[Skiboot] [PATCH skiboot] npu: Implement FLR

[Alexey Kardashevskiy]

As the comment in npu_dev_populate_pcie_cap() says,
"We should support FLR" and the NPU device advertises its
support. However, when the kernel issues FLR, skiboot does
nothing which leaves NPU in a state which does not allow
to use NV links again after GPU was reset.

This adds basic handling of FLR (function level reset).

Signed-off-by: Alexey Kardashevskiy <aik at ozlabs.ru>
---

With this fix, the bandwidthTest reports PINNED Memory Transfers
>= 20000MB/s, without it only first guest boot reports this much
of bandwidth, all consequent guests only report 6000MB/s maximum.


---
 hw/npu.c | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/hw/npu.c b/hw/npu.c
index 302a5518..6a535885 100644
--- a/hw/npu.c
+++ b/hw/npu.c
@@ -271,6 +271,26 @@ static int64_t npu_dev_cfg_bar(void *dev, struct pci_cfg_reg_filter *pcrf,
 	return npu_dev_cfg_bar_read(ndev, pcrf, offset, len, data);
 }
 
+static int64_t npu_dev_cfg_exp_devcap(void *dev,
+		struct pci_cfg_reg_filter *pcrf __unused,
+		uint32_t offset __unused, uint32_t size __unused,
+		uint32_t *data, bool write)
+{
+	struct pci_virt_device *pvd = dev;
+	struct npu_dev *ndev = pvd->data;
+
+	assert(write);
+
+	if (*data & PCICAP_EXP_DEVCTL_FUNC_RESET) { /* PCI_EXP_DEVCTL_BCR_FLR */
+		ndev->procedure_status = 0;
+		ndev->procedure_number = 0;
+		ndev->procedure_step = 0;
+		ndev->procedure_data = 0;
+	}
+
+	return OPAL_PARTIAL;
+}
+
 static struct npu_dev *bdfn_to_npu_dev(struct npu *p, uint32_t bdfn)
 {
 	struct pci_virt_device *pvd;
@@ -1198,6 +1218,10 @@ static void npu_dev_populate_pcie_cap(struct npu_dev_cap *cap)
 	       (PCICAP_EXP_DEVCAP_FUNC_RESET));
 	PCI_VIRT_CFG_INIT_RO(pvd, base + PCICAP_EXP_DEVCAP, 4, val);
 
+	pci_virt_add_filter(pvd, base + PCICAP_EXP_DEVCTL, 2,
+			    PCI_REG_FLAG_WRITE,
+			    npu_dev_cfg_exp_devcap, NULL);
+
 	/* 0x08 - Device control and status */
 	PCI_VIRT_CFG_INIT(pvd, base + PCICAP_EXP_DEVCTL, 4, 0x00002810,
 			 0xffff0000, 0x000f0000);
-- 
2.11.0