linux-libre/Documentation/x86/entry_64.txt

This file documents some of the kernel entries in
arch/x86/entry/entry_64.S.  A lot of this explanation is adapted from
an email from Ingo Molnar:

http://lkml.kernel.org/r/<20110529191055.GC9835%40elte.hu>

The x86 architecture has quite a few different ways to jump into
kernel code.  Most of these entry points are registered in
arch/x86/kernel/traps.c and implemented in arch/x86/entry/entry_64.S
for 64-bit, arch/x86/entry/entry_32.S for 32-bit and finally
arch/x86/entry/entry_64_compat.S which implements the 32-bit compatibility
syscall entry points and thus provides for 32-bit processes the
ability to execute syscalls when running on 64-bit kernels.

The IDT vector assignments are listed in arch/x86/include/asm/irq_vectors.h.

Some of these entries are:

 - system_call: syscall instruction from 64-bit code.

 - entry_INT80_compat: int 0x80 from 32-bit or 64-bit code; compat syscall
   either way.

 - entry_INT80_compat, ia32_sysenter: syscall and sysenter from 32-bit
   code

 - interrupt: An array of entries.  Every IDT vector that doesn't
   explicitly point somewhere else gets set to the corresponding
   value in interrupts.  These point to a whole array of
   magically-generated functions that make their way to do_IRQ with
   the interrupt number as a parameter.

 - APIC interrupts: Various special-purpose interrupts for things
   like TLB shootdown.

 - Architecturally-defined exceptions like divide_error.

There are a few complexities here.  The different x86-64 entries
have different calling conventions.  The syscall and sysenter
instructions have their own peculiar calling conventions.  Some of
the IDT entries push an error code onto the stack; others don't.
IDT entries using the IST alternative stack mechanism need their own
magic to get the stack frames right.  (You can find some
documentation in the AMD APM, Volume 2, Chapter 8 and the Intel SDM,
Volume 3, Chapter 6.)

Dealing with the swapgs instruction is especially tricky.  Swapgs
toggles whether gs is the kernel gs or the user gs.  The swapgs
instruction is rather fragile: it must nest perfectly and only in
single depth, it should only be used if entering from user mode to
kernel mode and then when returning to user-space, and precisely
so. If we mess that up even slightly, we crash.

So when we have a secondary entry, already in kernel mode, we *must
not* use SWAPGS blindly - nor must we forget doing a SWAPGS when it's
not switched/swapped yet.

Now, there's a secondary complication: there's a cheap way to test
which mode the CPU is in and an expensive way.

The cheap way is to pick this info off the entry frame on the kernel
stack, from the CS of the ptregs area of the kernel stack:

	xorl %ebx,%ebx
	testl $3,CS+8(%rsp)
	je error_kernelspace
	SWAPGS

The expensive (paranoid) way is to read back the MSR_GS_BASE value
(which is what SWAPGS modifies):

	movl $1,%ebx
	movl $MSR_GS_BASE,%ecx
	rdmsr
	testl %edx,%edx
	js 1f   /* negative -> in kernel */
	SWAPGS
	xorl %ebx,%ebx
1:	ret

If we are at an interrupt or user-trap/gate-alike boundary then we can
use the faster check: the stack will be a reliable indicator of
whether SWAPGS was already done: if we see that we are a secondary
entry interrupting kernel mode execution, then we know that the GS
base has already been switched. If it says that we interrupted
user-space execution then we must do the SWAPGS.

But if we are in an NMI/MCE/DEBUG/whatever super-atomic entry context,
which might have triggered right after a normal entry wrote CS to the
stack but before we executed SWAPGS, then the only safe way to check
for GS is the slower method: the RDMSR.

Therefore, super-atomic entries (except NMI, which is handled separately)
must use idtentry with paranoid=1 to handle gsbase correctly.  This
triggers three main behavior changes:

 - Interrupt entry will use the slower gsbase check.
 - Interrupt entry from user mode will switch off the IST stack.
 - Interrupt exit to kernel mode will not attempt to reschedule.

We try to only use IST entries and the paranoid entry code for vectors
that absolutely need the more expensive check for the GS base - and we
generate all 'normal' entry points with the regular (faster) paranoid=0
variant.