diff options
Diffstat (limited to 'drivers/misc/habanalabs/common')
| -rw-r--r-- | drivers/misc/habanalabs/common/command_submission.c | 77 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/device.c | 19 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/firmware_if.c | 65 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/habanalabs.h | 5 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/habanalabs_drv.c | 1 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/habanalabs_ioctl.c | 9 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/hw_queue.c | 5 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/memory.c | 10 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/mmu.c | 6 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/mmu_v1.c | 12 | ||||
| -rw-r--r-- | drivers/misc/habanalabs/common/pci.c | 28 |
11 files changed, 177 insertions, 60 deletions
diff --git a/drivers/misc/habanalabs/common/command_submission.c b/drivers/misc/habanalabs/common/command_submission.c index beb482310a58..b2b3d2b0f808 100644 --- a/drivers/misc/habanalabs/common/command_submission.c +++ b/drivers/misc/habanalabs/common/command_submission.c @@ -472,8 +472,11 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx, cntr = &hdev->aggregated_cs_counters; cs = kzalloc(sizeof(*cs), GFP_ATOMIC); - if (!cs) + if (!cs) { + atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt); + atomic64_inc(&cntr->out_of_mem_drop_cnt); return -ENOMEM; + } cs->ctx = ctx; cs->submitted = false; @@ -486,6 +489,8 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx, cs_cmpl = kmalloc(sizeof(*cs_cmpl), GFP_ATOMIC); if (!cs_cmpl) { + atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt); + atomic64_inc(&cntr->out_of_mem_drop_cnt); rc = -ENOMEM; goto free_cs; } @@ -513,6 +518,8 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx, cs->jobs_in_queue_cnt = kcalloc(hdev->asic_prop.max_queues, sizeof(*cs->jobs_in_queue_cnt), GFP_ATOMIC); if (!cs->jobs_in_queue_cnt) { + atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt); + atomic64_inc(&cntr->out_of_mem_drop_cnt); rc = -ENOMEM; goto free_fence; } @@ -562,7 +569,7 @@ void hl_cs_rollback_all(struct hl_device *hdev) for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++) flush_workqueue(hdev->cq_wq[i]); - /* Make sure we don't have leftovers in the H/W queues mirror list */ + /* Make sure we don't have leftovers in the CS mirror list */ list_for_each_entry_safe(cs, tmp, &hdev->cs_mirror_list, mirror_node) { cs_get(cs); cs->aborted = true; @@ -764,11 +771,14 @@ static int hl_cs_sanity_checks(struct hl_fpriv *hpriv, union hl_cs_args *args) static int hl_cs_copy_chunk_array(struct hl_device *hdev, struct hl_cs_chunk **cs_chunk_array, - void __user *chunks, u32 num_chunks) + void __user *chunks, u32 num_chunks, + struct hl_ctx *ctx) { u32 size_to_copy; if (num_chunks > HL_MAX_JOBS_PER_CS) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt); dev_err(hdev->dev, "Number of chunks can NOT be larger than %d\n", HL_MAX_JOBS_PER_CS); @@ -777,11 +787,16 @@ static int hl_cs_copy_chunk_array(struct hl_device *hdev, *cs_chunk_array = kmalloc_array(num_chunks, sizeof(**cs_chunk_array), GFP_ATOMIC); - if (!*cs_chunk_array) + if (!*cs_chunk_array) { + atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt); + atomic64_inc(&hdev->aggregated_cs_counters.out_of_mem_drop_cnt); return -ENOMEM; + } size_to_copy = num_chunks * sizeof(struct hl_cs_chunk); if (copy_from_user(*cs_chunk_array, chunks, size_to_copy)) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt); dev_err(hdev->dev, "Failed to copy cs chunk array from user\n"); kfree(*cs_chunk_array); return -EFAULT; @@ -797,6 +812,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, struct hl_device *hdev = hpriv->hdev; struct hl_cs_chunk *cs_chunk_array; struct hl_cs_counters_atomic *cntr; + struct hl_ctx *ctx = hpriv->ctx; struct hl_cs_job *job; struct hl_cs *cs; struct hl_cb *cb; @@ -805,7 +821,8 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, cntr = &hdev->aggregated_cs_counters; *cs_seq = ULLONG_MAX; - rc = hl_cs_copy_chunk_array(hdev, &cs_chunk_array, chunks, num_chunks); + rc = hl_cs_copy_chunk_array(hdev, &cs_chunk_array, chunks, num_chunks, + hpriv->ctx); if (rc) goto out; @@ -832,8 +849,8 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, rc = validate_queue_index(hdev, chunk, &queue_type, &is_kernel_allocated_cb); if (rc) { - atomic64_inc(&hpriv->ctx->cs_counters.parsing_drop_cnt); - atomic64_inc(&cntr->parsing_drop_cnt); + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); goto free_cs_object; } @@ -841,8 +858,8 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, cb = get_cb_from_cs_chunk(hdev, &hpriv->cb_mgr, chunk); if (!cb) { atomic64_inc( - &hpriv->ctx->cs_counters.parsing_drop_cnt); - atomic64_inc(&cntr->parsing_drop_cnt); + &ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); rc = -EINVAL; goto free_cs_object; } @@ -856,8 +873,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, job = hl_cs_allocate_job(hdev, queue_type, is_kernel_allocated_cb); if (!job) { - atomic64_inc( - &hpriv->ctx->cs_counters.out_of_mem_drop_cnt); + atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt); atomic64_inc(&cntr->out_of_mem_drop_cnt); dev_err(hdev->dev, "Failed to allocate a new job\n"); rc = -ENOMEM; @@ -891,7 +907,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, rc = cs_parser(hpriv, job); if (rc) { - atomic64_inc(&hpriv->ctx->cs_counters.parsing_drop_cnt); + atomic64_inc(&ctx->cs_counters.parsing_drop_cnt); atomic64_inc(&cntr->parsing_drop_cnt); dev_err(hdev->dev, "Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n", @@ -901,8 +917,8 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks, } if (int_queues_only) { - atomic64_inc(&hpriv->ctx->cs_counters.parsing_drop_cnt); - atomic64_inc(&cntr->parsing_drop_cnt); + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); dev_err(hdev->dev, "Reject CS %d.%llu because only internal queues jobs are present\n", cs->ctx->asid, cs->sequence); @@ -1042,7 +1058,7 @@ out: } static int cs_ioctl_extract_signal_seq(struct hl_device *hdev, - struct hl_cs_chunk *chunk, u64 *signal_seq) + struct hl_cs_chunk *chunk, u64 *signal_seq, struct hl_ctx *ctx) { u64 *signal_seq_arr = NULL; u32 size_to_copy, signal_seq_arr_len; @@ -1052,6 +1068,8 @@ static int cs_ioctl_extract_signal_seq(struct hl_device *hdev, /* currently only one signal seq is supported */ if (signal_seq_arr_len != 1) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt); dev_err(hdev->dev, "Wait for signal CS supports only one signal CS seq\n"); return -EINVAL; @@ -1060,13 +1078,18 @@ static int cs_ioctl_extract_signal_seq(struct hl_device *hdev, signal_seq_arr = kmalloc_array(signal_seq_arr_len, sizeof(*signal_seq_arr), GFP_ATOMIC); - if (!signal_seq_arr) + if (!signal_seq_arr) { + atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt); + atomic64_inc(&hdev->aggregated_cs_counters.out_of_mem_drop_cnt); return -ENOMEM; + } size_to_copy = chunk->num_signal_seq_arr * sizeof(*signal_seq_arr); if (copy_from_user(signal_seq_arr, u64_to_user_ptr(chunk->signal_seq_arr), size_to_copy)) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt); dev_err(hdev->dev, "Failed to copy signal seq array from user\n"); rc = -EFAULT; @@ -1153,6 +1176,7 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, struct hl_device *hdev = hpriv->hdev; struct hl_cs_compl *sig_waitcs_cmpl; u32 q_idx, collective_engine_id = 0; + struct hl_cs_counters_atomic *cntr; struct hl_fence *sig_fence = NULL; struct hl_ctx *ctx = hpriv->ctx; enum hl_queue_type q_type; @@ -1160,9 +1184,11 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, u64 signal_seq; int rc; + cntr = &hdev->aggregated_cs_counters; *cs_seq = ULLONG_MAX; - rc = hl_cs_copy_chunk_array(hdev, &cs_chunk_array, chunks, num_chunks); + rc = hl_cs_copy_chunk_array(hdev, &cs_chunk_array, chunks, num_chunks, + ctx); if (rc) goto out; @@ -1170,6 +1196,8 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, chunk = &cs_chunk_array[0]; if (chunk->queue_index >= hdev->asic_prop.max_queues) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); dev_err(hdev->dev, "Queue index %d is invalid\n", chunk->queue_index); rc = -EINVAL; @@ -1181,6 +1209,8 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, q_type = hw_queue_prop->type; if (!hw_queue_prop->supports_sync_stream) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); dev_err(hdev->dev, "Queue index %d does not support sync stream operations\n", q_idx); @@ -1190,6 +1220,8 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, if (cs_type == CS_TYPE_COLLECTIVE_WAIT) { if (!(hw_queue_prop->collective_mode == HL_COLLECTIVE_MASTER)) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); dev_err(hdev->dev, "Queue index %d is invalid\n", q_idx); rc = -EINVAL; @@ -1200,12 +1232,14 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, } if (cs_type == CS_TYPE_WAIT || cs_type == CS_TYPE_COLLECTIVE_WAIT) { - rc = cs_ioctl_extract_signal_seq(hdev, chunk, &signal_seq); + rc = cs_ioctl_extract_signal_seq(hdev, chunk, &signal_seq, ctx); if (rc) goto free_cs_chunk_array; sig_fence = hl_ctx_get_fence(ctx, signal_seq); if (IS_ERR(sig_fence)) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); dev_err(hdev->dev, "Failed to get signal CS with seq 0x%llx\n", signal_seq); @@ -1223,6 +1257,8 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, container_of(sig_fence, struct hl_cs_compl, base_fence); if (sig_waitcs_cmpl->type != CS_TYPE_SIGNAL) { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); dev_err(hdev->dev, "CS seq 0x%llx is not of a signal CS\n", signal_seq); @@ -1270,8 +1306,11 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type, else if (cs_type == CS_TYPE_COLLECTIVE_WAIT) rc = hdev->asic_funcs->collective_wait_create_jobs(hdev, ctx, cs, q_idx, collective_engine_id); - else + else { + atomic64_inc(&ctx->cs_counters.validation_drop_cnt); + atomic64_inc(&cntr->validation_drop_cnt); rc = -EINVAL; + } if (rc) goto free_cs_object; diff --git a/drivers/misc/habanalabs/common/device.c b/drivers/misc/habanalabs/common/device.c index 5871162a8442..69d04eca767f 100644 --- a/drivers/misc/habanalabs/common/device.c +++ b/drivers/misc/habanalabs/common/device.c @@ -17,12 +17,12 @@ enum hl_device_status hl_device_status(struct hl_device *hdev) { enum hl_device_status status; - if (hdev->disabled) - status = HL_DEVICE_STATUS_MALFUNCTION; - else if (atomic_read(&hdev->in_reset)) + if (atomic_read(&hdev->in_reset)) status = HL_DEVICE_STATUS_IN_RESET; else if (hdev->needs_reset) status = HL_DEVICE_STATUS_NEEDS_RESET; + else if (hdev->disabled) + status = HL_DEVICE_STATUS_MALFUNCTION; else status = HL_DEVICE_STATUS_OPERATIONAL; @@ -1037,7 +1037,7 @@ kill_processes: if (hard_reset) { /* Release kernel context */ - if (hl_ctx_put(hdev->kernel_ctx) == 1) + if (hdev->kernel_ctx && hl_ctx_put(hdev->kernel_ctx) == 1) hdev->kernel_ctx = NULL; hl_vm_fini(hdev); hl_mmu_fini(hdev); @@ -1092,6 +1092,7 @@ kill_processes: GFP_KERNEL); if (!hdev->kernel_ctx) { rc = -ENOMEM; + hl_mmu_fini(hdev); goto out_err; } @@ -1103,6 +1104,7 @@ kill_processes: "failed to init kernel ctx in hard reset\n"); kfree(hdev->kernel_ctx); hdev->kernel_ctx = NULL; + hl_mmu_fini(hdev); goto out_err; } } @@ -1485,6 +1487,15 @@ void hl_device_fini(struct hl_device *hdev) } } + /* Disable PCI access from device F/W so it won't send us additional + * interrupts. We disable MSI/MSI-X at the halt_engines function and we + * can't have the F/W sending us interrupts after that. We need to + * disable the access here because if the device is marked disable, the + * message won't be send. Also, in case of heartbeat, the device CPU is + * marked as disable so this message won't be sent + */ + hl_fw_send_pci_access_msg(hdev, CPUCP_PACKET_DISABLE_PCI_ACCESS); + /* Mark device as disabled */ hdev->disabled = true; diff --git a/drivers/misc/habanalabs/common/firmware_if.c b/drivers/misc/habanalabs/common/firmware_if.c index 0e1c629e9800..c9a12980218a 100644 --- a/drivers/misc/habanalabs/common/firmware_if.c +++ b/drivers/misc/habanalabs/common/firmware_if.c @@ -402,6 +402,10 @@ int hl_fw_cpucp_pci_counters_get(struct hl_device *hdev, } counters->rx_throughput = result; + memset(&pkt, 0, sizeof(pkt)); + pkt.ctl = cpu_to_le32(CPUCP_PACKET_PCIE_THROUGHPUT_GET << + CPUCP_PKT_CTL_OPCODE_SHIFT); + /* Fetch PCI tx counter */ pkt.index = cpu_to_le32(cpucp_pcie_throughput_tx); rc = hdev->asic_funcs->send_cpu_message(hdev, (u32 *) &pkt, sizeof(pkt), @@ -414,6 +418,7 @@ int hl_fw_cpucp_pci_counters_get(struct hl_device *hdev, counters->tx_throughput = result; /* Fetch PCI replay counter */ + memset(&pkt, 0, sizeof(pkt)); pkt.ctl = cpu_to_le32(CPUCP_PACKET_PCIE_REPLAY_CNT_GET << CPUCP_PKT_CTL_OPCODE_SHIFT); @@ -627,25 +632,38 @@ int hl_fw_read_preboot_status(struct hl_device *hdev, u32 cpu_boot_status_reg, security_status = RREG32(cpu_security_boot_status_reg); /* We read security status multiple times during boot: - * 1. preboot - we check if fw security feature is supported - * 2. boot cpu - we get boot cpu security status - * 3. FW application - we get FW application security status + * 1. preboot - a. Check whether the security status bits are valid + * b. Check whether fw security is enabled + * c. Check whether hard reset is done by preboot + * 2. boot cpu - a. Fetch boot cpu security status + * b. Check whether hard reset is done by boot cpu + * 3. FW application - a. Fetch fw application security status + * b. Check whether hard reset is done by fw app * * Preboot: * Check security status bit (CPU_BOOT_DEV_STS0_ENABLED), if it is set * check security enabled bit (CPU_BOOT_DEV_STS0_SECURITY_EN) */ if (security_status & CPU_BOOT_DEV_STS0_ENABLED) { - hdev->asic_prop.fw_security_status_valid = 1; - prop->fw_security_disabled = - !(security_status & CPU_BOOT_DEV_STS0_SECURITY_EN); + prop->fw_security_status_valid = 1; + + if (security_status & CPU_BOOT_DEV_STS0_SECURITY_EN) + prop->fw_security_disabled = false; + else + prop->fw_security_disabled = true; + + if (security_status & CPU_BOOT_DEV_STS0_FW_HARD_RST_EN) + prop->hard_reset_done_by_fw = true; } else { - hdev->asic_prop.fw_security_status_valid = 0; + prop->fw_security_status_valid = 0; prop->fw_security_disabled = true; } + dev_dbg(hdev->dev, "Firmware preboot hard-reset is %s\n", + prop->hard_reset_done_by_fw ? "enabled" : "disabled"); + dev_info(hdev->dev, "firmware-level security is %s\n", - prop->fw_security_disabled ? "disabled" : "enabled"); + prop->fw_security_disabled ? "disabled" : "enabled"); return 0; } @@ -655,6 +673,7 @@ int hl_fw_init_cpu(struct hl_device *hdev, u32 cpu_boot_status_reg, u32 cpu_security_boot_status_reg, u32 boot_err0_reg, bool skip_bmc, u32 cpu_timeout, u32 boot_fit_timeout) { + struct asic_fixed_properties *prop = &hdev->asic_prop; u32 status; int rc; @@ -723,11 +742,22 @@ int hl_fw_init_cpu(struct hl_device *hdev, u32 cpu_boot_status_reg, /* Read U-Boot version now in case we will later fail */ hdev->asic_funcs->read_device_fw_version(hdev, FW_COMP_UBOOT); + /* Clear reset status since we need to read it again from boot CPU */ + prop->hard_reset_done_by_fw = false; + /* Read boot_cpu security bits */ - if (hdev->asic_prop.fw_security_status_valid) - hdev->asic_prop.fw_boot_cpu_security_map = + if (prop->fw_security_status_valid) { + prop->fw_boot_cpu_security_map = RREG32(cpu_security_boot_status_reg); + if (prop->fw_boot_cpu_security_map & + CPU_BOOT_DEV_STS0_FW_HARD_RST_EN) + prop->hard_reset_done_by_fw = true; + } + + dev_dbg(hdev->dev, "Firmware boot CPU hard-reset is %s\n", + prop->hard_reset_done_by_fw ? "enabled" : "disabled"); + if (rc) { detect_cpu_boot_status(hdev, status); rc = -EIO; @@ -796,18 +826,21 @@ int hl_fw_init_cpu(struct hl_device *hdev, u32 cpu_boot_status_reg, goto out; } + /* Clear reset status since we need to read again from app */ + prop->hard_reset_done_by_fw = false; + /* Read FW application security bits */ - if (hdev->asic_prop.fw_security_status_valid) { - hdev->asic_prop.fw_app_security_map = + if (prop->fw_security_status_valid) { + prop->fw_app_security_map = RREG32(cpu_security_boot_status_reg); - if (hdev->asic_prop.fw_app_security_map & + if (prop->fw_app_security_map & CPU_BOOT_DEV_STS0_FW_HARD_RST_EN) - hdev->asic_prop.hard_reset_done_by_fw = true; + prop->hard_reset_done_by_fw = true; } - dev_dbg(hdev->dev, "Firmware hard-reset is %s\n", - hdev->asic_prop.hard_reset_done_by_fw ? "enabled" : "disabled"); + dev_dbg(hdev->dev, "Firmware application CPU hard-reset is %s\n", + prop->hard_reset_done_by_fw ? "enabled" : "disabled"); dev_info(hdev->dev, "Successfully loaded firmware to device\n"); diff --git a/drivers/misc/habanalabs/common/habanalabs.h b/drivers/misc/habanalabs/common/habanalabs.h index 571eda6ef5ab..60e16dc4bcac 100644 --- a/drivers/misc/habanalabs/common/habanalabs.h +++ b/drivers/misc/habanalabs/common/habanalabs.h @@ -944,7 +944,7 @@ struct hl_asic_funcs { u32 (*get_signal_cb_size)(struct hl_device *hdev); u32 (*get_wait_cb_size)(struct hl_device *hdev); u32 (*gen_signal_cb)(struct hl_device *hdev, void *data, u16 sob_id, - u32 size); + u32 size, bool eb); u32 (*gen_wait_cb)(struct hl_device *hdev, struct hl_gen_wait_properties *prop); void (*reset_sob)(struct hl_device *hdev, void *data); @@ -1000,6 +1000,7 @@ struct hl_va_range { * @queue_full_drop_cnt: dropped due to queue full * @device_in_reset_drop_cnt: dropped due to device in reset * @max_cs_in_flight_drop_cnt: dropped due to maximum CS in-flight + * @validation_drop_cnt: dropped due to error in validation */ struct hl_cs_counters_atomic { atomic64_t out_of_mem_drop_cnt; @@ -1007,6 +1008,7 @@ struct hl_cs_counters_atomic { atomic64_t queue_full_drop_cnt; atomic64_t device_in_reset_drop_cnt; atomic64_t max_cs_in_flight_drop_cnt; + atomic64_t validation_drop_cnt; }; /** @@ -2180,6 +2182,7 @@ void hl_mmu_v1_set_funcs(struct hl_device *hdev, struct hl_mmu_funcs *mmu); int hl_mmu_va_to_pa(struct hl_ctx *ctx, u64 virt_addr, u64 *phys_addr); int hl_mmu_get_tlb_info(struct hl_ctx *ctx, u64 virt_addr, struct hl_mmu_hop_info *hops); +bool hl_is_dram_va(struct hl_device *hdev, u64 virt_addr); int hl_fw_load_fw_to_device(struct hl_device *hdev, const char *fw_name, void __iomem *dst, u32 src_offset, u32 size); diff --git a/drivers/misc/habanalabs/common/habanalabs_drv.c b/drivers/misc/habanalabs/common/habanalabs_drv.c index 6bbb6bca6860..032d114f01ea 100644 --- a/drivers/misc/habanalabs/common/habanalabs_drv.c +++ b/drivers/misc/habanalabs/common/habanalabs_drv.c @@ -544,6 +544,7 @@ static struct pci_driver hl_pci_driver = { .id_table = ids, .probe = hl_pci_probe, .remove = hl_pci_remove, + .shutdown = hl_pci_remove, .driver.pm = &hl_pm_ops, .err_handler = &hl_pci_err_handler, }; diff --git a/drivers/misc/habanalabs/common/habanalabs_ioctl.c b/drivers/misc/habanalabs/common/habanalabs_ioctl.c index 32e6af1db4e3..d25892d61ec9 100644 --- a/drivers/misc/habanalabs/common/habanalabs_ioctl.c +++ b/drivers/misc/habanalabs/common/habanalabs_ioctl.c @@ -133,6 +133,8 @@ static int hw_idle(struct hl_device *hdev, struct hl_info_args *args) hw_idle.is_idle = hdev->asic_funcs->is_device_idle(hdev, &hw_idle.busy_engines_mask_ext, NULL); + hw_idle.busy_engines_mask = + lower_32_bits(hw_idle.busy_engines_mask_ext); return copy_to_user(out, &hw_idle, min((size_t) max_size, sizeof(hw_idle))) ? -EFAULT : 0; @@ -335,6 +337,8 @@ static int cs_counters_info(struct hl_fpriv *hpriv, struct hl_info_args *args) atomic64_read(&cntr->device_in_reset_drop_cnt); cs_counters.total_max_cs_in_flight_drop_cnt = atomic64_read(&cntr->max_cs_in_flight_drop_cnt); + cs_counters.total_validation_drop_cnt = + atomic64_read(&cntr->validation_drop_cnt); if (hpriv->ctx) { cs_counters.ctx_out_of_mem_drop_cnt = @@ -352,6 +356,9 @@ static int cs_counters_info(struct hl_fpriv *hpriv, struct hl_info_args *args) cs_counters.ctx_max_cs_in_flight_drop_cnt = atomic64_read( &hpriv->ctx->cs_counters.max_cs_in_flight_drop_cnt); + cs_counters.ctx_validation_drop_cnt = + atomic64_read( + &hpriv->ctx->cs_counters.validation_drop_cnt); } return copy_to_user(out, &cs_counters, @@ -406,7 +413,7 @@ static int total_energy_consumption_info(struct hl_fpriv *hpriv, static int pll_frequency_info(struct hl_fpriv *hpriv, struct hl_info_args *args) { struct hl_device *hdev = hpriv->hdev; - struct hl_pll_frequency_info freq_info = {0}; + struct hl_pll_frequency_info freq_info = { {0} }; u32 max_size = args->return_size; void __user *out = (void __user *) (uintptr_t) args->return_pointer; int rc; diff --git a/drivers/misc/habanalabs/common/hw_queue.c b/drivers/misc/habanalabs/common/hw_queue.c index 7caf868d1585..76217258780a 100644 --- a/drivers/misc/habanalabs/common/hw_queue.c +++ b/drivers/misc/habanalabs/common/hw_queue.c @@ -418,8 +418,11 @@ static void init_signal_cs(struct hl_device *hdev, "generate signal CB, sob_id: %d, sob val: 0x%x, q_idx: %d\n", cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val, q_idx); + /* we set an EB since we must make sure all oeprations are done + * when sending the signal + */ hdev->asic_funcs->gen_signal_cb(hdev, job->patched_cb, - cs_cmpl->hw_sob->sob_id, 0); + cs_cmpl->hw_sob->sob_id, 0, true); kref_get(&hw_sob->kref); diff --git a/drivers/misc/habanalabs/common/memory.c b/drivers/misc/habanalabs/common/memory.c index cbe9da4e0211..5d4fbdcaefe3 100644 --- a/drivers/misc/habanalabs/common/memory.c +++ b/drivers/misc/habanalabs/common/memory.c @@ -886,8 +886,10 @@ static void unmap_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr, { struct hl_device *hdev = ctx->hdev; u64 next_vaddr, i; + bool is_host_addr; u32 page_size; + is_host_addr = !hl_is_dram_va(hdev, vaddr); page_size = phys_pg_pack->page_size; next_vaddr = vaddr; @@ -900,9 +902,13 @@ static void unmap_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr, /* * unmapping on Palladium can be really long, so avoid a CPU * soft lockup bug by sleeping a little between unmapping pages + * + * In addition, when unmapping host memory we pass through + * the Linux kernel to unpin the pages and that takes a long + * time. Therefore, sleep every 32K pages to avoid soft lockup */ - if (hdev->pldm) - usleep_range(500, 1000); + if (hdev->pldm || (is_host_addr && (i & 0x7FFF) == 0)) + usleep_range(50, 200); } } diff --git a/drivers/misc/habanalabs/common/mmu.c b/drivers/misc/habanalabs/common/mmu.c index 33ae953d3a36..28a4638741d8 100644 --- a/drivers/misc/habanalabs/common/mmu.c +++ b/drivers/misc/habanalabs/common/mmu.c @@ -9,7 +9,7 @@ #include "habanalabs.h" -static bool is_dram_va(struct hl_device *hdev, u64 virt_addr) +bool hl_is_dram_va(struct hl_device *hdev, u64 virt_addr) { struct asic_fixed_properties *prop = &hdev->asic_prop; @@ -156,7 +156,7 @@ int hl_mmu_unmap_page(struct hl_ctx *ctx, u64 virt_addr, u32 page_size, if (!hdev->mmu_enable) return 0; - is_dram_addr = is_dram_va(hdev, virt_addr); + is_dram_addr = hl_is_dram_va(hdev, virt_addr); if (is_dram_addr) mmu_prop = &prop->dmmu; @@ -236,7 +236,7 @@ int hl_mmu_map_page(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr, if (!hdev->mmu_enable) return 0; - is_dram_addr = is_dram_va(hdev, virt_addr); + is_dram_addr = hl_is_dram_va(hdev, virt_addr); if (is_dram_addr) mmu_prop = &prop->dmmu; diff --git a/drivers/misc/habanalabs/common/mmu_v1.c b/drivers/misc/habanalabs/common/mmu_v1.c index 2ce6ea89d4fa..06d8a44dd5d4 100644 --- a/drivers/misc/habanalabs/common/mmu_v1.c +++ b/drivers/misc/habanalabs/common/mmu_v1.c @@ -467,8 +467,16 @@ static void hl_mmu_v1_fini(struct hl_device *hdev) { /* MMU H/W fini was already done in device hw_fini() */ - kvfree(hdev->mmu_priv.dr.mmu_shadow_hop0); - gen_pool_destroy(hdev->mmu_priv.dr.mmu_pgt_pool); + if (!ZERO_OR_NULL_PTR(hdev->mmu_priv.hr.mmu_shadow_hop0)) { + kvfree(hdev->mmu_priv.dr.mmu_shadow_hop0); + gen_pool_destroy(hdev->mmu_priv.dr.mmu_pgt_pool); + } + + /* Make sure that if we arrive here again without init was called we + * won't cause kernel panic. This can happen for example if we fail + * during hard reset code at certain points + */ + hdev->mmu_priv.dr.mmu_shadow_hop0 = NULL; } /** diff --git a/drivers/misc/habanalabs/common/pci.c b/drivers/misc/habanalabs/common/pci.c index 923b2606e29f..b4725e6101f6 100644 --- a/drivers/misc/habanalabs/common/pci.c +++ b/drivers/misc/habanalabs/common/pci.c @@ -130,10 +130,8 @@ static int hl_pci_elbi_write(struct hl_device *hdev, u64 addr, u32 data) if ((val & PCI_CONFIG_ELBI_STS_MASK) == PCI_CONFIG_ELBI_STS_DONE) return 0; - if (val & PCI_CONFIG_ELBI_STS_ERR) { - dev_err(hdev->dev, "Error writing to ELBI\n"); + if (val & PCI_CONFIG_ELBI_STS_ERR) return -EIO; - } if (!(val & PCI_CONFIG_ELBI_STS_MASK)) { dev_err(hdev->dev, "ELBI write didn't finish in time\n"); @@ -160,8 +158,12 @@ int hl_pci_iatu_write(struct hl_device *hdev, u32 addr, u32 data) dbi_offset = addr & 0xFFF; - rc = hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr, 0x00300000); - rc |= hl_pci_elbi_write(hdev, prop->pcie_dbi_base_address + dbi_offset, + /* Ignore result of writing to pcie_aux_dbi_reg_addr as it could fail + * in case the firmware security is enabled + */ + hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr, 0x00300000); + + rc = hl_pci_elbi_write(hdev, prop->pcie_dbi_base_address + dbi_offset, data); if (rc) @@ -244,9 +246,11 @@ int hl_pci_set_inbound_region(struct hl_device *hdev, u8 region, rc |= hl_pci_iatu_write(hdev, offset + 0x4, ctrl_reg_val); - /* Return the DBI window to the default location */ - rc |= hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr, 0); - rc |= hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr + 4, 0); + /* Return the DBI window to the default location + * Ignore result of writing to pcie_aux_dbi_reg_addr as it could fail + * in case the firmware security is enabled + */ + hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr, 0); if (rc) dev_err(hdev->dev, "failed to map bar %u to 0x%08llx\n", @@ -294,9 +298,11 @@ int hl_pci_set_outbound_region(struct hl_device *hdev, /* Enable */ rc |= hl_pci_iatu_write(hdev, 0x004, 0x80000000); - /* Return the DBI window to the default location */ - rc |= hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr, 0); - rc |= hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr + 4, 0); + /* Return the DBI window to the default location + * Ignore result of writing to pcie_aux_dbi_reg_addr as it could fail + * in case the firmware security is enabled + */ + hl_pci_elbi_write(hdev, prop->pcie_aux_dbi_reg_addr, 0); return rc; } |