msm: ipa3: tx optimization - caching tx_wrapper

Caching tx_wrapper for tx DP optimization.
This change modifies tx_wrapper allocation from completely
dynamic,to partially dynamic for the WAN and LAN producer pipes.

A new list of free tx_wrapper is added,this can
dynmically decrease and grow, with a thereshold
which is initially read from the dts, and can be
changed via debugfs on runtime.

Change-Id: I60032a67f9ae87baeb42848cc312bf211c84f4e0
Acked-by: Tal Gelbard <[email protected]>
Acked-by: Chemin Hsieh <[email protected]>
Signed-off-by: Jennifer L. Zenner <[email protected]>

drivers/platform/msm/ipa/ipa_v3/ipa.c

@@ -6562,6 +6562,13 @@ static inline void ipa3_enable_napi_netdev(void)
 	}
 }
 
+static u32 get_tx_wrapper_cache_size(u32 cache_size)
+{
+	if (cache_size <= IPA_TX_WRAPPER_CACHE_MAX_THRESHOLD)
+		return cache_size;
+	return IPA_TX_WRAPPER_CACHE_MAX_THRESHOLD;
+}
+
 /**
  * ipa3_pre_init() - Initialize the IPA Driver.
  * This part contains all initialization which doesn't require IPA HW, such
@@ -6672,6 +6679,8 @@ static int ipa3_pre_init(const struct ipa3_plat_drv_res *resource_p,
 		resource_p->do_ram_collection_on_crash;
 	ipa3_ctx->lan_rx_napi_enable = resource_p->lan_rx_napi_enable;
 	ipa3_ctx->rmnet_ctl_enable = resource_p->rmnet_ctl_enable;
+	ipa3_ctx->tx_wrapper_cache_max_size = get_tx_wrapper_cache_size(
+			resource_p->tx_wrapper_cache_max_size);
 
 	if (resource_p->gsi_fw_file_name) {
 		ipa3_ctx->gsi_fw_file_name =
@@ -7381,6 +7390,22 @@ static int get_ipa_dts_pm_info(struct platform_device *pdev,
 	return 0;
 }
 
+static void get_dts_tx_wrapper_cache_size(struct platform_device *pdev,
+		struct ipa3_plat_drv_res *ipa_drv_res)
+{
+	int result;
+
+	result = of_property_read_u32 (
+		pdev->dev.of_node,
+		"qcom,tx-wrapper-cache-max-size",
+		&ipa_drv_res->tx_wrapper_cache_max_size);
+	if (result)
+		ipa_drv_res->tx_wrapper_cache_max_size = 0;
+
+	IPADBG("tx_wrapper_cache_max_size is set to %d",
+		ipa_drv_res->tx_wrapper_cache_max_size);
+}
+
 static int get_ipa_dts_configuration(struct platform_device *pdev,
 		struct ipa3_plat_drv_res *ipa_drv_res)
 {
@@ -7843,6 +7868,8 @@ static int get_ipa_dts_configuration(struct platform_device *pdev,
 	IPADBG(": secure-debug-check-action = %d\n",
 		   ipa_drv_res->secure_debug_check_action);
 
+	get_dts_tx_wrapper_cache_size(pdev, ipa_drv_res);
+
 	return 0;
 }
 

drivers/platform/msm/ipa/ipa_v3/ipa_debugfs.c

@@ -3042,6 +3042,14 @@ void ipa3_debugfs_init(void)
 		goto fail;
 	}
 
+	file = debugfs_create_u32("tx_wrapper_cache_max_size",
+		IPA_READ_WRITE_MODE,
+		dent, &ipa3_ctx->tx_wrapper_cache_max_size);
+	if (!file) {
+		IPAERR("could not create tx_wrapper_cache_max_size file\n");
+		goto fail;
+	}
+
 	file = debugfs_create_u32("enable_napi_chain", IPA_READ_WRITE_MODE,
 		dent, &ipa3_ctx->enable_napi_chain);
 	if (!file) {

drivers/platform/msm/ipa/ipa_v3/ipa_dp.c

@@ -1,3 +1,4 @@
+
 // SPDX-License-Identifier: GPL-2.0-only
 /*
  * Copyright (c) 2012-2020, The Linux Foundation. All rights reserved.
@@ -140,6 +141,9 @@ static void ipa3_wq_write_done_common(struct ipa3_sys_context *sys,
 {
 	struct ipa3_tx_pkt_wrapper *next_pkt;
 	int i, cnt;
+	void *user1;
+	int user2;
+	void (*callback)(void *user1, int user2);
 
 	if (unlikely(tx_pkt == NULL)) {
 		IPAERR("tx_pkt is NULL\n");
@@ -157,7 +161,6 @@ static void ipa3_wq_write_done_common(struct ipa3_sys_context *sys,
 		next_pkt = list_next_entry(tx_pkt, link);
 		list_del(&tx_pkt->link);
 		sys->len--;
-		spin_unlock_bh(&sys->spinlock);
 		if (!tx_pkt->no_unmap_dma) {
 			if (tx_pkt->type != IPA_DATA_DESC_SKB_PAGED) {
 				dma_unmap_single(ipa3_ctx->pdev,
@@ -171,10 +174,22 @@ static void ipa3_wq_write_done_common(struct ipa3_sys_context *sys,
 					DMA_TO_DEVICE);
 			}
 		}
-		if (tx_pkt->callback)
-			tx_pkt->callback(tx_pkt->user1, tx_pkt->user2);
-
-		kmem_cache_free(ipa3_ctx->tx_pkt_wrapper_cache, tx_pkt);
+		callback = tx_pkt->callback;
+		user1 = tx_pkt->user1;
+		user2 = tx_pkt->user2;
+		if (sys->avail_tx_wrapper >=
+			ipa3_ctx->tx_wrapper_cache_max_size ||
+			sys->ep->client == IPA_CLIENT_APPS_CMD_PROD) {
+			kmem_cache_free(ipa3_ctx->tx_pkt_wrapper_cache,
+				tx_pkt);
+		} else {
+			list_add_tail(&tx_pkt->link,
+				&sys->avail_tx_wrapper_list);
+			sys->avail_tx_wrapper++;
+		}
+		spin_unlock_bh(&sys->spinlock);
+		if (callback)
+			(*callback)(user1, user2);
 		tx_pkt = next_pkt;
 	}
 }
@@ -239,12 +254,21 @@ static void ipa3_send_nop_desc(struct work_struct *work)
 	struct ipa3_tx_pkt_wrapper *tx_pkt;
 
 	IPADBG_LOW("gsi send NOP for ch: %lu\n", sys->ep->gsi_chan_hdl);
-
 	if (atomic_read(&sys->workqueue_flushed))
 		return;
-
-	tx_pkt = kmem_cache_zalloc(ipa3_ctx->tx_pkt_wrapper_cache, GFP_KERNEL);
+	spin_lock_bh(&sys->spinlock);
+	if (!list_empty(&sys->avail_tx_wrapper_list)) {
+		tx_pkt = list_first_entry(&sys->avail_tx_wrapper_list,
+				struct ipa3_tx_pkt_wrapper, link);
+		list_del(&tx_pkt->link);
+		sys->avail_tx_wrapper--;
+		memset(tx_pkt, 0, sizeof(struct ipa3_tx_pkt_wrapper));
+	} else {
+		tx_pkt = kmem_cache_zalloc(ipa3_ctx->tx_pkt_wrapper_cache,
+			GFP_KERNEL);
+	}
 	if (!tx_pkt) {
+		spin_unlock_bh(&sys->spinlock);
 		queue_work(sys->wq, &sys->work);
 		return;
 	}
@@ -253,7 +277,6 @@ static void ipa3_send_nop_desc(struct work_struct *work)
 	tx_pkt->cnt = 1;
 	tx_pkt->no_unmap_dma = true;
 	tx_pkt->sys = sys;
-	spin_lock_bh(&sys->spinlock);
 	if (unlikely(!sys->nop_pending)) {
 		spin_unlock_bh(&sys->spinlock);
 		kmem_cache_free(ipa3_ctx->tx_pkt_wrapper_cache, tx_pkt);
@@ -349,8 +372,18 @@ int ipa3_send(struct ipa3_sys_context *sys,
 	spin_lock_bh(&sys->spinlock);
 
 	for (i = 0; i < num_desc; i++) {
-		tx_pkt = kmem_cache_zalloc(ipa3_ctx->tx_pkt_wrapper_cache,
-					   GFP_ATOMIC);
+		if (!list_empty(&sys->avail_tx_wrapper_list)) {
+			tx_pkt = list_first_entry(&sys->avail_tx_wrapper_list,
+				struct ipa3_tx_pkt_wrapper, link);
+			list_del(&tx_pkt->link);
+			sys->avail_tx_wrapper--;
+
+			memset(tx_pkt, 0, sizeof(struct ipa3_tx_pkt_wrapper));
+		} else {
+			tx_pkt = kmem_cache_zalloc(
+				ipa3_ctx->tx_pkt_wrapper_cache,
+				GFP_ATOMIC);
+		}
 		if (!tx_pkt) {
 			IPAERR("failed to alloc tx wrapper\n");
 			result = -ENOMEM;
@@ -1035,6 +1068,8 @@ int ipa3_setup_sys_pipe(struct ipa_sys_connect_params *sys_in, u32 *clnt_hdl)
 
 		INIT_LIST_HEAD(&ep->sys->head_desc_list);
 		INIT_LIST_HEAD(&ep->sys->rcycl_list);
+		INIT_LIST_HEAD(&ep->sys->avail_tx_wrapper_list);
+		ep->sys->avail_tx_wrapper = 0;
 		spin_lock_init(&ep->sys->spinlock);
 		hrtimer_init(&ep->sys->db_timer, CLOCK_MONOTONIC,
 			HRTIMER_MODE_REL);
@@ -1282,6 +1317,22 @@ fail_gen:
 	return result;
 }
 
+static void delete_avail_tx_wrapper_list(struct ipa3_ep_context *ep)
+{
+	struct ipa3_tx_pkt_wrapper *tx_pkt_iterator = NULL;
+	struct ipa3_tx_pkt_wrapper *tx_pkt_temp = NULL;
+
+	spin_lock_bh(&ep->sys->spinlock);
+	list_for_each_entry_safe(tx_pkt_iterator, tx_pkt_temp,
+	    &ep->sys->avail_tx_wrapper_list, link) {
+		list_del(&tx_pkt_iterator->link);
+		kmem_cache_free(ipa3_ctx->tx_pkt_wrapper_cache, tx_pkt_iterator);
+		ep->sys->avail_tx_wrapper--;
+	}
+	ep->sys->avail_tx_wrapper = 0;
+	spin_unlock_bh(&ep->sys->spinlock);
+}
+
 /**
  * ipa3_teardown_sys_pipe() - Teardown the GPI pipe and cleanup IPA EP
  * @clnt_hdl:	[in] the handle obtained from ipa3_setup_sys_pipe
@@ -1318,6 +1369,8 @@ int ipa3_teardown_sys_pipe(u32 clnt_hdl)
 			else
 				break;
 		} while (1);
+
+		delete_avail_tx_wrapper_list(ep);
 	}
 
 	/* channel stop might fail on timeout if IPA is busy */

drivers/platform/msm/ipa/ipa_v3/ipa_i.h

@@ -210,6 +210,7 @@
 #define IPA3_ACTIVE_CLIENTS_LOG_HASHTABLE_SIZE 50
 #define IPA3_ACTIVE_CLIENTS_LOG_NAME_LEN 40
 #define SMEM_IPA_FILTER_TABLE 497
+#define IPA_TX_WRAPPER_CACHE_MAX_THRESHOLD 2000
 
 enum {
 	SMEM_APPS,
@@ -1066,6 +1067,8 @@ struct ipa3_sys_context {
 	struct ipa3_ep_context *ep;
 	struct list_head head_desc_list;
 	struct list_head rcycl_list;
+	struct list_head avail_tx_wrapper_list;
+	u32 avail_tx_wrapper;
 	spinlock_t spinlock;
 	struct hrtimer db_timer;
 	struct workqueue_struct *wq;
@@ -2062,6 +2065,7 @@ struct ipa3_context {
 	u32 icc_num_paths;
 	u32 icc_clk[IPA_ICC_LVL_MAX][IPA_ICC_PATH_MAX][IPA_ICC_TYPE_MAX];
 	struct ipahal_imm_cmd_pyld *coal_cmd_pyld;
+	u32 tx_wrapper_cache_max_size;
 	struct ipa3_app_clock_vote app_clock_vote;
 	bool clients_registered;
 	bool rmnet_ctl_enable;
@@ -2122,6 +2126,7 @@ struct ipa3_plat_drv_res {
 	bool rmnet_ctl_enable;
 	const char *gsi_fw_file_name;
 	const char *uc_fw_file_name;
+	u32 tx_wrapper_cache_max_size;
 };
 
 /**