On 2/17/22 10:53 PM, Zhang, Tianfei wrote:
-----Original Message-----
From: Tom Rix <trix@xxxxxxxxxx>
Sent: Tuesday, February 15, 2022 10:49 PM
To: Zhang, Tianfei <tianfei.zhang@xxxxxxxxx>; Wu, Hao <hao.wu@xxxxxxxxx>;
mdf@xxxxxxxxxx; Xu, Yilun <yilun.xu@xxxxxxxxx>; linux-fpga@xxxxxxxxxxxxxxx;
linux-doc@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
Cc: corbet@xxxxxxx
Subject: Re: [PATCH v1 2/7] fpga: dfl: check feature type before parse irq info
On 2/14/22 3:26 AM, Tianfei zhang wrote:
From: Tianfei Zhang <tianfei.zhang@xxxxxxxxx>
The feature ID of "Port User Interrupt" and the "PMCI Subsystem" are
identical, 0x12, but one is for FME, other is for Port. It should
check the feature type While parsing the irq info in
parse_feature_irqs().
This seems like a bug fix and not part of iofs feature.
Split this out of the patchset.
?
This is a workaround a hardware problem, there should be some comments to
the effect that you can't trust _this_ or _that_ feature id and some special
handling earlier.
The ambiguity of feature id is a problem, and this sort of bug will happen again.
What can be done to prevent this in the future ?
This patch is not workaround, this is a bug fix for DFL driver.
The root cause is that DLF driver miss check the feature type while parsing the interrupt information,
because some Feature IDs are identical between FME and Port, like PMCI in FME and "Port User Interrupt"
in Port.
The definition of Feature ID is here:
https://github.com/OPAE/linux-dfl-feature-id/blob/master/dfl-feature-ids.rst
Helpful but hidden. At least a link to this should be added to
Documentation/fpga/dfl.rst.
Signed-off-by: Tianfei Zhang <tianfei.zhang@xxxxxxxxx>
---
drivers/fpga/dfl.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/drivers/fpga/dfl.c b/drivers/fpga/dfl.c index
599bb21d86af..26f8cf890700 100644
--- a/drivers/fpga/dfl.c
+++ b/drivers/fpga/dfl.c
@@ -940,9 +940,14 @@ static int parse_feature_irqs(struct
build_feature_devs_info *binfo,
{
void __iomem *base = binfo->ioaddr + ofst;
unsigned int i, ibase, inr = 0;
+ enum dfl_id_type type;
int virq;
u64 v;
+ type = feature_dev_id_type(binfo->feature_dev);
+ if (type >= DFL_ID_MAX)
+ return -EINVAL;
+
/*
* Ideally DFL framework should only read info from DFL header, but
* current version DFL only provides mmio resources information for
@@ -959,16 +964,22 @@ static int parse_feature_irqs(struct
build_feature_devs_info *binfo,
*/
switch (fid) {
case PORT_FEATURE_ID_UINT:
+ if (type != PORT_ID)
+ break;
Instead of embedding a break in the switch, break the switch into fme switch
and port switch
if (type == PORT_ID)
port-switch
else if (type == FME_ID
fme-switch
Your suggestion is looks good for me, I will change on next version.
Tom
v = readq(base + PORT_UINT_CAP);
ibase = FIELD_GET(PORT_UINT_CAP_FST_VECT, v);
inr = FIELD_GET(PORT_UINT_CAP_INT_NUM, v);
break;
case PORT_FEATURE_ID_ERROR:
+ if (type != PORT_ID)
+ break;
v = readq(base + PORT_ERROR_CAP);
ibase = FIELD_GET(PORT_ERROR_CAP_INT_VECT, v);
inr = FIELD_GET(PORT_ERROR_CAP_SUPP_INT, v);
break;
case FME_FEATURE_ID_GLOBAL_ERR:
+ if (type != FME_ID)
+ break;
v = readq(base + FME_ERROR_CAP);
ibase = FIELD_GET(FME_ERROR_CAP_INT_VECT, v);
inr = FIELD_GET(FME_ERROR_CAP_SUPP_INT, v);
