| Message ID | 20210507131756.17028-2-patrice.chotard@foss.st.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | MTD: spinand: Add spi_mem_poll_status() support | expand |
On Fri, 7 May 2021 15:17:54 +0200 <patrice.chotard@foss.st.com> wrote: > From: Patrice Chotard <patrice.chotard@foss.st.com> > > With STM32 QSPI, it is possible to poll the status register of the device. > This could be done to offload the CPU during an operation (erase or > program a SPI NAND for example). > > spi_mem_poll_status API has been added to handle this feature. > This new function take care of the offload/non-offload cases. > > For the non-offload case, use read_poll_timeout() to poll the status in > order to release CPU during this phase. > > Signed-off-by: Patrice Chotard <patrice.chotard@foss.st.com> > Signed-off-by: Christophe Kerello <christophe.kerello@foss.st.com> > --- > Changes in v2: > - Indicates the spi_mem_poll_status() timeout unit > - Use 2-byte wide status register > - Add spi_mem_supports_op() call in spi_mem_poll_status() > - Add completion management in spi_mem_poll_status() > - Add offload/non-offload case mangement in spi_mem_poll_status() > - Optimize the non-offload case by using read_poll_timeout() > > drivers/spi/spi-mem.c | 71 +++++++++++++++++++++++++++++++++++++ > include/linux/spi/spi-mem.h | 10 ++++++ > 2 files changed, 81 insertions(+) > > diff --git a/drivers/spi/spi-mem.c b/drivers/spi/spi-mem.c > index 1513553e4080..3f29c604df7d 100644 > --- a/drivers/spi/spi-mem.c > +++ b/drivers/spi/spi-mem.c > @@ -6,6 +6,7 @@ > * Author: Boris Brezillon <boris.brezillon@bootlin.com> > */ > #include <linux/dmaengine.h> > +#include <linux/iopoll.h> > #include <linux/pm_runtime.h> > #include <linux/spi/spi.h> > #include <linux/spi/spi-mem.h> > @@ -743,6 +744,75 @@ static inline struct spi_mem_driver *to_spi_mem_drv(struct device_driver *drv) > return container_of(drv, struct spi_mem_driver, spidrv.driver); > } > > +/** > + * spi_mem_finalize_op - report completion of spi_mem_op > + * @ctlr: the controller reporting completion > + * > + * Called by SPI drivers using the spi-mem spi_mem_poll_status() > + * implementation to notify it that the current spi_mem_op has > + * finished. > + */ > +void spi_mem_finalize_op(struct spi_controller *ctlr) > +{ > + complete(&ctlr->xfer_completion); > +} > +EXPORT_SYMBOL_GPL(spi_mem_finalize_op); > + > +/** > + * spi_mem_poll_status() - Poll memory device status > + * @mem: SPI memory device > + * @op: the memory operation to execute > + * @mask: status bitmask to ckeck > + * @match: (status & mask) expected value > + * @timeout_ms: timeout in milliseconds > + * > + * This function send a polling status request to the controller driver > + * > + * Return: 0 in case of success, -ETIMEDOUT in case of error, > + * -EOPNOTSUPP if not supported. > + */ > +int spi_mem_poll_status(struct spi_mem *mem, > + const struct spi_mem_op *op, > + u16 mask, u16 match, u16 timeout_ms) > +{ > + struct spi_controller *ctlr = mem->spi->controller; > + unsigned long ms; > + int ret = -EOPNOTSUPP; > + int exec_op_ret; > + u16 *status; > + > + if (!spi_mem_supports_op(mem, op)) > + return ret; You should only test that in the SW-based polling path. The driver ->poll_status() method is expected to check the operation and return -EOPNOTSUPP if HW-based polling doesn't work for this op, no need to check things twice. > + > + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { > + ret = spi_mem_access_start(mem); > + if (ret) > + return ret; > + > + reinit_completion(&ctlr->xfer_completion); > + > + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, > + timeout_ms); > + > + ms = wait_for_completion_timeout(&ctlr->xfer_completion, > + msecs_to_jiffies(timeout_ms)); Why do you need to wait here? I'd expect the poll_status to take care of this wait. > + > + spi_mem_access_end(mem); > + if (!ms) > + return -ETIMEDOUT; > + } else { > + status = (u16 *)op->data.buf.in; Hm, I don't think it's safe, for 2 reasons: 1/ op->data.buf.in might be a 1byte buffer, but you're doing a 2byte check 2/ data is in big endian in the SPI buffer, which means your check won't work on LE architectures. You really need a dedicated spi_mem_read_status() function that's passed an u16 pointer: int spi_mem_read_status(struct spi_mem *mem, const struct spi_mem_op *op, u16 *status) { const u8 *bytes = (u8 *)op->data.buf.in; int ret; ret = spi_mem_exec_op(mem, op); if (ret) return ret; if (op->data.nbytes > 1) *status = ((u16)bytes[0] << 8) | bytes[1]; else *status = bytes[0]; return 0; } > + ret = read_poll_timeout(spi_mem_exec_op, exec_op_ret, > + ((*status) & mask) == match, 20, > + timeout_ms * 1000, false, mem, op); > + if (exec_op_ret) > + return exec_op_ret; > + } > + I would do something like this instead: int spi_mem_poll_status(struct spi_mem *mem, const struct spi_mem_op *op, u16 mask, u16 match, u16 timeout_ms) { struct spi_controller *ctlr = mem->spi->controller; int ret = -EOPNOTSUPP; if (op->data.nbytes < 1 || op->data.nbytes > 2) return -EINVAL; if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { ret = spi_mem_access_start(mem); if (ret) return ret; ret = ctlr->mem_ops->poll_status(mem, op, mask, match, timeout_ms); spi_mem_access_end(mem); } if (ret == -EOPNOTSUPP) { u16 status; int read_status_ret; if (!spi_mem_supports_op(mem, op)) return -EOPNOTSUPP; ret = read_poll_timeout(spi_mem_read_status, exec_op_ret, (read_status_ret || ((status & mask) == match), 20, timeout_ms * 1000, false, mem, op, &status); if (read_status_ret) return read_status_ret; } return ret; } > + return ret; > +} > +EXPORT_SYMBOL_GPL(spi_mem_poll_status); > + > static int spi_mem_probe(struct spi_device *spi) > { > struct spi_mem_driver *memdrv = to_spi_mem_drv(spi->dev.driver); > @@ -763,6 +833,7 @@ static int spi_mem_probe(struct spi_device *spi) > if (IS_ERR_OR_NULL(mem->name)) > return PTR_ERR_OR_ZERO(mem->name); > > + init_completion(&ctlr->xfer_completion); > spi_set_drvdata(spi, mem); > > return memdrv->probe(mem); > diff --git a/include/linux/spi/spi-mem.h b/include/linux/spi/spi-mem.h > index 2b65c9edc34e..0fbf5d0a3d31 100644 > --- a/include/linux/spi/spi-mem.h > +++ b/include/linux/spi/spi-mem.h > @@ -250,6 +250,7 @@ static inline void *spi_mem_get_drvdata(struct spi_mem *mem) > * the currently mapped area), and the caller of > * spi_mem_dirmap_write() is responsible for calling it again in > * this case. > + * @poll_status: poll memory device status > * > * This interface should be implemented by SPI controllers providing an > * high-level interface to execute SPI memory operation, which is usually the > @@ -274,6 +275,9 @@ struct spi_controller_mem_ops { > u64 offs, size_t len, void *buf); > ssize_t (*dirmap_write)(struct spi_mem_dirmap_desc *desc, > u64 offs, size_t len, const void *buf); > + int (*poll_status)(struct spi_mem *mem, > + const struct spi_mem_op *op, > + u16 mask, u16 match, unsigned long timeout); > }; > > /** > @@ -369,6 +373,12 @@ devm_spi_mem_dirmap_create(struct device *dev, struct spi_mem *mem, > void devm_spi_mem_dirmap_destroy(struct device *dev, > struct spi_mem_dirmap_desc *desc); > > +void spi_mem_finalize_op(struct spi_controller *ctlr); > + > +int spi_mem_poll_status(struct spi_mem *mem, > + const struct spi_mem_op *op, > + u16 mask, u16 match, u16 timeout); > + > int spi_mem_driver_register_with_owner(struct spi_mem_driver *drv, > struct module *owner); >
Hi Boris On 5/8/21 9:55 AM, Boris Brezillon wrote: > On Fri, 7 May 2021 15:17:54 +0200 > <patrice.chotard@foss.st.com> wrote: > >> From: Patrice Chotard <patrice.chotard@foss.st.com> >> >> With STM32 QSPI, it is possible to poll the status register of the device. >> This could be done to offload the CPU during an operation (erase or >> program a SPI NAND for example). >> >> spi_mem_poll_status API has been added to handle this feature. >> This new function take care of the offload/non-offload cases. >> >> For the non-offload case, use read_poll_timeout() to poll the status in >> order to release CPU during this phase. >> >> Signed-off-by: Patrice Chotard <patrice.chotard@foss.st.com> >> Signed-off-by: Christophe Kerello <christophe.kerello@foss.st.com> >> --- >> Changes in v2: >> - Indicates the spi_mem_poll_status() timeout unit >> - Use 2-byte wide status register >> - Add spi_mem_supports_op() call in spi_mem_poll_status() >> - Add completion management in spi_mem_poll_status() >> - Add offload/non-offload case mangement in spi_mem_poll_status() >> - Optimize the non-offload case by using read_poll_timeout() >> >> drivers/spi/spi-mem.c | 71 +++++++++++++++++++++++++++++++++++++ >> include/linux/spi/spi-mem.h | 10 ++++++ >> 2 files changed, 81 insertions(+) >> >> diff --git a/drivers/spi/spi-mem.c b/drivers/spi/spi-mem.c >> index 1513553e4080..3f29c604df7d 100644 >> --- a/drivers/spi/spi-mem.c >> +++ b/drivers/spi/spi-mem.c >> @@ -6,6 +6,7 @@ >> * Author: Boris Brezillon <boris.brezillon@bootlin.com> >> */ >> #include <linux/dmaengine.h> >> +#include <linux/iopoll.h> >> #include <linux/pm_runtime.h> >> #include <linux/spi/spi.h> >> #include <linux/spi/spi-mem.h> >> @@ -743,6 +744,75 @@ static inline struct spi_mem_driver *to_spi_mem_drv(struct device_driver *drv) >> return container_of(drv, struct spi_mem_driver, spidrv.driver); >> } >> >> +/** >> + * spi_mem_finalize_op - report completion of spi_mem_op >> + * @ctlr: the controller reporting completion >> + * >> + * Called by SPI drivers using the spi-mem spi_mem_poll_status() >> + * implementation to notify it that the current spi_mem_op has >> + * finished. >> + */ >> +void spi_mem_finalize_op(struct spi_controller *ctlr) >> +{ >> + complete(&ctlr->xfer_completion); >> +} >> +EXPORT_SYMBOL_GPL(spi_mem_finalize_op); >> + >> +/** >> + * spi_mem_poll_status() - Poll memory device status >> + * @mem: SPI memory device >> + * @op: the memory operation to execute >> + * @mask: status bitmask to ckeck >> + * @match: (status & mask) expected value >> + * @timeout_ms: timeout in milliseconds >> + * >> + * This function send a polling status request to the controller driver >> + * >> + * Return: 0 in case of success, -ETIMEDOUT in case of error, >> + * -EOPNOTSUPP if not supported. >> + */ >> +int spi_mem_poll_status(struct spi_mem *mem, >> + const struct spi_mem_op *op, >> + u16 mask, u16 match, u16 timeout_ms) >> +{ >> + struct spi_controller *ctlr = mem->spi->controller; >> + unsigned long ms; >> + int ret = -EOPNOTSUPP; >> + int exec_op_ret; >> + u16 *status; >> + >> + if (!spi_mem_supports_op(mem, op)) >> + return ret; > > You should only test that in the SW-based polling path. The driver > ->poll_status() method is expected to check the operation and > return -EOPNOTSUPP if HW-based polling doesn't work for this op, > no need to check things twice. Ok, i will fix this. > >> + >> + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { >> + ret = spi_mem_access_start(mem); >> + if (ret) >> + return ret; >> + >> + reinit_completion(&ctlr->xfer_completion); >> + >> + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, >> + timeout_ms); >> + >> + ms = wait_for_completion_timeout(&ctlr->xfer_completion, >> + msecs_to_jiffies(timeout_ms)); > > Why do you need to wait here? I'd expect the poll_status to take care > of this wait. It was a request from Mark Brown [1]. The idea is to implement similar mechanism already used in SPI framework. [1] https://patchwork.kernel.org/project/linux-arm-kernel/patch/20210426143934.25275-2-patrice.chotard@foss.st.com/#24140527 > >> + >> + spi_mem_access_end(mem); >> + if (!ms) >> + return -ETIMEDOUT; >> + } else { >> + status = (u16 *)op->data.buf.in; > > Hm, I don't think it's safe, for 2 reasons: > > 1/ op->data.buf.in might be a 1byte buffer, but you're doing a 2byte check > 2/ data is in big endian in the SPI buffer, which means your check > won't work on LE architectures. > > You really need a dedicated spi_mem_read_status() function that's passed > an u16 pointer: Yes, agree > > int spi_mem_read_status(struct spi_mem *mem, > const struct spi_mem_op *op, > u16 *status) > { > const u8 *bytes = (u8 *)op->data.buf.in; > int ret; > > ret = spi_mem_exec_op(mem, op); > if (ret) > return ret; > > if (op->data.nbytes > 1) > *status = ((u16)bytes[0] << 8) | bytes[1]; > else > *status = bytes[0]; > > return 0; > } > >> + ret = read_poll_timeout(spi_mem_exec_op, exec_op_ret, >> + ((*status) & mask) == match, 20, >> + timeout_ms * 1000, false, mem, op); >> + if (exec_op_ret) >> + return exec_op_ret; >> + } >> + > > I would do something like this instead: > > int spi_mem_poll_status(struct spi_mem *mem, > const struct spi_mem_op *op, > u16 mask, u16 match, u16 timeout_ms) > { > struct spi_controller *ctlr = mem->spi->controller; > int ret = -EOPNOTSUPP; > > if (op->data.nbytes < 1 || op->data.nbytes > 2) > return -EINVAL; > > if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { > ret = spi_mem_access_start(mem); > if (ret) > return ret; > > ret = ctlr->mem_ops->poll_status(mem, op, mask, match, > timeout_ms); > > spi_mem_access_end(mem); > } > > > if (ret == -EOPNOTSUPP) { > u16 status; > int read_status_ret; > > if (!spi_mem_supports_op(mem, op)) > return -EOPNOTSUPP; > > ret = read_poll_timeout(spi_mem_read_status, exec_op_ret, > (read_status_ret || ((status & mask) == match), 20, > timeout_ms * 1000, false, mem, op, &status); > > if (read_status_ret) > return read_status_ret; > } > > return ret; > } > >> + return ret; >> +} >> +EXPORT_SYMBOL_GPL(spi_mem_poll_status); >> + >> static int spi_mem_probe(struct spi_device *spi) >> { >> struct spi_mem_driver *memdrv = to_spi_mem_drv(spi->dev.driver); >> @@ -763,6 +833,7 @@ static int spi_mem_probe(struct spi_device *spi) >> if (IS_ERR_OR_NULL(mem->name)) >> return PTR_ERR_OR_ZERO(mem->name); >> >> + init_completion(&ctlr->xfer_completion); >> spi_set_drvdata(spi, mem); >> >> return memdrv->probe(mem); >> diff --git a/include/linux/spi/spi-mem.h b/include/linux/spi/spi-mem.h >> index 2b65c9edc34e..0fbf5d0a3d31 100644 >> --- a/include/linux/spi/spi-mem.h >> +++ b/include/linux/spi/spi-mem.h >> @@ -250,6 +250,7 @@ static inline void *spi_mem_get_drvdata(struct spi_mem *mem) >> * the currently mapped area), and the caller of >> * spi_mem_dirmap_write() is responsible for calling it again in >> * this case. >> + * @poll_status: poll memory device status >> * >> * This interface should be implemented by SPI controllers providing an >> * high-level interface to execute SPI memory operation, which is usually the >> @@ -274,6 +275,9 @@ struct spi_controller_mem_ops { >> u64 offs, size_t len, void *buf); >> ssize_t (*dirmap_write)(struct spi_mem_dirmap_desc *desc, >> u64 offs, size_t len, const void *buf); >> + int (*poll_status)(struct spi_mem *mem, >> + const struct spi_mem_op *op, >> + u16 mask, u16 match, unsigned long timeout); >> }; >> >> /** >> @@ -369,6 +373,12 @@ devm_spi_mem_dirmap_create(struct device *dev, struct spi_mem *mem, >> void devm_spi_mem_dirmap_destroy(struct device *dev, >> struct spi_mem_dirmap_desc *desc); >> >> +void spi_mem_finalize_op(struct spi_controller *ctlr); >> + >> +int spi_mem_poll_status(struct spi_mem *mem, >> + const struct spi_mem_op *op, >> + u16 mask, u16 match, u16 timeout); >> + >> int spi_mem_driver_register_with_owner(struct spi_mem_driver *drv, >> struct module *owner); >> > Thanks Patrice
On Mon, 10 May 2021 10:46:48 +0200 Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > > > >> + > >> + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { > >> + ret = spi_mem_access_start(mem); > >> + if (ret) > >> + return ret; > >> + > >> + reinit_completion(&ctlr->xfer_completion); > >> + > >> + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, > >> + timeout_ms); > >> + > >> + ms = wait_for_completion_timeout(&ctlr->xfer_completion, > >> + msecs_to_jiffies(timeout_ms)); > > > > Why do you need to wait here? I'd expect the poll_status to take care > > of this wait. > > It was a request from Mark Brown [1]. The idea is to implement > similar mechanism already used in SPI framework. Well, you have to choose, either you pass a timeout to ->poll_status() and let the driver wait for the status change (and return -ETIMEDOUT if it didn't happen in time), or you do it here and the driver only has to signal the core completion object. I think it's preferable to let the driver handle the timeout though, because you don't know how the status check will be implemented, and it's not like the reinit_completion()+wait_for_completion_timeout() done here would greatly simplify the drivers wait logic anyway.
Hi Boris On 5/10/21 11:22 AM, Boris Brezillon wrote: > On Mon, 10 May 2021 10:46:48 +0200 > Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > >>> >>>> + >>>> + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { >>>> + ret = spi_mem_access_start(mem); >>>> + if (ret) >>>> + return ret; >>>> + >>>> + reinit_completion(&ctlr->xfer_completion); >>>> + >>>> + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, >>>> + timeout_ms); >>>> + >>>> + ms = wait_for_completion_timeout(&ctlr->xfer_completion, >>>> + msecs_to_jiffies(timeout_ms)); >>> >>> Why do you need to wait here? I'd expect the poll_status to take care >>> of this wait. >> >> It was a request from Mark Brown [1]. The idea is to implement >> similar mechanism already used in SPI framework. > > Well, you have to choose, either you pass a timeout to ->poll_status() > and let the driver wait for the status change (and return -ETIMEDOUT if > it didn't happen in time), or you do it here and the driver only has to > signal the core completion object. I think it's preferable to let the > driver handle the timeout though, because you don't know how the > status check will be implemented, and it's not like the > reinit_completion()+wait_for_completion_timeout() done here would > greatly simplify the drivers wait logic anyway. > Ok i will remove the reinit/wait_completion() as you suggested. Thanks Patrice
On Fri, 7 May 2021 15:17:54 +0200 <patrice.chotard@foss.st.com> wrote: > +/** > + * spi_mem_poll_status() - Poll memory device status > + * @mem: SPI memory device > + * @op: the memory operation to execute > + * @mask: status bitmask to ckeck > + * @match: (status & mask) expected value > + * @timeout_ms: timeout in milliseconds > + * > + * This function send a polling status request to the controller driver > + * > + * Return: 0 in case of success, -ETIMEDOUT in case of error, > + * -EOPNOTSUPP if not supported. > + */ > +int spi_mem_poll_status(struct spi_mem *mem, > + const struct spi_mem_op *op, > + u16 mask, u16 match, u16 timeout_ms) Maybe you should pass a delay_us too, to poll the status at the right rate in the SW-based case (can also be used by drivers if they need to configure the polling rate). You could also add an initial_delay_us to avoid polling the status too early: an erase operation will take longer than a write which will take longer than a read. No need to check the status just after issuing the command, especially if the polling is done in SW. Those 2 arguments should also be passed to the driver. > +{ > + struct spi_controller *ctlr = mem->spi->controller; > + unsigned long ms; > + int ret = -EOPNOTSUPP; > + int exec_op_ret; > + u16 *status; > + > + if (!spi_mem_supports_op(mem, op)) > + return ret; > + > + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { > + ret = spi_mem_access_start(mem); > + if (ret) > + return ret; > + > + reinit_completion(&ctlr->xfer_completion); > + > + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, > + timeout_ms); > + > + ms = wait_for_completion_timeout(&ctlr->xfer_completion, > + msecs_to_jiffies(timeout_ms)); > + > + spi_mem_access_end(mem); > + if (!ms) > + return -ETIMEDOUT; > + } else { > + status = (u16 *)op->data.buf.in; > + ret = read_poll_timeout(spi_mem_exec_op, exec_op_ret, > + ((*status) & mask) == match, 20, > + timeout_ms * 1000, false, mem, op); > + if (exec_op_ret) > + return exec_op_ret; > + } > + > + return ret; > +} > +EXPORT_SYMBOL_GPL(spi_mem_poll_status); > +
Hi Boris On 5/17/21 9:41 AM, Boris Brezillon wrote: > On Fri, 7 May 2021 15:17:54 +0200 > <patrice.chotard@foss.st.com> wrote: > >> +/** >> + * spi_mem_poll_status() - Poll memory device status >> + * @mem: SPI memory device >> + * @op: the memory operation to execute >> + * @mask: status bitmask to ckeck >> + * @match: (status & mask) expected value >> + * @timeout_ms: timeout in milliseconds >> + * >> + * This function send a polling status request to the controller driver >> + * >> + * Return: 0 in case of success, -ETIMEDOUT in case of error, >> + * -EOPNOTSUPP if not supported. >> + */ >> +int spi_mem_poll_status(struct spi_mem *mem, >> + const struct spi_mem_op *op, >> + u16 mask, u16 match, u16 timeout_ms) > > Maybe you should pass a delay_us too, to poll the status at the right > rate in the SW-based case (can also be used by drivers if they need to Ok, i will add a polling_rate_us parameter to poll_status() callback, even if in STM32 driver case we will not use it, i agree it should be useful depending of driver's implementation. > configure the polling rate). You could also add an initial_delay_us to > avoid polling the status too early: an erase operation will take longer > than a write which will take longer than a read. No need to check the > status just after issuing the command, especially if the polling is > done in SW. Those 2 arguments should also be passed to the driver. Regarding the addition of an initial_delay_us. We got two solution: - use the same polling rate already used by read_poll_timeout() and set read_poll_timeout()'s sleep_before_read parameter to true (in our case 20 us will be used as initial delay and as polling rate). - add an udelay(initial_delay_us) or even better usleep_range(initial_delay_us, initial_delay_us + delta) before calling read_poll_timeout(). I imagine you prefer the second solution ? By adding polling_rate_us and initial_delay_us parameters to spi_mem_poll_status(), it implies to update all spinand_wait() calls for different operations (reset, read page, write page, erase) with respective initial_delay_us/polling_rate_us values for spi_mem_poll_status()'s parameters. Can you provide adequate initial_delay_us and polling rate_us for each operation type ?. Thanks Patrice > >> +{ >> + struct spi_controller *ctlr = mem->spi->controller; >> + unsigned long ms; >> + int ret = -EOPNOTSUPP; >> + int exec_op_ret; >> + u16 *status; >> + >> + if (!spi_mem_supports_op(mem, op)) >> + return ret; >> + >> + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { >> + ret = spi_mem_access_start(mem); >> + if (ret) >> + return ret; >> + >> + reinit_completion(&ctlr->xfer_completion); >> + >> + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, >> + timeout_ms); >> + >> + ms = wait_for_completion_timeout(&ctlr->xfer_completion, >> + msecs_to_jiffies(timeout_ms)); >> + >> + spi_mem_access_end(mem); >> + if (!ms) >> + return -ETIMEDOUT; >> + } else { >> + status = (u16 *)op->data.buf.in; >> + ret = read_poll_timeout(spi_mem_exec_op, exec_op_ret, >> + ((*status) & mask) == match, 20, >> + timeout_ms * 1000, false, mem, op); >> + if (exec_op_ret) >> + return exec_op_ret; >> + } >> + >> + return ret; >> +} >> +EXPORT_SYMBOL_GPL(spi_mem_poll_status); >> +
On Mon, 17 May 2021 11:24:25 +0200 Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > Hi Boris > > On 5/17/21 9:41 AM, Boris Brezillon wrote: > > On Fri, 7 May 2021 15:17:54 +0200 > > <patrice.chotard@foss.st.com> wrote: > > > >> +/** > >> + * spi_mem_poll_status() - Poll memory device status > >> + * @mem: SPI memory device > >> + * @op: the memory operation to execute > >> + * @mask: status bitmask to ckeck > >> + * @match: (status & mask) expected value > >> + * @timeout_ms: timeout in milliseconds > >> + * > >> + * This function send a polling status request to the controller driver > >> + * > >> + * Return: 0 in case of success, -ETIMEDOUT in case of error, > >> + * -EOPNOTSUPP if not supported. > >> + */ > >> +int spi_mem_poll_status(struct spi_mem *mem, > >> + const struct spi_mem_op *op, > >> + u16 mask, u16 match, u16 timeout_ms) > > > > Maybe you should pass a delay_us too, to poll the status at the right > > rate in the SW-based case (can also be used by drivers if they need to > > Ok, i will add a polling_rate_us parameter to poll_status() callback, > even if in STM32 driver case we will not use it, i agree it should be useful > depending of driver's implementation. > > > configure the polling rate). You could also add an initial_delay_us to > > avoid polling the status too early: an erase operation will take longer > > than a write which will take longer than a read. No need to check the > > status just after issuing the command, especially if the polling is > > done in SW. Those 2 arguments should also be passed to the driver. > > Regarding the addition of an initial_delay_us. We got two solution: > - use the same polling rate already used by read_poll_timeout() and > set read_poll_timeout()'s sleep_before_read parameter to true (in our case 20 us > will be used as initial delay and as polling rate). > > - add an udelay(initial_delay_us) or even better usleep_range(initial_delay_us, > initial_delay_us + delta) before calling read_poll_timeout(). > > I imagine you prefer the second solution ? Yep, you might want to use udelay() when the delay is small and usleep_range() otherwise. > > By adding polling_rate_us and initial_delay_us parameters to > spi_mem_poll_status(), it implies to update all spinand_wait() calls for > different operations (reset, read page, write page, erase) with respective > initial_delay_us/polling_rate_us values for spi_mem_poll_status()'s parameters. > > Can you provide adequate initial_delay_us and polling rate_us for each operation type ?. If I refer to the datasheets I have, tBERS (erase) 1ms to 4ms tPROG 300us to 400us tREAD 25us to 100us Let's assume we want to minimize the latency, I'd recommend dividing the min value by 4 for the initial delay, and dividing it by 20 for the poll delay, which gives: ERASE -> initial_delay = 250us, poll_delay = 50us PROG -> initial_delay = 100us, poll_delay = 20us READ -> initial_delay = 6us, poll_delay = 5us Of course, that'd be even better if we were able to extract this information from the NAND ID (or ONFI table), but I guess we can live with those optimistic values in the meantime.
Hi On 5/17/21 1:25 PM, Boris Brezillon wrote: > On Mon, 17 May 2021 11:24:25 +0200 > Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > >> Hi Boris >> >> On 5/17/21 9:41 AM, Boris Brezillon wrote: >>> On Fri, 7 May 2021 15:17:54 +0200 >>> <patrice.chotard@foss.st.com> wrote: >>> >>>> +/** >>>> + * spi_mem_poll_status() - Poll memory device status >>>> + * @mem: SPI memory device >>>> + * @op: the memory operation to execute >>>> + * @mask: status bitmask to ckeck >>>> + * @match: (status & mask) expected value >>>> + * @timeout_ms: timeout in milliseconds >>>> + * >>>> + * This function send a polling status request to the controller driver >>>> + * >>>> + * Return: 0 in case of success, -ETIMEDOUT in case of error, >>>> + * -EOPNOTSUPP if not supported. >>>> + */ >>>> +int spi_mem_poll_status(struct spi_mem *mem, >>>> + const struct spi_mem_op *op, >>>> + u16 mask, u16 match, u16 timeout_ms) >>> >>> Maybe you should pass a delay_us too, to poll the status at the right >>> rate in the SW-based case (can also be used by drivers if they need to >> >> Ok, i will add a polling_rate_us parameter to poll_status() callback, >> even if in STM32 driver case we will not use it, i agree it should be useful >> depending of driver's implementation. >> >>> configure the polling rate). You could also add an initial_delay_us to >>> avoid polling the status too early: an erase operation will take longer >>> than a write which will take longer than a read. No need to check the >>> status just after issuing the command, especially if the polling is >>> done in SW. Those 2 arguments should also be passed to the driver. >> >> Regarding the addition of an initial_delay_us. We got two solution: >> - use the same polling rate already used by read_poll_timeout() and >> set read_poll_timeout()'s sleep_before_read parameter to true (in our case 20 us >> will be used as initial delay and as polling rate). >> >> - add an udelay(initial_delay_us) or even better usleep_range(initial_delay_us, >> initial_delay_us + delta) before calling read_poll_timeout(). >> >> I imagine you prefer the second solution ? > > Yep, you might want to use udelay() when the delay is small and > usleep_range() otherwise. > >> >> By adding polling_rate_us and initial_delay_us parameters to >> spi_mem_poll_status(), it implies to update all spinand_wait() calls for >> different operations (reset, read page, write page, erase) with respective >> initial_delay_us/polling_rate_us values for spi_mem_poll_status()'s parameters. >> >> Can you provide adequate initial_delay_us and polling rate_us for each operation type ?. > > If I refer to the datasheets I have, > > tBERS (erase) 1ms to 4ms > tPROG 300us to 400us > tREAD 25us to 100us > > Let's assume we want to minimize the latency, I'd recommend dividing > the min value by 4 for the initial delay, and dividing it by 20 for the > poll delay, which gives: > > ERASE -> initial_delay = 250us, poll_delay = 50us > PROG -> initial_delay = 100us, poll_delay = 20us > READ -> initial_delay = 6us, poll_delay = 5us What about RESET ? we also need an initial and poll delay too (see spinand_reset_op() ) > > Of course, that'd be even better if we were able to extract this > information from the NAND ID (or ONFI table), but I guess we can live > with those optimistic values in the meantime. > Thanks Patrice
On 5/17/21 1:25 PM, Boris Brezillon wrote: > On Mon, 17 May 2021 11:24:25 +0200 > Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > >> Hi Boris >> >> On 5/17/21 9:41 AM, Boris Brezillon wrote: >>> On Fri, 7 May 2021 15:17:54 +0200 >>> <patrice.chotard@foss.st.com> wrote: >>> >>>> +/** >>>> + * spi_mem_poll_status() - Poll memory device status >>>> + * @mem: SPI memory device >>>> + * @op: the memory operation to execute >>>> + * @mask: status bitmask to ckeck >>>> + * @match: (status & mask) expected value >>>> + * @timeout_ms: timeout in milliseconds >>>> + * >>>> + * This function send a polling status request to the controller driver >>>> + * >>>> + * Return: 0 in case of success, -ETIMEDOUT in case of error, >>>> + * -EOPNOTSUPP if not supported. >>>> + */ >>>> +int spi_mem_poll_status(struct spi_mem *mem, >>>> + const struct spi_mem_op *op, >>>> + u16 mask, u16 match, u16 timeout_ms) >>> >>> Maybe you should pass a delay_us too, to poll the status at the right >>> rate in the SW-based case (can also be used by drivers if they need to >> >> Ok, i will add a polling_rate_us parameter to poll_status() callback, >> even if in STM32 driver case we will not use it, i agree it should be useful >> depending of driver's implementation. >> >>> configure the polling rate). You could also add an initial_delay_us to >>> avoid polling the status too early: an erase operation will take longer >>> than a write which will take longer than a read. No need to check the >>> status just after issuing the command, especially if the polling is >>> done in SW. Those 2 arguments should also be passed to the driver. >> >> Regarding the addition of an initial_delay_us. We got two solution: >> - use the same polling rate already used by read_poll_timeout() and >> set read_poll_timeout()'s sleep_before_read parameter to true (in our case 20 us >> will be used as initial delay and as polling rate). >> >> - add an udelay(initial_delay_us) or even better usleep_range(initial_delay_us, >> initial_delay_us + delta) before calling read_poll_timeout(). >> >> I imagine you prefer the second solution ? > > Yep, you might want to use udelay() when the delay is small and > usleep_range() otherwise. > >> >> By adding polling_rate_us and initial_delay_us parameters to >> spi_mem_poll_status(), it implies to update all spinand_wait() calls for >> different operations (reset, read page, write page, erase) with respective >> initial_delay_us/polling_rate_us values for spi_mem_poll_status()'s parameters. >> >> Can you provide adequate initial_delay_us and polling rate_us for each operation type ?. > > If I refer to the datasheets I have, > > tBERS (erase) 1ms to 4ms > tPROG 300us to 400us > tREAD 25us to 100us > > Let's assume we want to minimize the latency, I'd recommend dividing > the min value by 4 for the initial delay, and dividing it by 20 for the > poll delay, which gives: > > ERASE -> initial_delay = 250us, poll_delay = 50us > PROG -> initial_delay = 100us, poll_delay = 20us another remark, it should be: PROG -> initial_delay = 75 us (300 / 4) , poll_delay = 15us ( 300 / 20) Patrice > READ -> initial_delay = 6us, poll_delay = 5us > > Of course, that'd be even better if we were able to extract this > information from the NAND ID (or ONFI table), but I guess we can live > with those optimistic values in the meantime. >
On Mon, 17 May 2021 13:59:54 +0200 Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > Hi > > On 5/17/21 1:25 PM, Boris Brezillon wrote: > > On Mon, 17 May 2021 11:24:25 +0200 > > Patrice CHOTARD <patrice.chotard@foss.st.com> wrote: > > > >> Hi Boris > >> > >> On 5/17/21 9:41 AM, Boris Brezillon wrote: > >>> On Fri, 7 May 2021 15:17:54 +0200 > >>> <patrice.chotard@foss.st.com> wrote: > >>> > >>>> +/** > >>>> + * spi_mem_poll_status() - Poll memory device status > >>>> + * @mem: SPI memory device > >>>> + * @op: the memory operation to execute > >>>> + * @mask: status bitmask to ckeck > >>>> + * @match: (status & mask) expected value > >>>> + * @timeout_ms: timeout in milliseconds > >>>> + * > >>>> + * This function send a polling status request to the controller driver > >>>> + * > >>>> + * Return: 0 in case of success, -ETIMEDOUT in case of error, > >>>> + * -EOPNOTSUPP if not supported. > >>>> + */ > >>>> +int spi_mem_poll_status(struct spi_mem *mem, > >>>> + const struct spi_mem_op *op, > >>>> + u16 mask, u16 match, u16 timeout_ms) > >>> > >>> Maybe you should pass a delay_us too, to poll the status at the right > >>> rate in the SW-based case (can also be used by drivers if they need to > >> > >> Ok, i will add a polling_rate_us parameter to poll_status() callback, > >> even if in STM32 driver case we will not use it, i agree it should be useful > >> depending of driver's implementation. > >> > >>> configure the polling rate). You could also add an initial_delay_us to > >>> avoid polling the status too early: an erase operation will take longer > >>> than a write which will take longer than a read. No need to check the > >>> status just after issuing the command, especially if the polling is > >>> done in SW. Those 2 arguments should also be passed to the driver. > >> > >> Regarding the addition of an initial_delay_us. We got two solution: > >> - use the same polling rate already used by read_poll_timeout() and > >> set read_poll_timeout()'s sleep_before_read parameter to true (in our case 20 us > >> will be used as initial delay and as polling rate). > >> > >> - add an udelay(initial_delay_us) or even better usleep_range(initial_delay_us, > >> initial_delay_us + delta) before calling read_poll_timeout(). > >> > >> I imagine you prefer the second solution ? > > > > Yep, you might want to use udelay() when the delay is small and > > usleep_range() otherwise. > > > >> > >> By adding polling_rate_us and initial_delay_us parameters to > >> spi_mem_poll_status(), it implies to update all spinand_wait() calls for > >> different operations (reset, read page, write page, erase) with respective > >> initial_delay_us/polling_rate_us values for spi_mem_poll_status()'s parameters. > >> > >> Can you provide adequate initial_delay_us and polling rate_us for each operation type ?. > > > > If I refer to the datasheets I have, > > > > tBERS (erase) 1ms to 4ms > > tPROG 300us to 400us > > tREAD 25us to 100us > > > > Let's assume we want to minimize the latency, I'd recommend dividing > > the min value by 4 for the initial delay, and dividing it by 20 for the > > poll delay, which gives: > > > > ERASE -> initial_delay = 250us, poll_delay = 50us > > PROG -> initial_delay = 100us, poll_delay = 20us > > READ -> initial_delay = 6us, poll_delay = 5us > > > What about RESET ? we also need an initial and poll delay too (see spinand_reset_op() ) 5us/10us/500us if the device is respectively reading/programming/erasing when the RESET occurs. Since we always issue a RESET when the device is IDLE, I'd recommend going for 5us for both the initial_delay and poll_delay. > > > > > Of course, that'd be even better if we were able to extract this > > information from the NAND ID (or ONFI table), but I guess we can live > > with those optimistic values in the meantime. > > > > Thanks > Patrice
diff --git a/drivers/spi/spi-mem.c b/drivers/spi/spi-mem.c index 1513553e4080..3f29c604df7d 100644 --- a/drivers/spi/spi-mem.c +++ b/drivers/spi/spi-mem.c @@ -6,6 +6,7 @@ * Author: Boris Brezillon <boris.brezillon@bootlin.com> */ #include <linux/dmaengine.h> +#include <linux/iopoll.h> #include <linux/pm_runtime.h> #include <linux/spi/spi.h> #include <linux/spi/spi-mem.h> @@ -743,6 +744,75 @@ static inline struct spi_mem_driver *to_spi_mem_drv(struct device_driver *drv) return container_of(drv, struct spi_mem_driver, spidrv.driver); } +/** + * spi_mem_finalize_op - report completion of spi_mem_op + * @ctlr: the controller reporting completion + * + * Called by SPI drivers using the spi-mem spi_mem_poll_status() + * implementation to notify it that the current spi_mem_op has + * finished. + */ +void spi_mem_finalize_op(struct spi_controller *ctlr) +{ + complete(&ctlr->xfer_completion); +} +EXPORT_SYMBOL_GPL(spi_mem_finalize_op); + +/** + * spi_mem_poll_status() - Poll memory device status + * @mem: SPI memory device + * @op: the memory operation to execute + * @mask: status bitmask to ckeck + * @match: (status & mask) expected value + * @timeout_ms: timeout in milliseconds + * + * This function send a polling status request to the controller driver + * + * Return: 0 in case of success, -ETIMEDOUT in case of error, + * -EOPNOTSUPP if not supported. + */ +int spi_mem_poll_status(struct spi_mem *mem, + const struct spi_mem_op *op, + u16 mask, u16 match, u16 timeout_ms) +{ + struct spi_controller *ctlr = mem->spi->controller; + unsigned long ms; + int ret = -EOPNOTSUPP; + int exec_op_ret; + u16 *status; + + if (!spi_mem_supports_op(mem, op)) + return ret; + + if (ctlr->mem_ops && ctlr->mem_ops->poll_status) { + ret = spi_mem_access_start(mem); + if (ret) + return ret; + + reinit_completion(&ctlr->xfer_completion); + + ret = ctlr->mem_ops->poll_status(mem, op, mask, match, + timeout_ms); + + ms = wait_for_completion_timeout(&ctlr->xfer_completion, + msecs_to_jiffies(timeout_ms)); + + spi_mem_access_end(mem); + if (!ms) + return -ETIMEDOUT; + } else { + status = (u16 *)op->data.buf.in; + ret = read_poll_timeout(spi_mem_exec_op, exec_op_ret, + ((*status) & mask) == match, 20, + timeout_ms * 1000, false, mem, op); + if (exec_op_ret) + return exec_op_ret; + } + + return ret; +} +EXPORT_SYMBOL_GPL(spi_mem_poll_status); + static int spi_mem_probe(struct spi_device *spi) { struct spi_mem_driver *memdrv = to_spi_mem_drv(spi->dev.driver); @@ -763,6 +833,7 @@ static int spi_mem_probe(struct spi_device *spi) if (IS_ERR_OR_NULL(mem->name)) return PTR_ERR_OR_ZERO(mem->name); + init_completion(&ctlr->xfer_completion); spi_set_drvdata(spi, mem); return memdrv->probe(mem); diff --git a/include/linux/spi/spi-mem.h b/include/linux/spi/spi-mem.h index 2b65c9edc34e..0fbf5d0a3d31 100644 --- a/include/linux/spi/spi-mem.h +++ b/include/linux/spi/spi-mem.h @@ -250,6 +250,7 @@ static inline void *spi_mem_get_drvdata(struct spi_mem *mem) * the currently mapped area), and the caller of * spi_mem_dirmap_write() is responsible for calling it again in * this case. + * @poll_status: poll memory device status * * This interface should be implemented by SPI controllers providing an * high-level interface to execute SPI memory operation, which is usually the @@ -274,6 +275,9 @@ struct spi_controller_mem_ops { u64 offs, size_t len, void *buf); ssize_t (*dirmap_write)(struct spi_mem_dirmap_desc *desc, u64 offs, size_t len, const void *buf); + int (*poll_status)(struct spi_mem *mem, + const struct spi_mem_op *op, + u16 mask, u16 match, unsigned long timeout); }; /** @@ -369,6 +373,12 @@ devm_spi_mem_dirmap_create(struct device *dev, struct spi_mem *mem, void devm_spi_mem_dirmap_destroy(struct device *dev, struct spi_mem_dirmap_desc *desc); +void spi_mem_finalize_op(struct spi_controller *ctlr); + +int spi_mem_poll_status(struct spi_mem *mem, + const struct spi_mem_op *op, + u16 mask, u16 match, u16 timeout); + int spi_mem_driver_register_with_owner(struct spi_mem_driver *drv, struct module *owner);