Re: [PATCH v4 02/12] PCI: endpoint: pci-epf-vntb: Defer pci_epc_raise_irq() out of atomic context

On Wed, May 13, 2026 at 11:49:13AM +0900, Koichiro Den wrote:
> The NTB .peer_db_set() callback may be invoked from atomic context.
> pci-epf-vntb currently calls pci_epc_raise_irq() directly, but
> pci_epc_raise_irq() may sleep (it takes epc->lock).
>
> Avoid sleeping in atomic context by coalescing doorbell bits into an
> atomic64 pending mask and raising MSIs from a work item. Limit the
> amount of work per run to avoid monopolizing the workqueue under a
> doorbell storm.
>
> Clear stale pending bits before enabling the work item and after disabling
> it during cleanup. Also mask requested doorbells against the currently
> valid doorbell mask before queueing work, and iterate the pending u64 with
> __ffs64() so high doorbell bits are handled correctly.
>
> Fixes: e35f56bb0330 ("PCI: endpoint: Support NTB transfer between RC and EP")
> Signed-off-by: Koichiro Den <[email protected]>
> ---

Reviewed-by: Frank Li <[email protected]>

> Changes since v3:
>   - Clear peer_db_pending before enabling and after disabling the work item.
>   - Mask peer_db_set() requests with db_valid_mask before queueing work.
>   - Replace ffs() iteration with u64-safe __ffs64() iteration.
>   - Not pick Frank's Reviewed-by tag for v3 02/10 due to the added behavior.
>
>  drivers/pci/endpoint/functions/pci-epf-vntb.c | 112 +++++++++++++-----
>  1 file changed, 85 insertions(+), 27 deletions(-)
>
> diff --git a/drivers/pci/endpoint/functions/pci-epf-vntb.c b/drivers/pci/endpoint/functions/pci-epf-vntb.c
> index 668d25abc7f2..cc0b356973f3 100644
> --- a/drivers/pci/endpoint/functions/pci-epf-vntb.c
> +++ b/drivers/pci/endpoint/functions/pci-epf-vntb.c
> @@ -37,6 +37,7 @@
>   */
>
>  #include <linux/atomic.h>
> +#include <linux/bitops.h>
>  #include <linux/delay.h>
>  #include <linux/io.h>
>  #include <linux/module.h>
> @@ -69,6 +70,9 @@ static struct workqueue_struct *kpcintb_workqueue;
>  #define MAX_DB_COUNT			32
>  #define MAX_MW				4
>
> +/* Limit per-work execution to avoid monopolizing kworker on doorbell storms. */
> +#define VNTB_PEER_DB_WORK_BUDGET	5
> +
>  enum epf_ntb_bar {
>  	BAR_CONFIG,
>  	BAR_DB,
> @@ -129,6 +133,8 @@ struct epf_ntb {
>  	u32 spad_count;
>  	u64 mws_size[MAX_MW];
>  	atomic64_t db;
> +	atomic64_t peer_db_pending;
> +	struct work_struct peer_db_work;
>  	u32 vbus_number;
>  	u16 vntb_pid;
>  	u16 vntb_vid;
> @@ -972,6 +978,9 @@ static int epf_ntb_epc_init(struct epf_ntb *ntb)
>  	INIT_DELAYED_WORK(&ntb->cmd_handler, epf_ntb_cmd_handler);
>  	queue_work(kpcintb_workqueue, &ntb->cmd_handler.work);
>
> +	atomic64_set(&ntb->peer_db_pending, 0);
> +	enable_work(&ntb->peer_db_work);
> +
>  	return 0;
>
>  err_write_header:
> @@ -995,6 +1004,8 @@ static int epf_ntb_epc_init(struct epf_ntb *ntb)
>  static void epf_ntb_epc_cleanup(struct epf_ntb *ntb)
>  {
>  	disable_delayed_work_sync(&ntb->cmd_handler);
> +	disable_work_sync(&ntb->peer_db_work);
> +	atomic64_set(&ntb->peer_db_pending, 0);
>  	epf_ntb_mw_bar_clear(ntb, ntb->num_mws);
>  	epf_ntb_db_bar_clear(ntb);
>  	epf_ntb_config_sspad_bar_clear(ntb);
> @@ -1409,41 +1420,84 @@ static int vntb_epf_peer_spad_write(struct ntb_dev *ndev, int pidx, int idx, u32
>  	return 0;
>  }
>
> +static void vntb_epf_peer_db_work(struct work_struct *work)
> +{
> +	struct epf_ntb *ntb = container_of(work, struct epf_ntb, peer_db_work);
> +	struct pci_epf *epf = ntb->epf;
> +	unsigned int budget = VNTB_PEER_DB_WORK_BUDGET;
> +	u8 func_no, vfunc_no;
> +	unsigned int db_bit;
> +	u32 interrupt_num;
> +	u64 db_bits;
> +	int ret;
> +
> +	if (!epf || !epf->epc)
> +		return;
> +
> +	func_no = epf->func_no;
> +	vfunc_no = epf->vfunc_no;
> +
> +	/*
> +	 * Drain doorbells from peer_db_pending in snapshots (atomic64_xchg()).
> +	 * Limit the number of snapshots handled per run so we don't monopolize
> +	 * the workqueue under a doorbell storm.
> +	 */
> +	while (budget--) {
> +		db_bits = atomic64_xchg(&ntb->peer_db_pending, 0);
> +		if (!db_bits)
> +			return;
> +
> +		while (db_bits) {
> +			/*
> +			 * pci_epc_raise_irq() for MSI expects a 1-based
> +			 * interrupt number. db_bit is zero-based, so add 3 to
> +			 * preserve the historical slot offset.
> +			 *
> +			 * Legacy mapping (kept for compatibility):
> +			 *
> +			 *   MSI #1 : link event (reserved)
> +			 *   MSI #2 : unused (historical offset)
> +			 *   MSI #3 : doorbell bit 0 (DB#0)
> +			 *   MSI #4 : doorbell bit 1 (DB#1)
> +			 *   ...
> +			 *
> +			 * Do not change this mapping to avoid breaking
> +			 * interoperability with older peers.
> +			 */
> +			db_bit = __ffs64(db_bits);
> +			interrupt_num = db_bit + 3;
> +			db_bits &= ~BIT_ULL(db_bit);
> +
> +			ret = pci_epc_raise_irq(epf->epc, func_no, vfunc_no,
> +						PCI_IRQ_MSI, interrupt_num);
> +			if (ret)
> +				dev_err(&ntb->ntb.dev,
> +					"Failed to raise IRQ for interrupt_num %u: %d\n",
> +					interrupt_num, ret);
> +		}
> +	}
> +
> +	if (atomic64_read(&ntb->peer_db_pending))
> +		queue_work(kpcintb_workqueue, &ntb->peer_db_work);
> +}
> +
>  static int vntb_epf_peer_db_set(struct ntb_dev *ndev, u64 db_bits)
>  {
> -	u32 interrupt_num = ffs(db_bits) + 1;
>  	struct epf_ntb *ntb = ntb_ndev(ndev);
> -	u8 func_no, vfunc_no;
> -	int ret;
>
> -	func_no = ntb->epf->func_no;
> -	vfunc_no = ntb->epf->vfunc_no;
> +	db_bits &= vntb_epf_db_valid_mask(ndev);
> +	if (!db_bits)
> +		return 0;
>
>  	/*
> -	 * pci_epc_raise_irq() for MSI expects a 1-based interrupt number.
> -	 * ffs() returns a 1-based index (bit 0 -> 1). interrupt_num has already
> -	 * been computed as ffs(db_bits) + 1 above. Adding one more +1 when
> -	 * calling pci_epc_raise_irq() therefore results in:
> -	 *
> -	 *   doorbell bit 0 -> MSI #3
> -	 *
> -	 * Legacy mapping (kept for compatibility):
> -	 *
> -	 *   MSI #1 : link event (reserved)
> -	 *   MSI #2 : unused (historical offset)
> -	 *   MSI #3 : doorbell bit 0 (DB#0)
> -	 *   MSI #4 : doorbell bit 1 (DB#1)
> -	 *   ...
> -	 *
> -	 * Do not change this mapping to avoid breaking interoperability with
> -	 * older peers.
> +	 * .peer_db_set() may be called from atomic context. pci_epc_raise_irq()
> +	 * can sleep (it takes epc->lock), so defer MSI raising to process
> +	 * context. Doorbell requests are coalesced in peer_db_pending.
>  	 */
> -	ret = pci_epc_raise_irq(ntb->epf->epc, func_no, vfunc_no,
> -				PCI_IRQ_MSI, interrupt_num + 1);
> -	if (ret)
> -		dev_err(&ntb->ntb.dev, "Failed to raise IRQ\n");
> +	atomic64_or(db_bits, &ntb->peer_db_pending);
> +	queue_work(kpcintb_workqueue, &ntb->peer_db_work);
>
> -	return ret;
> +	return 0;
>  }
>
>  static u64 vntb_epf_db_read(struct ntb_dev *ndev)
> @@ -1690,6 +1744,10 @@ static int epf_ntb_probe(struct pci_epf *epf,
>  	ntb->epf = epf;
>  	ntb->vbus_number = 0xff;
>
> +	INIT_WORK(&ntb->peer_db_work, vntb_epf_peer_db_work);
> +	disable_work(&ntb->peer_db_work);
> +	atomic64_set(&ntb->peer_db_pending, 0);
> +
>  	/* Initially, no bar is assigned */
>  	for (i = 0; i < VNTB_BAR_NUM; i++)
>  		ntb->epf_ntb_bar[i] = NO_BAR;
> --
> 2.51.0
>