Re: [PATCH 3/3] mailbox: zynqmp: Enable Bufferless IPI usage on Versal-based SOC's
From: Ben Levinsky
Date: Wed Jan 03 2024 - 09:40:54 EST
Hi Jassi,
Please review when you can.
Thanks and Kind Regards,
Ben
On 12/20/23 9:42 AM, Ben Levinsky wrote:
> On 12/20/23 5:29 AM, Michal Simek wrote:
>>
>> On 12/14/23 22:13, Ben Levinsky wrote:
>>> On Xilinx-AMD Versal and Versal-NET, there exist both
>>> inter-processor-interrupts with corresponding message buffers and without
>>> such buffers.
>>>
>>> Add a routine that, if the corresponding DT compatible
>>> string "xlnx,versal-ipi-mailbox" is used then a Versal-based SOC
>>> can use a mailbox Device Tree entry where both host and remote
>>> can use either of the buffered or bufferless interrupts.
>>>
>>> Signed-off-by: Ben Levinsky <ben.levinsky@xxxxxxx>
>>> ---
>>> Note that the linked patch provides corresponding bindings.
>>> Depends on: https://lore.kernel.org/all/20231214054224.957336-3-tanmay.shah@xxxxxxx/T/
>>> ---
>>> drivers/mailbox/zynqmp-ipi-mailbox.c | 146 +++++++++++++++++++++++++--
>>> 1 file changed, 139 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/drivers/mailbox/zynqmp-ipi-mailbox.c b/drivers/mailbox/zynqmp-ipi-mailbox.c
>>> index edefb80a6e47..316d9406064e 100644
>>> --- a/drivers/mailbox/zynqmp-ipi-mailbox.c
>>> +++ b/drivers/mailbox/zynqmp-ipi-mailbox.c
>>> @@ -52,6 +52,13 @@
>>> #define IPI_MB_CHNL_TX 0 /* IPI mailbox TX channel */
>>> #define IPI_MB_CHNL_RX 1 /* IPI mailbox RX channel */
>>> +/* IPI Message Buffer Information */
>>> +#define RESP_OFFSET 0x20U
>>> +#define DEST_OFFSET 0x40U
>>> +#define IPI_BUF_SIZE 0x20U
>>> +#define DST_BIT_POS 9U
>>> +#define SRC_BITMASK GENMASK(11, 8)
>>> +
>>> /**
>>> * struct zynqmp_ipi_mchan - Description of a Xilinx ZynqMP IPI mailbox channel
>>> * @is_opened: indicate if the IPI channel is opened
>>> @@ -170,9 +177,11 @@ static irqreturn_t zynqmp_ipi_interrupt(int irq, void *data)
>>> if (ret > 0 && ret & IPI_MB_STATUS_RECV_PENDING) {
>>> if (mchan->is_opened) {
>>> msg = mchan->rx_buf;
>>> - msg->len = mchan->req_buf_size;
>>> - memcpy_fromio(msg->data, mchan->req_buf,
>>> - msg->len);
>>> + if (msg) {
>>> + msg->len = mchan->req_buf_size;
>>> + memcpy_fromio(msg->data, mchan->req_buf,
>>> + msg->len);
>>> + }
>>> mbox_chan_received_data(chan, (void *)msg);
>>> status = IRQ_HANDLED;
>>> }
>>> @@ -282,26 +291,26 @@ static int zynqmp_ipi_send_data(struct mbox_chan *chan, void *data)
>>> if (mchan->chan_type == IPI_MB_CHNL_TX) {
>>> /* Send request message */
>>> - if (msg && msg->len > mchan->req_buf_size) {
>>> + if (msg && msg->len > mchan->req_buf_size && mchan->req_buf) {
>>> dev_err(dev, "channel %d message length %u > max %lu\n",
>>> mchan->chan_type, (unsigned int)msg->len,
>>> mchan->req_buf_size);
>>> return -EINVAL;
>>> }
>>> - if (msg && msg->len)
>>> + if (msg && msg->len && mchan->req_buf)
>>> memcpy_toio(mchan->req_buf, msg->data, msg->len);
>>> /* Kick IPI mailbox to send message */
>>> arg0 = SMC_IPI_MAILBOX_NOTIFY;
>>> zynqmp_ipi_fw_call(ipi_mbox, arg0, 0, &res);
>>> } else {
>>> /* Send response message */
>>> - if (msg && msg->len > mchan->resp_buf_size) {
>>> + if (msg && msg->len > mchan->resp_buf_size && mchan->resp_buf) {
>>> dev_err(dev, "channel %d message length %u > max %lu\n",
>>> mchan->chan_type, (unsigned int)msg->len,
>>> mchan->resp_buf_size);
>>> return -EINVAL;
>>> }
>>> - if (msg && msg->len)
>>> + if (msg && msg->len && mchan->resp_buf)
>>> memcpy_toio(mchan->resp_buf, msg->data, msg->len);
>>> arg0 = SMC_IPI_MAILBOX_ACK;
>>> zynqmp_ipi_fw_call(ipi_mbox, arg0, IPI_SMC_ACK_EIRQ_MASK,
>>> @@ -640,6 +649,126 @@ static int zynqmp_ipi_setup(struct zynqmp_ipi_mbox *ipi_mbox,
>>> return 0;
>>> }
>>> +/**
>>> + * versal_ipi_setup - Set up IPIs to support mixed usage of
>>> + * Buffered and Bufferless IPIs.
>>> + *
>>> + * @ipi_mbox: pointer to IPI mailbox private data structure
>>> + * @node: IPI mailbox device node
>>> + *
>>> + * Return: 0 for success, negative value for failure
>>> + */
>>> +static int versal_ipi_setup(struct zynqmp_ipi_mbox *ipi_mbox,
>>> + struct device_node *node)
>>> +{
>>> + struct zynqmp_ipi_mchan *tx_mchan, *rx_mchan;
>>> + struct resource host_res, remote_res;
>>> + struct device_node *parent_node;
>>> + int host_idx, remote_idx;
>>> + struct device *mdev, *dev;
>>> +
>>> + tx_mchan = &ipi_mbox->mchans[IPI_MB_CHNL_TX];
>>> + rx_mchan = &ipi_mbox->mchans[IPI_MB_CHNL_RX];
>>> + parent_node = of_get_parent(node);
>>> + dev = ipi_mbox->pdata->dev;
>>> + mdev = &ipi_mbox->dev;
>>> +
>>> + host_idx = zynqmp_ipi_mbox_get_buf_res(parent_node, "msg", &host_res);
>>> + remote_idx = zynqmp_ipi_mbox_get_buf_res(node, "msg", &remote_res);
>>> +
>>> + /*
>>> + * Only set up buffers if both sides claim to have msg buffers.
>>> + * This is because each buffered IPI's corresponding msg buffers
>>> + * are reserved for use by other buffered IPI's.
>>> + */
>>> + if (!host_idx && !remote_idx) {
>>> + u32 host_src, host_dst, remote_src, remote_dst;
>>> + u32 buff_sz;
>>> +
>>> + buff_sz = resource_size(&host_res);
>>> +
>>> + host_src = host_res.start & SRC_BITMASK;
>>> + remote_src = remote_res.start & SRC_BITMASK;
>>> +
>>> + host_dst = (host_src >> DST_BIT_POS) * DEST_OFFSET;
>>> + remote_dst = (remote_src >> DST_BIT_POS) * DEST_OFFSET;
>>> +
>>> + /* Validate that IPI IDs is within IPI Message buffer space. */
>>> + if (host_dst >= buff_sz || remote_dst >= buff_sz) {
>>> + dev_err(mdev,
>>> + "Invalid IPI Message buffer values: %x %x\n",
>>> + host_dst, remote_dst);
>>> + return -EINVAL;
>>> + }
>>> +
>>> + tx_mchan->req_buf = devm_ioremap(mdev,
>>> + host_res.start | remote_dst,
>>> + IPI_BUF_SIZE);
>>> + if (!tx_mchan->req_buf) {
>>> + dev_err(mdev, "Unable to map IPI buffer I/O memory\n");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + tx_mchan->resp_buf = devm_ioremap(mdev,
>>> + (remote_res.start | host_dst) +
>>> + RESP_OFFSET, IPI_BUF_SIZE);
>>> + if (!tx_mchan->resp_buf) {
>>> + dev_err(mdev, "Unable to map IPI buffer I/O memory\n");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + rx_mchan->req_buf = devm_ioremap(mdev,
>>> + remote_res.start | host_dst,
>>> + IPI_BUF_SIZE);
>>> + if (!rx_mchan->req_buf) {
>>> + dev_err(mdev, "Unable to map IPI buffer I/O memory\n");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + rx_mchan->resp_buf = devm_ioremap(mdev,
>>> + (host_res.start | remote_dst) +
>>> + RESP_OFFSET, IPI_BUF_SIZE);
>>> + if (!rx_mchan->resp_buf) {
>>> + dev_err(mdev, "Unable to map IPI buffer I/O memory\n");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + tx_mchan->resp_buf_size = IPI_BUF_SIZE;
>>> + tx_mchan->req_buf_size = IPI_BUF_SIZE;
>>> + tx_mchan->rx_buf = devm_kzalloc(mdev, IPI_BUF_SIZE +
>>> + sizeof(struct zynqmp_ipi_message),
>>> + GFP_KERNEL);
>>> + if (!tx_mchan->rx_buf)
>>> + return -ENOMEM;
>>> +
>>> + rx_mchan->resp_buf_size = IPI_BUF_SIZE;
>>> + rx_mchan->req_buf_size = IPI_BUF_SIZE;
>>> + rx_mchan->rx_buf = devm_kzalloc(mdev, IPI_BUF_SIZE +
>>> + sizeof(struct zynqmp_ipi_message),
>>> + GFP_KERNEL);
>>> + if (!rx_mchan->rx_buf)
>>> + return -ENOMEM;
>>> + } else {
>>> + /*
>>> + * If here, then set up Bufferless IPI Channel because
>>> + * one or both of the IPI's is bufferless.
>>> + */
>>> + tx_mchan->req_buf = NULL;
>>> + tx_mchan->resp_buf = NULL;
>>> + tx_mchan->rx_buf = NULL;
>>> + tx_mchan->resp_buf_size = 0;
>>> + tx_mchan->req_buf_size = 0;
>>> +
>>> + rx_mchan->req_buf = NULL;
>>> + rx_mchan->resp_buf = NULL;
>>> + rx_mchan->rx_buf = NULL;
>>> + rx_mchan->resp_buf_size = 0;
>>> + rx_mchan->req_buf_size = 0;
>> Just curious if this is really needed. If none fills that values aren't they actually already 0/NULL because that location is cleared by kzalloc.
>>
>> Thanks,
>> Michal
> Confirmed. I removed the whole else condition and it still worked.
>
> Will update in next rev after I receive any pending feedback from Jassi
>