On 5/12/23 11:04 AM, Thomas Gleixner wrote:
On Thu, May 04 2023 at 14:50, Ross Philipson wrote:
+/* CPUID: leaf 1, ECX, SMX feature bit */
+#define X86_FEATURE_BIT_SMX (1 << 6)
+/* Can't include apiddef.h in asm */
Why not? All it needs is a #ifndef __ASSEMBLY__ guard around the C parts.
+#define XAPIC_ENABLE (1 << 11)
+#define X2APIC_ENABLE (1 << 10)
+/* Can't include traps.h in asm */
NMI_VECTOR is defined in irq_vectors.h which just has a include
<linux/threads.h> for no real good reason.
+#define X86_TRAP_NMI 2
+ * See the comment in head_64.S for detailed informatoin on what this macro
+ * is used for.
+#define rva(X) ((X) - sl_stub_entry)
I'm having a hard time to find that comment in head_64.S. At least it's
not in this patch.
+ /* Load the relocated AP IDT */
[ 11 more citation lines. Click/Enter to show. ]
+ lidt (sl_ap_idt_desc - sl_txt_ap_wake_begin)(%ecx)
+ /* Fixup MTRRs and misc enable MSR on APs too */
+ call sl_txt_load_regs
+ /* Enable SMI with GETSEC[SMCTRL] */
+ GETSEC $(SMX_X86_GETSEC_SMCTRL)
+ /* IRET-to-self can be used to enable NMIs which SENTER disabled */
+ leal rva(.Lnmi_enabled_ap)(%ebx), %eax
+ pushl $(__SL32_CS)
+ pushl %eax
So from here on any NMI which hits the AP before it can reach the wait
loop will corrupt EDX...
+/* This is the beginning of the relocated AP wake code block */
+ .global sl_txt_ap_wake_begin
[ 10 more citation lines. Click/Enter to show. ]
+ * Wait for NMI IPI in the relocated AP wake block which was provided
+ * and protected in the memory map by the prelaunch code. Leave all
+ * other interrupts masked since we do not expect anything but an NMI.
+ xorl %edx, %edx
+ testl %edx, %edx
+ jz 1b
This really makes me nervous. A stray NMI and the AP starts going.
Can't this NMI just bring the AP out of HLT w/o changing any state and
the AP evaluates a memory location which indicates whether it should
start up or not.
I have switched the existing code to use MONITOR/MWAIT and got rid of
the use of the NMIs here. I am currently using a monitor variable on the
stack of each AP but I think I may refactor that. The next step is to
rebase this work on top of your hotplug patchset. Is the code in your
devel repo on the hotplug branch still the latest bits?
I had a question - more a request for your thoughts on this. I am
currently assuming a cache line size and alignment of 64b for the
monitor variable location. Do you think this is sufficient for x86
platforms or do I need to dynamically find a way to read the CPUID
information for MONITOR and get my size/alignment values from there?
+ * This is the long absolute jump to the 32b Secure Launch protected
+ * mode stub code in the rmpiggy. The jump address will be fixed in
Providing an actual name for the stub might spare to rummage through
code to figure out where this is supposed to jump to.
+ * the SMP boot code when the first AP is brought up. This whole area
+ * is provided and protected in the memory map by the prelaunch code.
[ 2 more citation lines. Click/Enter to show. ]
+ .byte 0xea
+ .long 0x00000000
+ .word __SL32_CS