From: Shivesh Optimizes mailbox polling with exponential backoff rather than a fixed msleep(10) loop. Fixes define names. Eliminates coarse msleep(50) delays during IRQ teardown by replacing them with tight usleep_range polling. Signed-off-by: Shivesh --- .../broadcom/brcm80211/brcmfmac/pcie.c | 66 ++++++++++++++++--- 1 file changed, 58 insertions(+), 8 deletions(-) diff --git a/drivers/net/wireless/broadcom/brcm80211/brcmfmac/pcie.c b/drivers/net/wireless/broadcom/brcm80211/brcmfmac/pcie.c index 13662aa4b4ea..9338a5faa260 100644 --- a/drivers/net/wireless/broadcom/brcm80211/brcmfmac/pcie.c +++ b/drivers/net/wireless/broadcom/brcm80211/brcmfmac/pcie.c @@ -268,6 +268,27 @@ static const struct brcmf_firmware_mapping brcmf_pcie_fwnames[] = { #define BRCMF_PCIE_MBDATA_TIMEOUT msecs_to_jiffies(2000) +/* + * H2D mailbox poll timing parameters. + * + * The dongle typically clears the H2D mailbox register within a few + * hundred microseconds after the doorbell interrupt fires. The + * original code used msleep(10) * 100 iterations, meaning the + * minimum observable latency was 10ms even when the dongle was fast. + * + * We instead start with a short sleep and double it each iteration + * (exponential backoff) up to BRCMF_PCIE_MB_POLL_MAX_US, staying + * within the same 1-second absolute timeout. + * + * MIN_US / INITIAL_MAX_US : usleep_range bounds for the first iteration. + * MAX_US : cap on the per-iteration sleep (µs). + * TIMEOUT_US : total budget before giving up (1 second). + */ +#define BRCMF_PCIE_MB_POLL_MIN_US 40 +#define BRCMF_PCIE_MB_POLL_INITIAL_MAX_US 50 +#define BRCMF_PCIE_MB_POLL_MAX_US 5000 +#define BRCMF_PCIE_MB_POLL_TIMEOUT_US 1000000 + #define BRCMF_PCIE_CFGREG_STATUS_CMD 0x4 #define BRCMF_PCIE_CFGREG_PM_CSR 0x4C #define BRCMF_PCIE_CFGREG_MSI_CAP 0x58 @@ -766,7 +787,8 @@ brcmf_pcie_send_mb_data(struct brcmf_pciedev_info *devinfo, u32 htod_mb_data) struct brcmf_core *core; u32 addr; u32 cur_htod_mb_data; - u32 i; + u32 elapsed_us = 0; + u32 sleep_us = BRCMF_PCIE_MB_POLL_INITIAL_MAX_US; shared = &devinfo->shared; addr = shared->htod_mb_data_addr; @@ -776,12 +798,40 @@ brcmf_pcie_send_mb_data(struct brcmf_pciedev_info *devinfo, u32 htod_mb_data) brcmf_dbg(PCIE, "MB transaction is already pending 0x%04x\n", cur_htod_mb_data); - i = 0; + /* + * Wait for the dongle to consume the previous H2D mailbox message. + * + * There is no interrupt that signals when the dongle clears this + * register, so polling is unavoidable. The original code used + * msleep(10) per iteration, incurring at least 10ms of latency + * even when the dongle responded in microseconds. + * + * We use usleep_range() with exponential backoff instead: + * - First iteration sleeps ~50µs (fast path for responsive dongle). + * - Each subsequent iteration doubles the sleep, capped at 5ms, + * so long waits still yield the CPU without busy-spinning. + * - Total timeout matches the original 1-second limit. + * - We bail early if the device has gone down so that a dead + * dongle does not hold the caller for a full second. + */ while (cur_htod_mb_data != 0) { - msleep(10); - i++; - if (i > 100) + if (devinfo->state == BRCMFMAC_PCIE_STATE_DOWN) { + brcmf_dbg(PCIE, "Device down, aborting MB send\n"); return -EIO; + } + + if (elapsed_us >= BRCMF_PCIE_MB_POLL_TIMEOUT_US) { + brcmf_err("Timeout waiting for H2D MB slot after %u us\n", + elapsed_us); + return -EIO; + } + + usleep_range(BRCMF_PCIE_MB_POLL_MIN_US, sleep_us); + elapsed_us += sleep_us; + + /* Exponential backoff, capped at BRCMF_PCIE_MB_POLL_MAX_US */ + sleep_us = min(sleep_us * 2, (u32)BRCMF_PCIE_MB_POLL_MAX_US); + cur_htod_mb_data = brcmf_pcie_read_tcm32(devinfo, addr); } @@ -1001,10 +1051,10 @@ static void brcmf_pcie_release_irq(struct brcmf_pciedev_info *devinfo) free_irq(pdev->irq, devinfo); pci_disable_msi(pdev); - msleep(50); + usleep_range(1000, 2000); count = 0; - while ((devinfo->in_irq) && (count < 20)) { - msleep(50); + while ((devinfo->in_irq) && (count < 1000)) { + usleep_range(1000, 2000); count++; } if (devinfo->in_irq) -- 2.53.0