habanalabs/gaudi2: add PCI revision 2 support
PATCH 1/20 habanalabs/gaudi2: add PCI revision 2 support
From: Ofir Bitton <[email protected]>
Add support for Gaudi2 Device with PCI revision 2. Functionality is exactly the same as revision 1, the only difference is device name exposed to user.
--- drivers/misc/habanalabs/common/device.c | 4 +++ drivers/misc/habanalabs/common/habanalabs.h | 2 ++ .../misc/habanalabs/common/habanalabs_drv.c | 26 +++++++++++++------ .../misc/habanalabs/common/habanalabs_ioctl.c | 6 +++-- drivers/misc/habanalabs/common/mmu/mmu.c | 1 + drivers/misc/habanalabs/common/sysfs.c | 2 ++ drivers/misc/habanalabs/gaudi2/gaudi2.c | 6 +---- drivers/misc/habanalabs/gaudi2/gaudi2P.h | 2 -- .../include/hw_ip/pci/pci_general.h | 7 +++++ include/uapi/misc/habanalabs.h | 7 +++++ 10 files changed, 46 insertions(+), 17 deletions(-)
drivers/misc/habanalabs/common/device.c +4 -0
@@ -748,6 +748,10 @@ static int device_early_init(struct hl_device *hdev)gaudi2_set_asic_funcs (hdev);strscpy (hdev->asic_name,"GAUDI2" ,sizeof (hdev->asic_name));break ; +case ASIC_GAUDI2B : +gaudi2_set_asic_funcs (hdev); +strscpy (hdev->asic_name,"GAUDI2B" ,sizeof (hdev->asic_name)); +break ;break ;default : dev_err(hdev->dev, "Unrecognized ASIC type %d\n",
drivers/misc/habanalabs/common/habanalabs.h +2 -0
@@ -1192,6 +1192,7 @@ struct hl_dec {* @ASIC_GAUDI: Gaudi device (HL-2000). * @ASIC_GAUDI_SEC: Gaudi secured device (HL-2000). * @ASIC_GAUDI2: Gaudi2 device. +* @ASIC_GAUDI2B: Gaudi2B device. */ enum hl_asic_type {ASIC_INVALID ,
@@ -1199,6 +1200,7 @@ enum hl_asic_type {ASIC_GAUDI ,ASIC_GAUDI_SEC ,ASIC_GAUDI2 , + ASIC_GAUDI2B, };struct hl_cs_parser ;
drivers/misc/habanalabs/common/habanalabs_drv.c +18 -8
@@ -9,6 +9,7 @@#define pr_fmt (fmt) "habanalabs: " fmt#include "habanalabs.h" +#include "../include/hw_ip/pci/pci_general.h" #include <linux/pci.h> #include <linux/aer.h>
@@ -74,16 +75,17 @@ MODULE_DEVICE_TABLE(pci, ids);/* * get_asic_type - translate device id to asic type * -* @device: id of the PCI device +* @hdev: pointer to habanalabs device structure. * -* Translate device id to asic type. +* Translate device id and revision id to asic type. * In case of unidentified device, return -1 */ -static enum hl_asic_type get_asic_type (u16 device) +static enum hl_asic_type get_asic_type (struct hl_device *hdev) { -enum hl_asic_type asic_type; +struct pci_dev *pdev = hdev->pdev; +enum hl_asic_type asic_type =ASIC_INVALID ; -switch (device) { +switch (pdev->device) {case PCI_IDS_GOYA : asic_type =ASIC_GOYA ;break ;
@@ -94,10 +96,18 @@ static enum hl_asic_type get_asic_type(u16 device) asic_type =ASIC_GAUDI_SEC ;break ;case PCI_IDS_GAUDI2 : - asic_type =ASIC_GAUDI2 ; +switch (pdev->revision) { +case REV_ID_A : + asic_type =ASIC_GAUDI2 ; +break ; +case REV_ID_B : + asic_type =ASIC_GAUDI2B ; +break ; +default : +break ; + }break ;default : - asic_type =ASIC_INVALID ;break ; }
@@ -416,7 +426,7 @@ static int create_hdev(struct hl_device **dev, struct pci_dev *pdev)/* First, we must find out which ASIC are we handling. This is needed * to configure the behavior of the driver (kernel parameters) */ - hdev->asic_type =get_asic_type (pdev->device); + hdev->asic_type =get_asic_type (hdev);if (hdev->asic_type ==ASIC_INVALID ) {dev_err (&pdev->dev,"Unsupported ASIC\n" ); rc = -ENODEV ;
drivers/misc/habanalabs/common/habanalabs_ioctl.c +4 -2
@@ -10,10 +10,11 @@#include <uapi/misc/habanalabs.h> #include "habanalabs.h" -#include <linux/kernel.h> #include <linux/fs.h> -#include <linux/uaccess.h> +#include <linux/kernel.h> +#include <linux/pci.h> #include <linux/slab.h> +#include <linux/uaccess.h> #include <linux/vmalloc.h> static u32 hl_debug_struct_size[HL_DEBUG_OP_TIMESTAMP +1 ] = {
@@ -105,6 +106,7 @@ static int hw_ip_info(struct hl_device *hdev, struct hl_info_args *args) hw_ip.edma_enabled_mask = prop->edma_enabled_mask; hw_ip.server_type = prop->server_type; hw_ip.security_enabled = prop->fw_security_enabled; + hw_ip.revision_id = hdev->pdev->revision;return copy_to_user (out, &hw_ip,min ((size_t ) size,sizeof (hw_ip))) ? -EFAULT :0 ;
drivers/misc/habanalabs/common/mmu/mmu.c +1 -0
@@ -635,6 +635,7 @@ int hl_mmu_if_set_funcs(struct hl_device *hdev)hl_mmu_v1_set_funcs (hdev, &hdev->mmu_func[MMU_DR_PGT ]);break ;case ASIC_GAUDI2 : +case ASIC_GAUDI2B :/* MMUs in Gaudi2 are always host resident */ hl_mmu_v2_hr_set_funcs (hdev, &hdev->mmu_func[MMU_HR_PGT ]);break ;
drivers/misc/habanalabs/common/sysfs.c +2 -0
@@ -248,6 +248,8 @@ static ssize_t device_type_show(struct device *dev,case ASIC_GAUDI2 : str ="GAUDI2" ;break ; +case ASIC_GAUDI2B : + str ="GAUDI2B" ;break ;default : dev_err(hdev->dev, "Unrecognized ASIC type %d\n",
drivers/misc/habanalabs/gaudi2/gaudi2.c +1 -5
@@ -3968,11 +3968,7 @@ static void gaudi2_init_firmware_loader(struct hl_device *hdev) fw_loader->skip_bmc = false; fw_loader->sram_bar_id =SRAM_CFG_BAR_ID ; fw_loader->dram_bar_id =DRAM_BAR_ID ; - -if (hdev->asic_type ==ASIC_GAUDI2 ) - fw_loader->cpu_timeout =GAUDI2_CPU_TIMEOUT_USEC ; -else /* ASIC_GAUDI2_FPGA */ - fw_loader->cpu_timeout =GAUDI2_FPGA_CPU_TIMEOUT ; + fw_loader->cpu_timeout =GAUDI2_CPU_TIMEOUT_USEC ;/* here we update initial values for few specific dynamic regs (as * before reading the first descriptor from FW those value has to be
drivers/misc/habanalabs/gaudi2/gaudi2P.h +0 -2
@@ -23,8 +23,6 @@#define GAUDI2_CPU_TIMEOUT_USEC 30000000/* 30s */ -#define GAUDI2_FPGA_CPU_TIMEOUT 100000000/* 100s */ -#define NUMBER_OF_PDMA_QUEUES 2#define NUMBER_OF_EDMA_QUEUES 8#define NUMBER_OF_MME_QUEUES 4
drivers/misc/habanalabs/include/hw_ip/pci/pci_general.h +7 -0
@@ -20,4 +20,11 @@#define PCI_CONFIG_ELBI_STS_MASK (PCI_CONFIG_ELBI_STS_ERR | \ PCI_CONFIG_ELBI_STS_DONE) +enum hl_revision_id { +/* PCI revision ID 0 is not legal */ +REV_ID_INVALID =0x00 , +REV_ID_A =0x01 , +REV_ID_B =0x02 , +}; +#endif /* INCLUDE_PCI_GENERAL_H_ */
include/uapi/misc/habanalabs.h +7 -0
@@ -868,6 +868,7 @@ enum hl_server_type {* @number_of_user_interrupts: The number of interrupts that are available to the userspace * application to use. Relevant for Gaudi2 and later. * @device_mem_alloc_default_page_size: default page size used in device memory allocation. +* @revision_id: PCI revision ID of the ASIC. */ struct hl_info_hw_ip_info {__u64 sram_base_address;
@@ -898,6 +899,12 @@ struct hl_info_hw_ip_info {__u16 pad2;__u64 reserved4;__u64 device_mem_alloc_default_page_size; +__u64 reserved5; +__u64 reserved6; +__u32 reserved7; +__u8 reserved8; +__u8 revision_id; +__u8 pad[2 ]; };struct hl_info_dram_usage {
Signature · 1 line
2.25.1
PATCH 2/20 habanalabs/gaudi: add razwi notify event
From: Dani Liberman <[email protected]>
Each time razwi (read-only zero, write ignore) happens, besides capturing its data, also notify the user about it.
--- drivers/misc/habanalabs/common/device.c | 8 +++++ drivers/misc/habanalabs/common/habanalabs.h | 2 ++ drivers/misc/habanalabs/gaudi/gaudi.c | 37 +++++++++++---------- include/uapi/misc/habanalabs.h | 2 ++ 4 files changed, 31 insertions(+), 18 deletions(-)
drivers/misc/habanalabs/common/device.c +8 -0
@@ -2409,6 +2409,14 @@ void hl_capture_razwi(struct hl_device *hdev, u64 addr, u16 *engine_id, u16 num_ num_of_engines * sizeof(u16)); hdev->captured_err_info.razwi.flags = flags; } + +void hl_handle_razwi (struct hl_device *hdev,u64 addr,u16 *engine_id,u16 num_of_engines, +u8 flags,u64 *event_mask) +{ +hl_capture_razwi (hdev, addr, engine_id, num_of_engines, flags); + *event_mask |=HL_NOTIFIER_EVENT_RAZWI ; +} +static void hl_capture_user_mappings (struct hl_device *hdev,bool is_pmmu) {struct page_fault_info *pgf_info = &hdev->captured_err_info.pgf_info;
drivers/misc/habanalabs/common/habanalabs.h +2 -0
@@ -3812,6 +3812,8 @@ hl_mmap_mem_buf_alloc(struct hl_mem_mgr *mmg, __printf(2, 3) void hl_engine_data_sprintf(struct engines_data *e, const char *fmt, ...);void hl_capture_razwi (struct hl_device *hdev,u64 addr,u16 *engine_id,u16 num_of_engines,u8 flags); +void hl_handle_razwi (struct hl_device *hdev,u64 addr,u16 *engine_id,u16 num_of_engines, +u8 flags,u64 *event_mask);void hl_capture_page_fault (struct hl_device *hdev,u64 addr,u16 eng_id,bool is_pmmu);#ifdef CONFIG_DEBUG_FS
drivers/misc/habanalabs/gaudi/gaudi.c +19 -18
@@ -7301,7 +7301,7 @@ static void gaudi_handle_qman_err(struct hl_device *hdev, u16 event_type, u64 *e }static void gaudi_print_irq_info (struct hl_device *hdev,u16 event_type, -bool razwi) +bool razwi,u64 *event_mask) {bool is_read = false, is_write = false;u16 engine_id[2 ], num_of_razwi_eng =0 ;
@@ -7337,7 +7337,8 @@ static void gaudi_print_irq_info(struct hl_device *hdev, u16 event_type, num_of_razwi_eng =1 ; } -hl_capture_razwi (hdev, razwi_addr, engine_id, num_of_razwi_eng, razwi_flags); +hl_handle_razwi (hdev, razwi_addr, engine_id, num_of_razwi_eng, razwi_flags, + event_mask); } }
@@ -7675,7 +7676,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entr case GAUDI_EVENT_HBM_0_DERR ... GAUDI_EVENT_HBM_3_DERR:case GAUDI_EVENT_MMU_DERR : case GAUDI_EVENT_NIC0_CS_DBG_DERR ... GAUDI_EVENT_NIC4_CS_DBG_DERR: -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask);gaudi_handle_ecc_event (hdev, event_type, &eq_entry->ecc_data); event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ; fw_fatal_err_flag =HL_DRV_RESET_FW_FATAL_ERR ;
@@ -7685,7 +7686,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_AXI_ECC :case GAUDI_EVENT_L2_RAM_ECC : case GAUDI_EVENT_PLL0 ... GAUDI_EVENT_PLL17: -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask); fw_fatal_err_flag =HL_DRV_RESET_FW_FATAL_ERR ; event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ; goto reset_device;
@@ -7694,7 +7695,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_HBM1_SPI_0 :case GAUDI_EVENT_HBM2_SPI_0 :case GAUDI_EVENT_HBM3_SPI_0 : -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask);gaudi_hbm_read_interrupts (hdev,gaudi_hbm_event_to_dev (event_type), &eq_entry->hbm_ecc_data);
@@ -7706,7 +7707,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_HBM1_SPI_1 :case GAUDI_EVENT_HBM2_SPI_1 :case GAUDI_EVENT_HBM3_SPI_1 : -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask);gaudi_hbm_read_interrupts (hdev,gaudi_hbm_event_to_dev (event_type), &eq_entry->hbm_ecc_data);
@@ -7728,7 +7729,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entr* if the event is a TPC Assertion or a "real" TPC DEC. */ event_mask |=HL_NOTIFIER_EVENT_TPC_ASSERT ; -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask); reset_required =gaudi_tpc_read_interrupts (hdev,tpc_dec_event_to_tpc_id (event_type),"AXI_SLV_DEC_Error" );
@@ -7753,7 +7754,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_TPC5_KRN_ERR :case GAUDI_EVENT_TPC6_KRN_ERR :case GAUDI_EVENT_TPC7_KRN_ERR : -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask); reset_required =gaudi_tpc_read_interrupts (hdev,tpc_krn_event_to_tpc_id (event_type),"KRN_ERR" );
@@ -7792,7 +7793,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entr case GAUDI_EVENT_HBM_0_SERR ... GAUDI_EVENT_HBM_3_SERR: fallthrough;case GAUDI_EVENT_MMU_SERR : -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask);gaudi_handle_ecc_event (hdev, event_type, &eq_entry->ecc_data);hl_fw_unmask_irq (hdev, event_type); event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ;
@@ -7802,14 +7803,14 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_CPU_AXI_SPLITTER :case GAUDI_EVENT_PSOC_AXI_DEC :case GAUDI_EVENT_PSOC_PRSTN_FALL : -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask);hl_fw_unmask_irq (hdev, event_type); event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ;break ;case GAUDI_EVENT_MMU_PAGE_FAULT :case GAUDI_EVENT_MMU_WR_PERM : -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask);hl_fw_unmask_irq (hdev, event_type); event_mask |=HL_NOTIFIER_EVENT_USER_ENGINE_ERR ;break ;
@@ -7838,14 +7839,14 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_NIC4_QM1 : case GAUDI_EVENT_DMA0_CORE ... GAUDI_EVENT_DMA7_CORE: case GAUDI_EVENT_TPC0_QM ... GAUDI_EVENT_TPC7_QM: -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask);gaudi_handle_qman_err (hdev, event_type, &event_mask);hl_fw_unmask_irq (hdev, event_type); event_mask |= (HL_NOTIFIER_EVENT_USER_ENGINE_ERR |HL_NOTIFIER_EVENT_DEVICE_RESET );break ;case GAUDI_EVENT_RAZWI_OR_ADC_SW : -gaudi_print_irq_info (hdev, event_type, true); +gaudi_print_irq_info (hdev, event_type, true, &event_mask); event_mask |=HL_NOTIFIER_EVENT_USER_ENGINE_ERR ; goto reset_device;
@@ -7858,7 +7859,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrcase GAUDI_EVENT_TPC6_BMON_SPMU :case GAUDI_EVENT_TPC7_BMON_SPMU : case GAUDI_EVENT_DMA_BM_CH0 ... GAUDI_EVENT_DMA_BM_CH7: -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask);hl_fw_unmask_irq (hdev, event_type); event_mask |=HL_NOTIFIER_EVENT_USER_ENGINE_ERR ;break ;
@@ -7870,7 +7871,7 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrbreak ; case GAUDI_EVENT_DMA_IF_SEI_0 ... GAUDI_EVENT_DMA_IF_SEI_3: -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask);gaudi_print_sm_sei_info (hdev, event_type, &eq_entry->sm_sei_data); rc =hl_state_dump (hdev);
@@ -7899,18 +7900,18 @@ static void gaudi_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_entrbreak ;case GAUDI_EVENT_DEV_RESET_REQ : -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask); event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ; goto reset_device;case GAUDI_EVENT_PKT_QUEUE_OUT_SYNC : -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask);gaudi_print_out_of_sync_info (hdev, &eq_entry->pkt_sync_err); event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ; goto reset_device;case GAUDI_EVENT_FW_ALIVE_S : -gaudi_print_irq_info (hdev, event_type, false); +gaudi_print_irq_info (hdev, event_type, false, &event_mask);gaudi_print_fw_alive_info (hdev, &eq_entry->fw_alive); event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ; goto reset_device;
include/uapi/misc/habanalabs.h +2 -0
@@ -721,6 +721,7 @@ enum hl_server_type {* HL_NOTIFIER_EVENT_DEVICE_UNAVAILABLE - Indicates device is unavailable * HL_NOTIFIER_EVENT_USER_ENGINE_ERR - Indicates device engine in error state * HL_NOTIFIER_EVENT_GENERAL_HW_ERR - Indicates device HW error +* HL_NOTIFIER_EVENT_RAZWI - Indicates razwi happened */ #define HL_NOTIFIER_EVENT_TPC_ASSERT (1ULL << 0)#define HL_NOTIFIER_EVENT_UNDEFINED_OPCODE (1ULL << 1)
@@ -729,6 +730,7 @@ enum hl_server_type {#define HL_NOTIFIER_EVENT_DEVICE_UNAVAILABLE (1ULL << 4)#define HL_NOTIFIER_EVENT_USER_ENGINE_ERR (1ULL << 5)#define HL_NOTIFIER_EVENT_GENERAL_HW_ERR (1ULL << 6) +#define HL_NOTIFIER_EVENT_RAZWI (1ULL << 7)/* Opcode for management ioctl *
Signature · 1 line
2.25.1
PATCH 3/20 habanalabs: use single threaded WQ for event handling
In reply to Oded GabbayFrom: Dani Liberman <[email protected]>
Creating event queue workqueue using alloc_workqueue made it run in multi threaded mode, which caused parallel dumping of events as well as parallel events notifying to user, causing logs with multiple events to be out of order.
Fixed by creating event queue workqueue as single threaded work queue.
--- drivers/misc/habanalabs/common/device.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/misc/habanalabs/common/device.c +1 -1
@@ -787,7 +787,7 @@ static int device_early_init(struct hl_device *hdev) } } - hdev->eq_wq =alloc_workqueue ("hl-events" ,WQ_UNBOUND ,0 ); + hdev->eq_wq =create_singlethread_workqueue ("hl-events" );if (hdev->eq_wq ==NULL ) {dev_err (hdev->dev,"Failed to allocate EQ workqueue\n" ); rc = -ENOMEM ;
Signature · 1 line
2.25.1
PATCH 4/20 habanalabs/gaudi: add page fault notify event
In reply to Oded GabbayFrom: Dani Liberman <[email protected]>
Each time page fault happens, besides capturing its data, also notify the user about it.
--- drivers/misc/habanalabs/common/device.c | 9 +++++++++ drivers/misc/habanalabs/common/habanalabs.h | 2 ++ drivers/misc/habanalabs/gaudi/gaudi.c | 6 +++--- include/uapi/misc/habanalabs.h | 2 ++ 4 files changed, 16 insertions(+), 3 deletions(-)
drivers/misc/habanalabs/common/device.c +9 -0
@@ -2490,3 +2490,12 @@ void hl_capture_page_fault(struct hl_device *hdev, u64 addr, u16 eng_id, bool is hdev->captured_err_info.pgf_info.pgf.engine_id = eng_id;hl_capture_user_mappings (hdev, is_pmmu); } + +void hl_handle_page_fault (struct hl_device *hdev,u64 addr,u16 eng_id,bool is_pmmu, +u64 *event_mask) +{ +hl_capture_page_fault (hdev, addr, eng_id, is_pmmu); + +if (event_mask) + *event_mask |=HL_NOTIFIER_EVENT_PAGE_FAULT ; +}
drivers/misc/habanalabs/common/habanalabs.h +2 -0
@@ -3815,6 +3815,8 @@ void hl_capture_razwi(struct hl_device *hdev, u64 addr, u16 *engine_id, u16 num_void hl_handle_razwi (struct hl_device *hdev,u64 addr,u16 *engine_id,u16 num_of_engines,u8 flags,u64 *event_mask);void hl_capture_page_fault (struct hl_device *hdev,u64 addr,u16 eng_id,bool is_pmmu); +void hl_handle_page_fault (struct hl_device *hdev,u64 addr,u16 eng_id,bool is_pmmu, +u64 *event_mask);#ifdef CONFIG_DEBUG_FS
drivers/misc/habanalabs/gaudi/gaudi.c +3 -3
@@ -6740,7 +6740,7 @@ static void gaudi_print_and_get_razwi_info(struct hl_device *hdev, u16 *engine_i } } -static void gaudi_print_and_get_mmu_error_info (struct hl_device *hdev,u64 *addr) +static void gaudi_print_and_get_mmu_error_info (struct hl_device *hdev,u64 *addr,u64 *event_mask) {struct gaudi_device *gaudi = hdev->asic_specific;u32 val;
@@ -6755,7 +6755,7 @@ static void gaudi_print_and_get_mmu_error_info(struct hl_device *hdev, u64 *addr *addr |=RREG32 (mmMMU_UP_PAGE_ERROR_CAPTURE_VA);dev_err_ratelimited (hdev->dev,"MMU page fault on va 0x%llx\n" , *addr); -hl_capture_page_fault (hdev, *addr,0 , true); +hl_handle_page_fault (hdev, *addr,0 , true, event_mask);WREG32 (mmMMU_UP_PAGE_ERROR_CAPTURE,0 ); }
@@ -7323,7 +7323,7 @@ static void gaudi_print_irq_info(struct hl_device *hdev, u16 event_type,if (razwi) {gaudi_print_and_get_razwi_info (hdev, &engine_id[0 ], &engine_id[1 ], &is_read, &is_write); -gaudi_print_and_get_mmu_error_info (hdev, &razwi_addr); +gaudi_print_and_get_mmu_error_info (hdev, &razwi_addr, event_mask);if (is_read) razwi_flags |=HL_RAZWI_READ ;
include/uapi/misc/habanalabs.h +2 -0
@@ -722,6 +722,7 @@ enum hl_server_type {* HL_NOTIFIER_EVENT_USER_ENGINE_ERR - Indicates device engine in error state * HL_NOTIFIER_EVENT_GENERAL_HW_ERR - Indicates device HW error * HL_NOTIFIER_EVENT_RAZWI - Indicates razwi happened +* HL_NOTIFIER_EVENT_PAGE_FAULT - Indicates page fault happened */ #define HL_NOTIFIER_EVENT_TPC_ASSERT (1ULL << 0)#define HL_NOTIFIER_EVENT_UNDEFINED_OPCODE (1ULL << 1)
@@ -731,6 +732,7 @@ enum hl_server_type {#define HL_NOTIFIER_EVENT_USER_ENGINE_ERR (1ULL << 5)#define HL_NOTIFIER_EVENT_GENERAL_HW_ERR (1ULL << 6)#define HL_NOTIFIER_EVENT_RAZWI (1ULL << 7) +#define HL_NOTIFIER_EVENT_PAGE_FAULT (1ULL << 8)/* Opcode for management ioctl *
Signature · 1 line
2.25.1
PATCH 5/20 habanalabs/gaudi2: implement fp32 not supported event
In reply to Oded GabbayFrom: Ofir Bitton <[email protected]>
Due to binning, Gaudi2 does not always support fp32.
We add support for such an event in case fp32 is used by the user in such a device.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 5 +++++ drivers/misc/habanalabs/include/gaudi2/gaudi2_async_events.h | 1 + .../include/gaudi2/gaudi2_async_ids_map_extended.h | 4 +++- 3 files changed, 9 insertions(+), 1 deletion(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +5 -0
@@ -9148,6 +9148,11 @@ static void gaudi2_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_ent event_mask |=HL_NOTIFIER_EVENT_USER_ENGINE_ERR ;break ; +case GAUDI2_EVENT_CPU_FP32_NOT_SUPPORTED : + event_mask |=HL_NOTIFIER_EVENT_GENERAL_HW_ERR ; + is_critical = true; +break ; +default :if (gaudi2_irq_map_table[event_type].valid) dev_err_ratelimited(hdev->dev, "Cannot find handler for event %d\n",
drivers/misc/habanalabs/include/gaudi2/gaudi2_async_events.h +1 -0
@@ -957,6 +957,7 @@ enum gaudi2_async_event_id {GAUDI2_EVENT_CPU11_STATUS_NIC11_ENG0 =1317 ,GAUDI2_EVENT_CPU11_STATUS_NIC11_ENG1 =1318 ,GAUDI2_EVENT_ARC_DCCM_FULL =1319 , +GAUDI2_EVENT_CPU_FP32_NOT_SUPPORTED =1320 , GAUDI2_EVENT_SIZE, };
drivers/misc/habanalabs/include/gaudi2/gaudi2_async_ids_map_extended.h +3 -1
@@ -1,6 +1,6 @@/* SPDX-License-Identifier: GPL-2.0 * -* Copyright 2018-2021 HabanaLabs, Ltd. +* Copyright 2018-2022 HabanaLabs, Ltd. * All Rights Reserved. * */
@@ -2663,6 +2663,8 @@ static struct gaudi2_async_events_ids_map gaudi2_irq_map_table[] = { .msg = 1, .reset = 0, .name = "STATUS_NIC11_ENG1" }, { .fc_id = 1319, .cpu_id = 625, .valid = 1, .msg = 1, .reset = 0, .name = "ARC_DCCM_FULL" }, + { .fc_id = 1320, .cpu_id = 626, .valid = 1, + .msg = 1, .reset = 1, .name = "FP32_NOT_SUPPORTED" }, };#endif /* __GAUDI2_ASYNC_IDS_MAP_EVENTS_EXT_H_ */
Signature · 1 line
2.25.1
PATCH 6/20 habanalabs/gaudi2: add razwi notify event
In reply to Oded GabbayFrom: Dani Liberman <[email protected]>
Each time razwi (read-only zero, write ignored) event happens, besides capturing its data, also notify the user about it.
--- drivers/misc/habanalabs/common/device.c | 4 +- drivers/misc/habanalabs/gaudi2/gaudi2.c | 140 +++++++++++++----------- 2 files changed, 82 insertions(+), 62 deletions(-)
drivers/misc/habanalabs/common/device.c +3 -1
@@ -2414,7 +2414,9 @@ void hl_handle_razwi(struct hl_device *hdev, u64 addr, u16 *engine_id, u16 num_o u8 flags, u64 *event_mask) {hl_capture_razwi (hdev ,addr ,engine_id ,num_of_engines ,flags ); - *event_mask |=HL_NOTIFIER_EVENT_RAZWI ; + +if (event_mask) + *event_mask |=HL_NOTIFIER_EVENT_RAZWI ; }static void hl_capture_user_mappings (struct hl_device *hdev,bool is_pmmu)
drivers/misc/habanalabs/gaudi2/gaudi2.c +79 -61
441 lines not on this page · raw message
Signature · 1 line
2.25.1
PATCH 7/20 habanalabs: fix firmware descriptor copy operation
In reply to Oded GabbayFrom: farah kassabri <[email protected]>
This is needed to allow adding more data to the lkd_fw_comms_desc structure.
--- drivers/misc/habanalabs/common/firmware_if.c | 26 ++++++++++++++++++-- 1 file changed, 24 insertions(+), 2 deletions(-)
drivers/misc/habanalabs/common/firmware_if.c +24 -2
@@ -12,6 +12,7 @@#include <linux/crc32.h> #include <linux/slab.h> #include <linux/ctype.h> +#include <linux/vmalloc.h> #define FW_FILE_MAX_SIZE 0x1400000/* maximum size of 20MB */
@@ -1988,10 +1989,11 @@ static int hl_fw_dynamic_read_and_validate_descriptor(struct hl_device *hdev, struct fw_load_mgr *fw_loader) {struct lkd_fw_comms_desc *fw_desc; +void __iomem *src, *temp_fw_desc;struct pci_mem_region *region;struct fw_response *response; +u16 fw_data_size;enum pci_region region_id; -void __iomem *src;int rc; fw_desc = &fw_loader->dynamic_loader.comm_desc;
@@ -2018,9 +2020,29 @@ static int hl_fw_dynamic_read_and_validate_descriptor(struct hl_device *hdev, fw_loader->dynamic_loader.fw_desc_valid = false; src = hdev->pcie_bar[region->bar_id] + region->offset_in_bar + response->ram_offset; + +/* +* We do the copy of the fw descriptor in 2 phases: +* 1. copy the header + data info according to our lkd_fw_comms_desc definition. +* then we're able to read the actual data size provided by fw. +* this is needed for cases where data in descriptor was changed(add/remove) +* in embedded specs header file before updating lkd copy of the header file +* 2. copy descriptor to temporary buffer with aligned size and send it to validation +*/ memcpy_fromio (fw_desc, src,sizeof (struct lkd_fw_comms_desc )); + fw_data_size =le16_to_cpu (fw_desc->header.size); + + temp_fw_desc =vzalloc (sizeof (struct comms_desc_header ) + fw_data_size); +if (!temp_fw_desc) +return -ENOMEM ; + +memcpy_fromio (temp_fw_desc, src,sizeof (struct comms_desc_header ) + fw_data_size); -return hl_fw_dynamic_validate_descriptor (hdev, fw_loader, fw_desc); + rc =hl_fw_dynamic_validate_descriptor (hdev, fw_loader, + (struct lkd_fw_comms_desc *) temp_fw_desc); +vfree (temp_fw_desc); + +return rc; }/**
Signature · 1 line
2.25.1
PATCH 8/20 habanalabs: skip events info ioctl if not supported
In reply to Oded GabbayFrom: Ohad Sharabi <[email protected]>
Some ASICs haven't yet implemented this functionality and so the ioctl call should fail and the user should be notified of the reason.
--- drivers/misc/habanalabs/common/habanalabs_ioctl.c | 4 ++++ 1 file changed, 4 insertions(+)
drivers/misc/habanalabs/common/habanalabs_ioctl.c +4 -0
@@ -123,6 +123,10 @@ static int hw_events_info(struct hl_device *hdev, bool aggregate,return -EINVAL ; arr = hdev->asic_funcs->get_events_stat (hdev, aggregate, &size); +if (!arr) { +dev_err (hdev->dev,"Events info not supported\n" ); +return -EOPNOTSUPP ; + }return copy_to_user (out, arr,min (max_size, size)) ? -EFAULT :0 ; }
Signature · 1 line
2.25.1
PATCH 9/20 habanalabs/gaudi2: classify power/thermal events as info
In reply to Oded GabbayFrom: Ofir Bitton <[email protected]>
As power and thermal envelope events are pure informative and not indicating an error, we reduce the print level to info only.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 1 + 1 file changed, 1 insertion(+)
drivers/misc/habanalabs/gaudi2/gaudi2.c +1 -0
@@ -6828,6 +6828,7 @@ static inline bool is_info_event(u32 event) {switch (event) {case GAUDI2_EVENT_CPU_CPLD_SHUTDOWN_CAUSE : + case GAUDI2_EVENT_CPU_FIX_POWER_ENV_S ... GAUDI2_EVENT_CPU_FIX_THERMAL_ENV_E:return true;default :return false;
Signature · 1 line
2.25.1
PATCH 10/20 habanalabs/gaudi2: add page fault notify event
In reply to Oded GabbayFrom: Dani Liberman <[email protected]>
Each time page fault happens, besides capturing its data, also notify the user about it.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 17 +++++++++-------- 1 file changed, 9 insertions(+), 8 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +9 -8
@@ -8253,7 +8253,8 @@ static void gaudi2_handle_hif_fatal(struct hl_device *hdev, u16 event_type, u64 } } -static void gaudi2_handle_page_error (struct hl_device *hdev,u64 mmu_base,bool is_pmmu) +static void gaudi2_handle_page_error (struct hl_device *hdev,u64 mmu_base,bool is_pmmu, +u64 *event_mask) {u32 valid, val;u64 addr;
@@ -8270,7 +8271,7 @@ static void gaudi2_handle_page_error(struct hl_device *hdev, u64 mmu_base, booldev_err_ratelimited (hdev->dev,"%s page fault on va 0x%llx\n" , is_pmmu ?"PMMU" :"HMMU" , addr); -hl_capture_page_fault (hdev, addr,0 , is_pmmu); +hl_handle_page_fault (hdev, addr,0 , is_pmmu, event_mask);WREG32 (mmu_base +MMU_OFFSET (mmDCORE0_HMMU0_MMU_PAGE_ERROR_CAPTURE),0 ); }
@@ -8296,7 +8297,7 @@ static void gaudi2_handle_access_error(struct hl_device *hdev, u64 mmu_base, boo }static void gaudi2_handle_mmu_spi_sei_generic (struct hl_device *hdev,const char *mmu_name, -u64 mmu_base,bool is_pmmu) +u64 mmu_base,bool is_pmmu,u64 *event_mask) {u32 spi_sei_cause, interrupt_clr =0x0 ;int i;
@@ -8309,7 +8310,7 @@ static void gaudi2_handle_mmu_spi_sei_generic(struct hl_device *hdev, const char mmu_name, gaudi2_mmu_spi_sei[i].cause);if (i ==0 ) -gaudi2_handle_page_error (hdev, mmu_base, is_pmmu); +gaudi2_handle_page_error (hdev, mmu_base, is_pmmu, event_mask);else if (i ==1 )gaudi2_handle_access_error (hdev, mmu_base, is_pmmu);
@@ -8381,7 +8382,7 @@ static bool gaudi2_handle_sm_err(struct hl_device *hdev, u8 sm_index)return reset; } -static void gaudi2_handle_mmu_spi_sei_err (struct hl_device *hdev,u16 event_type) +static void gaudi2_handle_mmu_spi_sei_err (struct hl_device *hdev,u16 event_type,u64 *event_mask) {bool is_pmmu = false;char desc[32 ];
@@ -8439,7 +8440,7 @@ static void gaudi2_handle_mmu_spi_sei_err(struct hl_device *hdev, u16 event_typereturn ; } -gaudi2_handle_mmu_spi_sei_generic (hdev, desc, mmu_base, is_pmmu); +gaudi2_handle_mmu_spi_sei_generic (hdev, desc, mmu_base, is_pmmu, event_mask); }
@@ -8969,7 +8970,7 @@ static void gaudi2_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_ent case GAUDI2_EVENT_HMMU_0_AXI_ERR_RSP ... GAUDI2_EVENT_HMMU_12_AXI_ERR_RSP: case GAUDI2_EVENT_PMMU0_PAGE_FAULT_WR_PERM ... GAUDI2_EVENT_PMMU0_SECURITY_ERROR:case GAUDI2_EVENT_PMMU_AXI_ERR_RSP_0 : -gaudi2_handle_mmu_spi_sei_err (hdev, event_type); +gaudi2_handle_mmu_spi_sei_err (hdev, event_type, &event_mask); reset_flags |=HL_DRV_RESET_FW_FATAL_ERR ; event_mask |=HL_NOTIFIER_EVENT_USER_ENGINE_ERR ;break ;
@@ -10206,7 +10207,7 @@ static void gaudi2_ack_mmu_error(struct hl_device *hdev, u64 mmu_id)if (gaudi2_get_mmu_base (hdev, mmu_id, &mmu_base))return ; -gaudi2_handle_page_error (hdev, mmu_base, is_pmmu); +gaudi2_handle_page_error (hdev, mmu_base, is_pmmu,NULL );gaudi2_handle_access_error (hdev, mmu_base, is_pmmu); }
Signature · 1 line
2.25.1
PATCH 11/20 habanalabs: fix print for out-of-sync and pkt-failure events
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Add missing le32_to_cpu() conversions, and use %d for the value returned from atomic_read().
--- drivers/misc/habanalabs/gaudi/gaudi.c | 4 ++-- drivers/misc/habanalabs/gaudi2/gaudi2.c | 8 ++++---- drivers/misc/habanalabs/goya/goya.c | 4 ++-- 3 files changed, 8 insertions(+), 8 deletions(-)
drivers/misc/habanalabs/gaudi/gaudi.c +2 -2
@@ -7347,8 +7347,8 @@ static void gaudi_print_out_of_sync_info(struct hl_device *hdev, {struct hl_hw_queue *q = &hdev->kernel_queues[GAUDI_QUEUE_ID_CPU_PQ ]; -dev_err (hdev->dev,"Out of sync with FW, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%u\n" , - sync_err->pi, sync_err->ci, q->pi,atomic_read (&q->ci)); +dev_err (hdev->dev,"Out of sync with FW, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%d\n" , +le32_to_cpu (sync_err->pi),le32_to_cpu (sync_err->ci), q->pi,atomic_read (&q->ci)); } static void gaudi_print_fw_alive_info(struct hl_device *hdev,
drivers/misc/habanalabs/gaudi2/gaudi2.c +4 -4
@@ -8684,8 +8684,8 @@ static void gaudi2_print_out_of_sync_info(struct hl_device *hdev, {struct hl_hw_queue *q = &hdev->kernel_queues[GAUDI2_QUEUE_ID_CPU_PQ ]; -dev_err (hdev->dev,"Out of sync with FW, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%u\n" , - sync_err->pi, sync_err->ci, q->pi,atomic_read (&q->ci)); +dev_err (hdev->dev,"Out of sync with FW, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%d\n" , +le32_to_cpu (sync_err->pi),le32_to_cpu (sync_err->ci), q->pi,atomic_read (&q->ci)); }static void gaudi2_handle_pcie_p2p_msix (struct hl_device *hdev)
@@ -8751,8 +8751,8 @@ static void gaudi2_print_cpu_pkt_failure_info(struct hl_device *hdev,struct hl_hw_queue *q = &hdev->kernel_queues[GAUDI2_QUEUE_ID_CPU_PQ ];dev_warn (hdev->dev, -"FW reported sanity check failure, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%u\n" , - sync_err->pi, sync_err->ci, q->pi,atomic_read (&q->ci)); +"FW reported sanity check failure, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%d\n" , +le32_to_cpu (sync_err->pi),le32_to_cpu (sync_err->ci), q->pi,atomic_read (&q->ci)); } static void hl_arc_event_handle(struct hl_device *hdev,
drivers/misc/habanalabs/goya/goya.c +2 -2
@@ -4475,8 +4475,8 @@ static void goya_print_out_of_sync_info(struct hl_device *hdev, {struct hl_hw_queue *q = &hdev->kernel_queues[GOYA_QUEUE_ID_CPU_PQ ]; -dev_err (hdev->dev,"Out of sync with FW, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%u\n" , - sync_err->pi, sync_err->ci, q->pi,atomic_read (&q->ci)); +dev_err (hdev->dev,"Out of sync with FW, FW: pi=%u, ci=%u, LKD: pi=%u, ci=%d\n" , +le32_to_cpu (sync_err->pi),le32_to_cpu (sync_err->ci), q->pi,atomic_read (&q->ci)); } static void goya_print_irq_info(struct hl_device *hdev, u16 event_type,
Signature · 1 line
2.25.1
PATCH 12/20 habanalabs/gaudi: fix print for firmware-alive event
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Add missing le{32,64}_to_cpu conversions.
--- drivers/misc/habanalabs/gaudi/gaudi.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/misc/habanalabs/gaudi/gaudi.c +4 -3
@@ -7356,9 +7356,10 @@ static void gaudi_print_fw_alive_info(struct hl_device *hdev, {dev_err (hdev->dev,"FW alive report: severity=%s, process_id=%u, thread_id=%u, uptime=%llu seconds\n" , - (fw_alive->severity ==FW_ALIVE_SEVERITY_MINOR ) ? -"Minor" :"Critical" , fw_alive->process_id, - fw_alive->thread_id, fw_alive->uptime_seconds); + (fw_alive->severity ==FW_ALIVE_SEVERITY_MINOR ) ?"Minor" :"Critical" , +le32_to_cpu (fw_alive->process_id), +le32_to_cpu (fw_alive->thread_id), +le64_to_cpu (fw_alive->uptime_seconds)); } static void gaudi_print_nic_axi_irq_info(struct hl_device *hdev, u16 event_type,
Signature · 1 line
2.25.1
PATCH 13/20 habanalabs/gaudi2: remove redundant firmware version check
In reply to Oded GabbayFrom: farah kassabri <[email protected]>
Firmware 1.7 is the first official firmware, so no need to check if we are running a version below it.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +1 -2
@@ -10358,10 +10358,9 @@ int gaudi2_send_device_activity(struct hl_device *hdev, bool open) {struct gaudi2_device *gaudi2 = hdev->asic_specific; -if (!(gaudi2->hw_cap_initialized &HW_CAP_CPU_Q ) || hdev->fw_major_version <37 ) +if (!(gaudi2->hw_cap_initialized &HW_CAP_CPU_Q ))return 0 ; -/* TODO: add check for FW version using minor ver once it's known */ return hl_fw_send_device_activity (hdev, open); }
Signature · 1 line
2.25.1
PATCH 14/20 habanalabs/gaudi2: don't enable entries in the MSIX_GW table
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
User should use the virtual MSI-X doorbell to generate interrupts from the device, so there is no need to enable entries in the MSIX_GW table.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 26 ------------------------- 1 file changed, 26 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +0 -26
@@ -4695,30 +4695,6 @@ static void gaudi2_init_dec(struct hl_device *hdev) } } -static void gaudi2_init_msix_gw_table (struct hl_device *hdev) -{ -u32 first_reg_offset, last_reg_offset, msix_gw_table_base; -u8 first_bit, last_bit; -int i; - - msix_gw_table_base = mmPCIE_WRAP_MSIX_GW_TABLE_0; - first_reg_offset = (GAUDI2_IRQ_NUM_USER_FIRST >>5 ) <<2 ; - first_bit =GAUDI2_IRQ_NUM_USER_FIRST %32 ; - last_reg_offset = (GAUDI2_IRQ_NUM_USER_LAST >>5 ) <<2 ; - last_bit =GAUDI2_IRQ_NUM_USER_LAST %32 ; - -if (first_reg_offset == last_reg_offset) { -WREG32 (msix_gw_table_base + first_reg_offset,GENMASK (last_bit, first_bit)); -return ; - } - -WREG32 (msix_gw_table_base + first_reg_offset,GENMASK (31 , first_bit)); -WREG32 (msix_gw_table_base + last_reg_offset,GENMASK (last_bit,0 )); - -for (i = first_reg_offset +4 ; i < last_reg_offset ; i +=4 ) -WREG32 (msix_gw_table_base + i,0xFFFFFFFF ); -} -static int gaudi2_mmu_update_asid_hop0_addr (struct hl_device *hdev,u32 stlb_base,u32 asid,u64 phys_addr) {
@@ -5232,8 +5208,6 @@ static int gaudi2_hw_init(struct hl_device *hdev)return rc; } -gaudi2_init_msix_gw_table (hdev); -gaudi2_init_scrambler_hbm (hdev);gaudi2_init_kdma (hdev);
Signature · 1 line
2.25.1
PATCH 15/20 habanalabs/gaudi2: return to reset upon SM SEI BRESP error
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Due to a H/W issue in the LBW path to the PCIE_DBI MSI-X doorbell, there were false sporadic error responses in SM when it was configured to write to there, and hence no reset was done as part of handling the relevant event.
Now that the virtual MSI-X doorbell is used, such errors in SM are not expected and reset shouldn't be skipped.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 19 ++++++------------- 1 file changed, 6 insertions(+), 13 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +6 -13
@@ -8300,11 +8300,10 @@ static void gaudi2_handle_mmu_spi_sei_generic(struct hl_device *hdev, const charWREG32 (mmu_base +MMU_INTERRUPT_CLR_OFFSET , interrupt_clr); } -static bool gaudi2_handle_sm_err (struct hl_device *hdev,u8 sm_index) +static void gaudi2_handle_sm_err (struct hl_device *hdev,u8 sm_index) {u32 sei_cause_addr, sei_cause_val, sei_cause_cause, sei_cause_log;u32 cq_intr_addr, cq_intr_val, cq_intr_queue_index; -bool reset = true;int i; sei_cause_addr = mmDCORE0_SYNC_MNGR_GLBL_SM_SEI_CAUSE +DCORE_OFFSET * sm_index;
@@ -8329,10 +8328,6 @@ static bool gaudi2_handle_sm_err(struct hl_device *hdev, u8 sm_index) gaudi2_sm_sei_cause[i].cause_name, gaudi2_sm_sei_cause[i].log_name, sei_cause_log & gaudi2_sm_sei_cause[i].log_mask); - -/* Due to a potential H/W issue, do not reset upon BRESP errors */ -if (i ==2 ) - reset = false;break ; }
@@ -8352,8 +8347,6 @@ static bool gaudi2_handle_sm_err(struct hl_device *hdev, u8 sm_index)/* Clear CQ_INTR */ WREG32 (cq_intr_addr,0 ); } - -return reset; }static void gaudi2_handle_mmu_spi_sei_err (struct hl_device *hdev,u16 event_type,u64 *event_mask)
@@ -8755,8 +8748,8 @@ static void hl_arc_event_handle(struct hl_device *hdev,static void gaudi2_handle_eqe (struct hl_device *hdev,struct hl_eq_entry *eq_entry) { -bool reset_required = false, skip_reset = false, is_critical = false;struct gaudi2_device *gaudi2 = hdev->asic_specific; +bool reset_required = false, is_critical = false;u32 ctl, reset_flags =HL_DRV_RESET_HARD ;int index, sbte_index;u64 event_mask =0 ;
@@ -9113,7 +9106,7 @@ static void gaudi2_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_ent case GAUDI2_EVENT_SM0_AXI_ERROR_RESPONSE ... GAUDI2_EVENT_SM3_AXI_ERROR_RESPONSE: index = event_type -GAUDI2_EVENT_SM0_AXI_ERROR_RESPONSE ; - skip_reset = !gaudi2_handle_sm_err (hdev, index); +gaudi2_handle_sm_err (hdev, index); event_mask |=HL_NOTIFIER_EVENT_USER_ENGINE_ERR ;break ;
@@ -9153,9 +9146,9 @@ static void gaudi2_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_ent event_type); } -if ((gaudi2_irq_map_table[event_type].reset || reset_required) && !skip_reset && - (hdev->hard_reset_on_fw_events || - (hdev->asic_prop.fw_security_enabled && is_critical))) +if ((gaudi2_irq_map_table[event_type].reset || reset_required) && + (hdev->hard_reset_on_fw_events || + (hdev->asic_prop.fw_security_enabled && is_critical))) goto reset_device;/* Send unmask irq only for interrupts not classified as MSG */
Signature · 1 line
2.25.1
PATCH 16/20 habanalabs: reset device if still in use when released
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
If the device file is released while a context is still held, it won't be possible to reopen it until the context is eventually released. If that doesn't happen, only a device reset will revert it back to an operational state, i.e. need to wait for a CS timeout or an error, or to wait for an external intervention of injecting a reset via sysfs.
At this stage, after the device was released by user, context is held either because of CS which were left running on the device and are not relevant anymore, or due to missing cleanup steps from user side.
All of this is in any case handled in the device reset flow, so initiate the reset at this point instead of waiting for it.
--- drivers/misc/habanalabs/common/device.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/misc/habanalabs/common/device.c +4 -3
@@ -504,9 +504,10 @@ static int hl_device_release(struct inode *inode, struct file *filp) hdev->compute_ctx_in_release =1 ; -if (!hl_hpriv_put (hpriv)) -dev_notice (hdev->dev, -"User process closed FD but device still in use\n" ); +if (!hl_hpriv_put (hpriv)) { +dev_notice (hdev->dev,"User process closed FD but device still in use\n" ); +hl_device_reset (hdev,HL_DRV_RESET_HARD ); + } hdev->last_open_session_duration_jif = jiffies - hdev->last_successful_open_jif;
Signature · 1 line
2.25.1
PATCH 17/20 habanalabs: check schedule_hard_reset correctly
In reply to Oded Gabbayschedule_hard_reset can be true only if we didn't do hard-reset. Therefore, no point of checking it in case hard_reset is true.
--- drivers/misc/habanalabs/common/device.c | 25 +++++++++++++------------ 1 file changed, 13 insertions(+), 12 deletions(-)
drivers/misc/habanalabs/common/device.c +13 -12
@@ -1737,18 +1737,19 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)* the device will be operational although it shouldn't be */ hdev->asic_funcs->enable_events_from_fw (hdev); - }else if (!reset_upon_device_release) { - hdev->reset_info.compute_reset_cnt++; - } - -if (schedule_hard_reset) { -dev_info (hdev->dev,"Performing hard reset scheduled during compute reset\n" ); - flags = hdev->reset_info.hard_reset_schedule_flags; - hdev->reset_info.hard_reset_schedule_flags =0 ; - hdev->disabled = true; - hard_reset = true; -handle_reset_trigger (hdev, flags); - goto again; + }else { +if (!reset_upon_device_release) + hdev->reset_info.compute_reset_cnt++; + +if (schedule_hard_reset) { +dev_info (hdev->dev,"Performing hard reset scheduled during compute reset\n" ); + flags = hdev->reset_info.hard_reset_schedule_flags; + hdev->reset_info.hard_reset_schedule_flags =0 ; + hdev->disabled = true; + hard_reset = true; +handle_reset_trigger (hdev, flags); + goto again; + } }return 0 ;
Signature · 1 line
2.25.1
PATCH 18/20 habanalabs: extend process wait timeout in device fine
In reply to Oded GabbayProcesses that use our device are likely to use at the same time other devices such as remote storage.
In case our device is removed and a user process is still using the device, we need to kill the user process. However, if that process has a thread waiting for i/o to complete on remote storage, for example, the process won't terminate.
Let's give it enough time to terminate before giving up.
--- drivers/misc/habanalabs/common/device.c | 6 ++++-- drivers/misc/habanalabs/common/habanalabs.h | 11 ++++++++--- 2 files changed, 12 insertions(+), 5 deletions(-)
drivers/misc/habanalabs/common/device.c +4 -2
@@ -2300,14 +2300,16 @@ void hl_device_fini(struct hl_device *hdev)*/ dev_info (hdev->dev,"Waiting for all processes to exit (timeout of %u seconds)" , -HL_PENDING_RESET_LONG_SEC ); +HL_WAIT_PROCESS_KILL_ON_DEVICE_FINI ); - rc =device_kill_open_processes (hdev,HL_PENDING_RESET_LONG_SEC , false); + hdev->process_kill_trial_cnt =0 ; + rc =device_kill_open_processes (hdev,HL_WAIT_PROCESS_KILL_ON_DEVICE_FINI , false);if (rc) {dev_crit (hdev->dev,"Failed to kill all open processes\n" );device_disable_open_processes (hdev, false); } + hdev->process_kill_trial_cnt =0 ; rc =device_kill_open_processes (hdev,0 , true);if (rc) {dev_crit (hdev->dev,"Failed to kill all control device open processes\n" );
drivers/misc/habanalabs/common/habanalabs.h +8 -3
@@ -50,9 +50,14 @@ struct hl_fpriv;#define HL_MMAP_OFFSET_VALUE_MASK (0x1FFFFFFFFFFFull >> PAGE_SHIFT)#define HL_MMAP_OFFSET_VALUE_GET (off) (off & HL_MMAP_OFFSET_VALUE_MASK) -#define HL_PENDING_RESET_PER_SEC 10 -#define HL_PENDING_RESET_MAX_TRIALS 60/* 10 minutes */ -#define HL_PENDING_RESET_LONG_SEC 60 +#define HL_PENDING_RESET_PER_SEC 10 +#define HL_PENDING_RESET_MAX_TRIALS 60/* 10 minutes */ +#define HL_PENDING_RESET_LONG_SEC 60 +/* +* In device fini, wait 10 minutes for user processes to be terminated after we kill them. +* This is needed to prevent situation of clearing resources while user processes are still alive. +*/ +#define HL_WAIT_PROCESS_KILL_ON_DEVICE_FINI 600#define HL_HARD_RESET_MAX_TIMEOUT 120#define HL_PLDM_HARD_RESET_MAX_TIMEOUT (HL_HARD_RESET_MAX_TIMEOUT * 3)
Signature · 1 line
2.25.1
PATCH 19/20 habanalabs/gaudi2: change memory scrub mechanism
In reply to Oded GabbayFrom: farah kassabri <[email protected]>
Currently the scrubbing mechanism used the EDMA engines by directly setting the engine core registers to scrub a chunk of memory. Due to a sporadic failure with this mechanism, it was decided to initiate the engines via its QMAN using LIN-DMA packets.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 129 +++++++++++++++--------- 1 file changed, 83 insertions(+), 46 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +83 -46
@@ -9171,34 +9171,74 @@ static void gaudi2_handle_eqe(struct hl_device *hdev, struct hl_eq_entry *eq_enthl_device_cond_reset (hdev, reset_flags, event_mask); } +static int gaudi2_memset_memory_chunk_using_edma_qm (struct hl_device *hdev, +struct packet_lin_dma *lin_dma_pkt,dma_addr_t pkt_dma_addr, +u32 hw_queue_id,u32 size,u64 addr,u32 val) +{ +u32 ctl, pkt_size; +int rc =0 ; + + ctl =FIELD_PREP (GAUDI2_PKT_CTL_OPCODE_MASK ,PACKET_LIN_DMA ); + ctl |=FIELD_PREP (GAUDI2_PKT_LIN_DMA_CTL_MEMSET_MASK ,1 ); + ctl |=FIELD_PREP (GAUDI2_PKT_LIN_DMA_CTL_WRCOMP_MASK ,1 ); + ctl |=FIELD_PREP (GAUDI2_PKT_CTL_EB_MASK ,1 ); + + lin_dma_pkt->ctl =cpu_to_le32 (ctl); + lin_dma_pkt->src_addr =cpu_to_le64 (val); + lin_dma_pkt->dst_addr =cpu_to_le64 (addr); + lin_dma_pkt->tsize =cpu_to_le32 (size); + + pkt_size =sizeof (struct packet_lin_dma ); + + rc =hl_hw_queue_send_cb_no_cmpl (hdev, hw_queue_id, pkt_size, pkt_dma_addr); +if (rc) +dev_err (hdev->dev,"Failed to send lin dma packet to H/W queue %d\n" , + hw_queue_id); + +return rc; +} +static int gaudi2_memset_device_memory (struct hl_device *hdev,u64 addr,u64 size,u64 val) { -struct asic_fixed_properties *prop = &hdev->asic_prop; +u32 edma_queues_id[] = {GAUDI2_QUEUE_ID_DCORE0_EDMA_0_0 , +GAUDI2_QUEUE_ID_DCORE1_EDMA_0_0 , +GAUDI2_QUEUE_ID_DCORE2_EDMA_0_0 , +GAUDI2_QUEUE_ID_DCORE3_EDMA_0_0 }; +u32 chunk_size, dcore, edma_idx, sob_offset, sob_addr, comp_val, + old_mmubp, mmubp, num_of_pkts, busy, pkt_size;u64 comp_addr, cur_addr = addr, end_addr = addr + size; -u32 chunk_size, busy, dcore, edma_idx, sob_offset, sob_addr, comp_val, edma_commit; -u32 old_mmubp, mmubp; -int rc =0 ; +struct asic_fixed_properties *prop = &hdev->asic_prop; +void *lin_dma_pkts_arr; +dma_addr_t pkt_dma_addr; +int rc =0 , dma_num =0 ; + +if (prop->edma_enabled_mask ==0 ) { +dev_info (hdev->dev,"non of the EDMA engines is enabled - skip dram scrubbing\n" ); +return -EIO ; + } sob_offset = hdev->asic_prop.first_available_user_sob[0 ] *4 ; sob_addr = mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + sob_offset; comp_addr =CFG_BASE + sob_addr; comp_val =FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_INC_MASK ,1 ) |FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_VAL_MASK ,1 ); - - edma_commit =FIELD_PREP (ARC_FARM_KDMA_CTX_COMMIT_LIN_MASK ,1 ) | -FIELD_PREP (ARC_FARM_KDMA_CTX_COMMIT_MEM_SET_MASK ,1 ) | -FIELD_PREP (ARC_FARM_KDMA_CTX_COMMIT_WR_COMP_EN_MASK ,1 ); mmubp =FIELD_PREP (ARC_FARM_KDMA_CTX_AXUSER_HB_MMU_BP_WR_MASK ,1 ) |FIELD_PREP (ARC_FARM_KDMA_CTX_AXUSER_HB_MMU_BP_RD_MASK ,1 ); -if (prop->edma_enabled_mask ==0 ) { -dev_info (hdev->dev,"non of the EDMA engines is enabled - skip dram scrubbing\n" ); -return -EIO ; - } +/* Calculate how many lin dma pkts we'll need */ + num_of_pkts =div64_u64 (round_up (size,SZ_2G ),SZ_2G ); + pkt_size =sizeof (struct packet_lin_dma ); + + lin_dma_pkts_arr =hl_asic_dma_alloc_coherent (hdev, pkt_size * num_of_pkts, + &pkt_dma_addr,GFP_KERNEL ); +if (!lin_dma_pkts_arr) +return -ENOMEM ;/* * set mmu bypass for the scrubbing - all ddmas are configured the same so save * only the first one to restore later +* also set the sob addr for all edma cores for completion. +* set QM as trusted to allow it to access physical address with MMU bp. */ old_mmubp =RREG32 (mmDCORE0_EDMA0_CORE_CTX_AXUSER_HB_MMU_BP);for (dcore =0 ; dcore <NUM_OF_DCORES ; dcore++) {
@@ -9211,17 +9251,22 @@ static int gaudi2_memset_device_memory(struct hl_device *hdev, u64 addr, u64 sizWREG32 (mmDCORE0_EDMA0_CORE_CTX_AXUSER_HB_MMU_BP + edma_offset, mmubp); +WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_ADDR_LO + edma_offset, +lower_32_bits (comp_addr)); +WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_ADDR_HI + edma_offset, +upper_32_bits (comp_addr)); +WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_WDATA + edma_offset, + comp_val); +gaudi2_qman_set_test_mode (hdev, + edma_queues_id[dcore] +4 * edma_idx, true); } } -while (cur_addr < end_addr) { -int dma_num =0 ; +WREG32 (sob_addr,0 ); -WREG32 (sob_addr,0 ); +while (cur_addr < end_addr) {for (dcore =0 ; dcore <NUM_OF_DCORES ; dcore++) {for (edma_idx =0 ; edma_idx <NUM_OF_EDMA_PER_DCORE ; edma_idx++) { -u32 edma_offset = dcore *DCORE_OFFSET + - edma_idx *DCORE_EDMA_OFFSET ;u32 edma_bit = dcore *NUM_OF_EDMA_PER_DCORE + edma_idx;if (!(prop->edma_enabled_mask &BIT (edma_bit)))
@@ -9229,41 +9274,26 @@ static int gaudi2_memset_device_memory(struct hl_device *hdev, u64 addr, u64 siz chunk_size =min_t (u64,SZ_2G , end_addr - cur_addr); -WREG32 (mmDCORE0_EDMA0_CORE_CTX_SRC_BASE_LO + edma_offset, -lower_32_bits (val)); -WREG32 (mmDCORE0_EDMA0_CORE_CTX_SRC_BASE_HI + edma_offset, -upper_32_bits (val)); - -WREG32 (mmDCORE0_EDMA0_CORE_CTX_DST_BASE_LO + edma_offset, -lower_32_bits (cur_addr)); -WREG32 (mmDCORE0_EDMA0_CORE_CTX_DST_BASE_HI + edma_offset, -upper_32_bits (cur_addr)); - -WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_ADDR_LO + edma_offset, -lower_32_bits (comp_addr)); -WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_ADDR_HI + edma_offset, -upper_32_bits (comp_addr)); -WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_WDATA + edma_offset, - comp_val); - -WREG32 (mmDCORE0_EDMA0_CORE_CTX_DST_TSIZE_0 + edma_offset, - chunk_size); -WREG32 (mmDCORE0_EDMA0_CORE_CTX_COMMIT + edma_offset, edma_commit); + rc =gaudi2_memset_memory_chunk_using_edma_qm (hdev, + (struct packet_lin_dma *)lin_dma_pkts_arr + dma_num, + pkt_dma_addr + dma_num * pkt_size, + edma_queues_id[dcore] + edma_idx *4 , + chunk_size, cur_addr, val); +if (rc) + goto end; dma_num++; - cur_addr += chunk_size; -if (cur_addr == end_addr) - goto poll; +break ; } } -poll: - rc =hl_poll_timeout (hdev, sob_addr, busy, (busy == dma_num),1000 ,1000000 ); -if (rc) { -dev_err (hdev->dev,"DMA Timeout during HBM scrubbing\n" ); - goto end; - } + } + + rc =hl_poll_timeout (hdev, sob_addr, busy, (busy == dma_num),1000 ,1000000 ); +if (rc) { +dev_err (hdev->dev,"DMA Timeout during HBM scrubbing\n" ); + goto end; } end:for (dcore =0 ; dcore <NUM_OF_DCORES ; dcore++) {
@@ -9275,10 +9305,17 @@ static int gaudi2_memset_device_memory(struct hl_device *hdev, u64 addr, u64 sizcontinue ;WREG32 (mmDCORE0_EDMA0_CORE_CTX_AXUSER_HB_MMU_BP + edma_offset, old_mmubp); +WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_ADDR_LO + edma_offset,0 ); +WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_ADDR_HI + edma_offset,0 ); +WREG32 (mmDCORE0_EDMA0_CORE_CTX_WR_COMP_WDATA + edma_offset,0 ); +gaudi2_qman_set_test_mode (hdev, + edma_queues_id[dcore] +4 * edma_idx, false); } }WREG32 (sob_addr,0 ); +hl_asic_dma_free_coherent (hdev, pkt_size * num_of_pkts, lin_dma_pkts_arr, pkt_dma_addr); +return rc; }
Signature · 1 line
2.25.1
PATCH 20/20 habanalabs: increase the size of busy engines mask
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Increase the size of the busy engines mask in 'struct hl_info_hw_idle', for future ASICs with more than 128 engines.
--- drivers/misc/habanalabs/common/device.c | 9 +++++---- include/uapi/misc/habanalabs.h | 2 +- 2 files changed, 6 insertions(+), 5 deletions(-)
drivers/misc/habanalabs/common/device.c +5 -4
@@ -416,8 +416,9 @@ static void hpriv_release(struct kref *ref) device_is_idle = hdev->asic_funcs->is_device_idle (hdev, idle_mask,HL_BUSY_ENGINES_MASK_EXT_SIZE ,NULL );if (!device_is_idle) { -dev_err (hdev->dev,"device not idle after user context is closed (0x%llx_%llx)\n" , - idle_mask[1 ], idle_mask[0 ]); +dev_err (hdev->dev, +"device not idle after user context is closed (0x%llx_%llx_%llx_%llx)\n" , + idle_mask[3 ], idle_mask[2 ], idle_mask[1 ], idle_mask[0 ]); reset_device = true; }
@@ -1661,8 +1662,8 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)/* If device is not idle fail the reset process */ if (!hdev->asic_funcs->is_device_idle (hdev, idle_mask,HL_BUSY_ENGINES_MASK_EXT_SIZE ,NULL )) { -dev_err (hdev->dev,"device is not idle (mask 0x%llx_%llx) after reset\n" , - idle_mask[1 ], idle_mask[0 ]); +dev_err (hdev->dev,"device is not idle (mask 0x%llx_%llx_%llx_%llx) after reset\n" , + idle_mask[3 ], idle_mask[2 ], idle_mask[1 ], idle_mask[0 ]); rc = -EIO ; goto out_err; }
include/uapi/misc/habanalabs.h +1 -1
@@ -916,7 +916,7 @@ struct hl_info_dram_usage {__u64 ctx_dram_mem; }; -#define HL_BUSY_ENGINES_MASK_EXT_SIZE 2 +#define HL_BUSY_ENGINES_MASK_EXT_SIZE 4struct hl_info_hw_idle {__u32 is_idle;
Signature · 1 line
2.25.1