habanalabs: always try to use the hint address
PATCH 1/3 habanalabs: always try to use the hint address
From: farah kassabri <[email protected]>
Currently hint address is ignored in case va block page size is not power of 2. We need to support th user hint address also in this case, but only if the hint address is aligned to page size.
--- drivers/misc/habanalabs/common/memory.c | 133 ++++++------------------ 1 file changed, 33 insertions(+), 100 deletions(-)
drivers/misc/habanalabs/common/memory.c +33 -100
@@ -520,8 +520,8 @@ static inline int add_va_block(struct hl_device *hdev, }/** -* get_va_block_pow2() - get a virtual block for the given size and alignment -* where alignment is a power of 2. +* get_va_block() - get a virtual block for the given size and alignment. +* * @hdev: pointer to the habanalabs device structure. * @va_range: pointer to the virtual addresses range. * @size: requested block size.
@@ -530,11 +530,11 @@ static inline int add_va_block(struct hl_device *hdev,* * This function does the following: * - Iterate on the virtual block list to find a suitable virtual block for the -* given size and alignment. +* given size, hint address and alignment. * - Reserve the requested block and update the list. * - Return the start address of the virtual block. */ -static u64 get_va_block_pow2 (struct hl_device *hdev, +static u64 get_va_block (struct hl_device *hdev,struct hl_va_range *va_range,u64 size,u64 hint_addr,u32 va_block_align) {
@@ -542,22 +542,37 @@ static u64 get_va_block_pow2(struct hl_device *hdev,u64 valid_start, valid_size, prev_start, prev_end, align_mask, reserved_valid_start =0 , reserved_valid_size =0 ;bool add_prev = false; +bool is_align_pow_2 =is_power_of_2 (va_range->page_size); - align_mask = ~((u64 )va_block_align -1 ); +if (is_align_pow_2) + align_mask = ~((u64 )va_block_align -1 ); +else { +/* +* with non-power-of-2 range we work only with page granularity +* and the start address is page aligned, +* so no need for alignment checking. +*/ + size =DIV_ROUND_UP_ULL (size, va_range->page_size) * + va_range->page_size; + } -/* check if hint_addr is aligned */ -if (hint_addr & (va_block_align -1 )) +/* Check if we need to ignore hint address */ +if ((is_align_pow_2 && (hint_addr & (va_block_align -1 ))) || + (!is_align_pow_2 && (hint_addr % va_range->page_size))) { +dev_info (hdev->dev,"Hint address 0x%llx will be ignored\n" , + hint_addr); hint_addr =0 ; + }mutex_lock (&va_range->lock);print_va_list_locked (hdev, &va_range->list);list_for_each_entry (va_block, &va_range->list, node) { -/* calc the first possible aligned addr */ +/* Calc the first possible aligned addr */ valid_start = va_block->start; -if (valid_start & (va_block_align -1 )) { +if (is_align_pow_2 && (valid_start & (va_block_align -1 ))) { valid_start &= align_mask; valid_start += va_block_align;if (valid_start > va_block->end)
@@ -565,9 +580,11 @@ static u64 get_va_block_pow2(struct hl_device *hdev, } valid_size = va_block->end - valid_start; +if (valid_size < size) +continue ; -if (valid_size >= size && (!new_va_block || - valid_size < reserved_valid_size)) { +/* Pick the minimal length block which has the required size */ +if (!new_va_block || (valid_size < reserved_valid_size)) { new_va_block = va_block; reserved_valid_start = valid_start; reserved_valid_size = valid_size;
@@ -584,10 +601,14 @@ static u64 get_va_block_pow2(struct hl_device *hdev,if (!new_va_block) {dev_err (hdev->dev,"no available va block for size %llu\n" , - size); + size); goto out; } +/* +* Check if there is some leftover range due to reserving the new +* va block, then return it to the main virtual addresses list. +*/ if (reserved_valid_start > new_va_block->start) { prev_start = new_va_block->start; prev_end = reserved_valid_start -1 ;
@@ -617,94 +638,6 @@ static u64 get_va_block_pow2(struct hl_device *hdev,return reserved_valid_start; } -/** -* get_va_block_non_pow2() - get a virtual block for the given size and -* alignment where alignment is not a power of 2. -* @hdev: pointer to the habanalabs device structure. -* @va_range: pointer to the virtual addresses range. -* @size: requested block size. -* @hint_addr: hint for requested address by the user. -* @va_block_align: required alignment of the virtual block start address. -* -* This function does the following: -* - Iterate on the virtual block list to find a suitable virtual block for the -* given size and alignment. -* - Reserve the requested block and update the list. -* - Return the start address of the virtual block. -*/ -static u64 get_va_block_non_pow2 (struct hl_device *hdev, -struct hl_va_range *va_range, -u64 size,u64 hint_addr,u32 va_block_align) -{ -struct hl_vm_va_block *va_block, *new_va_block =NULL ; -u64 reserved_valid_start =0 ; - -/* -* with non-power-of-2 range we work only with page granularity and the -* start address is page aligned, so no need for alignment checking. -*/ - size =DIV_ROUND_UP_ULL (size, va_range->page_size) * - va_range->page_size; - -mutex_lock (&va_range->lock); - -print_va_list_locked (hdev, &va_range->list); - -list_for_each_entry (va_block, &va_range->list, node) { -if ((va_block->start + size) > va_block->end) -continue ; - - new_va_block = va_block; - reserved_valid_start = va_block->start; -break ; - } - -if (!new_va_block) { -dev_err (hdev->dev,"no available va block for size %llu\n" , - size); - goto out; - } - -if (new_va_block->size > size) { - new_va_block->start += size; - new_va_block->size = new_va_block->end - new_va_block->start; - }else { -list_del (&new_va_block->node); -kfree (new_va_block); - } - -print_va_list_locked (hdev, &va_range->list); -out: -mutex_unlock (&va_range->lock); - -return reserved_valid_start; -} - -/* -* get_va_block() - get a virtual block for the given size and alignment. -* @hdev: pointer to the habanalabs device structure. -* @va_range: pointer to the virtual addresses range. -* @size: requested block size. -* @hint_addr: hint for requested address by the user. -* @va_block_align: required alignment of the virtual block start address. -* -* This function does the following: -* - Iterate on the virtual block list to find a suitable virtual block for the -* given size and alignment. -* - Reserve the requested block and update the list. -* - Return the start address of the virtual block. -*/ -static u64 get_va_block (struct hl_device *hdev,struct hl_va_range *va_range, -u64 size,u64 hint_addr,u32 va_block_align) -{ -if (is_power_of_2 (va_range->page_size)) -return get_va_block_pow2 (hdev, va_range, - size, hint_addr, va_block_align); -else -return get_va_block_non_pow2 (hdev, va_range, - size, hint_addr, va_block_align); -} -/* * hl_reserve_va_block() - reserve a virtual block of a given size. * @hdev: pointer to the habanalabs device structure.
Signature · 1 line
2.25.1
PATCH 2/3 habanalabs: add user available interrupt to hw_ip
From: Ofir Bitton <[email protected]>
In order to support completions that arrive directly to the user, the driver needs to supply the user with the first available msix interrupt available.
--- drivers/misc/habanalabs/common/habanalabs.h | 3 +++ drivers/misc/habanalabs/common/habanalabs_ioctl.c | 2 ++ drivers/misc/habanalabs/gaudi/gaudi.c | 2 ++ drivers/misc/habanalabs/goya/goya.c | 2 ++ include/uapi/misc/habanalabs.h | 6 ++++-- 5 files changed, 13 insertions(+), 2 deletions(-)
drivers/misc/habanalabs/common/habanalabs.h +3 -0
@@ -408,6 +408,8 @@ struct hl_mmu_properties {* @sync_stream_first_mon: first monitor available for sync stream use * @first_available_user_sob: first sob available for the user * @first_available_user_mon: first monitor available for the user +* @first_available_user_msix_interrupt: first available msix interrupt +* reserved for the user * @tpc_enabled_mask: which TPCs are enabled. * @completion_queues_count: number of completion queues. * @fw_security_disabled: true if security measures are disabled in firmware,
@@ -469,6 +471,7 @@ struct asic_fixed_properties {u16 sync_stream_first_mon;u16 first_available_user_sob[HL_MAX_DCORES ];u16 first_available_user_mon[HL_MAX_DCORES ]; +u16 first_available_user_msix_interrupt;u8 tpc_enabled_mask;u8 completion_queues_count;u8 fw_security_disabled;
drivers/misc/habanalabs/common/habanalabs_ioctl.c +2 -0
@@ -90,6 +90,8 @@ static int hw_ip_info(struct hl_device *hdev, struct hl_info_args *args) hw_ip.psoc_pci_pll_od = prop->psoc_pci_pll_od; hw_ip.psoc_pci_pll_div_factor = prop->psoc_pci_pll_div_factor; + hw_ip.first_available_interrupt_id = + prop->first_available_user_msix_interrupt;return copy_to_user (out, &hw_ip,min ((size_t )size,sizeof (hw_ip))) ? -EFAULT :0 ; }
drivers/misc/habanalabs/gaudi/gaudi.c +2 -0
@@ -527,6 +527,8 @@ static int gaudi_get_fixed_properties(struct hl_device *hdev) prop->sync_stream_first_mon + (num_sync_stream_queues *HL_RSVD_MONS ); + prop->first_available_user_msix_interrupt =USHRT_MAX ; +/* disable fw security for now, set it in a later stage */ prop->fw_security_disabled = true; prop->fw_security_status_valid = false;
drivers/misc/habanalabs/goya/goya.c +2 -0
@@ -455,6 +455,8 @@ int goya_get_fixed_properties(struct hl_device *hdev) prop->max_pending_cs =GOYA_MAX_PENDING_CS ; + prop->first_available_user_msix_interrupt =USHRT_MAX ; +/* disable fw security for now, set it in a later stage */ prop->fw_security_disabled = true; prop->fw_security_status_valid = false;
include/uapi/misc/habanalabs.h +4 -2
@@ -309,7 +309,9 @@ struct hl_info_hw_ip_info {__u32 num_of_events;__u32 device_id;/* PCI Device ID */ __u32 module_id;/* For mezzanine cards in servers (From OCP spec.) */ -__u32 reserved[2 ]; +__u32 reserved; +__u16 first_available_interrupt_id; +__u16 reserved2;__u32 cpld_version;__u32 psoc_pci_pll_nr;__u32 psoc_pci_pll_nf;
@@ -320,7 +322,7 @@ struct hl_info_hw_ip_info {__u8 pad[2 ];__u8 cpucp_version[HL_INFO_VERSION_MAX_LEN ];__u8 card_name[HL_INFO_CARD_NAME_MAX_LEN ]; -__u64 reserved2; +__u64 reserved3;__u64 dram_page_size; };
Signature · 1 line
2.25.1
PATCH 3/3 habanalabs: fix MMU debugfs related nodes
In reply to Oded GabbayFrom: farah kassabri <[email protected]>
In mmu debugfs node show un-scrambled physical addresses. before read/write through data nodes, need to unscramble the physical address before using it for pci transaction.
--- drivers/misc/habanalabs/common/debugfs.c | 21 ++++-- drivers/misc/habanalabs/common/habanalabs.h | 14 +++- drivers/misc/habanalabs/common/mmu/mmu.c | 74 ++++++++++++++++++--- drivers/misc/habanalabs/gaudi/gaudi.c | 3 +- drivers/misc/habanalabs/goya/goya.c | 3 +- 5 files changed, 93 insertions(+), 22 deletions(-)
drivers/misc/habanalabs/common/debugfs.c +15 -6
@@ -310,8 +310,8 @@ static int mmu_show(struct seq_file *s, void *data)struct hl_dbg_device_entry *dev_entry = entry->dev_entry;struct hl_device *hdev = dev_entry->hdev;struct hl_ctx *ctx; -struct hl_mmu_hop_info hops_info; -u64 virt_addr = dev_entry->mmu_addr; +struct hl_mmu_hop_info hops_info = {0 }; +u64 virt_addr = dev_entry->mmu_addr, phys_addr;int i;if (!hdev->mmu_enable)
@@ -333,10 +333,19 @@ static int mmu_show(struct seq_file *s, void *data)return 0 ; } -seq_printf (s, -"asid: %u, virt_addr: 0x%llx, scrambled virt_addr: 0x%llx\n" , - dev_entry->mmu_asid, dev_entry->mmu_addr, - hops_info.scrambled_vaddr); + phys_addr = hops_info.hop_info[hops_info.used_hops -1 ].hop_pte_val; + +if (hops_info.scrambled_vaddr && + (dev_entry->mmu_addr != hops_info.scrambled_vaddr)) +seq_printf (s, +"asid: %u, virt_addr: 0x%llx, scrambled virt_addr: 0x%llx,\nphys_addr: 0x%llx, scrambled_phys_addr: 0x%llx\n" , + dev_entry->mmu_asid, dev_entry->mmu_addr, + hops_info.scrambled_vaddr, + hops_info.unscrambled_paddr, phys_addr); +else +seq_printf (s, +"asid: %u, virt_addr: 0x%llx, phys_addr: 0x%llx\n" , + dev_entry->mmu_asid, dev_entry->mmu_addr, phys_addr);for (i =0 ; i < hops_info.used_hops ; i++) { seq_printf(s, "hop%d_addr: 0x%llx\n",
drivers/misc/habanalabs/common/habanalabs.h +11 -3
@@ -851,8 +851,10 @@ enum div_select_defs {* @collective_wait_init_cs: Generate collective master/slave packets * and place them in the relevant cs jobs * @collective_wait_create_jobs: allocate collective wait cs jobs -* @scramble_vaddr: Routine to scramble the virtual address prior of mapping it +* @scramble_addr: Routine to scramble the address prior of mapping it * in the MMU. +* @descramble_addr: Routine to de-scramble the address prior of +* showing it to users. * @ack_protection_bits_errors: ack and dump all security violations */ struct hl_asic_funcs {
@@ -963,7 +965,8 @@ struct hl_asic_funcs {int (*collective_wait_create_jobs)(struct hl_device *hdev,struct hl_ctx *ctx,struct hl_cs *cs,u32 wait_queue_id,u32 collective_engine_id); -u64 (*scramble_vaddr)(struct hl_device *hdev,u64 virt_addr); +u64 (*scramble_addr)(struct hl_device *hdev,u64 addr); +u64 (*descramble_addr)(struct hl_device *hdev,u64 addr);void (*ack_protection_bits_errors)(struct hl_device *hdev); };
@@ -1726,13 +1729,17 @@ struct hl_mmu_per_hop_info {* @scrambled_vaddr: The value of the virtual address after scrambling. This * address replaces the original virtual-address when mapped * in the MMU tables. +* @unscrambled_paddr: The un-scrambled physical address. * @hop_info: Array holding the per-hop information used for the translation. * @used_hops: The number of hops used for the translation. +* @range_type: virtual address range type. */ struct hl_mmu_hop_info {u64 scrambled_vaddr; +u64 unscrambled_paddr;struct hl_mmu_per_hop_info hop_info[MMU_ARCH_5_HOPS ];u32 used_hops; +enum hl_va_range_type range_type; };/**
@@ -2222,7 +2229,8 @@ void hl_mmu_v1_set_funcs(struct hl_device *hdev, struct hl_mmu_funcs *mmu);int hl_mmu_va_to_pa (struct hl_ctx *ctx,u64 virt_addr,u64 *phys_addr);int hl_mmu_get_tlb_info (struct hl_ctx *ctx,u64 virt_addr,struct hl_mmu_hop_info *hops); -u64 hl_mmu_scramble_vaddr (struct hl_device *hdev,u64 virt_addr); +u64 hl_mmu_scramble_addr (struct hl_device *hdev,u64 addr); +u64 hl_mmu_descramble_addr (struct hl_device *hdev,u64 addr);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,
drivers/misc/habanalabs/common/mmu/mmu.c +63 -11
110 lines not on this page · raw message
drivers/misc/habanalabs/gaudi/gaudi.c +2 -1
@@ -8548,7 +8548,8 @@ static const struct hl_asic_funcs gaudi_funcs = { .get_device_time = gaudi_get_device_time, .collective_wait_init_cs = gaudi_collective_wait_init_cs, .collective_wait_create_jobs = gaudi_collective_wait_create_jobs, - .scramble_vaddr = hl_mmu_scramble_vaddr, + .scramble_addr = hl_mmu_scramble_addr, + .descramble_addr = hl_mmu_descramble_addr, .ack_protection_bits_errors = gaudi_ack_protection_bits_errors };
drivers/misc/habanalabs/goya/goya.c +2 -1
@@ -5459,7 +5459,8 @@ static const struct hl_asic_funcs goya_funcs = { .get_device_time = goya_get_device_time, .collective_wait_init_cs = goya_collective_wait_init_cs, .collective_wait_create_jobs = goya_collective_wait_create_jobs, - .scramble_vaddr = hl_mmu_scramble_vaddr, + .scramble_addr = hl_mmu_scramble_addr, + .descramble_addr = hl_mmu_descramble_addr, .ack_protection_bits_errors = goya_ack_protection_bits_errors };
Signature · 1 line
2.25.1