misc: tegra-profiler: send mapped memory regions

Tegra Profiler: send information about mapped
memory regions (on demand).

Bug 1394970

Change-Id: Idf9a3c5d2d9fac0fd622f99bd226632c810a8ab6
Signed-off-by: Igor Nabirushkin <inabirushkin@nvidia.com>
Reviewed-on: http://git-master/r/324093
(cherry picked from commit 9b684c4808301c89c8d479d562afc5830ff3835a)
Reviewed-on: http://git-master/r/340018
Reviewed-by: Harry Hong <hhong@nvidia.com>
Tested-by: Harry Hong <hhong@nvidia.com>

6 files changed, 146 insertions, 60 deletions

diff --git a/drivers/misc/tegra-profiler/hrt.c b/drivers/misc/tegra-profiler/hrt.c
index 8bc30727d9bb..fef37de8c288 100644
--- a/drivers/misc/tegra-profiler/hrt.c
+++ b/drivers/misc/tegra-profiler/hrt.c
@@ -222,14 +222,6 @@ static int get_sample_data(struct event_data *event,
 	return 0;
 }
 
-static char *get_mmap_data(struct pt_regs *regs,
-			   struct quadd_mmap_data *sample,
-			   unsigned int *extra_length)
-{
-	struct quadd_cpu_context *cpu_ctx = this_cpu_ptr(hrt.cpu_ctx);
-	return quadd_get_mmap(cpu_ctx, regs, sample, extra_length);
-}
-
 static void read_source(struct quadd_event_source_interface *source,
 			struct pt_regs *regs, pid_t pid)
 {
@@ -306,27 +298,13 @@ static void read_source(struct quadd_event_source_interface *source,
 
 static void read_all_sources(struct pt_regs *regs, pid_t pid)
 {
-	struct quadd_record_data record_data;
 	struct quadd_ctx *ctx = hrt.quadd_ctx;
-	unsigned int extra_length;
-	char *extra_data;
+	struct quadd_cpu_context *cpu_ctx = this_cpu_ptr(hrt.cpu_ctx);
 
 	if (!regs)
 		return;
 
-	extra_data = get_mmap_data(regs, &record_data.mmap, &extra_length);
-	if (extra_data && extra_length > 0) {
-		record_data.magic = QUADD_RECORD_MAGIC;
-		record_data.record_type = QUADD_RECORD_TYPE_MMAP;
-		record_data.cpu_mode = QUADD_CPU_MODE_USER;
-
-		record_data.mmap.filename_length = extra_length;
-		record_data.mmap.pid = pid > 0 ? pid : ctx->param.pids[0];
-
-		quadd_put_sample(&record_data, extra_data, extra_length);
-	} else {
-		record_data.mmap.filename_length = 0;
-	}
+	quadd_get_mmap_object(cpu_ctx, regs, pid);
 
 	if (ctx->pmu && ctx->pmu_info.active)
 		read_source(ctx->pmu, regs, pid);
@@ -473,9 +451,13 @@ static void reset_cpu_ctx(void)
 
 int quadd_hrt_start(void)
 {
+	int err;
 	u64 period;
 	long freq;
+	unsigned int extra;
 	struct quadd_ctx *ctx = hrt.quadd_ctx;
+	struct quadd_cpu_context *cpu_ctx = this_cpu_ptr(hrt.cpu_ctx);
+	struct quadd_parameters *param = &ctx->param;
 
 	freq = ctx->param.freq;
 	freq = max_t(long, QUADD_HRT_MIN_FREQ, freq);
@@ -493,6 +475,16 @@ int quadd_hrt_start(void)
 
 	put_header();
 
+	extra = param->reserved[QUADD_PARAM_IDX_EXTRA];
+
+	if (extra & QUADD_PARAM_IDX_EXTRA_GET_MMAP) {
+		err = quadd_get_current_mmap(cpu_ctx, param->pids[0]);
+		if (err) {
+			pr_err("error: quadd_get_current_mmap\n");
+			return err;
+		}
+	}
+
 	if (ctx->pl310)
 		ctx->pl310->start();
 
diff --git a/drivers/misc/tegra-profiler/main.c b/drivers/misc/tegra-profiler/main.c
index f1ed75cf5833..2b9b8ad1b59b 100644
--- a/drivers/misc/tegra-profiler/main.c
+++ b/drivers/misc/tegra-profiler/main.c
@@ -146,6 +146,9 @@ static int set_parameters(struct quadd_parameters *param, uid_t *debug_app_uid)
 	ctx.param.power_rate_freq = param->power_rate_freq;
 	ctx.param.debug_samples = param->debug_samples;
 
+	for (i = 0; i < QM_ARRAY_SIZE(param->reserved); i++)
+		ctx.param.reserved[i] = param->reserved[i];
+
 	/* Currently only one process */
 	if (param->nr_pids != 1)
 		return -EINVAL;
@@ -360,6 +363,8 @@ static void get_capabilities(struct quadd_comm_cap *cap)
 	cap->blocked_read = 1;
 
 	extra |= QUADD_COMM_CAP_EXTRA_BT_KERNEL_CTX;
+	extra |= QUADD_COMM_CAP_EXTRA_GET_MMAP;
+
 	cap->reserved[QUADD_COMM_CAP_IDX_EXTRA] = extra;
 }
 
diff --git a/drivers/misc/tegra-profiler/mmap.c b/drivers/misc/tegra-profiler/mmap.c
index 2b73e5cd9c0a..cf6f408a0d80 100644
--- a/drivers/misc/tegra-profiler/mmap.c
+++ b/drivers/misc/tegra-profiler/mmap.c
@@ -101,25 +101,61 @@ static int check_hash(u32 key)
 	return 1;
 }
 
-char *quadd_get_mmap(struct quadd_cpu_context *cpu_ctx,
-		     struct pt_regs *regs, struct quadd_mmap_data *sample,
-		     unsigned int *extra_length)
+static int find_file(const char *file_name, unsigned long addr,
+		     unsigned long len)
 {
 	u32 crc;
+	size_t length;
+
+	if (!file_name)
+		return 0;
+
+	length = strlen(file_name);
+
+	crc = crc32_le(~0, file_name, length);
+	crc = crc32_le(crc, (unsigned char *)&addr,
+		       sizeof(addr));
+	crc = crc32_le(crc, (unsigned char *)&len,
+		       sizeof(len));
+
+	return check_hash(crc);
+}
+
+static void
+put_mmap_sample(struct quadd_mmap_data *s, char *extra_data,
+		size_t extra_length)
+{
+	struct quadd_record_data r;
+
+	r.magic = QUADD_RECORD_MAGIC;
+	r.record_type = QUADD_RECORD_TYPE_MMAP;
+	r.cpu_mode = QUADD_CPU_MODE_USER;
+
+	memcpy(&r.mmap, s, sizeof(*s));
+	r.mmap.filename_length = extra_length;
+
+	pr_debug("MMAP: pid: %d, file_name: '%s', addr: %#x, length: %u",
+		 s->pid, extra_data, s->addr, extra_length);
+
+	quadd_put_sample(&r, extra_data, extra_length);
+}
+
+void quadd_get_mmap_object(struct quadd_cpu_context *cpu_ctx,
+			   struct pt_regs *regs, pid_t pid)
+{
 	unsigned long ip;
-	int length, length_aligned;
+	size_t length, length_aligned;
 	struct mm_struct *mm = current->mm;
 	struct vm_area_struct *vma;
 	struct file *vm_file;
 	struct path *path;
 	char *file_name = NULL;
+	struct quadd_mmap_data sample;
 	struct quadd_mmap_cpu_ctx *mm_cpu_ctx = this_cpu_ptr(mmap_ctx.cpu_ctx);
 	char *tmp_buf = mm_cpu_ctx->tmp_buf;
 
-	if (!mm) {
-		*extra_length = 0;
-		return NULL;
-	}
+	if (!mm)
+		return;
 
 	ip = instruction_pointer(regs);
 
@@ -138,9 +174,9 @@ char *quadd_get_mmap(struct quadd_cpu_context *cpu_ctx,
 			if (IS_ERR(file_name)) {
 				file_name = NULL;
 			} else {
-				sample->addr = vma->vm_start;
-				sample->len = vma->vm_end - vma->vm_start;
-				sample->pgoff =
+				sample.addr = vma->vm_start;
+				sample.len = vma->vm_end - vma->vm_start;
+				sample.pgoff =
 					(u64)vma->vm_pgoff << PAGE_SHIFT;
 			}
 			break;
@@ -156,38 +192,86 @@ char *quadd_get_mmap(struct quadd_cpu_context *cpu_ctx,
 		if (mod) {
 			file_name = mod->name;
 			if (file_name) {
-				sample->addr = (u32) mod->module_core;
-				sample->len = mod->core_size;
-				sample->pgoff = 0;
+				sample.addr = (u32) mod->module_core;
+				sample.len = mod->core_size;
+				sample.pgoff = 0;
 			}
 		}
 #endif
 	}
 
 	if (file_name) {
-		length = strlen(file_name);
-		if (length >= PATH_MAX) {
-			*extra_length = 0;
-			return NULL;
-		}
+		if (!find_file(file_name, sample.addr, sample.len)) {
+			length = strlen(file_name) + 1;
+			if (length > PATH_MAX)
+				return;
 
-		crc = crc32_le(~0, file_name, length);
-		crc = crc32_le(crc, (unsigned char *)&sample->addr,
-			       sizeof(sample->addr));
-		crc = crc32_le(crc, (unsigned char *)&sample->len,
-			       sizeof(sample->len));
+			sample.pid = pid;
 
-		if (!check_hash(crc)) {
 			strcpy(cpu_ctx->mmap_filename, file_name);
-			length_aligned = (length + 1 + 7) & (~7);
-			*extra_length = length_aligned;
+			length_aligned = ALIGN(length, 8);
 
-			return cpu_ctx->mmap_filename;
+			put_mmap_sample(&sample, cpu_ctx->mmap_filename,
+					length_aligned);
 		}
 	}
+}
+
+int quadd_get_current_mmap(struct quadd_cpu_context *cpu_ctx, pid_t pid)
+{
+	struct vm_area_struct *vma;
+	struct file *vm_file;
+	struct path *path;
+	char *file_name;
+	struct task_struct *task;
+	struct mm_struct *mm;
+	struct quadd_mmap_data sample;
+	size_t length, length_aligned;
+	struct quadd_mmap_cpu_ctx *mm_cpu_ctx = this_cpu_ptr(mmap_ctx.cpu_ctx);
+	char *tmp_buf = mm_cpu_ctx->tmp_buf;
+
+	rcu_read_lock();
+	task = pid_task(find_vpid(pid), PIDTYPE_PID);
+	rcu_read_unlock();
+	if (!task) {
+		pr_err("Process not found: %u\n", pid);
+		return -ESRCH;
+	}
+	mm = task->mm;
+
+	pr_info("Get mapped memory objects\n");
+
+	for (vma = mm->mmap; vma; vma = vma->vm_next) {
+		vm_file = vma->vm_file;
+		if (!vm_file)
+			continue;
 
-	*extra_length = 0;
-	return NULL;
+		path = &vm_file->f_path;
+
+		file_name = d_path(path, tmp_buf, PATH_MAX);
+		if (IS_ERR(file_name))
+			continue;
+
+		if (!(vma->vm_flags & VM_EXEC))
+			continue;
+
+		length = strlen(file_name) + 1;
+		if (length > PATH_MAX)
+			continue;
+
+		sample.pid = pid;
+		sample.addr = vma->vm_start;
+		sample.len = vma->vm_end - vma->vm_start;
+		sample.pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT;
+
+		if (!find_file(file_name, sample.addr, sample.len)) {
+			strcpy(cpu_ctx->mmap_filename, file_name);
+			length_aligned = ALIGN(length, 8);
+
+			put_mmap_sample(&sample, file_name, length_aligned);
+		}
+	}
+	return 0;
 }
 
 struct quadd_mmap_ctx *quadd_mmap_init(struct quadd_ctx *quadd_ctx)
diff --git a/drivers/misc/tegra-profiler/mmap.h b/drivers/misc/tegra-profiler/mmap.h
index dd750f956df9..4030ed216c43 100644
--- a/drivers/misc/tegra-profiler/mmap.h
+++ b/drivers/misc/tegra-profiler/mmap.h
@@ -39,10 +39,10 @@ struct quadd_mmap_ctx {
 	struct quadd_ctx *quadd_ctx;
 };
 
-char *quadd_get_mmap(struct quadd_cpu_context *cpu_ctx,
-		     struct pt_regs *regs, struct quadd_mmap_data *sample,
-		     unsigned int *extra_length);
+void quadd_get_mmap_object(struct quadd_cpu_context *cpu_ctx,
+			   struct pt_regs *regs, pid_t pid);
 
+int quadd_get_current_mmap(struct quadd_cpu_context *cpu_ctx, pid_t pid);
 
 struct quadd_mmap_ctx *quadd_mmap_init(struct quadd_ctx *quadd_ctx);
 void quadd_mmap_deinit(void);
diff --git a/drivers/misc/tegra-profiler/version.h b/drivers/misc/tegra-profiler/version.h
index 32b074609292..3dddadb66156 100644
--- a/drivers/misc/tegra-profiler/version.h
+++ b/drivers/misc/tegra-profiler/version.h
@@ -18,7 +18,7 @@
 #ifndef __QUADD_VERSION_H
 #define __QUADD_VERSION_H
 
-#define QUADD_MODULE_VERSION		"1.37"
+#define QUADD_MODULE_VERSION		"1.40"
 #define QUADD_MODULE_BRANCH		"Dev"
 
 #endif	/* __QUADD_VERSION_H */
diff --git a/include/linux/tegra_profiler.h b/include/linux/tegra_profiler.h
index ed21f75e9556..b122b97dd521 100644
--- a/include/linux/tegra_profiler.h
+++ b/include/linux/tegra_profiler.h
@@ -20,12 +20,13 @@
 #include <linux/ioctl.h>
 
 #define QUADD_SAMPLES_VERSION	17
-#define QUADD_IO_VERSION	8
+#define QUADD_IO_VERSION	9
 
 #define QUADD_IO_VERSION_DYNAMIC_RB		5
 #define QUADD_IO_VERSION_RB_MAX_FILL_COUNT	6
 #define QUADD_IO_VERSION_MOD_STATE_STATUS_FIELD	7
 #define QUADD_IO_VERSION_BT_KERNEL_CTX		8
+#define QUADD_IO_VERSION_GET_MMAP		9
 
 #define QUADD_SAMPLE_VERSION_THUMB_MODE_FLAG	17
 
@@ -236,9 +237,12 @@ struct quadd_record_data {
 #define QUADD_MAX_PACKAGE_NAME	320
 
 enum {
-	QUADD_PARAM_IDX_SIZE_OF_RB = 0,
+	QUADD_PARAM_IDX_SIZE_OF_RB	= 0,
+	QUADD_PARAM_IDX_EXTRA		= 1,
 };
 
+#define QUADD_PARAM_IDX_EXTRA_GET_MMAP	(1 << 0)
+
 struct quadd_parameters {
 	u32 freq;
 	u32 ma_freq;
@@ -281,6 +285,7 @@ enum {
 };
 
 #define QUADD_COMM_CAP_EXTRA_BT_KERNEL_CTX	(1 << 0)
+#define QUADD_COMM_CAP_EXTRA_GET_MMAP		(1 << 1)
 
 struct quadd_comm_cap {
 	u32	pmu:1,