On 16.02.2024 11:36 AM, Luca Weiss wrote: > On Mon Jan 8, 2024 at 11:45 PM CET, Dmitry Baryshkov wrote: >> On Mon, 8 Jan 2024 at 16:23, Luca Weiss <luca.weiss@xxxxxxxxxxxxx> wrote: >>> >>> On Mon Jan 8, 2024 at 3:18 PM CET, Konrad Dybcio wrote: >>>> On 8.01.2024 14:49, Luca Weiss wrote: >>>>> When num-channels and qcom,num-ees is not provided in devicetree, the >>>>> driver will try to read these values from the registers during probe but >>>>> this fails if the interconnect is not on and then crashes the system. >>>>> >>>>> So we can provide these properties in devicetree (queried after patching >>>>> BAM driver to enable the necessary interconnect) so we can probe >>>>> cryptobam without reading registers and then also use the QCE as >>>>> expected. >>>> >>>> This really feels a bit backwards.. Enable the resource to query the >>>> hardware for numbers, so that said resource can be enabled, but >>>> slightly later :/ >>> >>> If you think adding interconnect support to driver and dtsi is better, >>> let me know. >> >> I'd say, adding the proper interconnect is a better option. Otherwise >> we just depend on the QCE itself to set up the vote for us. > > Yes, currently we depend on that. > >> >>> >>> Stephan (+CC) mentioned it should be okay like this *shrug* >>> >>> For the record, this is the same way I got the values for sc7280[0] and >>> sm6350[1]. >>> >>> [0] https://lore.kernel.org/linux-arm-msm/20231229-sc7280-cryptobam-fixup-v1-1-bd8f68589b80@xxxxxxxxxxxxx/ >>> [1] https://lore.kernel.org/linux-arm-msm/20240105-sm6350-qce-v1-0-416e5c7319ac@xxxxxxxxxxxxx/ >>> >>> diff --git a/arch/arm64/boot/dts/qcom/sm8350.dtsi b/arch/arm64/boot/dts/qcom/sm8350.dtsi >>> index b46236235b7f..cd4dd9852d9e 100644 >>> --- a/arch/arm64/boot/dts/qcom/sm8350.dtsi >>> +++ b/arch/arm64/boot/dts/qcom/sm8350.dtsi >>> @@ -1756,8 +1756,8 @@ cryptobam: dma-controller@1dc4000 { >>> qcom,controlled-remotely; >>> iommus = <&apps_smmu 0x594 0x0011>, >>> <&apps_smmu 0x596 0x0011>; >>> - /* FIXME: Probing BAM DMA causes some abort and system hang */ >>> - status = "fail"; >>> + interconnects = <&aggre2_noc MASTER_CRYPTO 0 &mc_virt SLAVE_EBI1 0>; >>> + interconnect-names = "memory"; >>> }; >>> >>> crypto: crypto@1dfa000 { >>> diff --git a/drivers/dma/qcom/bam_dma.c b/drivers/dma/qcom/bam_dma.c >>> index 5e7d332731e0..9de28f615639 100644 >>> --- a/drivers/dma/qcom/bam_dma.c >>> +++ b/drivers/dma/qcom/bam_dma.c >>> @@ -40,6 +40,7 @@ >>> #include <linux/circ_buf.h> >>> #include <linux/clk.h> >>> #include <linux/dmaengine.h> >>> +#include <linux/interconnect.h> >>> #include <linux/pm_runtime.h> >>> >>> #include "../dmaengine.h" >>> @@ -394,6 +395,7 @@ struct bam_device { >>> const struct reg_offset_data *layout; >>> >>> struct clk *bamclk; >>> + struct icc_path *mem_path; >>> int irq; >>> >>> /* dma start transaction tasklet */ >>> @@ -1206,6 +1208,7 @@ static int bam_init(struct bam_device *bdev) >>> bdev->num_channels = val & BAM_NUM_PIPES_MASK; >>> } >>> >>> + printk(KERN_ERR "%s:%d DBG num_ees=%u num_channels=%u\n", __func__, __LINE__, bdev->num_ees, bdev->num_channels); >>> /* Reset BAM now if fully controlled locally */ >>> if (!bdev->controlled_remotely && !bdev->powered_remotely) >>> bam_reset(bdev); >>> @@ -1298,6 +1301,14 @@ static int bam_dma_probe(struct platform_device *pdev) >>> return ret; >>> } >>> >>> + bdev->mem_path = devm_of_icc_get(bdev->dev, "memory"); >>> + if (IS_ERR(bdev->mem_path)) >>> + return PTR_ERR(bdev->mem_path); >>> + >>> + ret = icc_set_bw(bdev->mem_path, 1, 1); >> >> Probably this needs some more sensible value. > > So downstream qcedev driver uses 384 for the interconnect. But this is > crypto-specific and probably different BAMs have different minimum > requirements? > > #define CRYPTO_AVG_BW 384 > #define CRYPTO_PEAK_BW 384 > https://github.com/xiaomi-sm8450-kernel/android_kernel_platform_msm-kernel/blob/lineage-20/drivers/crypto/msm/qce.h#L57 > > Do you have any suggestion what to use here? I'dve expected this to mean anything, but apparently not. My immediate guess is that the 384 was the lowest magic value that didn't result in the bus getting kicked offline.. 1 should be fine upstream due to commit 91e045b93db7 ("interconnect: qcom: Fix small BW votes being truncated to zero"). > > Also I'd assume that with pm_runtime suspended we'd need to clear the > votes in the driver so we don't keep the interconnect alive > unnecessarily? My naive understanding is that the power should only be necessary when the thing is in use, so early probe and pm-active sounds about sane.. Konrad
