habanalabs: fixes to the poll-timeout macros
Subsystems (2)
PATCH 1/12 habanalabs: fixes to the poll-timeout macros
From: Ohad Sharabi <[email protected]>
- use conventional internal macro variables (double underscore prefix)
- adjust address casting
- on register poll using ELBI use ELBI read rather than BAR read on error condition
- remove unused macro
--- drivers/misc/habanalabs/common/habanalabs.h | 119 +++++++++++++++----- 1 file changed, 90 insertions(+), 29 deletions(-)
drivers/misc/habanalabs/common/habanalabs.h +90 -29
@@ -2473,9 +2473,11 @@ void hl_wreg(struct hl_device *hdev, u32 reg, u32 val);/* Timeout should be longer when working with simulator but cap the * increased timeout to some maximum */ -#define hl_poll_timeout (hdev, addr, val, cond, sleep_us, timeout_us) \ +#define hl_poll_timeout_common (hdev, addr, val, cond, sleep_us, timeout_us, elbi) \ ({ \ ktime_t __timeout; \ + u32 __elbi_read; \ + int __rc = 0; \ if (hdev->pdev) \ __timeout = ktime_add_us(ktime_get(), timeout_us); \ else \
@@ -2484,19 +2486,103 @@ void hl_wreg(struct hl_device *hdev, u32 reg, u32 val); (u64) HL_SIM_MAX_TIMEOUT_US)); \might_sleep_if (sleep_us); \for (;;) { \ - (val) =RREG32 (addr); \ +if (elbi) { \ + __rc =hl_pci_elbi_read (hdev, addr, &__elbi_read); \ +if (__rc) \ +break ; \ + (val) = __elbi_read; \ + }else {\ + (val) =RREG32 ((u32 )addr); \ + } \if (cond) \break ; \if (timeout_us &&ktime_compare (ktime_get (), __timeout) >0 ) { \ - (val) =RREG32 (addr); \ +if (elbi) { \ + __rc =hl_pci_elbi_read (hdev, addr, &__elbi_read); \ +if (__rc) \ +break ; \ + (val) = __elbi_read; \ + }else {\ + (val) =RREG32 ((u32 )addr); \ + } \break ; \ } \if (sleep_us) \usleep_range ((sleep_us >>2 ) +1 , sleep_us); \ } \ - (cond) ?0 : -ETIMEDOUT ; \ + __rc ? __rc : ((cond) ?0 : -ETIMEDOUT ); \ }) +#define hl_poll_timeout (hdev, addr, val, cond, sleep_us, timeout_us) \ + hl_poll_timeout_common(hdev, addr, val, cond, sleep_us, timeout_us, false) + +#define hl_poll_timeout_elbi (hdev, addr, val, cond, sleep_us, timeout_us) \ + hl_poll_timeout_common(hdev, addr, val, cond, sleep_us, timeout_us, true) + +/* +* poll array of register addresses. +* condition is satisfied if all registers values match the expected value. +* once some register in the array satisfies the condition it will not be polled again, +* this is done both for efficiency and due to some registers are "clear on read". +* TODO: use read from PCI bar in other places in the code (SW-91406) +*/ +#define hl_poll_reg_array_timeout_common (hdev, addr_arr, arr_size, expected_val, sleep_us, \ + timeout_us, elbi) \ +({ \ + ktime_t __timeout; \ + u64 __elem_bitmask; \ + u32 __read_val; \ + u8 __arr_idx; \ + int __rc = 0; \ + \ + if (hdev->pdev) \ + __timeout = ktime_add_us(ktime_get(), timeout_us); \ + else \ + __timeout = ktime_add_us(ktime_get(),\ + min(((u64)timeout_us * 10), \ + (u64) HL_SIM_MAX_TIMEOUT_US)); \ + \ + might_sleep_if(sleep_us); \ + if (arr_size >= 64) \ + __rc = -EINVAL; \ + else \ + __elem_bitmask = BIT_ULL(arr_size) - 1; \ + for (;;) { \ + if (__rc) \ + break; \ + for (__arr_idx = 0; __arr_idx < (arr_size); __arr_idx++) { \ + if (!(__elem_bitmask & BIT_ULL(__arr_idx))) \ + continue; \ + if (elbi) { \ + __rc = hl_pci_elbi_read(hdev, (addr_arr)[__arr_idx], &__read_val); \ + if (__rc) \ + break; \ + } else { \ + __read_val = RREG32((u32)(addr_arr)[__arr_idx]); \ + } \ + if (__read_val == (expected_val)) \ + __elem_bitmask &= ~BIT_ULL(__arr_idx); \ + } \ + if (__rc || (__elem_bitmask == 0)) \ + break; \ + if (timeout_us && ktime_compare(ktime_get(), __timeout) > 0) \ + break; \ + if (sleep_us) \ + usleep_range((sleep_us >> 2) + 1, sleep_us); \ + } \ + __rc ? __rc : ((__elem_bitmask == 0) ? 0 : -ETIMEDOUT); \ +}) + +#define hl_poll_reg_array_timeout (hdev, addr_arr, arr_size, expected_val, sleep_us, \ + timeout_us) \ + hl_poll_reg_array_timeout_common(hdev, addr_arr, arr_size, expected_val, sleep_us, \ + timeout_us, false) + +#define hl_poll_reg_array_timeout_elbi (hdev, addr_arr, arr_size, expected_val, sleep_us, \ + timeout_us) \ + hl_poll_reg_array_timeout_common(hdev, addr_arr, arr_size, expected_val, sleep_us, \ + timeout_us, true) +/* * address in this macro points always to a memory location in the * host's (server's) memory. That location is updated asynchronously
@@ -2540,31 +2626,6 @@ void hl_wreg(struct hl_device *hdev, u32 reg, u32 val); (cond) ?0 : -ETIMEDOUT ; \ }) -#define hl_poll_timeout_device_memory (hdev, addr, val, cond, sleep_us, \ - timeout_us) \ -({ \ - ktime_t __timeout; \ - if (hdev->pdev) \ - __timeout = ktime_add_us(ktime_get(), timeout_us); \ - else \ - __timeout = ktime_add_us(ktime_get(),\ - min((u64)(timeout_us * 10), \ - (u64) HL_SIM_MAX_TIMEOUT_US)); \ - might_sleep_if(sleep_us); \ - for (;;) { \ - (val) = readl(addr); \ - if (cond) \ - break; \ - if (timeout_us && ktime_compare(ktime_get(), __timeout) > 0) { \ - (val) = readl(addr); \ - break; \ - } \ - if (sleep_us) \ - usleep_range((sleep_us >> 2) + 1, sleep_us); \ - } \ - (cond) ? 0 : -ETIMEDOUT; \ -}) -#define HL_USR_MAPPED_BLK_INIT (blk, base, sz) \ ({ \ struct user_mapped_block *p = blk; \
Signature · 1 line
2.25.1
PATCH 2/12 habanalabs: add a value field to hl_fw_send_pci_access_msg()
From: Tomer Tayar <[email protected]>
For gaudi2 we need to send a value to F/W as part of the PCI_ACCESS packet.
As a preparation, modify hl_fw_send_pci_access_msg() to have a 'value' field.
--- drivers/misc/habanalabs/common/device.c | 5 ++--- drivers/misc/habanalabs/common/firmware_if.c | 6 +++--- drivers/misc/habanalabs/common/habanalabs.h | 2 +- drivers/misc/habanalabs/gaudi/gaudi.c | 6 +++--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 6 +++--- drivers/misc/habanalabs/goya/goya.c | 4 ++-- 6 files changed, 14 insertions(+), 15 deletions(-)
drivers/misc/habanalabs/common/device.c +2 -3
@@ -1259,8 +1259,7 @@ static void handle_reset_trigger(struct hl_device *hdev, u32 flags)* of heartbeat, the device CPU is marked as disable * so this message won't be sent */ -if (hl_fw_send_pci_access_msg (hdev, -CPUCP_PACKET_DISABLE_PCI_ACCESS )) +if (hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 ))dev_warn (hdev->dev,"Failed to disable PCI access by F/W\n" ); }
@@ -2054,7 +2053,7 @@ void hl_device_fini(struct hl_device *hdev)* 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 ); +hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 );/* Mark device as disabled */ hdev->disabled = true;
drivers/misc/habanalabs/common/firmware_if.c +3 -3
@@ -242,14 +242,14 @@ int hl_fw_load_fw_to_device(struct hl_device *hdev, const char *fw_name,return rc; } -int hl_fw_send_pci_access_msg (struct hl_device *hdev,u32 opcode) +int hl_fw_send_pci_access_msg (struct hl_device *hdev,u32 opcode,u64 value) {struct cpucp_packet pkt = {}; pkt.ctl =cpu_to_le32 (opcode <<CPUCP_PKT_CTL_OPCODE_SHIFT ); + pkt.value =cpu_to_le64 (value); -return hdev->asic_funcs->send_cpu_message (hdev, (u32 *) &pkt, -sizeof (pkt),0 ,NULL ); +return hdev->asic_funcs->send_cpu_message (hdev, (u32 *) &pkt,sizeof (pkt),0 ,NULL ); } int hl_fw_send_cpu_message(struct hl_device *hdev, u32 hw_queue_id, u32 *msg,
drivers/misc/habanalabs/common/habanalabs.h +1 -1
@@ -3631,7 +3631,7 @@ 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); -int hl_fw_send_pci_access_msg (struct hl_device *hdev,u32 opcode); +int hl_fw_send_pci_access_msg (struct hl_device *hdev,u32 opcode,u64 value);int hl_fw_send_cpu_message (struct hl_device *hdev,u32 hw_queue_id,u32 *msg,u16 len,u32 timeout,u64 *result);int hl_fw_unmask_irq (struct hl_device *hdev,u16 event_type);
drivers/misc/habanalabs/gaudi/gaudi.c +3 -3
@@ -1632,7 +1632,7 @@ static int gaudi_late_init(struct hl_device *hdev) gaudi->hw_cap_initialized &= ~(HW_CAP_NIC0 |HW_CAP_NIC1 ); } - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ,0x0 );if (rc) {dev_err (hdev->dev,"Failed to enable PCI access from CPU\n" );return rc;
@@ -1677,7 +1677,7 @@ static int gaudi_late_init(struct hl_device *hdev)return 0 ; disable_pci_access: -hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ); +hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 );return rc; }
@@ -4236,7 +4236,7 @@ static int gaudi_suspend(struct hl_device *hdev) {int rc; - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 );if (rc)dev_err (hdev->dev,"Failed to disable PCI access from CPU\n" );
drivers/misc/habanalabs/gaudi2/gaudi2.c +3 -3
@@ -2675,7 +2675,7 @@ static int gaudi2_late_init(struct hl_device *hdev) {int rc; - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ,0x0 );if (rc) {dev_err (hdev->dev,"Failed to enable PCI access from CPU\n" );return rc;
@@ -2694,7 +2694,7 @@ static int gaudi2_late_init(struct hl_device *hdev)return 0 ; disable_pci_access: -hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ); +hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 );return rc; }
@@ -5323,7 +5323,7 @@ static int gaudi2_suspend(struct hl_device *hdev) {int rc; - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 );if (rc)dev_err (hdev->dev,"Failed to disable PCI access from CPU\n" );
drivers/misc/habanalabs/goya/goya.c +2 -2
@@ -888,7 +888,7 @@ int goya_late_init(struct hl_device *hdev)*/ WREG32 (mmMMU_LOG2_DDR_SIZE,ilog2 (prop->dram_size)); - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ,0x0 );if (rc) {dev_err (hdev->dev,"Failed to enable PCI access from CPU %d\n" , rc);
@@ -2878,7 +2878,7 @@ int goya_suspend(struct hl_device *hdev) {int rc; - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_DISABLE_PCI_ACCESS ,0x0 );if (rc)dev_err (hdev->dev,"Failed to disable PCI access from CPU\n" );
Signature · 1 line
2.25.1
PATCH 3/12 habanalabs/gaudi2: configure virtual MSI-X doorbell interface
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Due to a watchdog timer in the LBW path, writes to the MSI-X doorbell can return sporadic error responses.
To work-around this issue, a virtual MSI-X doorbell on the HBW path is configured, using the MSI-X AXI slave interface in the PCIe controller. Upon an access to a configured HBW host address, the controller will generate MSI-X interrupt instead of treating the access as regular host memory access.
This patch allocates the dedicate host memory page, and communicate the address to F/W, so it will configure the relevant address match registers in the controller, and will use this address to generate MSI-X interrupts for F/W events.
Following patches will handle other initiators in the device, to move them to use the virtual MSI-X doorbell.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 23 ++++++++++++++++--- drivers/misc/habanalabs/gaudi2/gaudi2P.h | 5 ++++ drivers/misc/habanalabs/gaudi2/gaudi2_masks.h | 3 +++ 3 files changed, 28 insertions(+), 3 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +20 -3
@@ -2673,9 +2673,11 @@ static void gaudi2_scrub_arcs_dccm(struct hl_device *hdev)static int gaudi2_late_init (struct hl_device *hdev) { +struct gaudi2_device *gaudi2 = hdev->asic_specific;int rc; - rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS ,0x0 ); + rc =hl_fw_send_pci_access_msg (hdev,CPUCP_PACKET_ENABLE_PCI_ACCESS , + gaudi2->virt_msix_db_dma_addr);if (rc) {dev_err (hdev->dev,"Failed to enable PCI access from CPU\n" );return rc;
@@ -2922,6 +2924,7 @@ static inline int gaudi2_get_non_zero_random_int(void)static int gaudi2_sw_init (struct hl_device *hdev) { +struct asic_fixed_properties *prop = &hdev->asic_prop;struct gaudi2_device *gaudi2;int i, rc;
@@ -2982,6 +2985,14 @@ static int gaudi2_sw_init(struct hl_device *hdev) goto free_cpu_accessible_dma_pool; } + gaudi2->virt_msix_db_cpu_addr =hl_cpu_accessible_dma_pool_alloc (hdev, prop->pmmu.page_size, + &gaudi2->virt_msix_db_dma_addr); +if (!gaudi2->virt_msix_db_cpu_addr) { +dev_err (hdev->dev,"Failed to allocate DMA memory for virtual MSI-X doorbell\n" ); + rc = -ENOMEM ; + goto free_cpu_accessible_dma_pool; + } +spin_lock_init (&gaudi2->hw_queues_lock);spin_lock_init (&gaudi2->kdma_lock);
@@ -2990,7 +3001,7 @@ static int gaudi2_sw_init(struct hl_device *hdev) GFP_KERNEL | __GFP_ZERO);if (!gaudi2->scratchpad_kernel_address) { rc = -ENOMEM ; - goto free_cpu_accessible_dma_pool; + goto free_virt_msix_db_mem; }gaudi2_user_mapped_blocks_init (hdev);
@@ -2999,15 +3010,18 @@ static int gaudi2_sw_init(struct hl_device *hdev)gaudi2_user_interrupt_setup (hdev); hdev->supports_coresight = true; - hdev->asic_prop.supports_soft_reset = true; hdev->supports_sync_stream = true; hdev->supports_cb_mapping = true; hdev->supports_wait_for_multi_cs = false; + prop->supports_soft_reset = true; + hdev->asic_funcs->set_pci_memory_regions (hdev);return 0 ; +free_virt_msix_db_mem: +hl_cpu_accessible_dma_pool_free (hdev, prop->pmmu.page_size, gaudi2->virt_msix_db_cpu_addr); free_cpu_accessible_dma_pool:gen_pool_destroy (hdev->cpu_accessible_dma_pool); free_cpu_dma_mem:
@@ -3022,8 +3036,11 @@ static int gaudi2_sw_init(struct hl_device *hdev)static int gaudi2_sw_fini (struct hl_device *hdev) { +struct asic_fixed_properties *prop = &hdev->asic_prop;struct gaudi2_device *gaudi2 = hdev->asic_specific; +hl_cpu_accessible_dma_pool_free (hdev, prop->pmmu.page_size, gaudi2->virt_msix_db_cpu_addr); +gen_pool_destroy (hdev->cpu_accessible_dma_pool); hl_asic_dma_free_coherent(hdev, HL_CPU_ACCESSIBLE_MEM_SIZE, hdev->cpu_accessible_dma_mem,
drivers/misc/habanalabs/gaudi2/gaudi2P.h +5 -0
@@ -439,6 +439,8 @@ struct dup_block_ctx {* currently used for HBW QMAN writes which is * redundant. * @scratchpad_bus_address: scratchpad bus address +* @virt_msix_db_cpu_addr: host memory page for the virtual MSI-X doorbell. +* @virt_msix_db_dma_addr: bus address of the page for the virtual MSI-X doorbell. * @dram_bar_cur_addr: current address of DRAM PCI bar. * @hw_cap_initialized: This field contains a bit per H/W engine. When that * engine is initialized, that bit is set by the driver to
@@ -499,6 +501,9 @@ struct gaudi2_device {void *scratchpad_kernel_address;dma_addr_t scratchpad_bus_address; +void *virt_msix_db_cpu_addr; +dma_addr_t virt_msix_db_dma_addr; +u64 dram_bar_cur_addr;u64 hw_cap_initialized;u64 active_hw_arc;
drivers/misc/habanalabs/gaudi2/gaudi2_masks.h +3 -0
@@ -132,4 +132,7 @@#define ROT_MSS_HALT_RSB_MASK BIT(1)#define ROT_MSS_HALT_MRSB_MASK BIT(2) +#define PCIE_DBI_MSIX_ADDRESS_MATCH_LOW_OFF_MSIX_ADDRESS_MATCH_EN_SHIFT 0 +#define PCIE_DBI_MSIX_ADDRESS_MATCH_LOW_OFF_MSIX_ADDRESS_MATCH_EN_MASK 0x1 +#endif /* GAUDI2_MASKS_H_ */
Signature · 1 line
2.25.1
PATCH 4/12 habanalabs/gaudi2: replace defines for reserved sob/mob with enums
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Following patches are going to add more reserved sync objects and monitors.
To make the counting of these reserved resources simpler, replace the existing RESERVED_* defines with enumerations.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 48 ++++++++++--------- drivers/misc/habanalabs/gaudi2/gaudi2P.h | 26 ++++++---- .../misc/habanalabs/gaudi2/gaudi2_security.c | 4 +- 3 files changed, 46 insertions(+), 32 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +26 -22
@@ -1867,8 +1867,8 @@ static int gaudi2_set_fixed_properties(struct hl_device *hdev) prop->user_dec_intr_count =NUMBER_OF_DEC ; prop->user_interrupt_count =GAUDI2_IRQ_NUM_USER_LAST -GAUDI2_IRQ_NUM_USER_FIRST +1 ; prop->completion_mode =HL_COMPLETION_MODE_CS ; - prop->sync_stream_first_sob =GAUDI2_RESERVED_SOBS ; - prop->sync_stream_first_mon =GAUDI2_RESERVED_MONITORS ; + prop->sync_stream_first_sob =GAUDI2_RESERVED_SOB_NUMBER ; + prop->sync_stream_first_mon =GAUDI2_RESERVED_MON_NUMBER ; prop->sram_base_address =SRAM_BASE_ADDR ; prop->sram_size =SRAM_SIZE ;
@@ -1988,10 +1988,10 @@ static int gaudi2_set_fixed_properties(struct hl_device *hdev) prop->mme_master_slave_mode =1 ; - prop->first_available_user_sob[0 ] =GAUDI2_RESERVED_SOBS + + prop->first_available_user_sob[0 ] =GAUDI2_RESERVED_SOB_NUMBER + (num_sync_stream_queues *HL_RSVD_SOBS ); - prop->first_available_user_mon[0 ] =GAUDI2_RESERVED_MONITORS + + prop->first_available_user_mon[0 ] =GAUDI2_RESERVED_MON_NUMBER + (num_sync_stream_queues *HL_RSVD_MONS ); prop->first_available_user_interrupt =GAUDI2_IRQ_NUM_USER_FIRST ;
@@ -3533,7 +3533,7 @@ static int gaudi2_enable_msix(struct hl_device *hdev) } irq =pci_irq_vector (hdev->pdev,GAUDI2_IRQ_NUM_COMPLETION ); - cq = &hdev->completion_queue[GAUDI2_RESERVED_CQ_COMPLETION ]; + cq = &hdev->completion_queue[GAUDI2_RESERVED_CQ_CS_COMPLETION ]; rc =request_irq (irq, hl_irq_handler_cq,0 ,gaudi2_irq_name (GAUDI2_IRQ_NUM_COMPLETION ), cq);if (rc) {dev_err (hdev->dev,"Failed to request IRQ %d" , irq);
@@ -3643,7 +3643,7 @@ static void gaudi2_disable_msix(struct hl_device *hdev) } irq =pci_irq_vector (hdev->pdev,GAUDI2_IRQ_NUM_COMPLETION ); - cq = &hdev->completion_queue[GAUDI2_RESERVED_CQ_COMPLETION ]; + cq = &hdev->completion_queue[GAUDI2_RESERVED_CQ_CS_COMPLETION ];free_irq (irq, cq);pci_free_irq_vectors (hdev->pdev);
@@ -4139,7 +4139,7 @@ static void gaudi2_init_qman(struct hl_device *hdev, u32 reg_base,u32 pq_id;for (pq_id =0 ; pq_id <NUM_OF_PQ_PER_QMAN ; pq_id++) - hdev->kernel_queues[queue_id_base + pq_id].cq_id =GAUDI2_RESERVED_CQ_COMPLETION ; + hdev->kernel_queues[queue_id_base + pq_id].cq_id =GAUDI2_RESERVED_CQ_CS_COMPLETION ;gaudi2_init_qman_pq (hdev, reg_base, queue_id_base);gaudi2_init_qman_cp (hdev, reg_base);
@@ -5762,31 +5762,33 @@ static void gaudi2_kdma_set_mmbp_asid(struct hl_device *hdev,WREG32 (mmARC_FARM_KDMA_CTX_AXUSER_HB_MMU_BP, rw_mmu_bp); } -static void gaudi2_arm_cq_monitor (struct hl_device *hdev,u32 index,u32 cq_id, +static void gaudi2_arm_cq_monitor (struct hl_device *hdev,u32 sob_id,u32 mon_id,u32 cq_id,u32 mon_payload,u32 sync_value) { -u32 sync_group_id, mode, mon_arm; -int offset = index *4 ; +u32 sob_offset, mon_offset, sync_group_id, mode, mon_arm;u8 mask; + sob_offset = sob_id *4 ; + mon_offset = mon_id *4 ; +/* Reset the SOB value */ -WREG32 (mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + offset,0 ); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + sob_offset,0 );/* Configure this address with CQ_ID 0 because CQ_EN is set */ -WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRL_0 + offset, cq_id); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRL_0 + mon_offset, cq_id);/* Configure this address with CS index because CQ_EN is set */ -WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_DATA_0 + offset, mon_payload); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_DATA_0 + mon_offset, mon_payload); - sync_group_id = index /8 ; - mask = ~(1 << (index &0x7 )); + sync_group_id = sob_id /8 ; + mask = ~(1 << (sob_id &0x7 )); mode =1 ;/* comparison mode is "equal to" */ mon_arm =FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_SOD_MASK , sync_value); mon_arm |=FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_SOP_MASK , mode); mon_arm |=FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_MASK_MASK , mask); mon_arm |=FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_SID_MASK , sync_group_id); -WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_ARM_0 + offset, mon_arm); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_ARM_0 + mon_offset, mon_arm); }/* This is an internal helper function used by gaudi2_send_job_to_kdma only */
@@ -5800,11 +5802,12 @@ static int gaudi2_send_job_to_kdma(struct hl_device *hdev,u64 comp_addr;int rc; -gaudi2_arm_cq_monitor (hdev,GAUDI2_RESERVED_SOB_KDMA_COMP , +gaudi2_arm_cq_monitor (hdev,GAUDI2_RESERVED_SOB_KDMA_COMPLETION , +GAUDI2_RESERVED_MON_KDMA_COMPLETION ,GAUDI2_RESERVED_CQ_KDMA_COMPLETION ,1 ,1 ); comp_addr =CFG_BASE + mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + - (GAUDI2_RESERVED_SOB_KDMA_COMP *sizeof (u32)); + (GAUDI2_RESERVED_SOB_KDMA_COMPLETION *sizeof (u32)); comp_val =FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_INC_MASK ,1 ) |FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_VAL_MASK ,1 );
@@ -9125,24 +9128,25 @@ static int gaudi2_pre_schedule_cs(struct hl_cs *cs) {struct hl_device *hdev = cs->ctx->hdev;int index = cs->sequence & (hdev->asic_prop.max_pending_cs -1 ); -u32 mon_payload; +u32 mon_payload, sob_id, mon_id;if (!cs_needs_completion (cs))return 0 ;/* -* First 1024 SOB/MON are reserved for driver for QMAN auto completion +* First 64 SOB/MON are reserved for driver for QMAN auto completion * mechanism. Each SOB/MON pair are used for a pending CS with the same * cyclic index. The SOB value is increased when each of the CS jobs is * completed. When the SOB reaches the number of CS jobs, the monitor * generates MSI-X interrupt. */ + sob_id = mon_id = index; mon_payload = (1 <<CQ_ENTRY_SHADOW_INDEX_VALID_SHIFT ) | (1 <<CQ_ENTRY_READY_SHIFT ) | index; -gaudi2_arm_cq_monitor (hdev, index,GAUDI2_RESERVED_CQ_COMPLETION , - mon_payload, cs->jobs_cnt); +gaudi2_arm_cq_monitor (hdev, sob_id, mon_id,GAUDI2_RESERVED_CQ_CS_COMPLETION , mon_payload, + cs->jobs_cnt);return 0 ; }
drivers/misc/habanalabs/gaudi2/gaudi2P.h +18 -8
@@ -80,12 +80,6 @@#define GAUDI2_MAX_PENDING_CS 64 -/* Sob/Mon per CS + Sob/Mon for KDMA completion */ -#define GAUDI2_RESERVED_SOBS (GAUDI2_MAX_PENDING_CS + 1) -#define GAUDI2_RESERVED_MONITORS (GAUDI2_MAX_PENDING_CS + 1) -#define GAUDI2_RESERVED_SOB_KDMA_COMP (GAUDI2_MAX_PENDING_CS) -#define GAUDI2_RESERVED_MON_KDMA_COMP (GAUDI2_MAX_PENDING_CS) -#if !IS_MAX_PENDING_CS_VALID (GAUDI2_MAX_PENDING_CS )#error "GAUDI2_MAX_PENDING_CS must be power of 2 and greater than 1"#endif
@@ -249,10 +243,26 @@#define GAUDI2_ARC_PCI_MSB_ADDR (addr) (((addr) & GENMASK_ULL(49, 28)) >> 28) +enum gaudi2_reserved_sob_id { +GAUDI2_RESERVED_SOB_CS_COMPLETION_FIRST , +GAUDI2_RESERVED_SOB_CS_COMPLETION_LAST = +GAUDI2_RESERVED_SOB_CS_COMPLETION_FIRST +GAUDI2_MAX_PENDING_CS -1 , +GAUDI2_RESERVED_SOB_KDMA_COMPLETION , +GAUDI2_RESERVED_SOB_NUMBER +}; + +enum gaudi2_reserved_mon_id { +GAUDI2_RESERVED_MON_CS_COMPLETION_FIRST , +GAUDI2_RESERVED_MON_CS_COMPLETION_LAST = +GAUDI2_RESERVED_MON_CS_COMPLETION_FIRST +GAUDI2_MAX_PENDING_CS -1 , +GAUDI2_RESERVED_MON_KDMA_COMPLETION , +GAUDI2_RESERVED_MON_NUMBER +}; +enum gaudi2_reserved_cq_id { -GAUDI2_RESERVED_CQ_COMPLETION , +GAUDI2_RESERVED_CQ_CS_COMPLETION ,GAUDI2_RESERVED_CQ_KDMA_COMPLETION , -GAUDI2_RESERVED_CQ_NUMBER , +GAUDI2_RESERVED_CQ_NUMBER };/*
drivers/misc/habanalabs/gaudi2/gaudi2_security.c +2 -2
@@ -2703,8 +2703,8 @@ static int gaudi2_init_pb_sm_objs(struct hl_device *hdev)if (!sec_array)return -ENOMEM ; - first_sob =GAUDI2_RESERVED_SOBS ; - first_mon =GAUDI2_RESERVED_MONITORS ; + first_sob =GAUDI2_RESERVED_SOB_NUMBER ; + first_mon =GAUDI2_RESERVED_MON_NUMBER ;/* 8192 SOB_OBJs skipping first GAUDI2_MAX_PENDING_CS of them */ for (j = i = first_sob ; i <DCORE_NUM_OF_SOB ; i++, j++)
Signature · 1 line
2.25.1
PATCH 5/12 habanalabs/gaudi2: modify CS completion CQ to use virtual MSI-X doorbell
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Modify the CQ which is used for CS completion, to use the virtual MSI-X doorbell.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 19 ++++++++++++++++--- 1 file changed, 16 insertions(+), 3 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +16 -3
@@ -4254,7 +4254,7 @@ static void gaudi2_init_edma(struct hl_device *hdev)static void gaudi2_init_sm (struct hl_device *hdev) { -u64 msix_db_reg =CFG_BASE + mmPCIE_DBI_MSIX_DOORBELL_OFF; +struct gaudi2_device *gaudi2 = hdev->asic_specific;u64 cq_address;u32 reg_val;int i;
@@ -4272,8 +4272,21 @@ static void gaudi2_init_sm(struct hl_device *hdev)/* Init CQ0 DB */ /* Configure the monitor to trigger MSI-X interrupt */ -WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_ADDR_L_0,lower_32_bits (msix_db_reg)); -WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_ADDR_H_0,upper_32_bits (msix_db_reg)); +/* TODO: +* Remove the if statement when virtual MSI-X doorbell is supported in simulator (SW-93022) +* and in F/W (SW-93024). +*/ +if (!hdev->pdev || hdev->asic_prop.fw_security_enabled) { +u64 msix_db_reg =CFG_BASE + mmPCIE_DBI_MSIX_DOORBELL_OFF; + +WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_ADDR_L_0,lower_32_bits (msix_db_reg)); +WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_ADDR_H_0,upper_32_bits (msix_db_reg)); + }else { +WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_ADDR_L_0, +lower_32_bits (gaudi2->virt_msix_db_dma_addr)); +WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_ADDR_H_0, +upper_32_bits (gaudi2->virt_msix_db_dma_addr)); + }WREG32 (mmDCORE0_SYNC_MNGR_GLBL_LBW_DATA_0,GAUDI2_IRQ_NUM_COMPLETION );for (i =0 ; i <GAUDI2_RESERVED_CQ_NUMBER ; i++) {
Signature · 1 line
2.25.1
PATCH 6/12 habanalabs/gaudi2: modify decoder to use virtual MSI-X doorbell
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Modify the decoder wrapper blocks to generate interrupts using the virtual MSI-X doorbell.
As a decoder wrapper block cannot write directly to HBW upon completion, it writes instead to SOB which is monitored by a master monitor. When resolved, this monitor will be the one to actually write to the virtual MSI-X doorbell.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 141 +++++++++++++++++- drivers/misc/habanalabs/gaudi2/gaudi2P.h | 15 ++ drivers/misc/habanalabs/gaudi2/gaudi2_masks.h | 3 + .../include/gaudi2/asic_reg/gaudi2_regs.h | 6 + 4 files changed, 160 insertions(+), 5 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +136 -5
@@ -4252,6 +4252,109 @@ static void gaudi2_init_edma(struct hl_device *hdev) } } +/* +* gaudi2_arm_monitors_for_virt_msix_db() - Arm monitors for writing to the virtual MSI-X doorbell. +* @hdev: pointer to habanalabs device structure. +* @sob_id: sync object ID. +* @first_mon_id: ID of first monitor out of 3 consecutive monitors. +* @interrupt_id: interrupt ID. +* +* Some initiators cannot have HBW address in their completion address registers, and thus cannot +* write directly to the HBW host memory of the virtual MSI-X doorbell. +* Instead, they are configured to LBW write to a sync object, and a monitor will do the HBW write. +* +* The mechanism in the sync manager block is composed of a master monitor with 3 messages. +* In addition to the HBW write, the other 2 messages are for preparing the monitor to next +* completion, by decrementing the sync object value and re-arming the monitor. +*/ +static void gaudi2_arm_monitors_for_virt_msix_db (struct hl_device *hdev,u32 sob_id, +u32 first_mon_id,u32 interrupt_id) +{ +u32 sob_offset, first_mon_offset, mon_offset, payload, sob_group, mode, arm, config; +struct gaudi2_device *gaudi2 = hdev->asic_specific; +u64 addr; +u8 mask; + +/* Reset the SOB value */ + sob_offset = sob_id *sizeof (u32); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + sob_offset,0 ); + +/* Configure 3 monitors: +* 1. Write interrupt ID to the virtual MSI-X doorbell (master monitor) +* 2. Decrement SOB value by 1. +* 3. Re-arm the master monitor. +*/ + + first_mon_offset = first_mon_id *sizeof (u32); + +/* 2nd monitor: Decrement SOB value by 1 */ + mon_offset = first_mon_offset +sizeof (u32); + + addr =CFG_BASE + mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + sob_offset; +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRL_0 + mon_offset,lower_32_bits (addr)); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRH_0 + mon_offset,upper_32_bits (addr)); + + payload =FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_VAL_MASK ,0x7FFF ) |/* "-1" */ +FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_SIGN_MASK ,1 ) | +FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_INC_MASK ,1 ); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_DATA_0 + mon_offset, payload); + +/* 3rd monitor: Re-arm the master monitor */ + mon_offset = first_mon_offset +2 *sizeof (u32); + + addr =CFG_BASE + mmDCORE0_SYNC_MNGR_OBJS_MON_ARM_0 + first_mon_offset; +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRL_0 + mon_offset,lower_32_bits (addr)); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRH_0 + mon_offset,upper_32_bits (addr)); + + sob_group = sob_id /8 ; + mask = ~BIT (sob_id &0x7 ); + mode =0 ;/* comparison mode is "greater than or equal to" */ + arm =FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_SID_MASK , sob_group) | +FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_MASK_MASK , mask) | +FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_SOP_MASK , mode) | +FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_ARM_SOD_MASK ,1 ); + + payload = arm; +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_DATA_0 + mon_offset, payload); + +/* 1st monitor (master): Write interrupt ID to the virtual MSI-X doorbell */ + mon_offset = first_mon_offset; + + config =FIELD_PREP (DCORE0_SYNC_MNGR_OBJS_MON_CONFIG_WR_NUM_MASK ,2 );/* "2": 3 writes */ +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_CONFIG_0 + mon_offset, config); + + addr = gaudi2->virt_msix_db_dma_addr; +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRL_0 + mon_offset,lower_32_bits (addr)); +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_ADDRH_0 + mon_offset,upper_32_bits (addr)); + + payload = interrupt_id; +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_PAY_DATA_0 + mon_offset, payload); + +WREG32 (mmDCORE0_SYNC_MNGR_OBJS_MON_ARM_0 + mon_offset, arm); +} + +static void gaudi2_prepare_sm_for_virt_msix_db (struct hl_device *hdev) +{ +u32 decoder_id, sob_id, first_mon_id, interrupt_id; +struct asic_fixed_properties *prop = &hdev->asic_prop; + +/* Decoder normal/abnormal interrupts */ +for (decoder_id =0 ; decoder_id <NUMBER_OF_DEC ; ++decoder_id) { +if (!(prop->decoder_enabled_mask &BIT (decoder_id))) +continue ; + + sob_id =GAUDI2_RESERVED_SOB_DEC_NRM_FIRST + decoder_id; + first_mon_id =GAUDI2_RESERVED_MON_DEC_NRM_FIRST +3 * decoder_id; + interrupt_id =GAUDI2_IRQ_NUM_DCORE0_DEC0_NRM +2 * decoder_id; +gaudi2_arm_monitors_for_virt_msix_db (hdev, sob_id, first_mon_id, interrupt_id); + + sob_id =GAUDI2_RESERVED_SOB_DEC_ABNRM_FIRST + decoder_id; + first_mon_id =GAUDI2_RESERVED_MON_DEC_ABNRM_FIRST +3 * decoder_id; + interrupt_id +=1 ; +gaudi2_arm_monitors_for_virt_msix_db (hdev, sob_id, first_mon_id, interrupt_id); + } +} +static void gaudi2_init_sm (struct hl_device *hdev) {struct gaudi2_device *gaudi2 = hdev->asic_specific;
@@ -4304,6 +4407,9 @@ static void gaudi2_init_sm(struct hl_device *hdev)/* Configure kernel ASID and MMU BP*/ WREG32 (mmDCORE0_SYNC_MNGR_GLBL_ASID_SEC,0x10000 );WREG32 (mmDCORE0_SYNC_MNGR_GLBL_ASID_NONE_SEC_PRIV,0 ); + +/* Initialize sync objects and monitors which are used for the virtual MSI-X doorbell */ +gaudi2_prepare_sm_for_virt_msix_db (hdev); }static void gaudi2_init_mme_acc (struct hl_device *hdev,u32 reg_base)
@@ -4452,10 +4558,35 @@ static void gaudi2_init_rotator(struct hl_device *hdev) } } -static void gaudi2_init_vdec_brdg_ctrl (struct hl_device *hdev,u64 base_addr,u32 msix_id) +static void gaudi2_init_vdec_brdg_ctrl (struct hl_device *hdev,u64 base_addr,u32 decoder_id) { -WREG32 (base_addr +BRDG_CTRL_NRM_MSIX_LBW_WDATA , msix_id); -WREG32 (base_addr +BRDG_CTRL_ABNRM_MSIX_LBW_WDATA , msix_id +1 ); +u32 sob_id; + +/* TODO: +* Remove when virtual MSI-X doorbell is supported in simulator (SW-93022) and in F/W +* (SW-93024). +*/ +if (!hdev->pdev || hdev->asic_prop.fw_security_enabled) { +u32 interrupt_id =GAUDI2_IRQ_NUM_DCORE0_DEC0_NRM +2 * decoder_id; + +WREG32 (base_addr +BRDG_CTRL_NRM_MSIX_LBW_AWADDR , mmPCIE_DBI_MSIX_DOORBELL_OFF); +WREG32 (base_addr +BRDG_CTRL_NRM_MSIX_LBW_WDATA , interrupt_id); +WREG32 (base_addr +BRDG_CTRL_ABNRM_MSIX_LBW_AWADDR , mmPCIE_DBI_MSIX_DOORBELL_OFF); +WREG32 (base_addr +BRDG_CTRL_ABNRM_MSIX_LBW_WDATA , interrupt_id +1 ); +return ; + } + +/* VCMD normal interrupt */ + sob_id =GAUDI2_RESERVED_SOB_DEC_NRM_FIRST + decoder_id; +WREG32 (base_addr +BRDG_CTRL_NRM_MSIX_LBW_AWADDR , + mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + sob_id *sizeof (u32)); +WREG32 (base_addr +BRDG_CTRL_NRM_MSIX_LBW_WDATA ,GAUDI2_SOB_INCREMENT_BY_ONE ); + +/* VCMD abnormal interrupt */ + sob_id =GAUDI2_RESERVED_SOB_DEC_ABNRM_FIRST + decoder_id; +WREG32 (base_addr +BRDG_CTRL_ABNRM_MSIX_LBW_AWADDR , + mmDCORE0_SYNC_MNGR_OBJS_SOB_OBJ_0 + sob_id *sizeof (u32)); +WREG32 (base_addr +BRDG_CTRL_ABNRM_MSIX_LBW_WDATA ,GAUDI2_SOB_INCREMENT_BY_ONE ); }static void gaudi2_init_dec (struct hl_device *hdev)
@@ -4485,7 +4616,7 @@ static void gaudi2_init_dec(struct hl_device *hdev) msix_id =GAUDI2_IRQ_NUM_DCORE0_DEC0_NRM + (dcore_id *NUM_OF_DEC_PER_DCORE + dec_id) *2 ; -gaudi2_init_vdec_brdg_ctrl (hdev, base_addr, msix_id); +gaudi2_init_vdec_brdg_ctrl (hdev, base_addr, dec_bit); gaudi2->dec_hw_cap_initialized |=BIT_ULL (HW_CAP_DEC_SHIFT + dec_bit); }
@@ -4500,7 +4631,7 @@ static void gaudi2_init_dec(struct hl_device *hdev) msix_id =GAUDI2_IRQ_NUM_SHARED_DEC0_NRM + (dec_id *2 ); -gaudi2_init_vdec_brdg_ctrl (hdev, base_addr, msix_id); +gaudi2_init_vdec_brdg_ctrl (hdev, base_addr, dec_bit); gaudi2->dec_hw_cap_initialized |=BIT_ULL (HW_CAP_DEC_SHIFT + dec_bit); }
drivers/misc/habanalabs/gaudi2/gaudi2P.h +15 -0
@@ -243,11 +243,20 @@#define GAUDI2_ARC_PCI_MSB_ADDR (addr) (((addr) & GENMASK_ULL(49, 28)) >> 28) +#define GAUDI2_SOB_INCREMENT_BY_ONE (FIELD_PREP(DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_VAL_MASK, 1) | \ + FIELD_PREP(DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_INC_MASK, 1)) +enum gaudi2_reserved_sob_id {GAUDI2_RESERVED_SOB_CS_COMPLETION_FIRST ,GAUDI2_RESERVED_SOB_CS_COMPLETION_LAST =GAUDI2_RESERVED_SOB_CS_COMPLETION_FIRST +GAUDI2_MAX_PENDING_CS -1 ,GAUDI2_RESERVED_SOB_KDMA_COMPLETION , +GAUDI2_RESERVED_SOB_DEC_NRM_FIRST , +GAUDI2_RESERVED_SOB_DEC_NRM_LAST = +GAUDI2_RESERVED_SOB_DEC_NRM_FIRST +NUMBER_OF_DEC -1 , +GAUDI2_RESERVED_SOB_DEC_ABNRM_FIRST , +GAUDI2_RESERVED_SOB_DEC_ABNRM_LAST = +GAUDI2_RESERVED_SOB_DEC_ABNRM_FIRST +NUMBER_OF_DEC -1 ,GAUDI2_RESERVED_SOB_NUMBER };
@@ -256,6 +265,12 @@ enum gaudi2_reserved_mon_id {GAUDI2_RESERVED_MON_CS_COMPLETION_LAST =GAUDI2_RESERVED_MON_CS_COMPLETION_FIRST +GAUDI2_MAX_PENDING_CS -1 ,GAUDI2_RESERVED_MON_KDMA_COMPLETION , +GAUDI2_RESERVED_MON_DEC_NRM_FIRST , +GAUDI2_RESERVED_MON_DEC_NRM_LAST = +GAUDI2_RESERVED_MON_DEC_NRM_FIRST +3 *NUMBER_OF_DEC -1 , +GAUDI2_RESERVED_MON_DEC_ABNRM_FIRST , +GAUDI2_RESERVED_MON_DEC_ABNRM_LAST = +GAUDI2_RESERVED_MON_DEC_ABNRM_FIRST +3 *NUMBER_OF_DEC -1 , GAUDI2_RESERVED_MON_NUMBER };
drivers/misc/habanalabs/gaudi2/gaudi2_masks.h +3 -0
@@ -135,4 +135,7 @@#define PCIE_DBI_MSIX_ADDRESS_MATCH_LOW_OFF_MSIX_ADDRESS_MATCH_EN_SHIFT 0#define PCIE_DBI_MSIX_ADDRESS_MATCH_LOW_OFF_MSIX_ADDRESS_MATCH_EN_MASK 0x1 +#define DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_SIGN_SHIFT 15 +#define DCORE0_SYNC_MNGR_OBJS_SOB_OBJ_SIGN_MASK 0x8000 +#endif /* GAUDI2_MASKS_H_ */
drivers/misc/habanalabs/include/gaudi2/asic_reg/gaudi2_regs.h +6 -0
@@ -415,9 +415,15 @@#define SFT_DCORE_OFFSET (mmSFT1_HBW_RTR_IF0_RTR_CTRL_BASE - mmSFT0_HBW_RTR_IF0_RTR_CTRL_BASE)#define SFT_IF_OFFSET (mmSFT0_HBW_RTR_IF1_RTR_CTRL_BASE - mmSFT0_HBW_RTR_IF0_RTR_CTRL_BASE) +#define BRDG_CTRL_NRM_MSIX_LBW_AWADDR \ + (mmDCORE0_VDEC0_BRDG_CTRL_NRM_MSIX_LBW_AWADDR - mmDCORE0_VDEC0_BRDG_CTRL_BASE) +#define BRDG_CTRL_NRM_MSIX_LBW_WDATA \ (mmDCORE0_VDEC0_BRDG_CTRL_NRM_MSIX_LBW_WDATA - mmDCORE0_VDEC0_BRDG_CTRL_BASE) +#define BRDG_CTRL_ABNRM_MSIX_LBW_AWADDR \ + (mmDCORE0_VDEC0_BRDG_CTRL_ABNRM_MSIX_LBW_AWADDR - mmDCORE0_VDEC0_BRDG_CTRL_BASE) +#define BRDG_CTRL_ABNRM_MSIX_LBW_WDATA \ (mmDCORE0_VDEC0_BRDG_CTRL_ABNRM_MSIX_LBW_WDATA - mmDCORE0_VDEC0_BRDG_CTRL_BASE)
Signature · 1 line
2.25.1
PATCH 7/12 habanalabs/gaudi2: map virtual MSI-X doorbell memory for user
In reply to Oded GabbayFrom: Tomer Tayar <[email protected]>
Upon the initialization of a user context, map the host memory page of the virtual MSI-X doorbell in the device MMU.
A reserved VA is used for this purpose, so user can use it directly without any allocation/map operation.
--- drivers/misc/habanalabs/gaudi2/gaudi2.c | 43 ++++++++++++++++++- .../misc/habanalabs/include/gaudi2/gaudi2.h | 3 ++ 2 files changed, 44 insertions(+), 2 deletions(-)
drivers/misc/habanalabs/gaudi2/gaudi2.c +41 -2
@@ -1920,7 +1920,7 @@ static int gaudi2_set_fixed_properties(struct hl_device *hdev) prop->pmmu.hop_table_size = prop->mmu_hop_table_size; prop->pmmu.hop0_tables_total_size = prop->mmu_hop0_tables_total_size; - prop->hints_host_reserved_va_range.start_addr =RESERVED_VA_RANGE_FOR_ARC_ON_HOST_START ; + prop->hints_host_reserved_va_range.start_addr =RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_START ; prop->hints_host_reserved_va_range.end_addr =RESERVED_VA_RANGE_FOR_ARC_ON_HOST_END ; prop->hints_host_hpage_reserved_va_range.start_addr =RESERVED_VA_RANGE_FOR_ARC_ON_HOST_HPAGE_START ;
@@ -9241,6 +9241,35 @@ static void gaudi2_restore_user_registers(struct hl_device *hdev)gaudi2_restore_user_qm_registers (hdev); } +static int gaudi2_map_virtual_msix_doorbell_memory (struct hl_ctx *ctx) +{ +struct hl_device *hdev = ctx->hdev; +struct asic_fixed_properties *prop = &hdev->asic_prop; +struct gaudi2_device *gaudi2 = hdev->asic_specific; +int rc; + + rc =hl_mmu_map_page (ctx,RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_START , + gaudi2->virt_msix_db_dma_addr, prop->pmmu.page_size, true); +if (rc) +dev_err (hdev->dev,"Failed to map VA %#llx for virtual MSI-X doorbell memory\n" , +RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_START ); + +return rc; +} + +static void gaudi2_unmap_virtual_msix_doorbell_memory (struct hl_ctx *ctx) +{ +struct hl_device *hdev = ctx->hdev; +struct asic_fixed_properties *prop = &hdev->asic_prop; +int rc; + + rc =hl_mmu_unmap_page (ctx,RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_START , + prop->pmmu.page_size, true); +if (rc) +dev_err (hdev->dev,"Failed to unmap VA %#llx of virtual MSI-X doorbell memory\n" , +RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_START ); +} +static int gaudi2_ctx_init (struct hl_ctx *ctx) {int rc;
@@ -9257,7 +9286,15 @@ static int gaudi2_ctx_init(struct hl_ctx *ctx)else gaudi2_restore_user_registers (ctx->hdev); -return gaudi2_internal_cb_pool_init (ctx->hdev, ctx); + rc =gaudi2_internal_cb_pool_init (ctx->hdev, ctx); +if (rc) +return rc; + + rc =gaudi2_map_virtual_msix_doorbell_memory (ctx); +if (rc) +gaudi2_internal_cb_pool_fini (ctx->hdev, ctx); + +return rc; }static void gaudi2_ctx_fini (struct hl_ctx *ctx)
@@ -9266,6 +9303,8 @@ static void gaudi2_ctx_fini(struct hl_ctx *ctx)return ;gaudi2_internal_cb_pool_fini (ctx->hdev, ctx); + +gaudi2_unmap_virtual_msix_doorbell_memory (ctx); }static int gaudi2_pre_schedule_cs (struct hl_cs *cs)
drivers/misc/habanalabs/include/gaudi2/gaudi2.h +3 -0
@@ -54,6 +54,9 @@#define RESERVED_VA_RANGE_FOR_ARC_ON_HBM_START 0x1001500000000000ull#define RESERVED_VA_RANGE_FOR_ARC_ON_HBM_END 0x10016FFFFFFFFFFFull +#define RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_START 0xFFF077FFFFFF0000ull +#define RESERVED_VA_FOR_VIRTUAL_MSIX_DOORBELL_END 0xFFF077FFFFFFFFFFull +#define RESERVED_VA_RANGE_FOR_ARC_ON_HOST_START 0xFFF0780000000000ull#define RESERVED_VA_RANGE_FOR_ARC_ON_HOST_END 0xFFF07FFFFFFFFFFFull
Signature · 1 line
2.25.1
PATCH 8/12 habanalabs: expose only valid debugfs nodes
In reply to Oded GabbayFrom: Ofir Bitton <[email protected]>
In case security is enabled on the device, some debugfs nodes will fail. Hence, we do not expose them.
--- .../ABI/testing/debugfs-driver-habanalabs | 24 +++-- drivers/misc/habanalabs/common/debugfs.c | 94 ++++++++++--------- 2 files changed, 66 insertions(+), 52 deletions(-)
Documentation/ABI/testing/debugfs-driver-habanalabs +16 -8
@@ -130,14 +130,16 @@ Date: Jan 2019 KernelVersion: 5.1 Contact: [email protected] Description: Sets I2C device address for I2C transaction that is generated - by the device's CPU + by the device's CPU, Not available when device is loaded with secured + firmware What: /sys/kernel/debug/habanalabs/hl<n>/i2c_bus Date: Jan 2019 KernelVersion: 5.1 Contact: [email protected] Description: Sets I2C bus address for I2C transaction that is generated by - the device's CPU + the device's CPU, Not available when device is loaded with secured + firmware What: /sys/kernel/debug/habanalabs/hl<n>/i2c_data Date: Jan 2019
@@ -145,39 +147,45 @@ KernelVersion: 5.1 Contact: [email protected] Description: Triggers an I2C transaction that is generated by the device's CPU. Writing to this file generates a write transaction while - reading from the file generates a read transaction + reading from the file generates a read transaction, Not available + when device is loaded with secured firmware What: /sys/kernel/debug/habanalabs/hl<n>/i2c_len Date: Dec 2021 KernelVersion: 5.17 Contact: [email protected] Description: Sets I2C length in bytes for I2C transaction that is generated by - the device's CPU + the device's CPU, Not available when device is loaded with secured + firmware What: /sys/kernel/debug/habanalabs/hl<n>/i2c_reg Date: Jan 2019 KernelVersion: 5.1 Contact: [email protected] Description: Sets I2C register id for I2C transaction that is generated by - the device's CPU + the device's CPU, Not available when device is loaded with secured + firmware What: /sys/kernel/debug/habanalabs/hl<n>/led0 Date: Jan 2019 KernelVersion: 5.1 Contact: [email protected] -Description: Sets the state of the first S/W led on the device +Description: Sets the state of the first S/W led on the device, Not available + when device is loaded with secured firmware What: /sys/kernel/debug/habanalabs/hl<n>/led1 Date: Jan 2019 KernelVersion: 5.1 Contact: [email protected] -Description: Sets the state of the second S/W led on the device +Description: Sets the state of the second S/W led on the device, Not available + when device is loaded with secured firmware What: /sys/kernel/debug/habanalabs/hl<n>/led2 Date: Jan 2019 KernelVersion: 5.1 Contact: [email protected] -Description: Sets the state of the third S/W led on the device +Description: Sets the state of the third S/W led on the device, Not available + when device is loaded with secured firmware What: /sys/kernel/debug/habanalabs/hl<n>/memory_scrub Date: May 2022
drivers/misc/habanalabs/common/debugfs.c +50 -44
112 lines not on this page · raw message
Signature · 1 line
2.25.1
PATCH 9/12 habanalabs: fix update of is_in_soft_reset
In reply to Oded Gabbayreset_info.is_in_soft_reset should be updated both before in_reset and inside the spin lock of the reset info structure.
The reasons are:
- When we are inside soft reset, it implies we are in reset. Therefore, if someone checks if we are in soft reset, he can deduce we are in reset, while the opposite is not correct and might be misleading.
- Both these flags are changed together so they must be changed inside the reset info spinlock.
--- drivers/misc/habanalabs/common/device.c | 20 +++++++++++++++----- 1 file changed, 15 insertions(+), 5 deletions(-)
drivers/misc/habanalabs/common/device.c +15 -5
@@ -1346,7 +1346,14 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)spin_unlock (&hdev->reset_info.lock);return 0 ; } + +/* This still allows the completion of some KDMA ops +* Update this before in_reset because is_in_soft_reset implies we are in reset +*/ + hdev->reset_info.is_in_soft_reset = !hard_reset; + hdev->reset_info.in_reset =1 ; +spin_unlock (&hdev->reset_info.lock);if (delay_reset)
@@ -1354,9 +1361,6 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)handle_reset_trigger (hdev, flags); -/* This still allows the completion of some KDMA ops */ - hdev->reset_info.is_in_soft_reset = !hard_reset; -/* This also blocks future CS/VM/JOB completion operations */ hdev->disabled = true;
@@ -1565,7 +1569,7 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) }spin_lock (&hdev->reset_info.lock); - hdev->reset_info.is_in_soft_reset = false; + hdev->reset_info.is_in_soft_reset =0 ;/* Schedule hard reset only if requested and if not already in hard reset. * We keep 'in_reset' enabled, so no other reset can go in during the hard
@@ -1612,18 +1616,22 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) out_err: hdev->disabled = true; - hdev->reset_info.is_in_soft_reset = false; + +spin_lock (&hdev->reset_info.lock); + hdev->reset_info.is_in_soft_reset =0 ;if (hard_reset) {dev_err (hdev->dev,"Failed to reset! Device is NOT usable\n" ); hdev->reset_info.hard_reset_cnt++; }else if (reset_upon_device_release) { +spin_unlock (&hdev->reset_info.lock);dev_err (hdev->dev,"Failed to reset device after user release\n" ); flags |=HL_DRV_RESET_HARD ; flags &= ~HL_DRV_RESET_DEV_RELEASE ; hard_reset = true; goto again; }else { +spin_unlock (&hdev->reset_info.lock);dev_err (hdev->dev,"Failed to do soft-reset\n" ); hdev->reset_info.soft_reset_cnt++; flags |=HL_DRV_RESET_HARD ;
@@ -1633,6 +1641,8 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) hdev->reset_info.in_reset =0 ; +spin_unlock (&hdev->reset_info.lock); +return rc; }
Signature · 1 line
2.25.1
PATCH 10/12 habanalabs: add status of reset after device release
In reply to Oded GabbayThe user might want to know the device is in reset after device release, which is not an erroneous event as a regular reset.
--- drivers/misc/habanalabs/common/device.c | 17 +++++++++++------ drivers/misc/habanalabs/common/habanalabs_drv.c | 6 +++++- include/uapi/misc/habanalabs.h | 5 ++++- 3 files changed, 20 insertions(+), 8 deletions(-)
drivers/misc/habanalabs/common/device.c +11 -6
@@ -271,16 +271,20 @@ enum hl_device_status hl_device_status(struct hl_device *hdev) {enum hl_device_status status; -if (hdev->reset_info.in_reset) - status =HL_DEVICE_STATUS_IN_RESET ; -else if (hdev->reset_info.needs_reset) +if (hdev->reset_info.in_reset) { +if (hdev->reset_info.is_in_soft_reset) + status =HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE ; +else + status =HL_DEVICE_STATUS_IN_RESET ; + }else if (hdev->reset_info.needs_reset) { status =HL_DEVICE_STATUS_NEEDS_RESET ; -else if (hdev->disabled) + }else if (hdev->disabled) { status =HL_DEVICE_STATUS_MALFUNCTION ; -else if (!hdev->init_done) + }else if (!hdev->init_done) { status =HL_DEVICE_STATUS_IN_DEVICE_CREATION ; -else + }else { status =HL_DEVICE_STATUS_OPERATIONAL ; + }return status; }
@@ -296,6 +300,7 @@ bool hl_device_operational(struct hl_device *hdev,switch (current_status) {case HL_DEVICE_STATUS_IN_RESET : +case HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE :case HL_DEVICE_STATUS_MALFUNCTION :case HL_DEVICE_STATUS_NEEDS_RESET :return false;
drivers/misc/habanalabs/common/habanalabs_drv.c +5 -1
@@ -165,7 +165,8 @@ int hl_device_open(struct inode *inode, struct file *filp)"Can't open %s because it is %s\n" , dev_name(hdev->dev), hdev->status[status]); -if (status ==HL_DEVICE_STATUS_IN_RESET ) +if (status ==HL_DEVICE_STATUS_IN_RESET || + status ==HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE ) rc = -EAGAIN ;else rc = -EPERM ;
@@ -395,6 +396,9 @@ static int create_hdev(struct hl_device **dev, struct pci_dev *pdev)strncpy (hdev->status[HL_DEVICE_STATUS_NEEDS_RESET ],"needs reset" ,HL_STR_MAX );strncpy (hdev->status[HL_DEVICE_STATUS_IN_DEVICE_CREATION ],"in device creation" ,HL_STR_MAX ); +strncpy (hdev->status[HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE ], +"in reset after device release" ,HL_STR_MAX ); +/* First, we must find out which ASIC are we handling. This is needed * to configure the behavior of the driver (kernel parameters)
include/uapi/misc/habanalabs.h +4 -1
@@ -684,6 +684,8 @@ enum hl_goya_dma_direction {* @HL_DEVICE_STATUS_NEEDS_RESET: Device needs reset because auto reset was disabled. * @HL_DEVICE_STATUS_IN_DEVICE_CREATION: Device is operational but its creation is still in * progress. +* @HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE: Device is currently during reset that was +* triggered because the user released the device * @HL_DEVICE_STATUS_LAST: Last status. */ enum hl_device_status {
@@ -692,7 +694,8 @@ enum hl_device_status {HL_DEVICE_STATUS_MALFUNCTION ,HL_DEVICE_STATUS_NEEDS_RESET ,HL_DEVICE_STATUS_IN_DEVICE_CREATION , -HL_DEVICE_STATUS_LAST =HL_DEVICE_STATUS_IN_DEVICE_CREATION +HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE , +HL_DEVICE_STATUS_LAST =HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE }; enum hl_server_type {
Signature · 1 line
2.25.1
PATCH 11/12 habanalabs: rename soft reset to compute reset
In reply to Oded GabbayDoing compute reset can be the traditional inference soft reset that is supported only in Goya.
Or it can be the new reset upon device release, which is supported in Gaudi2 and above.
Therefore, wherever suitable, use the terminology of compute reset instead of soft reset.
--- drivers/misc/habanalabs/common/device.c | 28 +++++++++---------- drivers/misc/habanalabs/common/firmware_if.c | 4 +-- drivers/misc/habanalabs/common/habanalabs.h | 16 +++++------ .../misc/habanalabs/common/habanalabs_ioctl.c | 2 +- drivers/misc/habanalabs/common/irq.c | 2 +- drivers/misc/habanalabs/common/sysfs.c | 2 +- drivers/misc/habanalabs/gaudi2/gaudi2.c | 4 +-- drivers/misc/habanalabs/goya/goya.c | 2 +- 8 files changed, 30 insertions(+), 30 deletions(-)
drivers/misc/habanalabs/common/device.c +14 -14
@@ -272,7 +272,7 @@ enum hl_device_status hl_device_status(struct hl_device *hdev)enum hl_device_status status;if (hdev->reset_info.in_reset) { -if (hdev->reset_info.is_in_soft_reset) +if (hdev->reset_info.in_compute_reset) status =HL_DEVICE_STATUS_IN_RESET_AFTER_DEVICE_RELEASE ;else status =HL_DEVICE_STATUS_IN_RESET ;
@@ -1306,7 +1306,7 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) skip_wq_flush = !!(flags &HL_DRV_RESET_DEV_RELEASE ); delay_reset = !!(flags &HL_DRV_RESET_DELAY ); -if (!hard_reset && !hdev->asic_prop.supports_soft_reset) { +if (!hard_reset && !hdev->asic_prop.supports_compute_reset) { hard_instead_soft = true; hard_reset = true; }
@@ -1329,7 +1329,7 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) }if (hard_instead_soft) -dev_dbg (hdev->dev,"Doing hard-reset instead of soft-reset\n" ); +dev_dbg (hdev->dev,"Doing hard-reset instead of compute reset\n" ); do_reset:/* Re-entry of reset thread */
@@ -1345,17 +1345,17 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)/* Block future CS/VM/JOB completion operations */ spin_lock (&hdev->reset_info.lock);if (hdev->reset_info.in_reset) { -/* We only allow scheduling of a hard reset during soft reset */ -if (hard_reset && hdev->reset_info.is_in_soft_reset) +/* We only allow scheduling of a hard reset during compute reset */ +if (hard_reset && hdev->reset_info.in_compute_reset) hdev->reset_info.hard_reset_schedule_flags = flags;spin_unlock (&hdev->reset_info.lock);return 0 ; }/* This still allows the completion of some KDMA ops -* Update this before in_reset because is_in_soft_reset implies we are in reset +* Update this before in_reset because in_compute_reset implies we are in reset */ - hdev->reset_info.is_in_soft_reset = !hard_reset; + hdev->reset_info.in_compute_reset = !hard_reset; hdev->reset_info.in_reset =1 ;
@@ -1562,7 +1562,7 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)dev_err (hdev->dev,"Failed late init in reset after device release\n" ); else - dev_err(hdev->dev, "Failed late init after soft reset\n"); + dev_err(hdev->dev, "Failed late init after compute reset\n"); goto out_err; } }
@@ -1574,7 +1574,7 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) }spin_lock (&hdev->reset_info.lock); - hdev->reset_info.is_in_soft_reset =0 ; + hdev->reset_info.in_compute_reset =0 ;/* Schedule hard reset only if requested and if not already in hard reset. * We keep 'in_reset' enabled, so no other reset can go in during the hard
@@ -1604,11 +1604,11 @@ int hl_device_reset(struct hl_device *hdev, u32 flags)*/ hdev->asic_funcs->enable_events_from_fw (hdev); }else if (!reset_upon_device_release) { - hdev->reset_info.soft_reset_cnt++; + hdev->reset_info.compute_reset_cnt++; }if (schedule_hard_reset) { -dev_info (hdev->dev,"Performing hard reset scheduled during soft reset\n" ); +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;
@@ -1623,7 +1623,7 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) hdev->disabled = true;spin_lock (&hdev->reset_info.lock); - hdev->reset_info.is_in_soft_reset =0 ; + hdev->reset_info.in_compute_reset =0 ;if (hard_reset) {dev_err (hdev->dev,"Failed to reset! Device is NOT usable\n" );
@@ -1637,8 +1637,8 @@ int hl_device_reset(struct hl_device *hdev, u32 flags) goto again; }else {spin_unlock (&hdev->reset_info.lock); -dev_err (hdev->dev,"Failed to do soft-reset\n" ); - hdev->reset_info.soft_reset_cnt++; +dev_err (hdev->dev,"Failed to do compute reset\n" ); + hdev->reset_info.compute_reset_cnt++; flags |=HL_DRV_RESET_HARD ; hard_reset = true; goto again;
drivers/misc/habanalabs/common/firmware_if.c +2 -2
@@ -275,7 +275,7 @@ int hl_fw_send_cpu_message(struct hl_device *hdev, u32 hw_queue_id, u32 *msg,mutex_lock (&hdev->send_cpu_message_lock);/* CPU-CP messages can be sent during soft-reset */ -if (hdev->disabled && !hdev->reset_info.is_in_soft_reset) { +if (hdev->disabled && !hdev->reset_info.in_compute_reset) { rc =0 ; goto out; }
@@ -314,7 +314,7 @@ int hl_fw_send_cpu_message(struct hl_device *hdev, u32 hw_queue_id, u32 *msg,/* If FW performed reset just before sending it a packet, we will get a timeout. * This is expected behavior, hence no need for error message. */ -if (!hl_device_operational (hdev,NULL ) && !hdev->reset_info.is_in_soft_reset) +if (!hl_device_operational (hdev,NULL ) && !hdev->reset_info.in_compute_reset)dev_dbg (hdev->dev,"Device CPU packet timeout (0x%x) due to FW reset\n" , tmp);else
drivers/misc/habanalabs/common/habanalabs.h +8 -8
@@ -647,7 +647,7 @@ struct hl_hints_range {* false otherwise. * @use_get_power_for_reset_history: To support backward compatibility for Goya * and Gaudi -* @supports_soft_reset: is soft reset supported. +* @supports_compute_reset: is a reset which is not a hard-reset supported by this asic. * @allow_inference_soft_reset: true if the ASIC supports soft reset that is * initiated by user or TDR. This is only true * in inference ASICs, as there is no real-world
@@ -760,7 +760,7 @@ struct asic_fixed_properties {u8 dynamic_fw_load;u8 gic_interrupts_enable;u8 use_get_power_for_reset_history; -u8 supports_soft_reset; +u8 supports_compute_reset;u8 allow_inference_soft_reset;u8 configurable_stop_on_err;u8 set_max_power_on_device_init;
@@ -2960,12 +2960,12 @@ struct last_error_session_info {/** * struct hl_reset_info - holds current device reset information. * @lock: lock to protect critical reset flows. -* @soft_reset_cnt: number of soft reset since the driver was loaded. -* @hard_reset_cnt: number of hard reset since the driver was loaded. -* @hard_reset_schedule_flags: hard reset is scheduled to after current soft reset, +* @compute_reset_cnt: number of compte resets since the driver was loaded. +* @hard_reset_cnt: number of hard resets since the driver was loaded. +* @hard_reset_schedule_flags: hard reset is scheduled to after current compute reset, * here we hold the hard reset flags. * @in_reset: is device in reset flow. -* @is_in_soft_reset: Device is currently in soft reset process. +* @in_compute_reset: Device is currently in reset but not in hard-reset. * @needs_reset: true if reset_on_lockup is false and device should be reset * due to lockup. * @hard_reset_pending: is there a hard reset work pending.
@@ -2980,11 +2980,11 @@ struct last_error_session_info {*/ struct hl_reset_info {spinlock_t lock; -u32 soft_reset_cnt; +u32 compute_reset_cnt;u32 hard_reset_cnt;u32 hard_reset_schedule_flags;u8 in_reset; -u8 is_in_soft_reset; +u8 in_compute_reset;u8 needs_reset;u8 hard_reset_pending;
drivers/misc/habanalabs/common/habanalabs_ioctl.c +1 -1
@@ -294,7 +294,7 @@ static int get_reset_count(struct hl_device *hdev, struct hl_info_args *args)return -EINVAL ; reset_count.hard_reset_cnt = hdev->reset_info.hard_reset_cnt; - reset_count.soft_reset_cnt = hdev->reset_info.soft_reset_cnt; + reset_count.soft_reset_cnt = hdev->reset_info.compute_reset_cnt;return copy_to_user (out, &reset_count,min ((size_t ) max_size,sizeof (reset_count))) ? -EFAULT :0 ;
drivers/misc/habanalabs/common/irq.c +1 -1
@@ -411,7 +411,7 @@ irqreturn_t hl_irq_handler_eq(int irq, void *arg)*/ dma_rmb (); -if (hdev->disabled && !hdev->reset_info.is_in_soft_reset) { +if (hdev->disabled && !hdev->reset_info.in_compute_reset) {dev_warn (hdev->dev,"Device disabled but received an EQ event\n" ); goto skip_irq; }
drivers/misc/habanalabs/common/sysfs.c +1 -1
@@ -291,7 +291,7 @@ static ssize_t soft_reset_cnt_show(struct device *dev, {struct hl_device *hdev =dev_get_drvdata (dev); -return sprintf (buf,"%d\n" , hdev->reset_info.soft_reset_cnt); +return sprintf (buf,"%d\n" , hdev->reset_info.compute_reset_cnt); } static ssize_t hard_reset_cnt_show(struct device *dev,
drivers/misc/habanalabs/gaudi2/gaudi2.c +2 -2
@@ -2392,7 +2392,7 @@ static int gaudi2_cpucp_info_get(struct hl_device *hdev)/* No point of asking this information again when not doing hard reset, as the device * CPU hasn't been reset */ -if (hdev->reset_info.is_in_soft_reset) +if (hdev->reset_info.in_compute_reset)return 0 ; rc = hl_fw_cpucp_handshake(hdev, mmCPU_BOOT_DEV_STS0, mmCPU_BOOT_DEV_STS1, mmCPU_BOOT_ERR0,
@@ -3014,7 +3014,7 @@ static int gaudi2_sw_init(struct hl_device *hdev) hdev->supports_cb_mapping = true; hdev->supports_wait_for_multi_cs = false; - prop->supports_soft_reset = true; + prop->supports_compute_reset = true; hdev->asic_funcs->set_pci_memory_regions (hdev);
drivers/misc/habanalabs/goya/goya.c +1 -1
@@ -1038,7 +1038,7 @@ static int goya_sw_init(struct hl_device *hdev)spin_lock_init (&goya->hw_queues_lock); hdev->supports_coresight = true; - hdev->asic_prop.supports_soft_reset = true; + hdev->asic_prop.supports_compute_reset = true; hdev->asic_prop.allow_inference_soft_reset = true; hdev->supports_wait_for_multi_cs = false;
Signature · 1 line
2.25.1
PATCH 12/12 habanalabs: move h/w dirty message to debug
In reply to Oded GabbayH/W being dirty during initialization is completely expected in case f/w tools are used before loading the driver. As it is not an error, and as it doesn't give any meaningful information to the user, no point of printing it.
--- drivers/misc/habanalabs/gaudi/gaudi.c | 3 +-- drivers/misc/habanalabs/gaudi2/gaudi2.c | 2 +- drivers/misc/habanalabs/goya/goya.c | 3 +-- 3 files changed, 3 insertions(+), 5 deletions(-)
drivers/misc/habanalabs/gaudi/gaudi.c +1 -2
@@ -870,8 +870,7 @@ static int gaudi_early_init(struct hl_device *hdev) }if (gaudi_get_hw_state (hdev) ==HL_DEVICE_HW_STATE_DIRTY ) { -dev_info (hdev->dev, -"H/W state is dirty, must reset before initializing\n" ); +dev_dbg (hdev->dev,"H/W state is dirty, must reset before initializing\n" ); hdev->asic_funcs->hw_fini (hdev, true, false); }
drivers/misc/habanalabs/gaudi2/gaudi2.c +1 -1
@@ -2536,7 +2536,7 @@ static int gaudi2_early_init(struct hl_device *hdev) }if (gaudi2_get_hw_state (hdev) ==HL_DEVICE_HW_STATE_DIRTY ) { -dev_info (hdev->dev,"H/W state is dirty, must reset before initializing\n" ); +dev_dbg (hdev->dev,"H/W state is dirty, must reset before initializing\n" ); hdev->asic_funcs->hw_fini (hdev, true, false); }
drivers/misc/habanalabs/goya/goya.c +1 -2
@@ -673,8 +673,7 @@ static int goya_early_init(struct hl_device *hdev) }if (goya_get_hw_state (hdev) ==HL_DEVICE_HW_STATE_DIRTY ) { -dev_info (hdev->dev, -"H/W state is dirty, must reset before initializing\n" ); +dev_dbg (hdev->dev,"H/W state is dirty, must reset before initializing\n" ); hdev->asic_funcs->hw_fini (hdev, true, false); }
Signature · 1 line
2.25.1