Description of the patch:
o avoids init to get a sigsegv from the signal code due oom
o improves the oom detection in the signal-frame code
o send the sigbus if we are trying to write beyond the end of
a shared mapping in the 386 case (__verify_write())
o on ia32 avoids the kernel to sigkill a task running with iopl > 0,
send a sigterm instead (so X _won't_ screw-up the graphic card
anymore)
o page-cache oom friendly by breaking the loop at the first GFP fail
o page-cache nopage operation fixed to correctly kill the process
instead of sending the SIGBUS when necessary
o send a sigkill to the current task it triggered an OOM condition
by reading the ptrace data. Theorically if an iopl application
would go oom while reading a process MM, then it could get
a sigkill too, but I am not worried about that since usually
the debuggers doesn't need iopl() privilegies ;).
o remove the obsoleted lowlevel oom() function and replace it
with a sigkill where necessary
diff -urN 2.3.17/arch/alpha/kernel/signal.c 2.3.17-oom/arch/alpha/kernel/signal.c
--- 2.3.17/arch/alpha/kernel/signal.c Sun Aug 1 18:11:07 1999
+++ 2.3.17-oom/arch/alpha/kernel/signal.c Wed Sep 8 19:50:32 1999
@@ -437,6 +437,8 @@
err |= __copy_to_user(frame->extramask, &set->sig[1],
sizeof(frame->extramask));
}
+ if (err)
+ goto give_sigsegv;
/* Set up to return from userspace. If provided, use a stub
already in userspace. */
@@ -499,6 +501,8 @@
err |= setup_sigcontext(&frame->uc.uc_mcontext, regs, sw,
set->sig[0], oldsp);
err |= __copy_to_user(&frame->uc.uc_sigmask, set, sizeof(*set));
+ if (err)
+ goto give_sigsegv;
/* Set up to return from userspace. If provided, use a stub
already in userspace. */
diff -urN 2.3.17/arch/alpha/mm/fault.c 2.3.17-oom/arch/alpha/mm/fault.c
--- 2.3.17/arch/alpha/mm/fault.c Fri Aug 20 17:42:19 1999
+++ 2.3.17-oom/arch/alpha/mm/fault.c Wed Sep 8 20:43:28 1999
@@ -130,13 +130,13 @@
* make sure we exit gracefully rather than endlessly redo
* the fault.
*/
+survive:
fault = handle_mm_fault(current, vma, address, cause > 0);
- up(&mm->mmap_sem);
-
if (fault < 0)
goto out_of_memory;
if (fault == 0)
goto do_sigbus;
+ up(&mm->mmap_sem);
return;
@@ -177,13 +177,23 @@
* us unable to handle the page fault gracefully.
*/
out_of_memory:
- printk(KERN_ALERT "VM: killing process %s(%d)\n",
- current->comm, current->pid);
- if (!user_mode(regs))
- goto no_context;
- do_exit(SIGKILL);
+ if (current->pid == 1)
+ {
+ current->policy |= SCHED_YIELD;
+ schedule();
+ goto survive;
+ }
+ up(&mm->mmap_sem);
+ if (user_mode(regs))
+ {
+ printk(KERN_ALERT "VM: killing process %s(%d)\n",
+ current->comm, current->pid);
+ do_exit(SIGKILL);
+ }
+ goto no_context;
do_sigbus:
+ up(&mm->mmap_sem);
/*
* Send a sigbus, regardless of whether we were in kernel
* or user mode.
diff -urN 2.3.17/arch/i386/kernel/signal.c 2.3.17-oom/arch/i386/kernel/signal.c
--- 2.3.17/arch/i386/kernel/signal.c Sun Aug 1 18:11:08 1999
+++ 2.3.17-oom/arch/i386/kernel/signal.c Wed Sep 8 19:58:06 1999
@@ -419,13 +419,19 @@
? current->exec_domain->signal_invmap[sig]
: sig),
&frame->sig);
+ if (err)
+ goto give_sigsegv;
err |= setup_sigcontext(&frame->sc, &frame->fpstate, regs, set->sig[0]);
+ if (err)
+ goto give_sigsegv;
if (_NSIG_WORDS > 1) {
err |= __copy_to_user(frame->extramask, &set->sig[1],
sizeof(frame->extramask));
}
+ if (err)
+ goto give_sigsegv;
/* Set up to return from userspace. If provided, use a stub
already in userspace. */
@@ -486,6 +492,8 @@
err |= __put_user(&frame->info, &frame->pinfo);
err |= __put_user(&frame->uc, &frame->puc);
err |= __copy_to_user(&frame->info, info, sizeof(*info));
+ if (err)
+ goto give_sigsegv;
/* Create the ucontext. */
err |= __put_user(0, &frame->uc.uc_flags);
@@ -497,6 +505,8 @@
err |= setup_sigcontext(&frame->uc.uc_mcontext, &frame->fpstate,
regs, set->sig[0]);
err |= __copy_to_user(&frame->uc.uc_sigmask, set, sizeof(*set));
+ if (err)
+ goto give_sigsegv;
/* Set up to return from userspace. If provided, use a stub
already in userspace. */
diff -urN 2.3.17/arch/i386/mm/fault.c 2.3.17-oom/arch/i386/mm/fault.c
--- 2.3.17/arch/i386/mm/fault.c Thu Aug 12 02:53:18 1999
+++ 2.3.17-oom/arch/i386/mm/fault.c Wed Sep 8 20:05:05 1999
@@ -31,6 +31,7 @@
{
struct vm_area_struct * vma;
unsigned long start = (unsigned long) addr;
+ int fault;
if (!size)
return 1;
@@ -50,8 +51,12 @@
start &= PAGE_MASK;
for (;;) {
- if (handle_mm_fault(current, vma, start, 1) <= 0)
- goto bad_area;
+survive:
+ fault = handle_mm_fault(current, vma, start, 1);
+ if (!fault)
+ goto do_sigbus;
+ if (fault < 0)
+ goto out_of_memory;
if (!size)
break;
size--;
@@ -74,6 +79,19 @@
bad_area:
return 0;
+
+do_sigbus:
+ force_sig(SIGBUS, current);
+ goto bad_area;
+
+out_of_memory:
+ if (current->pid == 1)
+ {
+ current->policy |= SCHED_YIELD;
+ schedule();
+ goto survive;
+ }
+ goto bad_area;
}
asmlinkage void do_invalid_op(struct pt_regs *, unsigned long);
@@ -163,6 +181,7 @@
* make sure we exit gracefully rather than endlessly redo
* the fault.
*/
+survive:
{
int fault = handle_mm_fault(tsk, vma, address, write);
if (fault < 0)
@@ -262,10 +281,33 @@
* us unable to handle the page fault gracefully.
*/
out_of_memory:
+ if (tsk->pid == 1)
+ {
+ tsk->policy |= SCHED_YIELD;
+ schedule();
+ goto survive;
+ }
up(&mm->mmap_sem);
- printk("VM: killing process %s\n", tsk->comm);
if (error_code & 4)
- do_exit(SIGKILL);
+ {
+ if (!((regs->eflags >> 12) & 3))
+ {
+ printk(KERN_ALERT "VM: killing process %s\n",
+ tsk->comm);
+ do_exit(SIGKILL);
+ }
+ else
+ {
+ /*
+ * The task is running with privilegies and so we
+ * trust it and we give it a chance to die gracefully.
+ */
+ printk(KERN_ALERT "VM: terminating process %s\n",
+ tsk->comm);
+ force_sig(SIGTERM, current);
+ return;
+ }
+ }
goto no_context;
do_sigbus:
diff -urN 2.3.17/include/linux/mm.h 2.3.17-oom/include/linux/mm.h
--- 2.3.17/include/linux/mm.h Wed Sep 8 18:18:56 1999
+++ 2.3.17-oom/include/linux/mm.h Wed Sep 8 22:30:54 1999
@@ -325,7 +325,6 @@
extern unsigned long paging_init(unsigned long start_mem, unsigned long end_mem);
extern void mem_init(unsigned long start_mem, unsigned long end_mem);
extern void show_mem(void);
-extern void oom(struct task_struct * tsk);
extern void si_meminfo(struct sysinfo * val);
extern void swapin_readahead(unsigned long);
diff -urN 2.3.17/kernel/ptrace.c 2.3.17-oom/kernel/ptrace.c
--- 2.3.17/kernel/ptrace.c Wed Sep 8 00:26:08 1999
+++ 2.3.17-oom/kernel/ptrace.c Wed Sep 8 19:36:04 1999
@@ -24,6 +24,7 @@
pmd_t * pgmiddle;
pte_t * pgtable;
unsigned long page;
+ int fault;
repeat:
pgdir = pgd_offset(vma->vm_mm, addr);
@@ -64,8 +65,12 @@
fault_in_page:
/* -1: out of memory. 0 - unmapped page */
- if (handle_mm_fault(tsk, vma, addr, write) > 0)
+ fault = handle_mm_fault(tsk, vma, addr, write);
+ if (fault > 0)
goto repeat;
+ if (fault < 0)
+ /* the out of memory is been triggered by the current task. */
+ force_sig(SIGKILL, current);
return 0;
bad_pgd:
diff -urN 2.3.17/mm/filemap.c 2.3.17-oom/mm/filemap.c
--- 2.3.17/mm/filemap.c Wed Sep 8 18:18:57 1999
+++ 2.3.17-oom/mm/filemap.c Wed Sep 8 23:18:53 1999
@@ -530,7 +530,7 @@
* This adds the requested page to the page cache if it isn't already there,
* and schedules an I/O to read in its contents from disk.
*/
-static inline void page_cache_read(struct file * file, unsigned long offset)
+static inline int page_cache_read(struct file * file, unsigned long offset)
{
unsigned long new_page;
struct inode *inode = file->f_dentry->d_inode;
@@ -541,17 +541,17 @@
page = __find_page_nolock(inode, offset, *hash);
spin_unlock(&pagecache_lock);
if (page)
- return;
+ return 1;
new_page = page_cache_alloc();
if (!new_page)
- return;
+ return 0;
page = page_cache_entry(new_page);
if (!add_to_page_cache_unique(page, inode, offset, hash)) {
inode->i_op->readpage(file, page);
page_cache_release(page);
- return;
+ return 1;
}
/*
@@ -559,14 +559,14 @@
* raced with us and added our page to the cache first.
*/
page_cache_free(new_page);
- return;
+ return 1;
}
/*
* Read in an entire cluster at once. A cluster is usually a 64k-
* aligned block that includes the address requested in "offset."
*/
-static void read_cluster_nonblocking(struct file * file,
+static int read_cluster_nonblocking(struct file * file,
unsigned long offset)
{
off_t filesize = file->f_dentry->d_inode->i_size;
@@ -574,11 +574,12 @@
offset = CLUSTER_OFFSET(offset);
while ((pages-- > 0) && (offset < filesize)) {
- page_cache_read(file, offset);
+ if (!page_cache_read(file, offset))
+ return 0;
offset += PAGE_CACHE_SIZE;
}
- return;
+ return 1;
}
/*
@@ -912,7 +913,8 @@
ahead = 0;
while (ahead < max_ahead) {
ahead += PAGE_CACHE_SIZE;
- page_cache_read(filp, raend + ahead);
+ if (!page_cache_read(filp, raend + ahead))
+ break;
}
/*
* If we tried to read ahead some pages,
@@ -1347,7 +1349,7 @@
flush_page_to_ram(new_page);
}
page_cache_release(page);
- return new_page;
+ return new_page ? : -1;
}
flush_page_to_ram(old_page);
@@ -1361,10 +1363,13 @@
* Otherwise, we're off the end of a privately mapped file,
* so we need to map a zero page.
*/
- if (offset < inode->i_size)
- read_cluster_nonblocking(file, offset);
- else
- page_cache_read(file, offset);
+ if (offset < inode->i_size) {
+ if (!read_cluster_nonblocking(file, offset))
+ return -1;
+ } else {
+ if (!page_cache_read(file, offset))
+ return -1;
+ }
/*
* The page we want has now been added to the page cache.
diff -urN 2.3.17/mm/memory.c 2.3.17-oom/mm/memory.c
--- 2.3.17/mm/memory.c Wed Sep 8 00:26:08 1999
+++ 2.3.17-oom/mm/memory.c Wed Sep 8 23:03:06 1999
@@ -69,16 +69,6 @@
mem_map_t * mem_map = NULL;
/*
- * oom() prints a message (so that the user knows why the process died),
- * and gives the process an untrappable SIGKILL.
- */
-void oom(struct task_struct * task)
-{
- printk("\nOut of memory for %s.\n", task->comm);
- force_sig(SIGKILL, task);
-}
-
-/*
* Note: this doesn't free the actual pages themselves. That
* has been handled earlier when unmapping all the memory regions.
*/
@@ -732,13 +722,13 @@
pmd = pmd_alloc(pgd, address);
if (!pmd) {
free_page(page);
- oom(tsk);
+ force_sig(SIGKILL, tsk);
return 0;
}
pte = pte_alloc(pmd, address);
if (!pte) {
free_page(page);
- oom(tsk);
+ force_sig(SIGKILL, tsk);
return 0;
}
if (!pte_none(*pte)) {
o swap_out() is now allowed to refresh the swap counter only at
once per loop. max_cnt is an unsigned long.
diff -urN 2.3.17/mm/vmscan.c 2.3.17-oom/mm/vmscan.c
--- 2.3.17/mm/vmscan.c Wed Sep 8 00:26:08 1999
+++ 2.3.17-oom/mm/vmscan.c Wed Sep 8 23:12:07 1999
@@ -327,6 +327,7 @@
struct task_struct * p;
int counter;
int __ret = 0;
+ int assign = 0;
lock_kernel();
/*
@@ -346,12 +347,9 @@
counter = nr_threads / (priority+1);
if (counter < 1)
counter = 1;
- if (counter > nr_threads)
- counter = nr_threads;
for (; counter >= 0; counter--) {
- int assign = 0;
- int max_cnt = 0;
+ unsigned long max_cnt = 0;
struct mm_struct *best = NULL;
int pid = 0;
select:
@@ -364,7 +362,7 @@
if (mm->rss <= 0)
continue;
/* Refresh swap_cnt? */
- if (assign)
+ if (assign == 1)
mm->swap_cnt = mm->rss;
if (mm->swap_cnt > max_cnt) {
max_cnt = mm->swap_cnt;
@@ -373,6 +371,8 @@
}
}
read_unlock(&tasklist_lock);
+ if (assign == 1)
+ assign = 2;
if (!best) {
if (!assign) {
assign = 1;
Andrea
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/