Re: [RFC,v2] Bluetooth: hci_qca: Collect controller memory dump during SSR

From: Marcel Holtmann
Date: Fri Jan 04 2019 - 04:11:09 EST


Hi Balakrishna,

>>> We will collect the ramdump of BT controller when hardware error event
>>> received before rebooting the HCI layer. Before restarting a subsystem
>>> or a process running on a subsystem, it is often required to request
>>> either a subsystem or a process to perform proper cache dump and
>>> software failure reason into a memory buffer which application
>>> processor can retrieve afterwards. SW developers can often provide
>>> initial investigation by looking into that debugging information.
>>> Signed-off-by: Balakrishna Godavarthi <bgodavar@xxxxxxxxxxxxxx>
>>> ---
>>> changes v2:
>>> * entirely an newer approach handling with an work queue.
>>> ---
>>> drivers/bluetooth/hci_qca.c | 289 ++++++++++++++++++++++++++++++++++--
>>> 1 file changed, 278 insertions(+), 11 deletions(-)
>>> diff --git a/drivers/bluetooth/hci_qca.c b/drivers/bluetooth/hci_qca.c
>>> index f036c8f98ea3..62b768bc32ec 100644
>>> --- a/drivers/bluetooth/hci_qca.c
>>> +++ b/drivers/bluetooth/hci_qca.c
>>> @@ -32,6 +32,7 @@
>>> #include <linux/clk.h>
>>> #include <linux/debugfs.h>
>>> #include <linux/delay.h>
>>> +#include <linux/devcoredump.h>
>>> #include <linux/device.h>
>>> #include <linux/gpio/consumer.h>
>>> #include <linux/mod_devicetable.h>
>> don't we have some crashdump core facility that could be utilized here?
>
> [Bala]: no i don't think so. that is reason calling devcoredump.h

what is wrong with using devcoredump.h?

>>> @@ -57,9 +58,10 @@
>>> /* Controller states */
>>> #define STATE_IN_BAND_SLEEP_ENABLED 1
>>> -#define IBS_WAKE_RETRANS_TIMEOUT_MS 100
>>> +#define IBS_WAKE_RETRANS_TIMEOUT_MS 100
>>> #define IBS_TX_IDLE_TIMEOUT_MS 2000
>>> -#define BAUDRATE_SETTLE_TIMEOUT_MS 300
>>> +#define BAUDRATE_SETTLE_TIMEOUT_MS 300
>>> +#define MEMDUMP_COLLECTION_TIMEOUT_MS 8000
>>> /* susclk rate */
>>> #define SUSCLK_RATE_32KHZ 32768
>>> @@ -67,6 +69,13 @@
>>> /* Controller debug log header */
>>> #define QCA_DEBUG_HANDLE 0x2EDC
>>> +/* Controller dump header */
>>> +#define QCA_SSR_DUMP_HANDLE 0x0108
>>> +#define QCA_DUMP_PACKET_SIZE 255
>>> +#define QCA_LAST_SEQUENCE_NUM 0xFFFF
>>> +#define QCA_CRASHBYTE_PACKET_LEN 1100
>>> +#define QCA_MEMDUMP_BYTE 0xFB
>>> +
>>> /* HCI_IBS transmit side sleep protocol states */
>>> enum tx_ibs_states {
>>> HCI_IBS_TX_ASLEEP,
>>> @@ -89,12 +98,41 @@ enum hci_ibs_clock_state_vote {
>>> HCI_IBS_RX_VOTE_CLOCK_OFF,
>>> };
>>> +/* Controller memory dump states */
>>> +enum qca_memdump_states {
>>> + QCA_MEMDUMP_IDLE,
>>> + QCA_MEMDUMP_COLLECTING,
>>> + QCA_MEMDUMP_COLLECTED,
>>> + QCA_MEMDUMP_TIMEOUT,
>>> +};
>>> +
>>> +struct qca_memdump_data {
>>> + char *memdump_buf_head;
>>> + char *memdump_buf_tail;
>>> + u32 current_seq_no;
>>> + u32 received_dump;
>>> +};
>>> +
>>> +struct qca_memdump_event_hdr {
>>> + __u8 evt;
>>> + __u8 plen;
>>> + __u16 opcode;
>>> + __u16 seq_no;
>>> + __u8 reserved;
>>> +} __packed;
>>> +
>>> +
>>> +struct qca_dump_size {
>>> + u32 dump_size;
>>> +} __packed;
>>> +
>>> struct qca_data {
>>> struct hci_uart *hu;
>>> struct sk_buff *rx_skb;
>>> struct sk_buff_head txq;
>>> - struct sk_buff_head tx_wait_q; /* HCI_IBS wait queue */
>>> - spinlock_t hci_ibs_lock; /* HCI_IBS state lock */
>>> + struct sk_buff_head tx_wait_q; /* HCI_IBS wait queue */
>>> + struct sk_buff_head rx_memdump_q; /* Memdump wait queue */
>>> + spinlock_t hci_ibs_lock; /* HCI_IBS state lock */
>>> u8 tx_ibs_state; /* HCI_IBS transmit side power state*/
>>> u8 rx_ibs_state; /* HCI_IBS receive side power state */
>>> bool tx_vote; /* Clock must be on for TX */
>>> @@ -103,12 +141,17 @@ struct qca_data {
>>> u32 tx_idle_delay;
>>> struct timer_list wake_retrans_timer;
>>> u32 wake_retrans;
>>> + struct timer_list memdump_timer;
>>> + u32 memdump_delay;
>>> struct workqueue_struct *workqueue;
>>> struct work_struct ws_awake_rx;
>>> struct work_struct ws_awake_device;
>>> struct work_struct ws_rx_vote_off;
>>> struct work_struct ws_tx_vote_off;
>>> + struct work_struct ctrl_memdump_evt;
>>> + struct qca_memdump_data *qca_memdump;
>>> unsigned long flags;
>>> + enum qca_memdump_states memdump_state;
>>> /* For debugging purpose */
>>> u64 ibs_sent_wacks;
>>> @@ -173,6 +216,7 @@ struct qca_serdev {
>>> static int qca_power_setup(struct hci_uart *hu, bool on);
>>> static void qca_power_shutdown(struct hci_uart *hu);
>>> static int qca_power_off(struct hci_dev *hdev);
>>> +static void qca_controller_memdump(struct work_struct *work);
>>> static void __serial_clock_on(struct tty_struct *tty)
>>> {
>>> @@ -446,6 +490,21 @@ static void hci_ibs_wake_retrans_timeout(struct timer_list *t)
>>> hci_uart_tx_wakeup(hu);
>>> }
>>> +static void hci_memdump_timeout(struct timer_list *t)
>>> +{
>>> + struct qca_data *qca = from_timer(qca, t, tx_idle_timer);
>>> + struct hci_uart *hu = qca->hu;
>>> + struct qca_memdump_data *qca_memdump = qca->qca_memdump;
>>> + char *memdump_buf = qca_memdump->memdump_buf_tail;
>>> +
>>> + bt_dev_err(hu->hdev, "clearing allocated memory due to memdump timeout");
>>> + kfree(memdump_buf);
>>> + kfree(qca_memdump);
>>> + qca->memdump_state = QCA_MEMDUMP_TIMEOUT;
>>> + del_timer(&qca->memdump_timer);
>>> + cancel_work_sync(&qca->ctrl_memdump_evt);
>>> +}
>>> +
>>> /* Initialize protocol */
>>> static int qca_open(struct hci_uart *hu)
>>> {
>>> @@ -461,6 +520,7 @@ static int qca_open(struct hci_uart *hu)
>>> skb_queue_head_init(&qca->txq);
>>> skb_queue_head_init(&qca->tx_wait_q);
>>> + skb_queue_head_init(&qca->rx_memdump_q);
>>> spin_lock_init(&qca->hci_ibs_lock);
>>> qca->workqueue = alloc_ordered_workqueue("qca_wq", 0);
>>> if (!qca->workqueue) {
>>> @@ -473,6 +533,7 @@ static int qca_open(struct hci_uart *hu)
>>> INIT_WORK(&qca->ws_awake_device, qca_wq_awake_device);
>>> INIT_WORK(&qca->ws_rx_vote_off, qca_wq_serial_rx_clock_vote_off);
>>> INIT_WORK(&qca->ws_tx_vote_off, qca_wq_serial_tx_clock_vote_off);
>>> + INIT_WORK(&qca->ctrl_memdump_evt, qca_controller_memdump);
>>> qca->hu = hu;
>>> @@ -500,7 +561,7 @@ static int qca_open(struct hci_uart *hu)
>>> qca->tx_votes_off = 0;
>>> qca->rx_votes_on = 0;
>>> qca->rx_votes_off = 0;
>>> -
>>> + qca->memdump_state = QCA_MEMDUMP_IDLE;
>>> hu->priv = qca;
>>> if (hu->serdev) {
>>> @@ -528,6 +589,9 @@ static int qca_open(struct hci_uart *hu)
>>> timer_setup(&qca->tx_idle_timer, hci_ibs_tx_idle_timeout, 0);
>>> qca->tx_idle_delay = IBS_TX_IDLE_TIMEOUT_MS;
>>> + timer_setup(&qca->memdump_timer, hci_memdump_timeout, 0);
>>> + qca->memdump_delay = MEMDUMP_COLLECTION_TIMEOUT_MS;
>>> +
>>> BT_DBG("HCI_UART_QCA open, tx_idle_delay=%u, wake_retrans=%u",
>>> qca->tx_idle_delay, qca->wake_retrans);
>>> @@ -605,8 +669,10 @@ static int qca_close(struct hci_uart *hu)
>>> skb_queue_purge(&qca->tx_wait_q);
>>> skb_queue_purge(&qca->txq);
>>> + skb_queue_purge(&qca->rx_memdump_q);
>>> del_timer(&qca->tx_idle_timer);
>>> del_timer(&qca->wake_retrans_timer);
>>> + del_timer(&qca->memdump_timer);
>>> destroy_workqueue(qca->workqueue);
>>> qca->hu = NULL;
>>> @@ -866,6 +932,141 @@ static int qca_recv_acl_data(struct hci_dev *hdev, struct sk_buff *skb)
>>> return hci_recv_frame(hdev, skb);
>>> }
>>> +static void qca_controller_memdump(struct work_struct *work)
>>> +{
>>> + struct qca_data *qca = container_of(work, struct qca_data,
>>> + ctrl_memdump_evt);
>>> + struct hci_uart *hu = qca->hu;
>>> + struct sk_buff *skb;
>>> + struct qca_memdump_event_hdr *cmd_hdr;
>>> + struct qca_memdump_data *qca_memdump = qca->qca_memdump;
>>> + struct qca_dump_size *dump;
>>> + char *memdump_buf;
>>> + char nullBuff[QCA_DUMP_PACKET_SIZE] = { 0 };
>>> + u16 opcode, seq_no;
>>> + u32 dump_size;
>>> +
>>> + while ((skb = skb_dequeue(&qca->rx_memdump_q))) {
>>> +
>>> + if (!qca_memdump) {
>>> + qca_memdump = kzalloc(sizeof(struct qca_memdump_data),
>>> + GFP_ATOMIC);
>>> + if (!qca_memdump)
>>> + return;
>>> +
>>> + qca->qca_memdump = qca_memdump;
>>> + }
>>> +
>>> + qca->memdump_state = QCA_MEMDUMP_COLLECTING;
>>> + cmd_hdr = (void *) skb->data;
>>> + opcode = __le16_to_cpu(cmd_hdr->opcode);
>>> + seq_no = __le16_to_cpu(cmd_hdr->seq_no);
>>> + skb_pull(skb, sizeof(struct qca_memdump_event_hdr));
>>> +
>>> + if (!seq_no) {
>>> +
>>> + /* This is the first frame of memdump packet from
>>> + * the controller, Disable IBS to recevie dump
>>> + * with out any interruption, ideally time required for
>>> + * the controller to send the dump is 8 seconds. let us
>>> + * start timer to handle this asynchronous activity.
>>> + */
>>> + clear_bit(STATE_IN_BAND_SLEEP_ENABLED, &qca->flags);
>>> +
>>> + dump = (void *) skb->data;
>>> + dump_size = __le32_to_cpu(dump->dump_size);
>>> + if (!(dump_size)) {
>>> + bt_dev_err(hu->hdev, "QCA invalid memdump size");
>>> + kfree_skb(skb);
>>> + return;
>>> + }
>>> +
>>> + bt_dev_info(hu->hdev, "QCA: collecting memdump started"
>>> + "of size:%u", dump_size);
>>> + mod_timer(&qca->memdump_timer, (jiffies +
>>> + msecs_to_jiffies(qca->memdump_delay)));
>>> +
>>> + skb_pull(skb, sizeof(dump_size));
>>> + memdump_buf = vmalloc(dump_size);
>>> + qca_memdump->memdump_buf_head = memdump_buf;
>>> + qca_memdump->memdump_buf_tail = memdump_buf;
>>> + }
>>> +
>>> + memdump_buf = qca_memdump->memdump_buf_tail;
>>> +
>>> + /* If sequence no 0 is missed then there is no point in
>>> + * accepting the other sequences.
>>> + */
>>> + if (!memdump_buf) {
>>> + bt_dev_err(hu->hdev, "QCA: Discarding other packets");
>>> + kfree(qca_memdump);
>>> + kfree_skb(skb);
>>> + qca->qca_memdump = NULL;
>>> + return;
>>> + }
>>> +
>>> + /* There could be chance of missing some packets from
>>> + * the controller. In such cases let us store the dummy
>>> + * packets in the buffer.
>>> + */
>>> + while ((seq_no > qca_memdump->current_seq_no + 1) &&
>>> + seq_no != QCA_LAST_SEQUENCE_NUM) {
>>> + bt_dev_err(hu->hdev, "QCA controller missed packet:%d",
>>> + qca_memdump->current_seq_no);
>>> + memcpy(memdump_buf, nullBuff, QCA_DUMP_PACKET_SIZE);
>>> + memdump_buf = memdump_buf + QCA_DUMP_PACKET_SIZE;
>>> + qca_memdump->received_dump += QCA_DUMP_PACKET_SIZE;
>>> + qca_memdump->current_seq_no++;
>>> + }
>>> +
>>> + memcpy(memdump_buf, (unsigned char *) skb->data, skb->len);
>>> + memdump_buf = memdump_buf + skb->len;
>>> + qca_memdump->memdump_buf_tail = memdump_buf;
>>> + qca_memdump->current_seq_no = seq_no + 1;
>>> + qca_memdump->received_dump += skb->len;
>>> + qca->qca_memdump = qca_memdump;
>>> + kfree_skb(skb);
>>> + if (seq_no == QCA_LAST_SEQUENCE_NUM) {
>>> + bt_dev_info(hu->hdev, "QCA writing crash dump of size %d bytes",
>>> + qca_memdump->received_dump);
>>> + memdump_buf = qca_memdump->memdump_buf_head;
>>> + dev_coredumpv(&hu->serdev->dev, memdump_buf,
>>> + qca_memdump->received_dump, GFP_KERNEL);
>>> + // Revisit for free(memdump) if needed.
>>> + del_timer(&qca->memdump_timer);
>>> + kfree(qca->qca_memdump);
>>> + qca->qca_memdump = NULL;
>>> + qca->memdump_state = QCA_MEMDUMP_COLLECTED;
>>> + }
>>> + }
>>> +
>>> +}
>>> +
>>> +int qca_controller_memdump_event(struct hci_dev *hdev, struct sk_buff *skb)
>>> +{
>>> + struct hci_uart *hu = hci_get_drvdata(hdev);
>>> + struct qca_data *qca = hu->priv;
>>> +
>>> + skb_queue_tail(&qca->rx_memdump_q, skb);
>>> + queue_work(qca->workqueue, &qca->ctrl_memdump_evt);
>>> +
>>> + return 0;
>>> +}
>>> +
>>> +static int qca_recv_event_data(struct hci_dev *hdev, struct sk_buff *skb)
>>> +{
>>> + /* We receive chip memory dump as an event packet, With a dedicated
>>> + * handler followed by a hardware error event. When this event is
>>> + * received we store dump into a file before closing hci. This
>>> + * dump will help in triaging the issues.
>>> + */
>>> + if ((skb->data[0] == HCI_VENDOR_PKT) &&
>>> + (get_unaligned_be16((skb->data) + 2) == QCA_SSR_DUMP_HANDLE))
>> The (skb->data) is pointless, just do skb->data.
> [Bala]: will update.
>
>>> + return qca_controller_memdump_event(hdev, skb);
>>> +
>>> + return hci_recv_frame(hdev, skb);
>>> +}
>>> +
>>> #define QCA_IBS_SLEEP_IND_EVENT \
>>> .type = HCI_IBS_SLEEP_IND, \
>>> .hlen = 0, \
>>> @@ -888,12 +1089,12 @@ static int qca_recv_acl_data(struct hci_dev *hdev, struct sk_buff *skb)
>>> .maxlen = HCI_MAX_IBS_SIZE
>>> static const struct h4_recv_pkt qca_recv_pkts[] = {
>>> - { H4_RECV_ACL, .recv = qca_recv_acl_data },
>>> - { H4_RECV_SCO, .recv = hci_recv_frame },
>>> - { H4_RECV_EVENT, .recv = hci_recv_frame },
>>> - { QCA_IBS_WAKE_IND_EVENT, .recv = qca_ibs_wake_ind },
>>> - { QCA_IBS_WAKE_ACK_EVENT, .recv = qca_ibs_wake_ack },
>>> - { QCA_IBS_SLEEP_IND_EVENT, .recv = qca_ibs_sleep_ind },
>>> + { H4_RECV_ACL, .recv = qca_recv_acl_data },
>>> + { H4_RECV_SCO, .recv = hci_recv_frame },
>>> + { H4_RECV_EVENT, .recv = qca_recv_event_data },
>> call it qca_recv_event and don't realign the whole table.
> [Bala]: will update.
>
>>> + { QCA_IBS_WAKE_IND_EVENT, .recv = qca_ibs_wake_ind },
>>> + { QCA_IBS_WAKE_ACK_EVENT, .recv = qca_ibs_wake_ack },
>>> + { QCA_IBS_SLEEP_IND_EVENT, .recv = qca_ibs_sleep_ind },
>>> };
>>> static int qca_recv(struct hci_uart *hu, const void *data, int count)
>>> @@ -1114,6 +1315,68 @@ static int qca_set_speed(struct hci_uart *hu, enum qca_speed_type speed_type)
>>> return 0;
>>> }
>>> +static int qca_send_crashbuffer(struct hci_uart *hu)
>>> +{
>>> + struct qca_data *qca = hu->priv;
>>> + struct sk_buff *skb;
>>> + unsigned char buf[QCA_CRASHBYTE_PACKET_LEN];
>>> +
>>> + bt_dev_dbg(hu->hdev, "sending soc crash buffer to controller");
>>> +
>>> + /* We forcefully crash the controller, by sending 0xfb byte for
>>> + * 1024 times. We also might have chance of losing data, To be
>>> + * on safer side we send 1100 bytes to the SoC.
>>> + */
>>> + memset(buf, QCA_MEMDUMP_BYTE, QCA_CRASHBYTE_PACKET_LEN);
>>> +
>>> + skb = bt_skb_alloc(QCA_CRASHBYTE_PACKET_LEN, GFP_KERNEL);
>>> + if (!skb) {
>>> + bt_dev_err(hu->hdev, "Failed to allocate memory for skb packet");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + skb_put_data(skb, buf, QCA_CRASHBYTE_PACKET_LEN);
>>> + hci_skb_pkt_type(skb) = HCI_COMMAND_PKT;
>>> +
>>> + skb_queue_tail(&qca->txq, skb);
>>> + hci_uart_tx_wakeup(hu);
>>> +
>>> + return 0;
>>> +}
>>> +
>>> +static void qca_hw_error(struct hci_dev *hdev, u8 code)
>>> +{
>>> + struct hci_uart *hu = hci_get_drvdata(hdev);
>>> + struct qca_data *qca = hu->priv;
>>> +
>>> + bt_dev_info(hdev, "mem_dump_status: %d", qca->memdump_state);
>>> +
>>> + if (qca->memdump_state == QCA_MEMDUMP_IDLE) {
>>> + /* If hardware error event received for other than QCA
>>> + * soc memory dump event, then we need to crash the SOC
>>> + * and wait here for 8 seconds to get the dump packets.
>>> + * This will block main thread to be on hold until we
>>> + * collect dump.
>>> + */
>>> + mod_timer(&qca->memdump_timer,
>>> + (jiffies + msecs_to_jiffies(qca->memdump_delay)));
>> Why a timer that then ends up in interrupt context?
> [Bala]: some times we only receive hardware error from the chip, so we are sending an special buffer to request the
> chip for memory dump. we are using an timer to check whether the data memdump is collected. if the timer
> over flows we are cleaning up all the assigned memory and unblock the main thread,
> in short this timer is a watch dog timer for the memory dump collection.

I saw that you used a bunch of timers actually. My question is why are these not delayed_wq instead? I wonder if we really want to keep using timers and handling interrupt context locking if we could move everything to wq instead.

>>> + qca_send_crashbuffer(hu);
>>> +
>>> + while (qca->memdump_state != QCA_MEMDUMP_COLLECTED &&
>>> + qca->memdump_state != QCA_MEMDUMP_TIMEOUT)
>>> + ;
>> What is this? Busy loop? I rather not do that.
> [Bala]: this busy required to collected the dump from the chip before we close an reopen the hci.
> >> + } else if (qca->memdump_state == QCA_MEMDUMP_COLLECTING) {

Why not use some wakeup handling here. I am really against any kind of busy loops in a driver.

>>> + /* Let us wait here until memory dump collected or
>>> + * memory dump timer expired.
>>> + */
>>> + bt_dev_info(hdev, "waiting for dump to complete");
>>> + while (qca->memdump_state == QCA_MEMDUMP_COLLECTING)
>>> + ;
>>> + }
>>> +
>>> + qca->memdump_state = QCA_MEMDUMP_IDLE;
>>> +}
>>> +
>>> static int qca_wcn3990_init(struct hci_uart *hu)
>>> {
>>> struct hci_dev *hdev = hu->hdev;
>>> @@ -1229,6 +1492,10 @@ static int qca_setup(struct hci_uart *hu)
>>> if (!ret) {
>>> set_bit(STATE_IN_BAND_SLEEP_ENABLED, &qca->flags);
>>> qca_debugfs_init(hdev);
>>> + /* We only get controller dump when fw download is
>>> + * successful.
>>> + */
>>> + hu->hdev->hw_error = qca_hw_error;
>> I would set this all the time and handle it with a flag if firmware
>> downloaded has succeeded or not.
> [Bala]: ok will update.
>
>>> } else if (ret == -ENOENT) {
>>> /* No patch/nvm-config found, run with original fw/config */
>>> ret = 0;

Regards

Marcel