Ronald G. Minnich (rminnich@gmail.com) just uploaded a new patch set to gerrit, which you can find at http://review.coreboot.org/877
-gerrit
commit 7ebde200d7acc1409b952b738c41be28ddf9a4e7 Author: Ron Minnich rminnich@gmail.com Date: Thu Apr 5 23:51:18 2012 -0700
Add multicore support to coreboot
This set of changes supports the ability for multicore support in coreboot. To give people a chance to examine the structure, and to make bisecting for bugs easier, we are implementing the patch in 3 stages. This first stage introduces the basic mechanism but does not make any visible change in behavior. The APs, instead of being immediately spun down, are asked to run one debug print function and then spun down.
The means by which APs are tasked was implemented in the NIX operating system in 2011, see: http://code.google.com/p/nix-os/ The APs come alive and spin on a memory location, contained in a per-AP structure. Tasking is accomplished by setting parameters into an argument array and then writing a function pointer into the memory location. The AP indicates completion by writing zero to the memory location. The BSP can wait for the AP to finish (synchronous) or going off to do other work (asynchronous). Wait times are in units of microseconds.
This way of tasking APs is incredibly cheap and fast: in the minimal case, one memory write suffices to launch an AP into doing work.
This code has been tested on a sandybridge system (chromebook) and on another 4-core sandybridge system.
Take the opportunity to remove the macro definition whose origins few people remember. Change-Id: I19f8587562fd499e98457aa9f11b52400c105697 Signed-off-by: Ron Minnich rminnich@gmail.com --- src/arch/x86/include/arch/cpu.h | 5 ++ src/arch/x86/lib/cpu.c | 30 +++++++-- src/cpu/x86/lapic/lapic_cpu_init.c | 132 +++++++++++++++++++++++++++++++++--- src/include/cpu/cpu.h | 10 ++- 4 files changed, 160 insertions(+), 17 deletions(-)
diff --git a/src/arch/x86/include/arch/cpu.h b/src/arch/x86/include/arch/cpu.h index 604abde..34d0856 100644 --- a/src/arch/x86/include/arch/cpu.h +++ b/src/arch/x86/include/arch/cpu.h @@ -157,9 +157,14 @@ struct cpu_driver { struct device; struct cpu_driver *find_cpu_driver(struct device *cpu);
+typedef u32 (*workfunc)(u32, u32, u32); + struct cpu_info { device_t cpu; unsigned long index; + workfunc work; + u32 params[3]; + u32 result; };
static inline struct cpu_info *cpu_info(void) diff --git a/src/arch/x86/lib/cpu.c b/src/arch/x86/lib/cpu.c index 98ede06..ad028e3 100644 --- a/src/arch/x86/lib/cpu.c +++ b/src/arch/x86/lib/cpu.c @@ -234,7 +234,7 @@ static void set_cpu_ops(struct device *cpu) cpu->ops = driver ? driver->ops : NULL; }
-void cpu_initialize(void) +void cpu_initialize(struct cpu_info *info) { /* Because we busy wait at the printk spinlock. * It is important to keep the number of printed messages @@ -242,12 +242,9 @@ void cpu_initialize(void) * disabled. */ struct device *cpu; - struct cpu_info *info; struct cpuinfo_x86 c;
- info = cpu_info(); - - printk(BIOS_INFO, "Initializing CPU #%ld\n", info->index); + printk(BIOS_INFO, "cpu_initialize: CPU #%ld\n", info->index);
cpu = info->cpu; if (!cpu) { @@ -289,3 +286,26 @@ void cpu_initialize(void) return; }
+void cpu_work(struct cpu_info *info) +{ + workfunc f; + volatile workfunc *ptr = &info->work; + volatile u32 *params = info->params; + + printk(BIOS_INFO, "CPU #%ld ready to work\n", info->index); + + while (!*ptr) + ; + f = *ptr; + + printk(BIOS_SPEW, "CPU #%ld is asked to do %p\n", info->index, f); + f(params[0], params[1], params[2]); + + printk(BIOS_SPEW, "CPU #%ld finishes %p, mark %p\n", info->index, f, ptr); + *ptr = 0; + + printk(BIOS_INFO, "CPU #%ld leaving cpu_work()\n", info->index); + + return; +} + diff --git a/src/cpu/x86/lapic/lapic_cpu_init.c b/src/cpu/x86/lapic/lapic_cpu_init.c index 2ac9093..aca429a 100644 --- a/src/cpu/x86/lapic/lapic_cpu_init.c +++ b/src/cpu/x86/lapic/lapic_cpu_init.c @@ -17,6 +17,21 @@ #include <cpu/intel/speedstep.h>
#if CONFIG_SMP == 1 + +/* we do not want this struct visible outside this file. + * The external interface is via functions. + */ + +struct apcore { + u32 stack[1024 - sizeof(struct cpu_info)]; + struct cpu_info info; +}; + +#define TOS(x) (&apcores[(x)].stack[ARRAY_SIZE(apcores[x].stack)-1]) +typedef struct apcore apcore_t; + +apcore_t apcores[CONFIG_MAX_CPUS]; + /* This is a lot more paranoid now, since Linux can NOT handle * being told there is a CPU when none exists. So any errors * will return 0, meaning no CPU. @@ -227,7 +242,6 @@ volatile unsigned long secondary_stack;
int start_cpu(device_t cpu) { - extern unsigned char _estack[]; struct cpu_info *info; unsigned long stack_end; unsigned long apicid; @@ -241,19 +255,25 @@ int start_cpu(device_t cpu) apicid = cpu->path.apic.apic_id;
/* Get an index for the new processor */ + if (last_cpu_index >= CONFIG_MAX_CPUS){ + printk(BIOS_ERR, "CONFIG_MAX_CPUS(%d) too small!\n", CONFIG_MAX_CPUS); + return 0; + } + index = ++last_cpu_index;
/* Find end of the new processors stack */ - stack_end = ((unsigned long)_estack) - (CONFIG_STACK_SIZE*index) - sizeof(struct cpu_info); + stack_end = (unsigned long) TOS(index);
/* Record the index and which cpu structure we are using */ - info = (struct cpu_info *)stack_end; + info = &apcores[index].info; info->index = index; info->cpu = cpu;
/* Advertise the new stack to start_cpu */ secondary_stack = stack_end; - + printk(BIOS_SPEW, "start_cpu CPU %ld secondary_stack %#lx info %p\n", + index, secondary_stack, info); /* Until the cpu starts up report the cpu is not enabled */ cpu->enabled = 0; cpu->initialized = 0; @@ -381,8 +401,11 @@ static __inline__ __attribute__((always_inline)) void writecr4(unsigned long Dat #endif
/* C entry point of secondary cpus */ -void secondary_cpu_init(void) +void secondary_cpu_init(u32 infoptr) { + /* note: we tried (((u32 *)&infoptr)[1]) but that failed. Compiler? */ + struct cpu_info *info = (struct cpu_info *)(((u8*)&infoptr)+sizeof(u32)); + atomic_inc(&active_cpus); #if CONFIG_SERIAL_CPU_INIT == 1 spin_lock(&start_cpu_lock); @@ -398,11 +421,11 @@ void secondary_cpu_init(void) cr4_val |= (1 << 9 | 1 << 10); writecr4(cr4_val); #endif - cpu_initialize(); + cpu_initialize(info); #if CONFIG_SERIAL_CPU_INIT == 1 spin_unlock(&start_cpu_lock); #endif - + cpu_work(info); atomic_dec(&active_cpus);
stop_this_cpu(); @@ -476,8 +499,6 @@ static void wait_other_cpus_stop(struct bus *cpu_bus) printk(BIOS_DEBUG, "All AP CPUs stopped (%ld loops)\n", loopcount); }
-#else /* CONFIG_SMP */ -#define initialize_other_cpus(root) do {} while(0) #endif /* CONFIG_SMP */
void initialize_cpus(struct bus *cpu_bus) @@ -505,6 +526,7 @@ void initialize_cpus(struct bus *cpu_bus) info->cpu = alloc_find_dev(cpu_bus, &cpu_path);
#if CONFIG_SMP == 1 + memset(apcores, 0, sizeof(*apcores)); copy_secondary_start_to_1m_below(); // why here? In case some day we can start core1 in amd_sibling_init #endif
@@ -522,7 +544,7 @@ void initialize_cpus(struct bus *cpu_bus) #endif
/* Initialize the bootstrap processor */ - cpu_initialize(); + cpu_initialize(info);
#if CONFIG_SMP == 1 #if CONFIG_SERIAL_CPU_INIT == 1 @@ -534,3 +556,93 @@ void initialize_cpus(struct bus *cpu_bus) #endif }
+#if CONFIG_SMP == 1 +/* work primitives */ + +/* start_work is only intended to be called by the bsp. */ +/* A built in assumption of this code is that you know what + * you're doing. This is firmware, not pthreads. You should + * not call this function for a core if: + * - the core does not exist + * - the core is not initialized + * - the core is busy + * Any of these return a -1, else the core is started and + * 0 is returned. + */ +int start_work(unsigned int core, workfunc f, u32 a, u32 b, u32 c) +{ + struct cpu_info *info; + volatile workfunc *ptr; + volatile u32 *params; + + if (core >= CONFIG_MAX_CPUS){ + printk(BIOS_EMERG, "start_work: invalid core %d\n", core); + return -1; + } + info = &apcores[core].info; + if (! info->cpu->initialized){ + printk(BIOS_EMERG, "start_work: core not initialized %d\n", core); + return -1; + } + ptr = &info->work; + if (*ptr){ + printk(BIOS_EMERG, "start_work: core is busy %d\n", core); + return -1; + } + params = (u32 *)&info->params; + params[0] = a; + params[1] = b; + params[2] = c; + printk(BIOS_INFO, "BSP starts work on core %d\n", core); + *ptr = f; + return 0; +} + +/* wait for the work to finish and return the result. Wait at + * most maxusec microseconds. + */ +int wait_work(unsigned int core, u32 *retval, unsigned int maxusec) +{ + unsigned int usec; + struct cpu_info *info; + volatile workfunc *ptr; + volatile u32 *result; + + if (core >= CONFIG_MAX_CPUS){ + printk(BIOS_EMERG, "start_work: invalid core %d\n", core); + return -1; + } + info = &apcores[core].info; + if (! info->cpu->initialized){ + printk(BIOS_EMERG, "start_work: core not initialized %d\n", core); + return -1; + } + ptr = &info->work; + for(usec = 0; usec < maxusec && *ptr; usec++) + udelay(1); + + /* N.B. since only the BSP starts cores, there is not + * problem checking the pointer since access to it is + * serialized by the single-threaded BSP code. + */ + if (*ptr){ + printk(BIOS_INFO, "core %d still running after %d microseconds\n", + core, usec); + return -1; + } + result = &info->result; + printk(BIOS_INFO, "info->work after result is %#lx\n", (unsigned long)*result); + if (retval) + *retval = *result; + return 0; +} + +/* start the work and wait for it to finish */ +int run_work(unsigned int core, workfunc f, u32 a, u32 b, u32 c, u32 *retval, + unsigned int timeout) +{ + if (start_work(core, f, a, b, c)) + return -1; + return wait_work(core, retval, timeout); +} +#endif /* CONFIG_SMP == 1 */ diff --git a/src/include/cpu/cpu.h b/src/include/cpu/cpu.h index cca2be1..0b7a748 100644 --- a/src/include/cpu/cpu.h +++ b/src/include/cpu/cpu.h @@ -5,9 +5,15 @@ struct device; struct bus; #include <arch/cpu.h>
-void cpu_initialize(void); +void cpu_initialize(struct cpu_info *info); void initialize_cpus(struct bus *cpu_bus); -void secondary_cpu_init(void); +void secondary_cpu_init(u32 unused); +int start_work(unsigned int core, workfunc f, u32 a, u32 b, u32 c); +int wait_work(unsigned int core, u32 *retval, unsigned int maxusec); +int run_work(unsigned int core, workfunc f, u32 a, u32 b, u32 c, u32 *retval, + unsigned int timeout); +void cpu_work(struct cpu_info *info); +
#if !CONFIG_WAIT_BEFORE_CPUS_INIT #define cpus_ready_for_init() do {} while(0)