Re: [RFC] KEXEC: allocate crash note buffers at boot time v4

[Andrew Cooper <andrew.cooper3@citrix.com>]

[-- Attachment #1: Type: text/plain, Size: 286 bytes --]

Version 4 attached.

Fixed the logic regarding locking and x{malloc,free}'ing, added a few
more comments in places, and changed the coding style to avoid
constant-first compares.

-- 
Andrew Cooper - Dom0 Kernel Engineer, Citrix XenServer
T: +44 (0)1223 225 900, http://www.citrix.com


[-- Attachment #2: KEXEC-setup-crash-notes-during-boot.patch --]
[-- Type: text/x-patch, Size: 10666 bytes --]

# HG changeset patch
# Parent b5ceec1ccccad6e79053a80820132303ff1e136f
KEXEC: Allocate crash notes on boot v4

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 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 <andrew.cooper3@citrix.com>

diff -r df7cec2c6c03 xen/common/kexec.c
--- a/xen/common/kexec.c
+++ b/xen/common/kexec.c
@@ -25,13 +25,18 @@
 #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. */
+static void ** 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;
 
@@ -165,13 +170,17 @@ 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];
+
     prstatus = (ELF_Prstatus *)ELFNOTE_DESC(note);
 
     note = ELFNOTE_NEXT(note);
@@ -257,13 +266,6 @@ 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;
@@ -280,6 +282,128 @@ static int sizeof_note(const char *name,
             ELFNOTE_ALIGN(descsz));
 }
 
+/* Allocate a crash note buffer for a newly onlined cpu. */
+static int kexec_init_cpu_notes(const unsigned long cpu)
+{
+    Elf_Note * note;
+    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] )
+        return ret;
+
+    /* All CPUs present a PRSTATUS and crash_xen_core note. */
+    nr_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 )
+        nr_bytes = nr_bytes +
+            sizeof_note("Xen", sizeof(crash_xen_info_t));
+
+    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. */
+    spin_lock(&crash_notes_lock);
+
+    /* If we are racing with another CPU and it has beaten us, give up
+     * gracefully. */
+    if ( crash_notes[cpu] )
+    {
+        spin_unlock(&crash_notes_lock);
+        /* Always return ok, because whether we successfully allocated or not,
+         * another CPU has successfully allocated. */
+        if ( note )
+            xfree(note);
+    }
+    else
+    {
+        crash_notes[cpu] = note;
+        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
+     * 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
+         * 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(void *, nr_cpu_ids);
+    if ( ! crash_notes )
+        return -ENOMEM;
+
+    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. */
+presmp_initcall(kexec_init);
+
 static int kexec_get_reserve(xen_kexec_range_t *range)
 {
     if ( kexec_crash_area.size > 0 && kexec_crash_area.start > 0) {
@@ -296,7 +420,14 @@ static int kexec_get_cpu(xen_kexec_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 || !cpu_online(nr) || !crash_notes )
+        return -EINVAL;
+
+    /* 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.
+     */
+    if ( kexec_init_cpu_notes(nr) )
         return -EINVAL;
 
     nr_bytes += sizeof_note("CORE", sizeof(ELF_Prstatus));
@@ -306,31 +437,7 @@ static int kexec_get_cpu(xen_kexec_range
     if ( nr == 0 )
         nr_bytes += sizeof_note("Xen", sizeof(crash_xen_info_t));
 
-    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));
-
-        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));
-        }
-    }
-
-    range->start = __pa((unsigned long)per_cpu(crash_notes, nr));
+    range->start = __pa((unsigned long)crash_notes[nr]);
     range->size = nr_bytes;
     return 0;
 }

[-- Attachment #3: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel