Re: [PATCH] cxl/pci: Get rid of pointer arithmetic reading CDAT table

[Ira Weiny]

Robert Richter wrote:
> On 17.11.23 21:09:18, Robert Richter wrote:
> > I will send an on-top patch for 6.8 that reworks that code area to
> > remove the pointer arithmetic.
> 
> Here it is:
> 
> From 13787f72c20b8c54754ae86015d982307eae0397 Mon Sep 17 00:00:00 2001
> From: Robert Richter <rrichter@xxxxxxx>
> Subject: [PATCH] cxl/pci: Get rid of pointer arithmetic reading CDAT table
> 
> Reading the CDAT table using DOE requires a Table Access Response
> Header in addition to the CDAT entry. In current implementation this
> has caused offsets with sizeof(__le32) to the actual buffers. This led
> to hardly readable code and even bugs (see fix of devm_kfree() in
> read_cdat_data()).
> 
> Rework code to avoid calculations with sizeof(__le32). Introduce
> struct cdat_doe for this which contains the Table Access Response
> Header and a variable payload size for various data structures
> afterwards to access the CDAT table and its CDAT Data Structures
> without recalculating buffer offsets.

Thanks for this.

> 
> Cc: Lukas Wunner <lukas@xxxxxxxxx>
> Cc: Dave Jiang <dave.jiang@xxxxxxxxx>
> Cc: Fan Ni <nifan.cxl@xxxxxxxxx>
> Signed-off-by: Robert Richter <rrichter@xxxxxxx>
> ---
>  drivers/cxl/core/pci.c | 80 ++++++++++++++++++++----------------------
>  drivers/cxl/cxlpci.h   | 19 ++++++++++
>  2 files changed, 57 insertions(+), 42 deletions(-)
> 
> diff --git a/drivers/cxl/core/pci.c b/drivers/cxl/core/pci.c
> index 5aaa0b36c42a..f900740c6dea 100644
> --- a/drivers/cxl/core/pci.c
> +++ b/drivers/cxl/core/pci.c
> @@ -517,14 +517,14 @@ EXPORT_SYMBOL_NS_GPL(cxl_hdm_decode_init, CXL);
>  	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE, (entry_handle)))
>  
>  static int cxl_cdat_get_length(struct device *dev,
> -			       struct pci_doe_mb *cdat_doe,
> +			       struct pci_doe_mb *doe_mb,

NIT: Why change the variable name here?

>  			       size_t *length)
>  {
>  	__le32 request = CDAT_DOE_REQ(0);
>  	__le32 response[2];
>  	int rc;
>  
> -	rc = pci_doe(cdat_doe, PCI_DVSEC_VENDOR_ID_CXL,
> +	rc = pci_doe(doe_mb, PCI_DVSEC_VENDOR_ID_CXL,
>  		     CXL_DOE_PROTOCOL_TABLE_ACCESS,
>  		     &request, sizeof(request),
>  		     &response, sizeof(response));
> @@ -542,56 +542,54 @@ static int cxl_cdat_get_length(struct device *dev,
>  }
>  
>  static int cxl_cdat_read_table(struct device *dev,
> -			       struct pci_doe_mb *cdat_doe,
> -			       void *cdat_table, size_t *cdat_length)
> +			       struct pci_doe_mb *doe_mb,
> +			       struct cdat_doe *doe, size_t *length)
>  {
> -	size_t length = *cdat_length + sizeof(__le32);
> -	__le32 *data = cdat_table;
> +	size_t received, remaining = *length;
>  	int entry_handle = 0;
>  	__le32 saved_dw = 0;
>  
>  	do {
>  		__le32 request = CDAT_DOE_REQ(entry_handle);
> -		struct cdat_entry_header *entry;
> -		size_t entry_dw;
>  		int rc;
>  
> -		rc = pci_doe(cdat_doe, PCI_DVSEC_VENDOR_ID_CXL,
> +		rc = pci_doe(doe_mb, PCI_DVSEC_VENDOR_ID_CXL,
>  			     CXL_DOE_PROTOCOL_TABLE_ACCESS,
>  			     &request, sizeof(request),
> -			     data, length);
> +			     doe, sizeof(*doe) + remaining);
>  		if (rc < 0) {
>  			dev_err(dev, "DOE failed: %d", rc);
>  			return rc;
>  		}
>  
> -		/* 1 DW Table Access Response Header + CDAT entry */
> -		entry = (struct cdat_entry_header *)(data + 1);
> +		if (rc < sizeof(*doe))
> +			return -EIO;
> +
> +		received = rc - sizeof(*doe);

This is not ideal.  See comments on struct cdat_doe below.

> +
>  		if ((entry_handle == 0 &&
> -		     rc != sizeof(__le32) + sizeof(struct cdat_header)) ||
> +		     received != sizeof(doe->header[0])) ||
>  		    (entry_handle > 0 &&
> -		     (rc < sizeof(__le32) + sizeof(*entry) ||
> -		      rc != sizeof(__le32) + le16_to_cpu(entry->length))))
> +		     (received < sizeof(doe->entry[0]) ||
> +		      received != le16_to_cpu(doe->entry->length))))
>  			return -EIO;
>  
>  		/* Get the CXL table access header entry handle */
>  		entry_handle = FIELD_GET(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE,
> -					 le32_to_cpu(data[0]));
> -		entry_dw = rc / sizeof(__le32);
> -		/* Skip Header */
> -		entry_dw -= 1;
> +					 le32_to_cpu(doe->doe_header));
> +
>  		/*
>  		 * Table Access Response Header overwrote the last DW of
>  		 * previous entry, so restore that DW
>  		 */
> -		*data = saved_dw;
> -		length -= entry_dw * sizeof(__le32);
> -		data += entry_dw;
> -		saved_dw = *data;
> +		doe->doe_header = saved_dw;
> +		remaining -= received;
> +		doe = (void *)doe + received;
> +		saved_dw = doe->doe_header;
>  	} while (entry_handle != CXL_DOE_TABLE_ACCESS_LAST_ENTRY);
>  
>  	/* Length in CDAT header may exceed concatenation of CDAT entries */
> -	*cdat_length -= length - sizeof(__le32);
> +	*length -= remaining;
>  
>  	return 0;
>  }
> @@ -616,11 +614,11 @@ void read_cdat_data(struct cxl_port *port)
>  {
>  	struct device *uport = port->uport_dev;
>  	struct device *dev = &port->dev;
> -	struct pci_doe_mb *cdat_doe;
> +	struct pci_doe_mb *doe_mb;
>  	struct pci_dev *pdev = NULL;
>  	struct cxl_memdev *cxlmd;
> -	size_t cdat_length;
> -	void *cdat_table, *cdat_buf;
> +	struct cdat_doe *doe;

Since we are trying to make this cleaner I would prefer a struct and
variable name of cdat_doe_rsp.

So...
	struct cdat_doe_rsp *cdat_doe_rsp;

> +	size_t length;
>  	int rc;
>  
>  	if (is_cxl_memdev(uport)) {
> @@ -637,40 +635,38 @@ void read_cdat_data(struct cxl_port *port)
>  	if (!pdev)
>  		return;
>  
> -	cdat_doe = pci_find_doe_mailbox(pdev, PCI_DVSEC_VENDOR_ID_CXL,
> -					CXL_DOE_PROTOCOL_TABLE_ACCESS);
> -	if (!cdat_doe) {
> +	doe_mb = pci_find_doe_mailbox(pdev, PCI_DVSEC_VENDOR_ID_CXL,
> +				      CXL_DOE_PROTOCOL_TABLE_ACCESS);
> +	if (!doe_mb) {
>  		dev_dbg(dev, "No CDAT mailbox\n");
>  		return;
>  	}
>  
>  	port->cdat_available = true;
>  
> -	if (cxl_cdat_get_length(dev, cdat_doe, &cdat_length)) {
> +	if (cxl_cdat_get_length(dev, doe_mb, &length)) {
>  		dev_dbg(dev, "No CDAT length\n");
>  		return;
>  	}
>  
> -	cdat_buf = devm_kzalloc(dev, cdat_length + sizeof(__le32),
> -				  GFP_KERNEL);
> -	if (!cdat_buf)
> -		return;
> +	doe = devm_kzalloc(dev, sizeof(*doe) + length, GFP_KERNEL);
> +	if (!doe)
> +		goto err;
>  
> -	rc = cxl_cdat_read_table(dev, cdat_doe, cdat_buf, &cdat_length);
> +	rc = cxl_cdat_read_table(dev, doe_mb, doe, &length);
>  	if (rc)
>  		goto err;
>  
> -	cdat_table = cdat_buf + sizeof(__le32);
> -	if (cdat_checksum(cdat_table, cdat_length))
> +	if (cdat_checksum(doe->table, length))
>  		goto err;
>  
> -	port->cdat.table = cdat_table;
> -	port->cdat.length = cdat_length;
> -	return;
> +	port->cdat.table = doe->table;

As an aside: the type of port->cdat may need to change at some point too.

> +	port->cdat.length = length;
>  
> +	return;
>  err:
>  	/* Don't leave table data allocated on error */
> -	devm_kfree(dev, cdat_buf);
> +	devm_kfree(dev, doe);
>  	dev_err(dev, "Failed to read/validate CDAT.\n");
>  }
>  EXPORT_SYMBOL_NS_GPL(read_cdat_data, CXL);
> diff --git a/drivers/cxl/cxlpci.h b/drivers/cxl/cxlpci.h
> index 0fa4799ea316..d12ed9d8dec1 100644
> --- a/drivers/cxl/cxlpci.h
> +++ b/drivers/cxl/cxlpci.h
> @@ -85,6 +85,25 @@ struct cdat_entry_header {
>  	__le16 length;
>  } __packed;
>  
> +/*
> + * Response contains the CDAT only response header of the DOE. The
> + * response payload is a CDAT structure (either CDAT header or entry),
> + * it may also mark the beginning of the CDAT table.
> + *
> + * Spec refs:
> + *
> + * CXL 3.1 Table 8-14: Read Entry Response
> + * CDAT Specification 1.03: 2 CDAT Data Structures
> + */
> +struct cdat_doe {
> +	__le32 doe_header;
> +	union {
> +		u8 table[0];

At a minimum we need to use flexible arrays here.  See:

.../Documentation/process/deprecated.rst
	'Zero-length and one-element arrays'

> +		struct cdat_header header[0];
> +		struct cdat_entry_header entry[0];

So this would need to be:

		DECLARE_FLEX_ARRAY(u8, table);
		DECLARE_FLEX_ARRAY(struct cdat_header, header);
		DECLARE_FLEX_ARRAY(struct cdat_entry_header, entry);

Also I think it would be best to use sizeof_field().

That said I got crossed up when this structure was used to represent not
only the individual DOE responses but also the cdat table as a whole.

I think it would be best to call this cdat_doe_rsp and somehow make a
distinction in read_cdat_data() when assigning to the entire port cdat
table.

Ira

> +	};
> +} __packed;
> +
>  int devm_cxl_port_enumerate_dports(struct cxl_port *port);
>  struct cxl_dev_state;
>  int cxl_hdm_decode_init(struct cxl_dev_state *cxlds, struct cxl_hdm *cxlhdm,
> -- 
> 2.39.2
> 
>