msm: ADSPRPC: add profiling counters

Added profiling counters for measuring RPC performance and
by default profiling is disabled.

Change-Id: I6a79dde638e751998ac37948f826639a161ff4a2
Acked-by: Viswanatham Paduchuri <vpaduchu@qti.qualcomm.com>
Signed-off-by: Sathish Ambley <sathishambley@codeaurora.org>

3 files changed, 167 insertions, 3 deletions

diff --git a/drivers/char/adsprpc.c b/drivers/char/adsprpc.c
index 896bd6ec90f6..82caf98491fa 100644
--- a/drivers/char/adsprpc.c
+++ b/drivers/char/adsprpc.c
@@ -70,6 +70,24 @@
 #define FASTRPC_LINK_CONNECTED    (0x3)
 #define FASTRPC_LINK_DISCONNECTING (0x7)
 
+#define PERF_KEYS "count:flush:map:copy:glink:getargs:putargs:invalidate:invoke"
+#define FASTRPC_STATIC_HANDLE_LISTENER (3)
+#define FASTRPC_STATIC_HANDLE_MAX (20)
+
+#define PERF_END (void)0
+
+#define PERF(enb, cnt, ff) \
+	{\
+		struct timespec startT = {0};\
+		if (enb) {\
+			getnstimeofday(&startT);\
+		} \
+		ff ;\
+		if (enb) {\
+			cnt += getnstimediff(&startT);\
+		} \
+	}
+
 static int fastrpc_glink_open(int cid);
 static void fastrpc_glink_close(void *chan, int cid);
 
@@ -237,6 +255,18 @@ struct fastrpc_mmap {
 	uintptr_t attr;
 };
 
+struct fastrpc_perf {
+	int64_t count;
+	int64_t flush;
+	int64_t map;
+	int64_t copy;
+	int64_t link;
+	int64_t getargs;
+	int64_t putargs;
+	int64_t invargs;
+	int64_t invoke;
+};
+
 struct fastrpc_file {
 	struct hlist_node hn;
 	spinlock_t hlock;
@@ -246,11 +276,13 @@ struct fastrpc_file {
 	struct fastrpc_session_ctx *sctx;
 	struct fastrpc_session_ctx *secsctx;
 	uint32_t mode;
+	uint32_t profile;
 	int tgid;
 	int cid;
 	int ssrcount;
 	int pd;
 	struct fastrpc_apps *apps;
+	struct fastrpc_perf perf;
 };
 
 static struct fastrpc_apps gfa;
@@ -286,6 +318,17 @@ static struct fastrpc_channel_ctx gcinfo[NUM_CHANNELS] = {
 	},
 };
 
+static inline int64_t getnstimediff(struct timespec *start)
+{
+	int64_t ns;
+	struct timespec ts, b;
+
+	getnstimeofday(&ts);
+	b = timespec_sub(ts, *start);
+	ns = timespec_to_ns(&b);
+	return ns;
+}
+
 static void fastrpc_buf_free(struct fastrpc_buf *buf, int cache)
 {
 	struct fastrpc_file *fl = buf == 0 ? 0 : buf->fl;
@@ -1075,6 +1118,7 @@ static int get_args(uint32_t kernel, struct smq_invoke_ctx *ctx)
 		ipage++;
 	}
 	/* map ion buffers */
+	PERF(ctx->fl->profile, ctx->fl->perf.map,
 	for (i = 0; i < inbufs + outbufs; ++i) {
 		struct fastrpc_mmap *map = ctx->maps[i];
 		uint64_t buf = ptr_to_uint64(lpra[i].buf.pv);
@@ -1110,7 +1154,10 @@ static int get_args(uint32_t kernel, struct smq_invoke_ctx *ctx)
 		}
 		rpra[i].buf.pv = buf;
 	}
+	PERF_END);
+
 	/* copy non ion buffers */
+	PERF(ctx->fl->profile, ctx->fl->perf.copy,
 	rlen = copylen - metalen;
 	for (oix = 0; oix < inbufs + outbufs; ++oix) {
 		int i = ctx->overps[oix]->raix;
@@ -1146,7 +1193,9 @@ static int get_args(uint32_t kernel, struct smq_invoke_ctx *ctx)
 		args = args + mlen;
 		rlen -= mlen;
 	}
+	PERF_END);
 
+	PERF(ctx->fl->profile, ctx->fl->perf.flush,
 	for (oix = 0; oix < inbufs + outbufs; ++oix) {
 		int i = ctx->overps[oix]->raix;
 		struct fastrpc_mmap *map = ctx->maps[i];
@@ -1163,14 +1212,20 @@ static int get_args(uint32_t kernel, struct smq_invoke_ctx *ctx)
 			dmac_flush_range(uint64_to_ptr(rpra[i].buf.pv),
 			uint64_to_ptr(rpra[i].buf.pv + rpra[i].buf.len));
 	}
+	PERF_END);
+
 	inh = inbufs + outbufs;
 	for (i = 0; i < REMOTE_SCALARS_INHANDLES(sc); i++) {
 		rpra[inh + i].buf.pv = ptr_to_uint64(ctx->lpra[inh + i].buf.pv);
 		rpra[inh + i].buf.len = ctx->lpra[inh + i].buf.len;
 		rpra[inh + i].h = ctx->lpra[inh + i].h;
 	}
-	if (!ctx->fl->sctx->smmu.coherent)
+
+	if (!ctx->fl->sctx->smmu.coherent) {
+		PERF(ctx->fl->profile, ctx->fl->perf.flush,
 		dmac_flush_range((char *)rpra, (char *)rpra + ctx->used);
+		PERF_END);
+	}
  bail:
 	return err;
 }
@@ -1390,7 +1445,10 @@ static int fastrpc_internal_invoke(struct fastrpc_file *fl, uint32_t mode,
 	int cid = fl->cid;
 	int interrupted = 0;
 	int err = 0;
+	struct timespec invoket;
 
+	if (fl->profile)
+		getnstimeofday(&invoket);
 	if (!kernel) {
 		VERIFY(err, 0 == context_restore_interrupted(fl, inv,
 								&ctx));
@@ -1409,19 +1467,30 @@ static int fastrpc_internal_invoke(struct fastrpc_file *fl, uint32_t mode,
 		goto bail;
 
 	if (REMOTE_SCALARS_LENGTH(ctx->sc)) {
+		PERF(fl->profile, fl->perf.getargs,
 		VERIFY(err, 0 == get_args(kernel, ctx));
+		PERF_END);
 		if (err)
 			goto bail;
 	}
 
+	PERF(fl->profile, fl->perf.invargs,
 	inv_args_pre(ctx);
 	if (mode == FASTRPC_MODE_SERIAL)
 		inv_args(ctx);
+	PERF_END);
+
+	PERF(fl->profile, fl->perf.link,
 	VERIFY(err, 0 == fastrpc_invoke_send(ctx, kernel, invoke->handle));
+	PERF_END);
+
 	if (err)
 		goto bail;
+
+	PERF(fl->profile, fl->perf.invargs,
 	if (mode == FASTRPC_MODE_PARALLEL)
 		inv_args(ctx);
+	PERF_END);
  wait:
 	if (kernel)
 		wait_for_completion(&ctx->work);
@@ -1434,7 +1503,10 @@ static int fastrpc_internal_invoke(struct fastrpc_file *fl, uint32_t mode,
 	VERIFY(err, 0 == (err = ctx->retval));
 	if (err)
 		goto bail;
+
+	PERF(fl->profile, fl->perf.putargs,
 	VERIFY(err, 0 == put_args(kernel, ctx, invoke->pra));
+	PERF_END);
 	if (err)
 		goto bail;
  bail:
@@ -1444,6 +1516,14 @@ static int fastrpc_internal_invoke(struct fastrpc_file *fl, uint32_t mode,
 		context_free(ctx);
 	if (fl->ssrcount != fl->apps->channel[cid].ssrcount)
 		err = ECONNRESET;
+
+	if (fl->profile && !interrupted) {
+		if (invoke->handle != FASTRPC_STATIC_HANDLE_LISTENER)
+			fl->perf.invoke += getnstimediff(&invoket);
+		if (!(invoke->handle >= 0 &&
+			invoke->handle <= FASTRPC_STATIC_HANDLE_MAX))
+			fl->perf.count++;
+	}
 	return err;
 }
 
@@ -2086,6 +2166,8 @@ static int fastrpc_device_open(struct inode *inode, struct file *filp)
 	fl->tgid = current->tgid;
 	fl->apps = me;
 	fl->cid = cid;
+	memset(&fl->perf, 0, sizeof(fl->perf));
+
 	VERIFY(err, !fastrpc_session_alloc_locked(&me->channel[cid], 0,
 							&fl->sctx));
 	if (err)
@@ -2156,6 +2238,7 @@ static long fastrpc_device_ioctl(struct file *file, unsigned int ioctl_num,
 		struct fastrpc_ioctl_mmap mmap;
 		struct fastrpc_ioctl_munmap munmap;
 		struct fastrpc_ioctl_init init;
+		struct fastrpc_ioctl_perf perf;
 	} p;
 	void *param = (char *)ioctl_param;
 	struct fastrpc_file *fl = (struct fastrpc_file *)file->private_data;
@@ -2211,11 +2294,36 @@ static long fastrpc_device_ioctl(struct file *file, unsigned int ioctl_num,
 		case FASTRPC_MODE_SERIAL:
 			fl->mode = (uint32_t)ioctl_param;
 			break;
+		case FASTRPC_MODE_PROFILE:
+			fl->profile = (uint32_t)ioctl_param;
+			break;
 		default:
 			err = -ENOTTY;
 			break;
 		}
 		break;
+	case FASTRPC_IOCTL_GETPERF:
+		VERIFY(err, 0 == copy_from_user(&p.perf,
+					param, sizeof(p.perf)));
+		if (err)
+			goto bail;
+		p.perf.numkeys = sizeof(struct fastrpc_perf)/sizeof(int64_t);
+		if (p.perf.keys) {
+			char *keys = PERF_KEYS;
+
+			VERIFY(err, 0 == copy_to_user((char *)p.perf.keys,
+						 keys, strlen(keys)+1));
+			if (err)
+				goto bail;
+		}
+		if (p.perf.data) {
+			VERIFY(err, 0 == copy_to_user((int64_t *)p.perf.data,
+						 &fl->perf, sizeof(fl->perf)));
+		}
+		VERIFY(err, 0 == copy_to_user(param, &p.perf, sizeof(p.perf)));
+		if (err)
+			goto bail;
+		break;
 	case FASTRPC_IOCTL_GETINFO:
 		VERIFY(err, 0 == (err = fastrpc_get_info(fl, &info)));
 		if (err)
diff --git a/drivers/char/adsprpc_compat.c b/drivers/char/adsprpc_compat.c
index 1e5649a8d4c4..a224ba7ed3e4 100644
--- a/drivers/char/adsprpc_compat.c
+++ b/drivers/char/adsprpc_compat.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2014-2016, The Linux Foundation. All rights reserved.
+ * Copyright (c) 2014-2017, The Linux Foundation. All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 and
@@ -32,6 +32,8 @@
 		_IOWR('R', 6, struct compat_fastrpc_ioctl_init)
 #define COMPAT_FASTRPC_IOCTL_INVOKE_ATTRS \
 		_IOWR('R', 7, struct compat_fastrpc_ioctl_invoke_attrs)
+#define COMPAT_FASTRPC_IOCTL_GETPERF \
+		_IOWR('R', 9, struct compat_fastrpc_ioctl_perf)
 
 struct compat_remote_buf {
 	compat_uptr_t pv;	/* buffer pointer */
@@ -83,6 +85,12 @@ struct compat_fastrpc_ioctl_init {
 	compat_int_t memfd;	/* ION fd for the mem */
 };
 
+struct compat_fastrpc_ioctl_perf {	/* kernel performance data */
+	compat_uptr_t  data;
+	compat_int_t numkeys;
+	compat_uptr_t keys;
+};
+
 static int compat_get_fastrpc_ioctl_invoke(
 			struct compat_fastrpc_ioctl_invoke_attrs __user *inv32,
 			struct fastrpc_ioctl_invoke_attrs __user **inva,
@@ -222,6 +230,21 @@ static int compat_get_fastrpc_ioctl_munmap(
 	return err;
 }
 
+static int compat_get_fastrpc_ioctl_perf(
+			struct compat_fastrpc_ioctl_perf __user *perf32,
+			struct fastrpc_ioctl_perf __user *perf)
+{
+	compat_uptr_t p;
+	int err;
+
+	err = get_user(p, &perf32->data);
+	err |= put_user(p, &perf->data);
+	err |= get_user(p, &perf32->keys);
+	err |= put_user(p, &perf->keys);
+
+	return err;
+}
+
 static int compat_get_fastrpc_ioctl_init(
 			struct compat_fastrpc_ioctl_init __user *init32,
 			struct fastrpc_ioctl_init __user *init)
@@ -356,6 +379,30 @@ long compat_fastrpc_device_ioctl(struct file *filp, unsigned int cmd,
 	case FASTRPC_IOCTL_SETMODE:
 		return filp->f_op->unlocked_ioctl(filp, cmd,
 						(unsigned long)compat_ptr(arg));
+	case COMPAT_FASTRPC_IOCTL_GETPERF:
+	{
+		struct compat_fastrpc_ioctl_perf __user *perf32;
+		struct fastrpc_ioctl_perf *perf;
+		compat_uint_t u;
+		long ret;
+
+		perf32 = compat_ptr(arg);
+		VERIFY(err, NULL != (perf = compat_alloc_user_space(
+							sizeof(*perf))));
+		if (err)
+			return -EFAULT;
+		VERIFY(err, 0 == compat_get_fastrpc_ioctl_perf(perf32,
+							perf));
+		if (err)
+			return err;
+		ret = filp->f_op->unlocked_ioctl(filp, FASTRPC_IOCTL_GETPERF,
+							(unsigned long)perf);
+		if (ret)
+			return ret;
+		err = get_user(u, &perf->numkeys);
+		err |= put_user(u, &perf32->numkeys);
+		return err;
+	}
 	default:
 		return -ENOIOCTLCMD;
 	}
diff --git a/drivers/char/adsprpc_shared.h b/drivers/char/adsprpc_shared.h
index 4fc9396bfd3a..9b24e2da489f 100644
--- a/drivers/char/adsprpc_shared.h
+++ b/drivers/char/adsprpc_shared.h
@@ -25,6 +25,7 @@
 #define FASTRPC_IOCTL_INVOKE_ATTRS \
 				_IOWR('R', 7, struct fastrpc_ioctl_invoke_attrs)
 #define FASTRPC_IOCTL_GETINFO	_IOWR('R', 8, uint32_t)
+#define FASTRPC_IOCTL_GETPERF	_IOWR('R', 9, struct fastrpc_ioctl_perf)
 
 #define FASTRPC_GLINK_GUID "fastrpcglink-apps-dsp"
 #define FASTRPC_SMD_GUID "fastrpcsmd-apps-dsp"
@@ -45,6 +46,9 @@
 /* Driver should operate in serial mode with the co-processor */
 #define FASTRPC_MODE_SERIAL      1
 
+/* Driver should operate in profile mode with the co-processor */
+#define FASTRPC_MODE_PROFILE     2
+
 /* INIT a new process or attach to guestos */
 #define FASTRPC_INIT_ATTACH      0
 #define FASTRPC_INIT_CREATE      1
@@ -159,7 +163,6 @@ struct fastrpc_ioctl_munmap {
 	ssize_t size;		/* size */
 };
 
-
 struct fastrpc_ioctl_mmap {
 	int fd;				/* ion fd */
 	uint32_t flags;			/* flags for dsp to map with */
@@ -168,6 +171,12 @@ struct fastrpc_ioctl_mmap {
 	uintptr_t vaddrout;		/* dsps virtual address */
 };
 
+struct fastrpc_ioctl_perf {			/* kernel performance data */
+	uintptr_t __user data;
+	uint32_t numkeys;
+	uintptr_t __user keys;
+};
+
 struct smq_null_invoke {
 	uint64_t ctx;			/* invoke caller context */
 	uint32_t handle;	    /* handle to invoke */