]> www.infradead.org Git - users/dwmw2/linux.git/commitdiff
x86/smpboot: Support parallel startup of secondary CPUs
authorDavid Woodhouse <dwmw@amazon.co.uk>
Thu, 23 Feb 2023 19:11:34 +0000 (19:11 +0000)
committerDavid Woodhouse <dwmw@amazon.co.uk>
Sat, 25 Feb 2023 10:09:18 +0000 (10:09 +0000)
Rework the real-mode startup code to allow for APs to be brought up in
parallel. This is in two parts:

1. Introduce a bit-spinlock to prevent them from all using the real
   mode stack at the same time.

2. Avoid needing to use the global smpboot_control variable to pass
   each AP its CPU#.

To achieve the latter, export the cpuid_to_apicid[] array so that each
AP can find its own CPU# by searching therein based on its APIC ID.

Introduce flags in the top bits of smpboot_control which indicate methods
by which an AP should find its CPU#. For a serialized bringup, the CPU#
is explicitly passed in the low bits of smpboot_control as before. For
parallel mode there are flags directing the AP to find its APIC ID in
CPUID leaf 0x0b (for X2APIC mode) or CPUID leaf 0x01 where 8 bits are
sufficient, then perform the cpuid_to_apicid[] lookup with that.

Parallel startup may be disabled by a command line option, and also if:
 • AMD SEV-ES is in use, since the AP may not use CPUID that early.
 • X2APIC is enabled, but CPUID leaf 0xb is not present and correect.
 • X2APIC is not enabled but not even CPUID leaf 0x01 exists.

Aside from the fact that APs will now look up their CPU# via the
newly-exported cpuid_to_apicid[] table, there is no behavioural change
intended yet, since new parallel CPUHP states have not — yet — been
added.

[ tglx: Initial proof of concept patch with bitlock and APIC ID lookup ]
[ dwmw2: Rework and testing, commit message, CPUID 0x1 and CPU0 support ]
[ seanc: Fix stray override of initial_gs in common_cpu_up() ]
[ Oleksandr Natalenko: reported suspend/resume issue fixed in
  x86_acpi_suspend_lowlevel ]
Co-developed-by: Thomas Gleixner <tglx@linutronix.de>
Co-developed-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Signed-off-by: Usama Arif <usama.arif@bytedance.com>
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Tested-by: Kim Phillips <kim.phillips@amd.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Documentation/admin-guide/kernel-parameters.txt
arch/x86/include/asm/realmode.h
arch/x86/include/asm/smp.h
arch/x86/kernel/acpi/sleep.c
arch/x86/kernel/apic/apic.c
arch/x86/kernel/head_64.S
arch/x86/kernel/smpboot.c
arch/x86/realmode/init.c
arch/x86/realmode/rm/trampoline_64.S

index 6cfa6e3996cf75ee6bb1e8b399daa4d5701ab92b..ee099b8aac6d419f341a011de3359b3441d43dcf 100644 (file)
 
        nomodule        Disable module load
 
+       no_parallel_bringup
+                       [X86,SMP] Disable parallel brinugp of secondary cores.
+
        nopat           [X86] Disable PAT (page attribute table extension of
                        pagetables) support.
 
index f6a1737c77be2dc128147978e7f9f85945aed7a6..87e5482acd0dca56c4169a8c5321057447bbf6a5 100644 (file)
@@ -52,6 +52,7 @@ struct trampoline_header {
        u64 efer;
        u32 cr4;
        u32 flags;
+       u32 lock;
 #endif
 };
 
@@ -64,6 +65,8 @@ extern unsigned long initial_stack;
 extern unsigned long initial_vc_handler;
 #endif
 
+extern u32 *trampoline_lock;
+
 extern unsigned char real_mode_blob[];
 extern unsigned char real_mode_relocs[];
 
index bf2c51df9e0b39f6f171696e6bda246d911cb5ce..1cf4f1e57570f379a97565ac343b58a9faf71a8a 100644 (file)
@@ -203,4 +203,10 @@ extern unsigned int smpboot_control;
 
 #endif /* !__ASSEMBLY__ */
 
+/* Control bits for startup_64 */
+#define STARTUP_APICID_CPUID_0B        0x80000000
+#define STARTUP_APICID_CPUID_01        0x40000000
+
+#define STARTUP_PARALLEL_MASK (STARTUP_APICID_CPUID_01 | STARTUP_APICID_CPUID_0B)
+
 #endif /* _ASM_X86_SMP_H */
index 9d2d88424c77e2f487478f5c88947e3fd585c8c2..5dcf5ca15383644c698f9957eadd693740718720 100644 (file)
@@ -16,6 +16,7 @@
 #include <asm/cacheflush.h>
 #include <asm/realmode.h>
 #include <asm/hypervisor.h>
+#include <asm/smp.h>
 
 #include <linux/ftrace.h>
 #include "../../realmode/rm/wakeup.h"
@@ -112,7 +113,13 @@ int x86_acpi_suspend_lowlevel(void)
 #else /* CONFIG_64BIT */
 #ifdef CONFIG_SMP
        current->thread.sp = (unsigned long)temp_stack + sizeof(temp_stack);
-       smpboot_control = smp_processor_id();
+       /*
+        * Ensure the CPU knows which one it is when it comes back, if
+        * it isn't in parallel mode and expected to work that out for
+        * itself.
+        */
+       if (!(smpboot_control & STARTUP_PARALLEL_MASK))
+               smpboot_control = smp_processor_id();
 #endif
        initial_code = (unsigned long)wakeup_long64;
        saved_magic = 0x123456789abcdef0L;
index 20d9a604da7c4b624f701182f0ead13244d1323f..ac1d7e5da1f233e1927f7a08d9b3db1b432cc4e2 100644 (file)
@@ -2377,7 +2377,7 @@ static int nr_logical_cpuids = 1;
 /*
  * Used to store mapping between logical CPU IDs and APIC IDs.
  */
-static int cpuid_to_apicid[] = {
+int cpuid_to_apicid[] = {
        [0 ... NR_CPUS - 1] = -1,
 };
 
index 069191e33490d967f065f3dab3379ce19f7fa0f7..17bdd6122dcaf3622a5cf55cf6a748519dcf659f 100644 (file)
@@ -25,6 +25,7 @@
 #include <asm/export.h>
 #include <asm/nospec-branch.h>
 #include <asm/fixmap.h>
+#include <asm/smp.h>
 
 /*
  * We are not able to switch in one step to the final KERNEL ADDRESS SPACE
@@ -234,8 +235,57 @@ SYM_INNER_LABEL(secondary_startup_64_no_verify, SYM_L_GLOBAL)
        ANNOTATE_NOENDBR // above
 
 #ifdef CONFIG_SMP
+       /*
+        * For parallel boot, the APIC ID is retrieved from CPUID, and then
+        * used to look up the CPU number.  For booting a single CPU, the
+        * CPU number is encoded in smpboot_control.
+        *
+        * Bit 31       STARTUP_APICID_CPUID_0B flag (use CPUID 0x0b)
+        * Bit 30       STARTUP_APICID_CPUID_01 flag (use CPUID 0x01)
+        * Bit 0-24     CPU# if STARTUP_APICID_CPUID_xx flags are not set
+        */
        movl    smpboot_control(%rip), %ecx
+       testl   $STARTUP_APICID_CPUID_0B, %ecx
+       jnz     .Luse_cpuid_0b
+       testl   $STARTUP_APICID_CPUID_01, %ecx
+       jnz     .Luse_cpuid_01
+       andl    $0x0FFFFFFF, %ecx
+       jmp     .Lsetup_cpu
+
+.Luse_cpuid_01:
+       mov     $0x01, %eax
+       cpuid
+       mov     %ebx, %edx
+       shr     $24, %edx
+       jmp     .Lsetup_AP
+
+.Luse_cpuid_0b:
+       mov     $0x0B, %eax
+       xorl    %ecx, %ecx
+       cpuid
+
+.Lsetup_AP:
+       /* EDX contains the APIC ID of the current CPU */
+       xorq    %rcx, %rcx
+       leaq    cpuid_to_apicid(%rip), %rbx
 
+.Lfind_cpunr:
+       cmpl    (%rbx,%rcx,4), %edx
+       jz      .Lsetup_cpu
+       inc     %ecx
+       cmpl    nr_cpu_ids(%rip), %ecx
+       jb      .Lfind_cpunr
+
+       /*  APIC ID not found in the table. Drop the trampoline lock and bail. */
+       movq    trampoline_lock(%rip), %rax
+       lock
+       btrl    $0, (%rax)
+
+1:     cli
+       hlt
+       jmp     1b
+
+.Lsetup_cpu:
        /* Get the per cpu offset for the given CPU# which is in ECX */
        movq    __per_cpu_offset(,%rcx,8), %rdx
 #else
@@ -244,7 +294,7 @@ SYM_INNER_LABEL(secondary_startup_64_no_verify, SYM_L_GLOBAL)
 
        /*
         * Setup a boot time stack - Any secondary CPU will have lost its stack
-        * by now because the cr3-switch above unmaps the real-mode stack.
+        * by now because the cr3-switch above unmaps the real-mode stack
         *
         * RDX contains the per-cpu offset
         */
@@ -293,6 +343,14 @@ SYM_INNER_LABEL(secondary_startup_64_no_verify, SYM_L_GLOBAL)
        shrq    $32, %rdx
        wrmsr
 
+       /* Drop the realmode protection. For the boot CPU the pointer is NULL! */
+       movq    trampoline_lock(%rip), %rax
+       testq   %rax, %rax
+       jz      .Lsetup_idt
+       lock
+       btrl    $0, (%rax)
+
+.Lsetup_idt:
        /* Setup and Load IDT */
        pushq   %rsi
        call    early_setup_idt
@@ -435,6 +493,8 @@ SYM_DATA(initial_code,      .quad x86_64_start_kernel)
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 SYM_DATA(initial_vc_handler,   .quad handle_vc_boot_ghcb)
 #endif
+
+SYM_DATA(trampoline_lock, .quad 0);
        __FINITDATA
 
        __INIT
@@ -665,7 +725,6 @@ SYM_DATA_END(level1_fixmap_pgt)
 
        .data
        .align 16
-
 SYM_DATA(smpboot_control,              .long 0)
 
        .align 16
index b04520085582baeee135d0722892c02f2f836c57..19b9b89b7458b410a013b02e445fa30e0672e39f 100644 (file)
@@ -797,6 +797,16 @@ static int __init cpu_init_udelay(char *str)
 }
 early_param("cpu_init_udelay", cpu_init_udelay);
 
+static bool do_parallel_bringup __ro_after_init = true;
+
+static int __init no_parallel_bringup(char *str)
+{
+       do_parallel_bringup = false;
+
+       return 0;
+}
+early_param("no_parallel_bringup", no_parallel_bringup);
+
 static void __init smp_quirk_init_udelay(void)
 {
        /* if cmdline changed it from default, leave it alone */
@@ -1113,7 +1123,7 @@ static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
        if (IS_ENABLED(CONFIG_X86_32)) {
                early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
                initial_stack  = idle->thread.sp;
-       } else {
+       } else if (!do_parallel_bringup) {
                smpboot_control = cpu;
        }
 
@@ -1515,6 +1525,47 @@ void __init native_smp_prepare_cpus(unsigned int max_cpus)
 
        speculative_store_bypass_ht_init();
 
+       /*
+        * We can do 64-bit AP bringup in parallel if the CPU reports
+        * its APIC ID in CPUID (either leaf 0x0B if we need the full
+        * APIC ID in X2APIC mode, or leaf 0x01 if 8 bits are
+        * sufficient). Otherwise it's too hard. And not for SEV-ES
+        * guests because they can't use CPUID that early.
+        */
+       if (IS_ENABLED(CONFIG_X86_32) || boot_cpu_data.cpuid_level < 1 ||
+           (x2apic_mode && boot_cpu_data.cpuid_level < 0xb) ||
+           cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
+               do_parallel_bringup = false;
+
+       if (do_parallel_bringup && x2apic_mode) {
+               unsigned int eax, ebx, ecx, edx;
+
+               /*
+                * To support parallel bringup in x2apic mode, the AP will need
+                * to obtain its APIC ID from CPUID 0x0B, since CPUID 0x01 has
+                * only 8 bits. Check that it is present and seems correct.
+                */
+               cpuid_count(0xb, 0, &eax, &ebx, &ecx, &edx);
+
+               /*
+                * AMD says that if executed with an umimplemented level in
+                * ECX, then it will return all zeroes in EAX. Intel says it
+                * will return zeroes in both EAX and EBX. Checking only EAX
+                * should be sufficient.
+                */
+               if (eax) {
+                       pr_debug("Using CPUID 0xb for parallel CPU startup\n");
+                       smpboot_control = STARTUP_APICID_CPUID_0B;
+               } else {
+                       pr_info("Disabling parallel bringup because CPUID 0xb looks untrustworthy\n");
+                       do_parallel_bringup = false;
+               }
+       } else if (do_parallel_bringup) {
+               /* Without X2APIC, what's in CPUID 0x01 should suffice. */
+               pr_debug("Using CPUID 0x1 for parallel CPU startup\n");
+               smpboot_control = STARTUP_APICID_CPUID_01;
+       }
+
        snp_set_wakeup_secondary_cpu();
 }
 
index af565816d2ba6aed5df706910d9e43e773b53fcb..788e5559549f39ab72573fb582d5964ce3490ed0 100644 (file)
@@ -154,6 +154,9 @@ static void __init setup_real_mode(void)
 
        trampoline_header->flags = 0;
 
+       trampoline_lock = &trampoline_header->lock;
+       *trampoline_lock = 0;
+
        trampoline_pgd = (u64 *) __va(real_mode_header->trampoline_pgd);
 
        /* Map the real mode stub as virtual == physical */
index e38d61d6562e4611c9150b935b018bee46717fe5..49ebc1636ffde1ae88d3709113292e9ef4778b10 100644 (file)
@@ -49,6 +49,19 @@ SYM_CODE_START(trampoline_start)
        mov     %ax, %es
        mov     %ax, %ss
 
+       /*
+        * Make sure only one CPU fiddles with the realmode stack
+        */
+.Llock_rm:
+       btl     $0, tr_lock
+       jnc     2f
+       pause
+       jmp     .Llock_rm
+2:
+       lock
+       btsl    $0, tr_lock
+       jc      .Llock_rm
+
        # Setup stack
        movl    $rm_stack_end, %esp
 
@@ -241,6 +254,7 @@ SYM_DATA_START(trampoline_header)
        SYM_DATA(tr_efer,               .space 8)
        SYM_DATA(tr_cr4,                .space 4)
        SYM_DATA(tr_flags,              .space 4)
+       SYM_DATA(tr_lock,               .space 4)
 SYM_DATA_END(trampoline_header)
 
 #include "trampoline_common.S"