On Tue, Sep 22, 2020 at 12:22 PM Michal Kubecek <mkube...@suse.cz> wrote:
>
> On Tue, Sep 22, 2020 at 11:24:24AM +0530, Vasundhara Volam wrote:
> > On Mon, Sep 21, 2020 at 2:48 PM Michal Kubecek <mkube...@suse.cz> wrote:
> > > > + return -1;
> > > > + }
> > > > +
> > > > + pcie_stats = (u16 *)(regs->data + BNXT_PXP_REG_LEN);
> > > > + fprintf(stdout, "PCIe statistics:\n");
> > > > + fprintf(stdout, "----------------\n");
> > > > + for (i = 0; i < ARRAY_SIZE(bnxt_pcie_stats); i++) {
> > > > + pcie_stat = 0;
> > > > + memcpy(&pcie_stat, &pcie_stats[stats[i].offset],
> > > > + stats[i].size * sizeof(u16));
> > >
> > > This will only work on little endian architectures.
> >
> > Data is already converted to host endian order by ETHTOOL_REGS, so it
> > will not be an issue.
>
> It does not work correctly. Assume we are on big endian architecture and
> are reading a 16-bit value (stats[i].size = 1) 0x1234 which is laid out
> in memory as
>
> ... 12 34 ...
>
> Copying that by memcpy() to the address of 64-bit pcie_stat, you get
>
> 12 34 00 00 00 00 00 00
>
> which represents 0x1234000000000000, not 0x1234. You will also have the
> same problem with 32-bit values (stats[i].size = 2).
You are right. I understood the issue now.
I will modify it to use different size variables based on the size and
convert it to a switch statement.
Thanks.
>
> Michal
