Re: [PATCH v7 4/6] cxl/region: Provide region info to the cxl_poison trace event

[Jonathan Cameron]

On Fri, 17 Feb 2023 18:18:52 -0800
alison.schofield@xxxxxxxxx wrote:

> From: Alison Schofield <alison.schofield@xxxxxxxxx>
> 
> User space may need to know which region, if any, maps the poison
> address(es) logged in a cxl_poison trace event. Since the mapping
> of DPAs (device physical addresses) to a region can change, the
> kernel must provide this information at the time the poison list
> is read. The event informs user space that at event <timestamp>
> this <region> mapped to this <DPA>, which is poisoned.
> 
> The cxl_poison trace event is already wired up to log the region
> name and uuid if it receives param 'struct cxl_region'.
> 
> In order to provide that cxl_region, add another method for gathering
> poison - by committed endpoint decoder mappings. This method is only
> available with CONFIG_CXL_REGION and is only used if a region actually
> maps the memdev where poison is being read. After the poison list is
> read for all the mapped resources, poison is read for the unmapped
> resources, and those events are logged without the region info.
> 
> Mixed mode decoders are not currently supported in Linux. Add a debug
> message to the poison request path. That will serve as an alert that
> poison list retrieval needs to add support for mixed mode.
> 
> The default method remains: read the poison by memdev resource.
> 
> Signed-off-by: Alison Schofield <alison.schofield@xxxxxxxxx>
Hi Alison,

I decided to give this a test run on to of a single emulated direct
attached EP with both persistent and volatile memory, but only
part of each added to a region (so we get a skip).

Without regions it all works as expected. With them not so much - see inline

region uuid is also a bit pointless for volatile regions as I think
current code makes it all 0s.

Jonathan


> ---
>  drivers/cxl/core/core.h   |  5 +++
>  drivers/cxl/core/memdev.c | 14 +++++-
>  drivers/cxl/core/region.c | 89 +++++++++++++++++++++++++++++++++++++++
>  3 files changed, 107 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/cxl/core/core.h b/drivers/cxl/core/core.h
> index cde475e13216..4f507cb85926 100644
> --- a/drivers/cxl/core/core.h
> +++ b/drivers/cxl/core/core.h
> @@ -25,7 +25,12 @@ void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled);
>  #define CXL_DAX_REGION_TYPE(x) (&cxl_dax_region_type)
>  int cxl_region_init(void);
>  void cxl_region_exit(void);
> +int cxl_get_poison_by_endpoint(struct device *dev, void *data);
>  #else
> +static inline int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> +{
> +	return 0;
> +}
>  static inline void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled)
>  {
>  }
> diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> index ea996057815e..c11b7bc253b4 100644
> --- a/drivers/cxl/core/memdev.c
> +++ b/drivers/cxl/core/memdev.c
> @@ -139,14 +139,26 @@ static ssize_t trigger_poison_list_store(struct device *dev,
>  					 const char *buf, size_t len)
>  {
>  	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
> +	struct cxl_port *port;
>  	bool trigger;
>  	int rc;
>  
>  	if (kstrtobool(buf, &trigger) || !trigger)
>  		return -EINVAL;
>  
> +	port = dev_get_drvdata(&cxlmd->dev);
> +	if (!port || !is_cxl_endpoint(port))
> +		return -EINVAL;
> +
>  	down_read(&cxl_dpa_rwsem);
> -	rc = cxl_get_poison_by_memdev(cxlmd);
> +	if (port->commit_end == -1)
> +		/* No regions mapped to this memdev */
> +		rc = cxl_get_poison_by_memdev(cxlmd);
> +	else
> +		/* Regions mapped, collect poison by endpoint */
> +		rc = device_for_each_child(&port->dev, port,
> +					   cxl_get_poison_by_endpoint);

So this had me confused for a while.  If I setup a couple of regions
on a mixed device with a skip between (may happen anyway)
then this returns 1 which is not what we are aiming for.

Chasing it through we are reaching the last part of
cxl_get_poison_by_endpoint() which has the comment
"... Return rc or 1 to stop iteration." which is exactly 
what it is doing.

so need an if (rc == 1) rc = 0; here



> +
>  	up_read(&cxl_dpa_rwsem);
>  
>  	return rc ? rc : len;
> diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
> index f29028148806..a055f8e36ef1 100644
> --- a/drivers/cxl/core/region.c
> +++ b/drivers/cxl/core/region.c
> @@ -2213,6 +2213,95 @@ struct cxl_pmem_region *to_cxl_pmem_region(struct device *dev)
>  }
>  EXPORT_SYMBOL_NS_GPL(to_cxl_pmem_region, CXL);
>  
> +int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> +{
> +	struct cxl_endpoint_decoder *cxled;
> +	struct cxl_port *port = data;
> +	struct cxl_dev_state *cxlds;
> +	struct cxl_memdev *cxlmd;
> +	u64 offset, length;
> +	int rc = 0;
> +
> +	down_read(&cxl_dpa_rwsem);
> +
> +	if (!is_endpoint_decoder(dev))
> +		goto out;
> +
> +	cxled = to_cxl_endpoint_decoder(dev);
> +	if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
> +		goto out;
> +
> +	/*
> +	 * Regions are only created with single mode decoders: pmem or ram.
> +	 * Linux does not currently support mixed mode decoders. This means
> +	 * that reading poison per endpoint decoder adheres to the spec
> +	 * requirement that poison reads of pmem and ram must be separated.
> +	 * CXL 3.0 Spec 8.2.9.8.4.1
> +	 *
> +	 * Watch for future support of mixed with a dev_dbg() msg.
> +	 */
> +	if (cxled->mode == CXL_DECODER_MIXED) {
> +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
> +		goto out;
> +	}
> +
> +	cxlmd = cxled_to_memdev(cxled);
> +	if (cxled->skip) {
> +		rc = cxl_mem_get_poison(cxlmd, 0, cxled->skip, NULL);

The base of this should be the top fo the previous decoder. I'm not immediately
sure how you can get hold of that.

In my test with 1GB volatile and 1GB persistent the queries that are sent are:
0 to 0x10000000 (correct)
0 to 0x30000000 (wrong start, should be 0x10000000)

0x40000000 to 0x50000000 (correct)
0x50000000 to 0x40000000 (not correct but at least the start is correct).
See below for this one.

> +		if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> +			rc = 0;
> +		if (rc)
> +			goto out;
> +	}
> +	length = cxled->dpa_res->end - cxled->dpa_res->start + 1;
> +	rc = cxl_mem_get_poison(cxlmd, cxled->dpa_res->start, length,
> +				cxled->cxld.region);
> +	if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> +		rc = 0;
> +	if (rc)
> +		goto out;
> +
> +	/* Iterate until commit_end is reached */
> +	if (cxled->cxld.id < port->commit_end)
> +		goto out;
> +
> +	/*
> +	 * Reach here with the last committed decoder only.
> +	 * Knowing that PMEM must always follow RAM, get poison
> +	 * for unmapped ranges based on the last decoder's mode:
> +	 *	ram: scan remains of ram range, then scan for pmem
> +	 *	pmem: scan remains of pmem range
> +	 */
> +	cxlds = cxlmd->cxlds;
> +
> +	if (cxled->mode == CXL_DECODER_RAM) {
> +		offset = cxled->dpa_res->end + 1;
> +		length = resource_size(&cxlds->ram_res) - offset;
> +		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> +		if (rc == -EFAULT)
> +			rc = 0;
> +		if (rc)
> +			goto out;
> +	}
> +	if (cxled->mode == CXL_DECODER_PMEM) {
> +		offset = cxled->dpa_res->end + 1;

This is fine, unless you have a ram region in which case
the offset computed here includes that as well as the
persistent bit.

> +		length = resource_size(&cxlds->pmem_res) - offset;

So you need to take that into account with something like
- (offset - cxlds->pmem_res.start);

> +	} else if (resource_size(&cxlds->pmem_res)) {
> +		offset = cxlds->pmem_res.start;
> +		length = resource_size(&cxlds->pmem_res);
> +	} else {
> +		rc = 1;
> +		goto out;
> +	}
> +	/* Final get poison call. Return rc or 1 to stop iteration. */
> +	rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> +	if (!rc)
> +		rc = 1;
> +out:
> +	up_read(&cxl_dpa_rwsem);
> +	return rc;
> +}
> +
>  static struct lock_class_key cxl_pmem_region_key;
>  
>  static struct cxl_pmem_region *cxl_pmem_region_alloc(struct cxl_region *cxlr)