Skip to contentSkip to conversation

x86 paravirtualization infrastructure

PATCH In review Chris Wright 13 participants 56 messages Oct 28, 2006 – Nov 18, 2006

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover x86 paravirtualization infrastructure

The following patches introduce the core infrastructure needed to paravirtualize the 32-bit x86 Linux kernel. This is done by moving virtualization sensitive insn's or code paths to a function table, paravirt_ops. This structure can be populated with hypervisor specific calls or native stubs and currently support running on bare metal, VMI, Xen, or Lhype. These patches apply to 2.6.19-rc2-mm2 plus the last set of paravirt prep patches that Rusty sent.

thanks,
-chris
--

#1 lore raw

PATCH 1/7 header and stubs for paravirtualizing critical operations

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

This patch does the dumbest possible replacement of paravirtualized
instructions: calls through a "paravirt_ops" structure. Currently these are function implementations of native hardware: hypervisors will override the ops structure with their own variants.

All the pv-ops functions are declared "fastcall" so that a specific register-based ABI is used, to make inlining assember easier.

Cc Jeremy Fitzhardinge <[email protected]>
Cc Zachary Amsden <[email protected]>
---
 arch/i386/Kconfig                          |   11 
 arch/i386/boot/compressed/misc.c           |    1 
 arch/i386/kernel/Makefile                  |    1 
 arch/i386/kernel/asm-offsets.c             |   10 
 arch/i386/kernel/entry.S                   |   34 +-
 arch/i386/kernel/i8259.c                   |    2 
 arch/i386/kernel/paravirt.c                |  410 +++++++++++++++++++++++++++++
 arch/i386/kernel/setup.c                   |    2 
 arch/i386/kernel/smpboot.c                 |    5 
 arch/i386/kernel/time.c                    |   15 -
 drivers/net/de600.c                        |    1 
 include/asm-i386/delay.h                   |    8 
 include/asm-i386/desc.h                    |    4 
 include/asm-i386/io.h                      |   10 
 include/asm-i386/irq.h                     |   10 
 include/asm-i386/irqflags.h                |   42 +-
 include/asm-i386/mach-default/setup_arch.h |    2 
 include/asm-i386/msr.h                     |    5 
 include/asm-i386/paravirt.h                |  291 ++++++++++++++++++++
 include/asm-i386/processor.h               |   15 -
 include/asm-i386/segment.h                 |    2 
 include/asm-i386/setup.h                   |    8 
 include/asm-i386/spinlock.h                |    4 
 include/asm-i386/system.h                  |   16 -
 include/asm-i386/time.h                    |   41 ++
 25 files changed, 902 insertions(+), 48 deletions(-)
arch/i386/Kconfig +11 -0
@@ -197,6 +197,17 @@ config X86_ES7000
 
 endchoice
 
+config PARAVIRT
+	bool "Paravirtualization support (EXPERIMENTAL)"
+	depends on EXPERIMENTAL
+	help
+	  Paravirtualization is a way of running multiple instances of
+	  Linux on the same machine, under a hypervisor.  This option
+	  changes the kernel so it can modify itself when it is run
+	  under a hypervisor, improving performance significantly.
+	  However, when run without a hypervisor the kernel is
+	  theoretically slower.  If in doubt, say N.
+
 config ACPI_SRAT
 	bool
 	default y
arch/i386/boot/compressed/misc.c +1 -0
@@ -9,6 +9,7 @@
  * High loaded stuff by Hans Lermen & Werner Almesberger, Feb. 1996
  */
 
+#undef CONFIG_PARAVIRT
 #include <linux/linkage.h>
 #include <linux/vmalloc.h>
 #include <linux/screen_info.h>
arch/i386/kernel/Makefile +1 -0
@@ -39,6 +39,7 @@ obj-$(CONFIG_VM86)		+= vm86.o
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 obj-$(CONFIG_HPET_TIMER) 	+= hpet.o
 obj-$(CONFIG_K8_NB)		+= k8.o
+obj-$(CONFIG_PARAVIRT)		+= paravirt.o
 
 EXTRA_AFLAGS   := -traditional
 
arch/i386/kernel/asm-offsets.c +10 -0
@@ -101,4 +101,14 @@ void foo(void)
 	BLANK();
  	OFFSET(PDA_cpu, i386_pda, cpu_number);
 	OFFSET(PDA_pcurrent, i386_pda, pcurrent);
+
+#ifdef CONFIG_PARAVIRT
+	BLANK();
+	OFFSET(PARAVIRT_enabled, paravirt_ops, paravirt_enabled);
+	OFFSET(PARAVIRT_irq_disable, paravirt_ops, irq_disable);
+	OFFSET(PARAVIRT_irq_enable, paravirt_ops, irq_enable);
+	OFFSET(PARAVIRT_irq_enable_sysexit, paravirt_ops, irq_enable_sysexit);
+	OFFSET(PARAVIRT_iret, paravirt_ops, iret);
+	OFFSET(PARAVIRT_read_cr0, paravirt_ops, read_cr0);
+#endif
 }
arch/i386/kernel/entry.S +27 -7
@@ -62,13 +62,6 @@ DF_MASK		= 0x00000400 
 NT_MASK		= 0x00004000
 VM_MASK		= 0x00020000
 
-/* These are replaces for paravirtualization */
-#define DISABLE_INTERRUPTS		cli
-#define ENABLE_INTERRUPTS		sti
-#define ENABLE_INTERRUPTS_SYSEXIT	sti; sysexit
-#define INTERRUPT_RETURN		iret
-#define GET_CR0_INTO_EAX		movl %cr0, %eax
-
 #ifdef CONFIG_PREEMPT
 #define preempt_stop		DISABLE_INTERRUPTS; TRACE_IRQS_OFF
 #else
@@ -416,6 +409,20 @@ ldt_ss:
 	jnz restore_nocheck
 	testl $0x00400000, %eax		# returning to 32bit stack?
 	jnz restore_nocheck		# allright, normal return
+
+#ifdef CONFIG_PARAVIRT
+	/* 
+	 * The kernel can't run on a non-flat stack if paravirt mode
+	 * is active.  Rather than try to fixup the high bits of
+	 * ESP, bypass this code entirely.  This may break DOSemu
+	 * and/or Wine support in a paravirt VM, although the option
+	 * is still available to implement the setting of the high
+	 * 16-bits in the INTERRUPT_RETURN paravirt-op.
+	 */
+	cmpl $0, paravirt_ops+PARAVIRT_enabled
+	jne restore_nocheck
+#endif
+	
 	/* If returning to userspace with 16bit stack,
 	 * try to fix the higher word of ESP, as the CPU
 	 * won't restore it.
@@ -831,6 +838,19 @@ nmi_espfix_stack:
 .previous
 KPROBE_END(nmi)
 
+#ifdef CONFIG_PARAVIRT
+ENTRY(native_iret)
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
+ENTRY(native_irq_enable_sysexit)
+	sti
+	sysexit
+#endif
+
 KPROBE_ENTRY(int3)
 	RING0_INT_FRAME
 	pushl $-1			# mark this as an int
arch/i386/kernel/i8259.c +1 -1
@@ -392,7 +392,7 @@ void __init init_ISA_irqs (void)
 	}
 }
 
-void __init init_IRQ(void)
+void __init native_init_IRQ(void)
 {
 	int i;
 
arch/i386/kernel/setup.c +1 -1
@@ -1404,7 +1404,7 @@ void __init setup_arch(char **cmdline_p)
 		efi_init();
 	else {
 		printk(KERN_INFO "BIOS-provided physical RAM map:\n");
-		print_memory_map(machine_specific_memory_setup());
+		print_memory_map(memory_setup());
 	}
 
 	copy_edd();
arch/i386/kernel/smpboot.c +5 -0
@@ -33,6 +33,11 @@
  *		Dave Jones	:	Report invalid combinations of Athlon CPUs.
 *		Rusty Russell	:	Hacked into shape for new "hotplug" boot process. */
 
+
+/* SMP boot always wants to use real time delay to allow sufficient time for
+ * the APs to come online */
+#define USE_REAL_TIME_DELAY
+
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/kernel.h>
arch/i386/kernel/time.c +5 -10
@@ -56,6 +56,7 @@
 #include <asm/uaccess.h>
 #include <asm/processor.h>
 #include <asm/timer.h>
+#include <asm/time.h>
 
 #include "mach_time.h"
 
@@ -116,10 +117,7 @@ static int set_rtc_mmss(unsigned long no
 	/* gets recalled with irq locally disabled */
 	/* XXX - does irqsave resolve this? -johnstul */
 	spin_lock_irqsave(&rtc_lock, flags);
-	if (efi_enabled)
-		retval = efi_set_rtc_mmss(nowtime);
-	else
-		retval = mach_set_rtc_mmss(nowtime);
+	retval = set_wallclock(nowtime);
 	spin_unlock_irqrestore(&rtc_lock, flags);
 
 	return retval;
@@ -211,10 +209,7 @@ unsigned long read_persistent_clock(void
 
 	spin_lock_irqsave(&rtc_lock, flags);
 
-	if (efi_enabled)
-		retval = efi_get_time();
-	else
-		retval = mach_get_cmos_time();
+	retval = get_wallclock();
 
 	spin_unlock_irqrestore(&rtc_lock, flags);
 
@@ -280,7 +275,7 @@ static void __init hpet_time_init(void)
 		printk("Using HPET for base-timer\n");
 	}
 
-	time_init_hook();
+	do_time_init();
 }
 #endif
 
@@ -296,5 +291,5 @@ void __init time_init(void)
 		return;
 	}
 #endif
-	time_init_hook();
+	do_time_init();
 }
drivers/net/de600.c +0 -1
@@ -43,7 +43,6 @@ static const char version[] = "de600.c: 
  * modify the following "#define": (see <asm/io.h> for more info)
 #define REALLY_SLOW_IO
  */
-#define SLOW_IO_BY_JUMPING /* Looks "better" than dummy write to port 0x80 :-) */
 
 /* use 0 for production, 1 for verification, >2 for debug */
 #ifdef DE600_DEBUG
include/asm-i386/delay.h +8 -0
@@ -15,6 +15,13 @@ extern void __ndelay(unsigned long nsecs
 extern void __const_udelay(unsigned long usecs);
 extern void __delay(unsigned long loops);
 
+#if defined(CONFIG_PARAVIRT) && !defined(USE_REAL_TIME_DELAY)
+#define udelay(n) paravirt_ops.const_udelay((n) * 0x10c7ul)
+	
+#define ndelay(n) paravirt_ops.const_udelay((n) * 5ul)
+
+#else /* !PARAVIRT || USE_REAL_TIME_DELAY */
+
 #define udelay(n) (__builtin_constant_p(n) ? \
 	((n) > 20000 ? __bad_udelay() : __const_udelay((n) * 0x10c7ul)) : \
 	__udelay(n))
@@ -22,6 +29,7 @@ extern void __delay(unsigned long loops)
 #define ndelay(n) (__builtin_constant_p(n) ? \
 	((n) > 20000 ? __bad_ndelay() : __const_udelay((n) * 5ul)) : \
 	__ndelay(n))
+#endif
 
 void use_tsc_delay(void);
 
include/asm-i386/desc.h +4 -0
@@ -55,6 +55,9 @@ static inline void pack_gate(u32 *low, u
 #define DESCTYPE_DPL3	0x60	/* DPL-3 */
 #define DESCTYPE_S	0x10	/* !system */
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 #define load_TR_desc() __asm__ __volatile__("ltr %w0"::"q" (GDT_ENTRY_TSS*8))
 
 #define load_gdt(dtr) __asm__ __volatile("lgdt %0"::"m" (*dtr))
@@ -104,6 +107,7 @@ static inline void set_ldt(void *addr, u
 		__asm__ __volatile__("lldt %w0"::"q" (GDT_ENTRY_LDT*8));
 	}
 }
+#endif /* CONFIG_PARAVIRT */
 
 static inline void _set_gate(int gate, unsigned int type, void *addr, unsigned short seg)
 {
include/asm-i386/io.h +6 -4
@@ -256,11 +256,11 @@ static inline void flush_write_buffers(v
 
 #endif /* __KERNEL__ */
 
-#ifdef SLOW_IO_BY_JUMPING
-#define __SLOW_DOWN_IO "jmp 1f; 1: jmp 1f; 1:"
-#else
+#if defined(CONFIG_PARAVIRT)
+#include <asm/paravirt.h>
+#else 
+
 #define __SLOW_DOWN_IO "outb %%al,$0x80;"
-#endif
 
 static inline void slow_down_io(void) {
 	__asm__ __volatile__(
@@ -271,6 +271,8 @@ static inline void slow_down_io(void) {
 		: : );
 }
 
+#endif
+
 #ifdef CONFIG_X86_NUMAQ
 extern void *xquad_portio;    /* Where the IO area was mapped */
 #define XQUAD_PORT_ADDR(port, quad) (xquad_portio + (XQUAD_PORTIO_QUAD*quad) + port)
include/asm-i386/irq.h +10 -0
@@ -41,4 +41,14 @@ extern int irqbalance_disable(char *str)
 extern void fixup_irqs(cpumask_t map);
 #endif
 
+void __init native_init_IRQ(void);
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+static inline void init_IRQ(void)
+{
+	native_init_IRQ();
+}
+#endif /* CONFIG_PARAVIRT */
+
 #endif /* _ASM_IRQ_H */
include/asm-i386/irqflags.h +27 -15

69 lines not on this page · raw message

include/asm-i386/mach-default/setup_arch.h +2 -0
@@ -2,4 +2,6 @@
 
 /* no action for generic */
 
+#ifndef ARCH_SETUP
 #define ARCH_SETUP
+#endif
include/asm-i386/msr.h +5 -0
@@ -1,6 +1,10 @@
 #ifndef __ASM_MSR_H
 #define __ASM_MSR_H
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+
 /*
  * Access to machine-specific registers (available on 586 and better only)
  * Note: the rd* operations modify the parameters directly (without using
@@ -77,6 +81,7 @@ static inline void wrmsrl (unsigned long
      __asm__ __volatile__("rdpmc" \
 			  : "=a" (low), "=d" (high) \
 			  : "c" (counter))
+#endif	/* !CONFIG_PARAVIRT */
 
 /* symbolic names for some interesting MSRs */
 /* Intel defined MSRs. */
include/asm-i386/processor.h +12 -3
@@ -146,8 +146,8 @@ static inline void detect_ht(struct cpui
 #define X86_EFLAGS_VIP	0x00100000 /* Virtual Interrupt Pending */
 #define X86_EFLAGS_ID	0x00200000 /* CPUID detection flag */
 
-static inline void __cpuid(unsigned int *eax, unsigned int *ebx,
-			   unsigned int *ecx, unsigned int *edx)
+static inline fastcall void native_cpuid(unsigned int *eax, unsigned int *ebx,
+					 unsigned int *ecx, unsigned int *edx)
 {
 	/* ecx is often an input as well as an output. */
 	__asm__("cpuid"
@@ -548,6 +548,12 @@ static inline void rep_nop(void)
 
 #define cpu_relax()	rep_nop()
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+#define paravirt_enabled() 0
+#define __cpuid native_cpuid
+
 static inline void load_esp0(struct tss_struct *tss, struct thread_struct *thread)
 {
 	tss->esp0 = thread->esp0;
@@ -570,10 +576,13 @@ static inline void load_esp0(struct tss_
 			: /* no output */			\
 			:"r" (value))
 
+#define set_iopl_mask native_set_iopl_mask
+#endif /* CONFIG_PARAVIRT */
+
 /*
  * Set IOPL bits in EFLAGS from given mask
  */
-static inline void set_iopl_mask(unsigned mask)
+static fastcall inline void native_set_iopl_mask(unsigned mask)
 {
 	unsigned int reg;
 	__asm__ __volatile__ ("pushfl;"
include/asm-i386/segment.h +2 -0
@@ -131,5 +131,7 @@
 #define SEGMENT_LDT		0x4
 #define SEGMENT_GDT		0x0
 
+#ifndef CONFIG_PARAVIRT
 #define get_kernel_rpl()  0
 #endif
+#endif
include/asm-i386/setup.h +8 -0
@@ -70,6 +70,14 @@ extern unsigned char boot_params[PARAM_S
 struct e820entry;
 
 char * __init machine_specific_memory_setup(void);
+#ifndef CONFIG_PARAVIRT
+static inline char *memory_setup(void)
+{
+	return machine_specific_memory_setup();
+}
+#else
+#include <asm/paravirt.h>
+#endif
 
 int __init copy_e820_map(struct e820entry * biosmap, int nr_map);
 int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map);
include/asm-i386/spinlock.h +4 -0
@@ -7,8 +7,12 @@
 #include <asm/processor.h>
 #include <linux/compiler.h>
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 #define CLI_STRING	"cli"
 #define STI_STRING	"sti"
+#endif /* CONFIG_PARAVIRT */
 
 /*
  * Your basic SMP spinlocks, allowing only a single CPU anywhere
include/asm-i386/system.h +10 -6
@@ -88,6 +88,9 @@ __asm__ __volatile__ ("movw %%dx,%1\n\t"
 #define savesegment(seg, value) \
 	asm volatile("mov %%" #seg ",%0":"=rm" (value))
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 #define read_cr0() ({ \
 	unsigned int __dummy; \
 	__asm__ __volatile__( \
@@ -139,17 +142,18 @@ __asm__ __volatile__ ("movw %%dx,%1\n\t"
 #define write_cr4(x) \
 	__asm__ __volatile__("movl %0,%%cr4": :"r" (x))
 
-/*
- * Clear and set 'TS' bit respectively
- */
+#define wbinvd() \
+	__asm__ __volatile__ ("wbinvd": : :"memory")
+
+/* Clear the 'TS' bit */
 #define clts() __asm__ __volatile__ ("clts")
+#endif/* CONFIG_PARAVIRT */
+
+/* Set the 'TS' bit */
 #define stts() write_cr0(8 | read_cr0())
 
 #endif	/* __KERNEL__ */
 
-#define wbinvd() \
-	__asm__ __volatile__ ("wbinvd": : :"memory")
-
 static inline unsigned long get_limit(unsigned long segment)
 {
 	unsigned long __limit;
arch/i386/kernel/paravirt.c new file +410 -0

410 lines not on this page · raw message

include/asm-i386/paravirt.h new file +291 -0

291 lines not on this page · raw message

include/asm-i386/time.h new file +41 -0

41 lines not on this page · raw message

--

#2 lore raw

Re: [PATCH 1/7] header and stubs for paravirtualizing critical operations

Can you add a high level comment to entry.S what paravirt mode is all about and perhaps a quick cheat sheet on the macros?

Quoted 3 lines

+/* SMP boot always wants to use real time delay to allow sufficient time for + * the APs to come online */
+#define USE_REAL_TIME_DELAY

That's ugly. Can't you call different wait functions for that case instead?

include/asm-i386/irq.h extern int irqbalance_disable(char *str) · line 52

Quoted 8 lines
include/asm-i386/irq.h +8 -0
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+static inline void init_IRQ(void)
+{
+	native_init_IRQ();
+}
+#endif /* CONFIG_PARAVIRT */

You could probably avoid a lot of ifdefs by strategic use of __attribute__((weak))

include/asm-i386/desc.h static inline void pack_gate(u32 *low, u · line 60

Quoted 3 lines
include/asm-i386/desc.h +3 -0
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else

This is probably a good candidate for rename to native + wrapper macros too. Otherwise we'll always have to hack two different places later.

-Andi

#3 lore raw

PATCH 2/7 Patch inline replacements for common paravirt operations.

In reply to Chris Wright

It turns out that the most called ops, by several orders of magnitude, are the interrupt manipulation ops. These are obvious candidates for patching, so mark them up and create infrastructure for it.

The method used is that the ops structure has a patch function, which is called for each place which needs to be patched: this returns a number of instructions (the rest are NOP-padded).

Usually we can spare a register (%eax) for the binary patched code to use, but in a couple of critical places in entry.S we can't: we make the clobbers explicit at the call site, and manually clobber the allowed registers in debug mode as an extra check.

---
 arch/i386/kernel/alternative.c |   66 ++++++++++++--
 arch/i386/kernel/entry.S       |   26 ++---
 arch/i386/kernel/module.c      |    9 +
 arch/i386/kernel/paravirt.c    |   44 +++++++++
 arch/i386/kernel/vmlinux.lds.S |    6 +
 include/asm-i386/alternative.h |    3 
 include/asm-i386/irqflags.h    |    4 
 include/asm-i386/paravirt.h    |  185 +++++++++++++++++++++++++++++++++--------
 include/asm-i386/spinlock.h    |    5 -
 9 files changed, 284 insertions(+), 64 deletions(-)
arch/i386/kernel/alternative.c +55 -11
@@ -123,6 +123,20 @@ static unsigned char** find_nop_table(vo
 
 #endif /* CONFIG_X86_64 */
 
+static void nop_out(void *insns, unsigned int len)
+{
+	unsigned char **noptable = find_nop_table();
+
+	while (len > 0) {
+		unsigned int noplen = len;
+		if (noplen > ASM_NOP_MAX)
+			noplen = ASM_NOP_MAX;
+		memcpy(insns, noptable[noplen], noplen);
+		insns += noplen;
+		len -= noplen;
+	}
+}
+
 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
 extern struct alt_instr __smp_alt_instructions[], __smp_alt_instructions_end[];
 extern u8 *__smp_locks[], *__smp_locks_end[];
@@ -137,10 +151,9 @@ extern u8 __smp_alt_begin[], __smp_alt_e
 
 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
 {
-	unsigned char **noptable = find_nop_table();
 	struct alt_instr *a;
 	u8 *instr;
-	int diff, i, k;
+	int diff;
 
 	DPRINTK("%s: alt table %p -> %p\n", __FUNCTION__, start, end);
 	for (a = start; a < end; a++) {
@@ -158,13 +171,7 @@ void apply_alternatives(struct alt_instr
 #endif
 		memcpy(instr, a->replacement, a->replacementlen);
 		diff = a->instrlen - a->replacementlen;
-		/* Pad the rest with nops */
-		for (i = a->replacementlen; diff > 0; diff -= k, i += k) {
-			k = diff;
-			if (k > ASM_NOP_MAX)
-				k = ASM_NOP_MAX;
-			memcpy(a->instr + i, noptable[k], k);
-		}
+		nop_out(instr + a->replacementlen, diff);
 	}
 }
 
@@ -208,7 +215,6 @@ static void alternatives_smp_lock(u8 **s
 
 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 {
-	unsigned char **noptable = find_nop_table();
 	u8 **ptr;
 
 	for (ptr = start; ptr < end; ptr++) {
@@ -216,7 +222,7 @@ static void alternatives_smp_unlock(u8 *
 			continue;
 		if (*ptr > text_end)
 			continue;
-		**ptr = noptable[1][0];
+		nop_out(*ptr, 1);
 	};
 }
 
@@ -342,6 +348,43 @@ void alternatives_smp_switch(int smp)
 
 #endif
 
+#ifdef CONFIG_PARAVIRT
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end)
+{
+	struct paravirt_patch *p;
+	int i;
+
+	for (p = start; p < end; p++) {
+		unsigned int used;
+
+		used = paravirt_ops.patch(p->instrtype, p->clobbers, p->instr,
+					  p->len);
+#ifdef CONFIG_DEBUG_KERNEL
+		/* Deliberately clobber regs using "not %reg" to find bugs. */
+		for (i = 0; i < 3; i++) {
+			if (p->len - used >= 2 && (p->clobbers & (1 << i))) {
+				memcpy(p->instr + used, "\xf7\xd0", 2);
+				p->instr[used+1] |= i;
+				used += 2;
+			}
+		}
+#endif
+		/* Pad the rest with nops */
+		nop_out(p->instr + used, p->len - used);
+	}
+
+	/* Sync to be conservative, in case we patched following instructions */
+	sync_core();
+}
+extern struct paravirt_patch __start_parainstructions[],
+	__stop_parainstructions[];
+#else
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end)
+{
+}
+extern struct paravirt_patch *__start_parainstructions, *__stop_parainstructions;
+#endif	/* CONFIG_PARAVIRT */
+
 void __init alternative_instructions(void)
 {
 	unsigned long flags;
@@ -389,5 +432,6 @@ void __init alternative_instructions(voi
 		alternatives_smp_switch(0);
 	}
 #endif
+ 	apply_paravirt(__start_parainstructions, __stop_parainstructions);
 	local_irq_restore(flags);
 }
arch/i386/kernel/entry.S +13 -13
@@ -63,9 +63,9 @@ NT_MASK		= 0x00004000
 VM_MASK		= 0x00020000
 
 #ifdef CONFIG_PREEMPT
-#define preempt_stop		DISABLE_INTERRUPTS; TRACE_IRQS_OFF
+#define preempt_stop(clobbers)	DISABLE_INTERRUPTS(clobbers); TRACE_IRQS_OFF
 #else
-#define preempt_stop
+#define preempt_stop(clobbers)
 #define resume_kernel		restore_nocheck
 #endif
 
@@ -226,7 +226,7 @@ ENTRY(ret_from_fork)
 	ALIGN
 	RING0_PTREGS_FRAME
 ret_from_exception:
-	preempt_stop
+	preempt_stop(CLBR_ANY)
 ret_from_intr:
 	GET_THREAD_INFO(%ebp)
 check_userspace:
@@ -237,7 +237,7 @@ check_userspace:
 	jb resume_kernel		# not returning to v8086 or userspace
 
 ENTRY(resume_userspace)
- 	DISABLE_INTERRUPTS		# make sure we don't miss an interrupt
+ 	DISABLE_INTERRUPTS(CLBR_ANY)	# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
 					# between sampling and the iret
 	movl TI_flags(%ebp), %ecx
@@ -248,7 +248,7 @@ ENTRY(resume_userspace)
 
 #ifdef CONFIG_PREEMPT
 ENTRY(resume_kernel)
-	DISABLE_INTERRUPTS
+	DISABLE_INTERRUPTS(CLBR_ANY)
 	cmpl $0,TI_preempt_count(%ebp)	# non-zero preempt_count ?
 	jnz restore_nocheck
 need_resched:
@@ -277,7 +277,7 @@ sysenter_past_esp:
 	 * No need to follow this irqs on/off section: the syscall
 	 * disabled irqs and here we enable it straight after entry:
 	 */
-	ENABLE_INTERRUPTS
+	ENABLE_INTERRUPTS(CLBR_NONE)
 	pushl $(__USER_DS)
 	CFI_ADJUST_CFA_OFFSET 4
 	/*CFI_REL_OFFSET ss, 0*/
@@ -322,7 +322,7 @@ sysenter_past_esp:
 	jae syscall_badsys
 	call *sys_call_table(,%eax,4)
 	movl %eax,PT_EAX(%esp)
-	DISABLE_INTERRUPTS
+	DISABLE_INTERRUPTS(CLBR_ECX|CLBR_EDX)
 	TRACE_IRQS_OFF
 	movl TI_flags(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx
@@ -364,7 +364,7 @@ syscall_call:
 	call *sys_call_table(,%eax,4)
 	movl %eax,PT_EAX(%esp)		# store the return value
 syscall_exit:
-	DISABLE_INTERRUPTS		# make sure we don't miss an interrupt
+	DISABLE_INTERRUPTS(CLBR_ANY)	# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
 					# between sampling and the iret
 	TRACE_IRQS_OFF
@@ -393,7 +393,7 @@ restore_nocheck_notrace:
 .section .fixup,"ax"
 iret_exc:
 	TRACE_IRQS_ON
-	ENABLE_INTERRUPTS
+	ENABLE_INTERRUPTS(CLBR_NONE)
 	pushl $0			# no error code
 	pushl $do_iret_error
 	jmp error_code
@@ -436,7 +436,7 @@ ldt_ss:
 	CFI_ADJUST_CFA_OFFSET 4
 	pushl %eax
 	CFI_ADJUST_CFA_OFFSET 4
-	DISABLE_INTERRUPTS
+	DISABLE_INTERRUPTS(CLBR_EAX)
 	TRACE_IRQS_OFF
 	lss (%esp), %esp
 	CFI_ADJUST_CFA_OFFSET -8
@@ -451,7 +451,7 @@ work_pending:
 	jz work_notifysig
 work_resched:
 	call schedule
-	DISABLE_INTERRUPTS		# make sure we don't miss an interrupt
+	DISABLE_INTERRUPTS(CLBR_ANY)	# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
 					# between sampling and the iret
 	TRACE_IRQS_OFF
@@ -507,7 +507,7 @@ syscall_exit_work:
 	testb $(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SINGLESTEP), %cl
 	jz work_pending
 	TRACE_IRQS_ON
-	ENABLE_INTERRUPTS		# could let do_syscall_trace() call
+	ENABLE_INTERRUPTS(CLBR_ANY)	# could let do_syscall_trace() call
 					# schedule() instead
 	movl %esp, %eax
 	movl $1, %edx
@@ -691,7 +691,7 @@ ENTRY(device_not_available)
 	GET_CR0_INTO_EAX
 	testl $0x4, %eax		# EM (math emulation bit)
 	jne device_not_available_emulate
-	preempt_stop
+	preempt_stop(CLBR_ANY)
 	call math_state_restore
 	jmp ret_from_exception
 device_not_available_emulate:
arch/i386/kernel/module.c +8 -1
@@ -109,7 +109,8 @@ int module_finalize(const Elf_Ehdr *hdr,
 		    const Elf_Shdr *sechdrs,
 		    struct module *me)
 {
-	const Elf_Shdr *s, *text = NULL, *alt = NULL, *locks = NULL;
+	const Elf_Shdr *s, *text = NULL, *alt = NULL, *locks = NULL,
+		*para = NULL;
 	char *secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
 
 	for (s = sechdrs; s < sechdrs + hdr->e_shnum; s++) { 
@@ -119,6 +120,8 @@ int module_finalize(const Elf_Ehdr *hdr,
 			alt = s;
 		if (!strcmp(".smp_locks", secstrings + s->sh_name))
 			locks= s;
+		if (!strcmp(".parainstructions", secstrings + s->sh_name))
+			para = s;
 	}
 
 	if (alt) {
@@ -133,6 +136,10 @@ int module_finalize(const Elf_Ehdr *hdr,
 					    lseg, lseg + locks->sh_size,
 					    tseg, tseg + text->sh_size);
 	}
+	if (para) {
+		void *pseg = (void *)para->sh_addr;
+		apply_paravirt(pseg, pseg + para->sh_size);
+	}
 
 	return module_bug_finalize(hdr, sechdrs, me);
 }
arch/i386/kernel/paravirt.c +44 -0

56 lines not on this page · raw message

arch/i386/kernel/vmlinux.lds.S +6 -0
@@ -154,6 +154,12 @@ SECTIONS
   .altinstr_replacement : AT(ADDR(.altinstr_replacement) - LOAD_OFFSET) {
 	*(.altinstr_replacement)
   }
+  . = ALIGN(4);
+  __start_parainstructions = .;
+  .parainstructions : AT(ADDR(.parainstructions) - LOAD_OFFSET) {
+	*(.parainstructions)
+  }
+  __stop_parainstructions = .;
   /* .exit.text is discard at runtime, not link time, to deal with references
      from .altinstructions and .eh_frame */
   .exit.text : AT(ADDR(.exit.text) - LOAD_OFFSET) { *(.exit.text) }
include/asm-i386/alternative.h +3 -0
@@ -118,4 +118,7 @@ static inline void alternatives_smp_swit
 #define LOCK_PREFIX ""
 #endif
 
+struct paravirt_patch;
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end);
+
 #endif /* _I386_ALTERNATIVE_H */
include/asm-i386/irqflags.h +2 -2
@@ -79,8 +79,8 @@ static inline unsigned long __raw_local_
 }
 
 #else
-#define DISABLE_INTERRUPTS		cli
-#define ENABLE_INTERRUPTS		sti
+#define DISABLE_INTERRUPTS(clobbers)	cli
+#define ENABLE_INTERRUPTS(clobbers)	sti
 #define ENABLE_INTERRUPTS_SYSEXIT	sti; sysexit
 #define INTERRUPT_RETURN		iret
 #define GET_CR0_INTO_EAX		movl %cr0, %eax
include/asm-i386/paravirt.h +149 -36

213 lines not on this page · raw message

include/asm-i386/spinlock.h +4 -1
@@ -12,6 +12,7 @@
 #else
 #define CLI_STRING	"cli"
 #define STI_STRING	"sti"
+#define CLI_STI_CLOBBERS
 #endif /* CONFIG_PARAVIRT */
 
 /*
@@ -75,7 +76,9 @@ static inline void __raw_spin_lock_flags
 		"jg 1b\n\t"
 		"jmp 4b\n"
 		"5:\n\t"
-		: "+m" (lock->slock) : "r" (flags) : "memory");
+		: "+m" (lock->slock)
+		: "r" (flags)
+		: "memory" CLI_STI_CLOBBERS);
 }
 #endif
 

--

#4 lore raw

PATCH 3/7 More generic paravirtualization entry point.

In reply to Chris Wright

1) Each hypervisor writes a probe function to detect whether we are running under that hypervisor. paravirt_probe() registers this function.

2) If vmlinux is booted with ring != 0, we call all the probe functions (with registers except %esp intact) in link order: the winner will not return.

Cc Jeremy Fitzhardinge <[email protected]>
Cc Zachary Amsden <[email protected]>
---
 arch/i386/kernel/Makefile      |    2 ++
 arch/i386/kernel/head.S        |   33 +++++++++++++++++++++++++++++++++
 arch/i386/kernel/paravirt.c    |    6 +++++-
 arch/i386/kernel/vmlinux.lds.S |    6 ++++++
 include/asm-i386/paravirt.h    |    5 +++++
 5 files changed, 51 insertions(+), 1 deletion(-)
arch/i386/kernel/Makefile +2 -0
@@ -39,6 +39,8 @@ obj-$(CONFIG_VM86)		+= vm86.o
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 obj-$(CONFIG_HPET_TIMER) 	+= hpet.o
 obj-$(CONFIG_K8_NB)		+= k8.o
+
+# Make sure this is linked after any other paravirt_ops structs: see head.S
 obj-$(CONFIG_PARAVIRT)		+= paravirt.o
 
 EXTRA_AFLAGS   := -traditional
arch/i386/kernel/head.S +33 -0
@@ -55,6 +55,12 @@
  */
 ENTRY(startup_32)
 
+#ifdef CONFIG_PARAVIRT
+        movl %cs, %eax
+        testl $0x3, %eax
+        jnz startup_paravirt
+#endif
+
 /*
  * Set segments to known values.
  */
@@ -486,6 +492,33 @@ ignore_int:
 #endif
 	iret
 
+#ifdef CONFIG_PARAVIRT
+startup_paravirt:
+	cld
+ 	movl $(init_thread_union+THREAD_SIZE),%esp
+
+	/* We take pains to preserve all the regs. */
+	pushl	%edx
+	pushl	%ecx
+	pushl	%eax
+
+	/* paravirt.o is last in link, and that probe fn never returns */
+	pushl	$__start_paravirtprobe
+1:
+	movl	0(%esp), %eax
+	pushl	(%eax)
+	movl	8(%esp), %eax
+	call	*(%esp)
+	popl	%eax
+
+	movl	4(%esp), %eax
+	movl	8(%esp), %ecx
+	movl	12(%esp), %edx
+
+	addl	$4, (%esp)
+	jmp	1b
+#endif
+
 /*
  * Real beginning of normal "text" segment
  */
arch/i386/kernel/paravirt.c +5 -1
@@ -392,7 +392,11 @@ static int __init print_banner(void)
 	return 0;
 }
 core_initcall(print_banner);
- 
+
+/* We simply declare start_kernel to be the paravirt probe of last resort. */
+asmlinkage void __init start_kernel(void);
+paravirt_probe(start_kernel);
+  
 struct paravirt_ops paravirt_ops = {
 	.name = "bare hardware",
 	.paravirt_enabled = 0,
arch/i386/kernel/vmlinux.lds.S +6 -0
@@ -60,6 +60,12 @@ SECTIONS
 	CONSTRUCTORS
 	} :data
 
+  __start_paravirtprobe = .;
+  .paravirtprobe : AT(ADDR(.paravirtprobe) - LOAD_OFFSET) {
+	*(.paravirtprobe)
+  }
+  __stop_paravirtprobe = .;
+
   . = ALIGN(4096);
   __nosave_begin = .;
   .data_nosave : AT(ADDR(.data_nosave) - LOAD_OFFSET) { *(.data.nosave) }
include/asm-i386/paravirt.h +5 -0
@@ -120,6 +120,11 @@ struct paravirt_ops
 	void (fastcall *iret)(void);
 };
 
+/* Mark a paravirt probe function. */
+#define paravirt_probe(fn)						\
+	static void (*__paravirtprobe_##fn)(void) __attribute_used__	\
+		__attribute__((__section__(".paravirtprobe"))) = fn
+
 extern struct paravirt_ops paravirt_ops;
 
 #define paravirt_enabled() (paravirt_ops.paravirt_enabled)

--

#5 lore raw

Re: [PATCH 3/7] More generic paravirtualization entry point.

Quoted 3 lines · Chris Wright
(quoted diff) +2 -0
+
+/* We simply declare start_kernel to be the paravirt probe of last resort.

*/ +asmlinkage void __init start_kernel(void);

Didn't rusty put that into a header recently?

-Andi

#6 lore raw

PATCH 4/7 Allow selected bug checks to be skipped by paravirt kernels

In reply to Chris Wright

Allow selected bug checks to be skipped by paravirt kernels. The two most important are the F00F workaround (which is either done by the hypervisor, or not required), and the 'hlt' instruction check, which can break under some hypervisors.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
---
 arch/i386/kernel/cpu/intel.c |    2 +-
 include/asm-i386/bugs.h      |    4 ++++
 2 files changed, 5 insertions(+), 1 deletion(-)
arch/i386/kernel/cpu/intel.c +1 -1
@@ -107,7 +107,7 @@ static void __cpuinit init_intel(struct 
 	 * Note that the workaround only should be initialized once...
 	 */
 	c->f00f_bug = 0;
-	if ( c->x86 == 5 ) {
+	if (!paravirt_enabled() && c->x86 == 5) {
 		static int f00f_workaround_enabled = 0;
 
 		c->f00f_bug = 1;
include/asm-i386/bugs.h +4 -0
@@ -21,6 +21,7 @@
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/msr.h>
+#include <asm/paravirt.h>
 
 static int __init no_halt(char *s)
 {
@@ -91,6 +92,9 @@ static void __init check_fpu(void)
 
 static void __init check_hlt(void)
 {
+	if (paravirt_enabled())
+		return;
+
 	printk(KERN_INFO "Checking 'hlt' instruction... ");
 	if (!boot_cpu_data.hlt_works_ok) {
 		printk("disabled\n");

--

#7 lore raw

Re: [PATCH 4/7] Allow selected bug checks to be skipped by paravirt kernels

Quoted 4 lines · Chris Wright

Allow selected bug checks to be skipped by paravirt kernels. The two most important are the F00F workaround (which is either done by the hypervisor, or not required), and the 'hlt' instruction check, which can break under some hypervisors.

How can hlt check break? It is hlt;hlt;hlt, IIRC, that looks fairly innocent to me.

arch/i386/kernel/cpu/intel.c static void __cpuinit init_intel(struct · line 110

Quoted 8 lines
arch/i386/kernel/cpu/intel.c +1 -1
@@ -107,7 +107,7 @@ static void __cpuinit init_intel(struct 
 	 * Note that the workaround only should be initialized once...
 	 */
 	c->f00f_bug = 0;
-	if ( c->x86 == 5 ) {
+	if (!paravirt_enabled() && c->x86 == 5) {

I'd do x86==5 check first... pentiums are not common any more.

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#8 lore raw

Re: [PATCH 4/7] Allow selected bug checks to be skipped by paravirt kernels

Quoted 9 lines · Pavel Machek

arch/i386/kernel/cpu/intel.c static void __cpuinit init_intel(struct · line 110

Quoted 7 lines

+++ linux-2.6-pv/arch/i386/kernel/cpu/intel.c

arch/i386/kernel/cpu/intel.c +1 -1
@@ -107,7 +107,7 @@ static void __cpuinit init_intel(struct 
 	 * Note that the workaround only should be initialized once...
 	 */
 	c->f00f_bug = 0;
-	if ( c->x86 == 5 ) {
+	if (!paravirt_enabled() && c->x86 == 5) {

I'd do x86==5 check first... pentiums are not common any more.

It's not like paravirt_enabled will be common-case either, and is this isn't exactly a performance critical piece of code, it doesn't really matter which way around the checks are done.

Dave

Signature · 1 line
http://www.codemonkey.org.uk
#9 lore raw

Re: [PATCH 4/7] Allow selected bug checks to be skipped by paravirt kernels

In reply to Pavel Machek
Quoted 11 lines · Pavel Machek
Quoted 5 lines · Chris Wright

Allow selected bug checks to be skipped by paravirt kernels. The two most important are the F00F workaround (which is either done by the hypervisor, or not required), and the 'hlt' instruction check, which can break under some hypervisors.

How can hlt check break? It is hlt;hlt;hlt, IIRC, that looks fairly innocent to me.

Not if you use tickless timers that don't generate interrupts to unhalt you, or if you delay ticks until the next scheduled timeout and you haven't yet scheduled any timeout. Both are likely in a hypervisor.

Zach

#10 lore raw

Re: [PATCH 4/7] Allow selected bug checks to be skipped by paravirt kernels

Hi!

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

Allow selected bug checks to be skipped by paravirt kernels. The two most important are the F00F workaround (which is either done by the hypervisor, or not required), and the 'hlt' instruction check, which can break under some hypervisors.

Quoted 7 lines
Quoted 3 lines

How can hlt check break? It is hlt;hlt;hlt, IIRC, that looks fairly innocent to me.

Not if you use tickless timers that don't generate interrupts to unhalt you, or if you delay ticks until the next scheduled timeout and you haven't yet scheduled any timeout. Both are likely in a hypervisor.

Well.. but you are working around problem, instead of fixing it.

Tickless kernels are possible on normal machines, too.

Please fix it properly... probably by requesting timer 10msec in advance or something.
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#11 lore raw

Re: [PATCH 4/7] Allow selected bug checks to be skipped by paravirt kernels

Quoted 7 lines · Pavel Machek
Quoted 6 lines
Quoted 4 lines

How can hlt check break? It is hlt;hlt;hlt, IIRC, that looks fairly innocent to me.

Not if you use tickless timers that don't generate interrupts to unhalt you, or if you delay ticks until the next scheduled timeout and you haven't yet scheduled any timeout. Both are likely in a hypervisor.

Quoted 8 lines

Well.. but you are working around problem, instead of fixing it.

Tickless kernels are possible on normal machines, too.

Please fix it properly... probably by requesting timer 10msec in advance or something.
Pavel

Well, I agree in spirit, but there is something to be said for keeping the code less complicated by removing these workarounds for broken processors. Preferably, we could remove the hlt check entirely, but then those people with these broken processors would not get the expected behavior of stalling during boot - that is the expected behavior of failure, correct? In any case, I added this workaround for the case when running under Xen. I would rather not add a dependence on timer scheduling to legacy bug checking code when the number of timer sources and tickless variations available is proliferating and the number of legacy processors that would even need this check is rapidly approaching zero.

Zach

#12 lore raw

PATCH 5/7 Allow disabling legacy power management modes with paravirt kernels

In reply to Chris Wright

Two legacy power management modes are much easier to just explicitly disable when running in paravirtualized mode - neither APM nor PnP is still relevant. The status of ACPI is still debatable, and noacpi is still a common enough boot parameter that it is not necessary to explicitly disable ACPI.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
---
 arch/i386/kernel/apm.c     |    3 ++-
 drivers/pnp/pnpbios/core.c |    3 ++-
 2 files changed, 4 insertions(+), 2 deletions(-)
arch/i386/kernel/apm.c +2 -1
@@ -231,6 +231,7 @@
 #include <asm/uaccess.h>
 #include <asm/desc.h>
 #include <asm/i8253.h>
+#include <asm/paravirt.h>
 
 #include "io_ports.h"
 
@@ -2191,7 +2192,7 @@ static int __init apm_init(void)
 
 	dmi_check_system(apm_dmi_table);
 
-	if (apm_info.bios.version == 0) {
+	if (apm_info.bios.version == 0 || paravirt_enabled()) {
 		printk(KERN_INFO "apm: BIOS not found.\n");
 		return -ENODEV;
 	}
drivers/pnp/pnpbios/core.c +2 -1
@@ -530,7 +530,8 @@ static int __init pnpbios_init(void)
 	if (check_legacy_ioport(PNPBIOS_BASE))
 		return -ENODEV;
 #endif
-	if (pnpbios_disabled || dmi_check_system(pnpbios_dmi_table)) {
+	if (pnpbios_disabled || dmi_check_system(pnpbios_dmi_table) ||
+	    paravirt_enabled()) {
 		printk(KERN_INFO "PnPBIOS: Disabled\n");
 		return -ENODEV;
 	}

--

#13 lore raw

PATCH 6/7 Add APIC accessors to paravirt-ops.

In reply to Chris Wright

Add APIC accessors to paravirt-ops. Unfortunately, we need two write functions, as some older broken hardware requires workarounds for Pentium APIC errata - this is the purpose of apic_write_atomic.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
---
 arch/i386/kernel/paravirt.c |   28 ++++++++++++++++++++++++++++
 include/asm-i386/apic.h     |    5 ++++-
 include/asm-i386/paravirt.h |   27 +++++++++++++++++++++++++++
 3 files changed, 59 insertions(+), 1 deletion(-)
arch/i386/kernel/paravirt.c +28 -0
@@ -28,6 +28,8 @@
 #include <asm/time.h>
 #include <asm/irq.h>
 #include <asm/delay.h>
+#include <asm/fixmap.h>
+#include <asm/apic.h>
 
 /* nop stub */
 static void native_nop(void)
@@ -382,6 +384,26 @@ static fastcall void native_io_delay(voi
 	asm volatile("outb %al,$0x80");
 }
 
+#ifdef CONFIG_X86_LOCAL_APIC
+/*
+ * Basic functions for reading and writing APIC registers
+ */
+static fastcall void native_apic_write(unsigned long reg, unsigned long v)
+{
+	*((volatile unsigned long *)(APIC_BASE+reg)) = v;
+}
+
+static fastcall void native_apic_write_atomic(unsigned long reg, unsigned long v)
+{
+	xchg((volatile unsigned long *)(APIC_BASE+reg), v);
+}
+
+static fastcall unsigned long native_apic_read(unsigned long reg)
+{
+	return *((volatile unsigned long *)(APIC_BASE+reg));
+}
+#endif /* CONFIG_X86_LOCAL_APIC */
+
 /* These are in entry.S */
 extern fastcall void native_iret(void);
 extern fastcall void native_irq_enable_sysexit(void);
@@ -452,6 +474,12 @@ struct paravirt_ops paravirt_ops = {
 	.io_delay = native_io_delay,
 	.const_udelay = __const_udelay,
 
+#ifdef CONFIG_X86_LOCAL_APIC
+	.apic_write = native_apic_write,
+	.apic_write_atomic = native_apic_write_atomic,
+	.apic_read = native_apic_read,
+#endif
+
 	.irq_enable_sysexit = native_irq_enable_sysexit,
 	.iret = native_iret,
 };
include/asm-i386/apic.h +4 -1
@@ -37,7 +37,9 @@ extern void generic_apic_probe(void);
 /*
  * Basic functions accessing APICs.
  */
-
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 static __inline void apic_write(unsigned long reg, unsigned long v)
 {
 	*((volatile unsigned long *)(APIC_BASE+reg)) = v;
@@ -52,6 +54,7 @@ static __inline unsigned long apic_read(
 {
 	return *((volatile unsigned long *)(APIC_BASE+reg));
 }
+#endif
 
 static __inline__ void apic_wait_icr_idle(void)
 {
include/asm-i386/paravirt.h +27 -0
@@ -115,6 +115,12 @@ struct paravirt_ops
 	void (fastcall *io_delay)(void);
 	void (*const_udelay)(unsigned long loops);
 
+#ifdef CONFIG_X86_LOCAL_APIC
+	void (fastcall *apic_write)(unsigned long reg, unsigned long v);
+	void (fastcall *apic_write_atomic)(unsigned long reg, unsigned long v);
+	unsigned long (fastcall *apic_read)(unsigned long reg);
+#endif
+
 	/* These two are jmp to, not actually called. */
 	void (fastcall *irq_enable_sysexit)(void);
 	void (fastcall *iret)(void);
@@ -280,6 +286,27 @@ static inline void slow_down_io(void) {
 #endif
 }
 
+#ifdef CONFIG_X86_LOCAL_APIC
+/*
+ * Basic functions accessing APICs.
+ */
+static __inline void apic_write(unsigned long reg, unsigned long v)
+{
+	paravirt_ops.apic_write(reg,v);
+}
+
+static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+{
+	paravirt_ops.apic_write_atomic(reg,v);
+}
+
+static __inline unsigned long apic_read(unsigned long reg)
+{
+	return paravirt_ops.apic_read(reg);
+}
+#endif
+
+
 /* These all sit in the .parainstructions section to tell us what to patch. */
 struct paravirt_patch {
 	u8 *instr; 		/* original instructions */

--

#14 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

arch/i386/kernel/paravirt.c static fastcall void native_io_delay(voi · line 387

Quoted 9 lines

/* nop stub */
static void native_nop(void)

arch/i386/kernel/paravirt.c +1 -0
@@ -382,6 +384,26 @@ static fastcall void native_io_delay(voi
 	asm volatile("outb %al,$0x80");
 }
 
+#ifdef CONFIG_X86_LOCAL_APIC

It would be nicer if you renamed the functions in apic.h to native_apic_* and then do

#ifdef CONFIG_PARAVIRT
#include <asm/paravirt.h>
#else
#define apic_read native_apic_read
...
#endif

This way we wouldn't get that much duplication.

This might apply to at least some of the other paravirt ops too.

-Andi

#15 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

Quoted 2 lines · Andi Kleen

It would be nicer if you renamed the functions in apic.h to native_apic_* and then do

...

Quoted 1 line

This might apply to at least some of the other paravirt ops too.

Yes. I've done the obvious candidates below (as well as responding to some of your other points). Many ops are one-liners, and I don't want to cause too much additional churn.

Cheers!
Rusty.

Subject: Paravirtualization Kleenups

1) Add "cheatsheet" comments to entry.S about macros.
2) Use weak alias for init_IRQ -> native_init_IRQ in !CONFIG_PARAVIRT case. This removes an #ifdef.
3) Use shiny new start_kernel.h rather than another declaration.
4) Avoid duplication in paravirt.c: rename set_ldt to native_set_ldt, and use macro in !PARAVIRT case.
5) Same trick for apic ops.

There are other cases where we could use a renaming+macro similar trick to avoid duplication, but they're generally one-liners.

arch/i386/kernel/entry.S +13 -0
diff -r ea3bae5ebb37 arch/i386/kernel/entry.S
@@ -52,6 +52,19 @@
 #include <asm/percpu.h>
 #include <asm/dwarf2.h>
 #include "irq_vectors.h"
+
+/*
+ * We use macros for low-level operations which need to be overridden
+ * for paravirtualization.  The following will never clobber any registers:
+ *   INTERRUPT_RETURN (aka. "iret")
+ *   GET_CR0_INTO_EAX (aka. "movl %cr0, %eax")
+ *   ENABLE_INTERRUPTS_SYSEXIT (aka "sti; sysexit").
+ *
+ * For DISABLE_INTERRUPTS/ENABLE_INTERRUPTS (aka "cli"/"sti"), you must
+ * specify what registers can be overwritten (CLBR_NONE, CLBR_EAX/EDX/ECX/ANY).
+ * Allowing a register to be clobbered can shrink the paravirt replacement
+ * enough to patch inline, increasing performance.
+ */
 
 #define nr_syscalls ((syscall_table_size)/4)
 
arch/i386/kernel/i8259.c +3 -0
diff -r ea3bae5ebb37 arch/i386/kernel/i8259.c
@@ -392,6 +392,9 @@ void __init init_ISA_irqs (void)
 	}
 }
 
+/* Overridden in paravirt.c */
+void init_IRQ(void) __attribute__((weak, alias("native_init_IRQ")));
+
 void __init native_init_IRQ(void)
 {
 	int i;
arch/i386/kernel/paravirt.c +6 -37

73 lines not on this page · raw message

include/asm-i386/apic.h +10 -4
diff -r ea3bae5ebb37 include/asm-i386/apic.h
@@ -40,21 +40,27 @@ extern void generic_apic_probe(void);
 #ifdef CONFIG_PARAVIRT
 #include <asm/paravirt.h>
 #else
-static __inline void apic_write(unsigned long reg, unsigned long v)
+#define apic_write native_apic_write
+#define apic_write_atomic native_apic_write_atomic
+#define apic_read native_apic_read
+#endif
+
+static __inline fastcall void native_apic_write(unsigned long reg,
+						unsigned long v)
 {
 	*((volatile unsigned long *)(APIC_BASE+reg)) = v;
 }
 
-static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+static __inline fastcall void native_apic_write_atomic(unsigned long reg,
+						       unsigned long v)
 {
 	xchg((volatile unsigned long *)(APIC_BASE+reg), v);
 }
 
-static __inline unsigned long apic_read(unsigned long reg)
+static __inline fastcall unsigned long native_apic_read(unsigned long reg)
 {
 	return *((volatile unsigned long *)(APIC_BASE+reg));
 }
-#endif
 
 static __inline__ void apic_wait_icr_idle(void)
 {
include/asm-i386/desc.h +5 -2
diff -r ea3bae5ebb37 include/asm-i386/desc.h
@@ -92,7 +92,11 @@ static inline void write_dt_entry(void *
 	lp[1] = entry_high;
 }
 
-static inline void set_ldt(void *addr, unsigned int entries)
+#define set_ldt native_set_ldt
+#endif /* CONFIG_PARAVIRT */
+
+static inline fastcall void native_set_ldt(const void *addr,
+					   unsigned int entries)
 {
 	if (likely(entries == 0))
 		__asm__ __volatile__("lldt %w0"::"q" (0));
@@ -107,7 +111,6 @@ static inline void set_ldt(void *addr, u
 		__asm__ __volatile__("lldt %w0"::"q" (GDT_ENTRY_LDT*8));
 	}
 }
-#endif /* CONFIG_PARAVIRT */
 
 static inline void _set_gate(int gate, unsigned int type, void *addr, unsigned short seg)
 {
include/asm-i386/irq.h +1 -8
diff -r ea3bae5ebb37 include/asm-i386/irq.h
@@ -41,14 +41,7 @@ extern void fixup_irqs(cpumask_t map);
 extern void fixup_irqs(cpumask_t map);
 #endif
 
+void init_IRQ(void);
 void __init native_init_IRQ(void);
-#ifdef CONFIG_PARAVIRT
-#include <asm/paravirt.h>
-#else
-static inline void init_IRQ(void)
-{
-	native_init_IRQ();
-}
-#endif /* CONFIG_PARAVIRT */
 
 #endif /* _ASM_IRQ_H */
include/asm-i386/paravirt.h +0 -5
diff -r ea3bae5ebb37 include/asm-i386/paravirt.h
@@ -153,11 +153,6 @@ extern struct paravirt_ops paravirt_ops;
 extern struct paravirt_ops paravirt_ops;
 
 #define paravirt_enabled() (paravirt_ops.paravirt_enabled)
-
-static inline void init_IRQ(void)
-{
-	paravirt_ops.init_IRQ();
-}
 
 static inline void load_esp0(struct tss_struct *tss,
 			     struct thread_struct *thread)
#16 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

Quoted 1 line

Subject: Paravirtualization Kleenups

Thanks.

Chris, can you please merge those into the original patchkit?

I could do it myself, but then retransmits from Chris would be difficult if anything else would need to be changed.

Also fixing that !-Os compile error in the original patches would be good.

-Andi

#17 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

Quoted 1 line · * Andi Kleen ([email protected]) wrote:

Chris, can you please merge those into the original patchkit?

Sure, I'll fold those in.

Quoted 4 lines

I could do it myself, but then retransmits from Chris would be difficult if anything else would need to be changed.

Also fixing that !-Os compile error in the original patches would be good.

Hmm, builds fine here. If you have a .config and/or error message I'll fix it up.

thanks,
-chris

#18 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

Quoted 7 lines · Chris Wright
Quoted 4 lines

I could do it myself, but then retransmits from Chris would be difficult if anything else would need to be changed.

Also fixing that !-Os compile error in the original patches would be good.

Hmm, builds fine here. If you have a .config and/or error message I'll fix it up.

I haven't tried it myself (my laptop was on battery all the time and I didn't want to drain it with a full rebuild ;-), there was just a report that it didn't work. Or maybe that was with an old patch. If it works it's fine.

-Andi

#19 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

Quoted 3 lines · * Andi Kleen ([email protected]) wrote:

I haven't tried it myself (my laptop was on battery all the time and I didn't want to drain it with a full rebuild ;-), there was just a report that it didn't work. Or maybe that was with an old patch. If it works it's fine.

Ah yes, I see the report, (it's against a patch that has been redone), but I'll double check.

thanks,
-chris

#20 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

In reply to Andi Kleen
Quoted 5 lines · Andi Kleen
Quoted 3 lines · Chris Wright
Quoted 3 lines

I could do it myself, but then retransmits from Chris would be difficult if anything else would need to be changed.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

Also fixing that !-Os compile error in the original patches would be good.

Hmm, builds fine here. If you have a .config and/or error message I'll fix it up.

I haven't tried it myself (my laptop was on battery all the time and I didn't want to drain it with a full rebuild ;-), there was just a report that it didn't work. Or maybe that was with an old patch. If it works it's fine.

The -Os thing was a red herring. It was a brokenpatch in the original 4 which for which I immediately sent a fixup to akpm. Here it is again below:

==
Move write_dt_entry back: moving it up breaks compile.

===================================================================

include/asm-i386/desc.h +11 -11
@@ -78,6 +78,17 @@ static inline void load_TLS(struct threa
 #undef C
 }
 
+#define write_ldt_entry(dt, entry, low, high) write_dt_entry(dt,entry,low,high)
+#define write_gdt_entry(dt, entry, low, high) write_dt_entry(dt,entry,low,high)
+#define write_idt_entry(dt, entry, low, high) write_dt_entry(dt,entry,low,high)
+
+static inline void write_dt_entry(void *dt, int entry, u32 entry_low, u32 entry_high)
+{
+	u32 *lp = (u32 *)((char *)dt + entry*8);
+	lp[0] = entry_low;
+	lp[1] = entry_high;
+}
+
 static inline void set_ldt(void *addr, unsigned int entries)
 {
 	if (likely(entries == 0))
@@ -92,17 +103,6 @@ static inline void set_ldt(void *addr, u
 		write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_LDT, low, high);
 		__asm__ __volatile__("lldt %w0"::"q" (GDT_ENTRY_LDT*8));
 	}
-}
-
-#define write_ldt_entry(dt, entry, low, high) write_dt_entry(dt,entry,low,high)
-#define write_gdt_entry(dt, entry, low, high) write_dt_entry(dt,entry,low,high)
-#define write_idt_entry(dt, entry, low, high) write_dt_entry(dt,entry,low,high)
-
-static inline void write_dt_entry(void *dt, int entry, u32 entry_low, u32 entry_high)
-{
-	u32 *lp = (u32 *)((char *)dt + entry*8);
-	lp[0] = entry_low;
-	lp[1] = entry_high;
 }
 
 static inline void _set_gate(int gate, unsigned int type, void *addr, unsigned short seg)
#21 lore raw

Re: [PATCH 6/7] Add APIC accessors to paravirt-ops.

In reply to Andi Kleen
Quoted 10 lines · Andi Kleen
Quoted 1 line

Subject: Paravirtualization Kleenups

Thanks.

Chris, can you please merge those into the original patchkit?

I could do it myself, but then retransmits from Chris would be difficult if anything else would need to be changed.

Also fixing that !-Os compile error in the original patches would be good.

That is "prep-for-paravirt-desch-clearer-parameter-names-fix.patch" in rc4-mm1.

I'll follow with the updated series, although the cleanup patch was pretty clear by itself...

Rusty.

#22 lore raw

PATCH 1/7 paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Andi Kleen

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

This patch does the dumbest possible replacement of paravirtualized
instructions: calls through a "paravirt_ops" structure. Currently these are function implementations of native hardware: hypervisors will override the ops structure with their own variants.

All the pv-ops functions are declared "fastcall" so that a specific register-based ABI is used, to make inlining assember easier.

Cc Jeremy Fitzhardinge <[email protected]>
Cc Zachary Amsden <[email protected]>

===================================================================

arch/i386/Kconfig +11 -0
@@ -196,6 +196,17 @@ config X86_ES7000
 	  should say N here.
 
 endchoice
+
+config PARAVIRT
+	bool "Paravirtualization support (EXPERIMENTAL)"
+	depends on EXPERIMENTAL
+	help
+	  Paravirtualization is a way of running multiple instances of
+	  Linux on the same machine, under a hypervisor.  This option
+	  changes the kernel so it can modify itself when it is run
+	  under a hypervisor, improving performance significantly.
+	  However, when run without a hypervisor the kernel is
+	  theoretically slower.  If in doubt, say N.
 
 config ACPI_SRAT
 	bool

===================================================================

arch/i386/boot/compressed/misc.c +1 -0
@@ -9,6 +9,7 @@
  * High loaded stuff by Hans Lermen & Werner Almesberger, Feb. 1996
  */
 
+#undef CONFIG_PARAVIRT
 #include <linux/linkage.h>
 #include <linux/vmalloc.h>
 #include <linux/screen_info.h>

===================================================================

arch/i386/kernel/Makefile +1 -0
@@ -39,6 +39,7 @@ obj-$(CONFIG_EARLY_PRINTK)	+= early_prin
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 obj-$(CONFIG_HPET_TIMER) 	+= hpet.o
 obj-$(CONFIG_K8_NB)		+= k8.o
+obj-$(CONFIG_PARAVIRT)		+= paravirt.o
 
 EXTRA_AFLAGS   := -traditional
 

===================================================================

arch/i386/kernel/asm-offsets.c +10 -0
@@ -101,4 +101,14 @@ void foo(void)
 	BLANK();
  	OFFSET(PDA_cpu, i386_pda, cpu_number);
 	OFFSET(PDA_pcurrent, i386_pda, pcurrent);
+
+#ifdef CONFIG_PARAVIRT
+	BLANK();
+	OFFSET(PARAVIRT_enabled, paravirt_ops, paravirt_enabled);
+	OFFSET(PARAVIRT_irq_disable, paravirt_ops, irq_disable);
+	OFFSET(PARAVIRT_irq_enable, paravirt_ops, irq_enable);
+	OFFSET(PARAVIRT_irq_enable_sysexit, paravirt_ops, irq_enable_sysexit);
+	OFFSET(PARAVIRT_iret, paravirt_ops, iret);
+	OFFSET(PARAVIRT_read_cr0, paravirt_ops, read_cr0);
+#endif
 }

===================================================================

arch/i386/kernel/entry.S +27 -7
@@ -61,13 +61,6 @@ DF_MASK		= 0x00000400
 DF_MASK		= 0x00000400 
 NT_MASK		= 0x00004000
 VM_MASK		= 0x00020000
-
-/* These are replaces for paravirtualization */
-#define DISABLE_INTERRUPTS		cli
-#define ENABLE_INTERRUPTS		sti
-#define ENABLE_INTERRUPTS_SYSEXIT	sti; sysexit
-#define INTERRUPT_RETURN		iret
-#define GET_CR0_INTO_EAX		movl %cr0, %eax
 
 #ifdef CONFIG_PREEMPT
 #define preempt_stop		DISABLE_INTERRUPTS; TRACE_IRQS_OFF
@@ -416,6 +409,20 @@ ldt_ss:
 	jnz restore_nocheck
 	testl $0x00400000, %eax		# returning to 32bit stack?
 	jnz restore_nocheck		# allright, normal return
+
+#ifdef CONFIG_PARAVIRT
+	/* 
+	 * The kernel can't run on a non-flat stack if paravirt mode
+	 * is active.  Rather than try to fixup the high bits of
+	 * ESP, bypass this code entirely.  This may break DOSemu
+	 * and/or Wine support in a paravirt VM, although the option
+	 * is still available to implement the setting of the high
+	 * 16-bits in the INTERRUPT_RETURN paravirt-op.
+	 */
+	cmpl $0, paravirt_ops+PARAVIRT_enabled
+	jne restore_nocheck
+#endif
+	
 	/* If returning to userspace with 16bit stack,
 	 * try to fix the higher word of ESP, as the CPU
 	 * won't restore it.
@@ -830,6 +837,19 @@ 1:	INTERRUPT_RETURN
 	.long 1b,iret_exc
 .previous
 KPROBE_END(nmi)
+
+#ifdef CONFIG_PARAVIRT
+ENTRY(native_iret)
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
+ENTRY(native_irq_enable_sysexit)
+	sti
+	sysexit
+#endif
 
 KPROBE_ENTRY(int3)
 	RING0_INT_FRAME

===================================================================

arch/i386/kernel/i8259.c +4 -1
@@ -392,7 +392,10 @@ void __init init_ISA_irqs (void)
 	}
 }
 
-void __init init_IRQ(void)
+/* Overridden in paravirt.c */
+void init_IRQ(void) __attribute__((weak, alias("native_init_IRQ")));
+
+void __init native_init_IRQ(void)
 {
 	int i;
 

===================================================================

arch/i386/kernel/setup.c +1 -1
@@ -1405,7 +1405,7 @@ void __init setup_arch(char **cmdline_p)
 		efi_init();
 	else {
 		printk(KERN_INFO "BIOS-provided physical RAM map:\n");
-		print_memory_map(machine_specific_memory_setup());
+		print_memory_map(memory_setup());
 	}
 
 	copy_edd();

===================================================================

arch/i386/kernel/smpboot.c +5 -0
@@ -33,6 +33,11 @@
  *		Dave Jones	:	Report invalid combinations of Athlon CPUs.
 *		Rusty Russell	:	Hacked into shape for new "hotplug" boot process. */
 
+
+/* SMP boot always wants to use real time delay to allow sufficient time for
+ * the APs to come online */
+#define USE_REAL_TIME_DELAY
+
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/kernel.h>

===================================================================

arch/i386/kernel/time.c +6 -11
@@ -56,6 +56,7 @@
 #include <asm/uaccess.h>
 #include <asm/processor.h>
 #include <asm/timer.h>
+#include <asm/time.h>
 
 #include "mach_time.h"
 
@@ -116,10 +117,7 @@ static int set_rtc_mmss(unsigned long no
 	/* gets recalled with irq locally disabled */
 	/* XXX - does irqsave resolve this? -johnstul */
 	spin_lock_irqsave(&rtc_lock, flags);
-	if (efi_enabled)
-		retval = efi_set_rtc_mmss(nowtime);
-	else
-		retval = mach_set_rtc_mmss(nowtime);
+	retval = set_wallclock(nowtime);
 	spin_unlock_irqrestore(&rtc_lock, flags);
 
 	return retval;
@@ -211,10 +209,7 @@ unsigned long read_persistent_clock(void
 
 	spin_lock_irqsave(&rtc_lock, flags);
 
-	if (efi_enabled)
-		retval = efi_get_time();
-	else
-		retval = mach_get_cmos_time();
+	retval = get_wallclock();
 
 	spin_unlock_irqrestore(&rtc_lock, flags);
 
@@ -280,7 +275,7 @@ static void __init hpet_time_init(void)
 		printk("Using HPET for base-timer\n");
 	}
 
-	time_init_hook();
+	do_time_init();
 }
 #endif
 
@@ -296,5 +291,5 @@ void __init time_init(void)
 		return;
 	}
 #endif
-	time_init_hook();
-}
+	do_time_init();
+}

===================================================================

drivers/net/de600.c +0 -1
@@ -43,7 +43,6 @@ static const char version[] = "de600.c: 
  * modify the following "#define": (see <asm/io.h> for more info)
 #define REALLY_SLOW_IO
  */
-#define SLOW_IO_BY_JUMPING /* Looks "better" than dummy write to port 0x80 :-) */
 
 /* use 0 for production, 1 for verification, >2 for debug */
 #ifdef DE600_DEBUG

===================================================================

include/asm-i386/delay.h +8 -0
@@ -15,6 +15,13 @@ extern void __const_udelay(unsigned long
 extern void __const_udelay(unsigned long usecs);
 extern void __delay(unsigned long loops);
 
+#if defined(CONFIG_PARAVIRT) && !defined(USE_REAL_TIME_DELAY)
+#define udelay(n) paravirt_ops.const_udelay((n) * 0x10c7ul)
+	
+#define ndelay(n) paravirt_ops.const_udelay((n) * 5ul)
+
+#else /* !PARAVIRT || USE_REAL_TIME_DELAY */
+
 #define udelay(n) (__builtin_constant_p(n) ? \
 	((n) > 20000 ? __bad_udelay() : __const_udelay((n) * 0x10c7ul)) : \
 	__udelay(n))
@@ -22,6 +29,7 @@ extern void __delay(unsigned long loops)
 #define ndelay(n) (__builtin_constant_p(n) ? \
 	((n) > 20000 ? __bad_ndelay() : __const_udelay((n) * 5ul)) : \
 	__ndelay(n))
+#endif
 
 void use_tsc_delay(void);
 

===================================================================

include/asm-i386/desc.h +8 -1
@@ -55,6 +55,9 @@ static inline void pack_gate(u32 *low, u
 #define DESCTYPE_DPL3	0x60	/* DPL-3 */
 #define DESCTYPE_S	0x10	/* !system */
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 #define load_TR_desc() __asm__ __volatile__("ltr %w0"::"q" (GDT_ENTRY_TSS*8))
 
 #define load_gdt(dtr) __asm__ __volatile("lgdt %0"::"m" (*dtr))
@@ -89,7 +92,11 @@ static inline void write_dt_entry(void *
 	lp[1] = entry_high;
 }
 
-static inline void set_ldt(void *addr, unsigned int entries)
+#define set_ldt native_set_ldt
+#endif /* CONFIG_PARAVIRT */
+
+static inline fastcall void native_set_ldt(const void *addr,
+					   unsigned int entries)
 {
 	if (likely(entries == 0))
 		__asm__ __volatile__("lldt %w0"::"q" (0));

===================================================================

include/asm-i386/io.h +6 -4
@@ -256,11 +256,11 @@ static inline void flush_write_buffers(v
 
 #endif /* __KERNEL__ */
 
-#ifdef SLOW_IO_BY_JUMPING
-#define __SLOW_DOWN_IO "jmp 1f; 1: jmp 1f; 1:"
-#else
+#if defined(CONFIG_PARAVIRT)
+#include <asm/paravirt.h>
+#else 
+
 #define __SLOW_DOWN_IO "outb %%al,$0x80;"
-#endif
 
 static inline void slow_down_io(void) {
 	__asm__ __volatile__(
@@ -270,6 +270,8 @@ static inline void slow_down_io(void) {
 #endif
 		: : );
 }
+
+#endif
 
 #ifdef CONFIG_X86_NUMAQ
 extern void *xquad_portio;    /* Where the IO area was mapped */

===================================================================

include/asm-i386/irq.h +3 -0
@@ -41,4 +41,7 @@ extern void fixup_irqs(cpumask_t map);
 extern void fixup_irqs(cpumask_t map);
 #endif
 
+void init_IRQ(void);
+void __init native_init_IRQ(void);
+
 #endif /* _ASM_IRQ_H */

===================================================================

include/asm-i386/irqflags.h +27 -15

69 lines not on this page · raw message

===================================================================

include/asm-i386/mach-default/setup_arch.h +2 -0
@@ -2,4 +2,6 @@
 
 /* no action for generic */
 
+#ifndef ARCH_SETUP
 #define ARCH_SETUP
+#endif

===================================================================

include/asm-i386/msr.h +5 -0
@@ -1,5 +1,9 @@
 #ifndef __ASM_MSR_H
 #define __ASM_MSR_H
+
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 
 /*
  * Access to machine-specific registers (available on 586 and better only)
@@ -77,6 +81,7 @@ static inline void wrmsrl (unsigned long
      __asm__ __volatile__("rdpmc" \
 			  : "=a" (low), "=d" (high) \
 			  : "c" (counter))
+#endif	/* !CONFIG_PARAVIRT */
 
 /* symbolic names for some interesting MSRs */
 /* Intel defined MSRs. */

===================================================================

include/asm-i386/processor.h +12 -3
@@ -146,8 +146,8 @@ static inline void detect_ht(struct cpui
 #define X86_EFLAGS_VIP	0x00100000 /* Virtual Interrupt Pending */
 #define X86_EFLAGS_ID	0x00200000 /* CPUID detection flag */
 
-static inline void __cpuid(unsigned int *eax, unsigned int *ebx,
-			   unsigned int *ecx, unsigned int *edx)
+static inline fastcall void native_cpuid(unsigned int *eax, unsigned int *ebx,
+					 unsigned int *ecx, unsigned int *edx)
 {
 	/* ecx is often an input as well as an output. */
 	__asm__("cpuid"
@@ -548,6 +548,12 @@ static inline void rep_nop(void)
 
 #define cpu_relax()	rep_nop()
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+#define paravirt_enabled() 0
+#define __cpuid native_cpuid
+
 static inline void load_esp0(struct tss_struct *tss, struct thread_struct *thread)
 {
 	tss->esp0 = thread->esp0;
@@ -570,10 +576,13 @@ static inline void load_esp0(struct tss_
 			: /* no output */			\
 			:"r" (value))
 
+#define set_iopl_mask native_set_iopl_mask
+#endif /* CONFIG_PARAVIRT */
+
 /*
  * Set IOPL bits in EFLAGS from given mask
  */
-static inline void set_iopl_mask(unsigned mask)
+static fastcall inline void native_set_iopl_mask(unsigned mask)
 {
 	unsigned int reg;
 	__asm__ __volatile__ ("pushfl;"

===================================================================

include/asm-i386/segment.h +2 -0
@@ -131,5 +131,7 @@
 #define SEGMENT_LDT		0x4
 #define SEGMENT_GDT		0x0
 
+#ifndef CONFIG_PARAVIRT
 #define get_kernel_rpl()  0
 #endif
+#endif

===================================================================

include/asm-i386/setup.h +8 -0
@@ -70,6 +70,14 @@ struct e820entry;
 struct e820entry;
 
 char * __init machine_specific_memory_setup(void);
+#ifndef CONFIG_PARAVIRT
+static inline char *memory_setup(void)
+{
+	return machine_specific_memory_setup();
+}
+#else
+#include <asm/paravirt.h>
+#endif
 
 int __init copy_e820_map(struct e820entry * biosmap, int nr_map);
 int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map);

===================================================================

include/asm-i386/spinlock.h +4 -0
@@ -7,8 +7,12 @@
 #include <asm/processor.h>
 #include <linux/compiler.h>
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 #define CLI_STRING	"cli"
 #define STI_STRING	"sti"
+#endif /* CONFIG_PARAVIRT */
 
 /*
  * Your basic SMP spinlocks, allowing only a single CPU anywhere

===================================================================

include/asm-i386/system.h +12 -8
@@ -88,6 +88,9 @@ __asm__ __volatile__ ("movw %%dx,%1\n\t"
 #define savesegment(seg, value) \
 	asm volatile("mov %%" #seg ",%0":"=rm" (value))
 
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
 #define read_cr0() ({ \
 	unsigned int __dummy; \
 	__asm__ __volatile__( \
@@ -139,16 +142,17 @@ __asm__ __volatile__ ("movw %%dx,%1\n\t"
 #define write_cr4(x) \
 	__asm__ __volatile__("movl %0,%%cr4": :"r" (x))
 
-/*
- * Clear and set 'TS' bit respectively
- */
-#define clts() __asm__ __volatile__ ("clts")
-#define stts() write_cr0(8 | read_cr0())
-
-#endif	/* __KERNEL__ */
-
 #define wbinvd() \
 	__asm__ __volatile__ ("wbinvd": : :"memory")
+
+/* Clear the 'TS' bit */
+#define clts() __asm__ __volatile__ ("clts")
+#endif/* CONFIG_PARAVIRT */
+
+/* Set the 'TS' bit */
+#define stts() write_cr0(8 | read_cr0())
+
+#endif	/* __KERNEL__ */
 
 static inline unsigned long get_limit(unsigned long segment)
 {

===================================================================

arch/i386/kernel/paravirt.c new file +399 -0

399 lines not on this page · raw message

===================================================================

include/asm-i386/paravirt.h new file +286 -0

286 lines not on this page · raw message

===================================================================

include/asm-i386/time.h new file +41 -0

41 lines not on this page · raw message

#23 lore raw

PATCH 2/7 paravirtualization: Patch inline replacements for common paravirt operations.

It turns out that the most called ops, by several orders of magnitude, are the interrupt manipulation ops. These are obvious candidates for patching, so mark them up and create infrastructure for it.

The method used is that the ops structure has a patch function, which is called for each place which needs to be patched: this returns a number of instructions (the rest are NOP-padded).

Usually we can spare a register (%eax) for the binary patched code to use, but in a couple of critical places in entry.S we can't: we make the clobbers explicit at the call site, and manually clobber the allowed registers in debug mode as an extra check.

===================================================================

arch/i386/kernel/alternative.c +55 -11
@@ -123,6 +123,20 @@ static unsigned char** find_nop_table(vo
 
 #endif /* CONFIG_X86_64 */
 
+static void nop_out(void *insns, unsigned int len)
+{
+	unsigned char **noptable = find_nop_table();
+
+	while (len > 0) {
+		unsigned int noplen = len;
+		if (noplen > ASM_NOP_MAX)
+			noplen = ASM_NOP_MAX;
+		memcpy(insns, noptable[noplen], noplen);
+		insns += noplen;
+		len -= noplen;
+	}
+}
+
 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
 extern struct alt_instr __smp_alt_instructions[], __smp_alt_instructions_end[];
 extern u8 *__smp_locks[], *__smp_locks_end[];
@@ -137,10 +151,9 @@ extern u8 __smp_alt_begin[], __smp_alt_e
 
 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
 {
-	unsigned char **noptable = find_nop_table();
 	struct alt_instr *a;
 	u8 *instr;
-	int diff, i, k;
+	int diff;
 
 	DPRINTK("%s: alt table %p -> %p\n", __FUNCTION__, start, end);
 	for (a = start; a < end; a++) {
@@ -158,13 +171,7 @@ void apply_alternatives(struct alt_instr
 #endif
 		memcpy(instr, a->replacement, a->replacementlen);
 		diff = a->instrlen - a->replacementlen;
-		/* Pad the rest with nops */
-		for (i = a->replacementlen; diff > 0; diff -= k, i += k) {
-			k = diff;
-			if (k > ASM_NOP_MAX)
-				k = ASM_NOP_MAX;
-			memcpy(a->instr + i, noptable[k], k);
-		}
+		nop_out(instr + a->replacementlen, diff);
 	}
 }
 
@@ -208,7 +215,6 @@ static void alternatives_smp_lock(u8 **s
 
 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 {
-	unsigned char **noptable = find_nop_table();
 	u8 **ptr;
 
 	for (ptr = start; ptr < end; ptr++) {
@@ -216,7 +222,7 @@ static void alternatives_smp_unlock(u8 *
 			continue;
 		if (*ptr > text_end)
 			continue;
-		**ptr = noptable[1][0];
+		nop_out(*ptr, 1);
 	};
 }
 
@@ -341,6 +347,43 @@ void alternatives_smp_switch(int smp)
 }
 
 #endif
+
+#ifdef CONFIG_PARAVIRT
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end)
+{
+	struct paravirt_patch *p;
+	int i;
+
+	for (p = start; p < end; p++) {
+		unsigned int used;
+
+		used = paravirt_ops.patch(p->instrtype, p->clobbers, p->instr,
+					  p->len);
+#ifdef CONFIG_DEBUG_KERNEL
+		/* Deliberately clobber regs using "not %reg" to find bugs. */
+		for (i = 0; i < 3; i++) {
+			if (p->len - used >= 2 && (p->clobbers & (1 << i))) {
+				memcpy(p->instr + used, "\xf7\xd0", 2);
+				p->instr[used+1] |= i;
+				used += 2;
+			}
+		}
+#endif
+		/* Pad the rest with nops */
+		nop_out(p->instr + used, p->len - used);
+	}
+
+	/* Sync to be conservative, in case we patched following instructions */
+	sync_core();
+}
+extern struct paravirt_patch __start_parainstructions[],
+	__stop_parainstructions[];
+#else
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end)
+{
+}
+extern struct paravirt_patch *__start_parainstructions, *__stop_parainstructions;
+#endif	/* CONFIG_PARAVIRT */
 
 void __init alternative_instructions(void)
 {
@@ -389,5 +432,6 @@ void __init alternative_instructions(voi
 		alternatives_smp_switch(0);
 	}
 #endif
+ 	apply_paravirt(__start_parainstructions, __stop_parainstructions);
 	local_irq_restore(flags);
 }

===================================================================

arch/i386/kernel/entry.S +26 -13
@@ -53,6 +53,19 @@
 #include <asm/dwarf2.h>
 #include "irq_vectors.h"
 
+/*
+ * We use macros for low-level operations which need to be overridden
+ * for paravirtualization.  The following will never clobber any registers:
+ *   INTERRUPT_RETURN (aka. "iret")
+ *   GET_CR0_INTO_EAX (aka. "movl %cr0, %eax")
+ *   ENABLE_INTERRUPTS_SYSEXIT (aka "sti; sysexit").
+ *
+ * For DISABLE_INTERRUPTS/ENABLE_INTERRUPTS (aka "cli"/"sti"), you must
+ * specify what registers can be overwritten (CLBR_NONE, CLBR_EAX/EDX/ECX/ANY).
+ * Allowing a register to be clobbered can shrink the paravirt replacement
+ * enough to patch inline, increasing performance.
+ */
+
 #define nr_syscalls ((syscall_table_size)/4)
 
 CF_MASK		= 0x00000001
@@ -63,9 +76,9 @@ VM_MASK		= 0x00020000
 VM_MASK		= 0x00020000
 
 #ifdef CONFIG_PREEMPT
-#define preempt_stop		DISABLE_INTERRUPTS; TRACE_IRQS_OFF
+#define preempt_stop(clobbers)	DISABLE_INTERRUPTS(clobbers); TRACE_IRQS_OFF
 #else
-#define preempt_stop
+#define preempt_stop(clobbers)
 #define resume_kernel		restore_nocheck
 #endif
 
@@ -226,7 +239,7 @@ ENTRY(ret_from_fork)
 	ALIGN
 	RING0_PTREGS_FRAME
 ret_from_exception:
-	preempt_stop
+	preempt_stop(CLBR_ANY)
 ret_from_intr:
 	GET_THREAD_INFO(%ebp)
 check_userspace:
@@ -237,7 +250,7 @@ check_userspace:
 	jb resume_kernel		# not returning to v8086 or userspace
 
 ENTRY(resume_userspace)
- 	DISABLE_INTERRUPTS		# make sure we don't miss an interrupt
+ 	DISABLE_INTERRUPTS(CLBR_ANY)	# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
 					# between sampling and the iret
 	movl TI_flags(%ebp), %ecx
@@ -248,7 +261,7 @@ ENTRY(resume_userspace)
 
 #ifdef CONFIG_PREEMPT
 ENTRY(resume_kernel)
-	DISABLE_INTERRUPTS
+	DISABLE_INTERRUPTS(CLBR_ANY)
 	cmpl $0,TI_preempt_count(%ebp)	# non-zero preempt_count ?
 	jnz restore_nocheck
 need_resched:
@@ -277,7 +290,7 @@ sysenter_past_esp:
 	 * No need to follow this irqs on/off section: the syscall
 	 * disabled irqs and here we enable it straight after entry:
 	 */
-	ENABLE_INTERRUPTS
+	ENABLE_INTERRUPTS(CLBR_NONE)
 	pushl $(__USER_DS)
 	CFI_ADJUST_CFA_OFFSET 4
 	/*CFI_REL_OFFSET ss, 0*/
@@ -322,7 +335,7 @@ 1:	movl (%ebp),%ebp
 	jae syscall_badsys
 	call *sys_call_table(,%eax,4)
 	movl %eax,PT_EAX(%esp)
-	DISABLE_INTERRUPTS
+	DISABLE_INTERRUPTS(CLBR_ECX|CLBR_EDX)
 	TRACE_IRQS_OFF
 	movl TI_flags(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx
@@ -364,7 +377,7 @@ syscall_call:
 	call *sys_call_table(,%eax,4)
 	movl %eax,PT_EAX(%esp)		# store the return value
 syscall_exit:
-	DISABLE_INTERRUPTS		# make sure we don't miss an interrupt
+	DISABLE_INTERRUPTS(CLBR_ANY)	# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
 					# between sampling and the iret
 	TRACE_IRQS_OFF
@@ -393,7 +406,7 @@ 1:	INTERRUPT_RETURN
 .section .fixup,"ax"
 iret_exc:
 	TRACE_IRQS_ON
-	ENABLE_INTERRUPTS
+	ENABLE_INTERRUPTS(CLBR_NONE)
 	pushl $0			# no error code
 	pushl $do_iret_error
 	jmp error_code
@@ -436,7 +449,7 @@ ldt_ss:
 	CFI_ADJUST_CFA_OFFSET 4
 	pushl %eax
 	CFI_ADJUST_CFA_OFFSET 4
-	DISABLE_INTERRUPTS
+	DISABLE_INTERRUPTS(CLBR_EAX)
 	TRACE_IRQS_OFF
 	lss (%esp), %esp
 	CFI_ADJUST_CFA_OFFSET -8
@@ -451,7 +464,7 @@ work_pending:
 	jz work_notifysig
 work_resched:
 	call schedule
-	DISABLE_INTERRUPTS		# make sure we don't miss an interrupt
+	DISABLE_INTERRUPTS(CLBR_ANY)	# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
 					# between sampling and the iret
 	TRACE_IRQS_OFF
@@ -507,7 +520,7 @@ syscall_exit_work:
 	testb $(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SINGLESTEP), %cl
 	jz work_pending
 	TRACE_IRQS_ON
-	ENABLE_INTERRUPTS		# could let do_syscall_trace() call
+	ENABLE_INTERRUPTS(CLBR_ANY)	# could let do_syscall_trace() call
 					# schedule() instead
 	movl %esp, %eax
 	movl $1, %edx
@@ -691,7 +704,7 @@ ENTRY(device_not_available)
 	GET_CR0_INTO_EAX
 	testl $0x4, %eax		# EM (math emulation bit)
 	jne device_not_available_emulate
-	preempt_stop
+	preempt_stop(CLBR_ANY)
 	call math_state_restore
 	jmp ret_from_exception
 device_not_available_emulate:

===================================================================

arch/i386/kernel/module.c +8 -1
@@ -109,7 +109,8 @@ int module_finalize(const Elf_Ehdr *hdr,
 		    const Elf_Shdr *sechdrs,
 		    struct module *me)
 {
-	const Elf_Shdr *s, *text = NULL, *alt = NULL, *locks = NULL;
+	const Elf_Shdr *s, *text = NULL, *alt = NULL, *locks = NULL,
+		*para = NULL;
 	char *secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
 
 	for (s = sechdrs; s < sechdrs + hdr->e_shnum; s++) { 
@@ -119,6 +120,8 @@ int module_finalize(const Elf_Ehdr *hdr,
 			alt = s;
 		if (!strcmp(".smp_locks", secstrings + s->sh_name))
 			locks= s;
+		if (!strcmp(".parainstructions", secstrings + s->sh_name))
+			para = s;
 	}
 
 	if (alt) {
@@ -133,6 +136,10 @@ int module_finalize(const Elf_Ehdr *hdr,
 					    lseg, lseg + locks->sh_size,
 					    tseg, tseg + text->sh_size);
 	}
+	if (para) {
+		void *pseg = (void *)para->sh_addr;
+		apply_paravirt(pseg, pseg + para->sh_size);
+	}
 
 	return module_bug_finalize(hdr, sechdrs, me);
 }

===================================================================

arch/i386/kernel/paravirt.c +44 -0

56 lines not on this page · raw message

===================================================================

arch/i386/kernel/vmlinux.lds.S +6 -0
@@ -154,6 +154,12 @@ SECTIONS
   .altinstr_replacement : AT(ADDR(.altinstr_replacement) - LOAD_OFFSET) {
 	*(.altinstr_replacement)
   }
+  . = ALIGN(4);
+  __start_parainstructions = .;
+  .parainstructions : AT(ADDR(.parainstructions) - LOAD_OFFSET) {
+	*(.parainstructions)
+  }
+  __stop_parainstructions = .;
   /* .exit.text is discard at runtime, not link time, to deal with references
      from .altinstructions and .eh_frame */
   .exit.text : AT(ADDR(.exit.text) - LOAD_OFFSET) { *(.exit.text) }

===================================================================

include/asm-i386/alternative.h +3 -0
@@ -118,4 +118,7 @@ static inline void alternatives_smp_swit
 #define LOCK_PREFIX ""
 #endif
 
+struct paravirt_patch;
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end);
+
 #endif /* _I386_ALTERNATIVE_H */

===================================================================

include/asm-i386/irqflags.h +2 -2
@@ -79,8 +79,8 @@ static inline unsigned long __raw_local_
 }
 
 #else
-#define DISABLE_INTERRUPTS		cli
-#define ENABLE_INTERRUPTS		sti
+#define DISABLE_INTERRUPTS(clobbers)	cli
+#define ENABLE_INTERRUPTS(clobbers)	sti
 #define ENABLE_INTERRUPTS_SYSEXIT	sti; sysexit
 #define INTERRUPT_RETURN		iret
 #define GET_CR0_INTO_EAX		movl %cr0, %eax

===================================================================

include/asm-i386/paravirt.h +150 -37

214 lines not on this page · raw message

===================================================================

include/asm-i386/spinlock.h +4 -1
@@ -12,6 +12,7 @@
 #else
 #define CLI_STRING	"cli"
 #define STI_STRING	"sti"
+#define CLI_STI_CLOBBERS
 #endif /* CONFIG_PARAVIRT */
 
 /*
@@ -75,7 +76,9 @@ static inline void __raw_spin_lock_flags
 		"jg 1b\n\t"
 		"jmp 4b\n"
 		"5:\n\t"
-		: "+m" (lock->slock) : "r" (flags) : "memory");
+		: "+m" (lock->slock)
+		: "r" (flags)
+		: "memory" CLI_STI_CLOBBERS);
 }
 #endif
 
#24 lore raw

PATCH 3/7 paravirtualization: More generic paravirtualization entry point.

1) Each hypervisor writes a probe function to detect whether we are running under that hypervisor. paravirt_probe() registers this function.

2) If vmlinux is booted with ring != 0, we call all the probe functions (with registers except %esp intact) in link order: the winner will not return.

Cc Jeremy Fitzhardinge <[email protected]>
Cc Zachary Amsden <[email protected]>

===================================================================

arch/i386/kernel/Makefile +2 -0
@@ -39,6 +39,8 @@ obj-$(CONFIG_EARLY_PRINTK)	+= early_prin
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 obj-$(CONFIG_HPET_TIMER) 	+= hpet.o
 obj-$(CONFIG_K8_NB)		+= k8.o
+
+# Make sure this is linked after any other paravirt_ops structs: see head.S
 obj-$(CONFIG_PARAVIRT)		+= paravirt.o
 
 EXTRA_AFLAGS   := -traditional

===================================================================

arch/i386/kernel/head.S +33 -0
@@ -54,6 +54,12 @@
  * can.
  */
 ENTRY(startup_32)
+
+#ifdef CONFIG_PARAVIRT
+        movl %cs, %eax
+        testl $0x3, %eax
+        jnz startup_paravirt
+#endif
 
 /*
  * Set segments to known values.
@@ -486,6 +492,33 @@ ignore_int:
 #endif
 	iret
 
+#ifdef CONFIG_PARAVIRT
+startup_paravirt:
+	cld
+ 	movl $(init_thread_union+THREAD_SIZE),%esp
+
+	/* We take pains to preserve all the regs. */
+	pushl	%edx
+	pushl	%ecx
+	pushl	%eax
+
+	/* paravirt.o is last in link, and that probe fn never returns */
+	pushl	$__start_paravirtprobe
+1:
+	movl	0(%esp), %eax
+	pushl	(%eax)
+	movl	8(%esp), %eax
+	call	*(%esp)
+	popl	%eax
+
+	movl	4(%esp), %eax
+	movl	8(%esp), %ecx
+	movl	12(%esp), %edx
+
+	addl	$4, (%esp)
+	jmp	1b
+#endif
+
 /*
  * Real beginning of normal "text" segment
  */

===================================================================

arch/i386/kernel/paravirt.c +5 -1
@@ -19,6 +19,7 @@
 #include <linux/module.h>
 #include <linux/efi.h>
 #include <linux/bcd.h>
+#include <linux/start_kernel.h>
 
 #include <asm/bug.h>
 #include <asm/paravirt.h>
@@ -381,7 +382,10 @@ static int __init print_banner(void)
 	return 0;
 }
 core_initcall(print_banner);
- 
+
+/* We simply declare start_kernel to be the paravirt probe of last resort. */
+paravirt_probe(start_kernel);
+  
 struct paravirt_ops paravirt_ops = {
 	.name = "bare hardware",
 	.paravirt_enabled = 0,

===================================================================

arch/i386/kernel/vmlinux.lds.S +6 -0
@@ -60,6 +60,12 @@ SECTIONS
 	CONSTRUCTORS
 	} :data
 
+  __start_paravirtprobe = .;
+  .paravirtprobe : AT(ADDR(.paravirtprobe) - LOAD_OFFSET) {
+	*(.paravirtprobe)
+  }
+  __stop_paravirtprobe = .;
+
   . = ALIGN(4096);
   __nosave_begin = .;
   .data_nosave : AT(ADDR(.data_nosave) - LOAD_OFFSET) { *(.data.nosave) }

===================================================================

include/asm-i386/paravirt.h +5 -0
@@ -119,6 +119,11 @@ struct paravirt_ops
 	void (fastcall *irq_enable_sysexit)(void);
 	void (fastcall *iret)(void);
 };
+
+/* Mark a paravirt probe function. */
+#define paravirt_probe(fn)						\
+	static void (*__paravirtprobe_##fn)(void) __attribute_used__	\
+		__attribute__((__section__(".paravirtprobe"))) = fn
 
 extern struct paravirt_ops paravirt_ops;
 
#25 lore raw

PATCH 4/7 paravirtualization: Allow selected bug checks to be skipped by paravirt kernels

Allow selected bug checks to be skipped by paravirt kernels. The two most important are the F00F workaround (which is either done by the hypervisor, or not required), and the 'hlt' instruction check, which can break under some hypervisors.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>

===================================================================

arch/i386/kernel/cpu/intel.c +1 -1
@@ -107,7 +107,7 @@ static void __cpuinit init_intel(struct 
 	 * Note that the workaround only should be initialized once...
 	 */
 	c->f00f_bug = 0;
-	if ( c->x86 == 5 ) {
+	if (!paravirt_enabled() && c->x86 == 5) {
 		static int f00f_workaround_enabled = 0;
 
 		c->f00f_bug = 1;

===================================================================

include/asm-i386/bugs.h +4 -0
@@ -21,6 +21,7 @@
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/msr.h>
+#include <asm/paravirt.h>
 
 static int __init no_halt(char *s)
 {
@@ -91,6 +92,9 @@ static void __init check_fpu(void)
 
 static void __init check_hlt(void)
 {
+	if (paravirt_enabled())
+		return;
+
 	printk(KERN_INFO "Checking 'hlt' instruction... ");
 	if (!boot_cpu_data.hlt_works_ok) {
 		printk("disabled\n");
#26 lore raw

PATCH 5/7 paravirtualization: Allow disabling legacy power management modes with paravirt kernels

Two legacy power management modes are much easier to just explicitly disable when running in paravirtualized mode - neither APM nor PnP is still relevant. The status of ACPI is still debatable, and noacpi is still a common enough boot parameter that it is not necessary to explicitly disable ACPI.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>

===================================================================

arch/i386/kernel/apm.c +2 -1
@@ -231,6 +231,7 @@
 #include <asm/uaccess.h>
 #include <asm/desc.h>
 #include <asm/i8253.h>
+#include <asm/paravirt.h>
 
 #include "io_ports.h"
 
@@ -2191,7 +2192,7 @@ static int __init apm_init(void)
 
 	dmi_check_system(apm_dmi_table);
 
-	if (apm_info.bios.version == 0) {
+	if (apm_info.bios.version == 0 || paravirt_enabled()) {
 		printk(KERN_INFO "apm: BIOS not found.\n");
 		return -ENODEV;
 	}

===================================================================

drivers/pnp/pnpbios/core.c +2 -1
@@ -530,7 +530,8 @@ static int __init pnpbios_init(void)
 	if (check_legacy_ioport(PNPBIOS_BASE))
 		return -ENODEV;
 #endif
-	if (pnpbios_disabled || dmi_check_system(pnpbios_dmi_table)) {
+	if (pnpbios_disabled || dmi_check_system(pnpbios_dmi_table) ||
+	    paravirt_enabled()) {
 		printk(KERN_INFO "PnPBIOS: Disabled\n");
 		return -ENODEV;
 	}
#27 lore raw

PATCH 6/7 paravirtualization: Add APIC accessors to paravirt-ops.

Add APIC accessors to paravirt-ops. Unfortunately, we need two write functions, as some older broken hardware requires workarounds for Pentium APIC errata - this is the purpose of apic_write_atomic.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>

===================================================================

arch/i386/kernel/paravirt.c +8 -0
@@ -29,6 +29,8 @@
 #include <asm/time.h>
 #include <asm/irq.h>
 #include <asm/delay.h>
+#include <asm/fixmap.h>
+#include <asm/apic.h>
 
 /* nop stub */
 static void native_nop(void)
@@ -441,6 +443,12 @@ struct paravirt_ops paravirt_ops = {
 	.io_delay = native_io_delay,
 	.const_udelay = __const_udelay,
 
+#ifdef CONFIG_X86_LOCAL_APIC
+	.apic_write = native_apic_write,
+	.apic_write_atomic = native_apic_write_atomic,
+	.apic_read = native_apic_read,
+#endif
+
 	.irq_enable_sysexit = native_irq_enable_sysexit,
 	.iret = native_iret,
 };

===================================================================

include/asm-i386/apic.h +12 -3
@@ -37,18 +37,27 @@ extern void generic_apic_probe(void);
 /*
  * Basic functions accessing APICs.
  */
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+#define apic_write native_apic_write
+#define apic_write_atomic native_apic_write_atomic
+#define apic_read native_apic_read
+#endif
 
-static __inline void apic_write(unsigned long reg, unsigned long v)
+static __inline fastcall void native_apic_write(unsigned long reg,
+						unsigned long v)
 {
 	*((volatile unsigned long *)(APIC_BASE+reg)) = v;
 }
 
-static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+static __inline fastcall void native_apic_write_atomic(unsigned long reg,
+						       unsigned long v)
 {
 	xchg((volatile unsigned long *)(APIC_BASE+reg), v);
 }
 
-static __inline unsigned long apic_read(unsigned long reg)
+static __inline fastcall unsigned long native_apic_read(unsigned long reg)
 {
 	return *((volatile unsigned long *)(APIC_BASE+reg));
 }

===================================================================

include/asm-i386/paravirt.h +27 -0
@@ -114,6 +114,12 @@ struct paravirt_ops
 
 	void (fastcall *io_delay)(void);
 	void (*const_udelay)(unsigned long loops);
+
+#ifdef CONFIG_X86_LOCAL_APIC
+	void (fastcall *apic_write)(unsigned long reg, unsigned long v);
+	void (fastcall *apic_write_atomic)(unsigned long reg, unsigned long v);
+	unsigned long (fastcall *apic_read)(unsigned long reg);
+#endif
 
 	/* These two are jmp to, not actually called. */
 	void (fastcall *irq_enable_sysexit)(void);
@@ -275,6 +281,27 @@ static inline void slow_down_io(void) {
 #endif
 }
 
+#ifdef CONFIG_X86_LOCAL_APIC
+/*
+ * Basic functions accessing APICs.
+ */
+static __inline void apic_write(unsigned long reg, unsigned long v)
+{
+	paravirt_ops.apic_write(reg,v);
+}
+
+static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+{
+	paravirt_ops.apic_write_atomic(reg,v);
+}
+
+static __inline unsigned long apic_read(unsigned long reg)
+{
+	return paravirt_ops.apic_read(reg);
+}
+#endif
+
+
 /* These all sit in the .parainstructions section to tell us what to patch. */
 struct paravirt_patch {
 	u8 *instr; 		/* original instructions */
#28 lore raw

PATCH 7/7 paravirtualization: Add mmu virtualization to paravirt-ops.

Add the three bare TLB accessor functions to paravirt-ops. Most amusingly, flush_tlb is redefined on SMP, so I can't call the paravirt op flush_tlb. Instead, I chose to indicate the actual flush type, kernel (global) vs. user (non-global). Global in this sense means using the global bit in the page table entry, which makes TLB entries persistent across CR3 reloads, not global as in the SMP sense of invoking remote shootdowns, so the term is confusingly overloaded.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>

===================================================================

arch/i386/kernel/paravirt.c +109 -0
@@ -31,6 +31,7 @@
 #include <asm/delay.h>
 #include <asm/fixmap.h>
 #include <asm/apic.h>
+#include <asm/tlbflush.h>
 
 /* nop stub */
 static void native_nop(void)
@@ -373,6 +374,97 @@ static fastcall void native_io_delay(voi
 {
 	asm volatile("outb %al,$0x80");
 }
+
+static fastcall void native_flush_tlb(void)
+{
+	__native_flush_tlb();
+}
+
+/*
+ * Global pages have to be flushed a bit differently. Not a real
+ * performance problem because this does not happen often.
+ */
+static fastcall void native_flush_tlb_global(void)
+{
+	__native_flush_tlb_global();
+}
+
+static fastcall void native_flush_tlb_single(u32 addr)
+{
+	__native_flush_tlb_single(addr);
+}
+
+#ifndef CONFIG_X86_PAE
+static fastcall void native_set_pte(pte_t *ptep, pte_t pteval)
+{
+	*ptep = pteval;
+}
+
+static fastcall void native_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pteval)
+{
+	*ptep = pteval;
+}
+
+static fastcall void native_set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	*pmdp = pmdval;
+}
+
+#else /* CONFIG_X86_PAE */
+
+static fastcall void native_set_pte(pte_t *ptep, pte_t pte)
+{
+	ptep->pte_high = pte.pte_high;
+	smp_wmb();
+	ptep->pte_low = pte.pte_low;
+}
+
+static fastcall void native_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pte)
+{
+	ptep->pte_high = pte.pte_high;
+	smp_wmb();
+	ptep->pte_low = pte.pte_low;
+}
+
+static fastcall void native_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
+{
+	ptep->pte_low = 0;
+	smp_wmb();
+	ptep->pte_high = pte.pte_high;
+	smp_wmb();
+	ptep->pte_low = pte.pte_low;
+}
+
+static fastcall void native_set_pte_atomic(pte_t *ptep, pte_t pteval)
+{
+	set_64bit((unsigned long long *)ptep,pte_val(pteval));
+}
+
+static fastcall void native_set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	set_64bit((unsigned long long *)pmdp,pmd_val(pmdval));
+}
+
+static fastcall void native_set_pud(pud_t *pudp, pud_t pudval)
+{
+	*pudp = pudval;
+}
+
+static fastcall void native_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
+{
+	ptep->pte_low = 0;
+	smp_wmb();
+	ptep->pte_high = 0;
+}
+
+static fastcall void native_pmd_clear(pmd_t *pmd)
+{
+	u32 *tmp = (u32 *)pmd;
+	*tmp = 0;
+	smp_wmb();
+	*(tmp + 1) = 0;
+}
+#endif /* CONFIG_X86_PAE */
 
 /* These are in entry.S */
 extern fastcall void native_iret(void);
@@ -449,6 +541,23 @@ struct paravirt_ops paravirt_ops = {
 	.apic_read = native_apic_read,
 #endif
 
+	.flush_tlb_user = native_flush_tlb,
+	.flush_tlb_kernel = native_flush_tlb_global,
+	.flush_tlb_single = native_flush_tlb_single,
+
+	.set_pte = native_set_pte,
+	.set_pte_at = native_set_pte_at,
+	.set_pmd = native_set_pmd,
+	.pte_update = (void *)native_nop,
+	.pte_update_defer = (void *)native_nop,
+#ifdef CONFIG_X86_PAE
+	.set_pte_atomic = native_set_pte_atomic,
+	.set_pte_present = native_set_pte_present,
+	.set_pud = native_set_pud,
+	.pte_clear = native_pte_clear,
+	.pmd_clear = native_pmd_clear,
+#endif
+
 	.irq_enable_sysexit = native_irq_enable_sysexit,
 	.iret = native_iret,
 };

===================================================================

arch/i386/mm/boot_ioremap.c +1 -0
@@ -16,6 +16,7 @@
  */
 
 #undef CONFIG_X86_PAE
+#undef CONFIG_PARAVIRT
 #include <asm/page.h>
 #include <asm/pgtable.h>
 #include <asm/tlbflush.h>

===================================================================

include/asm-i386/paravirt.h +75 -0

99 lines not on this page · raw message

===================================================================

include/asm-i386/pgtable-2level.h +4 -1
@@ -11,11 +11,14 @@
  * within a page table are directly modified.  Thus, the following
  * hook is made available.
  */
+#ifndef CONFIG_PARAVIRT
 #define set_pte(pteptr, pteval) (*(pteptr) = pteval)
 #define set_pte_at(mm,addr,ptep,pteval) set_pte(ptep,pteval)
+#define set_pmd(pmdptr, pmdval) (*(pmdptr) = (pmdval))
+#endif
+
 #define set_pte_atomic(pteptr, pteval) set_pte(pteptr,pteval)
 #define set_pte_present(mm,addr,ptep,pteval) set_pte_at(mm,addr,ptep,pteval)
-#define set_pmd(pmdptr, pmdval) (*(pmdptr) = (pmdval))
 
 #define pte_clear(mm,addr,xp)	do { set_pte_at(mm, addr, xp, __pte(0)); } while (0)
 #define pmd_clear(xp)	do { set_pmd(xp, __pmd(0)); } while (0)

===================================================================

include/asm-i386/pgtable-3level.h +28 -26

72 lines not on this page · raw message

===================================================================

include/asm-i386/pgtable.h +3 -1
@@ -15,6 +15,7 @@
 #include <asm/processor.h>
 #include <asm/fixmap.h>
 #include <linux/threads.h>
+#include <asm/paravirt.h>
 
 #ifndef _I386_BITOPS_H
 #include <asm/bitops.h>
@@ -246,6 +247,7 @@ static inline pte_t pte_mkhuge(pte_t pte
 # include <asm/pgtable-2level.h>
 #endif
 
+#ifndef CONFIG_PARAVIRT
 /*
  * Rules for using pte_update - it must be called after any PTE update which
  * has not been done using the set_pte / clear_pte interfaces.  It is used by
@@ -261,7 +263,7 @@ static inline pte_t pte_mkhuge(pte_t pte
  */
 #define pte_update(mm, addr, ptep)		do { } while (0)
 #define pte_update_defer(mm, addr, ptep)	do { } while (0)
-
+#endif
 
 /*
  * We only update the dirty/accessed state if we set

===================================================================

include/asm-i386/tlbflush.h +13 -5
@@ -4,7 +4,15 @@
 #include <linux/mm.h>
 #include <asm/processor.h>
 
-#define __flush_tlb()							\
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+#define __flush_tlb() __native_flush_tlb()
+#define __flush_tlb_global() __native_flush_tlb_global()
+#define __flush_tlb_single(addr) __native_flush_tlb_single(addr)
+#endif
+
+#define __native_flush_tlb()						\
 	do {								\
 		unsigned int tmpreg;					\
 									\
@@ -19,7 +27,7 @@
  * Global pages have to be flushed a bit differently. Not a real
  * performance problem because this does not happen often.
  */
-#define __flush_tlb_global()						\
+#define __native_flush_tlb_global()					\
 	do {								\
 		unsigned int tmpreg, cr4, cr4_orig;			\
 									\
@@ -36,6 +44,9 @@
 			: "memory");					\
 	} while (0)
 
+#define __native_flush_tlb_single(addr) 				\
+	__asm__ __volatile__("invlpg (%0)" ::"r" (addr) : "memory")
+
 # define __flush_tlb_all()						\
 	do {								\
 		if (cpu_has_pge)					\
@@ -45,9 +56,6 @@
 	} while (0)
 
 #define cpu_has_invlpg	(boot_cpu_data.x86 > 3)
-
-#define __flush_tlb_single(addr) \
-	__asm__ __volatile__("invlpg (%0)" ::"r" (addr) : "memory")
 
 #ifdef CONFIG_X86_INVLPG
 # define __flush_tlb_one(addr) __flush_tlb_single(addr)
#29 lore raw

Re: [PATCH 6/7] paravirtualization: Add APIC accessors to paravirt-ops.

In reply to Rusty Russell

include/asm-i386/paravirt.h static inline void slow_down_io(void) { · line 298

Quoted 3 lines · On Wed, 01 Nov 2006 21:32:30 +1100 Rusty Russell <[email protected]> wrote:
include/asm-i386/paravirt.h +3 -0
+static __inline void apic_write(unsigned long reg, unsigned long v)
+static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+static __inline unsigned long apic_read(unsigned long reg)

Just `inline', please.

#30 lore raw

Re: [PATCH 6/7] paravirtualization: Add APIC accessors to paravirt-ops.

Quoted 8 lines · Andrew Morton

Review

Quoted 3 lines · On Wed, 01 Nov 2006 21:32:30 +1100 Rusty Russell <[email protected]> wrote:
(quoted diff) +3 -0
+static __inline void apic_write(unsigned long reg, unsigned long v)
+static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+static __inline unsigned long apic_read(unsigned long reg)

Just `inline', please.

akpm says: "Just `inline', please."

include/asm-i386/paravirt.h +3 -3
diff -r 3a3bc9aed04c include/asm-i386/paravirt.h
@@ -304,17 +304,17 @@ static inline void slow_down_io(void) {
 /*
  * Basic functions accessing APICs.
  */
-static __inline void apic_write(unsigned long reg, unsigned long v)
+static inline void apic_write(unsigned long reg, unsigned long v)
 {
 	paravirt_ops.apic_write(reg,v);
 }
 
-static __inline void apic_write_atomic(unsigned long reg, unsigned long v)
+static inline void apic_write_atomic(unsigned long reg, unsigned long v)
 {
 	paravirt_ops.apic_write_atomic(reg,v);
 }
 
-static __inline unsigned long apic_read(unsigned long reg)
+static inline unsigned long apic_read(unsigned long reg)
 {
 	return paravirt_ops.apic_read(reg);
 }
#31 lore raw

Re: [PATCH 4/7] paravirtualization: Allow selected bug checks to be skipped by paravirt kernels

In reply to Rusty Russell

include/asm-i386/bugs.h line 24

Quoted 7 lines · On Wed, 01 Nov 2006 21:30:43 +1100 Rusty Russell <[email protected]> wrote:
include/asm-i386/bugs.h +1 -0
@@ -21,6 +21,7 @@
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/msr.h>
+#include <asm/paravirt.h>

In many other places you have

#ifdef CONFIG_PARAVIRT
#include <asm/paravirt.h>
...

But not here.

Making <asm/paravirt.h> invulnerable would be the more typical approach.

#32 lore raw

Re: [PATCH 4/7] paravirtualization: Allow selected bug checks to be skipped by paravirt kernels

Quoted 15 lines · Andrew Morton

include/asm-i386/bugs.h line 25

Quoted 8 lines · On Wed, 01 Nov 2006 21:30:43 +1100 Rusty Russell <[email protected]> wrote:
include/asm-i386/bugs.h +1 -0
@@ -21,6 +21,7 @@
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/msr.h>
+#include <asm/paravirt.h>
 

In many other places you have

Quoted 7 lines

#ifdef CONFIG_PARAVIRT
#include <asm/paravirt.h>
...

But not here.

Making <asm/paravirt.h> invulnerable would be the more typical approach.

CONFIG_PARAVIRT is not being used to guard asm/paravirt.h from multiple inclusion. In places where it is being used to guard #include <asm/paravirt.h>, the idea is that asm/paravirt.h defines various inlines/macros which would otherwise be defined in the header. So, for example, asm/desc.h would normally define load_gdt() in the !CONFIG_PARAVIRT case, but asm/paravirt.h defines it when CONFIG_PARAVIRT is enabled.

In this case, asm/paravirt.h included because we need the definition for paravirt_enabled(), not because it is replacing any of bugs.h's definitions.

J

#33 lore raw

Re: [PATCH 4/7] paravirtualization: Allow selected bug checks to be skipped by paravirt kernels

In reply to Andrew Morton
Quoted 13 lines · Andrew Morton

include/asm-i386/bugs.h line 24

Quoted 7 lines · On Wed, 01 Nov 2006 21:30:43 +1100 Rusty Russell <[email protected]> wrote:
include/asm-i386/bugs.h +1 -0
@@ -21,6 +21,7 @@
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/msr.h>
+#include <asm/paravirt.h>

In many other places you have

Quoted 7 lines

#ifdef CONFIG_PARAVIRT
#include <asm/paravirt.h>
...

But not here.

Making <asm/paravirt.h> invulnerable would be the more typical approach.

It *is* actually safe. The "#ifdef CONFIG_PARAVIRT / #include <asm/paravirt.h> / #else / <... native versions...>" is to give a big hint to the reader to look in paravirt.h for the real definitions.

Originally I had a noparavirt.h where all these lived, and people hated it. So we did it this way, which minimizes churn.

Rusty.

#34 lore raw

Re: [PATCH 2/7] paravirtualization: Patch inline replacements for common paravirt operations.

In reply to Rusty Russell

arch/i386/kernel/alternative.c void alternatives_smp_switch(int smp) · line 363

Quoted 12 lines · On Wed, 01 Nov 2006 21:28:13 +1100 Rusty Russell <[email protected]> wrote:
arch/i386/kernel/alternative.c +12 -0
+void apply_paravirt(struct paravirt_patch *start, struct paravirt_patch *end)
+{
+	struct paravirt_patch *p;
+	int i;
+
+	for (p = start; p < end; p++) {
+		unsigned int used;
+
+		used = paravirt_ops.patch(p->instrtype, p->clobbers, p->instr,
+					  p->len);
+#ifdef CONFIG_DEBUG_KERNEL
+		/* Deliberately clobber regs using "not %reg" to find bugs. */

That would be considered to be abusive of CONFIG_DEBUG_KERNEL. A CONFIG_DEBUG_PARAVIRT which depends on CONFIG_DEBUG_KERNEL would be more harmonious.

#35 lore raw

Re: [PATCH 2/7] paravirtualization: Patch inline replacements for common paravirt operations.

Quoted 8 lines · Andrew Morton

Review

Quoted 2 lines · On Wed, 01 Nov 2006 21:28:13 +1100 Rusty Russell <[email protected]> wrote:
(quoted diff) +2 -0
+#ifdef CONFIG_DEBUG_KERNEL
+		/* Deliberately clobber regs using "not %reg" to find bugs. */

That would be considered to be abusive of CONFIG_DEBUG_KERNEL. A CONFIG_DEBUG_PARAVIRT which depends on CONFIG_DEBUG_KERNEL would be more harmonious.

I wasn't sure. Making a config option for what is a one-liner seemed overkill.

==

Don't abuse CONFIG_DEBUG_KERNEL, add CONFIG_DEBUG_PARAVIRT.

arch/i386/Kconfig.debug +10 -0
diff -r 2707c89d72f0 arch/i386/Kconfig.debug
@@ -87,4 +87,14 @@ config DOUBLEFAULT
           option saves about 4k and might cause you much additional grey
           hair.
 
+config DEBUG_PARAVIRT
+	bool "Enable some paravirtualization debugging"
+	default y
+	depends on PARAVIRT && DEBUG_KERNEL
+	help
+	  Currently deliberately clobbers regs which are allowed to be
+	  clobbered in inlined paravirt hooks, even in native mode.
+	  If turning this off solves a problem, then DISABLE_INTERRUPTS() or
+	  ENABLE_INTERRUPTS() is lying about what registers can be clobbered.
+
 endmenu
arch/i386/kernel/alternative.c +1 -1
diff -r 2707c89d72f0 arch/i386/kernel/alternative.c
@@ -359,7 +359,7 @@ void apply_paravirt(struct paravirt_patc
 
 		used = paravirt_ops.patch(p->instrtype, p->clobbers, p->instr,
 					  p->len);
-#ifdef CONFIG_DEBUG_KERNEL
+#ifdef CONFIG_DEBUG_PARAVIRT
 		/* Deliberately clobber regs using "not %reg" to find bugs. */
 		for (i = 0; i < 3; i++) {
 			if (p->len - used >= 2 && (p->clobbers & (1 << i))) {
#36 lore raw

Re: [PATCH 2/7] paravirtualization: Patch inline replacements for common paravirt operations.

Quoted 13 lines · Rusty Russell
Quoted 10 lines · Andrew Morton

Review

Quoted 3 lines · On Wed, 01 Nov 2006 21:28:13 +1100 Rusty Russell <[email protected]> wrote:
(quoted diff) +2 -0
+#ifdef CONFIG_DEBUG_KERNEL
+		/* Deliberately clobber regs using "not %reg" to find bugs. */

That would be considered to be abusive of CONFIG_DEBUG_KERNEL. A CONFIG_DEBUG_PARAVIRT which depends on CONFIG_DEBUG_KERNEL would be more harmonious.

Quoted 3 lines

I wasn't sure. Making a config option for what is a one-liner seemed overkill.

I have further stuff in my vmi-debug patch that can use CONFIG_DEBUG_PARAVIRT as well :)

Zach

#37 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Rusty Russell
Quoted 11 lines · Rusty Russell

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

This patch does the dumbest possible replacement of paravirtualized
instructions: calls through a "paravirt_ops" structure. Currently these are function implementations of native hardware: hypervisors will override the ops structure with their own variants.

All the pv-ops functions are declared "fastcall" so that a specific register-based ABI is used, to make inlining assember easier.

this is a lot of infrastructure... do we have more than 1 user of this yet that wants to get merged in mainline?

Signature · 2 lines
if you want to mail me at work (you don't), use arjan (at) linux.intel.com
Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org
#38 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 2 lines · Arjan van de Ven

this is a lot of infrastructure... do we have more than 1 user of this yet that wants to get merged in mainline?

AFAIK xen, vmi, lhype (and native ops).

-Andi

#39 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Arjan van de Ven
Quoted 6 lines · Arjan van de Ven
Quoted 3 lines · Rusty Russell

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

this is a lot of infrastructure... do we have more than 1 user of this yet that wants to get merged in mainline?

Yep. Xen and VMI both have patches on top of this pending merge. I also have a toy hypervisor "lhype" based on this, but it's not ready for mainline. (It seems people expect consoles to do *input* as well as output).

Cheers,
Rusty.

#40 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Rusty Russell

This patch breaks `make headers_check' in mysterious ways:

  CHECK   include/linux/netfilter_ipv4/ip_conntrack_tcp.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_sctp.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_protocol.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_helper_h323_types.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_helper_h323_asn1.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_helper.h

make[2]: *** [/usr/src/devel/usr/include/asm/.check.setup.h] Error 1 make[2]: *** Waiting for unfinished jobs....
make[1]: *** [asm-i386] Error 2
make[1]: *** Waiting for unfinished jobs....
make: *** [headers_check] Error 2

#41 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 8 lines · Andrew Morton

This patch breaks `make headers_check' in mysterious ways:

  CHECK   include/linux/netfilter_ipv4/ip_conntrack_tcp.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_sctp.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_protocol.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_helper_h323_types.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_helper_h323_asn1.h
  CHECK   include/linux/netfilter_ipv4/ip_conntrack_helper.h
Quoted 5 lines

make[2]: *** [/usr/src/devel/usr/include/asm/.check.setup.h] Error 1 make[2]: *** Waiting for unfinished jobs....
make[1]: *** [asm-i386] Error 2
make[1]: *** Waiting for unfinished jobs....
make: *** [headers_check] Error 2

It seems like missing
"header-y += paravirt.h" in the "include/asm-i386/Kbuild".
____

#42 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Rusty Russell
Quoted 3 lines · Rusty Russell

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

Hmm, did this all ever compile in mainline? I had to do a few merges and in the end i get

/home/lsrc/quilt/linux/kernel/spinlock.c: In function ‘_spin_lock_irqsave’:
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
{standard input}: Assembler messages:
{standard input}:593: Error: undefined symbol paravirt_ops' in operation {standard input}:593: Error: undefined symbol PARAVIRT_irq_enable' in operation {standard input}:605: Error: undefined symbol paravirt_ops' in operation {standard input}:605: Error: undefined symbol PARAVIRT_irq_disable' in operatio n

and lots of new warnings like

/home/lsrc/quilt/linux/arch/i386/kernel/traps.c: In function ‘set_intr_gate’:
/home/lsrc/quilt/linux/arch/i386/kernel/traps.c:1165: warning: implicit declarat ion of function ‘_set_gate’ /home/lsrc/quilt/linux/arch/i386/kernel/cpu/common.c: In function ‘_cpu_init’:
/home/lsrc/quilt/linux/arch/i386/kernel/cpu/common.c:754: warning: implicit decl aration of function ‘__set_tss_desc’ /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c: In function ‘intel_mach ine_check’:
/home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.eax’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.ebx’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.ecx’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.edx’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.esi’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.edi’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.ebp’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.esp’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.eflag s’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.eip’ may be used uninitialized in this function

This is with i386 defconfig + CONFIG_PARAVIRT

-Andi

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/virtualization

#43 lore raw

[x86_64] Strange oprofile results on access to per_cpu data

Hi Andi

While doing some oprofile analysis, I got this result on ip_route_input() : one particular instruction seems to spend a lot of cycles.

machine is a dual core 285, 2.6 GHz

/*
* Command line: opannotate -a event:CPU_CLK_UNHALTED
/usr/src/linux-2.6.18/vmlinux
*
* Interpretation of command line:
* Output annotated assembly listing with samples
*
* CPU: AMD64 processors, speed 2600.01 MHz (estimated)
* Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit
mask of 0x00 (No unit mask) count 10000
*/

ffffffff803e9860 <ip_route_input>: /* ip_route_input total: 543098 2.5487 */

/* relevant extract from ip_route_input() */

   600  0.0028 :ffffffff803e98b3:       mov    $0xffffffff806375e0,%rsi
   883  0.0041 :ffffffff803e98ba:       mov    %rax,%rdx
     6 2.8e-05 :ffffffff803e98bd:       mov    %rsi,%rcx
  2281  0.0107 :ffffffff803e98c0:       cmp    0xf0(%rdx),%r12d
  9767  0.0458 :ffffffff803e98c7:       jne    ffffffff803e98f1 
<ip_route_input+0x91>
   108 5.1e-04 :ffffffff803e98c9:       cmp    0xf4(%rdx),%r14d
 41459  0.1946 :ffffffff803e98d0:       jne    ffffffff803e98f1 
<ip_route_input+0x91>
   549  0.0026 :ffffffff803e98d2:       cmp    0xec(%rdx),%ebx
 88604  0.4158 :ffffffff803e98d8:       jne    ffffffff803e98f1 
<ip_route_input+0x91>
   478  0.0022 :ffffffff803e98da:       mov    0xe8(%rdx),%eax
   315  0.0015 :ffffffff803e98e0:       test   %eax,%eax
   241  0.0011 :ffffffff803e98e2:       jne    ffffffff803e98f1 

<ip_route_input+0x91>
248 0.0012 :ffffffff803e98e4: cmp 0xfc(%rdx),%r13b

2314 0.0109 :ffffffff803e98eb: je ffffffff803ea3b3
################ BEGIN

   370  0.0017 :ffffffff803e98f1:       mov    %gs:0x8,%rax
222769  1.0454 :ffffffff803e98fa:       incl   0x38(%rcx,%rax,1)
################ END
     6 2.8e-05 :ffffffff803e98fe:       mov    (%rdx),%rdx
   833  0.0039 :ffffffff803e9901:       test   %rdx,%rdx

__raw_get_cpu_var(rt_cache_stat).field++ appears to be very expensive

(about 18000 RT_CACHE_STAT_INC(in_hlist_search); are done per second, not an impressive count in fact)

Are segment prefixes that expensive ?
Or is it only the first access to %gs:8 that is doing extra checks ? (because other RT_CACHE_STAT_INC() done in the same function dont have this cost) Or is it the loading of %rcx (done in ffffffff803e98bd) that is stalling ?

I was wondering if avoiding a dependancy would help :

As we dont have TLS support in kernel yet, I was considering trying (just for experimentation) to stick a struct rt_cache_stat in pda, since it avoids one step.

#if defined(RT_CACHE_STAT_IN_PDA)

# define RT_CACHE_STAT_INC(field) add_pda(rt_cache_stat.field, 1)
# define addr_of_rt_cache_stat(cpu) &cpu_pda(cpu)->rt_cache_stat
#else
  static DEFINE_PER_CPU(struct rt_cache_stat, rt_cache_stat);
# define RT_CACHE_STAT_INC(field) (__raw_get_cpu_var(rt_cache_stat).field++)
# define addr_of_rt_cache_stat(cpu) &per_cpu(rt_cache_stat, cpu)
#endif

so that RT_CACHE_STAT_INC(field) would map to

addl $1,%gs:OFFSET /* no register needed */

Thank you
Eric

#44 lore raw

Re: [x86_64] Strange oprofile results on access to per_cpu data

Quoted 6 lines · Eric Dumazet

Hi Andi

While doing some oprofile analysis, I got this result on ip_route_input() :
one particular instruction seems to spend a lot of cycles.

machine is a dual core 285, 2.6 GHz

Single socket?

Quoted 8 lines

################ BEGIN

    370  0.0017 :ffffffff803e98f1:       mov    %gs:0x8,%rax
222769  1.0454 :ffffffff803e98fa:       incl   0x38(%rcx,%rax,1)
################ END
      6 2.8e-05 :ffffffff803e98fe:       mov    (%rdx),%rdx
    833  0.0039 :ffffffff803e9901:       test   %rdx,%rdx

__raw_get_cpu_var(rt_cache_stat).field++ appears to be very expensive

First the profile events are not exact. It could be an earlier instruction. The reordering Window is relatively large (~80 macro ops)

But let's assume it is this one.

Weird. Maybe it's a cache miss. Can you check with DATA_CACHE_MISSES ? Or possible a TLB miss, although that's far less likely (L1_AND_L2_DTLB_MISSES)

Quoted 4 lines

(about 18000 RT_CACHE_STAT_INC(in_hlist_search); are done per second, not an impressive count in fact)

Are segment prefixes that expensive ?

No, they are supposed to be one cycle only.

Quoted 3 lines

Or is it only the first access to %gs:8 that is doing extra checks ? (because other RT_CACHE_STAT_INC() done in the same function dont have this cost) is it the loading of %rcx (done in ffffffff803e98bd) that is stalling ?

My first guess would be a cache miss of some sort.

Maybe the rest of the code needs enough cache to push this line out. Or since you got dual core with separated caches they are bouncing for some reason (they shouldn't, but maybe there is a bug)

Quoted 3 lines

so that RT_CACHE_STAT_INC(field) would map to

addl $1,%gs:OFFSET /* no register needed */

I doubt that will make much difference.

-Andi

#45 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Andi Kleen
Quoted 25 lines · Andi Kleen
Quoted 4 lines · Rusty Russell

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

Hmm, did this all ever compile in mainline? I had to do a few merges and in the end i get

/home/lsrc/quilt/linux/kernel/spinlock.c: In function ‘_spin_lock_irqsave’:
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter

Quoted 9 lines

include2/asm/spinlock.h:59: error: invalid 'asm': operand number missing after % -letter
{standard input}: Assembler messages:
{standard input}:593: Error: undefined symbol paravirt_ops' in operation {standard input}:593: Error: undefined symbol PARAVIRT_irq_enable' in operation {standard input}:605: Error: undefined symbol paravirt_ops' in operation {standard input}:605: Error: undefined symbol PARAVIRT_irq_disable' in operatio n

Not seeing that here (on 2.6.19-rc2-mm2 with gcc 4.0.2).

Quoted 9 lines

and lots of new warnings like

/home/lsrc/quilt/linux/arch/i386/kernel/traps.c: In function ‘set_intr_gate’:
/home/lsrc/quilt/linux/arch/i386/kernel/traps.c:1165: warning: implicit declarat ion of function ‘_set_gate’ /home/lsrc/quilt/linux/arch/i386/kernel/cpu/common.c: In function ‘_cpu_init’:
/home/lsrc/quilt/linux/arch/i386/kernel/cpu/common.c:754: warning: implicit decl aration of function ‘__set_tss_desc'

Sounds like desc.h got reordered. Somewhere, there was a broken patch once that did this, I thought we fixed that.

Quoted 5 lines

/home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c: In function ‘intel_mach
ine_check’:
/home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.eax’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.ebx’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.ecx’ may be used uninitialized in this function

Quoted 9 lines

/home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.edx’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.esi’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.edi’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.ebp’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.esp’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.eflag
s’ may be used uninitialized in this function /home/lsrc/quilt/linux/arch/i386/kernel/cpu/mcheck/p4.c:158: warning: ‘dbg.eip’ may be used uninitialized in this function

Those appear to be valid warnings, with or without paravirt, due to the tacky glued inline oddity of intel_get_extended_msrs.

Zach

#46 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 4 lines

Sounds like desc.h got reordered. Somewhere, there was a broken patch once that did this, I thought we fixed that.

I think I got Rusty's latest patches that I found in my mailbox.

I haven't looked at desc.h, but at least processor.h ordering was totally b0rken (e.g. #define __cpuid native_cpuid was after several uses). I fixed that to make at least the CONFIG_PARAVIRT not set case compile.

I can't see how this ever worked either.

Haven't attempted the CONFIG_PARAVIRT case which apparently needs more work (it is currently marked CONFIG_BROKEN)

Can someone double check this is the correct patchkit?

ftp://ftp.frstfloor.org/pub/ak/x86_64/quilt/patches/paravirt*

-Andi

#47 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 10 lines · Andi Kleen
Quoted 3 lines

Sounds like desc.h got reordered. Somewhere, there was a broken patch once that did this, I thought we fixed that.

I think I got Rusty's latest patches that I found in my mailbox.

I haven't looked at desc.h, but at least processor.h ordering was totally b0rken (e.g. #define __cpuid native_cpuid was after several uses). I fixed that to make at least the CONFIG_PARAVIRT not set case compile.

Quoted 8 lines

I can't see how this ever worked either.

Haven't attempted the CONFIG_PARAVIRT case which apparently needs more work (it is currently marked CONFIG_BROKEN)

Can someone double check this is the correct patchkit?

ftp://ftp.frstfloor.org/pub/ak/x86_64/quilt/patches/paravirt*

Andi, the patches work against Andrew's tree, and he's merged them in rc4-mm2. There are a few warnings to clean up, but it seems basically sound.

At this point I our think time is better spent on beating those patches up, rather than going back and figuring out why they don't work in your tree.

Sorry,
Rusty.

#48 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 8 lines · Rusty Russell

Andi, the patches work against Andrew's tree, and he's merged them in rc4-mm2. There are a few warnings to clean up, but it seems basically sound.

At this point I our think time is better spent on beating those patches up, rather than going back and figuring out why they don't work in your tree.

This begs the question - should we rebase the paravirt-ops patchset against -rc4-mm2? I almost did it today, but didn't want to stomp on anybody else's toes.

Zach

#49 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 5 lines · Zachary Amsden
Quoted 4 lines · Rusty Russell

Andi, the patches work against Andrew's tree, and he's merged them in rc4-mm2. There are a few warnings to clean up, but it seems basically sound.

Quoted 8 lines
Quoted 4 lines

At this point I our think time is better spent on beating those patches up, rather than going back and figuring out why they don't work in your tree.

This begs the question - should we rebase the paravirt-ops patchset against -rc4-mm2? I almost did it today, but didn't want to stomp on anybody else's toes.

Yes. Andrew has shot me a couple of warnings which people have found, and I'm preparing patches for them. Rebasing will make it easier.

If you're not awake now, I'll do it. If you are, see me on IRC.

Rusty.

#50 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Rusty Russell
Quoted 8 lines

Andi, the patches work against Andrew's tree, and he's merged them in rc4-mm2. There are a few warnings to clean up, but it seems basically sound.

At this point I our think time is better spent on beating those patches up, rather than going back and figuring out why they don't work in your tree.

My tree is basically mainline as base. Sure if you don't care about mainline merges we can ignore it there and keep it forever in -mm* until Andrew gets tired of it?

That's a possible strategy, but only if you want to keep it as a mm-only toy forever.

-Andi

#51 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 9 lines · On Sun, 5 Nov 2006 06:46:15 +0100 Andi Kleen <[email protected]> wrote:
Quoted 8 lines

Andi, the patches work against Andrew's tree, and he's merged them in rc4-mm2. There are a few warnings to clean up, but it seems basically sound.

At this point I our think time is better spent on beating those patches up, rather than going back and figuring out why they don't work in your tree.

Quoted 7 lines

My tree is basically mainline as base. Sure if you don't care about mainline merges we can ignore it there and keep it forever in -mm* until Andrew gets tired of it?

That's a possible strategy, but only if you want to keep it as a mm-only toy forever.

They're in my regular list-of-thing-to-spam-maintainers-with, so we can transfer them as-is next week sometime.

It would be better to sort out the various warnings and any other nasties first though.

#52 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Andi Kleen
Quoted 8 lines · Andi Kleen
Quoted 7 lines

Andi, the patches work against Andrew's tree, and he's merged them in rc4-mm2. There are a few warnings to clean up, but it seems basically sound.

At this point I our think time is better spent on beating those patches up, rather than going back and figuring out why they don't work in your tree.

Quoted 6 lines

My tree is basically mainline as base. Sure if you don't care about mainline merges we can ignore it there and keep it forever in -mm* until Andrew gets tired of it?

That's a possible strategy, but only if you want to keep it as a mm-only toy forever.

Andi, it's been simpler for us to get the code into Andrew's tree, in nice bit-size pieces. We've had trouble every time we've tried to get stuff into your tree. In addition, Andrew's tree gives the code exposure and testing.

If Andrew says we have to get those patches into mainline through you, then I'll spend all that time re-spinning the patches for you from the -mm tree until they go in. It doesn't seem like a good use of anyone's time though.

Rusty.

#53 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

Quoted 3 lines

If Andrew says we have to get those patches into mainline through you,

Well I'm mainline in this case.

Quoted 2 lines

then I'll spend all that time re-spinning the patches for you from the -mm tree until they go in.

I got it to compile now with this patch (+ one additional patch that is folded in). It then goes through kernel initialization and then init gets killed with "Inconsistency detected by rtld.c:1250:
Assertation ph_vaddr == _rtld_local.dl_sysinfo_vdso failed"

It looks like some of the ifdefs were placed completely wrong and in addition you were missing a patch to include asm/offset.h everywhere as assembly (I patched around that). And two macros were apparently never compiled in their current form.

But it seems it is dependent on even more -mm* magic than just that. If you can identify the missing patches that make init's rtld work again that would be useful.

-Andi

Get paravirt ops to compile

TBD should be folded into the original patches

Unfortunately still doesn't boot.

include/asm-i386/desc.h +3 -3
Index: linux/include/asm-i386/desc.h
===================================================================
@@ -92,6 +92,9 @@ static inline void write_dt_entry(void *
 #define write_gdt_entry(dt, entry, a, b) write_dt_entry(dt, entry, a, b)
 #define write_idt_entry(dt, entry, a, b) write_dt_entry(dt, entry, a, b)
 
+#define set_ldt native_set_ldt
+#endif /* CONFIG_PARAVIRT */
+
 static inline void _set_gate(int gate, unsigned int type, void *addr, unsigned short seg)
 {
 	__u32 a, b;
@@ -108,9 +111,6 @@ static inline void __set_tss_desc(unsign
 	write_gdt_entry(get_cpu_gdt_table(cpu), entry, a, b);
 }
 
-#define set_ldt native_set_ldt
-#endif /* CONFIG_PARAVIRT */
-
 static inline fastcall void native_set_ldt(const void *addr,
 					   unsigned int entries)
 {
include/asm-i386/paravirt.h +10 -6
Index: linux/include/asm-i386/paravirt.h
===================================================================
@@ -454,16 +454,20 @@ static inline unsigned long __raw_local_
 	return f;
 }
 
-#define CLI_STRING paravirt_alt("pushl %ecx; pushl %edx;"		\
-		     "call *paravirt_ops+PARAVIRT_irq_disable;"		\
-		     "popl %edx; popl %ecx",				\
+#define CLI_STRING paravirt_alt("pushl %%ecx; pushl %%edx;"		\
+		     "call *paravirt_ops+%c[irq_disable];"		\
+		     "popl %%edx; popl %%ecx",				\
 		     PARAVIRT_IRQ_DISABLE, CLBR_EAX)
 
-#define STI_STRING paravirt_alt("pushl %ecx; pushl %edx;"		\
-		     "call *paravirt_ops+PARAVIRT_irq_enable;"		\
-		     "popl %edx; popl %ecx",				\
+#define STI_STRING paravirt_alt("pushl %%ecx; pushl %%edx;"		\
+		     "call *paravirt_ops+%c[irq_enable];"		\
+		     "popl %%edx; popl %%ecx",				\
 		     PARAVIRT_IRQ_ENABLE, CLBR_EAX)
 #define CLI_STI_CLOBBERS , "%eax"
+#define CLI_STI_INPUT_ARGS \
+	,								\
+	[irq_disable] "i" (offsetof(struct paravirt_ops, irq_disable)),	\
+	[irq_enable] "i" (offsetof(struct paravirt_ops, irq_enable))
 
 #else  /* __ASSEMBLY__ */
 
include/asm-i386/spinlock.h +8 -6
Index: linux/include/asm-i386/spinlock.h
===================================================================
@@ -13,6 +13,7 @@
 #define CLI_STRING	"cli"
 #define STI_STRING	"sti"
 #define CLI_STI_CLOBBERS
+#define CLI_STI_INPUT_ARGS
 #endif /* CONFIG_PARAVIRT */
 
 /*
@@ -58,26 +59,27 @@ static inline void __raw_spin_lock_flags
 {
 	asm volatile(
 		"\n1:\t"
-		LOCK_PREFIX " ; decb %0\n\t"
+		LOCK_PREFIX " ; decb %[slock]\n\t"
 		"jns 5f\n"
 		"2:\t"
-		"testl $0x200, %1\n\t"
+		"testl $0x200, %[flags]\n\t"
 		"jz 4f\n\t"
 		STI_STRING "\n"
 		"3:\t"
 		"rep;nop\n\t"
-		"cmpb $0, %0\n\t"
+		"cmpb $0, %[slock]\n\t"
 		"jle 3b\n\t"
 		CLI_STRING "\n\t"
 		"jmp 1b\n"
 		"4:\t"
 		"rep;nop\n\t"
-		"cmpb $0, %0\n\t"
+		"cmpb $0, %[slock]\n\t"
 		"jg 1b\n\t"
 		"jmp 4b\n"
 		"5:\n\t"
-		: "+m" (lock->slock)
-		: "r" (flags)
+		: [slock] "+m" (lock->slock)
+		: [flags] "r" (flags) 
+	 	  CLI_STI_INPUT_ARGS
 		: "memory" CLI_STI_CLOBBERS);
 }
 #endif
include/asm-i386/processor.h +2 -0
Index: linux/include/asm-i386/processor.h
===================================================================
@@ -511,6 +511,7 @@ static inline void load_esp0(struct tss_
 		wrmsr(MSR_IA32_SYSENTER_CS, thread->sysenter_cs, 0);
 	}
 }
+#endif
 
 #define start_thread(regs, new_eip, new_esp) do {		\
 	__asm__("movl %0,%%fs": :"r" (0));			\
@@ -524,6 +525,7 @@ static inline void load_esp0(struct tss_
 	regs->esp = new_esp;					\
 } while (0)
 
+#ifndef CONFIG_PARAVIRT
 /*
  * These special macros can be used to get or set a debugging register
  */
#54 lore raw

Re: [PATCH 1/7] paravirtualization: header and stubs for paravirtualizing critical operations

In reply to Rusty Russell
Quoted 9 lines · Rusty Russell

Create a paravirt.h header for all the critical operations which need to be replaced with hypervisor calls, and include that instead of defining native operations, when CONFIG_PARAVIRT.

This patch does the dumbest possible replacement of paravirtualized
instructions: calls through a "paravirt_ops" structure. Currently these are function implementations of native hardware: hypervisors will override the ops structure with their own variants.

[snip]

Quoted 2 lines
arch/i386/kernel/paravirt.c new file +2 -0
+struct paravirt_ops paravirt_ops = {
+	.name = "bare hardware",

[snip]

Quoted 2 lines
arch/i386/kernel/paravirt.c new file +2 -0
+	.get_wallclock = native_get_wallclock,
+	.set_wallclock = native_set_wallclock,

[snip]

Quoted 29 lines
include/asm-i386/time.h new file +26 -0
@@ -0,0 +1,41 @@
+#ifndef _ASMi386_TIME_H
+#define _ASMi386_TIME_H
+
+#include <linux/efi.h>
+#include "mach_time.h"
+
+static inline unsigned long native_get_wallclock(void)
+{
+	unsigned long retval;
+
+	if (efi_enabled)
+		retval = efi_get_time();
+	else
+		retval = mach_get_cmos_time();
+
+	return retval;
+}
+
+static inline int native_set_wallclock(unsigned long nowtime)
+{
+	int retval;
+
+	if (efi_enabled)
+		retval = efi_set_rtc_mmss(nowtime);
+	else
+		retval = mach_set_rtc_mmss(nowtime);

include/asm-i386/time.h line 36

Quoted 10 lines
include/asm-i386/time.h new file +10 -0
+
+	return retval;
+}
+
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else /* !CONFIG_PARAVIRT */
+
+#define get_wallclock() native_get_wallclock()
+#define set_wallclock(x) native_set_wallclock(x)

Could a better name then "get/set_wallclock" be used here? Its too vague and would be easily confused with do_set/gettimeofday() functions.

My suggestion would be to use "persistent_clock" to describe the battery-backed CMOS/hardware clock. (I assume that is what you intend this paravirt_op to be, rather then get the high-resolution system timeofday)

thanks
-john

#55 lore raw

PATCH 7/7 Add mmu virtualization to paravirt-ops.

In reply to Chris Wright

Add the three bare TLB accessor functions to paravirt-ops. Most amusingly, flush_tlb is redefined on SMP, so I can't call the paravirt op flush_tlb. Instead, I chose to indicate the actual flush type, kernel (global) vs. user (non-global). Global in this sense means using the global bit in the page table entry, which makes TLB entries persistent across CR3 reloads, not global as in the SMP sense of invoking remote shootdowns, so the term is confusingly overloaded.

Cc Rusty Russell <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
---
 arch/i386/kernel/paravirt.c       |  109 ++++++++++++++++++++++++++++++++++++++
 arch/i386/mm/boot_ioremap.c       |    1 
 include/asm-i386/paravirt.h       |   75 ++++++++++++++++++++++++++
 include/asm-i386/pgtable-2level.h |    5 +
 include/asm-i386/pgtable-3level.h |   42 +++++++-------
 include/asm-i386/pgtable.h        |    4 +
 include/asm-i386/tlbflush.h       |   18 ++++--
 7 files changed, 227 insertions(+), 27 deletions(-)
arch/i386/kernel/paravirt.c +109 -0
@@ -30,6 +30,7 @@
 #include <asm/delay.h>
 #include <asm/fixmap.h>
 #include <asm/apic.h>
+#include <asm/tlbflush.h>
 
 /* nop stub */
 static void native_nop(void)
@@ -404,6 +405,97 @@ static fastcall unsigned long native_api
 }
 #endif /* CONFIG_X86_LOCAL_APIC */
 
+static fastcall void native_flush_tlb(void)
+{
+	__native_flush_tlb();
+}
+
+/*
+ * Global pages have to be flushed a bit differently. Not a real
+ * performance problem because this does not happen often.
+ */
+static fastcall void native_flush_tlb_global(void)
+{
+	__native_flush_tlb_global();
+}
+
+static fastcall void native_flush_tlb_single(u32 addr)
+{
+	__native_flush_tlb_single(addr);
+}
+
+#ifndef CONFIG_X86_PAE
+static fastcall void native_set_pte(pte_t *ptep, pte_t pteval)
+{
+	*ptep = pteval;
+}
+
+static fastcall void native_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pteval)
+{
+	*ptep = pteval;
+}
+
+static fastcall void native_set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	*pmdp = pmdval;
+}
+
+#else /* CONFIG_X86_PAE */
+
+static fastcall void native_set_pte(pte_t *ptep, pte_t pte)
+{
+	ptep->pte_high = pte.pte_high;
+	smp_wmb();
+	ptep->pte_low = pte.pte_low;
+}
+
+static fastcall void native_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pte)
+{
+	ptep->pte_high = pte.pte_high;
+	smp_wmb();
+	ptep->pte_low = pte.pte_low;
+}
+
+static fastcall void native_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
+{
+	ptep->pte_low = 0;
+	smp_wmb();
+	ptep->pte_high = pte.pte_high;
+	smp_wmb();
+	ptep->pte_low = pte.pte_low;
+}
+
+static fastcall void native_set_pte_atomic(pte_t *ptep, pte_t pteval)
+{
+	set_64bit((unsigned long long *)ptep,pte_val(pteval));
+}
+
+static fastcall void native_set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	set_64bit((unsigned long long *)pmdp,pmd_val(pmdval));
+}
+
+static fastcall void native_set_pud(pud_t *pudp, pud_t pudval)
+{
+	*pudp = pudval;
+}
+
+static fastcall void native_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
+{
+	ptep->pte_low = 0;
+	smp_wmb();
+	ptep->pte_high = 0;
+}
+
+static fastcall void native_pmd_clear(pmd_t *pmd)
+{
+	u32 *tmp = (u32 *)pmd;
+	*tmp = 0;
+	smp_wmb();
+	*(tmp + 1) = 0;
+}
+#endif /* CONFIG_X86_PAE */
+
 /* These are in entry.S */
 extern fastcall void native_iret(void);
 extern fastcall void native_irq_enable_sysexit(void);
@@ -480,6 +572,23 @@ struct paravirt_ops paravirt_ops = {
 	.apic_read = native_apic_read,
 #endif
 
+	.flush_tlb_user = native_flush_tlb,
+	.flush_tlb_kernel = native_flush_tlb_global,
+	.flush_tlb_single = native_flush_tlb_single,
+
+	.set_pte = native_set_pte,
+	.set_pte_at = native_set_pte_at,
+	.set_pmd = native_set_pmd,
+	.pte_update = (void *)native_nop,
+	.pte_update_defer = (void *)native_nop,
+#ifdef CONFIG_X86_PAE
+	.set_pte_atomic = native_set_pte_atomic,
+	.set_pte_present = native_set_pte_present,
+	.set_pud = native_set_pud,
+	.pte_clear = native_pte_clear,
+	.pmd_clear = native_pmd_clear,
+#endif
+
 	.irq_enable_sysexit = native_irq_enable_sysexit,
 	.iret = native_iret,
 };
arch/i386/mm/boot_ioremap.c +1 -0
@@ -16,6 +16,7 @@
  */
 
 #undef CONFIG_X86_PAE
+#undef CONFIG_PARAVIRT
 #include <asm/page.h>
 #include <asm/pgtable.h>
 #include <asm/tlbflush.h>
include/asm-i386/paravirt.h +75 -0

99 lines not on this page · raw message

include/asm-i386/pgtable-2level.h +4 -1
@@ -11,11 +11,14 @@
  * within a page table are directly modified.  Thus, the following
  * hook is made available.
  */
+#ifndef CONFIG_PARAVIRT
 #define set_pte(pteptr, pteval) (*(pteptr) = pteval)
 #define set_pte_at(mm,addr,ptep,pteval) set_pte(ptep,pteval)
+#define set_pmd(pmdptr, pmdval) (*(pmdptr) = (pmdval))
+#endif
+
 #define set_pte_atomic(pteptr, pteval) set_pte(pteptr,pteval)
 #define set_pte_present(mm,addr,ptep,pteval) set_pte_at(mm,addr,ptep,pteval)
-#define set_pmd(pmdptr, pmdval) (*(pmdptr) = (pmdval))
 
 #define pte_clear(mm,addr,xp)	do { set_pte_at(mm, addr, xp, __pte(0)); } while (0)
 #define pmd_clear(xp)	do { set_pmd(xp, __pmd(0)); } while (0)
include/asm-i386/pgtable-3level.h +22 -20

66 lines not on this page · raw message

include/asm-i386/pgtable.h +3 -1
@@ -15,6 +15,7 @@
 #include <asm/processor.h>
 #include <asm/fixmap.h>
 #include <linux/threads.h>
+#include <asm/paravirt.h>
 
 #ifndef _I386_BITOPS_H
 #include <asm/bitops.h>
@@ -246,6 +247,7 @@ static inline pte_t pte_mkhuge(pte_t pte
 # include <asm/pgtable-2level.h>
 #endif
 
+#ifndef CONFIG_PARAVIRT
 /*
  * Rules for using pte_update - it must be called after any PTE update which
  * has not been done using the set_pte / clear_pte interfaces.  It is used by
@@ -261,7 +263,7 @@ static inline pte_t pte_mkhuge(pte_t pte
  */
 #define pte_update(mm, addr, ptep)		do { } while (0)
 #define pte_update_defer(mm, addr, ptep)	do { } while (0)
-
+#endif
 
 /*
  * We only update the dirty/accessed state if we set
include/asm-i386/tlbflush.h +13 -5
@@ -4,7 +4,15 @@
 #include <linux/mm.h>
 #include <asm/processor.h>
 
-#define __flush_tlb()							\
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+#define __flush_tlb() __native_flush_tlb()
+#define __flush_tlb_global() __native_flush_tlb_global()
+#define __flush_tlb_single(addr) __native_flush_tlb_single(addr)
+#endif
+
+#define __native_flush_tlb()						\
 	do {								\
 		unsigned int tmpreg;					\
 									\
@@ -19,7 +27,7 @@
  * Global pages have to be flushed a bit differently. Not a real
  * performance problem because this does not happen often.
  */
-#define __flush_tlb_global()						\
+#define __native_flush_tlb_global()					\
 	do {								\
 		unsigned int tmpreg, cr4, cr4_orig;			\
 									\
@@ -36,6 +44,9 @@
 			: "memory");					\
 	} while (0)
 
+#define __native_flush_tlb_single(addr) 				\
+	__asm__ __volatile__("invlpg (%0)" ::"r" (addr) : "memory")
+
 # define __flush_tlb_all()						\
 	do {								\
 		if (cpu_has_pge)					\
@@ -46,9 +57,6 @@
 
 #define cpu_has_invlpg	(boot_cpu_data.x86 > 3)
 
-#define __flush_tlb_single(addr) \
-	__asm__ __volatile__("invlpg (%0)" ::"r" (addr) : "memory")
-
 #ifdef CONFIG_X86_INVLPG
 # define __flush_tlb_one(addr) __flush_tlb_single(addr)
 #else

--