On 4/8/21 11:02 AM, Alexander Motin wrote:
On 08.04.2021 13:36, John Baldwin wrote:
On 4/5/21 7:40 AM, Alexander Motin wrote:
+ rmps = pcie_read_config(root, PCIER_DEVICE_CTL, 2) &
+ PCIEM_CTL_MAX_PAYLOAD;
+ mps = pcie_read_config(dev, PCIER_DEVICE_CTL, 2) &
+ PCIEM_CTL_MAX_PAYLOAD;
+ if (mps == rmps)
+ return;
+ /* Check whether the device is capable of the root's MPS. */
+ mmps = (pcie_read_config(dev, PCIER_DEVICE_CAP, 2) &
+ PCIEM_CAP_MAX_PAYLOAD) << 5;
+ if (rmps > mmps) {
+ /*
+ * The device is unable to handle root's MPS. Limit root.
+ * XXX: We should traverse through all the tree, applying
+ * it to all the devices.
+ */
Hmmm, limiting the root seems very dubious here. Do you really need this?
All devices under the same root (at least ones that talk to each other)
must have the same MPS value, otherwise some may consider larger
transfer as an error. In case of direct PCIe connection the root is the
only other device, so this code should be sufficient.
I mean, had you seen any cases where you needed to adjust the root? I
do worry about this breaking other systems due to it not doing the walk
of the full sub-tree. Maybe only do it if the root port is the grandparent
of the device in question since that is the safe case here and punt if
there are switches in the middle?
If not, I'd put it behind a tunable sysctl that defaults to off. Ideally
what you'd do here is use an and of the two masks and select one of those
bits to choose the value rather than assuming the root can do the device's
value.
It is not a bitmask, it is a power-of-2 between 128 bytes and the
maximum device capability. So if the root is already configured for
higher value, then it must support the lover one too.
I thought the DEVICE_CAP is a bitmask, and that is what I was thinking of using
the and with. Nominally I think we should be doing the AND of the DEVICE_CAP
field for all the devices under a port and then programming them to some value
remaining in the mask. Most devices just DMA to/from memory though rather than
to each other, so generally you just care about the path from a device to a root
port.
--
John Baldwin
_______________________________________________
dev-commits-src-main@freebsd.org mailing list
https://lists.freebsd.org/mailman/listinfo/dev-commits-src-main
To unsubscribe, send any mail to "dev-commits-src-main-unsubscr...@freebsd.org"
