From 51676b14e8cfd5bbabf487f659f11704f17f6d0f Mon Sep 17 00:00:00 2001 From: Sven Schnelle Date: Sun, 29 Jul 2012 19:18:03 +0200 Subject: Revert "Use broadcast SIPI to startup siblings" This reverts commit 042c1461fb777e583e5de48edf9326e47ee5595f. It turned out that sending IPIs via broadcast doesn't work on Sandybridge. We tried to come up with a solution, but didn't found any so far. So revert the code for now until we have a working solution. Change-Id: I7dd1cba5a4c1e4b0af366b20e8263b1f6f4b9714 Signed-off-by: Sven Schnelle Reviewed-on: http://review.coreboot.org/1381 Tested-by: build bot (Jenkins) Reviewed-by: Ronald G. Minnich --- src/cpu/x86/lapic/lapic_cpu_init.c | 366 ++++++++++++++++++++++++++++++------- 1 file changed, 296 insertions(+), 70 deletions(-) (limited to 'src/cpu/x86/lapic/lapic_cpu_init.c') diff --git a/src/cpu/x86/lapic/lapic_cpu_init.c b/src/cpu/x86/lapic/lapic_cpu_init.c index b6dc560308..d61547dad0 100644 --- a/src/cpu/x86/lapic/lapic_cpu_init.c +++ b/src/cpu/x86/lapic/lapic_cpu_init.c @@ -60,28 +60,31 @@ static void copy_secondary_start_to_1m_below(void) printk(BIOS_DEBUG, "start_eip=0x%08lx, code_size=0x%08lx\n", (long unsigned int)AP_SIPI_VECTOR, code_size); } -static struct bus *current_cpu_bus; - -static int lapic_start_cpus(struct bus *cpu_bus) +static int lapic_start_cpu(unsigned long apicid) { int timeout; unsigned long send_status, accept_status; - int maxlvt; + int j, num_starts, maxlvt; /* * Starting actual IPI sequence... */ - current_cpu_bus = cpu_bus; - printk(BIOS_SPEW, "Asserting INIT.\n"); - /* Send INIT SIPI to target chip */ - lapic_write_around(LAPIC_ICR2, 0); - lapic_write_around(LAPIC_ICR, LAPIC_INT_ASSERT - | LAPIC_DM_INIT | LAPIC_DEST_ALLBUT); + /* + * Turn INIT on target chip + */ + lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(apicid)); + + /* + * Send IPI + */ + + lapic_write_around(LAPIC_ICR, LAPIC_INT_LEVELTRIG | LAPIC_INT_ASSERT + | LAPIC_DM_INIT); - printk(BIOS_DEBUG, "Waiting for send to finish...\n"); + printk(BIOS_SPEW, "Waiting for send to finish...\n"); timeout = 0; do { printk(BIOS_SPEW, "+"); @@ -89,66 +92,106 @@ static int lapic_start_cpus(struct bus *cpu_bus) send_status = lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY; } while (send_status && (timeout++ < 1000)); if (timeout >= 1000) { - printk(BIOS_DEBUG, "First apic write timed out. Disabling\n"); + printk(BIOS_ERR, "CPU %ld: First apic write timed out. Disabling\n", + apicid); // too bad. - printk(BIOS_DEBUG, "ESR is 0x%lx\n", lapic_read(LAPIC_ESR)); + printk(BIOS_ERR, "ESR is 0x%lx\n", lapic_read(LAPIC_ESR)); if (lapic_read(LAPIC_ESR)) { - printk(BIOS_DEBUG, "Try to reset ESR\n"); + printk(BIOS_ERR, "Try to reset ESR\n"); lapic_write_around(LAPIC_ESR, 0); - printk(BIOS_DEBUG, "ESR is 0x%lx\n", lapic_read(LAPIC_ESR)); + printk(BIOS_ERR, "ESR is 0x%lx\n", lapic_read(LAPIC_ESR)); } return 0; } +#if !CONFIG_CPU_AMD_MODEL_10XXX && !CONFIG_CPU_INTEL_MODEL_206AX + mdelay(10); +#endif - maxlvt = 4; - - printk(BIOS_SPEW, "Sending STARTUP.\n"); - lapic_read_around(LAPIC_SPIV); - lapic_write(LAPIC_ESR, 0); - lapic_read(LAPIC_ESR); - printk(BIOS_SPEW, "After apic_write.\n"); - - /* - * STARTUP IPI - */ + printk(BIOS_SPEW, "Deasserting INIT.\n"); /* Target chip */ - lapic_write_around(LAPIC_ICR2, 0); - - /* Boot on the stack */ - /* Kick the second */ - lapic_write_around(LAPIC_ICR, LAPIC_INT_ASSERT | LAPIC_DM_STARTUP | LAPIC_DEST_ALLBUT - | ((AP_SIPI_VECTOR >> 12) & 0xff)); + lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(apicid)); - /* - * Give the other CPU some time to accept the IPI. - */ - udelay(300); + /* Send IPI */ + lapic_write_around(LAPIC_ICR, LAPIC_INT_LEVELTRIG | LAPIC_DM_INIT); - printk(BIOS_DEBUG, "Startup point 1.\n"); - - printk(BIOS_DEBUG, "Waiting for send to finish...\n"); + printk(BIOS_SPEW, "Waiting for send to finish...\n"); timeout = 0; do { - printk(BIOS_DEBUG, "+"); + printk(BIOS_SPEW, "+"); udelay(100); send_status = lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY; } while (send_status && (timeout++ < 1000)); + if (timeout >= 1000) { + printk(BIOS_ERR, "CPU %ld: Second apic write timed out. Disabling\n", + apicid); + // too bad. + return 0; + } + +#if !CONFIG_CPU_AMD_MODEL_10XXX + num_starts = 2; +#else + num_starts = 1; +#endif /* - * Give the other CPU some time to accept the IPI. - */ - udelay(200); - /* - * Due to the Pentium erratum 3AP. + * Run STARTUP IPI loop. */ - if (maxlvt > 3) { + printk(BIOS_SPEW, "#startup loops: %d.\n", num_starts); + + maxlvt = 4; + + for (j = 1; j <= num_starts; j++) { + printk(BIOS_SPEW, "Sending STARTUP #%d to %lu.\n", j, apicid); lapic_read_around(LAPIC_SPIV); lapic_write(LAPIC_ESR, 0); + lapic_read(LAPIC_ESR); + printk(BIOS_SPEW, "After apic_write.\n"); + + /* + * STARTUP IPI + */ + + /* Target chip */ + lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(apicid)); + + /* Boot on the stack */ + /* Kick the second */ + lapic_write_around(LAPIC_ICR, LAPIC_DM_STARTUP + | (AP_SIPI_VECTOR >> 12)); + + /* + * Give the other CPU some time to accept the IPI. + */ + udelay(300); + + printk(BIOS_SPEW, "Startup point 1.\n"); + + printk(BIOS_SPEW, "Waiting for send to finish...\n"); + timeout = 0; + do { + printk(BIOS_SPEW, "+"); + udelay(100); + send_status = lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY; + } while (send_status && (timeout++ < 1000)); + + /* + * Give the other CPU some time to accept the IPI. + */ + udelay(200); + /* + * Due to the Pentium erratum 3AP. + */ + if (maxlvt > 3) { + lapic_read_around(LAPIC_SPIV); + lapic_write(LAPIC_ESR, 0); + } + accept_status = (lapic_read(LAPIC_ESR) & 0xEF); + if (send_status || accept_status) + break; } - accept_status = (lapic_read(LAPIC_ESR) & 0xEF); - - printk(BIOS_DEBUG, "After Startup.\n"); + printk(BIOS_SPEW, "After Startup.\n"); if (send_status) printk(BIOS_WARNING, "APIC never delivered???\n"); if (accept_status) @@ -158,34 +201,156 @@ static int lapic_start_cpus(struct bus *cpu_bus) return 1; } - /* Number of cpus that are currently running in coreboot */ static atomic_t active_cpus = ATOMIC_INIT(1); +/* start_cpu_lock covers last_cpu_index and secondary_stack. + * Only starting one cpu at a time let's me remove the logic + * for select the stack from assembly language. + * + * In addition communicating by variables to the cpu I + * am starting allows me to veryify it has started before + * start_cpu returns. + */ + +static spinlock_t start_cpu_lock = SPIN_LOCK_UNLOCKED; +static unsigned last_cpu_index = 0; volatile unsigned long secondary_stack; -extern unsigned char _estack[]; -static void stop_all_ap_cpus(void) +int start_cpu(device_t cpu) +{ + extern unsigned char _estack[]; + struct cpu_info *info; + unsigned long stack_end; + unsigned long apicid; + unsigned long index; + unsigned long count; + int result; + + spin_lock(&start_cpu_lock); + + /* Get the cpu's apicid */ + apicid = cpu->path.apic.apic_id; + + /* Get an index for the new processor */ + index = ++last_cpu_index; + + /* Find end of the new processors stack */ + stack_end = ((unsigned long)_estack) - (CONFIG_STACK_SIZE*index) - sizeof(struct cpu_info); + + /* Record the index and which cpu structure we are using */ + info = (struct cpu_info *)stack_end; + info->index = index; + info->cpu = cpu; + + /* Advertise the new stack to start_cpu */ + secondary_stack = stack_end; + + /* Until the cpu starts up report the cpu is not enabled */ + cpu->enabled = 0; + cpu->initialized = 0; + + /* Start the cpu */ + result = lapic_start_cpu(apicid); + + if (result) { + result = 0; + /* Wait 1s or until the new cpu calls in */ + for(count = 0; count < 100000 ; count++) { + if (secondary_stack == 0) { + result = 1; + break; + } + udelay(10); + } + } + secondary_stack = 0; + spin_unlock(&start_cpu_lock); + return result; +} + +#if CONFIG_AP_IN_SIPI_WAIT + +/** + * Sending INIT IPI to self is equivalent of asserting #INIT with a bit of delay. + * An undefined number of instruction cycles will complete. All global locks + * must be released before INIT IPI and no printk is allowed after this. + * De-asserting INIT IPI is a no-op on later Intel CPUs. + * + * If you set DEBUG_HALT_SELF to 1, printk's after INIT IPI are enabled + * but running thread may halt without releasing the lock and effectively + * deadlock other CPUs. + */ +#define DEBUG_HALT_SELF 0 + +/** + * Normally this function is defined in lapic.h as an always inline function + * that just keeps the CPU in a hlt() loop. This does not work on all CPUs. + * I think all hyperthreading CPUs might need this version, but I could only + * verify this on the Intel Core Duo + */ +void stop_this_cpu(void) { - unsigned long send_status; int timeout; - /* send an LAPIC INIT to all but myself */ - lapic_write_around(LAPIC_ICR2, 0); - lapic_write_around(LAPIC_ICR, LAPIC_INT_ASSERT | LAPIC_DM_INIT | LAPIC_DEST_ALLBUT); + unsigned long send_status; + unsigned long id; + + id = lapic_read(LAPIC_ID) >> 24; + + printk(BIOS_DEBUG, "CPU %ld going down...\n", id); + + /* send an LAPIC INIT to myself */ + lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(id)); + lapic_write_around(LAPIC_ICR, LAPIC_INT_LEVELTRIG | LAPIC_INT_ASSERT | LAPIC_DM_INIT); /* wait for the ipi send to finish */ +#if DEBUG_HALT_SELF printk(BIOS_SPEW, "Waiting for send to finish...\n"); +#endif timeout = 0; do { +#if DEBUG_HALT_SELF printk(BIOS_SPEW, "+"); +#endif udelay(100); send_status = lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY; } while (send_status && (timeout++ < 1000)); if (timeout >= 1000) { +#if DEBUG_HALT_SELF printk(BIOS_ERR, "timed out\n"); +#endif } mdelay(10); + +#if DEBUG_HALT_SELF + printk(BIOS_SPEW, "Deasserting INIT.\n"); +#endif + /* Deassert the LAPIC INIT */ + lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(id)); + lapic_write_around(LAPIC_ICR, LAPIC_INT_LEVELTRIG | LAPIC_DM_INIT); + +#if DEBUG_HALT_SELF + printk(BIOS_SPEW, "Waiting for send to finish...\n"); +#endif + timeout = 0; + do { +#if DEBUG_HALT_SELF + printk(BIOS_SPEW, "+"); +#endif + udelay(100); + send_status = lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY; + } while (send_status && (timeout++ < 1000)); + if (timeout >= 1000) { +#if DEBUG_HALT_SELF + printk(BIOS_ERR, "timed out\n"); +#endif + } + + while(1) { + hlt(); + } } +#endif #ifdef __SSE3__ static __inline__ __attribute__((always_inline)) unsigned long readcr4(void) @@ -208,21 +373,66 @@ static __inline__ __attribute__((always_inline)) void writecr4(unsigned long Dat #endif /* C entry point of secondary cpus */ -void secondary_cpu_init(int index) +void secondary_cpu_init(void) { + atomic_inc(&active_cpus); +#if CONFIG_SERIAL_CPU_INIT + spin_lock(&start_cpu_lock); +#endif + #ifdef __SSE3__ - /* - * Seems that CR4 was cleared when AP start via lapic_start_cpu() - * Turn on CR4.OSFXSR and CR4.OSXMMEXCPT when SSE options enabled - */ - u32 cr4_val; - cr4_val = readcr4(); - cr4_val |= (1 << 9 | 1 << 10); - writecr4(cr4_val); + /* + * Seems that CR4 was cleared when AP start via lapic_start_cpu() + * Turn on CR4.OSFXSR and CR4.OSXMMEXCPT when SSE options enabled + */ + u32 cr4_val; + cr4_val = readcr4(); + cr4_val |= (1 << 9 | 1 << 10); + writecr4(cr4_val); #endif - atomic_inc(&active_cpus); - cpu_initialize(current_cpu_bus, index); + cpu_initialize(); +#if CONFIG_SERIAL_CPU_INIT + spin_unlock(&start_cpu_lock); +#endif + atomic_dec(&active_cpus); + + stop_this_cpu(); +} + +static void start_other_cpus(struct bus *cpu_bus, device_t bsp_cpu) +{ + device_t cpu; + /* Loop through the cpus once getting them started */ + + for(cpu = cpu_bus->children; cpu ; cpu = cpu->sibling) { + if (cpu->path.type != DEVICE_PATH_APIC) { + continue; + } + #if !CONFIG_SERIAL_CPU_INIT + if(cpu==bsp_cpu) { + continue; + } + #endif + + if (!cpu->enabled) { + continue; + } + + if (cpu->initialized) { + continue; + } + + if (!start_cpu(cpu)) { + /* Record the error in cpu? */ + printk(BIOS_ERR, "CPU 0x%02x would not start!\n", + cpu->path.apic.apic_id); + } +#if CONFIG_SERIAL_CPU_INIT + udelay(10); +#endif + } + } static void wait_other_cpus_stop(struct bus *cpu_bus) @@ -255,7 +465,6 @@ static void wait_other_cpus_stop(struct bus *cpu_bus) cpu->path.apic.apic_id); } } - stop_all_ap_cpus(); printk(BIOS_DEBUG, "All AP CPUs stopped (%ld loops)\n", loopcount); } @@ -264,6 +473,10 @@ static void wait_other_cpus_stop(struct bus *cpu_bus) void initialize_cpus(struct bus *cpu_bus) { struct device_path cpu_path; + struct cpu_info *info; + + /* Find the info struct for this cpu */ + info = cpu_info(); #if NEED_LAPIC == 1 /* Ensure the local apic is enabled */ @@ -278,6 +491,9 @@ void initialize_cpus(struct bus *cpu_bus) cpu_path.cpu.id = 0; #endif + /* Find the device structure for the boot cpu */ + info->cpu = alloc_find_dev(cpu_bus, &cpu_path); + #if CONFIG_SMP copy_secondary_start_to_1m_below(); // why here? In case some day we can start core1 in amd_sibling_init #endif @@ -288,11 +504,21 @@ void initialize_cpus(struct bus *cpu_bus) cpus_ready_for_init(); +#if CONFIG_SMP + #if !CONFIG_SERIAL_CPU_INIT + /* start all aps at first, so we can init ECC all together */ + start_other_cpus(cpu_bus, info->cpu); + #endif +#endif + /* Initialize the bootstrap processor */ - cpu_initialize(cpu_bus, 0); + cpu_initialize(); #if CONFIG_SMP - lapic_start_cpus(cpu_bus); + #if CONFIG_SERIAL_CPU_INIT + start_other_cpus(cpu_bus, info->cpu); + #endif + /* Now wait the rest of the cpus stop*/ wait_other_cpus_stop(cpu_bus); #endif -- cgit v1.2.3