[Skiboot] [PATCH 5/9] fast-reboot: improve fast reboot sequence
Nicholas Piggin
npiggin at gmail.com
Thu Feb 27 05:34:04 AEDT 2020
The current fast reboot sequence is not as robust as it could be. It
is this:
- Fast reboot CPU stops all other threads with direct control xscoms;
- it disables ME (machine checks become checkstops);
- resets its SPRs (to get HID[HILE] for machine check interrupts) and
overwrites exception vectors with our vectors, with a special fast
reboot sreset vector that fixes endian (because OS owns HILE);
- then the fast reboot CPU enables ME.
At this point the fast reboot CPU can handle machine checks with the
skiboot handler, but no other cores can if the OS had switched HILE
(they'll execute garbled byte swapped instructions and crash badly).
- Then all CPUs run various cleanups, XIVE, resync TOD, etc.
- The boot CPU, which is not necessarily the same as the fast reboot
initiator CPU, runs xive_reset.
This is a lot of code to run, including locking and xscoms, with
machine check inoperable.
- Finally secondaries are released and everyone sets SPRs and enables
ME.
Secondaries on other cores don't wait for their thread 0 to set shared
SPRs before calling into the normal OPAL secondary code. This is
mostly okay because the boot CPU pauses here until all secondaries
reach their idle code, but it's not nice to release them out of the
fast reboot code in a state with various per-core SPRs in flux.
Fix this by having the fast reboot CPU not disable ME or reset its
SPRs, because machine checks can still be handled by the OS. Then
wait until all CPUs are called into fast reboot and spinning with
ME disabled, only then reset any SPRs, copy remaining exception
vectors, and now skiboot has taken over the machine check handling,
then the CPUs enable ME before cleaning up other things.
This way, the region with ME disabled and SPRs and exception vectors
in flux is kept absolutely minimal, with no xscoms, no MMIOs, and few
significant memory modifications, and all threads kept closely in step.
There are no windows where a machine check interrupt may execute
garbage due to mismatched HILE on any CPU.
Signed-off-by: Nicholas Piggin <npiggin at gmail.com>
---
core/fast-reboot.c | 239 +++++++++++++++++++++++----------------------
include/cpu.h | 1 +
2 files changed, 121 insertions(+), 119 deletions(-)
diff --git a/core/fast-reboot.c b/core/fast-reboot.c
index 410acfe63..df207e91e 100644
--- a/core/fast-reboot.c
+++ b/core/fast-reboot.c
@@ -26,6 +26,20 @@
/* Flag tested by the OPAL entry code */
static volatile bool fast_boot_release;
+static volatile bool spr_set_release;
+static volatile bool nmi_mce_release;
+
+static void wait_on(volatile bool *cond)
+{
+ sync();
+ if (!*cond) {
+ smt_lowest();
+ while (!*cond)
+ barrier();
+ smt_medium();
+ }
+ sync();
+}
static bool cpu_state_wait_all_others(enum cpu_thread_state state,
unsigned long timeout_tb)
@@ -131,6 +145,8 @@ void fast_reboot(void)
prlog(PR_NOTICE, "RESET: Initiating fast reboot %d...\n", ++fast_reboot_count);
fast_boot_release = false;
+ spr_set_release = false;
+ nmi_mce_release = false;
sync();
/* Put everybody in stop except myself */
@@ -150,50 +166,16 @@ void fast_reboot(void)
cpu_set_ipi_enable(false);
/*
- * There is no point clearing special wakeup or un-quiesce due to
- * failure after this point, because we will be going to full IPL.
- * Less cleanup work means less opportunity to fail.
- */
-
- /*
- * Move SPRs and exception vectors back to OPAL-mode while all
- * others are quiesced. MSR[ME] is disabled while these are switched,
- * but system reset can not be blocked -- in theory an sreset coming
- * from the BMC or SPE could crash here.
- */
- disable_machine_check();
-
- /*
- * Primarily we want to fix up the HID bits here to allow the OPAL
- * exception handlers to work. Machine check would be the important
- * one.
- *
- * This is one case where a thread other than thread0 of the core
- * may update the shared SPRs. All other threads are stopped, so
- * there should be no races.
- */
- init_shared_sprs();
- init_replicated_sprs();
-
- /* Restore skiboot vectors */
- copy_exception_vectors();
- patch_traps(true);
-
- /*
- * Secondaries may still have an issue with machine checks if they have
- * HILE set because the machine check exception does not FIXUP_ENDIAN.
- * Adding that would trash CFAR however. So we have a window where
- * if a secondary takes an interrupt before the HILE is fixed, it will
- * crash.
+ * The fast reboot sreset vector has FIXUP_ENDIAN, so secondaries can
+ * cope with a wrong HILE setting.
*/
- enable_machine_check();
- mtmsrd(MSR_RI, 1);
+ copy_sreset_vector_fast_reboot();
/*
- * sreset vector has a FIXUP_ENDIAN sequence at the start, so
- * secondaries can cope.
+ * There is no point clearing special wakeup or un-quiesce due to
+ * failure after this point, because we will be going to full IPL.
+ * Less cleanup work means less opportunity to fail.
*/
- copy_sreset_vector_fast_reboot();
/* Send everyone else to 0x100 */
if (sreset_all_others() != OPAL_SUCCESS) {
@@ -203,7 +185,7 @@ void fast_reboot(void)
}
/* Ensure all the sresets get through */
- if (!cpu_state_wait_all_others(cpu_state_present, msecs_to_tb(1000))) {
+ if (!cpu_state_wait_all_others(cpu_state_fast_reboot_entry, msecs_to_tb(1000))) {
prlog(PR_NOTICE, "RESET: Fast reboot timed out waiting for "
"secondaries to call in\n");
return;
@@ -217,48 +199,12 @@ void fast_reboot(void)
console_complete_flush();
+ mtmsrd(0, 1); /* Clear MSR[RI] for 0x100 reset */
asm volatile("ba 0x100\n\t" : : : "memory");
for (;;)
;
}
-static void cleanup_cpu_state(void)
-{
- struct cpu_thread *cpu = this_cpu();
-
- /* Per core cleanup */
- if (cpu_is_thread0(cpu)) {
- /* Shared SPRs whacked back to normal */
-
- /* XXX Update the SLW copies ! Also dbl check HIDs etc... */
- init_shared_sprs();
-
- if (proc_gen == proc_gen_p8) {
- /* If somebody was in fast_sleep, we may have a
- * workaround to undo
- */
- if (cpu->in_fast_sleep) {
- prlog(PR_DEBUG, "RESET: CPU 0x%04x in fast sleep"
- " undoing workarounds...\n", cpu->pir);
- fast_sleep_exit();
- }
-
- /* The TLB surely contains garbage.
- * P9 clears TLBs in cpu_fast_reboot_complete
- */
- cleanup_local_tlb();
- }
-
- /* And we might have lost TB sync */
- chiptod_wakeup_resync();
- }
-
- /* Per-thread additional cleanup */
- init_replicated_sprs();
-
- // XXX Cleanup SLW, check HIDs ...
-}
-
void __noreturn enter_nap(void);
static void check_split_core(void)
@@ -310,17 +256,46 @@ static void check_split_core(void)
}
}
+static void cleanup_cpu_state(void)
+{
+ struct cpu_thread *cpu = this_cpu();
+
+ if (proc_gen == proc_gen_p9)
+ xive_cpu_reset();
+
+ /* Per core cleanup */
+ if (cpu_is_thread0(cpu)) {
+ /* XXX should reset the SLW SPR restore values*/
+
+ if (proc_gen == proc_gen_p8) {
+ /* If somebody was in fast_sleep, we may have a
+ * workaround to undo
+ */
+ if (cpu->in_fast_sleep) {
+ prlog(PR_DEBUG, "RESET: CPU 0x%04x in fast sleep"
+ " undoing workarounds...\n", cpu->pir);
+ fast_sleep_exit();
+ }
+
+ /* The TLB surely contains garbage.
+ * P9 clears TLBs in cpu_fast_reboot_complete
+ */
+ cleanup_local_tlb();
+ }
+
+ /* And we might have lost TB sync */
+ chiptod_wakeup_resync();
+ }
+}
/* Entry from asm after a fast reset */
void __noreturn fast_reboot_entry(void);
void __noreturn fast_reboot_entry(void)
{
- prlog(PR_DEBUG, "RESET: CPU 0x%04x reset in\n", this_cpu()->pir);
+ struct cpu_thread *cpu = this_cpu();
- if (proc_gen == proc_gen_p9) {
- xive_cpu_reset();
- } else if (proc_gen == proc_gen_p8) {
+ if (proc_gen == proc_gen_p8) {
/* We reset our ICP first ! Otherwise we might get stray
* interrupts when unsplitting
*/
@@ -332,51 +307,87 @@ void __noreturn fast_reboot_entry(void)
check_split_core();
}
+ /* Until SPRs (notably HID[HILE]) are set and new exception vectors
+ * installed, nobody should take machine checks. Try to do minimal
+ * work between these points.
+ */
+ disable_machine_check();
+ mtmsrd(0, 1); /* Clear RI */
+
sync();
- this_cpu()->state = cpu_state_present;
+ cpu->state = cpu_state_fast_reboot_entry;
sync();
+ if (cpu == boot_cpu) {
+ cpu_state_wait_all_others(cpu_state_fast_reboot_entry, 0);
+ spr_set_release = true;
+ } else {
+ wait_on(&spr_set_release);
+ }
- /* Are we the original boot CPU ? If not, we spin waiting
- * for a relase signal from CPU 1, then we clean ourselves
- * up and go processing jobs.
- */
- if (this_cpu() != boot_cpu) {
- if (!fast_boot_release) {
- smt_lowest();
- while (!fast_boot_release)
- barrier();
- smt_medium();
- }
- sync();
- cleanup_cpu_state();
- enable_machine_check();
- mtmsrd(MSR_RI, 1);
- __secondary_cpu_entry();
+ /* Reset SPRs */
+ if (cpu_is_thread0(cpu))
+ init_shared_sprs();
+ init_replicated_sprs();
+
+ if (cpu == boot_cpu) {
+ /* Restore skiboot vectors */
+ copy_exception_vectors();
+ copy_sreset_vector();
+ patch_traps(true);
}
- prlog(PR_INFO, "RESET: Boot CPU waiting for everybody...\n");
+ /* Must wait for others to because shared SPRs like HID0 are only set
+ * by thread0, so can't enable machine checks until those have been
+ * set.
+ */
+ sync();
+ cpu->state = cpu_state_present;
+ sync();
+ if (cpu == boot_cpu) {
+ cpu_state_wait_all_others(cpu_state_present, 0);
+ nmi_mce_release = true;
+ } else {
+ wait_on(&nmi_mce_release);
+ }
+
+ /* At this point skiboot exception vectors are in place and all
+ * cores/threads have SPRs set for running skiboot.
+ */
+ enable_machine_check();
+ mtmsrd(MSR_RI, 1);
+
+ cleanup_cpu_state();
+
+ prlog(PR_DEBUG, "RESET: CPU 0x%04x reset in\n", cpu->pir);
- /* We are the original boot CPU, wait for secondaries to
- * be captured.
+ /* The original boot CPU (not the fast reboot initiator) takes
+ * command. Secondaries wait for the signal then go to their secondary
+ * entry point.
*/
- cpu_state_wait_all_others(cpu_state_present, 0);
+ if (cpu != boot_cpu) {
+ wait_on(&fast_boot_release);
+
+ __secondary_cpu_entry();
+ }
if (proc_gen == proc_gen_p9)
xive_reset();
+ /* Let the CPU layer do some last minute global cleanups */
+ cpu_fast_reboot_complete();
+
+ /* We can now do NAP mode */
+ cpu_set_sreset_enable(true);
+ cpu_set_ipi_enable(true);
+
prlog(PR_INFO, "RESET: Releasing secondaries...\n");
/* Release everybody */
sync();
fast_boot_release = true;
-
- /* Cleanup ourselves */
- cleanup_cpu_state();
-
- /* Set our state to active */
sync();
- this_cpu()->state = cpu_state_active;
+ cpu->state = cpu_state_active;
sync();
/* Wait for them to respond */
@@ -389,16 +400,6 @@ void __noreturn fast_reboot_entry(void)
/* Clear release flag for next time */
fast_boot_release = false;
- /* Let the CPU layer do some last minute global cleanups */
- cpu_fast_reboot_complete();
-
- /* Restore OPAL's sreset vector now that all CPUs have HILE clear */
- copy_sreset_vector();
-
- /* We can now do NAP mode */
- cpu_set_sreset_enable(true);
- cpu_set_ipi_enable(true);
-
if (!chip_quirk(QUIRK_MAMBO_CALLOUTS)) {
/*
* mem_region_clear_unused avoids these preload regions
diff --git a/include/cpu.h b/include/cpu.h
index 9b7f41dfb..8ef20e35b 100644
--- a/include/cpu.h
+++ b/include/cpu.h
@@ -21,6 +21,7 @@ enum cpu_thread_state {
cpu_state_no_cpu = 0, /* Nothing there */
cpu_state_unknown, /* In PACA, not called in yet */
cpu_state_unavailable, /* Not available */
+ cpu_state_fast_reboot_entry, /* Called back into OPAL, real mode */
cpu_state_present, /* Assumed to spin in asm entry */
cpu_state_active, /* Secondary called in */
cpu_state_os, /* Under OS control */
--
2.23.0
More information about the Skiboot
mailing list