+#define GDT_ENTRY_PERCPU		(GDT_ENTRY_KERNEL_BASE + 15)
+#define __KERNEL_PERCPU (GDT_ENTRY_PERCPU * 8)
 
 #define GDT_ENTRY_DOUBLEFAULT_TSS	31
 
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/smp.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/smp.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/smp.h	2006-09-19 14:55:56.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/smp.h	2006-09-19 15:27:59.000000000 +1000
@@ -8,7 +8,7 @@
 #include <linux/kernel.h>
 #include <linux/threads.h>
 #include <linux/cpumask.h>
-#include <asm/pda.h>
+#include <asm/percpu.h>
 #endif
 
 #ifdef CONFIG_X86_LOCAL_APIC
@@ -59,7 +59,8 @@ extern void cpu_uninit(void);
  * from the initial startup. We map APIC_BASE very early in page_setup(),
  * so this is correct in the x86 case.
  */
-#define raw_smp_processor_id() (read_pda(cpu_number))
+DECLARE_PER_CPU(unsigned int, processor_id);
+#define raw_smp_processor_id() (x86_read_percpu(processor_id))
 /* This is valid from the very earliest point in boot that we care
    about. */
 #define early_smp_processor_id() (current_thread_info()->cpu)
@@ -93,6 +94,8 @@ extern int __cpu_disable(void);
 extern void __cpu_die(unsigned int cpu);
 extern unsigned int num_processors;
 
+void setup_percpu_for_this_cpu(unsigned int cpu);
+
 #endif /* !__ASSEMBLY__ */
 
 #else /* CONFIG_SMP */
@@ -100,6 +103,7 @@ extern unsigned int num_processors;
 #define safe_smp_processor_id()		0
 #define cpu_physical_id(cpu)		boot_cpu_physical_apicid
 #define early_smp_processor_id()	0
+#define setup_percpu_for_this_cpu(cpu)
 
 #define NO_PROC_ID		0xFF		/* No processor magic marker */
 
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/dontdiff --minimal linux-2.6.18-rc6-mm2/include/asm-i386/unwind.h working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/unwind.h
--- linux-2.6.18-rc6-mm2/include/asm-i386/unwind.h	2006-09-19 14:55:56.000000000 +1000
+++ working-2.6.18-rc6-mm2-pda-to-percpu/include/asm-i386/unwind.h	2006-09-19 15:26:28.000000000 +1000
@@ -65,7 +65,7 @@ static inline void arch_unw_init_blocked
 	info->regs.xss = __KERNEL_DS;
 	info->regs.xds = __USER_DS;
 	info->regs.xes = __USER_DS;
-	info->regs.xgs = __KERNEL_PDA;
+	info->regs.xgs = __KERNEL_PERCPU;
 }
 
 extern asmlinkage int arch_unwind_init_running(struct unwind_frame_info *,

-- 
Help! Save Australia from the worst of the DMCA: http://linux.org.au/law

Re: Per-cpu patches on top of PDA stuff...

[Andi Kleen]

On Tuesday 19 September 2006 10:03, Rusty Russell wrote:
> On Tue, 2006-09-19 at 13:13 +1000, Rusty Russell wrote:
> > Hi Jeremy, all,
> > 
> > 	Sorry this took so long, spent last week in Japan at OSDL conf then
> > netconf.  After several false starts, I ended up with a very simple
> > implementation, which clashes significantly with your work since then
> > 8(.  I've pushed the patches anyway, but it's going to be significant
> > work for me to re-merge them, so I wanted your feedback first.
> 
> OK, here's a patch against 2.6.18-rc6-mm2.  Tested on UP and SMP.
> Crashes on hotplugging CPU, but crashes in same way as before the patch
> 8).

The incremental patches are madness and unmergeable. 
Can you please just do fresh patches with PDA stuff dropped first? 

-Andi

Re: Per-cpu patches on top of PDA stuff...

[Jeremy Fitzhardinge]

Rusty Russell wrote:
> This patch uses the "gs" segment register which Jeremy Fitzhardinge
> freed up for kernel use, for the per-cpu section.  This means that
> instead of having a special per-cpu struct which we can access in a
> single instruction, any per-cpu variable can be accessed in a single
> instruction.  In addition, it avoids introducing the concept of a
> "pda" into the kernel, in favour of the well-known "percpu" concept.
>   

The PDA is well established in the x86-64 tree; one of the reasons for 
having an x86-32 PDA was to make the trees a bit more similar.

> More cleanups/optimizations are possible:
> 1) Don't save/restore %gs on UP.  The cost is measurable, and we don't use it.
>   

You need to make sure the userspace %gs gets saved/restored for context 
switch.  If you don't do it in entry.S, you need to have two versions of 
the %gs-handling code in a number of places around the kernel 
(__switch_to, vm86, ptrace).

> 2) Remove early_smp_processor_id(), by setting up the per-cpu
>    processor_id field correctly before starting a CPU.
> 3) Similarly, get rid of early_current().
>   

I've already submitted a patch to do this.  In fact, it would help if 
you could rebase against my most recent patches (which I think Andrew 
has queued for the next -mm, and are in the paravirt patch queue).

Also, does this patch address the problem of percpu module variables?  
That was the only real sticking point with my previous percpu-in-pda patch.

I can't say this patch fills me with joy.

    J

Re: Per-cpu patches on top of PDA stuff...

[Jeremy Fitzhardinge]

Rusty Russell wrote:
> 	The first patch simply changes the GDTs to be a straight per-cpu
> variable.  I notice that you did a similar thing with your patches, but
> this is simpler and avoids wasting space in the UP case.  It's a bit
> tricky since we've never referred to per-cpu vars from asm before, but
> since we're only referring to the pre-setup versions, it's ok.
>   

The current mechanism was specifically introduced by James Bottomley a 
while back; I guess to deal with Voyager strangeness.

As far as setting up the PDA in head.S goes, it turns out to be very 
easy without having to access any per-cpu data, since the whole CPU 
bringup stuff depends on static variables anyway.

> 	The second patch changes gs to be the per-cpu offset, and by
> implication, avoids using it altogether on UP.  This avoids a special
> "pda" structure, instead allowing all per-cpu variables to be accessed
> this way.  It avoids __thread, which I gave up after creating a horribly
> complicated patch which still didn't quite work, and was no more
> efficient if we want the kernel to run under Xen anyway.
>
> I really think this is the way to go, and I'll start work on merging
> now.

Hm, now is not really a good time.  I'm still trying to get Xen 
basically working, and the percpu PDA stuff isn't really necessary for 
that.  The PDA stuff was enough of a problem in itself...

Also, the PDA patches are in -mm, so that's probably a better base for 
your patches.

    J

Re: Per-cpu patches on top of PDA stuff...

[Chris Wright]

* Jeremy Fitzhardinge (jeremy@goop.org) wrote:
> Rusty Russell wrote:
> > 	The first patch simply changes the GDTs to be a straight per-cpu
> > variable.  I notice that you did a similar thing with your patches, but
> > this is simpler and avoids wasting space in the UP case.  It's a bit
> > tricky since we've never referred to per-cpu vars from asm before, but
> > since we're only referring to the pre-setup versions, it's ok.
> 
> The current mechanism was specifically introduced by James Bottomley a 
> while back; I guess to deal with Voyager strangeness.

Yes, he reverted some changes because voyager boot cpu may not be cpu 0.

> As far as setting up the PDA in head.S goes, it turns out to be very 
> easy without having to access any per-cpu data, since the whole CPU 
> bringup stuff depends on static variables anyway.
> 
> > 	The second patch changes gs to be the per-cpu offset, and by
> > implication, avoids using it altogether on UP.  This avoids a special
> > "pda" structure, instead allowing all per-cpu variables to be accessed
> > this way.  It avoids __thread, which I gave up after creating a horribly
> > complicated patch which still didn't quite work, and was no more
> > efficient if we want the kernel to run under Xen anyway.
> >
> > I really think this is the way to go, and I'll start work on merging
> > now.
> 
> Hm, now is not really a good time.  I'm still trying to get Xen 
> basically working, and the percpu PDA stuff isn't really necessary for 
> that.  The PDA stuff was enough of a problem in itself...

I agree.  We're right in the middle of the last bit of Xen bring up.
PDA is not strictly needed, and destabilizing for this would be counter
productive.

thanks,
-chris

Re: Per-cpu patches on top of PDA stuff...

[Jeremy Fitzhardinge]

Chris Wright wrote:
> I agree.  We're right in the middle of the last bit of Xen bring up.
> PDA is not strictly needed, and destabilizing for this would be counter
> productive.
>   

Well, the basic PDA stuff is already in there, and seems OK now.  I 
don't want to revisit it until everything else looks sound though.

    J

Re: Per-cpu patches on top of PDA stuff...

[Rusty Russell]

On Tue, 2006-09-19 at 01:14 -0700, Jeremy Fitzhardinge wrote:
> Rusty Russell wrote:
> > 	The first patch simply changes the GDTs to be a straight per-cpu
> > variable.  I notice that you did a similar thing with your patches, but
> > this is simpler and avoids wasting space in the UP case.  It's a bit
> > tricky since we've never referred to per-cpu vars from asm before, but
> > since we're only referring to the pre-setup versions, it's ok.
> >   
> 
> The current mechanism was specifically introduced by James Bottomley a 
> while back; I guess to deal with Voyager strangeness.

(James CC'd).  I dislike the gratuitous copy: we have three gdts, the
boot GDT, the master GDT (cpu_gdt_table), then the per-cpu GDT.  Using
the per-cpu mechanisms already in place makes it simple, avoids manual
allocation, and the extra master GDT.  The extra GDT is particularly
embarrassing on UP, which doesn't want a per-cpu GDT anyway...

Seems that we can't assume boot CPU == 0.  I think I've removed the two
places where I assumed that, but will need testing.

> Hm, now is not really a good time.  I'm still trying to get Xen 
> basically working, and the percpu PDA stuff isn't really necessary for 
> that.  The PDA stuff was enough of a problem in itself...
> 
> Also, the PDA patches are in -mm, so that's probably a better base for 
> your patches.

Yes, it turned out to be easier to go straight to -mm anyway.  I'll redo
them as Andi requested....

Rusty.
-- 
Help! Save Australia from the worst of the DMCA: http://linux.org.au/law

Re: Per-cpu patches on top of PDA stuff...

[Andi Kleen]

> Seems that we can't assume boot CPU == 0.  I think I've removed the two
> places where I assumed that, but will need testing.

boot CPU == 0 should be true. You just can't assume anything about its
APIC ID, but we decouple APIC ID and logical processor id anyways.

Another issue is that with CPU hotunplug CPU #0 might disappear at some
point later.

-Andi

Re: Per-cpu patches on top of PDA stuff...

[James Bottomley]

On Wed, 2006-09-20 at 09:00 +0200, Andi Kleen wrote:
> boot CPU == 0 should be true. You just can't assume anything about its
> APIC ID, but we decouple APIC ID and logical processor id anyways.

No, it is not.  Voyager has fixed and immutable CPU IDs dependent on CPU
position in the system.  If CPU 0 is missing or logically deconfigured,
the boot CPU is definitely non zero.

James

Re: Per-cpu patches on top of PDA stuff...

[Andi Kleen]

On Wed, Sep 20, 2006 at 08:54:28AM -0400, James Bottomley wrote:
> On Wed, 2006-09-20 at 09:00 +0200, Andi Kleen wrote:
> > boot CPU == 0 should be true. You just can't assume anything about its
> > APIC ID, but we decouple APIC ID and logical processor id anyways.
> 
> No, it is not.  Voyager has fixed and immutable CPU IDs dependent on CPU
> position in the system.  If CPU 0 is missing or logically deconfigured,
> the boot CPU is definitely non zero.

As APIC ID (hard_smp_processor_id()) possible, but surely not as linux logical
 processor id (smp_processor_id()). That always starts with 0 and goes up.

-Andi

Re: Per-cpu patches on top of PDA stuff...

[James Bottomley]

On Wed, 2006-09-20 at 18:09 +0200, Andi Kleen wrote:
> As APIC ID (hard_smp_processor_id()) possible, but surely not as linux logical
>  processor id (smp_processor_id()). That always starts with 0 and goes up.

Voyager is not APIC based.  The SMP hal is constructed so there's no
such thing as processor translation.  hard and soft processor IDs return
the same thing.

James

Re: Per-cpu patches on top of PDA stuff...

[Andi Kleen]

On Wednesday 20 September 2006 18:15, James Bottomley wrote:
> On Wed, 2006-09-20 at 18:09 +0200, Andi Kleen wrote:
> > As APIC ID (hard_smp_processor_id()) possible, but surely not as linux logical
> >  processor id (smp_processor_id()). That always starts with 0 and goes up.
> 
> Voyager is not APIC based.  The SMP hal is constructed so there's no
> such thing as processor translation.  hard and soft processor IDs return
> the same thing.

Well that's your problem then. Just add the remapping array.

-Andi

Re: Per-cpu patches on top of PDA stuff...

[James Bottomley]

On Wed, 2006-09-20 at 18:22 +0200, Andi Kleen wrote:
> Well that's your problem then. Just add the remapping array.

I don't see why I should be the one to add unnecessary obfuscation.
Using the current ID scheme, the user always knows exactly where to find
CPU<n> in the chassis (and they are separately installable and
removable).  Why should I make up a new identity scheme that would have
no relation to the hardware?

James

Re: Per-cpu patches on top of PDA stuff...

[Andi Kleen]

On Wed, Sep 20, 2006 at 12:42:17PM -0400, James Bottomley wrote:
> On Wed, 2006-09-20 at 18:22 +0200, Andi Kleen wrote:
> > Well that's your problem then. Just add the remapping array.
> 
> I don't see why I should be the one to add unnecessary obfuscation.
> Using the current ID scheme, the user always knows exactly where to find
> CPU<n> in the chassis (and they are separately installable and
> removable).  Why should I make up a new identity scheme that would have
> no relation to the hardware?

Because it makes it easier to write other code? We don't really
want any unnecessary limiting assumptions in arch/i386 just because
of some obscure machine with one user.


-Andi

Re: Per-cpu patches on top of PDA stuff...

[James Bottomley]

On Wed, 2006-09-20 at 19:49 +0200, Andi Kleen wrote:
> Because it makes it easier to write other code? We don't really
> want any unnecessary limiting assumptions in arch/i386 just because
> of some obscure machine with one user.

Really?  I don't see it that way.  I did a lot of work back in 2000/2001
to break x86 of its remapped CPU assumptions ... it's been operating
nicely for 6 years, I don't see a reason to break it now.

Also, as we enter the era of hotplug CPUs, when a non voyager x86 system
tells you CPU3 is overheating, how to you find out exactly which CPU to
hot unplug?

James

Re: Per-cpu patches on top of PDA stuff...

[Andi Kleen]

On Wed, Sep 20, 2006 at 02:10:31PM -0400, James Bottomley wrote:
> On Wed, 2006-09-20 at 19:49 +0200, Andi Kleen wrote:
> > Because it makes it easier to write other code? We don't really
> > want any unnecessary limiting assumptions in arch/i386 just because
> > of some obscure machine with one user.
> 
> Really?  I don't see it that way.  I did a lot of work back in 2000/2001
> to break x86 of its remapped CPU assumptions ... it's been operating
> nicely for 6 years, I don't see a reason to break it now.

There are already CPU #0 assumptions in various paths, although
they might not affect you.

> 
> Also, as we enter the era of hotplug CPUs, when a non voyager x86 system
> tells you CPU3 is overheating, how to you find out exactly which CPU to
> hot unplug?

Right now you can't anyways, but at some point I would expect this
information to be in the SMBIOS (mapped from APIC-ID) and giving
some label that is printed on the motherboard.

We already support this for DIMMs BTW (although it doesn't work
everywhere due to very creative spec interpretation of some vendors) 

-Andi

Re: Per-cpu patches on top of PDA stuff...

[Rusty Russell]

On Wed, 2006-09-20 at 14:10 -0400, James Bottomley wrote:
> On Wed, 2006-09-20 at 19:49 +0200, Andi Kleen wrote:
> > Because it makes it easier to write other code? We don't really
> > want any unnecessary limiting assumptions in arch/i386 just because
> > of some obscure machine with one user.
> 
> Really?  I don't see it that way.  I did a lot of work back in 2000/2001
> to break x86 of its remapped CPU assumptions ... it's been operating
> nicely for 6 years, I don't see a reason to break it now.

I agree with James, it's just that I can't test it.  I'll send you my
patch set when I'm done, please test...

Thanks!
Rusty.
-- 
Help! Save Australia from the worst of the DMCA: http://linux.org.au/law

Re: Per-cpu patches on top of PDA stuff...

[Chris Wright]

* Rusty Russell (rusty@rustcorp.com.au) wrote:
> 8(.  I've pushed the patches anyway, but it's going to be significant
> work for me to re-merge them, so I wanted your feedback first.

Seems you or Jeremy pushed an unmerged head to the queue.  Please don't
do that, it makes life going forward a mess.  I'm merging it, will push
the result so best double check it.

thanks,
-chris

Re: Per-cpu patches on top of PDA stuff...

[Jeremy Fitzhardinge]

Chris Wright wrote:
> Seems you or Jeremy pushed an unmerged head to the queue.  Please don't
> do that, it makes life going forward a mess.  I'm merging it, will push
> the result so best double check it.
>   


I don't think we should merge it now.  There are enough other things 
broken before adding more complexity.

    J

Re: Per-cpu patches on top of PDA stuff...

[Chris Wright]

* Jeremy Fitzhardinge (jeremy@goop.org) wrote:
> Chris Wright wrote:
> > Seems you or Jeremy pushed an unmerged head to the queue.  Please don't
> > do that, it makes life going forward a mess.  I'm merging it, will push
> > the result so best double check it.
> 
> I don't think we should merge it now.  There are enough other things 
> broken before adding more complexity.

No, I just mean merge in the queue (which means I'll drop the series
changes that Rusty made, otherwise it should have zero overlap).
Otherwise mercurial gets unhappy with unmerged heads.

thanks,
-chris