diff options
Diffstat (limited to 'arch/powerpc/oprofile')
| -rw-r--r-- | arch/powerpc/oprofile/Makefile | 8 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/backtrace.c | 2 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/cell/pr_util.h | 26 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/cell/spu_profiler.c | 65 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/cell/spu_task_sync.c | 254 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/cell/vma_map.c | 5 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/common.c | 54 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/op_model_7450.c | 23 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/op_model_cell.c | 776 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/op_model_fsl_emb.c | 61 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/op_model_pa6t.c | 7 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/op_model_power4.c | 147 | ||||
| -rw-r--r-- | arch/powerpc/oprofile/op_model_rs64.c | 2 |
13 files changed, 1158 insertions, 272 deletions
diff --git a/arch/powerpc/oprofile/Makefile b/arch/powerpc/oprofile/Makefile index 2ef6b0dddd8..751ec7bd501 100644 --- a/arch/powerpc/oprofile/Makefile +++ b/arch/powerpc/oprofile/Makefile @@ -1,6 +1,6 @@ -ifeq ($(CONFIG_PPC64),y) -EXTRA_CFLAGS += -mno-minimal-toc -endif +subdir-ccflags-$(CONFIG_PPC_WERROR) := -Werror + +ccflags-$(CONFIG_PPC64) := $(NO_MINIMAL_TOC) obj-$(CONFIG_OPROFILE) += oprofile.o @@ -14,6 +14,6 @@ oprofile-y := $(DRIVER_OBJS) common.o backtrace.o oprofile-$(CONFIG_OPROFILE_CELL) += op_model_cell.o \ cell/spu_profiler.o cell/vma_map.o \ cell/spu_task_sync.o -oprofile-$(CONFIG_PPC64) += op_model_rs64.o op_model_power4.o op_model_pa6t.o +oprofile-$(CONFIG_PPC_BOOK3S_64) += op_model_rs64.o op_model_power4.o op_model_pa6t.o oprofile-$(CONFIG_FSL_EMB_PERFMON) += op_model_fsl_emb.o oprofile-$(CONFIG_6xx) += op_model_7450.o diff --git a/arch/powerpc/oprofile/backtrace.c b/arch/powerpc/oprofile/backtrace.c index b4278cfd1f8..f75301f2c85 100644 --- a/arch/powerpc/oprofile/backtrace.c +++ b/arch/powerpc/oprofile/backtrace.c @@ -105,7 +105,7 @@ void op_powerpc_backtrace(struct pt_regs * const regs, unsigned int depth) } } else { #ifdef CONFIG_PPC64 - if (!test_thread_flag(TIF_32BIT)) { + if (!is_32bit_task()) { while (depth--) { sp = user_getsp64(sp, first_frame); if (!sp) diff --git a/arch/powerpc/oprofile/cell/pr_util.h b/arch/powerpc/oprofile/cell/pr_util.h index 22e4e8d4eb2..964b93974d8 100644 --- a/arch/powerpc/oprofile/cell/pr_util.h +++ b/arch/powerpc/oprofile/cell/pr_util.h @@ -24,6 +24,15 @@ #define SKIP_GENERIC_SYNC 0 #define SYNC_START_ERROR -1 #define DO_GENERIC_SYNC 1 +#define SPUS_PER_NODE 8 +#define DEFAULT_TIMER_EXPIRE (HZ / 10) + +extern struct delayed_work spu_work; +extern int spu_prof_running; + +#define TRACE_ARRAY_SIZE 1024 + +extern spinlock_t oprof_spu_smpl_arry_lck; struct spu_overlay_info { /* map of sections within an SPU overlay */ unsigned int vma; /* SPU virtual memory address from elf */ @@ -62,11 +71,19 @@ struct vma_to_fileoffset_map { /* map of sections within an SPU program */ }; +struct spu_buffer { + int last_guard_val; + int ctx_sw_seen; + unsigned long *buff; + unsigned int head, tail; +}; + + /* The three functions below are for maintaining and accessing * the vma-to-fileoffset map. */ struct vma_to_fileoffset_map *create_vma_map(const struct spu *spu, - u64 objectid); + unsigned long objectid); unsigned int vma_map_lookup(struct vma_to_fileoffset_map *map, unsigned int vma, const struct spu *aSpu, int *grd_val); @@ -76,10 +93,11 @@ void vma_map_free(struct vma_to_fileoffset_map *map); * Entry point for SPU profiling. * cycles_reset is the SPU_CYCLES count value specified by the user. */ -int start_spu_profiling(unsigned int cycles_reset); - -void stop_spu_profiling(void); +int start_spu_profiling_cycles(unsigned int cycles_reset); +void start_spu_profiling_events(void); +void stop_spu_profiling_cycles(void); +void stop_spu_profiling_events(void); /* add the necessary profiling hooks */ int spu_sync_start(void); diff --git a/arch/powerpc/oprofile/cell/spu_profiler.c b/arch/powerpc/oprofile/cell/spu_profiler.c index 380d7e21753..b129d007e7f 100644 --- a/arch/powerpc/oprofile/cell/spu_profiler.c +++ b/arch/powerpc/oprofile/cell/spu_profiler.c @@ -16,24 +16,34 @@ #include <linux/smp.h> #include <linux/slab.h> #include <asm/cell-pmu.h> +#include <asm/time.h> #include "pr_util.h" -#define TRACE_ARRAY_SIZE 1024 #define SCALE_SHIFT 14 static u32 *samples; -static int spu_prof_running; +/* spu_prof_running is a flag used to indicate if spu profiling is enabled + * or not. It is set by the routines start_spu_profiling_cycles() and + * start_spu_profiling_events(). The flag is cleared by the routines + * stop_spu_profiling_cycles() and stop_spu_profiling_events(). These + * routines are called via global_start() and global_stop() which are called in + * op_powerpc_start() and op_powerpc_stop(). These routines are called once + * per system as a result of the user starting/stopping oprofile. Hence, only + * one CPU per user at a time will be changing the value of spu_prof_running. + * In general, OProfile does not protect against multiple users trying to run + * OProfile at a time. + */ +int spu_prof_running; static unsigned int profiling_interval; #define NUM_SPU_BITS_TRBUF 16 #define SPUS_PER_TB_ENTRY 4 -#define SPUS_PER_NODE 8 #define SPU_PC_MASK 0xFFFF -static DEFINE_SPINLOCK(sample_array_lock); -unsigned long sample_array_lock_flags; +DEFINE_SPINLOCK(oprof_spu_smpl_arry_lck); +unsigned long oprof_spu_smpl_arry_lck_flags; void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset) { @@ -50,7 +60,7 @@ void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_rese * of precision. This is close enough for the purpose at hand. * * The value of the timeout should be small enough that the hw - * trace buffer will not get more then about 1/3 full for the + * trace buffer will not get more than about 1/3 full for the * maximum user specified (the LFSR value) hw sampling frequency. * This is to ensure the trace buffer will never fill even if the * kernel thread scheduling varies under a heavy system load. @@ -146,13 +156,13 @@ static enum hrtimer_restart profile_spus(struct hrtimer *timer) * sample array must be loaded and then processed for a given * cpu. The sample array is not per cpu. */ - spin_lock_irqsave(&sample_array_lock, - sample_array_lock_flags); + spin_lock_irqsave(&oprof_spu_smpl_arry_lck, + oprof_spu_smpl_arry_lck_flags); num_samples = cell_spu_pc_collection(cpu); if (num_samples == 0) { - spin_unlock_irqrestore(&sample_array_lock, - sample_array_lock_flags); + spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck, + oprof_spu_smpl_arry_lck_flags); continue; } @@ -163,8 +173,8 @@ static enum hrtimer_restart profile_spus(struct hrtimer *timer) num_samples); } - spin_unlock_irqrestore(&sample_array_lock, - sample_array_lock_flags); + spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck, + oprof_spu_smpl_arry_lck_flags); } smp_wmb(); /* insure spu event buffer updates are written */ @@ -183,20 +193,20 @@ static enum hrtimer_restart profile_spus(struct hrtimer *timer) static struct hrtimer timer; /* - * Entry point for SPU profiling. + * Entry point for SPU cycle profiling. * NOTE: SPU profiling is done system-wide, not per-CPU. * * cycles_reset is the count value specified by the user when * setting up OProfile to count SPU_CYCLES. */ -int start_spu_profiling(unsigned int cycles_reset) +int start_spu_profiling_cycles(unsigned int cycles_reset) { ktime_t kt; pr_debug("timer resolution: %lu\n", TICK_NSEC); kt = ktime_set(0, profiling_interval); hrtimer_init(&timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); - timer.expires = kt; + hrtimer_set_expires(&timer, kt); timer.function = profile_spus; /* Allocate arrays for collecting SPU PC samples */ @@ -208,14 +218,35 @@ int start_spu_profiling(unsigned int cycles_reset) spu_prof_running = 1; hrtimer_start(&timer, kt, HRTIMER_MODE_REL); + schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE); return 0; } -void stop_spu_profiling(void) +/* + * Entry point for SPU event profiling. + * NOTE: SPU profiling is done system-wide, not per-CPU. + * + * cycles_reset is the count value specified by the user when + * setting up OProfile to count SPU_CYCLES. + */ +void start_spu_profiling_events(void) +{ + spu_prof_running = 1; + schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE); + + return; +} + +void stop_spu_profiling_cycles(void) { spu_prof_running = 0; hrtimer_cancel(&timer); kfree(samples); - pr_debug("SPU_PROF: stop_spu_profiling issued\n"); + pr_debug("SPU_PROF: stop_spu_profiling_cycles issued\n"); +} + +void stop_spu_profiling_events(void) +{ + spu_prof_running = 0; } diff --git a/arch/powerpc/oprofile/cell/spu_task_sync.c b/arch/powerpc/oprofile/cell/spu_task_sync.c index 2a9b4a04932..28f1af2db1f 100644 --- a/arch/powerpc/oprofile/cell/spu_task_sync.c +++ b/arch/powerpc/oprofile/cell/spu_task_sync.c @@ -26,6 +26,7 @@ #include <linux/notifier.h> #include <linux/numa.h> #include <linux/oprofile.h> +#include <linux/slab.h> #include <linux/spinlock.h> #include "pr_util.h" @@ -35,7 +36,102 @@ static DEFINE_SPINLOCK(buffer_lock); static DEFINE_SPINLOCK(cache_lock); static int num_spu_nodes; int spu_prof_num_nodes; -int last_guard_val[MAX_NUMNODES * 8]; + +struct spu_buffer spu_buff[MAX_NUMNODES * SPUS_PER_NODE]; +struct delayed_work spu_work; +static unsigned max_spu_buff; + +static void spu_buff_add(unsigned long int value, int spu) +{ + /* spu buff is a circular buffer. Add entries to the + * head. Head is the index to store the next value. + * The buffer is full when there is one available entry + * in the queue, i.e. head and tail can't be equal. + * That way we can tell the difference between the + * buffer being full versus empty. + * + * ASSUPTION: the buffer_lock is held when this function + * is called to lock the buffer, head and tail. + */ + int full = 1; + + if (spu_buff[spu].head >= spu_buff[spu].tail) { + if ((spu_buff[spu].head - spu_buff[spu].tail) + < (max_spu_buff - 1)) + full = 0; + + } else if (spu_buff[spu].tail > spu_buff[spu].head) { + if ((spu_buff[spu].tail - spu_buff[spu].head) + > 1) + full = 0; + } + + if (!full) { + spu_buff[spu].buff[spu_buff[spu].head] = value; + spu_buff[spu].head++; + + if (spu_buff[spu].head >= max_spu_buff) + spu_buff[spu].head = 0; + } else { + /* From the user's perspective make the SPU buffer + * size management/overflow look like we are using + * per cpu buffers. The user uses the same + * per cpu parameter to adjust the SPU buffer size. + * Increment the sample_lost_overflow to inform + * the user the buffer size needs to be increased. + */ + oprofile_cpu_buffer_inc_smpl_lost(); + } +} + +/* This function copies the per SPU buffers to the + * OProfile kernel buffer. + */ +void sync_spu_buff(void) +{ + int spu; + unsigned long flags; + int curr_head; + + for (spu = 0; spu < num_spu_nodes; spu++) { + /* In case there was an issue and the buffer didn't + * get created skip it. + */ + if (spu_buff[spu].buff == NULL) + continue; + + /* Hold the lock to make sure the head/tail + * doesn't change while spu_buff_add() is + * deciding if the buffer is full or not. + * Being a little paranoid. + */ + spin_lock_irqsave(&buffer_lock, flags); + curr_head = spu_buff[spu].head; + spin_unlock_irqrestore(&buffer_lock, flags); + + /* Transfer the current contents to the kernel buffer. + * data can still be added to the head of the buffer. + */ + oprofile_put_buff(spu_buff[spu].buff, + spu_buff[spu].tail, + curr_head, max_spu_buff); + + spin_lock_irqsave(&buffer_lock, flags); + spu_buff[spu].tail = curr_head; + spin_unlock_irqrestore(&buffer_lock, flags); + } + +} + +static void wq_sync_spu_buff(struct work_struct *work) +{ + /* move data from spu buffers to kernel buffer */ + sync_spu_buff(); + + /* only reschedule if profiling is not done */ + if (spu_prof_running) + schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE); +} /* Container for caching information about an active SPU task. */ struct cached_info { @@ -202,13 +298,13 @@ static inline unsigned long fast_get_dcookie(struct path *path) { unsigned long cookie; - if (path->dentry->d_cookie) + if (path->dentry->d_flags & DCACHE_COOKIE) return (unsigned long)path->dentry; get_dcookie(path, &cookie); return cookie; } -/* Look up the dcookie for the task's first VM_EXECUTABLE mapping, +/* Look up the dcookie for the task's mm->exe_file, * which corresponds loosely to "application name". Also, determine * the offset for the SPU ELF object. If computed offset is * non-zero, it implies an embedded SPU object; otherwise, it's a @@ -225,7 +321,6 @@ get_exec_dcookie_and_offset(struct spu *spu, unsigned int *offsetp, { unsigned long app_cookie = 0; unsigned int my_offset = 0; - struct file *app = NULL; struct vm_area_struct *vma; struct mm_struct *mm = spu->mm; @@ -234,16 +329,10 @@ get_exec_dcookie_and_offset(struct spu *spu, unsigned int *offsetp, down_read(&mm->mmap_sem); - for (vma = mm->mmap; vma; vma = vma->vm_next) { - if (!vma->vm_file) - continue; - if (!(vma->vm_flags & VM_EXECUTABLE)) - continue; - app_cookie = fast_get_dcookie(&vma->vm_file->f_path); + if (mm->exe_file) { + app_cookie = fast_get_dcookie(&mm->exe_file->f_path); pr_debug("got dcookie for %s\n", - vma->vm_file->f_dentry->d_name.name); - app = vma->vm_file; - break; + mm->exe_file->f_dentry->d_name.name); } for (vma = mm->mmap; vma; vma = vma->vm_next) { @@ -305,14 +394,21 @@ static int process_context_switch(struct spu *spu, unsigned long objectId) /* Record context info in event buffer */ spin_lock_irqsave(&buffer_lock, flags); - add_event_entry(ESCAPE_CODE); - add_event_entry(SPU_CTX_SWITCH_CODE); - add_event_entry(spu->number); - add_event_entry(spu->pid); - add_event_entry(spu->tgid); - add_event_entry(app_dcookie); - add_event_entry(spu_cookie); - add_event_entry(offset); + spu_buff_add(ESCAPE_CODE, spu->number); + spu_buff_add(SPU_CTX_SWITCH_CODE, spu->number); + spu_buff_add(spu->number, spu->number); + spu_buff_add(spu->pid, spu->number); + spu_buff_add(spu->tgid, spu->number); + spu_buff_add(app_dcookie, spu->number); + spu_buff_add(spu_cookie, spu->number); + spu_buff_add(offset, spu->number); + + /* Set flag to indicate SPU PC data can now be written out. If + * the SPU program counter data is seen before an SPU context + * record is seen, the postprocessing will fail. + */ + spu_buff[spu->number].ctx_sw_seen = 1; + spin_unlock_irqrestore(&buffer_lock, flags); smp_wmb(); /* insure spu event buffer updates are written */ /* don't want entries intermingled... */ @@ -360,6 +456,47 @@ static int number_of_online_nodes(void) return nodes; } +static int oprofile_spu_buff_create(void) +{ + int spu; + + max_spu_buff = oprofile_get_cpu_buffer_size(); + + for (spu = 0; spu < num_spu_nodes; spu++) { + /* create circular buffers to store the data in. + * use locks to manage accessing the buffers + */ + spu_buff[spu].head = 0; + spu_buff[spu].tail = 0; + + /* + * Create a buffer for each SPU. Can't reliably + * create a single buffer for all spus due to not + * enough contiguous kernel memory. + */ + + spu_buff[spu].buff = kzalloc((max_spu_buff + * sizeof(unsigned long)), + GFP_KERNEL); + + if (!spu_buff[spu].buff) { + printk(KERN_ERR "SPU_PROF: " + "%s, line %d: oprofile_spu_buff_create " + "failed to allocate spu buffer %d.\n", + __func__, __LINE__, spu); + + /* release the spu buffers that have been allocated */ + while (spu >= 0) { + kfree(spu_buff[spu].buff); + spu_buff[spu].buff = 0; + spu--; + } + return -ENOMEM; + } + } + return 0; +} + /* The main purpose of this function is to synchronize * OProfile with SPUFS by registering to be notified of * SPU task switches. @@ -372,20 +509,35 @@ static int number_of_online_nodes(void) */ int spu_sync_start(void) { - int k; + int spu; int ret = SKIP_GENERIC_SYNC; int register_ret; unsigned long flags = 0; spu_prof_num_nodes = number_of_online_nodes(); num_spu_nodes = spu_prof_num_nodes * 8; + INIT_DELAYED_WORK(&spu_work, wq_sync_spu_buff); + + /* create buffer for storing the SPU data to put in + * the kernel buffer. + */ + ret = oprofile_spu_buff_create(); + if (ret) + goto out; spin_lock_irqsave(&buffer_lock, flags); - add_event_entry(ESCAPE_CODE); - add_event_entry(SPU_PROFILING_CODE); - add_event_entry(num_spu_nodes); + for (spu = 0; spu < num_spu_nodes; spu++) { + spu_buff_add(ESCAPE_CODE, spu); + spu_buff_add(SPU_PROFILING_CODE, spu); + spu_buff_add(num_spu_nodes, spu); + } spin_unlock_irqrestore(&buffer_lock, flags); + for (spu = 0; spu < num_spu_nodes; spu++) { + spu_buff[spu].ctx_sw_seen = 0; + spu_buff[spu].last_guard_val = 0; + } + /* Register for SPU events */ register_ret = spu_switch_event_register(&spu_active); if (register_ret) { @@ -393,8 +545,6 @@ int spu_sync_start(void) goto out; } - for (k = 0; k < (MAX_NUMNODES * 8); k++) - last_guard_val[k] = 0; pr_debug("spu_sync_start -- running.\n"); out: return ret; @@ -446,13 +596,20 @@ void spu_sync_buffer(int spu_num, unsigned int *samples, * use. We need to discard samples taken during the time * period which an overlay occurs (i.e., guard value changes). */ - if (grd_val && grd_val != last_guard_val[spu_num]) { - last_guard_val[spu_num] = grd_val; + if (grd_val && grd_val != spu_buff[spu_num].last_guard_val) { + spu_buff[spu_num].last_guard_val = grd_val; /* Drop the rest of the samples. */ break; } - add_event_entry(file_offset | spu_num_shifted); + /* We must ensure that the SPU context switch has been written + * out before samples for the SPU. Otherwise, the SPU context + * information is not available and the postprocessing of the + * SPU PC will fail with no available anonymous map information. + */ + if (spu_buff[spu_num].ctx_sw_seen) + spu_buff_add((file_offset | spu_num_shifted), + spu_num); } spin_unlock(&buffer_lock); out: @@ -463,20 +620,41 @@ out: int spu_sync_stop(void) { unsigned long flags = 0; - int ret = spu_switch_event_unregister(&spu_active); - if (ret) { + int ret; + int k; + + ret = spu_switch_event_unregister(&spu_active); + + if (ret) printk(KERN_ERR "SPU_PROF: " - "%s, line %d: spu_switch_event_unregister returned %d\n", - __func__, __LINE__, ret); - goto out; - } + "%s, line %d: spu_switch_event_unregister " \ + "returned %d\n", + __func__, __LINE__, ret); + + /* flush any remaining data in the per SPU buffers */ + sync_spu_buff(); spin_lock_irqsave(&cache_lock, flags); ret = release_cached_info(RELEASE_ALL); spin_unlock_irqrestore(&cache_lock, flags); -out: + + /* remove scheduled work queue item rather then waiting + * for every queued entry to execute. Then flush pending + * system wide buffer to event buffer. + */ + cancel_delayed_work(&spu_work); + + for (k = 0; k < num_spu_nodes; k++) { + spu_buff[k].ctx_sw_seen = 0; + + /* + * spu_sys_buff will be null if there was a problem + * allocating the buffer. Only delete if it exists. + */ + kfree(spu_buff[k].buff); + spu_buff[k].buff = 0; + } pr_debug("spu_sync_stop -- done.\n"); return ret; } - diff --git a/arch/powerpc/oprofile/cell/vma_map.c b/arch/powerpc/oprofile/cell/vma_map.c index fff66662d02..c579b16845d 100644 --- a/arch/powerpc/oprofile/cell/vma_map.c +++ b/arch/powerpc/oprofile/cell/vma_map.c @@ -20,6 +20,7 @@ #include <linux/string.h> #include <linux/uaccess.h> #include <linux/elf.h> +#include <linux/slab.h> #include "pr_util.h" @@ -185,7 +186,7 @@ struct vma_to_fileoffset_map *create_vma_map(const struct spu *aSpu, goto fail; if (shdr_str.sh_type != SHT_STRTAB) - goto fail;; + goto fail; for (j = 0; j < shdr.sh_size / sizeof (sym); j++) { if (copy_from_user(&sym, spu_elf_start + @@ -229,7 +230,7 @@ struct vma_to_fileoffset_map *create_vma_map(const struct spu *aSpu, */ overlay_tbl_offset = vma_map_lookup(map, ovly_table_sym, aSpu, &grd_val); - if (overlay_tbl_offset < 0) { + if (overlay_tbl_offset > 0x10000000) { printk(KERN_ERR "SPU_PROF: " "%s, line %d: Error finding SPU overlay table\n", __func__, __LINE__); diff --git a/arch/powerpc/oprofile/common.c b/arch/powerpc/oprofile/common.c index 17807acb05d..c77348c5d46 100644 --- a/arch/powerpc/oprofile/common.c +++ b/arch/powerpc/oprofile/common.c @@ -18,7 +18,6 @@ #include <linux/smp.h> #include <linux/errno.h> #include <asm/ptrace.h> -#include <asm/system.h> #include <asm/pmc.h> #include <asm/cputable.h> #include <asm/oprofile_impl.h> @@ -120,7 +119,7 @@ static void op_powerpc_stop(void) model->global_stop(); } -static int op_powerpc_create_files(struct super_block *sb, struct dentry *root) +static int op_powerpc_create_files(struct dentry *root) { int i; @@ -129,9 +128,31 @@ static int op_powerpc_create_files(struct super_block *sb, struct dentry *root) * There is one mmcr0, mmcr1 and mmcra for setting the events for * all of the counters. */ - oprofilefs_create_ulong(sb, root, "mmcr0", &sys.mmcr0); - oprofilefs_create_ulong(sb, root, "mmcr1", &sys.mmcr1); - oprofilefs_create_ulong(sb, root, "mmcra", &sys.mmcra); + oprofilefs_create_ulong(root, "mmcr0", &sys.mmcr0); + oprofilefs_create_ulong(root, "mmcr1", &sys.mmcr1); + oprofilefs_create_ulong(root, "mmcra", &sys.mmcra); +#ifdef CONFIG_OPROFILE_CELL + /* create a file the user tool can check to see what level of profiling + * support exits with this kernel. Initialize bit mask to indicate + * what support the kernel has: + * bit 0 - Supports SPU event profiling in addition to PPU + * event and cycles; and SPU cycle profiling + * bits 1-31 - Currently unused. + * + * If the file does not exist, then the kernel only supports SPU + * cycle profiling, PPU event and cycle profiling. + */ + oprofilefs_create_ulong(root, "cell_support", &sys.cell_support); + sys.cell_support = 0x1; /* Note, the user OProfile tool must check + * that this bit is set before attempting to + * user SPU event profiling. Older kernels + * will not have this file, hence the user + * tool is not allowed to do SPU event + * profiling on older kernels. Older kernels + * will accept SPU events but collected data + * is garbage. + */ +#endif #endif for (i = 0; i < model->num_counters; ++i) { @@ -139,11 +160,11 @@ static int op_powerpc_create_files(struct super_block *sb, struct dentry *root) char buf[4]; snprintf(buf, sizeof buf, "%d", i); - dir = oprofilefs_mkdir(sb, root, buf); + dir = oprofilefs_mkdir(root, buf); - oprofilefs_create_ulong(sb, dir, "enabled", &ctr[i].enabled); - oprofilefs_create_ulong(sb, dir, "event", &ctr[i].event); - oprofilefs_create_ulong(sb, dir, "count", &ctr[i].count); + oprofilefs_create_ulong(dir, "enabled", &ctr[i].enabled); + oprofilefs_create_ulong(dir, "event", &ctr[i].event); + oprofilefs_create_ulong(dir, "count", &ctr[i].count); /* * Classic PowerPC doesn't support per-counter @@ -152,14 +173,14 @@ static int op_powerpc_create_files(struct super_block *sb, struct dentry *root) * Book-E style performance monitors, we do * support them. */ - oprofilefs_create_ulong(sb, dir, "kernel", &ctr[i].kernel); - oprofilefs_create_ulong(sb, dir, "user", &ctr[i].user); + oprofilefs_create_ulong(dir, "kernel", &ctr[i].kernel); + oprofilefs_create_ulong(dir, "user", &ctr[i].user); - oprofilefs_create_ulong(sb, dir, "unit_mask", &ctr[i].unit_mask); + oprofilefs_create_ulong(dir, "unit_mask", &ctr[i].unit_mask); } - oprofilefs_create_ulong(sb, root, "enable_kernel", &sys.enable_kernel); - oprofilefs_create_ulong(sb, root, "enable_user", &sys.enable_user); + oprofilefs_create_ulong(root, "enable_kernel", &sys.enable_kernel); + oprofilefs_create_ulong(root, "enable_user", &sys.enable_user); /* Default to tracing both kernel and user */ sys.enable_kernel = 1; @@ -173,11 +194,8 @@ int __init oprofile_arch_init(struct oprofile_operations *ops) if (!cur_cpu_spec->oprofile_cpu_type) return -ENODEV; - if (firmware_has_feature(FW_FEATURE_ISERIES)) - return -ENODEV; - switch (cur_cpu_spec->oprofile_type) { -#ifdef CONFIG_PPC64 +#ifdef CONFIG_PPC_BOOK3S_64 #ifdef CONFIG_OPROFILE_CELL case PPC_OPROFILE_CELL: if (firmware_has_feature(FW_FEATURE_LPAR)) diff --git a/arch/powerpc/oprofile/op_model_7450.c b/arch/powerpc/oprofile/op_model_7450.c index cc599eb8768..d29b6e4e5e7 100644 --- a/arch/powerpc/oprofile/op_model_7450.c +++ b/arch/powerpc/oprofile/op_model_7450.c @@ -16,10 +16,8 @@ */ #include <linux/oprofile.h> -#include <linux/init.h> #include <linux/smp.h> #include <asm/ptrace.h> -#include <asm/system.h> #include <asm/processor.h> #include <asm/cputable.h> #include <asm/page.h> @@ -29,7 +27,7 @@ static unsigned long reset_value[OP_MAX_COUNTER]; static int oprofile_running; -static u32 mmcr0_val, mmcr1_val, mmcr2_val; +static u32 mmcr0_val, mmcr1_val, mmcr2_val, num_pmcs; #define MMCR0_PMC1_SHIFT 6 #define MMCR0_PMC2_SHIFT 0 @@ -88,13 +86,12 @@ static int fsl7450_cpu_setup(struct op_counter_config *ctr) mtspr(SPRN_MMCR0, mmcr0_val); mtspr(SPRN_MMCR1, mmcr1_val); - mtspr(SPRN_MMCR2, mmcr2_val); + if (num_pmcs > 4) + mtspr(SPRN_MMCR2, mmcr2_val); return 0; } -#define NUM_CTRS 6 - /* Configures the global settings for the countes on all CPUs. */ static int fsl7450_reg_setup(struct op_counter_config *ctr, struct op_system_config *sys, @@ -102,12 +99,13 @@ static int fsl7450_reg_setup(struct op_counter_config *ctr, { int i; + num_pmcs = num_ctrs; /* Our counters count up, and "count" refers to * how much before the next interrupt, and we interrupt * on overflow. So we calculate the starting value * which will give us "count" until overflow. * Then we set the events on the enabled counters */ - for (i = 0; i < NUM_CTRS; ++i) + for (i = 0; i < num_ctrs; ++i) reset_value[i] = 0x80000000UL - ctr[i].count; /* Set events for Counters 1 & 2 */ @@ -123,9 +121,10 @@ static int fsl7450_reg_setup(struct op_counter_config *ctr, /* Set events for Counters 3-6 */ mmcr1_val = mmcr1_event3(ctr[2].event) - | mmcr1_event4(ctr[3].event) - | mmcr1_event5(ctr[4].event) - | mmcr1_event6(ctr[5].event); + | mmcr1_event4(ctr[3].event); + if (num_ctrs > 4) + mmcr1_val |= mmcr1_event5(ctr[4].event) + | mmcr1_event6(ctr[5].event); mmcr2_val = 0; @@ -139,7 +138,7 @@ static int fsl7450_start(struct op_counter_config *ctr) mtmsr(mfmsr() | MSR_PMM); - for (i = 0; i < NUM_CTRS; ++i) { + for (i = 0; i < num_pmcs; ++i) { if (ctr[i].enabled) classic_ctr_write(i, reset_value[i]); else @@ -184,7 +183,7 @@ static void fsl7450_handle_interrupt(struct pt_regs *regs, pc = mfspr(SPRN_SIAR); is_kernel = is_kernel_addr(pc); - for (i = 0; i < NUM_CTRS; ++i) { + for (i = 0; i < num_pmcs; ++i) { val = classic_ctr_read(i); if (val < 0) { if (oprofile_running && ctr[i].enabled) { diff --git a/arch/powerpc/oprofile/op_model_cell.c b/arch/powerpc/oprofile/op_model_cell.c index 5ff4de3eb3b..863d89386f6 100644 --- a/arch/powerpc/oprofile/op_model_cell.c +++ b/arch/powerpc/oprofile/op_model_cell.c @@ -16,7 +16,6 @@ #include <linux/cpufreq.h> #include <linux/delay.h> -#include <linux/init.h> #include <linux/jiffies.h> #include <linux/kthread.h> #include <linux/oprofile.h> @@ -34,20 +33,20 @@ #include <asm/ptrace.h> #include <asm/reg.h> #include <asm/rtas.h> -#include <asm/system.h> #include <asm/cell-regs.h> #include "../platforms/cell/interrupt.h" #include "cell/pr_util.h" -static void cell_global_stop_spu(void); +#define PPU_PROFILING 0 +#define SPU_PROFILING_CYCLES 1 +#define SPU_PROFILING_EVENTS 2 -/* - * spu_cycle_reset is the number of cycles between samples. - * This variable is used for SPU profiling and should ONLY be set - * at the beginning of cell_reg_setup; otherwise, it's read-only. - */ -static unsigned int spu_cycle_reset; +#define SPU_EVENT_NUM_START 4100 +#define SPU_EVENT_NUM_STOP 4399 +#define SPU_PROFILE_EVENT_ADDR 4363 /* spu, address trace, decimal */ +#define SPU_PROFILE_EVENT_ADDR_MASK_A 0x146 /* sub unit set to zero */ +#define SPU_PROFILE_EVENT_ADDR_MASK_B 0x186 /* sub unit set to zero */ #define NUM_SPUS_PER_NODE 8 #define SPU_CYCLES_EVENT_NUM 2 /* event number for SPU_CYCLES */ @@ -66,6 +65,21 @@ static unsigned int spu_cycle_reset; #define MAX_SPU_COUNT 0xFFFFFF /* maximum 24 bit LFSR value */ +/* Minimum HW interval timer setting to send value to trace buffer is 10 cycle. + * To configure counter to send value every N cycles set counter to + * 2^32 - 1 - N. + */ +#define NUM_INTERVAL_CYC 0xFFFFFFFF - 10 + +/* + * spu_cycle_reset is the number of cycles between samples. + * This variable is used for SPU profiling and should ONLY be set + * at the beginning of cell_reg_setup; otherwise, it's read-only. + */ +static unsigned int spu_cycle_reset; +static unsigned int profiling_mode; +static int spu_evnt_phys_spu_indx; + struct pmc_cntrl_data { unsigned long vcntr; unsigned long evnts; @@ -105,6 +119,8 @@ struct pm_cntrl { u16 trace_mode; u16 freeze; u16 count_mode; + u16 spu_addr_trace; + u8 trace_buf_ovflw; }; static struct { @@ -122,7 +138,7 @@ static struct { #define GET_INPUT_CONTROL(x) ((x & 0x00000004) >> 2) static DEFINE_PER_CPU(unsigned long[NR_PHYS_CTRS], pmc_values); - +static unsigned long spu_pm_cnt[MAX_NUMNODES * NUM_SPUS_PER_NODE]; static struct pmc_cntrl_data pmc_cntrl[NUM_THREADS][NR_PHYS_CTRS]; /* @@ -152,6 +168,7 @@ static u32 hdw_thread; static u32 virt_cntr_inter_mask; static struct timer_list timer_virt_cntr; +static struct timer_list timer_spu_event_swap; /* * pm_signal needs to be global since it is initialized in @@ -165,7 +182,7 @@ static int spu_rtas_token; /* token for SPU cycle profiling */ static u32 reset_value[NR_PHYS_CTRS]; static int num_counters; static int oprofile_running; -static DEFINE_SPINLOCK(virt_cntr_lock); +static DEFINE_SPINLOCK(cntr_lock); static u32 ctr_enabled; @@ -229,7 +246,7 @@ static int pm_rtas_activate_signals(u32 node, u32 count) * There is no debug setup required for the cycles event. * Note that only events in the same group can be used. * Otherwise, there will be conflicts in correctly routing - * the signals on the debug bus. It is the responsiblity + * the signals on the debug bus. It is the responsibility * of the OProfile user tool to check the events are in * the same group. */ @@ -336,13 +353,13 @@ static void set_pm_event(u32 ctr, int event, u32 unit_mask) for (i = 0; i < NUM_DEBUG_BUS_WORDS; i++) { if (bus_word & (1 << i)) { pm_regs.debug_bus_control |= - (bus_type << (30 - (2 * i))); + (bus_type << (30 - (2 * i))); for (j = 0; j < NUM_INPUT_BUS_WORDS; j++) { if (input_bus[j] == 0xff) { input_bus[j] = i; pm_regs.group_control |= - (i << (30 - (2 * j))); + (i << (30 - (2 * j))); break; } @@ -367,12 +384,16 @@ static void write_pm_cntrl(int cpu) if (pm_regs.pm_cntrl.stop_at_max == 1) val |= CBE_PM_STOP_AT_MAX; - if (pm_regs.pm_cntrl.trace_mode == 1) + if (pm_regs.pm_cntrl.trace_mode != 0) val |= CBE_PM_TRACE_MODE_SET(pm_regs.pm_cntrl.trace_mode); + if (pm_regs.pm_cntrl.trace_buf_ovflw == 1) + val |= CBE_PM_TRACE_BUF_OVFLW(pm_regs.pm_cntrl.trace_buf_ovflw); if (pm_regs.pm_cntrl.freeze == 1) val |= CBE_PM_FREEZE_ALL_CTRS; + val |= CBE_PM_SPU_ADDR_TRACE_SET(pm_regs.pm_cntrl.spu_addr_trace); + /* * Routine set_count_mode must be called previously to set * the count mode based on the user selection of user and kernel. @@ -404,7 +425,7 @@ set_count_mode(u32 kernel, u32 user) } } -static inline void enable_ctr(u32 cpu, u32 ctr, u32 * pm07_cntrl) +static inline void enable_ctr(u32 cpu, u32 ctr, u32 *pm07_cntrl) { pm07_cntrl[ctr] |= CBE_PM_CTR_ENABLE; @@ -441,7 +462,7 @@ static void cell_virtual_cntr(unsigned long data) * not both playing with the counters on the same node. */ - spin_lock_irqsave(&virt_cntr_lock, flags); + spin_lock_irqsave(&cntr_lock, flags); prev_hdw_thread = hdw_thread; @@ -480,7 +501,7 @@ static void cell_virtual_cntr(unsigned long data) cbe_disable_pm_interrupts(cpu); for (i = 0; i < num_counters; i++) { per_cpu(pmc_values, cpu + prev_hdw_thread)[i] - = cbe_read_ctr(cpu, i); + = cbe_read_ctr(cpu, i); if (per_cpu(pmc_values, cpu + next_hdw_thread)[i] == 0xFFFFFFFF) @@ -527,7 +548,7 @@ static void cell_virtual_cntr(unsigned long data) cbe_enable_pm(cpu); } - spin_unlock_irqrestore(&virt_cntr_lock, flags); + spin_unlock_irqrestore(&cntr_lock, flags); mod_timer(&timer_virt_cntr, jiffies + HZ / 10); } @@ -541,38 +562,146 @@ static void start_virt_cntrs(void) add_timer(&timer_virt_cntr); } -/* This function is called once for all cpus combined */ -static int cell_reg_setup(struct op_counter_config *ctr, +static int cell_reg_setup_spu_cycles(struct op_counter_config *ctr, struct op_system_config *sys, int num_ctrs) { - int i, j, cpu; - spu_cycle_reset = 0; + spu_cycle_reset = ctr[0].count; - if (ctr[0].event == SPU_CYCLES_EVENT_NUM) { - spu_cycle_reset = ctr[0].count; + /* + * Each node will need to make the rtas call to start + * and stop SPU profiling. Get the token once and store it. + */ + spu_rtas_token = rtas_token("ibm,cbe-spu-perftools"); + + if (unlikely(spu_rtas_token == RTAS_UNKNOWN_SERVICE)) { + printk(KERN_ERR + "%s: rtas token ibm,cbe-spu-perftools unknown\n", + __func__); + return -EIO; + } + return 0; +} + +/* Unfortunately, the hardware will only support event profiling + * on one SPU per node at a time. Therefore, we must time slice + * the profiling across all SPUs in the node. Note, we do this + * in parallel for each node. The following routine is called + * periodically based on kernel timer to switch which SPU is + * being monitored in a round robbin fashion. + */ +static void spu_evnt_swap(unsigned long data) +{ + int node; + int cur_phys_spu, nxt_phys_spu, cur_spu_evnt_phys_spu_indx; + unsigned long flags; + int cpu; + int ret; + u32 interrupt_mask; + + + /* enable interrupts on cntr 0 */ + interrupt_mask = CBE_PM_CTR_OVERFLOW_INTR(0); + + hdw_thread = 0; + + /* Make sure spu event interrupt handler and spu event swap + * don't access the counters simultaneously. + */ + spin_lock_irqsave(&cntr_lock, flags); + + cur_spu_evnt_phys_spu_indx = spu_evnt_phys_spu_indx; + + if (++(spu_evnt_phys_spu_indx) == NUM_SPUS_PER_NODE) + spu_evnt_phys_spu_indx = 0; + + pm_signal[0].sub_unit = spu_evnt_phys_spu_indx; + pm_signal[1].sub_unit = spu_evnt_phys_spu_indx; + pm_signal[2].sub_unit = spu_evnt_phys_spu_indx; + + /* switch the SPU being profiled on each node */ + for_each_online_cpu(cpu) { + if (cbe_get_hw_thread_id(cpu)) + continue; + + node = cbe_cpu_to_node(cpu); + cur_phys_spu = (node * NUM_SPUS_PER_NODE) + + cur_spu_evnt_phys_spu_indx; + nxt_phys_spu = (node * NUM_SPUS_PER_NODE) + + spu_evnt_phys_spu_indx; /* - * Each node will need to make the rtas call to start - * and stop SPU profiling. Get the token once and store it. + * stop counters, save counter values, restore counts + * for previous physical SPU */ - spu_rtas_token = rtas_token("ibm,cbe-spu-perftools"); + cbe_disable_pm(cpu); + cbe_disable_pm_interrupts(cpu); - if (unlikely(spu_rtas_token == RTAS_UNKNOWN_SERVICE)) { - printk(KERN_ERR - "%s: rtas token ibm,cbe-spu-perftools unknown\n", - __func__); - return -EIO; - } + spu_pm_cnt[cur_phys_spu] + = cbe_read_ctr(cpu, 0); + + /* restore previous count for the next spu to sample */ + /* NOTE, hardware issue, counter will not start if the + * counter value is at max (0xFFFFFFFF). + */ + if (spu_pm_cnt[nxt_phys_spu] >= 0xFFFFFFFF) + cbe_write_ctr(cpu, 0, 0xFFFFFFF0); + else + cbe_write_ctr(cpu, 0, spu_pm_cnt[nxt_phys_spu]); + + pm_rtas_reset_signals(cbe_cpu_to_node(cpu)); + + /* setup the debug bus measure the one event and + * the two events to route the next SPU's PC on + * the debug bus + */ + ret = pm_rtas_activate_signals(cbe_cpu_to_node(cpu), 3); + if (ret) + printk(KERN_ERR "%s: pm_rtas_activate_signals failed, " + "SPU event swap\n", __func__); + + /* clear the trace buffer, don't want to take PC for + * previous SPU*/ + cbe_write_pm(cpu, trace_address, 0); + + enable_ctr(cpu, 0, pm_regs.pm07_cntrl); + + /* Enable interrupts on the CPU thread that is starting */ + cbe_enable_pm_interrupts(cpu, hdw_thread, + interrupt_mask); + cbe_enable_pm(cpu); } - pm_rtas_token = rtas_token("ibm,cbe-perftools"); + spin_unlock_irqrestore(&cntr_lock, flags); + + /* swap approximately every 0.1 seconds */ + mod_timer(&timer_spu_event_swap, jiffies + HZ / 25); +} + +static void start_spu_event_swap(void) +{ + init_timer(&timer_spu_event_swap); + timer_spu_event_swap.function = spu_evnt_swap; + timer_spu_event_swap.data = 0UL; + timer_spu_event_swap.expires = jiffies + HZ / 25; + add_timer(&timer_spu_event_swap); +} + +static int cell_reg_setup_spu_events(struct op_counter_config *ctr, + struct op_system_config *sys, int num_ctrs) +{ + int i; + + /* routine is called once for all nodes */ + spu_evnt_phys_spu_indx = 0; /* - * For all events excetp PPU CYCLEs, each node will need to make + * For all events except PPU CYCLEs, each node will need to make * the rtas cbe-perftools call to setup and reset the debug bus. * Make the token lookup call once and store it in the global * variable pm_rtas_token. */ + pm_rtas_token = rtas_token("ibm,cbe-perftools"); + if (unlikely(pm_rtas_token == RTAS_UNKNOWN_SERVICE)) { printk(KERN_ERR "%s: rtas token ibm,cbe-perftools unknown\n", @@ -580,16 +709,67 @@ static int cell_reg_setup(struct op_counter_config *ctr, return -EIO; } - num_counters = num_ctrs; + /* setup the pm_control register settings, + * settings will be written per node by the + * cell_cpu_setup() function. + */ + pm_regs.pm_cntrl.trace_buf_ovflw = 1; - pm_regs.group_control = 0; - pm_regs.debug_bus_control = 0; + /* Use the occurrence trace mode to have SPU PC saved + * to the trace buffer. Occurrence data in trace buffer + * is not used. Bit 2 must be set to store SPU addresses. + */ + pm_regs.pm_cntrl.trace_mode = 2; + + pm_regs.pm_cntrl.spu_addr_trace = 0x1; /* using debug bus + event 2 & 3 */ + + /* setup the debug bus event array with the SPU PC routing events. + * Note, pm_signal[0] will be filled in by set_pm_event() call below. + */ + pm_signal[1].signal_group = SPU_PROFILE_EVENT_ADDR / 100; + pm_signal[1].bus_word = GET_BUS_WORD(SPU_PROFILE_EVENT_ADDR_MASK_A); + pm_signal[1].bit = SPU_PROFILE_EVENT_ADDR % 100; + pm_signal[1].sub_unit = spu_evnt_phys_spu_indx; + + pm_signal[2].signal_group = SPU_PROFILE_EVENT_ADDR / 100; + pm_signal[2].bus_word = GET_BUS_WORD(SPU_PROFILE_EVENT_ADDR_MASK_B); + pm_signal[2].bit = SPU_PROFILE_EVENT_ADDR % 100; + pm_signal[2].sub_unit = spu_evnt_phys_spu_indx; + + /* Set the user selected spu event to profile on, + * note, only one SPU profiling event is supported + */ + num_counters = 1; /* Only support one SPU event at a time */ + set_pm_event(0, ctr[0].event, ctr[0].unit_mask); - /* setup the pm_control register */ - memset(&pm_regs.pm_cntrl, 0, sizeof(struct pm_cntrl)); - pm_regs.pm_cntrl.stop_at_max = 1; - pm_regs.pm_cntrl.trace_mode = 0; - pm_regs.pm_cntrl.freeze = 1; + reset_value[0] = 0xFFFFFFFF - ctr[0].count; + + /* global, used by cell_cpu_setup */ + ctr_enabled |= 1; + + /* Initialize the count for each SPU to the reset value */ + for (i=0; i < MAX_NUMNODES * NUM_SPUS_PER_NODE; i++) + spu_pm_cnt[i] = reset_value[0]; + + return 0; +} + +static int cell_reg_setup_ppu(struct op_counter_config *ctr, + struct op_system_config *sys, int num_ctrs) +{ + /* routine is called once for all nodes */ + int i, j, cpu; + + num_counters = num_ctrs; + + if (unlikely(num_ctrs > NR_PHYS_CTRS)) { + printk(KERN_ERR + "%s: Oprofile, number of specified events " \ + "exceeds number of physical counters\n", + __func__); + return -EIO; + } set_count_mode(sys->enable_kernel, sys->enable_user); @@ -658,6 +838,63 @@ static int cell_reg_setup(struct op_counter_config *ctr, } +/* This function is called once for all cpus combined */ +static int cell_reg_setup(struct op_counter_config *ctr, + struct op_system_config *sys, int num_ctrs) +{ + int ret=0; + spu_cycle_reset = 0; + + /* initialize the spu_arr_trace value, will be reset if + * doing spu event profiling. + */ + pm_regs.group_control = 0; + pm_regs.debug_bus_control = 0; + pm_regs.pm_cntrl.stop_at_max = 1; + pm_regs.pm_cntrl.trace_mode = 0; + pm_regs.pm_cntrl.freeze = 1; + pm_regs.pm_cntrl.trace_buf_ovflw = 0; + pm_regs.pm_cntrl.spu_addr_trace = 0; + + /* + * For all events except PPU CYCLEs, each node will need to make + * the rtas cbe-perftools call to setup and reset the debug bus. + * Make the token lookup call once and store it in the global + * variable pm_rtas_token. + */ + pm_rtas_token = rtas_token("ibm,cbe-perftools"); + + if (unlikely(pm_rtas_token == RTAS_UNKNOWN_SERVICE)) { + printk(KERN_ERR + "%s: rtas token ibm,cbe-perftools unknown\n", + __func__); + return -EIO; + } + + if (ctr[0].event == SPU_CYCLES_EVENT_NUM) { + profiling_mode = SPU_PROFILING_CYCLES; + ret = cell_reg_setup_spu_cycles(ctr, sys, num_ctrs); + } else if ((ctr[0].event >= SPU_EVENT_NUM_START) && + (ctr[0].event <= SPU_EVENT_NUM_STOP)) { + profiling_mode = SPU_PROFILING_EVENTS; + spu_cycle_reset = ctr[0].count; + + /* for SPU event profiling, need to setup the + * pm_signal array with the events to route the + * SPU PC before making the FW call. Note, only + * one SPU event for profiling can be specified + * at a time. + */ + cell_reg_setup_spu_events(ctr, sys, num_ctrs); + } else { + profiling_mode = PPU_PROFILING; + ret = cell_reg_setup_ppu(ctr, sys, num_ctrs); + } + + return ret; +} + + /* This function is called once for each cpu */ static int cell_cpu_setup(struct op_counter_config *cntr) @@ -665,8 +902,13 @@ static int cell_cpu_setup(struct op_counter_config *cntr) u32 cpu = smp_processor_id(); u32 num_enabled = 0; int i; + int ret; - if (spu_cycle_reset) + /* Cycle based SPU profiling does not use the performance + * counters. The trace array is configured to collect + * the data. + */ + if (profiling_mode == SPU_PROFILING_CYCLES) return 0; /* There is one performance monitor per processor chip (i.e. node), @@ -679,7 +921,6 @@ static int cell_cpu_setup(struct op_counter_config *cntr) cbe_disable_pm(cpu); cbe_disable_pm_interrupts(cpu); - cbe_write_pm(cpu, pm_interval, 0); cbe_write_pm(cpu, pm_start_stop, 0); cbe_write_pm(cpu, group_control, pm_regs.group_control); cbe_write_pm(cpu, debug_bus_control, pm_regs.debug_bus_control); @@ -696,7 +937,20 @@ static int cell_cpu_setup(struct op_counter_config *cntr) * The pm_rtas_activate_signals will return -EIO if the FW * call failed. */ - return pm_rtas_activate_signals(cbe_cpu_to_node(cpu), num_enabled); + if (profiling_mode == SPU_PROFILING_EVENTS) { + /* For SPU event profiling also need to setup the + * pm interval timer + */ + ret = pm_rtas_activate_signals(cbe_cpu_to_node(cpu), + num_enabled+2); + /* store PC from debug bus to Trace buffer as often + * as possible (every 10 cycles) + */ + cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC); + return ret; + } else + return pm_rtas_activate_signals(cbe_cpu_to_node(cpu), + num_enabled); } #define ENTRIES 303 @@ -821,7 +1075,7 @@ static int calculate_lfsr(int n) index = ENTRIES-1; /* make sure index is valid */ - if ((index > ENTRIES) || (index < 0)) + if ((index >= ENTRIES) || (index < 0)) index = ENTRIES-1; return initial_lfsr[index]; @@ -830,13 +1084,13 @@ static int calculate_lfsr(int n) static int pm_rtas_activate_spu_profiling(u32 node) { int ret, i; - struct pm_signal pm_signal_local[NR_PHYS_CTRS]; + struct pm_signal pm_signal_local[NUM_SPUS_PER_NODE]; /* * Set up the rtas call to configure the debug bus to * route the SPU PCs. Setup the pm_signal for each SPU */ - for (i = 0; i < NUM_SPUS_PER_NODE; i++) { + for (i = 0; i < ARRAY_SIZE(pm_signal_local); i++) { pm_signal_local[i].cpu = node; pm_signal_local[i].signal_group = 41; /* spu i on word (i/2) */ @@ -848,7 +1102,7 @@ static int pm_rtas_activate_spu_profiling(u32 node) ret = rtas_ibm_cbe_perftools(SUBFUNC_ACTIVATE, PASSTHRU_ENABLE, pm_signal_local, - (NUM_SPUS_PER_NODE + (ARRAY_SIZE(pm_signal_local) * sizeof(struct pm_signal))); if (unlikely(ret)) { @@ -867,8 +1121,7 @@ oprof_cpufreq_notify(struct notifier_block *nb, unsigned long val, void *data) int ret = 0; struct cpufreq_freqs *frq = data; if ((val == CPUFREQ_PRECHANGE && frq->old < frq->new) || - (val == CPUFREQ_POSTCHANGE && frq->old > frq->new) || - (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) + (val == CPUFREQ_POSTCHANGE && frq->old > frq->new)) set_spu_profiling_frequency(frq->new, spu_cycle_reset); return ret; } @@ -878,7 +1131,122 @@ static struct notifier_block cpu_freq_notifier_block = { }; #endif -static int cell_global_start_spu(struct op_counter_config *ctr) +/* + * Note the generic OProfile stop calls do not support returning + * an error on stop. Hence, will not return an error if the FW + * calls fail on stop. Failure to reset the debug bus is not an issue. + * Failure to disable the SPU profiling is not an issue. The FW calls + * to enable the performance counters and debug bus will work even if + * the hardware was not cleanly reset. + */ +static void cell_global_stop_spu_cycles(void) +{ + int subfunc, rtn_value; + unsigned int lfsr_value; + int cpu; + + oprofile_running = 0; + smp_wmb(); + +#ifdef CONFIG_CPU_FREQ + cpufreq_unregister_notifier(&cpu_freq_notifier_block, + CPUFREQ_TRANSITION_NOTIFIER); +#endif + + for_each_online_cpu(cpu) { + if (cbe_get_hw_thread_id(cpu)) + continue; + + subfunc = 3; /* + * 2 - activate SPU tracing, + * 3 - deactivate + */ + lfsr_value = 0x8f100000; + + rtn_value = rtas_call(spu_rtas_token, 3, 1, NULL, + subfunc, cbe_cpu_to_node(cpu), + lfsr_value); + + if (unlikely(rtn_value != 0)) { + printk(KERN_ERR + "%s: rtas call ibm,cbe-spu-perftools " \ + "failed, return = %d\n", + __func__, rtn_value); + } + + /* Deactivate the signals */ + pm_rtas_reset_signals(cbe_cpu_to_node(cpu)); + } + + stop_spu_profiling_cycles(); +} + +static void cell_global_stop_spu_events(void) +{ + int cpu; + oprofile_running = 0; + + stop_spu_profiling_events(); + smp_wmb(); + + for_each_online_cpu(cpu) { + if (cbe_get_hw_thread_id(cpu)) + continue; + + cbe_sync_irq(cbe_cpu_to_node(cpu)); + /* Stop the counters */ + cbe_disable_pm(cpu); + cbe_write_pm07_control(cpu, 0, 0); + + /* Deactivate the signals */ + pm_rtas_reset_signals(cbe_cpu_to_node(cpu)); + + /* Deactivate interrupts */ + cbe_disable_pm_interrupts(cpu); + } + del_timer_sync(&timer_spu_event_swap); +} + +static void cell_global_stop_ppu(void) +{ + int cpu; + + /* + * This routine will be called once for the system. + * There is one performance monitor per node, so we + * only need to perform this function once per node. + */ + del_timer_sync(&timer_virt_cntr); + oprofile_running = 0; + smp_wmb(); + + for_each_online_cpu(cpu) { + if (cbe_get_hw_thread_id(cpu)) + continue; + + cbe_sync_irq(cbe_cpu_to_node(cpu)); + /* Stop the counters */ + cbe_disable_pm(cpu); + + /* Deactivate the signals */ + pm_rtas_reset_signals(cbe_cpu_to_node(cpu)); + + /* Deactivate interrupts */ + cbe_disable_pm_interrupts(cpu); + } +} + +static void cell_global_stop(void) +{ + if (profiling_mode == PPU_PROFILING) + cell_global_stop_ppu(); + else if (profiling_mode == SPU_PROFILING_EVENTS) + cell_global_stop_spu_events(); + else + cell_global_stop_spu_cycles(); +} + +static int cell_global_start_spu_cycles(struct op_counter_config *ctr) { int subfunc; unsigned int lfsr_value; @@ -944,18 +1312,18 @@ static int cell_global_start_spu(struct op_counter_config *ctr) /* start profiling */ ret = rtas_call(spu_rtas_token, 3, 1, NULL, subfunc, - cbe_cpu_to_node(cpu), lfsr_value); + cbe_cpu_to_node(cpu), lfsr_value); if (unlikely(ret != 0)) { printk(KERN_ERR - "%s: rtas call ibm,cbe-spu-perftools failed, return = %d\n", - __func__, ret); + "%s: rtas call ibm,cbe-spu-perftools failed, " \ + "return = %d\n", __func__, ret); rtas_error = -EIO; goto out; } } - rtas_error = start_spu_profiling(spu_cycle_reset); + rtas_error = start_spu_profiling_cycles(spu_cycle_reset); if (rtas_error) goto out_stop; @@ -963,11 +1331,74 @@ static int cell_global_start_spu(struct op_counter_config *ctr) return 0; out_stop: - cell_global_stop_spu(); /* clean up the PMU/debug bus */ + cell_global_stop_spu_cycles(); /* clean up the PMU/debug bus */ out: return rtas_error; } +static int cell_global_start_spu_events(struct op_counter_config *ctr) +{ + int cpu; + u32 interrupt_mask = 0; + int rtn = 0; + + hdw_thread = 0; + + /* spu event profiling, uses the performance counters to generate + * an interrupt. The hardware is setup to store the SPU program + * counter into the trace array. The occurrence mode is used to + * enable storing data to the trace buffer. The bits are set + * to send/store the SPU address in the trace buffer. The debug + * bus must be setup to route the SPU program counter onto the + * debug bus. The occurrence data in the trace buffer is not used. + */ + + /* This routine gets called once for the system. + * There is one performance monitor per node, so we + * only need to perform this function once per node. + */ + + for_each_online_cpu(cpu) { + if (cbe_get_hw_thread_id(cpu)) + continue; + + /* + * Setup SPU event-based profiling. + * Set perf_mon_control bit 0 to a zero before + * enabling spu collection hardware. + * + * Only support one SPU event on one SPU per node. + */ + if (ctr_enabled & 1) { + cbe_write_ctr(cpu, 0, reset_value[0]); + enable_ctr(cpu, 0, pm_regs.pm07_cntrl); + interrupt_mask |= + CBE_PM_CTR_OVERFLOW_INTR(0); + } else { + /* Disable counter */ + cbe_write_pm07_control(cpu, 0, 0); + } + + cbe_get_and_clear_pm_interrupts(cpu); + cbe_enable_pm_interrupts(cpu, hdw_thread, interrupt_mask); + cbe_enable_pm(cpu); + + /* clear the trace buffer */ + cbe_write_pm(cpu, trace_address, 0); + } + + /* Start the timer to time slice collecting the event profile + * on each of the SPUs. Note, can collect profile on one SPU + * per node at a time. + */ + start_spu_event_swap(); + start_spu_profiling_events(); + oprofile_running = 1; + smp_wmb(); + + return rtn; +} + static int cell_global_start_ppu(struct op_counter_config *ctr) { u32 cpu, i; @@ -987,8 +1418,7 @@ static int cell_global_start_ppu(struct op_counter_config *ctr) if (ctr_enabled & (1 << i)) { cbe_write_ctr(cpu, i, reset_value[i]); enable_ctr(cpu, i, pm_regs.pm07_cntrl); - interrupt_mask |= - CBE_PM_CTR_OVERFLOW_INTR(i); + interrupt_mask |= CBE_PM_CTR_OVERFLOW_INTR(i); } else { /* Disable counter */ cbe_write_pm07_control(cpu, i, 0); @@ -1017,99 +1447,162 @@ static int cell_global_start_ppu(struct op_counter_config *ctr) static int cell_global_start(struct op_counter_config *ctr) { - if (spu_cycle_reset) - return cell_global_start_spu(ctr); + if (profiling_mode == SPU_PROFILING_CYCLES) + return cell_global_start_spu_cycles(ctr); + else if (profiling_mode == SPU_PROFILING_EVENTS) + return cell_global_start_spu_events(ctr); else return cell_global_start_ppu(ctr); } -/* - * Note the generic OProfile stop calls do not support returning - * an error on stop. Hence, will not return an error if the FW - * calls fail on stop. Failure to reset the debug bus is not an issue. - * Failure to disable the SPU profiling is not an issue. The FW calls - * to enable the performance counters and debug bus will work even if - * the hardware was not cleanly reset. + +/* The SPU interrupt handler + * + * SPU event profiling works as follows: + * The pm_signal[0] holds the one SPU event to be measured. It is routed on + * the debug bus using word 0 or 1. The value of pm_signal[1] and + * pm_signal[2] contain the necessary events to route the SPU program + * counter for the selected SPU onto the debug bus using words 2 and 3. + * The pm_interval register is setup to write the SPU PC value into the + * trace buffer at the maximum rate possible. The trace buffer is configured + * to store the PCs, wrapping when it is full. The performance counter is + * initialized to the max hardware count minus the number of events, N, between + * samples. Once the N events have occurred, a HW counter overflow occurs + * causing the generation of a HW counter interrupt which also stops the + * writing of the SPU PC values to the trace buffer. Hence the last PC + * written to the trace buffer is the SPU PC that we want. Unfortunately, + * we have to read from the beginning of the trace buffer to get to the + * last value written. We just hope the PPU has nothing better to do then + * service this interrupt. The PC for the specific SPU being profiled is + * extracted from the trace buffer processed and stored. The trace buffer + * is cleared, interrupts are cleared, the counter is reset to max - N. + * A kernel timer is used to periodically call the routine spu_evnt_swap() + * to switch to the next physical SPU in the node to profile in round robbin + * order. This way data is collected for all SPUs on the node. It does mean + * that we need to use a relatively small value of N to ensure enough samples + * on each SPU are collected each SPU is being profiled 1/8 of the time. + * It may also be necessary to use a longer sample collection period. */ -static void cell_global_stop_spu(void) +static void cell_handle_interrupt_spu(struct pt_regs *regs, + struct op_counter_config *ctr) { - int subfunc, rtn_value; - unsigned int lfsr_value; - int cpu; + u32 cpu, cpu_tmp; + u64 trace_entry; + u32 interrupt_mask; + u64 trace_buffer[2]; + u64 last_trace_buffer; + u32 sample; + u32 trace_addr; + unsigned long sample_array_lock_flags; + int spu_num; + unsigned long flags; - oprofile_running = 0; + /* Make sure spu event interrupt handler and spu event swap + * don't access the counters simultaneously. + */ + cpu = smp_processor_id(); + spin_lock_irqsave(&cntr_lock, flags); -#ifdef CONFIG_CPU_FREQ - cpufreq_unregister_notifier(&cpu_freq_notifier_block, - CPUFREQ_TRANSITION_NOTIFIER); -#endif + cpu_tmp = cpu; + cbe_disable_pm(cpu); - for_each_online_cpu(cpu) { - if (cbe_get_hw_thread_id(cpu)) - continue; + interrupt_mask = cbe_get_and_clear_pm_interrupts(cpu); - subfunc = 3; /* - * 2 - activate SPU tracing, - * 3 - deactivate - */ - lfsr_value = 0x8f100000; + sample = 0xABCDEF; + trace_entry = 0xfedcba; + last_trace_buffer = 0xdeadbeaf; - rtn_value = rtas_call(spu_rtas_token, 3, 1, NULL, - subfunc, cbe_cpu_to_node(cpu), - lfsr_value); + if ((oprofile_running == 1) && (interrupt_mask != 0)) { + /* disable writes to trace buff */ + cbe_write_pm(cpu, pm_interval, 0); - if (unlikely(rtn_value != 0)) { - printk(KERN_ERR - "%s: rtas call ibm,cbe-spu-perftools failed, return = %d\n", - __func__, rtn_value); + /* only have one perf cntr being used, cntr 0 */ + if ((interrupt_mask & CBE_PM_CTR_OVERFLOW_INTR(0)) + && ctr[0].enabled) + /* The SPU PC values will be read + * from the trace buffer, reset counter + */ + + cbe_write_ctr(cpu, 0, reset_value[0]); + + trace_addr = cbe_read_pm(cpu, trace_address); + + while (!(trace_addr & CBE_PM_TRACE_BUF_EMPTY)) { + /* There is data in the trace buffer to process + * Read the buffer until you get to the last + * entry. This is the value we want. + */ + + cbe_read_trace_buffer(cpu, trace_buffer); + trace_addr = cbe_read_pm(cpu, trace_address); } - /* Deactivate the signals */ - pm_rtas_reset_signals(cbe_cpu_to_node(cpu)); - } + /* SPU Address 16 bit count format for 128 bit + * HW trace buffer is used for the SPU PC storage + * HDR bits 0:15 + * SPU Addr 0 bits 16:31 + * SPU Addr 1 bits 32:47 + * unused bits 48:127 + * + * HDR: bit4 = 1 SPU Address 0 valid + * HDR: bit5 = 1 SPU Address 1 valid + * - unfortunately, the valid bits don't seem to work + * + * Note trace_buffer[0] holds bits 0:63 of the HW + * trace buffer, trace_buffer[1] holds bits 64:127 + */ - stop_spu_profiling(); -} + trace_entry = trace_buffer[0] + & 0x00000000FFFF0000; -static void cell_global_stop_ppu(void) -{ - int cpu; + /* only top 16 of the 18 bit SPU PC address + * is stored in trace buffer, hence shift right + * by 16 -2 bits */ + sample = trace_entry >> 14; + last_trace_buffer = trace_buffer[0]; - /* - * This routine will be called once for the system. - * There is one performance monitor per node, so we - * only need to perform this function once per node. - */ - del_timer_sync(&timer_virt_cntr); - oprofile_running = 0; - smp_wmb(); + spu_num = spu_evnt_phys_spu_indx + + (cbe_cpu_to_node(cpu) * NUM_SPUS_PER_NODE); - for_each_online_cpu(cpu) { - if (cbe_get_hw_thread_id(cpu)) - continue; + /* make sure only one process at a time is calling + * spu_sync_buffer() + */ + spin_lock_irqsave(&oprof_spu_smpl_arry_lck, + sample_array_lock_flags); + spu_sync_buffer(spu_num, &sample, 1); + spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck, + sample_array_lock_flags); - cbe_sync_irq(cbe_cpu_to_node(cpu)); - /* Stop the counters */ - cbe_disable_pm(cpu); + smp_wmb(); /* insure spu event buffer updates are written + * don't want events intermingled... */ - /* Deactivate the signals */ - pm_rtas_reset_signals(cbe_cpu_to_node(cpu)); + /* The counters were frozen by the interrupt. + * Reenable the interrupt and restart the counters. + */ + cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC); + cbe_enable_pm_interrupts(cpu, hdw_thread, + virt_cntr_inter_mask); - /* Deactivate interrupts */ - cbe_disable_pm_interrupts(cpu); - } -} + /* clear the trace buffer, re-enable writes to trace buff */ + cbe_write_pm(cpu, trace_address, 0); + cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC); -static void cell_global_stop(void) -{ - if (spu_cycle_reset) - cell_global_stop_spu(); - else - cell_global_stop_ppu(); + /* The writes to the various performance counters only writes + * to a latch. The new values (interrupt setting bits, reset + * counter value etc.) are not copied to the actual registers + * until the performance monitor is enabled. In order to get + * this to work as desired, the performance monitor needs to + * be disabled while writing to the latches. This is a + * HW design issue. + */ + write_pm_cntrl(cpu); + cbe_enable_pm(cpu); + } + spin_unlock_irqrestore(&cntr_lock, flags); } -static void cell_handle_interrupt(struct pt_regs *regs, - struct op_counter_config *ctr) +static void cell_handle_interrupt_ppu(struct pt_regs *regs, + struct op_counter_config *ctr) { u32 cpu; u64 pc; @@ -1125,7 +1618,7 @@ static void cell_handle_interrupt(struct pt_regs *regs, * routine are not running at the same time. See the * cell_virtual_cntr() routine for additional comments. */ - spin_lock_irqsave(&virt_cntr_lock, flags); + spin_lock_irqsave(&cntr_lock, flags); /* * Need to disable and reenable the performance counters @@ -1160,7 +1653,7 @@ static void cell_handle_interrupt(struct pt_regs *regs, * The counters were frozen by the interrupt. * Reenable the interrupt and restart the counters. * If there was a race between the interrupt handler and - * the virtual counter routine. The virutal counter + * the virtual counter routine. The virtual counter * routine may have cleared the interrupts. Hence must * use the virt_cntr_inter_mask to re-enable the interrupts. */ @@ -1172,13 +1665,22 @@ static void cell_handle_interrupt(struct pt_regs *regs, * to a latch. The new values (interrupt setting bits, reset * counter value etc.) are not copied to the actual registers * until the performance monitor is enabled. In order to get - * this to work as desired, the permormance monitor needs to + * this to work as desired, the performance monitor needs to * be disabled while writing to the latches. This is a * HW design issue. */ cbe_enable_pm(cpu); } - spin_unlock_irqrestore(&virt_cntr_lock, flags); + spin_unlock_irqrestore(&cntr_lock, flags); +} + +static void cell_handle_interrupt(struct pt_regs *regs, + struct op_counter_config *ctr) +{ + if (profiling_mode == PPU_PROFILING) + cell_handle_interrupt_ppu(regs, ctr); + else + cell_handle_interrupt_spu(regs, ctr); } /* @@ -1188,7 +1690,8 @@ static void cell_handle_interrupt(struct pt_regs *regs, */ static int cell_sync_start(void) { - if (spu_cycle_reset) + if ((profiling_mode == SPU_PROFILING_CYCLES) || + (profiling_mode == SPU_PROFILING_EVENTS)) return spu_sync_start(); else return DO_GENERIC_SYNC; @@ -1196,7 +1699,8 @@ static int cell_sync_start(void) static int cell_sync_stop(void) { - if (spu_cycle_reset) + if ((profiling_mode == SPU_PROFILING_CYCLES) || + (profiling_mode == SPU_PROFILING_EVENTS)) return spu_sync_stop(); else return 1; diff --git a/arch/powerpc/oprofile/op_model_fsl_emb.c b/arch/powerpc/oprofile/op_model_fsl_emb.c index 91596f6ba1f..14cf86fddda 100644 --- a/arch/powerpc/oprofile/op_model_fsl_emb.c +++ b/arch/powerpc/oprofile/op_model_fsl_emb.c @@ -2,7 +2,7 @@ * Freescale Embedded oprofile support, based on ppc64 oprofile support * Copyright (C) 2004 Anton Blanchard <anton@au.ibm.com>, IBM * - * Copyright (c) 2004 Freescale Semiconductor, Inc + * Copyright (c) 2004, 2010 Freescale Semiconductor, Inc * * Author: Andy Fleming * Maintainer: Kumar Gala <galak@kernel.crashing.org> @@ -14,10 +14,8 @@ */ #include <linux/oprofile.h> -#include <linux/init.h> #include <linux/smp.h> #include <asm/ptrace.h> -#include <asm/system.h> #include <asm/processor.h> #include <asm/cputable.h> #include <asm/reg_fsl_emb.h> @@ -47,6 +45,12 @@ static inline u32 get_pmlca(int ctr) case 3: pmlca = mfpmr(PMRN_PMLCA3); break; + case 4: + pmlca = mfpmr(PMRN_PMLCA4); + break; + case 5: + pmlca = mfpmr(PMRN_PMLCA5); + break; default: panic("Bad ctr number\n"); } @@ -69,6 +73,12 @@ static inline void set_pmlca(int ctr, u32 pmlca) case 3: mtpmr(PMRN_PMLCA3, pmlca); break; + case 4: + mtpmr(PMRN_PMLCA4, pmlca); + break; + case 5: + mtpmr(PMRN_PMLCA5, pmlca); + break; default: panic("Bad ctr number\n"); } @@ -85,6 +95,10 @@ static inline unsigned int ctr_read(unsigned int i) return mfpmr(PMRN_PMC2); case 3: return mfpmr(PMRN_PMC3); + case 4: + return mfpmr(PMRN_PMC4); + case 5: + return mfpmr(PMRN_PMC5); default: return 0; } @@ -105,6 +119,12 @@ static inline void ctr_write(unsigned int i, unsigned int val) case 3: mtpmr(PMRN_PMC3, val); break; + case 4: + mtpmr(PMRN_PMC4, val); + break; + case 5: + mtpmr(PMRN_PMC5, val); + break; default: break; } @@ -134,6 +154,14 @@ static void init_pmc_stop(int ctr) mtpmr(PMRN_PMLCA3, pmlca); mtpmr(PMRN_PMLCB3, pmlcb); break; + case 4: + mtpmr(PMRN_PMLCA4, pmlca); + mtpmr(PMRN_PMLCB4, pmlcb); + break; + case 5: + mtpmr(PMRN_PMLCA5, pmlca); + mtpmr(PMRN_PMLCB5, pmlcb); + break; default: panic("Bad ctr number!\n"); } @@ -228,20 +256,6 @@ static void pmc_stop_ctrs(void) mtpmr(PMRN_PMGC0, pmgc0); } -static void dump_pmcs(void) -{ - printk("pmgc0: %x\n", mfpmr(PMRN_PMGC0)); - printk("pmc\t\tpmlca\t\tpmlcb\n"); - printk("%8x\t%8x\t%8x\n", mfpmr(PMRN_PMC0), - mfpmr(PMRN_PMLCA0), mfpmr(PMRN_PMLCB0)); - printk("%8x\t%8x\t%8x\n", mfpmr(PMRN_PMC1), - mfpmr(PMRN_PMLCA1), mfpmr(PMRN_PMLCB1)); - printk("%8x\t%8x\t%8x\n", mfpmr(PMRN_PMC2), - mfpmr(PMRN_PMLCA2), mfpmr(PMRN_PMLCB2)); - printk("%8x\t%8x\t%8x\n", mfpmr(PMRN_PMC3), - mfpmr(PMRN_PMLCA3), mfpmr(PMRN_PMLCB3)); -} - static int fsl_emb_cpu_setup(struct op_counter_config *ctr) { int i; @@ -335,9 +349,6 @@ static void fsl_emb_handle_interrupt(struct pt_regs *regs, int val; int i; - /* set the PMM bit (see comment below) */ - mtmsr(mfmsr() | MSR_PMM); - pc = regs->nip; is_kernel = is_kernel_addr(pc); @@ -354,9 +365,13 @@ static void fsl_emb_handle_interrupt(struct pt_regs *regs, } /* The freeze bit was set by the interrupt. */ - /* Clear the freeze bit, and reenable the interrupt. - * The counters won't actually start until the rfi clears - * the PMM bit */ + /* Clear the freeze bit, and reenable the interrupt. The + * counters won't actually start until the rfi clears the PMM + * bit. The PMM bit should not be set until after the interrupt + * is cleared to avoid it getting lost in some hypervisor + * environments. + */ + mtmsr(mfmsr() | MSR_PMM); pmc_start_ctrs(1); } diff --git a/arch/powerpc/oprofile/op_model_pa6t.c b/arch/powerpc/oprofile/op_model_pa6t.c index c40de461fd4..a114a7c22d4 100644 --- a/arch/powerpc/oprofile/op_model_pa6t.c +++ b/arch/powerpc/oprofile/op_model_pa6t.c @@ -22,7 +22,6 @@ */ #include <linux/oprofile.h> -#include <linux/init.h> #include <linux/smp.h> #include <linux/percpu.h> #include <asm/processor.h> @@ -132,7 +131,7 @@ static int pa6t_reg_setup(struct op_counter_config *ctr, for (pmc = 0; pmc < cur_cpu_spec->num_pmcs; pmc++) { /* counters are 40 bit. Move to cputable at some point? */ reset_value[pmc] = (0x1UL << 39) - ctr[pmc].count; - pr_debug("reset_value for pmc%u inited to 0x%lx\n", + pr_debug("reset_value for pmc%u inited to 0x%llx\n", pmc, reset_value[pmc]); } @@ -177,7 +176,7 @@ static int pa6t_start(struct op_counter_config *ctr) oprofile_running = 1; - pr_debug("start on cpu %d, mmcr0 %lx\n", smp_processor_id(), mmcr0); + pr_debug("start on cpu %d, mmcr0 %llx\n", smp_processor_id(), mmcr0); return 0; } @@ -193,7 +192,7 @@ static void pa6t_stop(void) oprofile_running = 0; - pr_debug("stop on cpu %d, mmcr0 %lx\n", smp_processor_id(), mmcr0); + pr_debug("stop on cpu %d, mmcr0 %llx\n", smp_processor_id(), mmcr0); } /* handle the perfmon overflow vector */ diff --git a/arch/powerpc/oprofile/op_model_power4.c b/arch/powerpc/oprofile/op_model_power4.c index 446a8bbb847..962fe7b3e3f 100644 --- a/arch/powerpc/oprofile/op_model_power4.c +++ b/arch/powerpc/oprofile/op_model_power4.c @@ -10,11 +10,9 @@ */ #include <linux/oprofile.h> -#include <linux/init.h> #include <linux/smp.h> #include <asm/firmware.h> #include <asm/ptrace.h> -#include <asm/system.h> #include <asm/processor.h> #include <asm/cputable.h> #include <asm/rtas.h> @@ -22,15 +20,78 @@ #include <asm/reg.h> #define dbg(args...) +#define OPROFILE_PM_PMCSEL_MSK 0xffULL +#define OPROFILE_PM_UNIT_SHIFT 60 +#define OPROFILE_PM_UNIT_MSK 0xfULL +#define OPROFILE_MAX_PMC_NUM 3 +#define OPROFILE_PMSEL_FIELD_WIDTH 8 +#define OPROFILE_UNIT_FIELD_WIDTH 4 +#define MMCRA_SIAR_VALID_MASK 0x10000000ULL static unsigned long reset_value[OP_MAX_COUNTER]; static int oprofile_running; +static int use_slot_nums; /* mmcr values are set in power4_reg_setup, used in power4_cpu_setup */ static u32 mmcr0_val; static u64 mmcr1_val; static u64 mmcra_val; +static u32 cntr_marked_events; + +static int power7_marked_instr_event(u64 mmcr1) +{ + u64 psel, unit; + int pmc, cntr_marked_events = 0; + + /* Given the MMCR1 value, look at the field for each counter to + * determine if it is a marked event. Code based on the function + * power7_marked_instr_event() in file arch/powerpc/perf/power7-pmu.c. + */ + for (pmc = 0; pmc < 4; pmc++) { + psel = mmcr1 & (OPROFILE_PM_PMCSEL_MSK + << (OPROFILE_MAX_PMC_NUM - pmc) + * OPROFILE_PMSEL_FIELD_WIDTH); + psel = (psel >> ((OPROFILE_MAX_PMC_NUM - pmc) + * OPROFILE_PMSEL_FIELD_WIDTH)) & ~1ULL; + unit = mmcr1 & (OPROFILE_PM_UNIT_MSK + << (OPROFILE_PM_UNIT_SHIFT + - (pmc * OPROFILE_PMSEL_FIELD_WIDTH ))); + unit = unit >> (OPROFILE_PM_UNIT_SHIFT + - (pmc * OPROFILE_PMSEL_FIELD_WIDTH)); + + switch (psel >> 4) { + case 2: + cntr_marked_events |= (pmc == 1 || pmc == 3) << pmc; + break; + case 3: + if (psel == 0x3c) { + cntr_marked_events |= (pmc == 0) << pmc; + break; + } + + if (psel == 0x3e) { + cntr_marked_events |= (pmc != 1) << pmc; + break; + } + + cntr_marked_events |= 1 << pmc; + break; + case 4: + case 5: + cntr_marked_events |= (unit == 0xd) << pmc; + break; + case 6: + if (psel == 0x64) + cntr_marked_events |= (pmc >= 2) << pmc; + break; + case 8: + cntr_marked_events |= (unit == 0xd) << pmc; + break; + } + } + return cntr_marked_events; +} static int power4_reg_setup(struct op_counter_config *ctr, struct op_system_config *sys, @@ -47,6 +108,23 @@ static int power4_reg_setup(struct op_counter_config *ctr, mmcr1_val = sys->mmcr1; mmcra_val = sys->mmcra; + /* Power 7+ and newer architectures: + * Determine which counter events in the group (the group of events is + * specified by the bit settings in the MMCR1 register) are marked + * events for use in the interrupt handler. Do the calculation once + * before OProfile starts. Information is used in the interrupt + * handler. Starting with Power 7+ we only record the sample for + * marked events if the SIAR valid bit is set. For non marked events + * the sample is always recorded. + */ + if (pvr_version_is(PVR_POWER7p)) + cntr_marked_events = power7_marked_instr_event(mmcr1_val); + else + cntr_marked_events = 0; /* For older processors, set the bit map + * to zero so the sample will always be + * be recorded. + */ + for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) reset_value[i] = 0x80000000UL - ctr[i].count; @@ -61,10 +139,16 @@ static int power4_reg_setup(struct op_counter_config *ctr, else mmcr0_val |= MMCR0_PROBLEM_DISABLE; + if (pvr_version_is(PVR_POWER4) || pvr_version_is(PVR_POWER4p) || + pvr_version_is(PVR_970) || pvr_version_is(PVR_970FX) || + pvr_version_is(PVR_970MP) || pvr_version_is(PVR_970GX) || + pvr_version_is(PVR_POWER5) || pvr_version_is(PVR_POWER5p)) + use_slot_nums = 1; + return 0; } -extern void ppc64_enable_pmcs(void); +extern void ppc_enable_pmcs(void); /* * Older CPUs require the MMCRA sample bit to be always set, but newer @@ -78,9 +162,9 @@ extern void ppc64_enable_pmcs(void); */ static inline int mmcra_must_set_sample(void) { - if (__is_processor(PV_POWER4) || __is_processor(PV_POWER4p) || - __is_processor(PV_970) || __is_processor(PV_970FX) || - __is_processor(PV_970MP) || __is_processor(PV_970GX)) + if (pvr_version_is(PVR_POWER4) || pvr_version_is(PVR_POWER4p) || + pvr_version_is(PVR_970) || pvr_version_is(PVR_970FX) || + pvr_version_is(PVR_970MP) || pvr_version_is(PVR_970GX)) return 1; return 0; @@ -91,7 +175,7 @@ static int power4_cpu_setup(struct op_counter_config *ctr) unsigned int mmcr0 = mmcr0_val; unsigned long mmcra = mmcra_val; - ppc64_enable_pmcs(); + ppc_enable_pmcs(); /* set the freeze bit */ mmcr0 |= MMCR0_FC; @@ -200,13 +284,13 @@ static unsigned long get_pc(struct pt_regs *regs) unsigned long mmcra; unsigned long slot; - /* Cant do much about it */ + /* Can't do much about it */ if (!cur_cpu_spec->oprofile_mmcra_sihv) return pc; mmcra = mfspr(SPRN_MMCRA); - if (mmcra & MMCRA_SAMPLE_ENABLE) { + if (use_slot_nums && (mmcra & MMCRA_SAMPLE_ENABLE)) { slot = ((mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT); if (slot > 1) pc += 4 * (slot - 1); @@ -254,6 +338,28 @@ static int get_kernel(unsigned long pc, unsigned long mmcra) return is_kernel; } +static bool pmc_overflow(unsigned long val) +{ + if ((int)val < 0) + return true; + + /* + * Events on POWER7 can roll back if a speculative event doesn't + * eventually complete. Unfortunately in some rare cases they will + * raise a performance monitor exception. We need to catch this to + * ensure we reset the PMC. In all cases the PMC will be 256 or less + * cycles from overflow. + * + * We only do this if the first pass fails to find any overflowing + * PMCs because a user might set a period of less than 256 and we + * don't want to mistakenly reset them. + */ + if (pvr_version_is(PVR_POWER7) && ((0x80000000 - val) <= 256)) + return true; + + return false; +} + static void power4_handle_interrupt(struct pt_regs *regs, struct op_counter_config *ctr) { @@ -263,6 +369,7 @@ static void power4_handle_interrupt(struct pt_regs *regs, int i; unsigned int mmcr0; unsigned long mmcra; + bool siar_valid = false; mmcra = mfspr(SPRN_MMCRA); @@ -272,11 +379,29 @@ static void power4_handle_interrupt(struct pt_regs *regs, /* set the PMM bit (see comment below) */ mtmsrd(mfmsr() | MSR_PMM); + /* Check that the SIAR valid bit in MMCRA is set to 1. */ + if ((mmcra & MMCRA_SIAR_VALID_MASK) == MMCRA_SIAR_VALID_MASK) + siar_valid = true; + for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) { val = classic_ctr_read(i); - if (val < 0) { + if (pmc_overflow(val)) { if (oprofile_running && ctr[i].enabled) { - oprofile_add_ext_sample(pc, regs, i, is_kernel); + /* Power 7+ and newer architectures: + * If the event is a marked event, then only + * save the sample if the SIAR valid bit is + * set. If the event is not marked, then + * always save the sample. + * Note, the Sample enable bit in the MMCRA + * register must be set to 1 if the group + * contains a marked event. + */ + if ((siar_valid && + (cntr_marked_events & (1 << i))) + || !(cntr_marked_events & (1 << i))) + oprofile_add_ext_sample(pc, regs, i, + is_kernel); + classic_ctr_write(i, reset_value[i]); } else { classic_ctr_write(i, 0); diff --git a/arch/powerpc/oprofile/op_model_rs64.c b/arch/powerpc/oprofile/op_model_rs64.c index a20afe45d93..7e5b8ed3a1b 100644 --- a/arch/powerpc/oprofile/op_model_rs64.c +++ b/arch/powerpc/oprofile/op_model_rs64.c @@ -8,10 +8,8 @@ */ #include <linux/oprofile.h> -#include <linux/init.h> #include <linux/smp.h> #include <asm/ptrace.h> -#include <asm/system.h> #include <asm/processor.h> #include <asm/cputable.h> #include <asm/oprofile_impl.h> |