Skip to contentSkip to conversation

Introduce Protection Domain Restart (PDR) Helpers

PATCH Superseded Sibi Sankar 2 participants 5 messages Dec 20, 2019 – Dec 21, 2019

Subsystems (4)

v2 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v2 cover Introduce Protection Domain Restart (PDR) Helpers

Qualcomm SoCs (starting with MSM8998) allow for multiple protection domains (PDs) to run on the same Q6 sub-system. This allows for services like AVS AUDIO to have their own separate address space and crash/recover without disrupting the other PDs running on the same Q6 ADSP. This patch series introduces pdr helper library and adds PD tracking functionality for "avs/audio" allowing apr services to register themselves asynchronously once the dependent PDs are up.

V2:
* fixup pd_status callback to return void.
* return 0 from pdr_get_domain_list on success.
* introduce status_lock to linearize the pd_status reported back to the clients.
* use the correct service name length across various string operations performed.
* service locator will now schedule the pending lookups registered when it comes up.
* other minor cleanups that Bjorn suggested.
* use pr_warn to indicate that the wait for service locator timed out.
* add Bjorn/Srini's "Reviewed-by" for the dt-bindings.

To Do:
* Add support for pd tracking in fastrpc driver.

Sibi Sankar (3):
soc: qcom: Introduce Protection Domain Restart helpers
dt-bindings: soc: qcom: apr: Add protection domain bindings
soc: qcom: apr: Add avs/audio tracking functionality

 .../devicetree/bindings/soc/qcom/qcom,apr.txt |  59 ++
 drivers/soc/qcom/Kconfig                      |   6 +
 drivers/soc/qcom/Makefile                     |   1 +
 drivers/soc/qcom/apr.c                        | 100 ++-
 drivers/soc/qcom/pdr_interface.c              | 691 ++++++++++++++++++
 drivers/soc/qcom/pdr_internal.h               | 375 ++++++++++
 include/linux/soc/qcom/apr.h                  |   1 +
 include/linux/soc/qcom/pdr.h                  | 105 +++
 include/linux/soc/qcom/qmi.h                  |   1 +
 9 files changed, 1328 insertions(+), 11 deletions(-)
 create mode 100644 drivers/soc/qcom/pdr_interface.c
 create mode 100644 drivers/soc/qcom/pdr_internal.h
 create mode 100644 include/linux/soc/qcom/pdr.h
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#1 lore raw

PATCH v2 1/3 soc: qcom: Introduce Protection Domain Restart helpers

Qualcomm SoCs (starting with MSM8998) allow for multiple protection domains to run on the same Q6 sub-system. This allows for services like ATH10K WLAN FW to have their own separate address space and crash/recover without disrupting the modem and other PDs running on the same sub-system. The PDR helpers introduces an abstraction that allows for tracking/controlling the life cycle of protection domains running on various Q6 sub-systems.

---
 drivers/soc/qcom/Kconfig         |   5 +
 drivers/soc/qcom/Makefile        |   1 +
 drivers/soc/qcom/pdr_interface.c | 691 +++++++++++++++++++++++++++++++
 drivers/soc/qcom/pdr_internal.h  | 375 +++++++++++++++++
 include/linux/soc/qcom/pdr.h     | 105 +++++
 include/linux/soc/qcom/qmi.h     |   1 +
 6 files changed, 1178 insertions(+)
 create mode 100644 drivers/soc/qcom/pdr_interface.c
 create mode 100644 drivers/soc/qcom/pdr_internal.h
 create mode 100644 include/linux/soc/qcom/pdr.h
drivers/soc/qcom/Kconfig +5 -0
index 79d826553ac82..5c4e76837f59b 100644
@@ -76,6 +76,11 @@ config QCOM_OCMEM
           requirements. This is typically used by the GPU, camera/video, and
           audio components on some Snapdragon SoCs.
 
+config QCOM_PDR_HELPERS
+	tristate
+	depends on ARCH_QCOM || COMPILE_TEST
+	select QCOM_QMI_HELPERS
+
 config QCOM_PM
 	bool "Qualcomm Power Management"
 	depends on ARCH_QCOM && !ARM64
drivers/soc/qcom/Makefile +1 -0
index 9fb35c8a495e1..5d6b83dc58e82 100644
@@ -7,6 +7,7 @@ obj-$(CONFIG_QCOM_GLINK_SSR) +=	glink_ssr.o
 obj-$(CONFIG_QCOM_GSBI)	+=	qcom_gsbi.o
 obj-$(CONFIG_QCOM_MDT_LOADER)	+= mdt_loader.o
 obj-$(CONFIG_QCOM_OCMEM)	+= ocmem.o
+obj-$(CONFIG_QCOM_PDR_HELPERS)	+= pdr_interface.o
 obj-$(CONFIG_QCOM_PM)	+=	spm.o
 obj-$(CONFIG_QCOM_QMI_HELPERS)	+= qmi_helpers.o
 qmi_helpers-y	+= qmi_encdec.o qmi_interface.o
drivers/soc/qcom/pdr_interface.c new file +691 -0

691 lines not on this page · raw message

drivers/soc/qcom/pdr_internal.h new file +375 -0

375 lines not on this page · raw message

include/linux/soc/qcom/pdr.h new file +105 -0

105 lines not on this page · raw message

include/linux/soc/qcom/qmi.h +1 -0
index 5efa2b67fa557..e712f94b89fcc 100644
@@ -88,6 +88,7 @@ struct qmi_elem_info {
 #define QMI_ERR_CLIENT_IDS_EXHAUSTED_V01	5
 #define QMI_ERR_INVALID_ID_V01			41
 #define QMI_ERR_ENCODING_V01			58
+#define QMI_ERR_DISABLED_V01                    69
 #define QMI_ERR_INCOMPATIBLE_STATE_V01		90
 #define QMI_ERR_NOT_SUPPORTED_V01		94
 
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#2 lore raw

Re: [PATCH v2 1/3] soc: qcom: Introduce Protection Domain Restart helpers

Quoted 78 lines · Sibi Sankar

Qualcomm SoCs (starting with MSM8998) allow for multiple protection domains to run on the same Q6 sub-system. This allows for services like ATH10K WLAN FW to have their own separate address space and crash/recover without disrupting the modem and other PDs running on the same sub-system. The PDR helpers introduces an abstraction that allows for tracking/controlling the life cycle of protection domains running on various Q6 sub-systems.

---
 drivers/soc/qcom/Kconfig         |   5 +
 drivers/soc/qcom/Makefile        |   1 +
 drivers/soc/qcom/pdr_interface.c | 691 +++++++++++++++++++++++++++++++
 drivers/soc/qcom/pdr_internal.h  | 375 +++++++++++++++++
 include/linux/soc/qcom/pdr.h     | 105 +++++
 include/linux/soc/qcom/qmi.h     |   1 +
 6 files changed, 1178 insertions(+)
 create mode 100644 drivers/soc/qcom/pdr_interface.c
 create mode 100644 drivers/soc/qcom/pdr_internal.h
 create mode 100644 include/linux/soc/qcom/pdr.h
drivers/soc/qcom/Kconfig +5 -0
index 79d826553ac82..5c4e76837f59b 100644
@@ -76,6 +76,11 @@ config QCOM_OCMEM
           requirements. This is typically used by the GPU, camera/video, and
           audio components on some Snapdragon SoCs.
 
+config QCOM_PDR_HELPERS
+	tristate
+	depends on ARCH_QCOM || COMPILE_TEST
+	select QCOM_QMI_HELPERS
+
 config QCOM_PM
 	bool "Qualcomm Power Management"
 	depends on ARCH_QCOM && !ARM64
drivers/soc/qcom/Makefile +1 -0
index 9fb35c8a495e1..5d6b83dc58e82 100644
@@ -7,6 +7,7 @@ obj-$(CONFIG_QCOM_GLINK_SSR) +=	glink_ssr.o
 obj-$(CONFIG_QCOM_GSBI)	+=	qcom_gsbi.o
 obj-$(CONFIG_QCOM_MDT_LOADER)	+= mdt_loader.o
 obj-$(CONFIG_QCOM_OCMEM)	+= ocmem.o
+obj-$(CONFIG_QCOM_PDR_HELPERS)	+= pdr_interface.o
 obj-$(CONFIG_QCOM_PM)	+=	spm.o
 obj-$(CONFIG_QCOM_QMI_HELPERS)	+= qmi_helpers.o
 qmi_helpers-y	+= qmi_encdec.o qmi_interface.o
drivers/soc/qcom/pdr_interface.c new file +24 -0
new file mode 100644
index 0000000000000..2cfe012edd2f2
@@ -0,0 +1,691 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2019 The Linux Foundation. All rights reserved.
+ */
+
+#include <linux/completion.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/qrtr.h>
+#include <linux/string.h>
+#include <linux/workqueue.h>
+
+#include "pdr_internal.h"
+
+struct pdr_list_node {
+	u16 transaction_id;
+	struct pdr_service *pds;
+	struct list_head node;
+};
+
+static int servreg_locator_new_server(struct qmi_handle *qmi,
+				      struct qmi_service *svc)
+{
+	struct pdr_handle *pdr = container_of(qmi, struct pdr_handle,

drivers/soc/qcom/pdr_interface.c line 34

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+					      servloc_client);
+
+	/* Create a Local client port for QMI communication */
+	pdr->servloc_addr.sq_family = AF_QIPCRTR;
+	pdr->servloc_addr.sq_node = svc->node;
+	pdr->servloc_addr.sq_port = svc->port;
+
+	complete_all(&pdr->locator_available);
+
+	/* Restart service locator work to service pending lookup requests */

I think you should drop first half of this sentence, it's captured quite well by "Service pending lookup requests".

drivers/soc/qcom/pdr_interface.c line 45

Quoted 11 lines
drivers/soc/qcom/pdr_interface.c new file +11 -0
+	schedule_work(&pdr->servloc_work);
+
+	return 0;
+}
+
+static void servreg_locator_del_server(struct qmi_handle *qmi,
+				       struct qmi_service *svc)
+{
+	struct pdr_handle *pdr = container_of(qmi, struct pdr_handle,
+					      servloc_client);
+

Perhaps invalidate pdr->servloc_addr?

Quoted 28 lines
drivers/soc/qcom/pdr_interface.c new file +28 -0
+	reinit_completion(&pdr->locator_available);
+}
+
+static struct qmi_ops service_locator_ops = {
+	.new_server = servreg_locator_new_server,
+	.del_server = servreg_locator_del_server,
+};
+
+static int pdr_register_listener(struct pdr_handle *pdr,
+				 struct pdr_service *pds,
+				 bool enable)
+{
+	struct servreg_register_listener_resp resp;
+	struct servreg_register_listener_req req;
+	struct qmi_txn txn;
+	int ret;
+
+	ret = qmi_txn_init(&pdr->servreg_client, &txn,
+			   servreg_register_listener_resp_ei,
+			   &resp);
+
+	if (ret < 0)
+		return ret;
+
+	req.enable = enable;
+	strcpy(req.service_path, pds->service_path);
+
+	ret = qmi_send_request(&pdr->servreg_client, &pdr->servreg_addr,

drivers/soc/qcom/pdr_interface.c line 83

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+			       &txn, SERVREG_REGISTER_LISTENER_REQ,
+			       SERVREG_REGISTER_LISTENER_REQ_LEN,
+			       servreg_register_listener_req_ei,
+			       &req);
+	if (ret < 0) {
+		qmi_txn_cancel(&txn);
+		return ret;
+	}
+
+	ret = qmi_txn_wait(&txn, 5 * HZ);

We should fix the helpers so that it's possible to handle this asynchronously, there's no reason to do these in sequence and risk having a faulty remote service block all other registrations.

But we can do that in an incremental effort after landing this.

Quoted 19 lines
drivers/soc/qcom/pdr_interface.c new file +19 -0
+	if (ret < 0) {
+		pr_err("PDR: %s register listener txn wait failed: %d\n",
+		       pds->service_path, ret);
+		return ret;
+	}
+
+	/* Check the response */
+	if (resp.resp.result != QMI_RESULT_SUCCESS_V01) {
+		pr_err("PDR: %s register listener failed: 0x%x\n",
+		       pds->service_path, resp.resp.error);
+		return ret;
+	}
+
+	if ((int)resp.curr_state < INT_MIN || (int)resp.curr_state > INT_MAX)
+		pr_err("PDR: %s notification state invalid: 0x%x\n",
+		       pds->service_path, resp.curr_state);
+
+	pds->state = resp.curr_state;
+

drivers/soc/qcom/pdr_interface.c line 112

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+	return 0;
+}
+
+static void pdr_servreg_work(struct work_struct *work)
+{
+	struct pdr_handle *pdr = container_of(work, struct pdr_handle,
+					      servreg_work);
+	struct pdr_list_node *servreg, *tmp;
+	struct pdr_service *pds;
+

You're holding list_lock for insertions into servreg_list, but not during traversal and removal of entries.

drivers/soc/qcom/pdr_interface.c line 121

Quoted 9 lines
drivers/soc/qcom/pdr_interface.c new file +9 -0
+	list_for_each_entry_safe(servreg, tmp, &pdr->servreg_list, node) {
+		pds = servreg->pds;
+		pdr_register_listener(pdr, pds, true);
+		mutex_lock(&pdr->status_lock);
+		pdr->status(pdr, pds);
+		mutex_unlock(&pdr->status_lock);
+		list_del(&servreg->node);
+		kfree(servreg);
+	}

So you have servloc_list for tracking pds entries that needs to be looked up when the service locator becomes available and then you have servreg_list for tracking pds entries where the remote service has shown up.

As argued below in pdr_add_lookup() I think it would be better to have a single list of pds entries and keep track of when service_connected is going high and a SERVREG_REGISTER_LISTENER_REQ needs to be sent out.

Quoted 61 lines
drivers/soc/qcom/pdr_interface.c new file +61 -0
+}
+
+static int servreg_notifier_new_server(struct qmi_handle *qmi,
+				       struct qmi_service *svc)
+{
+	struct pdr_handle *pdr = container_of(qmi, struct pdr_handle,
+					      servreg_client);
+	struct pdr_list_node *servreg;
+	struct pdr_service *pds, *tmp;
+
+	/* Create a Local client port for QMI communication */
+	pdr->servreg_addr.sq_family = AF_QIPCRTR;
+	pdr->servreg_addr.sq_node = svc->node;
+	pdr->servreg_addr.sq_port = svc->port;
+
+	servreg = kzalloc(sizeof(*servreg), GFP_KERNEL);
+	if (!servreg)
+		return -ENOMEM;
+
+	list_for_each_entry_safe(pds, tmp, &pdr->lookups, node) {
+		if (pds->service == svc->service &&
+		    pds->instance == svc->instance) {
+			pds->service_connected = true;
+			servreg->pds = pds;
+			mutex_lock(&pdr->list_lock);
+			list_add_tail(&servreg->node, &pdr->servreg_list);
+			mutex_unlock(&pdr->list_lock);
+			break;
+		}
+	}
+
+	queue_work(pdr->servreg_wq, &pdr->servreg_work);
+
+	return 0;
+}
+
+static void pdr_servdel_work(struct work_struct *work)
+{
+	struct pdr_handle *pdr = container_of(work, struct pdr_handle,
+					      servdel_work);
+	struct pdr_list_node *servreg, *tmp;
+	struct pdr_service *pds;
+
+	list_for_each_entry_safe(servreg, tmp, &pdr->servdel_list, node) {
+		pds = servreg->pds;
+		pds->service_connected = false;
+		pds->state = SERVREG_SERVICE_STATE_DOWN;
+		mutex_lock(&pdr->status_lock);
+		pdr->status(pdr, pds);
+		mutex_unlock(&pdr->status_lock);
+		list_del(&servreg->node);
+		kfree(servreg);
+	}
+}
+
+static void servreg_notifier_del_server(struct qmi_handle *qmi,
+					struct qmi_service *svc)
+{
+	struct pdr_handle *pdr = container_of(qmi, struct pdr_handle,
+					      servreg_client);
+	struct pdr_list_node *servreg;

drivers/soc/qcom/pdr_interface.c line 192

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+	struct pdr_service *pds, *tmp;
+
+	servreg = kzalloc(sizeof(*servreg), GFP_KERNEL);
+	if (!servreg)
+		return;
+
+	list_for_each_entry_safe(pds, tmp, &pdr->lookups, node) {
+		if (pds->service == svc->service &&
+		    pds->instance == svc->instance) {
+			servreg->pds = pds;

Again, I believe that updating the state of the notifier and then schedule a single worker would simplify the driver.

Quoted 42 lines
drivers/soc/qcom/pdr_interface.c new file +42 -0
+			mutex_lock(&pdr->list_lock);
+			list_add_tail(&servreg->node, &pdr->servdel_list);
+			mutex_unlock(&pdr->list_lock);
+			break;
+		}
+	}
+
+	queue_work(pdr->servreg_wq, &pdr->servdel_work);
+}
+
+static struct qmi_ops service_notifier_ops = {
+	.new_server = servreg_notifier_new_server,
+	.del_server = servreg_notifier_del_server,
+};
+
+static int pdr_send_indack_msg(struct pdr_handle *pdr, struct pdr_service *pds,
+			       u16 tid)
+{
+	struct servreg_set_ack_resp resp;
+	struct servreg_set_ack_req req;
+	struct qmi_txn txn;
+	int ret;
+
+	ret = qmi_txn_init(&pdr->servreg_client, &txn, servreg_set_ack_resp_ei,
+			   &resp);
+	if (ret < 0)
+		return ret;
+
+	req.transaction_id = tid;
+	strcpy(req.service_path, pds->service_path);
+
+	ret = qmi_send_request(&pdr->servreg_client, &pdr->servreg_addr,
+			       &txn, SERVREG_SET_ACK_REQ,
+			       SERVREG_SET_ACK_REQ_LEN,
+			       servreg_set_ack_req_ei,
+			       &req);
+	if (ret < 0) {
+		qmi_txn_cancel(&txn);
+		return ret;
+	}
+
+	ret = qmi_txn_wait(&txn, 5 * HZ);

drivers/soc/qcom/pdr_interface.c line 244

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+	if (ret < 0) {
+		pr_err("PDR: %s indication ack txn wait failed: %d\n",
+		       pds->service_path, ret);
+		return ret;
+	}
+
+	/* Check the response */
+	if (resp.resp.result != QMI_RESULT_SUCCESS_V01)
+		pr_err("PDR: %s indication ack failed: 0x%x\n",
+		       pds->service_path, resp.resp.error);

What happens if we don't get a successful result back in a timely manner here? Should we draw the conclusion that the remote side is dead?

If we're just "ignoring" the reply I think you should skip waiting for it... I.e. just cancel the txn and return immediately...

Quoted 173 lines
drivers/soc/qcom/pdr_interface.c new file +173 -0
+
+	return 0;
+}
+
+static void pdr_indack_work(struct work_struct *work)
+{
+	struct pdr_handle *pdr = container_of(work, struct pdr_handle,
+					      indack_work);
+	struct pdr_list_node *ind, *tmp;
+	struct pdr_service *pds;
+
+	list_for_each_entry_safe(ind, tmp, &pdr->indack_list, node) {
+		pds = ind->pds;
+		pdr_send_indack_msg(pdr, pds, ind->transaction_id);
+		mutex_lock(&pdr->status_lock);
+		pdr->status(pdr, pds);
+		mutex_unlock(&pdr->status_lock);
+		list_del(&ind->node);
+		kfree(ind);
+	}
+}
+
+static void pdr_servreg_ind_cb(struct qmi_handle *qmi,
+			       struct sockaddr_qrtr *sq,
+			       struct qmi_txn *txn, const void *data)
+{
+	struct pdr_handle *pdr = container_of(qmi, struct pdr_handle,
+					      servreg_client);
+	const struct servreg_state_updated_ind *ind_msg = data;
+	struct pdr_list_node *ind;
+	struct pdr_service *pds;
+
+	if (!ind_msg || !ind_msg->service_path ||
+	    strlen(ind_msg->service_path) > SERVREG_NAME_LENGTH)
+		return;
+
+	list_for_each_entry(pds, &pdr->lookups, node) {
+		if (!strcmp(pds->service_path, ind_msg->service_path))
+			goto found;
+	}
+	return;
+
+found:
+	pds->state = ind_msg->curr_state;
+
+	ind = kzalloc(sizeof(*ind), GFP_KERNEL);
+	if (!ind)
+		return;
+
+	pr_info("PDR: Indication received from %s, state: 0x%x, trans-id: %d\n",
+		ind_msg->service_path, ind_msg->curr_state,
+		ind_msg->transaction_id);
+
+	ind->transaction_id = ind_msg->transaction_id;
+	ind->pds = pds;
+
+	mutex_lock(&pdr->list_lock);
+	list_add_tail(&ind->node, &pdr->indack_list);
+	mutex_unlock(&pdr->list_lock);
+
+	queue_work(pdr->indack_wq, &pdr->indack_work);
+}
+
+static struct qmi_msg_handler qmi_indication_handler[] = {
+	{
+		.type = QMI_INDICATION,
+		.msg_id = SERVREG_STATE_UPDATED_IND_ID,
+		.ei = servreg_state_updated_ind_ei,
+		.decoded_size = sizeof(struct servreg_state_updated_ind),
+		.fn = pdr_servreg_ind_cb,
+	},
+	{}
+};
+
+static int pdr_get_domain_list(struct servreg_get_domain_list_req *req,
+			       struct servreg_get_domain_list_resp *resp,
+			       struct pdr_handle *pdr)
+{
+	struct qmi_txn txn;
+	int ret;
+
+	ret = qmi_txn_init(&pdr->servloc_client, &txn,
+			   servreg_get_domain_list_resp_ei, resp);
+	if (ret < 0)
+		return ret;
+
+	ret = qmi_send_request(&pdr->servloc_client,
+			       &pdr->servloc_addr,
+			       &txn, SERVREG_GET_DOMAIN_LIST_REQ,
+			       SERVREG_GET_DOMAIN_LIST_REQ_MAX_LEN,
+			       servreg_get_domain_list_req_ei,
+			       req);
+	if (ret < 0) {
+		qmi_txn_cancel(&txn);
+		return ret;
+	}
+
+	ret = qmi_txn_wait(&txn, 5 * HZ);
+	if (ret < 0) {
+		pr_err("PDR: %s get domain list txn wait failed: %d\n",
+		       req->service_name, ret);
+		return ret;
+	}
+
+	/* Check the response */
+	if (resp->resp.result != QMI_RESULT_SUCCESS_V01) {
+		pr_err("PDR: %s get domain list failed: 0x%x\n",
+		       req->service_name, resp->resp.error);
+		return -EREMOTEIO;
+	}
+
+	return 0;
+}
+
+static int pdr_locate_service(struct pdr_handle *pdr, struct pdr_service *pds)
+{
+	struct servreg_get_domain_list_resp *resp = NULL;
+	struct servreg_get_domain_list_req req;
+	int db_rev_count = 0, domains_read = 0;
+	struct servreg_location_entry *entry;
+	int ret, i;
+
+	resp = kzalloc(sizeof(*resp), GFP_KERNEL);
+	if (!resp)
+		return -ENOMEM;
+
+	/* Prepare req message */
+	strcpy(req.service_name, pds->service_name);
+	req.domain_offset_valid = true;
+	req.domain_offset = 0;
+
+	do {
+		req.domain_offset = domains_read;
+		ret = pdr_get_domain_list(&req, resp, pdr);
+		if (ret < 0)
+			goto out;
+
+		if (!domains_read)
+			db_rev_count = resp->db_rev_count;
+
+		if (db_rev_count != resp->db_rev_count) {
+			ret = -EAGAIN;
+			goto out;
+		}
+
+		for (i = domains_read; i < resp->domain_list_len; i++) {
+			entry = &resp->domain_list[i];
+
+			if (strnlen(entry->name, sizeof(entry->name)) == sizeof(entry->name))
+				continue;
+
+			if (!strcmp(entry->name, pds->service_path)) {
+				pds->service_data_valid = entry->service_data_valid;
+				pds->service_data = entry->service_data;
+				pds->instance = entry->instance;
+				goto out;
+			}
+		}
+
+		/* Update ret to indicate that the service is not yet found */
+		ret = -EINVAL;
+
+		/* Always read total_domains from the response msg */
+		if (resp->domain_list_len > resp->total_domains)
+			resp->domain_list_len = resp->total_domains;
+
+		domains_read += resp->domain_list_len;
+	} while (domains_read < resp->total_domains);
+out:
+	kfree(resp);
+	return ret;
+}
+

drivers/soc/qcom/pdr_interface.c line 427

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+static void pdr_servloc_work(struct work_struct *work)
+{
+	struct pdr_handle *pdr = container_of(work, struct pdr_handle,
+					      servloc_work);
+	struct pdr_list_node *servloc, *tmp;
+	struct pdr_service *pds;
+	int ret;
+
+	/* Wait for PD Mapper to come up */
+	ret = wait_for_completion_timeout(&pdr->locator_available, 5 * HZ);

servloc_work is scheduled either when the service locator is coming up or when a new client is registered.

In the first case locator_available will be completed right before, so the only case where this won't just go through is if the service locator disappears before this function is scheduled. In this case you'll sit here and wait for 5 seconds, for an add notification, which would schedule the worker again.

In the other case, where a pdr client is registered before the service locator is registered, either the service locator is becoming available before the worker is scheduled, or you'll wait for up to 5 seconds on it being registered. But it being registered would schedule this function anyways.

So I think rather than the completion you could check if the service locator is present and fail immediately otherwise.

drivers/soc/qcom/pdr_interface.c line 432

Quoted 5 lines
drivers/soc/qcom/pdr_interface.c new file +5 -0
+	if (!ret) {
+		pr_warn("PDR: SERVICE LOCATOR service wait timed out\n");
+		return;
+	}
+

You're holding list_lock when inserting into servloc_list, but not when traversing or removing entries from it.

drivers/soc/qcom/pdr_interface.c line 440

Quoted 8 lines
drivers/soc/qcom/pdr_interface.c new file +8 -0
+	list_for_each_entry_safe(servloc, tmp, &pdr->servloc_list, node) {
+		pds = servloc->pds;
+
+		ret = pdr_locate_service(pdr, pds);
+		if (ret < 0) {
+			/* Notify Lookup failed */
+			mutex_lock(&pdr->status_lock);
+			pdr->status(pdr, pds);

Is the purpose of this to notify the driver that we have a service locator, but it does not know about the particular service?

I think you need to distinguish between the "we don't have information about this service in the json files" and "something went wrong in the lookup" - in which case you should try to recover/retry.

Quoted 52 lines
drivers/soc/qcom/pdr_interface.c new file +52 -0
+			mutex_unlock(&pdr->status_lock);
+			pr_err("PDR: service lookup for %s failed: %d\n",
+			       pds->service_name, ret);
+			list_del(&pds->node);
+			kfree(pds);
+		} else {
+			qmi_add_lookup(&pdr->servreg_client, pds->service, 1,
+				       pds->instance);
+		}
+
+		list_del(&servloc->node);
+		kfree(servloc);
+	}
+}
+
+/**
+ * pdr_add_lookup() - register a tracking request for a PD
+ * @pdr:		PDR client handle
+ * @service_name:	service name of the tracking request
+ * @service_path:	service path of the tracking request
+ *
+ * Registering a pdr lookup allows for tracking the life cycle of the PD.
+ *
+ * Return: 0 on success, negative errno on failure.
+ */
+int pdr_add_lookup(struct pdr_handle *pdr, const char *service_name,
+		   const char *service_path)
+{
+	struct pdr_service *pds, *pds_iter, *tmp;
+	struct pdr_list_node *servloc;
+	int ret;
+
+	if (!service_name || strlen(service_name) > SERVREG_NAME_LENGTH ||
+	    !service_path || strlen(service_path) > SERVREG_NAME_LENGTH)
+		return -EINVAL;
+
+	servloc = kzalloc(sizeof(*servloc), GFP_KERNEL);
+	if (!servloc)
+		return -ENOMEM;
+
+	pds = kzalloc(sizeof(*pds), GFP_KERNEL);
+	if (!pds) {
+		ret = -ENOMEM;
+		goto err;
+	}
+
+	pds->service = SERVREG_NOTIFIER_SERVICE;
+	strcpy(pds->service_name, service_name);
+	strcpy(pds->service_path, service_path);
+	servloc->pds = pds;
+
+	mutex_lock(&pdr->list_lock);

drivers/soc/qcom/pdr_interface.c line 502

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+	list_for_each_entry_safe(pds_iter, tmp, &pdr->lookups, node) {
+		if (!strcmp(pds_iter->service_path, service_path)) {
+			mutex_unlock(&pdr->list_lock);
+			ret = -EALREADY;
+			goto err;
+		}
+	}
+
+	list_add(&pds->node, &pdr->lookups);
+	list_add(&servloc->node, &pdr->servloc_list);

Rather than having servloc_list contain the list of entries from the lookups list that needs to be locked up, how about just keeping some state flag in the pds; e.g. need-servreg-locator-lookup, need-qmi-lookup?

I believe this would simplify the reasoning about this driver and would allow you to express cases such as servreg-location succeeds but qmi_add_lookup() for some reason failed and needs to be retried.

Quoted 164 lines
drivers/soc/qcom/pdr_interface.c new file +164 -0
+	mutex_unlock(&pdr->list_lock);
+
+	schedule_work(&pdr->servloc_work);
+
+	return 0;
+err:
+	kfree(pds);
+	kfree(servloc);
+
+	return ret;
+}
+EXPORT_SYMBOL(pdr_add_lookup);
+
+/**
+ * pdr_restart_pd() - restart PD
+ * @pdr:		PDR client handle
+ * @service_path:	service path of restart request
+ *
+ * Restarts the PD tracked by the PDR client handle for a given service path.
+ *
+ * Return: 0 on success, negative errno on failure.
+ */
+int pdr_restart_pd(struct pdr_handle *pdr, const char *service_path)
+{
+	struct servreg_restart_pd_req req;
+	struct servreg_restart_pd_resp resp;
+	struct pdr_service *pds = NULL, *pds_iter, *tmp;
+	struct qmi_txn txn;
+	int ret;
+
+	if (!service_path || strlen(service_path) > SERVREG_NAME_LENGTH)
+		return -EINVAL;
+
+	list_for_each_entry_safe(pds_iter, tmp, &pdr->lookups, node) {
+		if (!pds_iter->service_connected)
+			continue;
+
+		if (!strcmp(pds_iter->service_path, service_path)) {
+			pds = pds_iter;
+			break;
+		}
+	}
+
+	if (!pds)
+		return -EINVAL;
+
+	/* Prepare req message */
+	strcpy(req.service_path, pds->service_path);
+
+	ret = qmi_txn_init(&pdr->servreg_client, &txn,
+			   servreg_restart_pd_resp_ei,
+			   &resp);
+	if (ret < 0)
+		return ret;
+
+	ret = qmi_send_request(&pdr->servreg_client, &pdr->servreg_addr,
+			       &txn, SERVREG_RESTART_PD_REQ,
+			       SERVREG_RESTART_PD_REQ_MAX_LEN,
+			       servreg_restart_pd_req_ei, &req);
+	if (ret < 0) {
+		qmi_txn_cancel(&txn);
+		return ret;
+	}
+
+	ret = qmi_txn_wait(&txn, 5 * HZ);
+	if (ret < 0) {
+		pr_err("PDR: %s PD restart txn wait failed: %d\n",
+		       pds->service_path, ret);
+		return ret;
+	}
+
+	/* Check response if PDR is disabled */
+	if (resp.resp.result == QMI_RESULT_FAILURE_V01 &&
+	    resp.resp.error == QMI_ERR_DISABLED_V01) {
+		pr_err("PDR: %s PD restart is disabled: 0x%x\n",
+		       pds->service_path, resp.resp.error);
+		return -EOPNOTSUPP;
+	}
+
+	/* Check the response for other error case*/
+	if (resp.resp.result != QMI_RESULT_SUCCESS_V01) {
+		pr_err("PDR: %s request for PD restart failed: 0x%x\n",
+		       pds->service_path, resp.resp.error);
+		return -EREMOTEIO;
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL(pdr_restart_pd);
+
+/**
+ * pdr_handle_init() - initialize the PDR client handle
+ * @pdr:	PDR client handle
+ * @status:	function to be called on PD state change
+ *
+ * Initializes the PDR client handle to allow for tracking/restart of PDs.
+ *
+ * Return: 0 on success, negative errno on failure.
+ */
+int pdr_handle_init(struct pdr_handle *pdr,
+		    void (*status)(struct pdr_handle *pdr,
+				   struct pdr_service *pds))
+{
+	int ret;
+
+	if (!status)
+		return -EINVAL;
+
+	pdr->status = *status;
+
+	mutex_init(&pdr->list_lock);
+	mutex_init(&pdr->status_lock);
+	init_completion(&pdr->locator_available);
+
+	INIT_LIST_HEAD(&pdr->lookups);
+	INIT_LIST_HEAD(&pdr->servloc_list);
+	INIT_LIST_HEAD(&pdr->servreg_list);
+	INIT_LIST_HEAD(&pdr->servdel_list);
+	INIT_LIST_HEAD(&pdr->indack_list);
+
+	INIT_WORK(&pdr->servloc_work, pdr_servloc_work);
+	INIT_WORK(&pdr->servreg_work, pdr_servreg_work);
+	INIT_WORK(&pdr->servdel_work, pdr_servdel_work);
+	INIT_WORK(&pdr->indack_work, pdr_indack_work);
+
+	pdr->servreg_wq = create_singlethread_workqueue("pdr_servreg_wq");
+	if (!pdr->servreg_wq)
+		return -ENOMEM;
+
+	pdr->indack_wq = alloc_ordered_workqueue("pdr_indack_wq", WQ_HIGHPRI);
+	if (!pdr->indack_wq) {
+		ret = -ENOMEM;
+		goto destroy_servreg;
+	}
+
+	ret = qmi_handle_init(&pdr->servloc_client,
+			      SERVREG_GET_DOMAIN_LIST_RESP_MAX_LEN,
+			      &service_locator_ops, NULL);
+	if (ret < 0)
+		goto destroy_indack;
+
+	ret = qmi_handle_init(&pdr->servreg_client,
+			      SERVREG_STATE_UPDATED_IND_MAX_LEN,
+			      &service_notifier_ops, qmi_indication_handler);
+	if (ret < 0)
+		goto release_handle;
+
+	qmi_add_lookup(&pdr->servloc_client, SERVREG_LOCATOR_SERVICE, 1, 1);
+
+	return 0;
+
+release_handle:
+	qmi_handle_release(&pdr->servloc_client);
+destroy_indack:
+	destroy_workqueue(pdr->indack_wq);
+destroy_servreg:
+	destroy_workqueue(pdr->servreg_wq);
+
+	return ret;
+}
+EXPORT_SYMBOL(pdr_handle_init);
+
+/**
+ * pdr_handle_release() - release the PDR client handle

drivers/soc/qcom/pdr_interface.c line 676

Quoted 10 lines
drivers/soc/qcom/pdr_interface.c new file +10 -0
+ * @pdr:	PDR client handle
+ *
+ * Cleans up pending tracking requests and releases the underlying qmi handles.
+ */
+void pdr_handle_release(struct pdr_handle *pdr)
+{
+	struct pdr_service *pds, *tmp;
+
+	qmi_handle_release(&pdr->servloc_client);
+	qmi_handle_release(&pdr->servreg_client);

Aren't the workers operating on the qmi handles until you've cancelled them below?

drivers/soc/qcom/pdr_interface.c line 689

Quoted 13 lines
drivers/soc/qcom/pdr_interface.c new file +13 -0
+
+	cancel_work_sync(&pdr->servloc_work);
+	cancel_work_sync(&pdr->servreg_work);
+	cancel_work_sync(&pdr->servdel_work);
+	cancel_work_sync(&pdr->indack_work);
+
+	destroy_workqueue(pdr->servreg_wq);
+	destroy_workqueue(pdr->indack_wq);
+
+	list_for_each_entry_safe(pds, tmp, &pdr->lookups, node) {
+		list_del(&pds->node);
+		kfree(pds);
+	}

So wouldn't it make sense to drop all pds entries, then cancel all the work, then release the qmi handles?

drivers/soc/qcom/pdr_interface.c line 690

Quoted 1 line
drivers/soc/qcom/pdr_interface.c new file +1 -0
+}

Regards,
Bjorn

#3 lore raw

PATCH v2 2/3 dt-bindings: soc: qcom: apr: Add protection domain bindings

In reply to Sibi Sankar

Add optional "qcom,protection-domain" bindings for APR services. This helps to capture the dependencies between APR services and the PD on which each apr service run.

---
 .../devicetree/bindings/soc/qcom/qcom,apr.txt | 59 +++++++++++++++++++
 1 file changed, 59 insertions(+)
Documentation/devicetree/bindings/soc/qcom/qcom,apr.txt +59 -0
index db501269f47b8..f87c0b2a48de4 100644
@@ -45,6 +45,12 @@ by the individual bindings for the specific service
 			12 - Ultrasound stream manager.
 			13 - Listen stream manager.
 
+- qcom,protection-domain
+	Usage: optional
+	Value type: <stringlist>
+	Definition: Must list the protection domain service name and path
+		    that the particular apr service has a dependency on.
+
 = EXAMPLE
 The following example represents a QDSP based sound card on a MSM8996 device
 which uses apr as communication between Apps and QDSP.
@@ -82,3 +88,56 @@ which uses apr as communication between Apps and QDSP.
 			...
 		};
 	};
+
+= EXAMPLE 2
+The following example represents a QDSP based sound card on SDM845 device.
+Here the apr services are dependent on "avs/audio" service running on AUDIO
+Protection Domain hosted on ADSP remote processor.
+
+	apr {
+		compatible = "qcom,apr-v2";
+		qcom,glink-channels = "apr_audio_svc";
+		qcom,apr-domain = <APR_DOMAIN_ADSP>;
+
+		q6core {
+			compatible = "qcom,q6core";
+			reg = <APR_SVC_ADSP_CORE>;
+			qcom,protection-domain = "avs/audio", "msm/adsp/audio_pd";
+		};
+
+		q6afe: q6afe {
+			compatible = "qcom,q6afe";
+			reg = <APR_SVC_AFE>;
+			qcom,protection-domain = "avs/audio", "msm/adsp/audio_pd";
+			q6afedai: dais {
+				compatible = "qcom,q6afe-dais";
+				#sound-dai-cells = <1>;
+
+				qi2s@22 {
+					reg = <22>;
+					qcom,sd-lines = <3>;
+				};
+			};
+		};
+
+		q6asm: q6asm {
+			compatible = "qcom,q6asm";
+			reg = <APR_SVC_ASM>;
+			qcom,protection-domain = "avs/audio", "msm/adsp/audio_pd";
+			q6asmdai: dais {
+				compatible = "qcom,q6asm-dais";
+				#sound-dai-cells = <1>;
+				iommus = <&apps_smmu 0x1821 0x0>;
+			};
+		};
+
+		q6adm: q6adm {
+			compatible = "qcom,q6adm";
+			reg = <APR_SVC_ADM>;
+			qcom,protection-domain = "avs/audio", "msm/adsp/audio_pd";
+			q6routing: routing {
+				compatible = "qcom,q6adm-routing";
+				#sound-dai-cells = <0>;
+			};
+		};
+	};
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#4 lore raw

PATCH v2 3/3 soc: qcom: apr: Add avs/audio tracking functionality

In reply to Sibi Sankar

Use PDR helper functions to track the protection domains that the apr services are dependent upon on SDM845 SoC, specifically the "avs/audio" service running on ADSP Q6.

---
 drivers/soc/qcom/Kconfig     |   1 +
 drivers/soc/qcom/apr.c       | 100 +++++++++++++++++++++++++++++++----
 include/linux/soc/qcom/apr.h |   1 +
 3 files changed, 91 insertions(+), 11 deletions(-)
drivers/soc/qcom/Kconfig +1 -0
index 5c4e76837f59b..cacfed945b275 100644
@@ -202,6 +202,7 @@ config QCOM_APR
 	tristate "Qualcomm APR Bus (Asynchronous Packet Router)"
 	depends on ARCH_QCOM || COMPILE_TEST
 	depends on RPMSG
+	select QCOM_PDR_HELPERS
 	help
           Enable APR IPC protocol support between
           application processor and QDSP6. APR is
drivers/soc/qcom/apr.c +89 -11

159 lines not on this page · raw message

include/linux/soc/qcom/apr.h +1 -0
index c5d52e2cb275f..7f0bc3cf4d610 100644
@@ -85,6 +85,7 @@ struct apr_device {
 	uint16_t	domain_id;
 	uint32_t	version;
 	char name[APR_NAME_SIZE];
+	const char *service_path;
 	spinlock_t	lock;
 	struct list_head node;
 };
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project