KEXEC: Allocate crash notes on boot

KEXEC: Allocate crash notes on boot

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.  This has certain problems including not being able to
allocate the crash buffers if the host is out of memory when crashing.

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 v6:
 *  Tweak kexec_init() to use xzmalloc_array(), and to defer
 registering the
    crashdump keyhandler until the crash notes have been successfully
    allocated.

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 <andrew.cooper3@citrix.com>
Committed-by: Keir Fraser <keir@xen.org>