Thank you for your contribution! Sashiko AI review found 3 potential issues to consider: - [Medium] accel/amd-ai-engine: unflushed pm_ops on error in aiemlv2_part_init() - [Medium] accel/amd-ai-engine: unchecked 2-column NMU range in aiemlv2_part_init() - [Low] accel/amd-ai-engine: open-coded polling in aie_part_maskpoll_register() --
commit 7ceb805ff0b7740b5882eeff9e935665efc44bc8 Author: Gregory Williams <[email protected]> accel/amd-ai-engine: Add reset operations This commit adds partition initialization and teardown callbacks for AI Engine partitions, configuring resets, memory zeroization, NMU routing, ECC scrubbing, error reporting, and tile clocks. > diff --git a/drivers/accel/amd-ai-engine/ai-engine-aiemlv2.c > b/drivers/accel/amd-ai-engine/ai-engine-aiemlv2.c > index 26dd1d726bfad..910a9bda23055 100644 > --- a/drivers/accel/amd-ai-engine/ai-engine-aiemlv2.c > +++ b/drivers/accel/amd-ai-engine/ai-engine-aiemlv2.c [ ... ] > @@ -173,10 +177,250 @@ static int aiemlv2_set_part_clocks(struct > aie_partition *apart) [ ... ] > +static int aiemlv2_part_init(struct aie_partition *apart, > + struct aie_partition_init_args *args) > +{ > + u32 opts; > + int ret; [ ... ] > + if (args->init_opts & AIE_PART_INIT_OPT_NMU_CONFIG) { > + struct aie_range range = {0}; > + > + opts |= AIE_PART_INIT_OPT_NMU_CONFIG; > + if (apart->range.start.col == 0) { > + range.size.col = 2; > + ret = aie_part_pm_ops(apart, NULL, > AIE_PART_INIT_OPT_NMU_CONFIG, > + range, 0); > + if (ret) > + goto out; > + } > + } [Severity: Medium] Can this configure column 1 outside the partition's allocated columns? If a 1-column partition is requested at column 0 (apart->range.start.col == 0, apart->range.size.col == 1), setting range.size.col = 2 without checking apart->range.size.col >= 2 programs NMU routing on column 1 as well, which may belong to another partition. > + > + if (args->init_opts & AIE_PART_INIT_OPT_HW_ERR_INT) { > + u16 data = 0; > + > + opts |= AIE_PART_INIT_OPT_HW_ERR_INT; > + ret = aie_part_pm_ops(apart, &data, > AIE_PART_INIT_OPT_HW_ERR_INT, > + apart->range, 0); > + if (ret) > + goto out; > + } > + > + if (args->init_opts & AIE_PART_INIT_OPT_HW_ERR_MASK) { > + u16 data = 0x2; > + > + opts |= AIE_PART_INIT_OPT_HW_ERR_MASK; > + ret = aie_part_pm_ops(apart, &data, > AIE_PART_INIT_OPT_HW_ERR_MASK, > + apart->range, 0); > + if (ret) > + goto out; > + } > + > + if (args->init_opts & AIE_PART_INIT_OPT_SET_ECC_SCRUB_PERIOD) { > + u16 data; > + > + if (args->ecc_scrub > U16_MAX) { > + ret = -EINVAL; > + goto out; > + } > + > + data = args->ecc_scrub; > + opts |= AIE_PART_INIT_OPT_SET_ECC_SCRUB_PERIOD; > + ret = aie_part_pm_ops(apart, &data, > + AIE_PART_INIT_OPT_SET_ECC_SCRUB_PERIOD, > + apart->range, 0); > + if (ret) > + goto out; > + } > + > + if (args->init_opts & AIE_PART_INIT_OPT_UC_ENB_MEM_PRIV) { > + opts |= AIE_PART_INIT_OPT_UC_ENB_MEM_PRIV; > + > + ret = aie_part_pm_ops(apart, NULL, > AIE_PART_INIT_OPT_UC_ENB_MEM_PRIV, > + apart->range, 0); > + if (ret) > + goto out; > + } > + > + /* Request tile locations */ > + ret = aie_part_request_tiles(apart, args->num_tiles, args->locs); > + if (ret) > + goto out; > + > + if (args->init_opts & AIE_PART_INIT_OPT_DIS_TLAST_ERROR) { > + u16 data = 0; > + > + opts |= AIE_PART_INIT_OPT_DIS_TLAST_ERROR; > + ret = aie_part_pm_ops(apart, &data, > + AIE_PART_INIT_OPT_DIS_TLAST_ERROR, > + apart->range, 1); > + if (ret) > + goto out; > + } > + > + if (opts != args->init_opts) > + dev_warn(&apart->dev, "Invalid init_opts: 0x%x\n", opts ^ > args->init_opts); > + > +out: > + mutex_unlock(&apart->mlock); > + return ret; > +} [Severity: Medium] Does the error path leave unflushed commands queued in apart->pm_ops? The operations from AIE_PART_INIT_OPT_NMU_CONFIG through AIE_PART_INIT_OPT_UC_ENB_MEM_PRIV are queued with flush = 0 and rely on aie_part_request_tiles() -> aiemlv2_set_part_clocks() -> aie_part_pm_ops_flush() to flush and reset the buffer. If args->ecc_scrub > U16_MAX or aie_part_request_tiles() fails (for example, when args->num_tiles > 0 and args->locs is NULL), jumping to out leaves those descriptors in apart->pm_ops.pkt_va, so they will be sent to firmware ahead of the DMA pause operations on the next initialize or teardown call. Validating args before modifying hardware state and resetting apart->pm_ops on failure would avoid both the stale command buffer and leaving the partition partially reset on -EINVAL. > diff --git a/drivers/accel/amd-ai-engine/ai-engine-part.c > b/drivers/accel/amd-ai-engine/ai-engine-part.c > index 35451d1ed88a4..2cd93c02246e5 100644 > --- a/drivers/accel/amd-ai-engine/ai-engine-part.c > +++ b/drivers/accel/amd-ai-engine/ai-engine-part.c [ ... ] > @@ -106,3 +107,34 @@ struct aie_partition *aie_part_create(struct aie_device > *adev, > device_unregister(&apart->dev); > return ERR_PTR(ret); > } > + > +/** > + * aie_part_maskpoll_register() - Mask polls an address for a data in a > partition. > + * @apart: AI engine partition. > + * @offset: AI engine register offset. > + * @data: data to poll. > + * @mask: if it is non 0, it is mask read. > + * @timeout: Timeout in microseconds. > + * > + * Return: zero for success, or -EBUSY for failure > + */ > +int aie_part_maskpoll_register(struct aie_partition *apart, u32 offset, u32 > data, > + u32 mask, u32 timeout) > +{ > + u32 usleep_time = 10; > + u32 value; > + u32 i; > + > + for (i = 0; i < timeout; i += usleep_time) { > + value = readl(apart->adev->base + offset); > + if ((value & mask) == data) > + return 0; > + usleep_range(usleep_time, usleep_time + 10); > + } > + > + value = readl(apart->adev->base + offset); > + if ((value & mask) == data) > + return 0; > + > + return -EBUSY; > +} [Severity: Low] Can this use readl_poll_timeout() from linux/iopoll.h instead of an open-coded loop? Advancing i by 10 us per iteration while calling usleep_range(10, 20) does not account for sleeps up to 20 us or scheduler wakeup overhead, so a 2000000 us timeout can block for much longer than 2 seconds per column while holding apart->mlock. -- Sashiko AI review ยท https://sashiko.dev/#/patchset/[email protected]?part=15
