Re: [PATCH 2/2] fpga: dfl: look for vendor specific capability
From: matthew . gerlach
Date: Tue Nov 17 2020 - 14:41:31 EST
On Tue, 17 Nov 2020, Xu Yilun wrote:
On Mon, Nov 16, 2020 at 05:25:52PM -0800, matthew.gerlach@xxxxxxxxxxxxxxx wrote:
From: Matthew Gerlach <matthew.gerlach@xxxxxxxxxxxxxxx>
A DFL may not begin at offset 0 of BAR 0. A PCIe vendor
specific capability can be used to specify the start of a
number of DFLs.
Signed-off-by: Matthew Gerlach <matthew.gerlach@xxxxxxxxxxxxxxx>
---
Documentation/fpga/dfl.rst | 10 +++++
drivers/fpga/dfl-pci.c | 88 +++++++++++++++++++++++++++++++++++++-
2 files changed, 97 insertions(+), 1 deletion(-)
diff --git a/Documentation/fpga/dfl.rst b/Documentation/fpga/dfl.rst
index 0404fe6ffc74..c81ceb1e79e2 100644
--- a/Documentation/fpga/dfl.rst
+++ b/Documentation/fpga/dfl.rst
@@ -501,6 +501,16 @@ Developer only needs to provide a sub feature driver with matched feature id.
FME Partial Reconfiguration Sub Feature driver (see drivers/fpga/dfl-fme-pr.c)
could be a reference.
+Location of DFLs on PCI bus
+===========================
+The start of the DFL is assumed to be offset 0 of bar 0.
+Alternatively, a vendor specific capability structure can be used to
+specify the location of one or more DFLs. Intel has reserved the
+vendor specific id of 0x43 for this purpose. The vendor specific
+data begins with a 4 byte count of the number of DFLs followed 4 byte
+Offset/BIR fields for each DFL. Bits 2:0 of Offset/BIR field indicates
+the BAR, and bits 31:3 form the 8 byte aligned offset where bits 2:0 are
+zero.
Open discussion
===============
diff --git a/drivers/fpga/dfl-pci.c b/drivers/fpga/dfl-pci.c
index b1b157b41942..5418e8bf2496 100644
--- a/drivers/fpga/dfl-pci.c
+++ b/drivers/fpga/dfl-pci.c
@@ -27,6 +27,13 @@
#define DRV_VERSION "0.8"
#define DRV_NAME "dfl-pci"
+#define PCI_VNDR_ID_DFLS 0x43
+
+#define PCI_VNDR_DFLS_CNT_OFFSET 8
+#define PCI_VNDR_DFLS_RES_OFFSET 0x0c
+
+#define PCI_VND_DFLS_RES_BAR_MASK 0x7
We could define the mask by GENMASK().
Also another macro PCI_VND_DFLS_RES_OFFSET_MASK is needed.
I will use GENMASK and and add PCI_VND_DFLS_RES_OFFSET_MASK in v2.
+
struct cci_drvdata {
struct dfl_fpga_cdev *cdev; /* container device */
};
@@ -119,6 +126,82 @@ static int *cci_pci_create_irq_table(struct pci_dev *pcidev, unsigned int nvec)
return table;
}
+static int find_dfl_in_cfg(struct pci_dev *pcidev,
+ struct dfl_fpga_enum_info *info)
+{
+ u32 bar, offset, vndr_hdr, dfl_cnt, dfl_res;
+ int dfl_res_off, i, voff = 0;
+ resource_size_t start, len;
+
+ while ((voff = pci_find_next_ext_capability(pcidev, voff, PCI_EXT_CAP_ID_VNDR))) {
+
+ pci_read_config_dword(pcidev, voff + PCI_VNDR_HEADER, &vndr_hdr);
+
+ dev_dbg(&pcidev->dev,
+ "vendor-specific capability id 0x%x, rev 0x%x len 0x%x\n",
+ PCI_VNDR_HEADER_ID(vndr_hdr),
+ PCI_VNDR_HEADER_REV(vndr_hdr),
+ PCI_VNDR_HEADER_LEN(vndr_hdr));
+
+ if (PCI_VNDR_HEADER_ID(vndr_hdr) == PCI_VNDR_ID_DFLS)
+ break;
+ }
+
+ if (!voff) {
+ dev_dbg(&pcidev->dev, "%s no VSEC found\n", __func__);
+ return -ENODEV;
+ }
+
+ pci_read_config_dword(pcidev, voff + PCI_VNDR_DFLS_CNT_OFFSET, &dfl_cnt);
+ dev_info(&pcidev->dev, "dfl_cnt %d\n", dfl_cnt);
dev_dbg() is better?
I will change to dev_dbg in v2.
+ for (i = 0; i < dfl_cnt; i++) {
+ dfl_res_off = voff + PCI_VNDR_DFLS_RES_OFFSET +
+ (i * sizeof(dfl_res));
+ pci_read_config_dword(pcidev, dfl_res_off, &dfl_res);
+
+ dev_dbg(&pcidev->dev, "dfl_res 0x%x\n", dfl_res);
+
+ bar = dfl_res & PCI_VND_DFLS_RES_BAR_MASK;
FIELD_GET is better?
I think & will the GENMASK will be better because it will be
symetrical to the & below for the offset.
+
+ if (bar >= PCI_STD_NUM_BARS) {
+ dev_err(&pcidev->dev, "%s bad bar number %d\n",
+ __func__, bar);
+ return -EINVAL;
+ }
+
+ len = pci_resource_len(pcidev, bar);
+
+ if (len == 0) {
+ dev_err(&pcidev->dev, "%s unmapped bar number %d\n",
+ __func__, bar);
+ return -EINVAL;
+ }
+
+ offset = dfl_res & ~PCI_VND_DFLS_RES_BAR_MASK;
ditto
We don't want to use FIELD_GET here because we don't the shifting.
+
+ if (offset >= len) {
+ dev_err(&pcidev->dev, "%s bad offset %u >= %llu\n",
+ __func__, offset, len);
+ return -EINVAL;
+ }
+
+ dev_info(&pcidev->dev, "%s BAR %d offset 0x%x\n", __func__, bar, offset);
dev_dbg()?
I will change to dev_dbg in v2.
+
+ start = pci_resource_start(pcidev, bar) + offset;
+ len -= offset;
With these code, I have the following assumption:
1. There is only one DFL in one bar, multiple DFLs requires multiple
bars.
2. The DFL region is from the "offset" to the end of the bar.
Are they correct? If yes maybe we should specify them clearly in Doc.
This code would have the same assumptions as the existing code for finding
the dfls. The len value is only used during the walk of the DFL to
prevent walking too far. So I think one could have more than one DFL
on a particular bar as long as the start of the DFLs are different.
+
+ if (!PAGE_ALIGNED(start)) {
+ dev_err(&pcidev->dev, "%s unaliged start 0x%llx\n",
+ __func__, start);
+ return -EINVAL;
+ }
+
+ dfl_fpga_enum_info_add_dfl(info, start, len);
Do we need some region overlapping check in this func? So we could find
the HW problem (e.g. same bar num for multiple DFLs) in early stage.
I think whatever overlapping check would also need to be in the existing
code because the logic is the same.
+ }
+
+ return 0;
+}
+
static int find_dfl_in_bar0(struct pci_dev *pcidev,
struct dfl_fpga_enum_info *info)
{
@@ -221,7 +304,10 @@ static int cci_enumerate_feature_devs(struct pci_dev *pcidev)
goto irq_free_exit;
}
- ret = find_dfl_in_bar0(pcidev, info);
+ ret = find_dfl_in_cfg(pcidev, info);
+
+ if (ret)
+ ret = find_dfl_in_bar0(pcidev, info);
The patch is more than the relocation support for DFL. Actually it
introduced a different way of DFL finding.
Previously it starts at bar0 offset 0, find dfl fme first, then find
dfl port according to fme header registers. Now it enumerates every DFL
by PCIe VSEC.
Maybe we should add more description about the change and why.
I will highlight this difference in the documentation in v2.
Thanks,
Yilun
if (ret)
goto irq_free_exit;
--
2.25.2