git: f3aa0098a82e - main - Use mtx_lock_spin in the gic driver
Andrew Turner
andrew at fubar.geek.nz
Wed Sep 29 15:43:03 UTC 2021
> On 28 Sep 2021, at 16:36, Marcin Wojtas <mw at semihalf.com> wrote:
>
> Hi Andrew,
>
> wt., 28 wrz 2021 o 14:43 Andrew Turner <andrew at freebsd.org> napisał(a):
>>
>> The branch main has been updated by andrew:
>>
>> URL: https://cgit.FreeBSD.org/src/commit/?id=f3aa0098a82ebf7712aa13716d794aa7e4ac59cd
>>
>> commit f3aa0098a82ebf7712aa13716d794aa7e4ac59cd
>> Author: Andrew Turner <andrew at FreeBSD.org>
>> AuthorDate: 2021-09-28 11:36:42 +0000
>> Commit: Andrew Turner <andrew at FreeBSD.org>
>> CommitDate: 2021-09-28 11:42:06 +0000
>>
>> Use mtx_lock_spin in the gic driver
>>
>> The mutex was changed to a spin lock when the MSI/MSI-X handling was
>> moved from the gicv2m to the gic driver. Update the calls to lock
>> and unlock the mutex to the spin variant.
>>
>> Submitted by: jrtc27 ("Change all the mtx_(un)lock(&sc->mutex) to be the _spin versions.")
>> Reported by: mw, antranigv at freebsd.am
>> Sponsored by: The FreeBSD Foundation
>> ---
>> sys/arm/arm/gic.c | 20 ++++++++++----------
>> 1 file changed, 10 insertions(+), 10 deletions(-)
>>
>> diff --git a/sys/arm/arm/gic.c b/sys/arm/arm/gic.c
>> index d7edd7885404..89db4e324600 100644
>> --- a/sys/arm/arm/gic.c
>> +++ b/sys/arm/arm/gic.c
>> @@ -1056,7 +1056,7 @@ arm_gic_alloc_msi(device_t dev, device_t child, int count, int maxcount,
>>
>> sc = device_get_softc(dev);
>>
>> - mtx_lock(&sc->mutex);
>> + mtx_lock_spin(&sc->mutex);
>>
>> found = false;
>> for (irq = sc->sc_spi_start; irq < sc->sc_spi_end; irq++) {
>> @@ -1091,7 +1091,7 @@ arm_gic_alloc_msi(device_t dev, device_t child, int count, int maxcount,
>>
>> /* Not enough interrupts were found */
>> if (!found || irq == sc->sc_spi_end) {
>> - mtx_unlock(&sc->mutex);
>> + mtx_unlock_spin(&sc->mutex);
>> return (ENXIO);
>> }
>>
>> @@ -1099,7 +1099,7 @@ arm_gic_alloc_msi(device_t dev, device_t child, int count, int maxcount,
>> /* Mark the interrupt as used */
>> sc->gic_irqs[irq + i].gi_flags |= GI_FLAG_MSI_USED;
>> }
>> - mtx_unlock(&sc->mutex);
>> + mtx_unlock_spin(&sc->mutex);
>>
>> for (i = 0; i < count; i++)
>> srcs[i] = (struct intr_irqsrc *)&sc->gic_irqs[irq + i];
>> @@ -1118,7 +1118,7 @@ arm_gic_release_msi(device_t dev, device_t child, int count,
>>
>> sc = device_get_softc(dev);
>>
>> - mtx_lock(&sc->mutex);
>> + mtx_lock_spin(&sc->mutex);
>> for (i = 0; i < count; i++) {
>> gi = (struct gic_irqsrc *)isrc[i];
>>
>> @@ -1128,7 +1128,7 @@ arm_gic_release_msi(device_t dev, device_t child, int count,
>>
>> gi->gi_flags &= ~GI_FLAG_MSI_USED;
>> }
>> - mtx_unlock(&sc->mutex);
>> + mtx_unlock_spin(&sc->mutex);
>>
>> return (0);
>> }
>> @@ -1142,7 +1142,7 @@ arm_gic_alloc_msix(device_t dev, device_t child, device_t *pic,
>>
>> sc = device_get_softc(dev);
>>
>> - mtx_lock(&sc->mutex);
>> + mtx_lock_spin(&sc->mutex);
>> /* Find an unused interrupt */
>> for (irq = sc->sc_spi_start; irq < sc->sc_spi_end; irq++) {
>> KASSERT((sc->gic_irqs[irq].gi_flags & GI_FLAG_MSI) != 0,
>> @@ -1152,13 +1152,13 @@ arm_gic_alloc_msix(device_t dev, device_t child, device_t *pic,
>> }
>> /* No free interrupt was found */
>> if (irq == sc->sc_spi_end) {
>> - mtx_unlock(&sc->mutex);
>> + mtx_unlock_spin(&sc->mutex);
>> return (ENXIO);
>> }
>>
>> /* Mark the interrupt as used */
>> sc->gic_irqs[irq].gi_flags |= GI_FLAG_MSI_USED;
>> - mtx_unlock(&sc->mutex);
>> + mtx_unlock_spin(&sc->mutex);
>>
>> *isrcp = (struct intr_irqsrc *)&sc->gic_irqs[irq];
>> *pic = dev;
>> @@ -1178,9 +1178,9 @@ arm_gic_release_msix(device_t dev, device_t child, struct intr_irqsrc *isrc)
>> KASSERT((gi->gi_flags & GI_FLAG_MSI_USED) == GI_FLAG_MSI_USED,
>> ("%s: Trying to release an unused MSI-X interrupt", __func__));
>>
>> - mtx_lock(&sc->mutex);
>> + mtx_lock_spin(&sc->mutex);
>> gi->gi_flags &= ~GI_FLAG_MSI_USED;
>> - mtx_unlock(&sc->mutex);
>> + mtx_unlock_spin(&sc->mutex);
>>
>> return (0);
>> }
>
> Thank you for the patch, it fixes the MSI-X in my setup, but in order
> to operate properly on Marvell SoCs (equipped with a standard GICv2m),
> I have to remove the asserts, which were added in c6f3076d4405 (Move
> the GICv2m msi handling to the parent):
>
> diff --git a/sys/arm/arm/gic.c b/sys/arm/arm/gic.c
> index 89db4e324600..b5d72a2a6c49 100644
> --- a/sys/arm/arm/gic.c
> +++ b/sys/arm/arm/gic.c
> @@ -528,8 +528,6 @@ arm_gic_write_ivar(device_t dev, device_t child,
> int which, uintptr_t value)
> * GIC_IVAR_MBI_START must be set once and first. This allows
> * us to reserve the registers when GIC_IVAR_MBI_COUNT is set.
> */
> - MPASS(sc->sc_spi_start == 0);
> - MPASS(sc->sc_spi_count == 0);
> MPASS(value >= GIC_FIRST_SPI);
> MPASS(value < sc->nirqs);
>
> @@ -537,7 +535,6 @@ arm_gic_write_ivar(device_t dev, device_t child,
> int which, uintptr_t value)
> return (0);
> case GIC_IVAR_MBI_COUNT:
> MPASS(sc->sc_spi_start != 0);
> - MPASS(sc->sc_spi_count == 0);
>
> sc->sc_spi_count = value;
> sc->sc_spi_end = sc->sc_spi_start + sc->sc_spi_count;
>
> Any thoughts?
Can you try the patch in https://reviews.freebsd.org/D32224. The above change is to check there is only a single mbi range, however it appears your hardware has multiple gicv2m devices so will try to reserve multiple mbi ranges.
Andrew
More information about the dev-commits-src-all
mailing list