On Fri, Aug 15, 2025 at 06:20:54PM GMT, Frank Li wrote:
> Add MSI doorbell support to reduce latency between PCI host and EP.
>
> Before this change:
> ping 169.254.172.137
> 64 bytes from 169.254.172.137: icmp_seq=1 ttl=64 time=0.575 ms
> 64 bytes from 169.254.172.137: icmp_seq=2 ttl=64 time=1.80 ms
> 64 bytes from 169.254.172.137: icmp_seq=3 ttl=64 time=8.19 ms
> 64 bytes from 169.254.172.137: icmp_seq=4 ttl=64 time=2.00 ms
>
> After this change:
> ping 169.254.144.71
> 64 bytes from 169.254.144.71: icmp_seq=1 ttl=64 time=0.215 ms
> 64 bytes from 169.254.144.71: icmp_seq=2 ttl=64 time=0.456 ms
> 64 bytes from 169.254.144.71: icmp_seq=3 ttl=64 time=0.448 ms
>
> Change u64 db to atomic_64 because difference doorbell may happen at the
> same time.
>
Only the atomicity of db variable is enough?
> Signed-off-by: Frank Li <Frank.Li@xxxxxxx>
> ---
> drivers/pci/endpoint/functions/pci-epf-vntb.c | 153 +++++++++++++++++++++++---
> 1 file changed, 136 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/pci/endpoint/functions/pci-epf-vntb.c b/drivers/pci/endpoint/functions/pci-epf-vntb.c
> index 83e9ab10f9c4fc2b485d5463faa2172500f12999..1c586205835fe9c7c5352e74819bccb4ece84438 100644
> --- a/drivers/pci/endpoint/functions/pci-epf-vntb.c
> +++ b/drivers/pci/endpoint/functions/pci-epf-vntb.c
> @@ -36,11 +36,13 @@
> * PCIe Root Port PCI EP
> */
>
> +#include <linux/atomic.h>
> #include <linux/delay.h>
> #include <linux/io.h>
> #include <linux/module.h>
> #include <linux/slab.h>
>
> +#include <linux/pci-ep-msi.h>
> #include <linux/pci-epc.h>
> #include <linux/pci-epf.h>
> #include <linux/ntb.h>
> @@ -126,12 +128,13 @@ struct epf_ntb {
> u32 db_count;
> u32 spad_count;
> u64 mws_size[MAX_MW];
> - u64 db;
> + atomic64_t db;
> u32 vbus_number;
> u16 vntb_pid;
> u16 vntb_vid;
>
> bool linkup;
> + bool msi_doorbell;
> u32 spad_size;
>
> enum pci_barno epf_ntb_bar[VNTB_BAR_NUM];
> @@ -258,9 +261,9 @@ static void epf_ntb_cmd_handler(struct work_struct *work)
>
> ntb = container_of(work, struct epf_ntb, cmd_handler.work);
>
> - for (i = 1; i < ntb->db_count; i++) {
> + for (i = 1; i < ntb->db_count && !ntb->msi_doorbell; i++) {
> if (ntb->epf_db[i]) {
> - ntb->db |= 1 << (i - 1);
> + atomic64_or(1 << (i - 1), &ntb->db);
> ntb_db_event(&ntb->ntb, i);
> ntb->epf_db[i] = 0;
> }
> @@ -319,7 +322,24 @@ static void epf_ntb_cmd_handler(struct work_struct *work)
>
> reset_handler:
> queue_delayed_work(kpcintb_workqueue, &ntb->cmd_handler,
> - msecs_to_jiffies(5));
> + ntb->msi_doorbell ? msecs_to_jiffies(500) : msecs_to_jiffies(5));
> +}
> +
> +static irqreturn_t epf_ntb_doorbell_handler(int irq, void *data)
> +{
> + struct epf_ntb *ntb = data;
> + int i = 0;
> +
> + for (i = 1; i < ntb->db_count; i++)
> + if (irq == ntb->epf->db_msg[i].virq) {
> + atomic64_or(1 << (i - 1), &ntb->db);
> + ntb_db_event(&ntb->ntb, i);
> + }
> +
> + if (irq == ntb->epf->db_msg[0].virq)
> + queue_delayed_work(kpcintb_workqueue, &ntb->cmd_handler, 0);
> +
> + return IRQ_HANDLED;
> }
>
> /**
> @@ -500,6 +520,90 @@ static int epf_ntb_configure_interrupt(struct epf_ntb *ntb)
> return 0;
> }
>
> +static int epf_ntb_db_bar_init_msi_doorbell(struct epf_ntb *ntb,
> + struct pci_epf_bar *db_bar,
> + const struct pci_epc_features *epc_features,
> + enum pci_barno barno)
> +{
> + struct pci_epf *epf = ntb->epf;
> + dma_addr_t low, high;
> + struct msi_msg *msg;
> + size_t sz;
> + int ret;
> + int i;
> +
> + ret = pci_epf_alloc_doorbell(epf, ntb->db_count);
> + if (ret)
> + return ret;
> +
> + for (i = 0; i < ntb->db_count; i++) {
> + ret = request_irq(epf->db_msg[i].virq, epf_ntb_doorbell_handler,
> + 0, "vntb_db", ntb);
> +
> + if (ret) {
> + dev_err(&epf->dev,
> + "Failed to request doorbell IRQ: %d\n",
> + epf->db_msg[i].virq);
> + goto err_request_irq;
> + }
> + }
> +
> + msg = &epf->db_msg[0].msg;
> +
> + high = 0;
> + low = (u64)msg->address_hi << 32 | msg->address_lo;
Can you remind me when the 'address_{hi/lo}' pairs are set?
Rest looks OK to me.
- Mani
--
மணிவண்ணன் சதாசிவம்
