[PATCH v3 10/11] cxl: EEH support

Cyril Bur cyrilbur at gmail.com
Wed Aug 12 16:40:37 AEST 2015


On Wed, 12 Aug 2015 10:48:19 +1000
Daniel Axtens <dja at axtens.net> wrote:

> EEH (Enhanced Error Handling) allows a driver to recover from the
> temporary failure of an attached PCI card. Enable basic CXL support
> for EEH.
> 

Looks like the only change since was the removal of the #ifdef, if that is
correct.

Reviewed-by: Cyril Bur <cyrilbur at gmail.com>

> Signed-off-by: Daniel Axtens <dja at axtens.net>
> ---
>  drivers/misc/cxl/cxl.h  |   1 +
>  drivers/misc/cxl/pci.c  | 252 ++++++++++++++++++++++++++++++++++++++++++++++++
>  drivers/misc/cxl/vphb.c |   8 ++
>  3 files changed, 261 insertions(+)
> 
> diff --git a/drivers/misc/cxl/cxl.h b/drivers/misc/cxl/cxl.h
> index cda02412b01e..6f5386653dae 100644
> --- a/drivers/misc/cxl/cxl.h
> +++ b/drivers/misc/cxl/cxl.h
> @@ -726,6 +726,7 @@ int cxl_psl_purge(struct cxl_afu *afu);
>  
>  void cxl_stop_trace(struct cxl *cxl);
>  int cxl_pci_vphb_add(struct cxl_afu *afu);
> +void cxl_pci_vphb_reconfigure(struct cxl_afu *afu);
>  void cxl_pci_vphb_remove(struct cxl_afu *afu);
>  
>  extern struct pci_driver cxl_pci_driver;
> diff --git a/drivers/misc/cxl/pci.c b/drivers/misc/cxl/pci.c
> index b4a68a896a33..1eb26a357ce0 100644
> --- a/drivers/misc/cxl/pci.c
> +++ b/drivers/misc/cxl/pci.c
> @@ -24,6 +24,7 @@
>  #include <asm/io.h>
>  
>  #include "cxl.h"
> +#include <misc/cxl.h>
>  
>  
>  #define CXL_PCI_VSEC_ID	0x1280
> @@ -1246,10 +1247,261 @@ static void cxl_remove(struct pci_dev *dev)
>  	cxl_remove_adapter(adapter);
>  }
>  
> +static pci_ers_result_t cxl_vphb_error_detected(struct cxl_afu *afu,
> +						pci_channel_state_t state)
> +{
> +	struct pci_dev *afu_dev;
> +	pci_ers_result_t result = PCI_ERS_RESULT_NEED_RESET;
> +	pci_ers_result_t afu_result = PCI_ERS_RESULT_NEED_RESET;
> +
> +	/* There should only be one entry, but go through the list
> +	 * anyway
> +	 */
> +	list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
> +		if (!afu_dev->driver)
> +			continue;
> +
> +		afu_dev->error_state = state;
> +
> +		if (afu_dev->driver->err_handler)
> +			afu_result = afu_dev->driver->err_handler->error_detected(afu_dev,
> +										  state);
> +		/* Disconnect trumps all, NONE trumps NEED_RESET */
> +		if (afu_result == PCI_ERS_RESULT_DISCONNECT)
> +			result = PCI_ERS_RESULT_DISCONNECT;
> +		else if ((afu_result == PCI_ERS_RESULT_NONE) &&
> +			 (result == PCI_ERS_RESULT_NEED_RESET))
> +			result = PCI_ERS_RESULT_NONE;
> +	}
> +	return result;
> +}
> +
> +static pci_ers_result_t cxl_pci_error_detected(struct pci_dev *pdev,
> +					       pci_channel_state_t state)
> +{
> +	struct cxl *adapter = pci_get_drvdata(pdev);
> +	struct cxl_afu *afu;
> +	pci_ers_result_t result = PCI_ERS_RESULT_NEED_RESET;
> +	int i;
> +
> +	/* At this point, we could still have an interrupt pending.
> +	 * Let's try to get them out of the way before they do
> +	 * anything we don't like.
> +	 */
> +	schedule();
> +
> +	/* If we're permanently dead, give up. */
> +	if (state == pci_channel_io_perm_failure) {
> +		/* Tell the AFU drivers; but we don't care what they
> +		 * say, we're going away.
> +		 */
> +		for (i = 0; i < adapter->slices; i++) {
> +			afu = adapter->afu[i];
> +			cxl_vphb_error_detected(afu, state);
> +		}
> +		return PCI_ERS_RESULT_DISCONNECT;
> +	}
> +
> +	/* Are we reflashing?
> +	 *
> +	 * If we reflash, we could come back as something entirely
> +	 * different, including a non-CAPI card. As such, by default
> +	 * we don't participate in the process. We'll be unbound and
> +	 * the slot re-probed. (TODO: check EEH doesn't blindly rebind
> +	 * us!)
> +	 *
> +	 * However, this isn't the entire story: for reliablity
> +	 * reasons, we usually want to reflash the FPGA on PERST in
> +	 * order to get back to a more reliable known-good state.
> +	 *
> +	 * This causes us a bit of a problem: if we reflash we can't
> +	 * trust that we'll come back the same - we could have a new
> +	 * image and been PERSTed in order to load that
> +	 * image. However, most of the time we actually *will* come
> +	 * back the same - for example a regular EEH event.
> +	 *
> +	 * Therefore, we allow the user to assert that the image is
> +	 * indeed the same and that we should continue on into EEH
> +	 * anyway.
> +	 */
> +	if (adapter->perst_loads_image && !adapter->perst_same_image) {
> +		/* TODO take the PHB out of CXL mode */
> +		dev_info(&pdev->dev, "reflashing, so opting out of EEH!\n");
> +		return PCI_ERS_RESULT_NONE;
> +	}
> +
> +	/*
> +	 * At this point, we want to try to recover.  We'll always
> +	 * need a complete slot reset: we don't trust any other reset.
> +	 *
> +	 * Now, we go through each AFU:
> +	 *  - We send the driver, if bound, an error_detected callback.
> +	 *    We expect it to clean up, but it can also tell us to give
> +	 *    up and permanently detach the card. To simplify things, if
> +	 *    any bound AFU driver doesn't support EEH, we give up on EEH.
> +	 *
> +	 *  - We detach all contexts associated with the AFU. This
> +	 *    does not free them, but puts them into a CLOSED state
> +	 *    which causes any the associated files to return useful
> +	 *    errors to userland. It also unmaps, but does not free,
> +	 *    any IRQs.
> +	 *
> +	 *  - We clean up our side: releasing and unmapping resources we hold
> +	 *    so we can wire them up again when the hardware comes back up.
> +	 *
> +	 * Driver authors should note:
> +	 *
> +	 *  - Any contexts you create in your kernel driver (except
> +	 *    those associated with anonymous file descriptors) are
> +	 *    your responsibility to free and recreate. Likewise with
> +	 *    any attached resources.
> +	 *
> +	 *  - We will take responsibility for re-initialising the
> +	 *    device context (the one set up for you in
> +	 *    cxl_pci_enable_device_hook and accessed through
> +	 *    cxl_get_context). If you've attached IRQs or other
> +	 *    resources to it, they remains yours to free.
> +	 *
> +	 * All calls you make into cxl that normally touch the
> +	 * hardware will not touch the hardware during recovery. So
> +	 * you can call the same functions to release resources as you
> +	 * normally would.
> +	 *
> +	 * Two examples:
> +	 *
> +	 * 1) If you normally free all your resources at the end of
> +	 *    each request, or if you use anonymous FDs, your
> +	 *    error_detected callback can simply set a flag to tell
> +	 *    your driver not to start any new calls. You can then
> +	 *    clear the flag in the resume callback.
> +	 *
> +	 * 2) If you normally allocate your resources on startup:
> +	 *     * Set a flag in error_detected as above.
> +	 *     * Let CXL detach your contexts.
> +	 *     * In slot_reset, free the old resources and allocate new ones.
> +	 *     * In resume, clear the flag to allow things to start.
> +	 */
> +	for (i = 0; i < adapter->slices; i++) {
> +		afu = adapter->afu[i];
> +
> +		result = cxl_vphb_error_detected(afu, state);
> +
> +		/* Only continue if everyone agrees on NEED_RESET */
> +		if (result != PCI_ERS_RESULT_NEED_RESET)
> +			return result;
> +
> +		cxl_context_detach_all(afu);
> +		cxl_afu_deactivate_mode(afu);
> +		cxl_deconfigure_afu(afu);
> +	}
> +	cxl_deconfigure_adapter(adapter);
> +
> +	return result;
> +}
> +
> +static pci_ers_result_t cxl_pci_slot_reset(struct pci_dev *pdev)
> +{
> +	struct cxl *adapter = pci_get_drvdata(pdev);
> +	struct cxl_afu *afu;
> +	struct cxl_context *ctx;
> +	struct pci_dev *afu_dev;
> +	pci_ers_result_t afu_result = PCI_ERS_RESULT_RECOVERED;
> +	pci_ers_result_t result = PCI_ERS_RESULT_RECOVERED;
> +	int i;
> +
> +	if (cxl_configure_adapter(adapter, pdev))
> +		goto err;
> +
> +	for (i = 0; i < adapter->slices; i++) {
> +		afu = adapter->afu[i];
> +
> +		if (cxl_configure_afu(afu, adapter, pdev))
> +			goto err;
> +
> +		if (cxl_afu_select_best_mode(afu))
> +			goto err;
> +
> +		cxl_pci_vphb_reconfigure(afu);
> +
> +		list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
> +			/* Reset the device context.
> +			 * TODO: make this less disruptive
> +			 */
> +			ctx = cxl_get_context(afu_dev);
> +
> +			if (ctx && cxl_release_context(ctx))
> +				goto err;
> +
> +			ctx = cxl_dev_context_init(afu_dev);
> +			if (!ctx)
> +				goto err;
> +
> +			afu_dev->dev.archdata.cxl_ctx = ctx;
> +
> +			if (cxl_afu_check_and_enable(afu))
> +				goto err;
> +
> +			afu_dev->error_state = pci_channel_io_normal;
> +
> +			/* If there's a driver attached, allow it to
> +			 * chime in on recovery. Drivers should check
> +			 * if everything has come back OK.
> +			 */
> +			if (!afu_dev->driver)
> +				continue;
> +
> +			if (afu_dev->driver->err_handler &&
> +			    afu_dev->driver->err_handler->slot_reset)
> +				afu_result = afu_dev->driver->err_handler->slot_reset(afu_dev);
> +
> +			if (afu_result == PCI_ERS_RESULT_DISCONNECT)
> +				result = PCI_ERS_RESULT_DISCONNECT;
> +		}
> +	}
> +	return result;
> +
> +err:
> +	/* All the bits that happen in both error_detected and cxl_remove
> +	 * should be idempotent, so we don't need to worry about leaving a mix
> +	 * of unconfigured and reconfigured resources.
> +	 */
> +	dev_err(&pdev->dev, "EEH recovery failed. Asking to be disconnected.\n");
> +	return PCI_ERS_RESULT_DISCONNECT;
> +}
> +
> +static void cxl_pci_resume(struct pci_dev *pdev)
> +{
> +	struct cxl *adapter = pci_get_drvdata(pdev);
> +	struct cxl_afu *afu;
> +	struct pci_dev *afu_dev;
> +	int i;
> +
> +	/* Everything is back now. Drivers should restart work now.
> +	 * This is not the place to be checking if everything came back up
> +	 * properly, because there's no return value: do that in slot_reset.
> +	 */
> +	for (i = 0; i < adapter->slices; i++) {
> +		afu = adapter->afu[i];
> +
> +		list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
> +			if (afu_dev->driver && afu_dev->driver->err_handler &&
> +			    afu_dev->driver->err_handler->resume)
> +				afu_dev->driver->err_handler->resume(afu_dev);
> +		}
> +	}
> +}
> +
> +static const struct pci_error_handlers cxl_err_handler = {
> +	.error_detected = cxl_pci_error_detected,
> +	.slot_reset = cxl_pci_slot_reset,
> +	.resume = cxl_pci_resume,
> +};
> +
>  struct pci_driver cxl_pci_driver = {
>  	.name = "cxl-pci",
>  	.id_table = cxl_pci_tbl,
>  	.probe = cxl_probe,
>  	.remove = cxl_remove,
>  	.shutdown = cxl_remove,
> +	.err_handler = &cxl_err_handler,
>  };
> diff --git a/drivers/misc/cxl/vphb.c b/drivers/misc/cxl/vphb.c
> index d1c3509c925d..af336ff6655e 100644
> --- a/drivers/misc/cxl/vphb.c
> +++ b/drivers/misc/cxl/vphb.c
> @@ -266,6 +266,14 @@ int cxl_pci_vphb_add(struct cxl_afu *afu)
>  	return 0;
>  }
>  
> +void cxl_pci_vphb_reconfigure(struct cxl_afu *afu)
> +{
> +	/* When we are reconfigured, the AFU's MMIO space is unmapped
> +	 * and remapped. We need to reflect this in the PHB's view of
> +	 * the world.
> +	 */
> +	afu->phb->cfg_addr = afu->afu_desc_mmio + afu->crs_offset;
> +}
>  
>  void cxl_pci_vphb_remove(struct cxl_afu *afu)
>  {



More information about the Linuxppc-dev mailing list