[PATCH] kexec: add kexec syscalls
authorEric W. Biederman <ebiederm@xmission.com>
Sat, 25 Jun 2005 21:57:52 +0000 (14:57 -0700)
committerLinus Torvalds <torvalds@ppc970.osdl.org>
Sat, 25 Jun 2005 23:24:48 +0000 (16:24 -0700)
This patch introduces the architecture independent implementation the
sys_kexec_load, the compat_sys_kexec_load system calls.

Kexec on panic support has been integrated into the core patch and is
relatively clean.

In addition the hopefully architecture independent option
crashkernel=size@location has been docuemented.  It's purpose is to reserve
space for the panic kernel to live, and where no DMA transfer will ever be
setup to access.

Signed-off-by: Eric Biederman <ebiederm@xmission.com>
Signed-off-by: Alexander Nyberg <alexn@telia.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Documentation/kernel-parameters.txt
MAINTAINERS
include/linux/kexec.h [new file with mode: 0644]
include/linux/reboot.h
include/linux/syscalls.h
kernel/Makefile
kernel/kexec.c [new file with mode: 0644]
kernel/panic.c
kernel/sys.c
kernel/sys_ni.c

index 86db43fd6b0f26559f8a6ce9ab5eae16e05a49b5..560ff5ae3fd990369b182527cd3d3499e3036cc0 100644 (file)
@@ -358,6 +358,10 @@ running once the system is up.
        cpia_pp=        [HW,PPT]
                        Format: { parport<nr> | auto | none }
 
+       crashkernel=nn[KMG]@ss[KMG]
+                       [KNL] Reserve a chunk of physical memory to
+                       hold a kernel to switch to with kexec on panic.
+
        cs4232=         [HW,OSS]
                        Format: <io>,<irq>,<dma>,<dma2>,<mpuio>,<mpuirq>
 
index dbdd8494b2e6f26b26f0191fe857dc00ed2a1012..81728572799e23d9a14d476ff2e034c4c6ac433e 100644 (file)
@@ -1330,6 +1330,16 @@ M:       rml@novell.com
 L:     linux-kernel@vger.kernel.org
 S:     Maintained
 
+KEXEC
+P:     Eric Biederman
+P:     Randy Dunlap
+M:     ebiederm@xmission.com
+M:     rddunlap@osdl.org
+W:     http://www.xmission.com/~ebiederm/files/kexec/
+L:     linux-kernel@vger.kernel.org
+L:     fastboot@osdl.org
+S:     Maintained
+
 LANMEDIA WAN CARD DRIVER
 P:     Andrew Stanley-Jones
 M:     asj@lanmedia.com
diff --git a/include/linux/kexec.h b/include/linux/kexec.h
new file mode 100644 (file)
index 0000000..e3fc35f
--- /dev/null
@@ -0,0 +1,127 @@
+#ifndef LINUX_KEXEC_H
+#define LINUX_KEXEC_H
+
+#ifdef CONFIG_KEXEC
+#include <linux/types.h>
+#include <linux/list.h>
+#include <linux/linkage.h>
+#include <linux/compat.h>
+#include <asm/kexec.h>
+
+/* Verify architecture specific macros are defined */
+
+#ifndef KEXEC_SOURCE_MEMORY_LIMIT
+#error KEXEC_SOURCE_MEMORY_LIMIT not defined
+#endif
+
+#ifndef KEXEC_DESTINATION_MEMORY_LIMIT
+#error KEXEC_DESTINATION_MEMORY_LIMIT not defined
+#endif
+
+#ifndef KEXEC_CONTROL_MEMORY_LIMIT
+#error KEXEC_CONTROL_MEMORY_LIMIT not defined
+#endif
+
+#ifndef KEXEC_CONTROL_CODE_SIZE
+#error KEXEC_CONTROL_CODE_SIZE not defined
+#endif
+
+#ifndef KEXEC_ARCH
+#error KEXEC_ARCH not defined
+#endif
+
+/*
+ * This structure is used to hold the arguments that are used when loading
+ * kernel binaries.
+ */
+
+typedef unsigned long kimage_entry_t;
+#define IND_DESTINATION  0x1
+#define IND_INDIRECTION  0x2
+#define IND_DONE         0x4
+#define IND_SOURCE       0x8
+
+#define KEXEC_SEGMENT_MAX 8
+struct kexec_segment {
+       void __user *buf;
+       size_t bufsz;
+       unsigned long mem;      /* User space sees this as a (void *) ... */
+       size_t memsz;
+};
+
+#ifdef CONFIG_COMPAT
+struct compat_kexec_segment {
+       compat_uptr_t buf;
+       compat_size_t bufsz;
+       compat_ulong_t mem;     /* User space sees this as a (void *) ... */
+       compat_size_t memsz;
+};
+#endif
+
+struct kimage {
+       kimage_entry_t head;
+       kimage_entry_t *entry;
+       kimage_entry_t *last_entry;
+
+       unsigned long destination;
+
+       unsigned long start;
+       struct page *control_code_page;
+
+       unsigned long nr_segments;
+       struct kexec_segment segment[KEXEC_SEGMENT_MAX];
+
+       struct list_head control_pages;
+       struct list_head dest_pages;
+       struct list_head unuseable_pages;
+
+       /* Address of next control page to allocate for crash kernels. */
+       unsigned long control_page;
+
+       /* Flags to indicate special processing */
+       unsigned int type : 1;
+#define KEXEC_TYPE_DEFAULT 0
+#define KEXEC_TYPE_CRASH   1
+};
+
+
+
+/* kexec interface functions */
+extern NORET_TYPE void machine_kexec(struct kimage *image) ATTRIB_NORET;
+extern int machine_kexec_prepare(struct kimage *image);
+extern void machine_kexec_cleanup(struct kimage *image);
+extern asmlinkage long sys_kexec_load(unsigned long entry,
+       unsigned long nr_segments, struct kexec_segment __user *segments,
+       unsigned long flags);
+#ifdef CONFIG_COMPAT
+extern asmlinkage long compat_sys_kexec_load(unsigned long entry,
+       unsigned long nr_segments, struct compat_kexec_segment __user *segments,
+       unsigned long flags);
+#endif
+extern struct page *kimage_alloc_control_pages(struct kimage *image, unsigned int order);
+extern void crash_kexec(void);
+extern struct kimage *kexec_image;
+
+#define KEXEC_ON_CRASH  0x00000001
+#define KEXEC_ARCH_MASK 0xffff0000
+
+/* These values match the ELF architecture values.
+ * Unless there is a good reason that should continue to be the case.
+ */
+#define KEXEC_ARCH_DEFAULT ( 0 << 16)
+#define KEXEC_ARCH_386     ( 3 << 16)
+#define KEXEC_ARCH_X86_64  (62 << 16)
+#define KEXEC_ARCH_PPC     (20 << 16)
+#define KEXEC_ARCH_PPC64   (21 << 16)
+#define KEXEC_ARCH_IA_64   (50 << 16)
+
+#define KEXEC_FLAGS    (KEXEC_ON_CRASH)  /* List of defined/legal kexec flags */
+
+/* Location of a reserved region to hold the crash kernel.
+ */
+extern struct resource crashk_res;
+
+#else /* !CONFIG_KEXEC */
+static inline void crash_kexec(void) { }
+#endif /* CONFIG_KEXEC */
+#endif /* LINUX_KEXEC_H */
index d60fafc8bdc584e73662a1e51f2876b09b13c9f1..c5a05e16edb20b0bedc0bd873d1c2be1552a20dc 100644 (file)
@@ -51,6 +51,9 @@ extern void machine_restart(char *cmd);
 extern void machine_halt(void);
 extern void machine_power_off(void);
 
+extern void machine_shutdown(void);
+extern void machine_crash_shutdown(void);
+
 #endif
 
 #endif /* _LINUX_REBOOT_H */
index c39f6f72cbbc3f0a53ae8e8a31b0749e2c8ea747..7ba8f8f747aab55029f0f393a06d8c4be979f2af 100644 (file)
@@ -159,8 +159,9 @@ asmlinkage long sys_shutdown(int, int);
 asmlinkage long sys_reboot(int magic1, int magic2, unsigned int cmd,
                                void __user *arg);
 asmlinkage long sys_restart_syscall(void);
-asmlinkage long sys_kexec_load(void *entry, unsigned long nr_segments,
-                       struct kexec_segment *segments, unsigned long flags);
+asmlinkage long sys_kexec_load(unsigned long entry,
+       unsigned long nr_segments, struct kexec_segment __user *segments,
+       unsigned long flags);
 
 asmlinkage long sys_exit(int error_code);
 asmlinkage void sys_exit_group(int error_code);
index b01d26fe8db71abdd0397bcb25eb4e47d3f8c45e..cfc8b0dea950c572e03ea61ec2757b75b7bc8b9f 100644 (file)
@@ -17,6 +17,7 @@ obj-$(CONFIG_MODULES) += module.o
 obj-$(CONFIG_KALLSYMS) += kallsyms.o
 obj-$(CONFIG_PM) += power/
 obj-$(CONFIG_BSD_PROCESS_ACCT) += acct.o
+obj-$(CONFIG_KEXEC) += kexec.o
 obj-$(CONFIG_COMPAT) += compat.o
 obj-$(CONFIG_CPUSETS) += cpuset.o
 obj-$(CONFIG_IKCONFIG) += configs.o
diff --git a/kernel/kexec.c b/kernel/kexec.c
new file mode 100644 (file)
index 0000000..def9c73
--- /dev/null
@@ -0,0 +1,1036 @@
+/*
+ * kexec.c - kexec system call
+ * Copyright (C) 2002-2004 Eric Biederman  <ebiederm@xmission.com>
+ *
+ * This source code is licensed under the GNU General Public License,
+ * Version 2.  See the file COPYING for more details.
+ */
+
+#include <linux/mm.h>
+#include <linux/file.h>
+#include <linux/slab.h>
+#include <linux/fs.h>
+#include <linux/kexec.h>
+#include <linux/spinlock.h>
+#include <linux/list.h>
+#include <linux/highmem.h>
+#include <linux/syscalls.h>
+#include <linux/reboot.h>
+#include <linux/syscalls.h>
+#include <linux/ioport.h>
+#include <asm/page.h>
+#include <asm/uaccess.h>
+#include <asm/io.h>
+#include <asm/system.h>
+#include <asm/semaphore.h>
+
+/* Location of the reserved area for the crash kernel */
+struct resource crashk_res = {
+       .name  = "Crash kernel",
+       .start = 0,
+       .end   = 0,
+       .flags = IORESOURCE_BUSY | IORESOURCE_MEM
+};
+
+/*
+ * When kexec transitions to the new kernel there is a one-to-one
+ * mapping between physical and virtual addresses.  On processors
+ * where you can disable the MMU this is trivial, and easy.  For
+ * others it is still a simple predictable page table to setup.
+ *
+ * In that environment kexec copies the new kernel to its final
+ * resting place.  This means I can only support memory whose
+ * physical address can fit in an unsigned long.  In particular
+ * addresses where (pfn << PAGE_SHIFT) > ULONG_MAX cannot be handled.
+ * If the assembly stub has more restrictive requirements
+ * KEXEC_SOURCE_MEMORY_LIMIT and KEXEC_DEST_MEMORY_LIMIT can be
+ * defined more restrictively in <asm/kexec.h>.
+ *
+ * The code for the transition from the current kernel to the
+ * the new kernel is placed in the control_code_buffer, whose size
+ * is given by KEXEC_CONTROL_CODE_SIZE.  In the best case only a single
+ * page of memory is necessary, but some architectures require more.
+ * Because this memory must be identity mapped in the transition from
+ * virtual to physical addresses it must live in the range
+ * 0 - TASK_SIZE, as only the user space mappings are arbitrarily
+ * modifiable.
+ *
+ * The assembly stub in the control code buffer is passed a linked list
+ * of descriptor pages detailing the source pages of the new kernel,
+ * and the destination addresses of those source pages.  As this data
+ * structure is not used in the context of the current OS, it must
+ * be self-contained.
+ *
+ * The code has been made to work with highmem pages and will use a
+ * destination page in its final resting place (if it happens
+ * to allocate it).  The end product of this is that most of the
+ * physical address space, and most of RAM can be used.
+ *
+ * Future directions include:
+ *  - allocating a page table with the control code buffer identity
+ *    mapped, to simplify machine_kexec and make kexec_on_panic more
+ *    reliable.
+ */
+
+/*
+ * KIMAGE_NO_DEST is an impossible destination address..., for
+ * allocating pages whose destination address we do not care about.
+ */
+#define KIMAGE_NO_DEST (-1UL)
+
+static int kimage_is_destination_range(
+       struct kimage *image, unsigned long start, unsigned long end);
+static struct page *kimage_alloc_page(struct kimage *image, unsigned int gfp_mask, unsigned long dest);
+
+static int do_kimage_alloc(struct kimage **rimage, unsigned long entry,
+       unsigned long nr_segments, struct kexec_segment __user *segments)
+{
+       size_t segment_bytes;
+       struct kimage *image;
+       unsigned long i;
+       int result;
+
+       /* Allocate a controlling structure */
+       result = -ENOMEM;
+       image = kmalloc(sizeof(*image), GFP_KERNEL);
+       if (!image) {
+               goto out;
+       }
+       memset(image, 0, sizeof(*image));
+       image->head = 0;
+       image->entry = &image->head;
+       image->last_entry = &image->head;
+       image->control_page = ~0; /* By default this does not apply */
+       image->start = entry;
+       image->type = KEXEC_TYPE_DEFAULT;
+
+       /* Initialize the list of control pages */
+       INIT_LIST_HEAD(&image->control_pages);
+
+       /* Initialize the list of destination pages */
+       INIT_LIST_HEAD(&image->dest_pages);
+
+       /* Initialize the list of unuseable pages */
+       INIT_LIST_HEAD(&image->unuseable_pages);
+
+       /* Read in the segments */
+       image->nr_segments = nr_segments;
+       segment_bytes = nr_segments * sizeof(*segments);
+       result = copy_from_user(image->segment, segments, segment_bytes);
+       if (result)
+               goto out;
+
+       /*
+        * Verify we have good destination addresses.  The caller is
+        * responsible for making certain we don't attempt to load
+        * the new image into invalid or reserved areas of RAM.  This
+        * just verifies it is an address we can use.
+        *
+        * Since the kernel does everything in page size chunks ensure
+        * the destination addreses are page aligned.  Too many
+        * special cases crop of when we don't do this.  The most
+        * insidious is getting overlapping destination addresses
+        * simply because addresses are changed to page size
+        * granularity.
+        */
+       result = -EADDRNOTAVAIL;
+       for (i = 0; i < nr_segments; i++) {
+               unsigned long mstart, mend;
+               mstart = image->segment[i].mem;
+               mend   = mstart + image->segment[i].memsz;
+               if ((mstart & ~PAGE_MASK) || (mend & ~PAGE_MASK))
+                       goto out;
+               if (mend >= KEXEC_DESTINATION_MEMORY_LIMIT)
+                       goto out;
+       }
+
+       /* Verify our destination addresses do not overlap.
+        * If we alloed overlapping destination addresses
+        * through very weird things can happen with no
+        * easy explanation as one segment stops on another.
+        */
+       result = -EINVAL;
+       for(i = 0; i < nr_segments; i++) {
+               unsigned long mstart, mend;
+               unsigned long j;
+               mstart = image->segment[i].mem;
+               mend   = mstart + image->segment[i].memsz;
+               for(j = 0; j < i; j++) {
+                       unsigned long pstart, pend;
+                       pstart = image->segment[j].mem;
+                       pend   = pstart + image->segment[j].memsz;
+                       /* Do the segments overlap ? */
+                       if ((mend > pstart) && (mstart < pend))
+                               goto out;
+               }
+       }
+
+       /* Ensure our buffer sizes are strictly less than
+        * our memory sizes.  This should always be the case,
+        * and it is easier to check up front than to be surprised
+        * later on.
+        */
+       result = -EINVAL;
+       for(i = 0; i < nr_segments; i++) {
+               if (image->segment[i].bufsz > image->segment[i].memsz)
+                       goto out;
+       }
+
+
+       result = 0;
+ out:
+       if (result == 0) {
+               *rimage = image;
+       } else {
+               kfree(image);
+       }
+       return result;
+
+}
+
+static int kimage_normal_alloc(struct kimage **rimage, unsigned long entry,
+       unsigned long nr_segments, struct kexec_segment __user *segments)
+{
+       int result;
+       struct kimage *image;
+
+       /* Allocate and initialize a controlling structure */
+       image = NULL;
+       result = do_kimage_alloc(&image, entry, nr_segments, segments);
+       if (result) {
+               goto out;
+       }
+       *rimage = image;
+
+       /*
+        * Find a location for the control code buffer, and add it
+        * the vector of segments so that it's pages will also be
+        * counted as destination pages.
+        */
+       result = -ENOMEM;
+       image->control_code_page = kimage_alloc_control_pages(image,
+               get_order(KEXEC_CONTROL_CODE_SIZE));
+       if (!image->control_code_page) {
+               printk(KERN_ERR "Could not allocate control_code_buffer\n");
+               goto out;
+       }
+
+       result = 0;
+ out:
+       if (result == 0) {
+               *rimage = image;
+       } else {
+               kfree(image);
+       }
+       return result;
+}
+
+static int kimage_crash_alloc(struct kimage **rimage, unsigned long entry,
+       unsigned long nr_segments, struct kexec_segment *segments)
+{
+       int result;
+       struct kimage *image;
+       unsigned long i;
+
+       image = NULL;
+       /* Verify we have a valid entry point */
+       if ((entry < crashk_res.start) || (entry > crashk_res.end)) {
+               result = -EADDRNOTAVAIL;
+               goto out;
+       }
+
+       /* Allocate and initialize a controlling structure */
+       result = do_kimage_alloc(&image, entry, nr_segments, segments);
+       if (result) {
+               goto out;
+       }
+
+       /* Enable the special crash kernel control page
+        * allocation policy.
+        */
+       image->control_page = crashk_res.start;
+       image->type = KEXEC_TYPE_CRASH;
+
+       /*
+        * Verify we have good destination addresses.  Normally
+        * the caller is responsible for making certain we don't
+        * attempt to load the new image into invalid or reserved
+        * areas of RAM.  But crash kernels are preloaded into a
+        * reserved area of ram.  We must ensure the addresses
+        * are in the reserved area otherwise preloading the
+        * kernel could corrupt things.
+        */
+       result = -EADDRNOTAVAIL;
+       for (i = 0; i < nr_segments; i++) {
+               unsigned long mstart, mend;
+               mstart = image->segment[i].mem;
+               mend = mstart + image->segment[i].memsz;
+               /* Ensure we are within the crash kernel limits */
+               if ((mstart < crashk_res.start) || (mend > crashk_res.end))
+                       goto out;
+       }
+
+
+       /*
+        * Find a location for the control code buffer, and add
+        * the vector of segments so that it's pages will also be
+        * counted as destination pages.
+        */
+       result = -ENOMEM;
+       image->control_code_page = kimage_alloc_control_pages(image,
+               get_order(KEXEC_CONTROL_CODE_SIZE));
+       if (!image->control_code_page) {
+               printk(KERN_ERR "Could not allocate control_code_buffer\n");
+               goto out;
+       }
+
+       result = 0;
+ out:
+       if (result == 0) {
+               *rimage = image;
+       } else {
+               kfree(image);
+       }
+       return result;
+}
+
+static int kimage_is_destination_range(
+       struct kimage *image, unsigned long start, unsigned long end)
+{
+       unsigned long i;
+
+       for (i = 0; i < image->nr_segments; i++) {
+               unsigned long mstart, mend;
+               mstart = image->segment[i].mem;
+               mend   = mstart + image->segment[i].memsz;
+               if ((end > mstart) && (start < mend)) {
+                       return 1;
+               }
+       }
+       return 0;
+}
+
+static struct page *kimage_alloc_pages(unsigned int gfp_mask, unsigned int order)
+{
+       struct page *pages;
+       pages = alloc_pages(gfp_mask, order);
+       if (pages) {
+               unsigned int count, i;
+               pages->mapping = NULL;
+               pages->private = order;
+               count = 1 << order;
+               for(i = 0; i < count; i++) {
+                       SetPageReserved(pages + i);
+               }
+       }
+       return pages;
+}
+
+static void kimage_free_pages(struct page *page)
+{
+       unsigned int order, count, i;
+       order = page->private;
+       count = 1 << order;
+       for(i = 0; i < count; i++) {
+               ClearPageReserved(page + i);
+       }
+       __free_pages(page, order);
+}
+
+static void kimage_free_page_list(struct list_head *list)
+{
+       struct list_head *pos, *next;
+       list_for_each_safe(pos, next, list) {
+               struct page *page;
+
+               page = list_entry(pos, struct page, lru);
+               list_del(&page->lru);
+
+               kimage_free_pages(page);
+       }
+}
+
+static struct page *kimage_alloc_normal_control_pages(
+       struct kimage *image, unsigned int order)
+{
+       /* Control pages are special, they are the intermediaries
+        * that are needed while we copy the rest of the pages
+        * to their final resting place.  As such they must
+        * not conflict with either the destination addresses
+        * or memory the kernel is already using.
+        *
+        * The only case where we really need more than one of
+        * these are for architectures where we cannot disable
+        * the MMU and must instead generate an identity mapped
+        * page table for all of the memory.
+        *
+        * At worst this runs in O(N) of the image size.
+        */
+       struct list_head extra_pages;
+       struct page *pages;
+       unsigned int count;
+
+       count = 1 << order;
+       INIT_LIST_HEAD(&extra_pages);
+
+       /* Loop while I can allocate a page and the page allocated
+        * is a destination page.
+        */
+       do {
+               unsigned long pfn, epfn, addr, eaddr;
+               pages = kimage_alloc_pages(GFP_KERNEL, order);
+               if (!pages)
+                       break;
+               pfn   = page_to_pfn(pages);
+               epfn  = pfn + count;
+               addr  = pfn << PAGE_SHIFT;
+               eaddr = epfn << PAGE_SHIFT;
+               if ((epfn >= (KEXEC_CONTROL_MEMORY_LIMIT >> PAGE_SHIFT)) ||
+                       kimage_is_destination_range(image, addr, eaddr))
+               {
+                       list_add(&pages->lru, &extra_pages);
+                       pages = NULL;
+               }
+       } while(!pages);
+       if (pages) {
+               /* Remember the allocated page... */
+               list_add(&pages->lru, &image->control_pages);
+
+               /* Because the page is already in it's destination
+                * location we will never allocate another page at
+                * that address.  Therefore kimage_alloc_pages
+                * will not return it (again) and we don't need
+                * to give it an entry in image->segment[].
+                */
+       }
+       /* Deal with the destination pages I have inadvertently allocated.
+        *
+        * Ideally I would convert multi-page allocations into single
+        * page allocations, and add everyting to image->dest_pages.
+        *
+        * For now it is simpler to just free the pages.
+        */
+       kimage_free_page_list(&extra_pages);
+       return pages;
+
+}
+
+static struct page *kimage_alloc_crash_control_pages(
+       struct kimage *image, unsigned int order)
+{
+       /* Control pages are special, they are the intermediaries
+        * that are needed while we copy the rest of the pages
+        * to their final resting place.  As such they must
+        * not conflict with either the destination addresses
+        * or memory the kernel is already using.
+        *
+        * Control pages are also the only pags we must allocate
+        * when loading a crash kernel.  All of the other pages
+        * are specified by the segments and we just memcpy
+        * into them directly.
+        *
+        * The only case where we really need more than one of
+        * these are for architectures where we cannot disable
+        * the MMU and must instead generate an identity mapped
+        * page table for all of the memory.
+        *
+        * Given the low demand this implements a very simple
+        * allocator that finds the first hole of the appropriate
+        * size in the reserved memory region, and allocates all
+        * of the memory up to and including the hole.
+        */
+       unsigned long hole_start, hole_end, size;
+       struct page *pages;
+       pages = NULL;
+       size = (1 << order) << PAGE_SHIFT;
+       hole_start = (image->control_page + (size - 1)) & ~(size - 1);
+       hole_end   = hole_start + size - 1;
+       while(hole_end <= crashk_res.end) {
+               unsigned long i;
+               if (hole_end > KEXEC_CONTROL_MEMORY_LIMIT) {
+                       break;
+               }
+               if (hole_end > crashk_res.end) {
+                       break;
+               }
+               /* See if I overlap any of the segments */
+               for(i = 0; i < image->nr_segments; i++) {
+                       unsigned long mstart, mend;
+                       mstart = image->segment[i].mem;
+                       mend   = mstart + image->segment[i].memsz - 1;
+                       if ((hole_end >= mstart) && (hole_start <= mend)) {
+                               /* Advance the hole to the end of the segment */
+                               hole_start = (mend + (size - 1)) & ~(size - 1);
+                               hole_end   = hole_start + size - 1;
+                               break;
+                       }
+               }
+               /* If I don't overlap any segments I have found my hole! */
+               if (i == image->nr_segments) {
+                       pages = pfn_to_page(hole_start >> PAGE_SHIFT);
+                       break;
+               }
+       }
+       if (pages) {
+               image->control_page = hole_end;
+       }
+       return pages;
+}
+
+
+struct page *kimage_alloc_control_pages(
+       struct kimage *image, unsigned int order)
+{
+       struct page *pages = NULL;
+       switch(image->type) {
+       case KEXEC_TYPE_DEFAULT:
+               pages = kimage_alloc_normal_control_pages(image, order);
+               break;
+       case KEXEC_TYPE_CRASH:
+               pages = kimage_alloc_crash_control_pages(image, order);
+               break;
+       }
+       return pages;
+}
+
+static int kimage_add_entry(struct kimage *image, kimage_entry_t entry)
+{
+       if (*image->entry != 0) {
+               image->entry++;
+       }
+       if (image->entry == image->last_entry) {
+               kimage_entry_t *ind_page;
+               struct page *page;
+               page = kimage_alloc_page(image, GFP_KERNEL, KIMAGE_NO_DEST);
+               if (!page) {
+                       return -ENOMEM;
+               }
+               ind_page = page_address(page);
+               *image->entry = virt_to_phys(ind_page) | IND_INDIRECTION;
+               image->entry = ind_page;
+               image->last_entry =
+                       ind_page + ((PAGE_SIZE/sizeof(kimage_entry_t)) - 1);
+       }
+       *image->entry = entry;
+       image->entry++;
+       *image->entry = 0;
+       return 0;
+}
+
+static int kimage_set_destination(
+       struct kimage *image, unsigned long destination)
+{
+       int result;
+
+       destination &= PAGE_MASK;
+       result = kimage_add_entry(image, destination | IND_DESTINATION);
+       if (result == 0) {
+               image->destination = destination;
+       }
+       return result;
+}
+
+
+static int kimage_add_page(struct kimage *image, unsigned long page)
+{
+       int result;
+
+       page &= PAGE_MASK;
+       result = kimage_add_entry(image, page | IND_SOURCE);
+       if (result == 0) {
+               image->destination += PAGE_SIZE;
+       }
+       return result;
+}
+
+
+static void kimage_free_extra_pages(struct kimage *image)
+{
+       /* Walk through and free any extra destination pages I may have */
+       kimage_free_page_list(&image->dest_pages);
+
+       /* Walk through and free any unuseable pages I have cached */
+       kimage_free_page_list(&image->unuseable_pages);
+
+}
+static int kimage_terminate(struct kimage *image)
+{
+       if (*image->entry != 0) {
+               image->entry++;
+       }
+       *image->entry = IND_DONE;
+       return 0;
+}
+
+#define for_each_kimage_entry(image, ptr, entry) \
+       for (ptr = &image->head; (entry = *ptr) && !(entry & IND_DONE); \
+               ptr = (entry & IND_INDIRECTION)? \
+                       phys_to_virt((entry & PAGE_MASK)): ptr +1)
+
+static void kimage_free_entry(kimage_entry_t entry)
+{
+       struct page *page;
+
+       page = pfn_to_page(entry >> PAGE_SHIFT);
+       kimage_free_pages(page);
+}
+
+static void kimage_free(struct kimage *image)
+{
+       kimage_entry_t *ptr, entry;
+       kimage_entry_t ind = 0;
+
+       if (!image)
+               return;
+       kimage_free_extra_pages(image);
+       for_each_kimage_entry(image, ptr, entry) {
+               if (entry & IND_INDIRECTION) {
+                       /* Free the previous indirection page */
+                       if (ind & IND_INDIRECTION) {
+                               kimage_free_entry(ind);
+                       }
+                       /* Save this indirection page until we are
+                        * done with it.
+                        */
+                       ind = entry;
+               }
+               else if (entry & IND_SOURCE) {
+                       kimage_free_entry(entry);
+               }
+       }
+       /* Free the final indirection page */
+       if (ind & IND_INDIRECTION) {
+               kimage_free_entry(ind);
+       }
+
+       /* Handle any machine specific cleanup */
+       machine_kexec_cleanup(image);
+
+       /* Free the kexec control pages... */
+       kimage_free_page_list(&image->control_pages);
+       kfree(image);
+}
+
+static kimage_entry_t *kimage_dst_used(struct kimage *image, unsigned long page)
+{
+       kimage_entry_t *ptr, entry;
+       unsigned long destination = 0;
+
+       for_each_kimage_entry(image, ptr, entry) {
+               if (entry & IND_DESTINATION) {
+                       destination = entry & PAGE_MASK;
+               }
+               else if (entry & IND_SOURCE) {
+                       if (page == destination) {
+                               return ptr;
+                       }
+                       destination += PAGE_SIZE;
+               }
+       }
+       return 0;
+}
+
+static struct page *kimage_alloc_page(struct kimage *image, unsigned int gfp_mask, unsigned long destination)
+{
+       /*
+        * Here we implement safeguards to ensure that a source page
+        * is not copied to its destination page before the data on
+        * the destination page is no longer useful.
+        *
+        * To do this we maintain the invariant that a source page is
+        * either its own destination page, or it is not a
+        * destination page at all.
+        *
+        * That is slightly stronger than required, but the proof
+        * that no problems will not occur is trivial, and the
+        * implementation is simply to verify.
+        *
+        * When allocating all pages normally this algorithm will run
+        * in O(N) time, but in the worst case it will run in O(N^2)
+        * time.   If the runtime is a problem the data structures can
+        * be fixed.
+        */
+       struct page *page;
+       unsigned long addr;
+
+       /*
+        * Walk through the list of destination pages, and see if I
+        * have a match.
+        */
+       list_for_each_entry(page, &image->dest_pages, lru) {
+               addr = page_to_pfn(page) << PAGE_SHIFT;
+               if (addr == destination) {
+                       list_del(&page->lru);
+                       return page;
+               }
+       }
+       page = NULL;
+       while (1) {
+               kimage_entry_t *old;
+
+               /* Allocate a page, if we run out of memory give up */
+               page = kimage_alloc_pages(gfp_mask, 0);
+               if (!page) {
+                       return 0;
+               }
+               /* If the page cannot be used file it away */
+               if (page_to_pfn(page) > (KEXEC_SOURCE_MEMORY_LIMIT >> PAGE_SHIFT)) {
+                       list_add(&page->lru, &image->unuseable_pages);
+                       continue;
+               }
+               addr = page_to_pfn(page) << PAGE_SHIFT;
+
+               /* If it is the destination page we want use it */
+               if (addr == destination)
+                       break;
+
+               /* If the page is not a destination page use it */
+               if (!kimage_is_destination_range(image, addr, addr + PAGE_SIZE))
+                       break;
+
+               /*
+                * I know that the page is someones destination page.
+                * See if there is already a source page for this
+                * destination page.  And if so swap the source pages.
+                */
+               old = kimage_dst_used(image, addr);
+               if (old) {
+                       /* If so move it */
+                       unsigned long old_addr;
+                       struct page *old_page;
+
+                       old_addr = *old & PAGE_MASK;
+                       old_page = pfn_to_page(old_addr >> PAGE_SHIFT);
+                       copy_highpage(page, old_page);
+                       *old = addr | (*old & ~PAGE_MASK);
+
+                       /* The old page I have found cannot be a
+                        * destination page, so return it.
+                        */
+                       addr = old_addr;
+                       page = old_page;
+                       break;
+               }
+               else {
+                       /* Place the page on the destination list I
+                        * will use it later.
+                        */
+                       list_add(&page->lru, &image->dest_pages);
+               }
+       }
+       return page;
+}
+
+static int kimage_load_normal_segment(struct kimage *image,
+       struct kexec_segment *segment)
+{
+       unsigned long maddr;
+       unsigned long ubytes, mbytes;
+       int result;
+       unsigned char *buf;
+
+       result = 0;
+       buf = segment->buf;
+       ubytes = segment->bufsz;
+       mbytes = segment->memsz;
+       maddr = segment->mem;
+
+       result = kimage_set_destination(image, maddr);
+       if (result < 0) {
+               goto out;
+       }
+       while(mbytes) {
+               struct page *page;
+               char *ptr;
+               size_t uchunk, mchunk;
+               page = kimage_alloc_page(image, GFP_HIGHUSER, maddr);
+               if (page == 0) {
+                       result  = -ENOMEM;
+                       goto out;
+               }
+               result = kimage_add_page(image, page_to_pfn(page) << PAGE_SHIFT);
+               if (result < 0) {
+                       goto out;
+               }
+               ptr = kmap(page);
+               /* Start with a clear page */
+               memset(ptr, 0, PAGE_SIZE);
+               ptr += maddr & ~PAGE_MASK;
+               mchunk = PAGE_SIZE - (maddr & ~PAGE_MASK);
+               if (mchunk > mbytes) {
+                       mchunk = mbytes;
+               }
+               uchunk = mchunk;
+               if (uchunk > ubytes) {
+                       uchunk = ubytes;
+               }
+               result = copy_from_user(ptr, buf, uchunk);
+               kunmap(page);
+               if (result) {
+                       result = (result < 0) ? result : -EIO;
+                       goto out;
+               }
+               ubytes -= uchunk;
+               maddr  += mchunk;
+               buf    += mchunk;
+               mbytes -= mchunk;
+       }
+ out:
+       return result;
+}
+
+static int kimage_load_crash_segment(struct kimage *image,
+       struct kexec_segment *segment)
+{
+       /* For crash dumps kernels we simply copy the data from
+        * user space to it's destination.
+        * We do things a page at a time for the sake of kmap.
+        */
+       unsigned long maddr;
+       unsigned long ubytes, mbytes;
+       int result;
+       unsigned char *buf;
+
+       result = 0;
+       buf = segment->buf;
+       ubytes = segment->bufsz;
+       mbytes = segment->memsz;
+       maddr = segment->mem;
+       while(mbytes) {
+               struct page *page;
+               char *ptr;
+               size_t uchunk, mchunk;
+               page = pfn_to_page(maddr >> PAGE_SHIFT);
+               if (page == 0) {
+                       result  = -ENOMEM;
+                       goto out;
+               }
+               ptr = kmap(page);
+               ptr += maddr & ~PAGE_MASK;
+               mchunk = PAGE_SIZE - (maddr & ~PAGE_MASK);
+               if (mchunk > mbytes) {
+                       mchunk = mbytes;
+               }
+               uchunk = mchunk;
+               if (uchunk > ubytes) {
+                       uchunk = ubytes;
+                       /* Zero the trailing part of the page */
+                       memset(ptr + uchunk, 0, mchunk - uchunk);
+               }
+               result = copy_from_user(ptr, buf, uchunk);
+               kunmap(page);
+               if (result) {
+                       result = (result < 0) ? result : -EIO;
+                       goto out;
+               }
+               ubytes -= uchunk;
+               maddr  += mchunk;
+               buf    += mchunk;
+               mbytes -= mchunk;
+       }
+ out:
+       return result;
+}
+
+static int kimage_load_segment(struct kimage *image,
+       struct kexec_segment *segment)
+{
+       int result = -ENOMEM;
+       switch(image->type) {
+       case KEXEC_TYPE_DEFAULT:
+               result = kimage_load_normal_segment(image, segment);
+               break;
+       case KEXEC_TYPE_CRASH:
+               result = kimage_load_crash_segment(image, segment);
+               break;
+       }
+       return result;
+}
+
+/*
+ * Exec Kernel system call: for obvious reasons only root may call it.
+ *
+ * This call breaks up into three pieces.
+ * - A generic part which loads the new kernel from the current
+ *   address space, and very carefully places the data in the
+ *   allocated pages.
+ *
+ * - A generic part that interacts with the kernel and tells all of
+ *   the devices to shut down.  Preventing on-going dmas, and placing
+ *   the devices in a consistent state so a later kernel can
+ *   reinitialize them.
+ *
+ * - A machine specific part that includes the syscall number
+ *   and the copies the image to it's final destination.  And
+ *   jumps into the image at entry.
+ *
+ * kexec does not sync, or unmount filesystems so if you need
+ * that to happen you need to do that yourself.
+ */
+struct kimage *kexec_image = NULL;
+static struct kimage *kexec_crash_image = NULL;
+/*
+ * A home grown binary mutex.
+ * Nothing can wait so this mutex is safe to use
+ * in interrupt context :)
+ */
+static int kexec_lock = 0;
+
+asmlinkage long sys_kexec_load(unsigned long entry,
+       unsigned long nr_segments, struct kexec_segment __user *segments,
+       unsigned long flags)
+{
+       struct kimage **dest_image, *image;
+       int locked;
+       int result;
+
+       /* We only trust the superuser with rebooting the system. */
+       if (!capable(CAP_SYS_BOOT))
+               return -EPERM;
+
+       /*
+        * Verify we have a legal set of flags
+        * This leaves us room for future extensions.
+        */
+       if ((flags & KEXEC_FLAGS) != (flags & ~KEXEC_ARCH_MASK))
+               return -EINVAL;
+
+       /* Verify we are on the appropriate architecture */
+       if (((flags & KEXEC_ARCH_MASK) != KEXEC_ARCH) &&
+               ((flags & KEXEC_ARCH_MASK) != KEXEC_ARCH_DEFAULT))
+       {
+               return -EINVAL;
+       }
+
+       /* Put an artificial cap on the number
+        * of segments passed to kexec_load.
+        */
+       if (nr_segments > KEXEC_SEGMENT_MAX)
+               return -EINVAL;
+
+       image = NULL;
+       result = 0;
+
+       /* Because we write directly to the reserved memory
+        * region when loading crash kernels we need a mutex here to
+        * prevent multiple crash  kernels from attempting to load
+        * simultaneously, and to prevent a crash kernel from loading
+        * over the top of a in use crash kernel.
+        *
+        * KISS: always take the mutex.
+        */
+       locked = xchg(&kexec_lock, 1);
+       if (locked) {
+               return -EBUSY;
+       }
+       dest_image = &kexec_image;
+       if (flags & KEXEC_ON_CRASH) {
+               dest_image = &kexec_crash_image;
+       }
+       if (nr_segments > 0) {
+               unsigned long i;
+               /* Loading another kernel to reboot into */
+               if ((flags & KEXEC_ON_CRASH) == 0) {
+                       result = kimage_normal_alloc(&image, entry, nr_segments, segments);
+               }
+               /* Loading another kernel to switch to if this one crashes */
+               else if (flags & KEXEC_ON_CRASH) {
+                       /* Free any current crash dump kernel before
+                        * we corrupt it.
+                        */
+                       kimage_free(xchg(&kexec_crash_image, NULL));
+                       result = kimage_crash_alloc(&image, entry, nr_segments, segments);
+               }
+               if (result) {
+                       goto out;
+               }
+               result = machine_kexec_prepare(image);
+               if (result) {
+                       goto out;
+               }
+               for(i = 0; i < nr_segments; i++) {
+                       result = kimage_load_segment(image, &image->segment[i]);
+                       if (result) {
+                               goto out;
+                       }
+               }
+               result = kimage_terminate(image);
+               if (result) {
+                       goto out;
+               }
+       }
+       /* Install the new kernel, and  Uninstall the old */
+       image = xchg(dest_image, image);
+
+ out:
+       xchg(&kexec_lock, 0); /* Release the mutex */
+       kimage_free(image);
+       return result;
+}
+
+#ifdef CONFIG_COMPAT
+asmlinkage long compat_sys_kexec_load(unsigned long entry,
+       unsigned long nr_segments, struct compat_kexec_segment __user *segments,
+       unsigned long flags)
+{
+       struct compat_kexec_segment in;
+       struct kexec_segment out, __user *ksegments;
+       unsigned long i, result;
+
+       /* Don't allow clients that don't understand the native
+        * architecture to do anything.
+        */
+       if ((flags & KEXEC_ARCH_MASK) == KEXEC_ARCH_DEFAULT) {
+               return -EINVAL;
+       }
+
+       if (nr_segments > KEXEC_SEGMENT_MAX) {
+               return -EINVAL;
+       }
+
+       ksegments = compat_alloc_user_space(nr_segments * sizeof(out));
+       for (i=0; i < nr_segments; i++) {
+               result = copy_from_user(&in, &segments[i], sizeof(in));
+               if (result) {
+                       return -EFAULT;
+               }
+
+               out.buf   = compat_ptr(in.buf);
+               out.bufsz = in.bufsz;
+               out.mem   = in.mem;
+               out.memsz = in.memsz;
+
+               result = copy_to_user(&ksegments[i], &out, sizeof(out));
+               if (result) {
+                       return -EFAULT;
+               }
+       }
+
+       return sys_kexec_load(entry, nr_segments, ksegments, flags);
+}
+#endif
+
+void crash_kexec(void)
+{
+       struct kimage *image;
+       int locked;
+
+
+       /* Take the kexec_lock here to prevent sys_kexec_load
+        * running on one cpu from replacing the crash kernel
+        * we are using after a panic on a different cpu.
+        *
+        * If the crash kernel was not located in a fixed area
+        * of memory the xchg(&kexec_crash_image) would be
+        * sufficient.  But since I reuse the memory...
+        */
+       locked = xchg(&kexec_lock, 1);
+       if (!locked) {
+               image = xchg(&kexec_crash_image, NULL);
+               if (image) {
+                       machine_crash_shutdown();
+                       machine_kexec(image);
+               }
+               xchg(&kexec_lock, 0);
+       }
+}
index 081f7465fc8dec17d3e69a5c9ff7c29f8165ad5b..66f43d33cd80a7cdab94c83fbff18212f6d46cd5 100644 (file)
@@ -18,6 +18,7 @@
 #include <linux/sysrq.h>
 #include <linux/interrupt.h>
 #include <linux/nmi.h>
+#include <linux/kexec.h>
 
 int panic_timeout;
 int panic_on_oops;
@@ -63,6 +64,13 @@ NORET_TYPE void panic(const char * fmt, ...)
         unsigned long caller = (unsigned long) __builtin_return_address(0);
 #endif
 
+       /*
+        * It's possible to come here directly from a panic-assertion and not
+        * have preempt disabled. Some functions called from here want
+        * preempt to be disabled. No point enabling it later though...
+        */
+       preempt_disable();
+
        bust_spinlocks(1);
        va_start(args, fmt);
        vsnprintf(buf, sizeof(buf), fmt, args);
@@ -70,7 +78,19 @@ NORET_TYPE void panic(const char * fmt, ...)
        printk(KERN_EMERG "Kernel panic - not syncing: %s\n",buf);
        bust_spinlocks(0);
 
+       /*
+        * If we have crashed and we have a crash kernel loaded let it handle
+        * everything else.
+        * Do we want to call this before we try to display a message?
+        */
+       crash_kexec();
+
 #ifdef CONFIG_SMP
+       /*
+        * Note smp_send_stop is the usual smp shutdown function, which
+        * unfortunately means it may not be hardened to work in a panic
+        * situation.
+        */
        smp_send_stop();
 #endif
 
@@ -79,8 +99,7 @@ NORET_TYPE void panic(const char * fmt, ...)
        if (!panic_blink)
                panic_blink = no_blink;
 
-       if (panic_timeout > 0)
-       {
+       if (panic_timeout > 0) {
                /*
                 * Delay timeout seconds before rebooting the machine. 
                 * We can't use the "normal" timers since we just panicked..
index dac10161ca2384c59340d4fab96eae4237f2f8af..9a24374c23bc4f69116d264991689b1aaa9c38fe 100644 (file)
@@ -16,6 +16,8 @@
 #include <linux/init.h>
 #include <linux/highuid.h>
 #include <linux/fs.h>
+#include <linux/kernel.h>
+#include <linux/kexec.h>
 #include <linux/workqueue.h>
 #include <linux/device.h>
 #include <linux/key.h>
@@ -439,6 +441,24 @@ asmlinkage long sys_reboot(int magic1, int magic2, unsigned int cmd, void __user
                machine_restart(buffer);
                break;
 
+#ifdef CONFIG_KEXEC
+       case LINUX_REBOOT_CMD_KEXEC:
+       {
+               struct kimage *image;
+               image = xchg(&kexec_image, 0);
+               if (!image) {
+                       unlock_kernel();
+                       return -EINVAL;
+               }
+               notifier_call_chain(&reboot_notifier_list, SYS_RESTART, NULL);
+               system_state = SYSTEM_RESTART;
+               device_shutdown();
+               printk(KERN_EMERG "Starting new kernel\n");
+               machine_shutdown();
+               machine_kexec(image);
+               break;
+       }
+#endif
 #ifdef CONFIG_SOFTWARE_SUSPEND
        case LINUX_REBOOT_CMD_SW_SUSPEND:
                {
index 6f15bea7d1a83dc363826c5bcdb2c0013feaebc0..29196ce9b40f09a9398d78e1c4ee555dd93f6c04 100644 (file)
@@ -18,6 +18,8 @@ cond_syscall(sys_acct);
 cond_syscall(sys_lookup_dcookie);
 cond_syscall(sys_swapon);
 cond_syscall(sys_swapoff);
+cond_syscall(sys_kexec_load);
+cond_syscall(compat_sys_kexec_load);
 cond_syscall(sys_init_module);
 cond_syscall(sys_delete_module);
 cond_syscall(sys_socketpair);