Skip to contentSkip to conversation

larger kernel stack (8k->16k) per task (fwd)

PATCH Balbir Singh 2 participants 2 messages Feb 8, 2002

lore
Share & tools
lore
#0 lore raw

Thanks, I managed to come up with a patch of my own.

The patch looks good, some comments

1. We do not really have to change the size of INIT_TASK since that is architecture independent code in sched.h PARISC is managing with init_task stack of 8K and other stacks of 16K.
2. I think you missed getuser.S in arch/i386/kernel/lib. All the __get_user_x should change to

3. I verified that the instance of GET_CURRENT in hw-irq.h is not being used by anybody and can safely be removed.

__get_user_1:

       movl %esp,%edx
       andl $~(THREAD_SIZE - 1),%edx
       cmpl addr_limit(%edx),%eax

I have a patch that lets the user select any stack size from 8K to 64K, it can be configured. And yes, it works on my system.

I do not have the /proc entry that u have though in my patch.

Would you like to merge both the patches or would you like me to do it and send out a new version.

The patch is attached along with this email. It
is againt 2.4.17

I believe that using 8K as stack size is correct. My patch is to help people debug stack overflows and if they wanted to experiment or port code from other OS'es (where the code uses a stack > 8K - sizeof (struct task_struct)) then they can use this patch.

Comments,
Balbir

Quoted 324 lines

Hi Balbir,

Here is my patch that does what you want but only on i386 architecture. I am forwarding it to you in case you are not subscribed to linux-kernel list.

Regards
Tigran

---------- Forwarded message ----------
Date: Fri, 8 Feb 2002 15:20:19 +0000 (GMT)
From: Tigran Aivazian <[email protected]>
To: [email protected]
Cc: Hugh Dickins <[email protected]>
Subject: [patch] larger kernel stack (8k->16k) per task

Hi,

In the light of some talks here about increasing kernel stack, here is my patch for i386 architecture that some may find useful. It also has a nice extra (/proc/stack) implemented by Hugh Dickins which helps to find major offenders.

It is against 2.4.9 but should be easy to port in any direction. (One way the patch could be improved is by making the size CONFIG_ option instead of hardcoding). Oh btw, please don't tell me "but now you'd need _four_ physically-contiguous pages to create a task instead of two!" because I know it (and think it's not too bad).

Regards,
Tigran

diff -urN 249-vx/arch/i386/kernel/entry.S 249-vx-bigstack/arch/i386/kernel/entry.S

arch/i386/kernel/entry.S +3 -3
@@ -130,7 +130,7 @@
 .previous
 
 #define GET_CURRENT(reg) \
-	movl $-8192, reg; \
+	movl $-16384, reg; \
 	andl %esp, reg
 
 ENTRY(lcall7)
@@ -145,7 +145,7 @@
 	movl %ecx,CS(%esp)	#
 	movl %esp,%ebx
 	pushl %ebx
-	andl $-8192,%ebx	# GET_CURRENT
+	andl $-16384,%ebx	# GET_CURRENT
 	movl exec_domain(%ebx),%edx	# Get the execution domain
 	movl 4(%edx),%edx	# Get the lcall7 handler for the domain
 	pushl $0x7
@@ -166,7 +166,7 @@
 	movl %ecx,CS(%esp)	#
 	movl %esp,%ebx
 	pushl %ebx
-	andl $-8192,%ebx	# GET_CURRENT
+	andl $-16384,%ebx	# GET_CURRENT
 	movl exec_domain(%ebx),%edx	# Get the execution domain
 	movl 4(%edx),%edx	# Get the lcall7 handler for the domain
 	pushl $0x27

diff -urN 249-vx/arch/i386/kernel/head.S 249-vx-bigstack/arch/i386/kernel/head.S

arch/i386/kernel/head.S +1 -1
@@ -320,7 +320,7 @@
 	ret
 
 ENTRY(stack_start)
-	.long SYMBOL_NAME(init_task_union)+8192
+	.long SYMBOL_NAME(init_task_union)+16384
 	.long __KERNEL_DS
 
 /* This is the default interrupt "handler" :-) */

diff -urN 249-vx/arch/i386/kernel/process.c 249-vx-bigstack/arch/i386/kernel/process.c

arch/i386/kernel/process.c +2 -2
@@ -757,12 +757,12 @@
 		return 0;
 	stack_page = (unsigned long)p;
 	esp = p->thread.esp;
-	if (!stack_page || esp < stack_page || esp > 8188+stack_page)
+	if (!stack_page || esp < stack_page || esp > 16380+stack_page)
 		return 0;
 	/* include/asm-i386/system.h:switch_to() pushes ebp last. */
 	ebp = *(unsigned long *) esp;
 	do {
-		if (ebp < stack_page || ebp > 8184+stack_page)
+		if (ebp < stack_page || ebp > 16376+stack_page)
 			return 0;
 		eip = *(unsigned long *) (ebp+4);
 		if (eip < first_sched || eip >= last_sched)

diff -urN 249-vx/arch/i386/kernel/smpboot.c 249-vx-bigstack/arch/i386/kernel/smpboot.c

arch/i386/kernel/smpboot.c +1 -1
@@ -577,7 +577,7 @@
 
 	/* So we see what's up   */
 	printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
-	stack_start.esp = (void *) (1024 + PAGE_SIZE + (char *)idle);
+	stack_start.esp = (void *) idle->thread.esp;
 
 	/*
 	 * This grunge runs the startup process for

diff -urN 249-vx/arch/i386/kernel/traps.c 249-vx-bigstack/arch/i386/kernel/traps.c

arch/i386/kernel/traps.c +1 -1
@@ -160,7 +160,7 @@
 	unsigned long esp = tsk->thread.esp;
 
 	/* User space on another CPU? */
-	if ((esp ^ (unsigned long)tsk) & (PAGE_MASK<<1))
+	if ((esp ^ (unsigned long)tsk) & ~(THREAD_SIZE-1))
 		return;
 	show_trace((unsigned long *)esp);
 }

diff -urN 249-vx/arch/i386/lib/getuser.S 249-vx-bigstack/arch/i386/lib/getuser.S

arch/i386/lib/getuser.S +3 -3
@@ -28,7 +28,7 @@
 .globl __get_user_1
 __get_user_1:
 	movl %esp,%edx
-	andl $0xffffe000,%edx
+	andl $-16384,%edx
 	cmpl addr_limit(%edx),%eax
 	jae bad_get_user
 1:	movzbl (%eax),%edx
@@ -41,7 +41,7 @@
 	addl $1,%eax
 	movl %esp,%edx
 	jc bad_get_user
-	andl $0xffffe000,%edx
+	andl $-16384,%edx
 	cmpl addr_limit(%edx),%eax
 	jae bad_get_user
 2:	movzwl -1(%eax),%edx
@@ -54,7 +54,7 @@
 	addl $3,%eax
 	movl %esp,%edx
 	jc bad_get_user
-	andl $0xffffe000,%edx
+	andl $-16384,%edx
 	cmpl addr_limit(%edx),%eax
 	jae bad_get_user
 3:	movl -3(%eax),%edx

diff -urN 249-vx/arch/i386/vmlinux.lds 249-vx-bigstack/arch/i386/vmlinux.lds

arch/i386/vmlinux.lds +1 -1
@@ -36,7 +36,7 @@
 
   _edata = .;			/* End of data section */
 
-  . = ALIGN(8192);		/* init_task */
+  . = ALIGN(16384);		/* init_task */
   .data.init_task : { *(.data.init_task) }
 
   . = ALIGN(4096);		/* Init code and data */

diff -urN 249-vx/include/asm-i386/current.h 249-vx-bigstack/include/asm-i386/current.h

include/asm-i386/current.h +1 -1
@@ -6,7 +6,7 @@
 static inline struct task_struct * get_current(void)
 {
 	struct task_struct *current;
-	__asm__("andl %%esp,%0; ":"=r" (current) : "0" (~8191UL));
+	__asm__("andl %%esp,%0; ":"=r" (current) : "0" (~16383UL));
 	return current;
  }
 

diff -urN 249-vx/include/asm-i386/hw_irq.h 249-vx-bigstack/include/asm-i386/hw_irq.h

include/asm-i386/hw_irq.h +1 -1
@@ -115,7 +115,7 @@
 
 #define GET_CURRENT \
 	"movl %esp, %ebx\n\t" \
-	"andl $-8192, %ebx\n\t"
+	"andl $-16384, %ebx\n\t"
 
 /*
  *	SMP has a few special interrupts for IPI messages

diff -urN 249-vx/include/asm-i386/processor.h 249-vx-bigstack/include/asm-i386/processor.h

include/asm-i386/processor.h +19 -6
@@ -445,13 +445,27 @@
 }
 
 unsigned long get_wchan(struct task_struct *p);
-#define KSTK_EIP(tsk)	(((unsigned long *)(4096+(unsigned 
 long)(tsk)))[1019])
-#define KSTK_ESP(tsk)	(((unsigned long *)(4096+(unsigned 
 long)(tsk)))[1022])
 
-#define THREAD_SIZE (2*PAGE_SIZE)
-#define alloc_task_struct() ((struct task_struct *) 
 __get_free_pages(GFP_KERNEL,1))
-#define free_task_struct(p) free_pages((unsigned long) (p), 1)
-#define get_task_struct(tsk)      atomic_inc(&virt_to_page(tsk)->count)
+#define THREAD_ORDER	2
+#define THREAD_SIZE	(PAGE_SIZE << THREAD_ORDER)
+#define THREAD_LONGS	(THREAD_SIZE / sizeof(unsigned long))
+
+#define KSTK_EIP(tsk)	(((unsigned long *)(tsk))[THREAD_LONGS-5])
+#define KSTK_ESP(tsk)	(((unsigned long *)(tsk))[THREAD_LONGS-2])
+
+#define alloc_task_struct()						\
+    ({									\
+	unsigned long tsk = __get_free_pages(GFP_KERNEL, THREAD_ORDER);	\
+	if (tsk != 0) {							\
+		/* Alt-SysRq-T show_task() show if stack went deep */	\
+		memset((char*)tsk+sizeof(struct task_struct), 0,	\
+			THREAD_SIZE-sizeof(struct task_struct));	\
+	}								\
+	(struct task_struct *) tsk;					\
+    })
+
+#define free_task_struct(tsk) free_pages((unsigned long) (tsk), 

THREAD_ORDER)
+#define get_task_struct(tsk) atomic_inc(&virt_to_page(tsk)->count)

#define init_task (init_task_union.task)
#define init_stack (init_task_union.stack)
diff -urN 249-vx/include/linux/sched.h 249-vx-bigstack/include/linux/sched.h

include/linux/sched.h +1 -1
@@ -484,7 +484,7 @@
 
 
 #ifndef INIT_TASK_SIZE
-# define INIT_TASK_SIZE	2048*sizeof(long)
+# define INIT_TASK_SIZE	4096*sizeof(long)
 #endif
 
 union task_union {

diff -urN 249-vx/kernel/exit.c 249-vx-bigstack/kernel/exit.c

kernel/exit.c +69 -0
@@ -11,6 +11,7 @@
 #include <linux/module.h>
 #include <linux/completion.h>
 #include <linux/tty.h>
+#include <linux/proc_fs.h>
 #ifdef CONFIG_BSD_PROCESS_ACCT
 #include <linux/acct.h>
 #endif
@@ -24,6 +25,74 @@
 
 int getrusage(struct task_struct *, int, struct rusage *);
 
+#ifdef CONFIG_PROC_FS
+static int max_stack_depth;
+static pid_t max_stack_pid;
+static struct task_struct *max_stack_task; /* kept for kdb */
+static spinlock_t max_stack_lock = SPIN_LOCK_UNLOCKED;
+static void check_stack_depth(struct task_struct *);
+
+static int stack_read_proc(char *page, char **start, off_t off,
+				int count, int *eof, void *data)
+{
+	struct task_struct *p;
+	int len;
+
+	read_lock(&tasklist_lock);
+	for_each_task(p)
+		check_stack_depth(p);
+	read_unlock(&tasklist_lock);
+
+	spin_lock(&max_stack_lock);
+	len = sprintf(page,
+		"Deepest stack is 0x%04x (+ task 0x%04x = 0x%04x) for pid %d %.16s\n",
+		max_stack_depth, sizeof(*p), sizeof(*p) + max_stack_depth,
+		max_stack_pid, max_stack_task->comm);
+	spin_unlock(&max_stack_lock);
+
+	*eof = (len <= off + count);
+	*start = page + off;
+	len -= off;
+	if (len > count)
+		len = count;
+	if (len < 0)
+		len = 0;
+	return len;
+}
+
+static void check_stack_depth(struct task_struct *p)
+{
+	int depth;
+	struct task_struct *old_stack_task;
+	unsigned long *sp = (unsigned long *)(p + 1);
+
+	while (!*sp)
+		sp++;
+	depth = (int)p + THREAD_SIZE - (int)sp;
+	if (depth <= max_stack_depth)
+		return;
+
+	old_stack_task = NULL;
+	spin_lock(&max_stack_lock);
+	if (depth > max_stack_depth) {
+		old_stack_task = max_stack_task;
+		max_stack_depth = depth;
+		max_stack_pid = p->pid;
+		max_stack_task = p;
+		get_task_struct(p);
+	}
+	spin_unlock(&max_stack_lock);
+
+	if (old_stack_task)
+		free_task_struct(old_stack_task);
+	else
+		create_proc_read_entry("stack",
+			0444, NULL, stack_read_proc, NULL);
+}
+#else  /* ndef CONFIG_PROC_FS */
+#define check_stack_depth(p)	do { } while (0)
+#endif /* ndef CONFIG_PROC_FS */
+
 static void release_task(struct task_struct * p)
 {
 	if (p != current) {

kernel/exit.c line 138

Quoted 7 lines
kernel/exit.c +1 -0
@@ -63,6 +132,7 @@
 		current->counter += p->counter;
 		if (current->counter >= MAX_COUNTER)
 			current->counter = MAX_COUNTER;
+		check_stack_depth(p);
 		p->pid = 0;
 		free_task_struct(p);
 	} else {

_________________________________________________________________
Chat with friends online, try MSN Messenger: http://messenger.msn.com

arch/i386/config.in +21 -1
diff -Naur linux/arch/i386/config.in linux-new/arch/i386/config.in
@@ -413,6 +413,26 @@
    bool '  Magic SysRq key' CONFIG_MAGIC_SYSRQ
    bool '  Spinlock debugging' CONFIG_DEBUG_SPINLOCK
    bool '  Verbose BUG() reporting (adds 70K)' CONFIG_DEBUG_BUGVERBOSE
-fi
+   choice 'Bigger Stack Size Support' \
+	"off    CONFIG_NOBIGSTACK \
+	 16KB   CONFIG_STACK_SIZE_16KB \
+	 32KB   CONFIG_STACK_SIZE_32KB \
+	 64KB   CONFIG_STACK_SIZE_64KB" off
 
+  if [ "$CONFIG_NOBIGSTACK" = "y" ]; then
+     define_int CONFIG_STACK_SIZE_SHIFT 1
+  else
+    if [ "$CONFIG_STACK_SIZE_16KB" = "y" ]; then
+       define_int CONFIG_STACK_SIZE_SHIFT 2
+    else
+      if [ "$CONFIG_STACK_SIZE_32KB" = "y" ]; then
+        define_int CONFIG_STACK_SIZE_SHIFT 3
+      else
+        if [ "$CONFIG_STACK_SIZE_64KB" = "y" ]; then
+          define_int CONFIG_STACK_SIZE_SHIF 4
+        fi
+      fi
+    fi
+  fi
+fi
 endmenu
arch/i386/kernel/entry.S +3 -6
diff -Naur linux/arch/i386/kernel/entry.S linux-new/arch/i386/kernel/entry.S
@@ -45,6 +45,7 @@
 #include <linux/linkage.h>
 #include <asm/segment.h>
 #include <asm/smp.h>
+#include <asm/current.h>
 
 EBX		= 0x00
 ECX		= 0x04
@@ -128,10 +129,6 @@
 	.long 3b,6b;	\
 .previous
 
-#define GET_CURRENT(reg) \
-	movl $-8192, reg; \
-	andl %esp, reg
-
 ENTRY(lcall7)
 	pushfl			# We get a different stack layout with call gates,
 	pushl %eax		# which has to be cleaned up later..
@@ -144,7 +141,7 @@
 	movl %ecx,CS(%esp)	#
 	movl %esp,%ebx
 	pushl %ebx
-	andl $-8192,%ebx	# GET_CURRENT
+	andl $-THREAD_SIZE,%ebx	# GET_CURRENT
 	movl exec_domain(%ebx),%edx	# Get the execution domain
 	movl 4(%edx),%edx	# Get the lcall7 handler for the domain
 	pushl $0x7
@@ -165,7 +162,7 @@
 	movl %ecx,CS(%esp)	#
 	movl %esp,%ebx
 	pushl %ebx
-	andl $-8192,%ebx	# GET_CURRENT
+	andl $-THREAD_SIZE,%ebx	# GET_CURRENT
 	movl exec_domain(%ebx),%edx	# Get the execution domain
 	movl 4(%edx),%edx	# Get the lcall7 handler for the domain
 	pushl $0x27
arch/i386/kernel/init_task.c +1 -1
diff -Naur linux/arch/i386/kernel/init_task.c linux-new/arch/i386/kernel/init_task.c
@@ -14,7 +14,7 @@
 /*
  * Initial task structure.
  *
- * We need to make sure that this is 8192-byte aligned due to the
+ * We need to make sure that this is 16384-byte aligned due to the
  * way process stacks are handled. This is done by having a special
  * "init_task" linker map entry..
  */
arch/i386/lib/getuser.S +7 -3
diff -Naur linux/arch/i386/lib/getuser.S linux-new/arch/i386/lib/getuser.S
@@ -21,6 +21,10 @@
  * as they get called from within inline assembly.
  */
 
+/* Duplicated from asm/processor.h */
+#include <asm/current.h>
+#include <linux/config.h>
+
 addr_limit = 12
 
 .text
@@ -28,7 +32,7 @@
 .globl __get_user_1
 __get_user_1:
 	movl %esp,%edx
-	andl $0xffffe000,%edx
+	andl $~(THREAD_SIZE - 1),%edx
 	cmpl addr_limit(%edx),%eax
 	jae bad_get_user
 1:	movzbl (%eax),%edx
@@ -41,7 +45,7 @@
 	addl $1,%eax
 	movl %esp,%edx
 	jc bad_get_user
-	andl $0xffffe000,%edx
+	andl $~(THREAD_SIZE - 1),%edx
 	cmpl addr_limit(%edx),%eax
 	jae bad_get_user
 2:	movzwl -1(%eax),%edx
@@ -54,7 +58,7 @@
 	addl $3,%eax
 	movl %esp,%edx
 	jc bad_get_user
-	andl $0xffffe000,%edx
+	andl $~(THREAD_SIZE - 1),%edx
 	cmpl addr_limit(%edx),%eax
 	jae bad_get_user
 3:	movl -3(%eax),%edx
include/asm-i386/current.h +32 -3
diff -Naur linux/include/asm-i386/current.h linux-new/include/asm-i386/current.h
@@ -1,15 +1,44 @@
 #ifndef _I386_CURRENT_H
 #define _I386_CURRENT_H
 
+#include <asm/page.h>
+
+/*
+ * Configurable page sizes on i386, mainly for debugging purposes.
+ * (c) Balbir Singh
+ */ 
+
+#ifdef __ASSEMBLY__
+
+#define PAGE_SIZE	4096	/* as cannot handle 1UL << 12 */
+#define THREAD_SIZE ((1 << CONFIG_STACK_SIZE_SHIFT) * PAGE_SIZE)
+
+#else	/* __ASSEMBLY__ */
+
+#define THREAD_SIZE ((1 << CONFIG_STACK_SIZE_SHIFT) * PAGE_SIZE)
+#define alloc_task_struct() \
+  ((struct task_struct *) __get_free_pages(GFP_KERNEL,CONFIG_STACK_SIZE_SHIFT))
+
+#define free_task_struct(p) \
+  free_pages((unsigned long) (p), CONFIG_STACK_SIZE_SHIFT)
+
 struct task_struct;
 
 static inline struct task_struct * get_current(void)
 {
 	struct task_struct *current;
-	__asm__("andl %%esp,%0; ":"=r" (current) : "0" (~8191UL));
+	__asm__("andl %%esp,%0; ":"=r" (current) : "0" (~(THREAD_SIZE - 1)));
 	return current;
- }
- 
+}
+
 #define current get_current()
 
+#endif /* __ASSEMBLY__ */
+
+#define GET_CURRENT(reg) \
+        movl $-THREAD_SIZE, (reg); \
+        andl %esp, (reg)
+
+
 #endif /* !(_I386_CURRENT_H) */
+#define THREAD_SIZE ((1 << CONFIG_STACK_SIZE_SHIFT) * PAGE_SIZE)
include/asm-i386/hw_irq.h +1 -4
diff -Naur linux/include/asm-i386/hw_irq.h linux-new/include/asm-i386/hw_irq.h
@@ -15,6 +15,7 @@
 #include <linux/config.h>
 #include <asm/atomic.h>
 #include <asm/irq.h>
+#include <asm/current.h>
 
 /*
  * IDT vectors usable for external interrupt sources start
@@ -113,10 +114,6 @@
 #define IRQ_NAME2(nr) nr##_interrupt(void)
 #define IRQ_NAME(nr) IRQ_NAME2(IRQ##nr)
 
-#define GET_CURRENT \
-	"movl %esp, %ebx\n\t" \
-	"andl $-8192, %ebx\n\t"
-
 /*
  *	SMP has a few special interrupts for IPI messages
  */
include/asm-i386/processor.h +1 -3
diff -Naur linux/include/asm-i386/processor.h linux-new/include/asm-i386/processor.h
@@ -14,6 +14,7 @@
 #include <asm/types.h>
 #include <asm/sigcontext.h>
 #include <asm/cpufeature.h>
+#include <asm/current.h>
 #include <linux/cache.h>
 #include <linux/config.h>
 #include <linux/threads.h>
@@ -447,9 +448,6 @@
 #define KSTK_EIP(tsk)	(((unsigned long *)(4096+(unsigned long)(tsk)))[1019])
 #define KSTK_ESP(tsk)	(((unsigned long *)(4096+(unsigned long)(tsk)))[1022])
 
-#define THREAD_SIZE (2*PAGE_SIZE)
-#define alloc_task_struct() ((struct task_struct *) __get_free_pages(GFP_KERNEL,1))
-#define free_task_struct(p) free_pages((unsigned long) (p), 1)
 #define get_task_struct(tsk)      atomic_inc(&virt_to_page(tsk)->count)
 
 #define init_task	(init_task_union.task)
#1 lore raw

Hi Balbir,

Quoted 2 lines · Balbir Singh

2. I think you missed getuser.S in arch/i386/kernel/lib. All the __get_user_x should change to

no, I didn't miss them. If you read the patch again you will see them.

Quoted 3 lines

3. I verified that the instance of GET_CURRENT in hw-irq.h is not being used by anybody and can safely be removed.

yes, I also verified and came to the same conclusion but left the change in the patch on purpose (so if anyone does start using it, it is already correct)

Quoted 13 lines

__get_user_1:

        movl %esp,%edx
        andl $~(THREAD_SIZE - 1),%edx
        cmpl addr_limit(%edx),%eax

I have a patch that lets the user select any stack size from 8K to 64K, it can be configured. And yes, it works on my system.

I do not have the /proc entry that u have though in my patch.

Quoted 6 lines

Would you like to merge both the patches or would you like me to do it and send out a new version.

The patch is attached along with this email. It
is againt 2.4.17

The serious problem with your patch is that you missed quite a few places (e.g. smpboot.c and traps.c). Most importantly, you missed the alignment in vmlinux.lds so I guess your machine boots by pure luck :) In the early stages (first hours of writing it) I missed that one too and was puzzled by random panics on boot...

Actually, the patch I sent is only part of the "complete piece", the other part being changes to kdb to work correctly with large stack. I can separate those from kdb patch that I use and send out if there was enough interest.

Regards,
Tigran