| Message ID | 1506604306-20739-12-git-send-email-sudeep.holla@arm.com |
|---|---|
| State | New |
| Headers | show |
| Series | [v3,01/22] dt-bindings: mailbox: add support for mailbox client shared memory | expand |
On 04/10/17 12:13, Arnd Bergmann wrote: > On Thu, Sep 28, 2017 at 3:11 PM, Sudeep Holla <sudeep.holla@arm.com> wrote: >> It would be useful to have options to perform some SCMI transfers >> atomically by polling for the completion flag instead of interrupt >> driven. The SCMI specification has option to disable the interrupt and >> poll for the completion flag in the shared memory. >> >> This patch adds support for polling based SCMI transfers using that >> option. This might be used for uninterrupted/atomic DVFS operations >> from the scheduler context. > > multi-millisecond timeouts from inside the scheduler sound like a > really bad idea. Could this maybe get changed to an asynchronous > operation? > We already support asynchronous version. This is mainly to support fast DVFS switching done at context switch. I can reduce the timeouts to few uS as the whole idea of fast dvfs is we can't return or schedule out of this function. Typically the remote should see the request in < 10 uS and acknowledge it. >> + if (xfer->hdr.poll_completion) { >> + timeout = info->desc->max_rx_timeout_ms * 100; >> + while (!scmi_xfer_poll_done(info, xfer) && timeout--) >> + udelay(10); > > The timeout calculation is bad as well, since both the > scmi_xfer_poll_done() call and udelay() can take much longer > than the 10 microsecond delay that you use for the calculation. > Ah, agreed. I will change it to few uS as that's what is expected. > If you want to do a timeout check like this, it should generally > be done using ktime_get()/ktime_add()/ktime_before(). > Good idea, will try to use that. -- Regards, Sudeep
diff --git a/drivers/firmware/arm_scmi/driver.c b/drivers/firmware/arm_scmi/driver.c index a1abcf2049ca..e1abedf15cb0 100644 --- a/drivers/firmware/arm_scmi/driver.c +++ b/drivers/firmware/arm_scmi/driver.c @@ -26,6 +26,7 @@ */ #include <linux/bitmap.h> +#include <linux/delay.h> #include <linux/export.h> #include <linux/io.h> #include <linux/kernel.h> @@ -369,6 +370,20 @@ void scmi_one_xfer_put(const struct scmi_handle *handle, struct scmi_xfer *xfer) up(&minfo->sem_xfer_count); } +static bool +scmi_xfer_poll_done(const struct scmi_info *info, struct scmi_xfer *xfer) +{ + struct scmi_shared_mem *mem = info->tx_payload; + u16 xfer_id = MSG_XTRACT_TOKEN(le32_to_cpu(mem->msg_header)); + + if (xfer->hdr.seq != xfer_id) + return false; + + return le32_to_cpu(mem->channel_status) & + (SCMI_SHMEM_CHAN_STAT_CHANNEL_ERROR | + SCMI_SHMEM_CHAN_STAT_CHANNEL_FREE); +} + /** * scmi_do_xfer() - Do one transfer * @@ -395,14 +410,24 @@ int scmi_do_xfer(const struct scmi_handle *handle, struct scmi_xfer *xfer) /* mbox_send_message returns non-negative value on success, so reset */ ret = 0; - /* And we wait for the response. */ - timeout = msecs_to_jiffies(info->desc->max_rx_timeout_ms); - if (!wait_for_completion_timeout(&xfer->done, timeout)) { - dev_err(dev, "mbox timed out in resp(caller: %pF)\n", - (void *)_RET_IP_); - ret = -ETIMEDOUT; - } else if (xfer->hdr.status) { - ret = scmi_to_linux_errno(xfer->hdr.status); + if (xfer->hdr.poll_completion) { + timeout = info->desc->max_rx_timeout_ms * 100; + while (!scmi_xfer_poll_done(info, xfer) && timeout--) + udelay(10); + if (timeout) + scmi_fetch_response(xfer, info->tx_payload); + else + ret = -ETIMEDOUT; + } else { + /* And we wait for the response. */ + timeout = msecs_to_jiffies(info->desc->max_rx_timeout_ms); + if (!wait_for_completion_timeout(&xfer->done, timeout)) { + dev_err(dev, "mbox timed out in resp(caller: %pF)\n", + (void *)_RET_IP_); + ret = -ETIMEDOUT; + } else if (xfer->hdr.status) { + ret = scmi_to_linux_errno(xfer->hdr.status); + } } /* * NOTE: we might prefer not to need the mailbox ticker to manage the
It would be useful to have options to perform some SCMI transfers atomically by polling for the completion flag instead of interrupt driven. The SCMI specification has option to disable the interrupt and poll for the completion flag in the shared memory. This patch adds support for polling based SCMI transfers using that option. This might be used for uninterrupted/atomic DVFS operations from the scheduler context. Cc: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Sudeep Holla <sudeep.holla@arm.com> --- drivers/firmware/arm_scmi/driver.c | 41 ++++++++++++++++++++++++++++++-------- 1 file changed, 33 insertions(+), 8 deletions(-) -- 2.7.4