KEXEC fixes

Quoted 1 line

Quoted 1 line · Andrew Cooper

Currently, KEXEC_CMD_kexec_get_range in compat mode, or with 32bit Xen, will truncate 64bit pointers to 32bits, leading to problems on machines with more than 4GiB of RAM. As 32bit dom0 kernels tend to be PAE capable, this is especially wasteful, as most structures currently limited to <4GiB could easily be <64GiB instead.

Therefore, introduce a new hypercall 'KEXEC_CMD_kexec64_get_range' which passes similar information as KEXEC_CMD_kexec_get_range, but which uses a structure with explicitly specified field widths, causing it to be identical in the compat and regular case. This new structure, xen_kexec64_range_t, will be the same as xen_kexec_range_t if Xen is compiled for x86_64, but will still use 64bit integers if Xen is compiled for x86_32.

To fix 32bit Xen which uses 32bit integers for addresses and sizes, change the internals to use xen_kexec64_range_t which will use 64bit integers instead. This also involves changing several casts to explicitly use uint64_ts rather than unsigned longs.

In addition, the hypercall entry points need updating to be able to cater for all possibilities.

|Xen/dom0 bits|          Bit width of addresses in structure for        |
+------+------+---------------------------+-----------------------------+
| Xen  | dom0 | KEXEC_CMD_kexec_get_range | KEXEC_CMD_kexec64_get_range |
+------+------+---------------------------+-----------------------------+
|  32  |  32  |             32            |              64             |
|  64  |  32  |             32            |              64             |
|  64  |  64  |             64            |              64             |
+------+------+---------------------------+-----------------------------+

This has been solved by splitting do_kexec_op_internal back into do_kexec_op{,_compat} and changing kexec_get_range{,_compat} to kexec_get_range{32,64} which now exist irrespective of CONFIG_COMPAT or not.

The knock-on effect of removing do_kexec_op_internal means that kexec_load_unload_compat is only ever called from inside an #ifdef CONFIG_COMPAT codepath, which does not exist on Xen x86_32. Therefore, move the #ifdef CONFIG_COMPAT guards to include the entire function.

Finally, and a little unrelatedly, fix KEXEC_CMD_kexec_{load,unload} to return -EBUSY instead of EOK if a kexec is in progress.

Changes since v1:
* Fix check for pointer truncation to work when Xen is compiled for 32 bit mode as well.

Signed-off-by Andrew Cooper <[email protected]>

diff -r e56500f95b6a -r f571dc8e4368 xen/arch/ia64/xen/machine_kexec.c

xen/arch/ia64/xen/machine_kexec.c +8 -8

@@ -102,10 +102,10 @@ void machine_reboot_kexec(xen_kexec_imag
 	machine_kexec(image);
 }
 
-int machine_kexec_get_xen(xen_kexec_range_t *range)
+int machine_kexec_get_xen(xen_kexec64_range_t *range)
 {
 	range->start = ia64_tpa(_text);
-	range->size = (unsigned long)_end - (unsigned long)_text;
+	range->size = (uint64_t)_end - (uint64_t)_text;
 	return 0;
 }
 

@@ -113,31 +113,31 @@ int machine_kexec_get_xen(xen_kexec_rang
 #define ELF_PAGE_SIZE  (__IA64_UL_CONST(1) << ELF_PAGE_SHIFT)
 #define ELF_PAGE_MASK  (~(ELF_PAGE_SIZE - 1))
 
-static int machine_kexec_get_xenheap(xen_kexec_range_t *range)
+static int machine_kexec_get_xenheap(xen_kexec64_range_t *range)
 {
 	range->start = (ia64_tpa(_end) + (ELF_PAGE_SIZE - 1)) & ELF_PAGE_MASK;
 	range->size =
-		(((unsigned long)range->start + KERNEL_TR_PAGE_SIZE) &
+		(((uint64_t)range->start + KERNEL_TR_PAGE_SIZE) &
          ~(KERNEL_TR_PAGE_SIZE - 1))
-		- (unsigned long)range->start;
+		- (uint64_t)range->start;
 	return 0;
 }
 
-static int machine_kexec_get_boot_param(xen_kexec_range_t *range)
+static int machine_kexec_get_boot_param(xen_kexec64_range_t *range)
 {
 	range->start = __pa(ia64_boot_param);
 	range->size = sizeof(*ia64_boot_param);
 	return 0;
 }
 
-static int machine_kexec_get_efi_memmap(xen_kexec_range_t *range)
+static int machine_kexec_get_efi_memmap(xen_kexec64_range_t *range)
 {
 	range->start = ia64_boot_param->efi_memmap;
 	range->size = ia64_boot_param->efi_memmap_size;
 	return 0;
 }
 
-int machine_kexec_get(xen_kexec_range_t *range)
+int machine_kexec_get(xen_kexec64_range_t *range)
 {
 	switch (range->range) {
 	case KEXEC_RANGE_MA_XEN:

diff -r e56500f95b6a -r f571dc8e4368 xen/arch/x86/machine_kexec.c

xen/arch/x86/machine_kexec.c +1 -1

@@ -120,7 +120,7 @@ void machine_kexec(xen_kexec_image_t *im
     }
 }
 
-int machine_kexec_get(xen_kexec_range_t *range)
+int machine_kexec_get(xen_kexec64_range_t *range)
 {
 	if (range->range != KEXEC_RANGE_MA_XEN)
 		return -EINVAL;

diff -r e56500f95b6a -r f571dc8e4368 xen/arch/x86/x86_32/machine_kexec.c

xen/arch/x86/x86_32/machine_kexec.c +3 -3

@@ -11,11 +11,11 @@
 #include <asm/page.h>
 #include <public/kexec.h>
 
-int machine_kexec_get_xen(xen_kexec_range_t *range)
+int machine_kexec_get_xen(xen_kexec64_range_t *range)
 {
         range->start = virt_to_maddr(_start);
-        range->size = (unsigned long)xenheap_phys_end -
-                      (unsigned long)range->start;
+        range->size = (uint64_t)xenheap_phys_end -
+                      (uint64_t)range->start;
         return 0;
 }
 

diff -r e56500f95b6a -r f571dc8e4368 xen/arch/x86/x86_64/machine_kexec.c

xen/arch/x86/x86_64/machine_kexec.c +2 -2

@@ -11,10 +11,10 @@
 #include <asm/page.h>
 #include <public/kexec.h>
 
-int machine_kexec_get_xen(xen_kexec_range_t *range)
+int machine_kexec_get_xen(xen_kexec64_range_t *range)
 {
         range->start = virt_to_maddr(_start);
-        range->size = virt_to_maddr(_end) - (unsigned long)range->start;
+        range->size = virt_to_maddr(_end) - (uint64_t)range->start;
         return 0;
 }
 

diff -r e56500f95b6a -r f571dc8e4368 xen/common/kexec.c

xen/common/kexec.c +37 -18

@@ -56,6 +56,14 @@ static struct {
     unsigned long size;
 } ranges[16] __initdata;
 
+/* This XLAT macro is required now even without CONFIG_COMPAT. */
+#define TRANSLATE_kexec_range(_d_, _s_) do { \
+    (_d_)->range = (_s_)->range; \
+    (_d_)->nr = (_s_)->nr; \
+    (_d_)->size = (_s_)->size; \
+    (_d_)->start = (_s_)->start; \
+} while (0)
+
 /*
  * Parse command lines in the format
  *

@@ -280,7 +288,7 @@ static int sizeof_note(const char *name,
             ELFNOTE_ALIGN(descsz));
 }
 
-static int kexec_get_reserve(xen_kexec_range_t *range)
+static int kexec_get_reserve(xen_kexec64_range_t *range)
 {
     if ( kexec_crash_area.size > 0 && kexec_crash_area.start > 0) {
         range->start = kexec_crash_area.start;

@@ -291,7 +299,7 @@ static int kexec_get_reserve(xen_kexec_r
     return 0;
 }
 
-static int kexec_get_cpu(xen_kexec_range_t *range)
+static int kexec_get_cpu(xen_kexec64_range_t *range)
 {
     int nr = range->nr;
     int nr_bytes = 0;

@@ -335,14 +343,14 @@ static int kexec_get_cpu(xen_kexec_range
     return 0;
 }
 
-static int kexec_get_vmcoreinfo(xen_kexec_range_t *range)
+static int kexec_get_vmcoreinfo(xen_kexec64_range_t *range)
 {
     range->start = __pa((unsigned long)vmcoreinfo_data);
     range->size = VMCOREINFO_BYTES;
     return 0;
 }
 
-static int kexec_get_range_internal(xen_kexec_range_t *range)
+static int kexec_get_range_internal(xen_kexec64_range_t *range)
 {
     int ret = -EINVAL;
 

@@ -365,9 +373,14 @@ static int kexec_get_range_internal(xen_
     return ret;
 }
 
-static int kexec_get_range(XEN_GUEST_HANDLE(void) uarg)
+/* This function can be invoked from either a KEXEC_CMD_kexec64_get_range
+ * hypercall, or from a KEXEC_CMD_kexec_get_range hypercall with 64bit dom0
+ * on 64bit Xen.  In the first case, the guest structure is a 
+ * xen_kexec64_range_t, and in the second case, the xen_kexec_range_t guest
+ * structure is identical to xen_kexec64_range_t. */
+static int kexec_get_range64(XEN_GUEST_HANDLE(void) uarg)
 {
-    xen_kexec_range_t range;
+    xen_kexec64_range_t range;
     int ret = -EINVAL;
 
     if ( unlikely(copy_from_guest(&range, uarg, 1)) )

@@ -381,34 +394,40 @@ static int kexec_get_range(XEN_GUEST_HAN
     return ret;
 }
 
-static int kexec_get_range_compat(XEN_GUEST_HANDLE(void) uarg)
+/* This function can be invoked from either a KEXEC_CMD_kexec_get_range
+ * compat hypercall for 32bit dom0 on 64bit Xen, or from the same hypercall
+ * on 32bit Xen.  In both cases, the guest argument uses 32bit integers, so 
 
+ * translate them to 64bit for use by kexec_get_range_internal.  The
+ * preprocessor guards are to choose the correct 32bit structure, as the
+ * compat_* structures dont exist in 32bit Xen. */
+static int kexec_get_range32(XEN_GUEST_HANDLE(void) uarg)
 {
+    xen_kexec64_range_t range64;
 #ifdef CONFIG_COMPAT
-    xen_kexec_range_t range;
-    compat_kexec_range_t compat_range;
+    compat_kexec_range_t range32;
+#else
+    xen_kexec_range_t range32;
+#endif
     int ret = -EINVAL;
 
-    if ( unlikely(copy_from_guest(&compat_range, uarg, 1)) )
+    if ( unlikely(copy_from_guest(&range32, uarg, 1)) )
         return -EFAULT;
 
-    XLAT_kexec_range(&range, &compat_range);
+    TRANSLATE_kexec_range(&range64, &range32);
 
-    ret = kexec_get_range_internal(&range);
+    ret = kexec_get_range_internal(&range64);
 
     /* Dont silently truncate physical addresses or sizes. */
-    if ( (range.start | range.size) & ~(unsigned long)(~0u) )
+    if ( (range64.start | range64.size) & 0xffffffff00000000ULL )
         return -ERANGE;
 
     if ( ret == 0 ) {
-        XLAT_kexec_range(&compat_range, &range);
-        if ( unlikely(copy_to_guest(uarg, &compat_range, 1)) )
+        TRANSLATE_kexec_range(&range32, &range64);
+        if ( unlikely(copy_to_guest(uarg, &range32, 1)) )
              return -EFAULT;
     }
 
     return ret;
-#else /* CONFIG_COMPAT */
-    return 0;
-#endif /* CONFIG_COMPAT */
 }
 

static int kexec_load_get_bits(int type, int *base, int *bit)

xen/common/kexec.c +55 -27

@@ -543,10 +562,10 @@ static int kexec_load_unload(unsigned lo
     return kexec_load_unload_internal(op, &load);
 }
 
+#ifdef CONFIG_COMPAT
 static int kexec_load_unload_compat(unsigned long op,
                                     XEN_GUEST_HANDLE(void) uarg)
 {
-#ifdef CONFIG_COMPAT
     compat_kexec_load_t compat_load;
     xen_kexec_load_t load;
 

@@ -564,10 +583,8 @@ static int kexec_load_unload_compat(unsi
     XLAT_kexec_image(&load.image, &compat_load.image);
 
     return kexec_load_unload_internal(op, &load);
-#else /* CONFIG_COMPAT */
-    return 0;
+}
 #endif /* CONFIG_COMPAT */
-}
 
 static int kexec_exec(XEN_GUEST_HANDLE(void) uarg)
 {

@@ -601,8 +618,51 @@ static int kexec_exec(XEN_GUEST_HANDLE(v
     return -EINVAL; /* never reached */
 }
 
-int do_kexec_op_internal(unsigned long op, XEN_GUEST_HANDLE(void) uarg,
-                           int compat)
+long do_kexec_op(unsigned long op, XEN_GUEST_HANDLE(void) uarg)
+{
+    unsigned long flags;
+    int ret = -EINVAL;
+
+    if ( !IS_PRIV(current->domain) )
+        return -EPERM;
+
+    ret = xsm_kexec();
+    if ( ret )
+        return ret;
+
+    switch ( op )
+    {
+#ifdef __i386__
+    case KEXEC_CMD_kexec_get_range:
+        ret = kexec_get_range32(uarg);
+        break;
+#else
+    case KEXEC_CMD_kexec_get_range:
+#endif
+    case KEXEC_CMD_kexec64_get_range:
+        ret = kexec_get_range64(uarg);
+        break;
+
+    case KEXEC_CMD_kexec_load:
+    case KEXEC_CMD_kexec_unload:
+        spin_lock_irqsave(&kexec_lock, flags);
+        if (!test_bit(KEXEC_FLAG_IN_PROGRESS, &kexec_flags))
+            ret = kexec_load_unload(op, uarg);
+        else
+            ret = -EBUSY;
+        spin_unlock_irqrestore(&kexec_lock, flags);
+        break;
+
+    case KEXEC_CMD_kexec:
+        ret = kexec_exec(uarg);
+        break;
+    }
+
+    return ret;
+}
+
+#ifdef CONFIG_COMPAT
+int compat_kexec_op(unsigned long op, XEN_GUEST_HANDLE(void) uarg)
 {
     unsigned long flags;
     int ret = -EINVAL;

@@ -617,23 +677,21 @@ int do_kexec_op_internal(unsigned long o
     switch ( op )
     {
     case KEXEC_CMD_kexec_get_range:
-        if (compat)
-                ret = kexec_get_range_compat(uarg);
-        else
-                ret = kexec_get_range(uarg);
+        ret = kexec_get_range32(uarg);
+        break;
+    case KEXEC_CMD_kexec64_get_range:
+        ret = kexec_get_range64(uarg);
         break;
     case KEXEC_CMD_kexec_load:
     case KEXEC_CMD_kexec_unload:
         spin_lock_irqsave(&kexec_lock, flags);
         if (!test_bit(KEXEC_FLAG_IN_PROGRESS, &kexec_flags))
-        {
-                if (compat)
-                        ret = kexec_load_unload_compat(op, uarg);
-                else
-                        ret = kexec_load_unload(op, uarg);
-        }
+            ret = kexec_load_unload_compat(op, uarg);
+        else
+            ret = -EBUSY;
         spin_unlock_irqrestore(&kexec_lock, flags);
         break;
+
     case KEXEC_CMD_kexec:
         ret = kexec_exec(uarg);
         break;

@@ -641,17 +699,6 @@ int do_kexec_op_internal(unsigned long o
 
     return ret;
 }
-
-long do_kexec_op(unsigned long op, XEN_GUEST_HANDLE(void) uarg)
-{
-    return do_kexec_op_internal(op, uarg, 0);
-}
-
-#ifdef CONFIG_COMPAT
-int compat_kexec_op(unsigned long op, XEN_GUEST_HANDLE(void) uarg)
-{
-    return do_kexec_op_internal(op, uarg, 1);
-}
 #endif
 
 /*

diff -r e56500f95b6a -r f571dc8e4368 xen/include/public/kexec.h

xen/include/public/kexec.h +9 -0

@@ -155,6 +155,15 @@ typedef struct xen_kexec_range {
     unsigned long start;
 } xen_kexec_range_t;
 
+#define KEXEC_CMD_kexec64_get_range      4
+/* xen_kexec_range_t using explicit sizes for fields. */
+typedef struct xen_kexec64_range {
+    int32_t range;
+    int32_t nr;
+    uint64_t size;
+    uint64_t start;
+} xen_kexec64_range_t;
+
 #endif /* _XEN_PUBLIC_KEXEC_H */
 
 /*

diff -r e56500f95b6a -r f571dc8e4368 xen/include/xen/kexec.h

xen/include/xen/kexec.h +2 -2

@@ -34,8 +34,8 @@ void kexec_crash(void);
 void kexec_crash_save_cpu(void);
 crash_xen_info_t *kexec_crash_save_info(void);
 void machine_crash_shutdown(void);
-int machine_kexec_get(xen_kexec_range_t *range);
-int machine_kexec_get_xen(xen_kexec_range_t *range);
+int machine_kexec_get(xen_kexec64_range_t *range);
+int machine_kexec_get_xen(xen_kexec64_range_t *range);
 
 void compat_machine_kexec(unsigned long rnk,
                           unsigned long indirection_page,

diff -r e56500f95b6a -r f571dc8e4368 xen/include/xlat.lst

xen/include/xlat.lst +0 -2

@@ -50,9 +50,7 @@
 ?	grant_entry_v1			grant_table.h
 ?       grant_entry_header              grant_table.h
 ?	grant_entry_v2			grant_table.h
-?	kexec_exec			kexec.h
 !	kexec_image			kexec.h
-!	kexec_range			kexec.h
 !	add_to_physmap			memory.h
 !	foreign_memory_map		memory.h
 !	memory_exchange			memory.h

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Quoted 15 lines

Quoted 1 line

Quoted 1 line · Andrew Cooper

Currently, KEXEC_CMD_kexec_get_range in compat mode, or with 32bit Xen, will truncate 64bit pointers to 32bits, leading to problems on machines with more than 4GiB of RAM. As 32bit dom0 kernels tend to be PAE capable, this is especially wasteful, as most structures currently limited to <4GiB could easily be <64GiB instead.

Therefore, introduce a new hypercall 'KEXEC_CMD_kexec64_get_range' which passes similar information as KEXEC_CMD_kexec_get_range, but which uses a structure with explicitly specified field widths, causing it to be identical in the compat and regular case. This new structure, xen_kexec64_range_t, will be the same as xen_kexec_range_t if Xen is compiled for x86_64, but will still use 64bit integers if Xen is compiled for x86_32.

Quoted 4 lines

To fix 32bit Xen which uses 32bit integers for addresses and sizes, change the internals to use xen_kexec64_range_t which will use 64bit integers instead. This also involves changing several casts to explicitly use uint64_ts rather than unsigned longs.

Just for the record - I continue to be opposed to doing this for the 32-bit hypervisor. All relevant allocations are below 4G there, so there's simply no need to decrease code readability for no benefit.

Jan

Quoted 5 lines

Just for the record - I continue to be opposed to doing this for the 32-bit hypervisor. All relevant allocations are below 4G there, so there's simply no need to decrease code readability for no benefit.

Jan

I don't mind reducing the scope to ignore the 32bit Xen case. Any object Keir?

Quoted 1 line

Quoted 1 line · Andrew Cooper

Currently, the buffers for crash notes are allocated per CPU when a KEXEC_CMD_kexec_get_range hypercall is made, referencing the CPU in question.

Although it certainly works in general, there are a few edge case problems:

1) There appears to be no guarentee that dom0 will make this hypercall for each pcpu on the system. There appears to be variable behaviour depending on how many cpus dom0 is compiled for, and how many vcpus Xen gives to dom0.

2) The allocation of these buffers occur at the whim of dom0. While this is typically very early on dom0 boot, but not guarenteed.

3) It is possible (although not sensible) for a crash kernel to be loaded without these (or some of these) hypercalls being made. Under these circumstances, a crash would cause the crash note code path will suffer a slew of null pointer deferences.

4) If the hypercalls are made after late in the day, it is possible for the hypercall to return -ENOMEM. As code tends to be more fragile once memory is enhausted, the likelyhood of us needing the crash kernel is greater.

In addition, my forthcoming code to support 32bit kdump kernels on 64bit Xen on large (>64GB) boxes will require some guarentees as to where the crash note buffers are actually allocated in physical memory. This is far easier to sort out at boot time, rather than after dom0 has been booted and potentially using the physical memory required.

Therefore, allocate the crash note buffers at boot time.

Changes since v5:
* Introduce sizeof_cpu_notes to move calculation of note size into a separate location.
* Tweak sizeof_note() to return size_t rather than int, as it is a function based upon sizeof().

Changes since v4:

* Replace the current cpu crash note scheme of using void pointers and hand calculating the size each time is needed, by a range structure containing a pointer and a size. This removes duplicate times where the size is calculated.
* Tweak kexec_get_cpu(). Don't fail if a cpu is offline because it may already have crash notes, and may be up by the time a crash happens. Split the error conditions up to return ERANGE for an out-of-range cpu request rather than EINVAL. Finally, returning a range of zeros is acceptable, so do this in preference to failing.

Changes since v3:

* Alter the spinlocks to avoid calling xmalloc/xfree while holding the lock.
* Tidy up the coding style used.

Changes since v2:

* Allocate crash_notes dynamically using nr_cpu_ids at boot time, rather than statically using NR_CPUS.
* Fix the incorrect use of signed integers for cpu id.
* Fix collateral damage to do_crashdump_trigger() and crashdump_trigger_handler caused by reordering sizeof_note() and setup_note()
* Tweak the issue about returing -ENOMEM from kexec_init_cpu_note(). No functional change.
* Change kexec_get_cpu() to attempt to allocate crash note buffers in case we have more free memory now than when the pcpu came up.
* Now that there are two codepaths possibly allocating crash notes, protect the allocation itself with a spinlock.

Changes since v1:

* Use cpu hotplug notifiers to handle allocating of the notes buffers rather than assuming the boot state of cpus will be the same as the crash state.
* Move crash_notes from being per_cpu. This is because the kdump kernel elf binary put in the crash area is hard coded to physical addresses which the dom0 kernel typically obtains at boot time. If a cpu is offlined, its buffer should not be deallocated because the kdump kernel would read junk when trying to get the crash notes. Similarly, the same problem would occur if the cpu was re-onlined later and its crash notes buffer was allocated elsewhere.
* Only attempt to allocate buffers if a crash area has been specified. Else, allocating crash note buffers is a waste of space. Along with this, change the test in kexec_get_cpu to return -EINVAL if no buffers have been allocated.

Signed-off-by Andrew Cooper <[email protected]>

diff -r f571dc8e4368 -r 3da37c68284f xen/common/kexec.c

xen/common/kexec.c +45 -10

@@ -25,13 +25,19 @@
 #include <xen/kexec.h>
 #include <public/elfnote.h>
 #include <xsm/xsm.h>
+#include <xen/cpu.h>
 #ifdef CONFIG_COMPAT
 #include <compat/kexec.h>
 #endif
 
 bool_t kexecing = FALSE;
 
-static DEFINE_PER_CPU_READ_MOSTLY(void *, crash_notes);
+/* Memory regions to store the per cpu register state etc. on a crash. */
+typedef struct { Elf_Note * start; size_t size; } crash_note_range_t;
+static crash_note_range_t * crash_notes;
+
+/* Lock to prevent race conditions when allocating the crash note buffers. 
 */
+static DEFINE_SPINLOCK(crash_notes_lock);
 
 static Elf_Note *xen_crash_note;

@@ -173,13 +179,17 @@ static void one_cpu_only(void)
 void kexec_crash_save_cpu(void)
 {
     int cpu = smp_processor_id();
-    Elf_Note *note = per_cpu(crash_notes, cpu);
+    Elf_Note *note;
     ELF_Prstatus *prstatus;
     crash_xen_core_t *xencore;
 
+    BUG_ON ( ! crash_notes );
+
     if ( cpumask_test_and_set_cpu(cpu, &crash_saved_cpus) )
         return;
 
+    note = crash_notes[cpu].start;
+
     prstatus = (ELF_Prstatus *)ELFNOTE_DESC(note);
 
     note = ELFNOTE_NEXT(note);

@@ -265,13 +275,6 @@ static struct keyhandler crashdump_trigg
     .desc = "trigger a crashdump"
 };
 
-static __init int register_crashdump_trigger(void)
-{
-    register_keyhandler('C', &crashdump_trigger_keyhandler);
-    return 0;
-}
-__initcall(register_crashdump_trigger);
-
 static void setup_note(Elf_Note *n, const char *name, int type, int descsz)
 {
     int l = strlen(name) + 1;

@@ -281,13 +284,144 @@ static void setup_note(Elf_Note *n, cons
     n->type = type;
 }
 
-static int sizeof_note(const char *name, int descsz)
+static size_t sizeof_note(const char *name, int descsz)
 {
     return (sizeof(Elf_Note) +
             ELFNOTE_ALIGN(strlen(name)+1) +
             ELFNOTE_ALIGN(descsz));
 }
 
+static size_t sizeof_cpu_notes(const unsigned long cpu)
+{
+    /* All CPUs present a PRSTATUS and crash_xen_core note. */
+    size_t bytes =
+        + sizeof_note("CORE", sizeof(ELF_Prstatus)) +
+        + sizeof_note("Xen", sizeof(crash_xen_core_t));
+
+    /* CPU0 also presents the crash_xen_info note. */
+    if ( ! cpu )
+        bytes = bytes +
+            sizeof_note("Xen", sizeof(crash_xen_info_t));
+
+    return bytes;
+}
+
+/* Allocate a crash note buffer for a newly onlined cpu. */
+static int kexec_init_cpu_notes(const unsigned long cpu)
+{
+    Elf_Note * note = NULL;
+    int ret = 0;
+    int nr_bytes = 0;
+
+    BUG_ON( cpu >= nr_cpu_ids || ! crash_notes );
+
+    /* If already allocated, nothing to do. */
+    if ( crash_notes[cpu].start )
+        return ret;
+
+    nr_bytes = sizeof_cpu_notes(cpu);
+    note = xmalloc_bytes(nr_bytes);
+
+    /* Protect the write into crash_notes[] with a spinlock, as this 

function

xen/common/kexec.c +9 -0

+     * is on a hotplug path and a hypercall path. */
+    spin_lock(&crash_notes_lock);
+
+    /* If we are racing with another CPU and it has beaten us, give up
+     * gracefully. */
+    if ( crash_notes[cpu].start )
+    {
+        spin_unlock(&crash_notes_lock);
+        /* Always return ok, because whether we successfully allocated or 

not,

xen/common/kexec.c +44 -0

+         * another CPU has successfully allocated. */
+        if ( note )
+            xfree(note);
+    }
+    else
+    {
+        crash_notes[cpu].start = note;
+        crash_notes[cpu].size = nr_bytes;
+        spin_unlock(&crash_notes_lock);
+
+        /* If the allocation failed, and another CPU did not beat us, give
+         * up with ENOMEM. */
+        if ( ! note )
+            ret = -ENOMEM;
+        /* else all is good so lets set up the notes. */
+        else
+        {
+            /* Set up CORE note. */
+            setup_note(note, "CORE", NT_PRSTATUS, sizeof(ELF_Prstatus));
+            note = ELFNOTE_NEXT(note);
+
+            /* Set up Xen CORE note. */
+            setup_note(note, "Xen", XEN_ELFNOTE_CRASH_REGS,
+                       sizeof(crash_xen_core_t));
+
+            if ( ! cpu )
+            {
+                /* Set up Xen Crash Info note. */
+                xen_crash_note = note = ELFNOTE_NEXT(note);
+                setup_note(note, "Xen", XEN_ELFNOTE_CRASH_INFO,
+                           sizeof(crash_xen_info_t));
+            }
+        }
+    }
+
+    return ret;
+}
+
+static int cpu_callback(
+    struct notifier_block *nfb, unsigned long action, void *hcpu)
+{
+    unsigned long cpu = (unsigned long)hcpu;
+
+    /* Only hook on CPU_UP_PREPARE because once a crash_note has been 

reported
+ * to dom0, it must keep it around in case of a crash, as the crash
kernel

xen/common/kexec.c +6 -0

+     * will be hard coded to the original physical address reported. */
+    switch ( action )
+    {
+    case CPU_UP_PREPARE:
+        /* Ignore return value.  If this boot time, -ENOMEM will cause all
+         * manner of problems elsewhere very soon, and if it is during 

runtime,
+ * then failing to allocate crash notes is not a good enough reason
to

xen/common/kexec.c +35 -0

+         * fail the CPU_UP_PREPARE */
+        kexec_init_cpu_notes(cpu);
+        break;
+    default:
+        break;
+    }
+    return NOTIFY_DONE;
+}
+
+static struct notifier_block cpu_nfb = {
+    .notifier_call = cpu_callback
+};
+
+static int __init kexec_init(void)
+{
+    void *cpu = (void *)(unsigned long)smp_processor_id();
+
+    /* If no crash area, no need to allocate space for notes. */
+    if ( !kexec_crash_area.size )
+        return 0;
+
+    register_keyhandler('C', &crashdump_trigger_keyhandler);
+
+    crash_notes = xmalloc_array(crash_note_range_t, nr_cpu_ids);
+    if ( ! crash_notes )
+        return -ENOMEM;
+
+    memset(crash_notes, 0, sizeof(crash_note_range_t) * nr_cpu_ids);
+
+    cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
+    register_cpu_notifier(&cpu_nfb);
+    return 0;
+}
+/* The reason for this to be a presmp_initcall as opposed to a regular
+ * __initcall is to allow the setup of the cpu hotplug handler before APs 

are
+ * brought up. */

xen/common/kexec.c +7 -4

+presmp_initcall(kexec_init);
+
 static int kexec_get_reserve(xen_kexec64_range_t *range)
 {
     if ( kexec_crash_area.size > 0 && kexec_crash_area.start > 0) {

@@ -302,44 +436,29 @@ static int kexec_get_reserve(xen_kexec64
 static int kexec_get_cpu(xen_kexec64_range_t *range)
 {
     int nr = range->nr;
-    int nr_bytes = 0;
 
-    if ( nr < 0 || nr >= nr_cpu_ids || !cpu_online(nr) )
+    if ( nr < 0 || nr >= nr_cpu_ids )
+        return -ERANGE;
+
+    if ( ! crash_notes )
         return -EINVAL;
 
-    nr_bytes += sizeof_note("CORE", sizeof(ELF_Prstatus));
-    nr_bytes += sizeof_note("Xen", sizeof(crash_xen_core_t));
+    /* Try once again to allocate room for the crash notes.  It is just 

possible
+ * that more space has become available since we last tried. If space
has
+ * already been allocated, kexec_init_cpu_notes() will return early
with 0.

xen/common/kexec.c +3 -3

+     */
+    kexec_init_cpu_notes(nr);
 
-    /* The Xen info note is included in CPU0's range. */
-    if ( nr == 0 )
-        nr_bytes += sizeof_note("Xen", sizeof(crash_xen_info_t));
+    /* In the case of still not having enough memory to allocate buffer 

room,

xen/common/kexec.c +8 -15

+     * returning a range of 0,0 is still valid. */
+    if ( crash_notes[nr].start )
+    {
+        range->start = __pa(crash_notes[nr].start);
+        range->size = crash_notes[nr].size;
+    }
+    else
+        range->start = range->size = 0;
 
-    if ( per_cpu(crash_notes, nr) == NULL )
-    {
-        Elf_Note *note;
-
-        note = per_cpu(crash_notes, nr) = xmalloc_bytes(nr_bytes);
-
-        if ( note == NULL )
-            return -ENOMEM;
-
-        /* Setup CORE note. */
-        setup_note(note, "CORE", NT_PRSTATUS, sizeof(ELF_Prstatus));
-
-        /* Setup Xen CORE note. */
-        note = ELFNOTE_NEXT(note);
-        setup_note(note, "Xen", XEN_ELFNOTE_CRASH_REGS, 

sizeof(crash_xen_core_t));

xen/common/kexec.c +0 -6

-
-        if (nr == 0)
-        {
-            /* Setup system wide Xen info note. */
-            xen_crash_note = note = ELFNOTE_NEXT(note);
-            setup_note(note, "Xen", XEN_ELFNOTE_CRASH_INFO, 

sizeof(crash_xen_info_t));

xen/common/kexec.c +0 -5

-        }
-    }
-
-    range->start = __pa((unsigned long)per_cpu(crash_notes, nr));
-    range->size = nr_bytes;
     return 0;
 }

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

On 64bit Xen with 32bit dom0 and crashkernel, xmalloc'ing items such as the CPU crash notes will go into the xenheap, which tends to be in upper memory. This causes problems on machines with more than 64GB (or 4GB if no PAE support) of ram as the crashkernel physically cant access the crash notes.

xen/common/kexec.c +5 -0

     /* Protect the write into crash_notes[] with a spinlock, as this function
      * is on a hotplug path and a hypercall path. */

@@ -338,6 +426,11 @@ static int kexec_init_cpu_notes(const un
     }
     else
     {
+        /* If we care about memory possition, alloc from the crash heap,
+         * also protected by the crash_notes_lock. */
+        if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+            note = alloc_from_crash_heap(nr_bytes);
+
         crash_notes[cpu].start = note;
         crash_notes[cpu].size = nr_bytes;
         spin_unlock(&crash_notes_lock);

xen/common/kexec.c +10 -0

+
 static int __init kexec_init(void)
 {
     void *cpu = (void *)(unsigned long)smp_processor_id();

@@ -407,6 +518,29 @@ static int __init kexec_init(void)
 
     register_keyhandler('C', &crashdump_trigger_keyhandler);
 
+    if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+    {
+        size_t crash_heap_size;
+
+        /* This calculation is safe even if the machine is booted in 
+         * uniprocessor mode. */
+        crash_heap_size = sizeof_cpu_notes(0) +
+            sizeof_cpu_notes(1) * (nr_cpu_ids - 1);
+        crash_heap_size = PAGE_ALIGN(crash_heap_size);

xen/common/kexec.c +4 -0

+
+    /* crash_notes may be allocated anywhere Xen can reach in memory.
+       Only the individual CPU crash notes themselves must be allocated
+       in lower memory if requested. */
     crash_notes = xmalloc_array(crash_note_range_t, nr_cpu_ids);
     if ( ! crash_notes )
         return -ENOMEM;

diff -r 3da37c68284f -r 23c31d59ffb1 xen/drivers/char/console.c

xen/drivers/char/console.c +1 -0

@@ -584,6 +584,7 @@ void __init console_init_postirq(void)
 {
     char *ring;
     unsigned int i, order;
+    u64 memflags;
 
     serial_init_postirq();
 

Quoted 5 lines · Andrew Cooper

On 64bit Xen with 32bit dom0 and crashkernel, xmalloc'ing items such as the CPU crash notes will go into the xenheap, which tends to be in upper memory. This causes problems on machines with more than 64GB (or 4GB if no PAE support) of ram as the crashkernel physically cant access the crash notes.

I've never been entirely convinced that this was the correct approach. It seems that using a 64-bit crash kernel would be an easier solution.

xen/common/kexec.c static int kexec_init_cpu_notes(const un · line 411

Quoted 9 lines

xen/common/kexec.c +4 -1

@@ -320,7 +405,10 @@ static int kexec_init_cpu_notes(const un
         return ret;
 
     nr_bytes = sizeof_cpu_notes(cpu);
-    note = xmalloc_bytes(nr_bytes);
+
+    /* If we dont care about the position of allocation, malloc. */
+    if ( low_crashinfo_mode == LOW_CRASHINFO_NONE )
+        note = xmalloc_bytes(nr_bytes);

Suggest refactoring this (and the other similar places) so that the check for the regular/crash head occurs in one wrapper alloc function.

Quoted 18 lines

xen/common/kexec.c +5 -0

     /* Protect the write into crash_notes[] with a spinlock, as this function
      * is on a hotplug path and a hypercall path. */

@@ -338,6 +426,11 @@ static int kexec_init_cpu_notes(const un
     }
     else
     {
+        /* If we care about memory possition, alloc from the crash heap,
+         * also protected by the crash_notes_lock. */
+        if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+            note = alloc_from_crash_heap(nr_bytes);
+
         crash_notes[cpu].start = note;
         crash_notes[cpu].size = nr_bytes;
         spin_unlock(&crash_notes_lock);

@@ -397,6 +490,24 @@ static struct notifier_block cpu_nfb = {
     .notifier_call = cpu_callback
 };
 

xen/common/kexec.c static struct notifier_block cpu_nfb = { · line 499

Quoted 7 lines

xen/common/kexec.c +7 -0

+void __init kexec_early_calculations(void)
+{
+    /* If low_crashinfo_mode is still INVALID, neither "low_crashinfo" nor
+     * "crashinfo_maxaddr" have been specified on the command line, so
+     * explicitly set to NONE. */
+    if ( low_crashinfo_mode == LOW_CRASHINFO_INVALID )
+        low_crashinfo_mode = LOW_CRASHINFO_NONE;

Why not initialize low_crash_info_mode to NONE?

xen/common/kexec.c static struct notifier_block cpu_nfb = { · line 509

Quoted 10 lines

xen/common/kexec.c +10 -0

+
+    crashinfo_maxaddr_bits = 0;
+    if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+    {
+        paddr_t tmp = crashinfo_maxaddr;
+
+        while ( tmp >>= 1 )
+            crashinfo_maxaddr_bits++;
+    }
+}

Do these during the parsing of the command line?

These will allow you to remove this unhelpfully named function.

Quoted 20 lines

xen/common/kexec.c +13 -0

+
 static int __init kexec_init(void)
 {
     void *cpu = (void *)(unsigned long)smp_processor_id();

@@ -407,6 +518,29 @@ static int __init kexec_init(void)
 
     register_keyhandler('C', &crashdump_trigger_keyhandler);
 
+    if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+    {
+        size_t crash_heap_size;
+
+        /* This calculation is safe even if the machine is booted in 
+         * uniprocessor mode. */
+        crash_heap_size = sizeof_cpu_notes(0) +
+            sizeof_cpu_notes(1) * (nr_cpu_ids - 1);
+        crash_heap_size = PAGE_ALIGN(crash_heap_size);
+
+        crash_heap_current = alloc_xenheap_pages(
+            get_order_from_bytes(crash_heap_size),

xen/common/kexec.c static int __init kexec_init(void) · line 539

Quoted 7 lines

xen/common/kexec.c +7 -0

+            MEMF_bits(crashinfo_maxaddr_bits) );
+
+        if ( crash_heap_current )
+            crash_heap_end = crash_heap_current + crash_heap_size;
+        else
+            return -ENOMEM;
+    }

Suggest moving this into a crash_heap_setup() function.

Quoted 18 lines

xen/common/kexec.c +4 -0

+
+    /* crash_notes may be allocated anywhere Xen can reach in memory.
+       Only the individual CPU crash notes themselves must be allocated
+       in lower memory if requested. */
     crash_notes = xmalloc_array(crash_note_range_t, nr_cpu_ids);
     if ( ! crash_notes )
         return -ENOMEM;

diff -r 3da37c68284f -r 23c31d59ffb1 xen/drivers/char/console.c

xen/drivers/char/console.c +1 -0

@@ -584,6 +584,7 @@ void __init console_init_postirq(void)
 {
     char *ring;
     unsigned int i, order;
+    u64 memflags;
 
     serial_init_postirq();
 

xen/drivers/char/console.c void __init console_init_postirq(void) · line 595

Quoted 6 lines

xen/drivers/char/console.c +1 -1

@@ -591,7 +592,8 @@ void __init console_init_postirq(void)
         opt_conring_size = num_present_cpus() << (9 + xenlog_lower_thresh);
 
     order = get_order_from_bytes(max(opt_conring_size, conring_size));
-    while ( (ring = alloc_xenheap_pages(order, 0)) == NULL )
+    memflags = low_crashinfo_mode > LOW_CRASHINFO_NONE ? MEMF_bits(crashinfo_maxaddr_bits) : 0;

If you set crashinfo_maxaddr_bits to 64 if low_crashinfo_mode isn't used you wouldn't need this test.

David

Quoted 1 line

Quoted 1 line · Andrew Cooper

Quoted 2 lines · David Vrabel

Quoted 1 line · Andrew Cooper

+ memflags = low_crashinfo_mode > LOW_CRASHINFO_NONE ?

MEMF_bits(crashinfo_maxaddr_bits) : 0;

Quoted 2 lines

If you set crashinfo_maxaddr_bits to 64 if low_crashinfo_mode isn't used you wouldn't need this test.

I am not familiar enough with the intricacies of alloc_xenheap_pages to know whether that is safe. cc'ing Tim for his expertise.

Quoted 1 line

Quoted 1 line · Andrew Cooper

On 64bit Xen with 32bit dom0 and crashkernel, xmalloc'ing items such as the CPU crash notes will go into the xenheap, which tends to be in upper memory. This causes problems on machines with more than 64GB (or 4GB if no PAE support) of ram as the crashkernel physically cant access the crash notes.

The solution is to force Xen to allocate certain structures in lower memory. This is achieved by introducing two new command line parameters; low_crashinfo and crashinfo_maxaddr. Because of the potential impact on 32bit PV guests, and that this problem does not exist for 64bit dom0 on 64bit Xen, this new functionality defaults to the codebase's previous behavior, requiring the user to explicitly add extra command line parameters to change the behavior.

This patch consists of 3 logically distinct but closely related changes.
1) Add the two new command line parameters.

2) Change crash note allocation to use lower memory when instructed.
3) Change the conring buffer to use lower memory when instructed.

There result is that the crash notes and console ring will be placed in lower memory so useful information can be recovered in the case of a crash.

Signed-off-by Andrew Cooper <[email protected]>

diff -r 3da37c68284f -r 23c31d59ffb1 xen/arch/x86/setup.c

xen/arch/x86/setup.c +0 -0

     parse_video_info();
 
     set_current((struct vcpu *)0xfffff000); /* debug sanity */

diff -r 3da37c68284f -r 23c31d59ffb1 xen/common/kexec.c

xen/common/kexec.c +3 -1

@@ -36,7 +36,9 @@ bool_t kexecing = FALSE;
 typedef struct { Elf_Note * start; size_t size; } crash_note_range_t;
 static crash_note_range_t * crash_notes;
 
-/* Lock to prevent race conditions when allocating the crash note buffers. 
 */
+/* Lock to prevent race conditions when allocating the crash note buffers.
+ * It also serves to protect calls to alloc_from_crash_heap when allocating
+ * crash note buffers in lower memory. */
 static DEFINE_SPINLOCK(crash_notes_lock);
 

static Elf_Note *xen_crash_note;

xen/common/kexec.c +6 -0

@@ -70,6 +72,23 @@ static struct {
     (_d_)->start = (_s_)->start; \
 } while (0)
 
+/* Low crashinfo mode.  Start as INVALID so serveral codepaths can set up
+ * defaults without needing to know the state of the others. */
+enum low_crashinfo low_crashinfo_mode = LOW_CRASHINFO_INVALID;
+
+/* This value is only considered if low_crash_mode is set to MIN or ALL, so
+ * setting a default here is safe. Default to 4GB.  This is because the 

current
+ * KEXEC_CMD_get_range compat hypercall trucates 64bit pointers to 32 bits. The
+ * typical usecase for crashinfo_maxaddr will be for 64bit Xen with 32bit dom0
+ * and 32bit crash kernel. */
+static paddr_t crashinfo_maxaddr = 4ULL << 30;

xen/common/kexec.c +13 -0

+
+/* = log base 2 of crashinfo_maxaddr after checking for sanity. */
+paddr_t crashinfo_maxaddr_bits = 0;
+
+/* Pointers to keep track of the crash heap region. */
+static void *crash_heap_current = NULL, *crash_heap_end = NULL;
+
 /*
  * Parse command lines in the format
  *

@@ -148,6 +167,58 @@ static void __init parse_crashkernel(con
 }
 custom_param("crashkernel", parse_crashkernel);
 
+/* Parse command lines in the format:
+ *
+ *   low_crashinfo=[none,min,all]
+ * 
+ * - none disables the low allocation of crash info.
+ * - min will allocate enough low information for the crash kernel to be 

able
+ * to extract the hypervisor and dom0 message ring buffers.
+ * - all will allocate additional structures such as domain and vcpu structs
+ * low so the crash kernel can perform an extended analysis of state.
+ */

xen/common/kexec.c +15 -0

+static void __init parse_low_crashinfo(const char * str)
+{
+
+    if ( !strlen(str) )
+        /* default to min if user just specifies "low_crashinfo" */
+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+    else if ( !strcmp(str, "none" ) )
+        low_crashinfo_mode = LOW_CRASHINFO_NONE;
+    else if ( !strcmp(str, "min" ) )
+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+    else if ( !strcmp(str, "all" ) )
+        low_crashinfo_mode = LOW_CRASHINFO_ALL;
+    else
+    {
+        printk("Unknown low_crashinfo parameter '%s'.  Defaulting to 

min.\n", str);

xen/common/kexec.c +45 -1

+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+    }
+}
+custom_param("low_crashinfo", parse_low_crashinfo);
+
+/* Parse command lines in the format:
+ * 
+ *   crashinfo_maxaddr=<addr>
+ *
+ * <addr> will be rounded down to the nearest power of two.  Defaults to 64G
+ */
+static void __init parse_crashinfo_maxaddr(const char * str)
+{
+    u64 addr;
+
+    /* if low_crashinfo_mode is unset, default to min. */
+    if ( low_crashinfo_mode == LOW_CRASHINFO_INVALID )
+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+
+    if ( (addr = parse_size_and_unit(str, NULL)) )
+        crashinfo_maxaddr = addr;
+    else
+        printk("Unable to parse crashinfo_maxaddr. Defaulting to %p\n",
+               (void*)crashinfo_maxaddr);
+}
+custom_param("crashinfo_maxaddr", parse_crashinfo_maxaddr);
+
 void __init set_kexec_crash_area_size(u64 system_ram)
 {
     unsigned int idx;

@@ -306,6 +377,20 @@ static size_t sizeof_cpu_notes(const uns
     return bytes;
 }
 
+/* Allocate size_t bytes of space from the previously allocated
+ * crash heap if the user has requested that crash notes be allocated
+ * in lower memory.  There is currently no case where the crash notes
+ * should be free()'d. */
+static void * alloc_from_crash_heap(const size_t bytes)
+{
+    void * ret;
+    if ( crash_heap_current + bytes > crash_heap_end )
+        return NULL;
+    ret = (void*)crash_heap_current;
+    crash_heap_current += bytes;
+    return ret;
+}
+
 /* Allocate a crash note buffer for a newly onlined cpu. */
 static int kexec_init_cpu_notes(const unsigned long cpu)
 {

@@ -320,7 +405,10 @@ static int kexec_init_cpu_notes(const un
         return ret;
 
     nr_bytes = sizeof_cpu_notes(cpu);
-    note = xmalloc_bytes(nr_bytes);
+
+    /* If we dont care about the position of allocation, malloc. */
+    if ( low_crashinfo_mode == LOW_CRASHINFO_NONE )
+        note = xmalloc_bytes(nr_bytes);
 
     /* Protect the write into crash_notes[] with a spinlock, as this 

function
* is on a hotplug path and a hypercall path. */

xen/common/kexec.c +10 -0

@@ -338,6 +426,11 @@ static int kexec_init_cpu_notes(const un
     }
     else
     {
+        /* If we care about memory possition, alloc from the crash heap,
+         * also protected by the crash_notes_lock. */
+        if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+            note = alloc_from_crash_heap(nr_bytes);
+
         crash_notes[cpu].start = note;
         crash_notes[cpu].size = nr_bytes;
         spin_unlock(&crash_notes_lock);

@@ -397,6 +490,24 @@ static struct notifier_block cpu_nfb = {
     .notifier_call = cpu_callback
 };
 
+void __init kexec_early_calculations(void)
+{
+    /* If low_crashinfo_mode is still INVALID, neither "low_crashinfo" nor
+     * "crashinfo_maxaddr" have been specified on the command line, so
+     * explicitly set to NONE. */

xen/common/kexec.c +11 -0

+    }
+}
+
 static int __init kexec_init(void)
 {
     void *cpu = (void *)(unsigned long)smp_processor_id();

@@ -407,6 +518,29 @@ static int __init kexec_init(void)
 
     register_keyhandler('C', &crashdump_trigger_keyhandler);
 
+    if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+    {
+        size_t crash_heap_size;
+
+        /* This calculation is safe even if the machine is booted in 
+         * uniprocessor mode. */
+        crash_heap_size = sizeof_cpu_notes(0) +
+            sizeof_cpu_notes(1) * (nr_cpu_ids - 1);

xen/common/kexec.c +5 -0

+    }
+
+    /* crash_notes may be allocated anywhere Xen can reach in memory.
+       Only the individual CPU crash notes themselves must be allocated
+       in lower memory if requested. */
     crash_notes = xmalloc_array(crash_note_range_t, nr_cpu_ids);
     if ( ! crash_notes )
         return -ENOMEM;

diff -r 3da37c68284f -r 23c31d59ffb1 xen/drivers/char/console.c

xen/drivers/char/console.c +3 -1

@@ -584,6 +584,7 @@ void __init console_init_postirq(void)
 {
     char *ring;
     unsigned int i, order;
+    u64 memflags;
 
     serial_init_postirq();
 

@@ -591,7 +592,8 @@ void __init console_init_postirq(void)
         opt_conring_size = num_present_cpus() << (9 + xenlog_lower_thresh);
 
     order = get_order_from_bytes(max(opt_conring_size, conring_size));
-    while ( (ring = alloc_xenheap_pages(order, 0)) == NULL )
+    memflags = low_crashinfo_mode > LOW_CRASHINFO_NONE ? 
 MEMF_bits(crashinfo_maxaddr_bits) : 0;
+    while ( (ring = alloc_xenheap_pages(order, memflags)) == NULL )
     {
         BUG_ON(order == 0);

         order--;

diff -r 3da37c68284f -r 23c31d59ffb1 xen/include/xen/kexec.h

xen/include/xen/kexec.h +13 -0

@@ -25,6 +25,19 @@ void set_kexec_crash_area_size(u64 syste
 #define KEXEC_IMAGE_CRASH_BASE   2
 #define KEXEC_IMAGE_NR           4
 
+enum low_crashinfo {
+    LOW_CRASHINFO_INVALID = 0,
+    LOW_CRASHINFO_NONE = 1,
+    LOW_CRASHINFO_MIN = 2,
+    LOW_CRASHINFO_ALL = 3
+};
+
+/* Low crashinfo mode.  Start as INVALID so serveral codepaths can set up
+ * defaults without needing to know the state of the others. */
+extern enum low_crashinfo low_crashinfo_mode;
+extern paddr_t crashinfo_maxaddr_bits;
+void kexec_early_calculations(void);
+
 int machine_kexec_load(int type, int slot, xen_kexec_image_t *image);
 void machine_kexec_unload(int type, int slot, xen_kexec_image_t *image);
 void machine_kexec_reserved(xen_kexec_reserve_t *reservation);

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Quoted 6 lines

Quoted 1 line

Quoted 1 line · Andrew Cooper

On 64bit Xen with 32bit dom0 and crashkernel, xmalloc'ing items such as the CPU crash notes will go into the xenheap, which tends to be in upper memory. This causes problems on machines with more than 64GB (or 4GB if no PAE support) of ram as the crashkernel physically cant access the crash notes.

What use is a crash dump taken with a 32-bit kernel on a machine with more than 64G (or more than 4G is the crash kernel doesn't support PAE)?

Quoted 8 lines

The solution is to force Xen to allocate certain structures in lower memory. This is achieved by introducing two new command line parameters; low_crashinfo and crashinfo_maxaddr. Because of the potential impact on 32bit PV guests, and that this problem does not exist for 64bit dom0 on 64bit Xen, this new functionality defaults to the codebase's previous behavior, requiring the user to explicitly add extra command line parameters to change the behavior.

Quoted 3 lines

This patch consists of 3 logically distinct but closely related changes.
1) Add the two new command line parameters.

Do you really need both? Just being able to set the max address would seem sufficient...

Quoted 2 lines

2) Change crash note allocation to use lower memory when instructed.
3) Change the conring buffer to use lower memory when instructed.

Why does this one need moving, but none of the other Xen data structures? If anything, I would have expected you to limit the Xen heap altogether, so that automatically all allocations get done in a way so that at least Xen's data structures are available in the (truncated) dump.

Quoted 13 lines

There result is that the crash notes and console ring will be placed in lower memory so useful information can be recovered in the case of a crash.

Signed-off-by Andrew Cooper <[email protected]>

diff -r 3da37c68284f -r 23c31d59ffb1 xen/arch/x86/setup.c

xen/arch/x86/setup.c +0 -0

@@ -586,6 +586,10 @@ void __init __start_xen(unsigned long mb
     }
     cmdline_parse(cmdline);
 

xen/arch/x86/setup.c void __init __start_xen(unsigned long mb · line 592

Quoted 4 lines

xen/arch/x86/setup.c +4 -0

+    /* Must be after command line argument parsing and before 
+     * allocing any xenheap structures wanted in lower memory. */
+    kexec_early_calculations();
+

But does it really need to be *this* early?

Quoted 22 lines

xen/arch/x86/setup.c +0 -0

     parse_video_info();
 
     set_current((struct vcpu *)0xfffff000); /* debug sanity */

diff -r 3da37c68284f -r 23c31d59ffb1 xen/common/kexec.c

xen/common/kexec.c +3 -1

@@ -36,7 +36,9 @@ bool_t kexecing = FALSE;
 typedef struct { Elf_Note * start; size_t size; } crash_note_range_t;
 static crash_note_range_t * crash_notes;
 
-/* Lock to prevent race conditions when allocating the crash note buffers. 
 */
+/* Lock to prevent race conditions when allocating the crash note buffers.
+ * It also serves to protect calls to alloc_from_crash_heap when allocating
+ * crash note buffers in lower memory. */
 static DEFINE_SPINLOCK(crash_notes_lock);
 

static Elf_Note *xen_crash_note;

xen/common/kexec.c +0 -0

@@ -70,6 +72,23 @@ static struct {
     (_d_)->start = (_s_)->start; \
 } while (0)
 

Quoted 6 lines

xen/common/kexec.c +6 -0

+/* Low crashinfo mode.  Start as INVALID so serveral codepaths can set up
+ * defaults without needing to know the state of the others. */
+enum low_crashinfo low_crashinfo_mode = LOW_CRASHINFO_INVALID;
+
+/* This value is only considered if low_crash_mode is set to MIN or ALL, so
+ * setting a default here is safe. Default to 4GB.  This is because the 

Quoted 7 lines

current
+ * KEXEC_CMD_get_range compat hypercall trucates 64bit pointers to 32 bits. The
+ * typical usecase for crashinfo_maxaddr will be for 64bit Xen with 32bit dom0
+ * and 32bit crash kernel. */
+static paddr_t crashinfo_maxaddr = 4ULL << 30;

__initdata

Quoted 121 lines

xen/common/kexec.c +13 -0

+
+/* = log base 2 of crashinfo_maxaddr after checking for sanity. */
+paddr_t crashinfo_maxaddr_bits = 0;
+
+/* Pointers to keep track of the crash heap region. */
+static void *crash_heap_current = NULL, *crash_heap_end = NULL;
+
 /*
  * Parse command lines in the format
  *

@@ -148,6 +167,58 @@ static void __init parse_crashkernel(con
 }
 custom_param("crashkernel", parse_crashkernel);
 
+/* Parse command lines in the format:
+ *
+ *   low_crashinfo=[none,min,all]
+ * 
+ * - none disables the low allocation of crash info.
+ * - min will allocate enough low information for the crash kernel to be 

able
+ * to extract the hypervisor and dom0 message ring buffers.
+ * - all will allocate additional structures such as domain and vcpu structs
+ * low so the crash kernel can perform an extended analysis of state.
+ */

xen/common/kexec.c +15 -0

+static void __init parse_low_crashinfo(const char * str)
+{
+
+    if ( !strlen(str) )
+        /* default to min if user just specifies "low_crashinfo" */
+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+    else if ( !strcmp(str, "none" ) )
+        low_crashinfo_mode = LOW_CRASHINFO_NONE;
+    else if ( !strcmp(str, "min" ) )
+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+    else if ( !strcmp(str, "all" ) )
+        low_crashinfo_mode = LOW_CRASHINFO_ALL;
+    else
+    {
+        printk("Unknown low_crashinfo parameter '%s'.  Defaulting to 

min.\n", str);

xen/common/kexec.c +45 -1

+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+    }
+}
+custom_param("low_crashinfo", parse_low_crashinfo);
+
+/* Parse command lines in the format:
+ * 
+ *   crashinfo_maxaddr=<addr>
+ *
+ * <addr> will be rounded down to the nearest power of two.  Defaults to 64G
+ */
+static void __init parse_crashinfo_maxaddr(const char * str)
+{
+    u64 addr;
+
+    /* if low_crashinfo_mode is unset, default to min. */
+    if ( low_crashinfo_mode == LOW_CRASHINFO_INVALID )
+        low_crashinfo_mode = LOW_CRASHINFO_MIN;
+
+    if ( (addr = parse_size_and_unit(str, NULL)) )
+        crashinfo_maxaddr = addr;
+    else
+        printk("Unable to parse crashinfo_maxaddr. Defaulting to %p\n",
+               (void*)crashinfo_maxaddr);
+}
+custom_param("crashinfo_maxaddr", parse_crashinfo_maxaddr);
+
 void __init set_kexec_crash_area_size(u64 system_ram)
 {
     unsigned int idx;

@@ -306,6 +377,20 @@ static size_t sizeof_cpu_notes(const uns
     return bytes;
 }
 
+/* Allocate size_t bytes of space from the previously allocated
+ * crash heap if the user has requested that crash notes be allocated
+ * in lower memory.  There is currently no case where the crash notes
+ * should be free()'d. */
+static void * alloc_from_crash_heap(const size_t bytes)
+{
+    void * ret;
+    if ( crash_heap_current + bytes > crash_heap_end )
+        return NULL;
+    ret = (void*)crash_heap_current;
+    crash_heap_current += bytes;
+    return ret;
+}
+
 /* Allocate a crash note buffer for a newly onlined cpu. */
 static int kexec_init_cpu_notes(const unsigned long cpu)
 {

@@ -320,7 +405,10 @@ static int kexec_init_cpu_notes(const un
         return ret;
 
     nr_bytes = sizeof_cpu_notes(cpu);
-    note = xmalloc_bytes(nr_bytes);
+
+    /* If we dont care about the position of allocation, malloc. */
+    if ( low_crashinfo_mode == LOW_CRASHINFO_NONE )
+        note = xmalloc_bytes(nr_bytes);
 
     /* Protect the write into crash_notes[] with a spinlock, as this 

function
* is on a hotplug path and a hypercall path. */

xen/common/kexec.c +5 -0

@@ -338,6 +426,11 @@ static int kexec_init_cpu_notes(const un
     }
     else
     {
+        /* If we care about memory possition, alloc from the crash heap,
+         * also protected by the crash_notes_lock. */
+        if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+            note = alloc_from_crash_heap(nr_bytes);
+
         crash_notes[cpu].start = note;
         crash_notes[cpu].size = nr_bytes;
         spin_unlock(&crash_notes_lock);

@@ -397,6 +490,24 @@ static struct notifier_block cpu_nfb = {
     .notifier_call = cpu_callback
 };
 

Quoted 5 lines

xen/common/kexec.c +5 -0

+void __init kexec_early_calculations(void)
+{
+    /* If low_crashinfo_mode is still INVALID, neither "low_crashinfo" nor
+     * "crashinfo_maxaddr" have been specified on the command line, so
+     * explicitly set to NONE. */

xen/common/kexec.c static struct notifier_block cpu_nfb = { · line 507

Quoted 10 lines

xen/common/kexec.c +10 -0

+    if ( low_crashinfo_mode == LOW_CRASHINFO_INVALID )
+        low_crashinfo_mode = LOW_CRASHINFO_NONE;
+
+    crashinfo_maxaddr_bits = 0;
+    if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+    {
+        paddr_t tmp = crashinfo_maxaddr;
+
+        while ( tmp >>= 1 )
+            crashinfo_maxaddr_bits++;

fls() or some of its cousins

Quoted 18 lines

xen/common/kexec.c +11 -0

+    }
+}
+
 static int __init kexec_init(void)
 {
     void *cpu = (void *)(unsigned long)smp_processor_id();

@@ -407,6 +518,29 @@ static int __init kexec_init(void)
 
     register_keyhandler('C', &crashdump_trigger_keyhandler);
 
+    if ( low_crashinfo_mode > LOW_CRASHINFO_NONE )
+    {
+        size_t crash_heap_size;
+
+        /* This calculation is safe even if the machine is booted in 
+         * uniprocessor mode. */
+        crash_heap_size = sizeof_cpu_notes(0) +
+            sizeof_cpu_notes(1) * (nr_cpu_ids - 1);

xen/common/kexec.c static int __init kexec_init(void) · line 538

Quoted 10 lines

xen/common/kexec.c +10 -0

+        crash_heap_size = PAGE_ALIGN(crash_heap_size);
+
+        crash_heap_current = alloc_xenheap_pages(
+            get_order_from_bytes(crash_heap_size),
+            MEMF_bits(crashinfo_maxaddr_bits) );
+
+        if ( crash_heap_current )
+            crash_heap_end = crash_heap_current + crash_heap_size;
+        else
+            return -ENOMEM;

Constructs like this generally look bogus to me, as

        if ( !crash_heap_current )
            return -ENOMEM;
        ...

is shorter and (to me at least) more clear.

Quoted 58 lines

xen/common/kexec.c +5 -0

+    }
+
+    /* crash_notes may be allocated anywhere Xen can reach in memory.
+       Only the individual CPU crash notes themselves must be allocated
+       in lower memory if requested. */
     crash_notes = xmalloc_array(crash_note_range_t, nr_cpu_ids);
     if ( ! crash_notes )
         return -ENOMEM;

diff -r 3da37c68284f -r 23c31d59ffb1 xen/drivers/char/console.c

xen/drivers/char/console.c +3 -1

@@ -584,6 +584,7 @@ void __init console_init_postirq(void)
 {
     char *ring;
     unsigned int i, order;
+    u64 memflags;
 
     serial_init_postirq();
 

@@ -591,7 +592,8 @@ void __init console_init_postirq(void)
         opt_conring_size = num_present_cpus() << (9 + xenlog_lower_thresh);
 
     order = get_order_from_bytes(max(opt_conring_size, conring_size));
-    while ( (ring = alloc_xenheap_pages(order, 0)) == NULL )
+    memflags = low_crashinfo_mode > LOW_CRASHINFO_NONE ? 
 MEMF_bits(crashinfo_maxaddr_bits) : 0;
+    while ( (ring = alloc_xenheap_pages(order, memflags)) == NULL )
     {
         BUG_ON(order == 0);

         order--;

diff -r 3da37c68284f -r 23c31d59ffb1 xen/include/xen/kexec.h

xen/include/xen/kexec.h +13 -0

@@ -25,6 +25,19 @@ void set_kexec_crash_area_size(u64 syste
 #define KEXEC_IMAGE_CRASH_BASE   2
 #define KEXEC_IMAGE_NR           4
 
+enum low_crashinfo {
+    LOW_CRASHINFO_INVALID = 0,
+    LOW_CRASHINFO_NONE = 1,
+    LOW_CRASHINFO_MIN = 2,
+    LOW_CRASHINFO_ALL = 3
+};
+
+/* Low crashinfo mode.  Start as INVALID so serveral codepaths can set up
+ * defaults without needing to know the state of the others. */
+extern enum low_crashinfo low_crashinfo_mode;
+extern paddr_t crashinfo_maxaddr_bits;
+void kexec_early_calculations(void);
+
 int machine_kexec_load(int type, int slot, xen_kexec_image_t *image);
 void machine_kexec_unload(int type, int slot, xen_kexec_image_t *image);
 void machine_kexec_reserved(xen_kexec_reserve_t *reservation);

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel