Add new apis: wd_blkpool_new; wd_blkpool_delete; wd_blkpool_phy; wd_blkpool_alloc; wd_blkpool_free; wd_blkpool_setup; wd_blkpool_destroy_mem; wd_blkpool_create_sglpool; wd_blkpool_destroy_sglpool; App only use two apis after setup blkpool wd_blkpool_alloc; wd_blkpool_free; Signed-off-by: Zhangfei Gao <zhangfei.gao@linaro.org> --- Makefile.am | 4 +- include/wd_bmm.h | 60 ++++ libwd.map | 10 + wd_bmm.c | 727 +++++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 800 insertions(+), 1 deletion(-) create mode 100644 include/wd_bmm.h create mode 100644 wd_bmm.c diff --git a/Makefile.am b/Makefile.am index c4b9c52..16704d3 100644 --- a/Makefile.am +++ b/Makefile.am @@ -36,7 +36,8 @@ pkginclude_HEADERS = include/wd.h include/wd_cipher.h include/wd_aead.h \ include/wd_comp.h include/wd_dh.h include/wd_digest.h \ include/wd_rsa.h include/uacce.h include/wd_alg_common.h \ include/wd_ecc.h include/wd_sched.h include/wd_alg.h \ - include/wd_zlibwrapper.h include/wd_dae.h include/wd_agg.h + include/wd_zlibwrapper.h include/wd_dae.h include/wd_agg.h \ + include/wd_bmm.h nobase_pkginclude_HEADERS = v1/wd.h v1/wd_cipher.h v1/wd_aead.h v1/uacce.h v1/wd_dh.h \ v1/wd_digest.h v1/wd_rsa.h v1/wd_bmm.h @@ -60,6 +61,7 @@ libwd_la_SOURCES=wd.c wd_mempool.c wd.h wd_alg.c wd_alg.h \ v1/wd_bmm.c v1/wd_bmm.h \ v1/wd_ecc.c v1/wd_ecc.h \ v1/wd_sgl.c v1/wd_sgl.h \ + wd_bmm.c \ aes.h sm4.h galois.h \ lib/crypto/aes.c lib/crypto/sm4.c lib/crypto/galois.c \ v1/drv/hisi_qm_udrv.c v1/drv/hisi_qm_udrv.h \ diff --git a/include/wd_bmm.h b/include/wd_bmm.h new file mode 100644 index 0000000..0730cf5 --- /dev/null +++ b/include/wd_bmm.h @@ -0,0 +1,60 @@ +/* SPDX-License-Identifier: Apache-2.0 + * Copyright 2025-2026 Huawei Technologies Co.,Ltd. All rights reserved. + * Copyright 2025-2026 Linaro ltd. + */ + +#ifndef _WD_BMM_H +#define _WD_BMM_H + +#include <asm/types.h> +#include "wd.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define DEFAULT_BLOCK_NM 16384 +#define DEFAULT_ALIGN_SIZE 0x40 +#define DEFAULT_BLOCK_SIZE (1024 * 8) + +/* memory APIs for Algorithm Layer */ +typedef void *(*wd_alloc)(void *usr, size_t size); +typedef void (*wd_free)(void *usr, void *va); + + /* memory VA to DMA address map */ +typedef void *(*wd_map)(void *usr, void *va, size_t sz); +typedef __u32 (*wd_bufsize)(void *usr); + +/* Memory from user, it is given at ctx creating. */ +struct wd_mm_br { + wd_alloc alloc; /* Memory allocation */ + wd_free free; /* Memory free */ + wd_map iova_map; /* get iova from user space VA */ + void *usr; /* data for the above operations */ + wd_bufsize get_bufsize; /* optional */ +}; + +/* Memory pool creating parameters */ +struct wd_blkpool_setup { + __u32 block_size; /* Block buffer size */ + __u32 block_num; /* Block buffer number */ + __u32 align_size; /* Block buffer starting address align size */ + struct wd_mm_br br; /* memory from user if don't use WD memory */ +}; + + +void *wd_blkpool_new(handle_t h_ctx); +void wd_blkpool_delete(void *pool); +int wd_blkpool_setup(void *pool, struct wd_blkpool_setup *setup); +void wd_blkpool_destroy_mem(void *pool); +void *wd_blkpool_alloc(void *pool, size_t size); +void wd_blkpool_free(void *pool, void *va); +void *wd_blkpool_phy(void *pool, void *va); +handle_t wd_blkpool_create_sglpool(void *pool); +void wd_blkpool_destroy_sglpool(void *pool, handle_t sgl_pool); + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/libwd.map b/libwd.map index 5522ec0..e884671 100644 --- a/libwd.map +++ b/libwd.map @@ -49,5 +49,15 @@ global: wd_enable_drv; wd_disable_drv; wd_get_alg_head; + + wd_blkpool_new; + wd_blkpool_delete; + wd_blkpool_phy; + wd_blkpool_alloc; + wd_blkpool_free; + wd_blkpool_setup; + wd_blkpool_destroy_mem; + wd_blkpool_create_sglpool; + wd_blkpool_destroy_sglpool; local: *; }; diff --git a/wd_bmm.c b/wd_bmm.c new file mode 100644 index 0000000..d252c74 --- /dev/null +++ b/wd_bmm.c @@ -0,0 +1,727 @@ +// SPDX-License-Identifier: Apache-2.0 +/* + * Copyright 2025-2026 Huawei Technologies Co.,Ltd. All rights reserved. + * Copyright 2025-2026 Linaro ltd. + */ + +/* Block Memory Management (lib): A block memory algorithm */ +#include <asm/byteorder.h> +#include <stdio.h> +#include <string.h> +#include <stdint.h> +#include <unistd.h> +#include <stdlib.h> +#include <errno.h> +#include <sys/queue.h> +#include <sys/mman.h> +#include <pthread.h> + +#include "wd.h" +#include "wd_bmm.h" + +#define __ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask)) +#define ALIGN(x, a) __ALIGN_MASK((uintptr_t)(x), (uintptr_t)(a)-1) + +#define TAG_FREE 0x12345678 /* block is free */ +#define TAG_USED 0x87654321 /* block is busy */ +#define MAX_ALIGN_SIZE 0x1000 /* 4KB */ +#define MAX_BLOCK_SIZE 0x10000000 /* 256MB */ +#define BLK_BALANCE_SZ 0x100000ul +#define NUM_TIMES(x) (87 * (x) / 100) + +#define WD_UACCE_GRAN_SIZE 0x10000ull +#define WD_UACCE_GRAN_SHIFT 16 +#define WD_UACCE_GRAN_NUM_MASK 0xfffull + +#define DEFAULT_BLKSIZE_ALIGN 0x1000 + +/* the max sge num in one sgl */ +#define SGE_NUM_IN_SGL 255 + +/* the max sge num in on BD, QM user it be the sgl pool size */ +#define SGL_NUM_IN_BD 256 + +struct wd_blk_hd { + unsigned int blk_tag; + void *blk_dma; + void *blk; + + TAILQ_ENTRY(wd_blk_hd) next; +}; + +TAILQ_HEAD(wd_blk_list, wd_blk_hd); + +struct wd_ss_region { + void *va; + unsigned long long pa; + size_t size; + + TAILQ_ENTRY(wd_ss_region) next; +}; + +TAILQ_HEAD(wd_ss_region_list, wd_ss_region); + +struct wd_blkpool { + pthread_spinlock_t lock; + unsigned int free_blk_num; + unsigned int blk_num; + unsigned int alloc_failures; + struct wd_blk_list head; + void *act_start; + unsigned int hd_sz; + unsigned int blk_sz; + struct wd_blkpool_setup setup; + + handle_t ctx; + void *mem; + unsigned long size; + struct wd_ss_region_list ss_list; + struct wd_ss_region_list *ss_head; + struct hisi_sgl_pool *sgl_pool; + void *sgl_mem; + size_t sgl_size; +}; + +struct hisi_sge { + uintptr_t buff; + void *page_ctrl; + __le32 len; + __le32 pad; + __le32 pad0; + __le32 pad1; +}; + +/* use default hw sgl head size 64B, in little-endian */ +struct hisi_sgl { + /* the next sgl address */ + uintptr_t next_dma; + /* the sge num of all the sgl */ + __le16 entry_sum_in_chain; + /* valid sge(has buff) num in this sgl */ + __le16 entry_sum_in_sgl; + /* the sge num in this sgl */ + __le16 entry_length_in_sgl; + __le16 pad0; + __le64 pad1[5]; + /* valid sge buffs total size */ + __le64 entry_size_in_sgl; + struct hisi_sge sge_entries[]; +}; + +struct hisi_sgl_pool { + /* the addr64 align offset base sgl */ + void **sgl_align; + /* the sgl src address array */ + void **sgl; + /* the sgl pool stack depth */ + __u32 depth; + __u32 top; + __u32 sge_num; + __u32 sgl_num; + pthread_spinlock_t lock; + void **phys; +}; + +static struct wd_blk_hd *wd_blk_head(struct wd_blkpool *pool, void *blk) +{ + unsigned long offset = (unsigned long)((uintptr_t)blk - + (uintptr_t)pool->act_start); + unsigned long sz = pool->hd_sz + pool->blk_sz; + unsigned long blk_idx = offset / sz; + + return (struct wd_blk_hd *)((uintptr_t)pool->act_start + blk_idx * sz); +} + +static int pool_params_check(struct wd_blkpool_setup *setup) +{ + if (!setup->block_size || + setup->block_size > MAX_BLOCK_SIZE) { + WD_ERR("Invalid block_size (%x)!\n", + setup->block_size); + return -WD_EINVAL; + } + + /* check parameters, and align_size must be 2^N */ + if (setup->align_size == 0x1 || setup->align_size > MAX_ALIGN_SIZE || + setup->align_size & (setup->align_size - 0x1)) { + WD_ERR("Invalid align_size.\n"); + return -WD_EINVAL; + } + + return WD_SUCCESS; +} + +static inline int calculate_sgl_size(void) +{ + int sgl_size = sizeof(struct hisi_sgl) + + SGE_NUM_IN_SGL * sizeof(struct hisi_sge); + + return ALIGN(sgl_size, DEFAULT_ALIGN_SIZE); +} + +static inline size_t calculate_extra_sgl_size(void) +{ + return SGL_NUM_IN_BD * calculate_sgl_size(); +} + +static int wd_pool_pre_layout(struct wd_blkpool *p, + struct wd_blkpool_setup *sp) +{ + size_t extra_sgl_size = calculate_extra_sgl_size(); + unsigned int asz; + int ret; + + ret = pool_params_check(sp); + if (ret) + return ret; + + asz = sp->align_size; + + /* Get actual value by align */ + p->hd_sz = ALIGN(sizeof(struct wd_blk_hd), asz); + p->blk_sz = ALIGN(sp->block_size, asz); + + if (p->size == 0 && !p->mem) { + p->size = (p->hd_sz + p->blk_sz) * + (unsigned long)sp->block_num + asz + + extra_sgl_size; + + /* Make sure memory map granularity size align */ + if (wd_is_noiommu(p->ctx)) + p->size = ALIGN(p->size, WD_UACCE_GRAN_SIZE); + } + + return WD_SUCCESS; +} + +static void *wd_get_phys(struct wd_blkpool *pool, void *va) +{ + struct wd_ss_region *rgn; + + TAILQ_FOREACH(rgn, pool->ss_head, next) { + if (rgn->va <= va && va < rgn->va + rgn->size) + return (void *)(uintptr_t)(rgn->pa + + ((uintptr_t)va - (uintptr_t)rgn->va)); + } + + return NULL; +} + +static int wd_pool_init(struct wd_blkpool *p) +{ + __u32 blk_size = p->setup.block_size; + size_t extra_sgl_size = calculate_extra_sgl_size(); + void *dma_start, *dma_end, *va; + struct wd_blk_hd *hd = NULL; + unsigned int dma_num = 0; + unsigned int i, act_num; + unsigned long loss; + + p->act_start = (void *)ALIGN((uintptr_t)p->mem, p->setup.align_size); + loss = p->act_start - p->mem; + + /* ignore sgl */ + act_num = (p->size - loss - extra_sgl_size) / (p->hd_sz + p->blk_sz); + + /* get dma address and initialize blocks */ + for (i = 0; i < act_num; i++) { + va = (void *)((uintptr_t)p->act_start + p->hd_sz + + (unsigned long)(p->hd_sz + + p->blk_sz) * i); + dma_start = wd_get_phys(p, va); + dma_end = wd_get_phys(p, va + blk_size - 1); + if (!dma_start || !dma_end) { + WD_ERR("wd_get_phys err.\n"); + return -WD_ENOMEM; + } + + if ((uintptr_t)dma_end - (uintptr_t)dma_start != blk_size - 1) + continue; + + hd = (void *)((uintptr_t)va - p->hd_sz); + hd->blk_dma = dma_start; + hd->blk = va; + hd->blk_tag = TAG_FREE; + TAILQ_INSERT_TAIL(&p->head, hd, next); + + dma_num++; + } + + p->free_blk_num = dma_num; + p->blk_num = dma_num; + + return WD_SUCCESS; +} + +static int usr_pool_init(struct wd_blkpool *p) +{ + struct wd_blkpool_setup *sp = &p->setup; + size_t extra_sgl_size = calculate_extra_sgl_size(); + __u32 blk_size = sp->block_size; + struct wd_blk_hd *hd = NULL; + unsigned long loss; + unsigned int i, act_num; + + p->act_start = (void *)ALIGN((uintptr_t)p->mem, sp->align_size); + loss = p->act_start - p->mem; + /* ignore sgl */ + act_num = (p->size - loss - extra_sgl_size) / (p->hd_sz + p->blk_sz); + + for (i = 0; i < act_num; i++) { + hd = (void *)((uintptr_t)p->act_start + (p->hd_sz + p->blk_sz) * i); + hd->blk = (void *)((uintptr_t)hd + p->hd_sz); + hd->blk_dma = sp->br.iova_map(sp->br.usr, hd->blk, blk_size); + if (!hd->blk_dma) { + WD_ERR("failed to map usr blk.\n"); + return -WD_ENOMEM; + } + hd->blk_tag = TAG_FREE; + TAILQ_INSERT_TAIL(&p->head, hd, next); + } + + p->free_blk_num = act_num; + p->blk_num = p->free_blk_num; + + return WD_SUCCESS; +} + +static void drv_free_slice(struct wd_blkpool *p) +{ + struct wd_ss_region *rgn; + + while (true) { + rgn = TAILQ_FIRST(&p->ss_list); + if (!rgn) + break; + TAILQ_REMOVE(&p->ss_list, rgn, next); + free(rgn); + } +} + +static void drv_add_slice(struct wd_blkpool *p, struct wd_ss_region *rgn) +{ + struct wd_ss_region *rg; + + rg = TAILQ_LAST(&p->ss_list, wd_ss_region_list); + if (rg) { + if (rg->pa + rg->size == rgn->pa) { + rg->size += rgn->size; + free(rgn); + return; + } + } + + TAILQ_INSERT_TAIL(&p->ss_list, rgn, next); +} + +static void *pool_reserve_mem(struct wd_blkpool *p, size_t size) +{ + struct wd_ss_region *rgn = NULL; + unsigned long info = 0; + size_t tmp = 0; + unsigned long i = 0; + void *ptr = NULL; + int ret = 1; + + if (!p->ctx) + return NULL; + + if (p->mem) + return NULL; + + ptr = wd_reserve_mem(p->ctx, size); + if (!ptr) + return NULL; + + p->ss_head = &p->ss_list; + TAILQ_INIT(&p->ss_list); + + while (ret > 0) { + info = i; + ret = wd_ctx_set_io_cmd(p->ctx, UACCE_CMD_GET_SS_DMA, &info); + if (ret < 0) { + WD_ERR("get DMA fail!\n"); + goto err_out; + } + rgn = malloc(sizeof(*rgn)); + if (!rgn) { + WD_ERR("alloc ss region fail!\n"); + goto err_out; + } + memset(rgn, 0, sizeof(*rgn)); + + if (wd_is_noiommu(p->ctx)) + rgn->size = (info & WD_UACCE_GRAN_NUM_MASK) << + WD_UACCE_GRAN_SHIFT; + else + rgn->size = p->size; + rgn->pa = info & (~WD_UACCE_GRAN_NUM_MASK); + rgn->va = ptr + tmp; + tmp += rgn->size; + drv_add_slice(p, rgn); + i++; + } + + return ptr; + +err_out: + drv_free_slice(p); + munmap(p->mem, size); + + return NULL; +} + +static int pool_init(struct wd_blkpool *pool, + struct wd_blkpool_setup *setup) +{ + void *addr = NULL; + + /* use user's memory, and its br alloc function */ + if (setup->br.alloc && setup->br.free) { + if (!pool->mem) { + addr = setup->br.alloc(setup->br.usr, pool->size); + if (!addr) { + WD_ERR("failed to allocate memory in user pool.\n"); + return -EINVAL; + } + pool->mem = addr; + } + if (usr_pool_init(pool)) { + WD_ERR("failed to initialize user pool.\n"); + setup->br.free(setup->br.usr, addr); + return -EINVAL; + } + } else { + if (!pool->mem) { + /* use wd to reserve memory */ + addr = pool_reserve_mem(pool, pool->size); + if (!addr) { + WD_ERR("wd pool failed to reserve memory.\n"); + return -EINVAL; + } + pool->mem = addr; + } + + if (wd_pool_init(pool)) { + WD_ERR("failed to initialize wd pool.\n"); + wd_blkpool_destroy_mem(pool); + return -EINVAL; + } + } + + return 0; +} + +void *wd_blkpool_new(handle_t h_ctx) +{ + struct wd_blkpool *pool; + + if (wd_is_sva(h_ctx)) + return NULL; + + pool = calloc(1, sizeof(*pool)); + if (!pool) { + WD_ERR("failed to malloc pool.\n"); + return NULL; + } + pool->ctx = h_ctx; + + if (pthread_spin_init(&pool->lock, PTHREAD_PROCESS_SHARED) != 0) { + free(pool); + return NULL; + } + return pool; +} + +int wd_blkpool_setup(void *pool, struct wd_blkpool_setup *setup) +{ + struct wd_blkpool *p = pool; + int ret = 0; + + if (!p || !setup) + return -EINVAL; + + pthread_spin_lock(&p->lock); + if (p->mem && p->size != 0) { + if (p->setup.block_size == setup->block_size || + p->blk_sz == ALIGN(setup->block_size, setup->align_size)) + goto out; + + /* re-org blk_size, no need reserve mem */ + if (p->free_blk_num != p->blk_num) { + WD_ERR("Can not reset blk pool, as it's in use.\n"); + ret = -EINVAL; + goto out; + } + } + + memcpy(&p->setup, setup, sizeof(p->setup)); + + ret = wd_pool_pre_layout(p, setup); + if (ret) + goto out; + + TAILQ_INIT(&p->head); + + ret = pool_init(p, setup); + +out: + pthread_spin_unlock(&p->lock); + return ret; +} + +void *wd_blkpool_alloc(void *pool, size_t size) +{ + struct wd_blkpool *p = pool; + struct wd_blk_hd *hd; + int ret; + + if (!p) + return NULL; + + if (!p->mem || size > p->blk_sz) { + struct wd_blkpool_setup setup; + /* + * if empty pool, will reserve mem and init pool + * if size > blk_size, will re-org as align 4K if free pool + */ + + memset(&setup, 0, sizeof(setup)); + setup.block_size = ALIGN(size, DEFAULT_BLKSIZE_ALIGN); + setup.block_num = DEFAULT_BLOCK_NM; + setup.align_size = DEFAULT_ALIGN_SIZE; + ret = wd_blkpool_setup(p, &setup); + if (ret) + return NULL; + } + + pthread_spin_lock(&p->lock); + hd = TAILQ_LAST(&p->head, wd_blk_list); + if (unlikely(!hd || hd->blk_tag != TAG_FREE)) { + p->alloc_failures++; + goto out; + } + + /* Delete the block buffer from free list */ + TAILQ_REMOVE(&p->head, hd, next); + p->free_blk_num--; + hd->blk_tag = TAG_USED; + pthread_spin_unlock(&p->lock); + + return hd->blk; + +out: + pthread_spin_unlock(&p->lock); + WD_ERR("Failed to malloc blk.\n"); + + return NULL; +} + +void wd_blkpool_free(void *pool, void *va) +{ + struct wd_blkpool *p = pool; + struct wd_blk_hd *hd; + + if (!p || !va) + return; + + hd = wd_blk_head(p, va); + if (unlikely(hd->blk_tag != TAG_USED)) { + WD_ERR("free block fail!\n"); + return; + } + + pthread_spin_lock(&p->lock); + TAILQ_INSERT_TAIL(&p->head, hd, next); + p->free_blk_num++; + hd->blk_tag = TAG_FREE; + pthread_spin_unlock(&p->lock); +} + +void *wd_blkpool_phy(void *pool, void *va) +{ + struct wd_blkpool *p = pool; + struct wd_blk_hd *hd; + unsigned long off, idx; + + if (!pool || !va) + return NULL; + + if (p->sgl_mem != 0 && va >= p->sgl_mem) { + off = (unsigned long) (va - p->sgl_mem); + idx = off / p->sgl_size; + + return p->sgl_pool->phys[idx]; + } + + hd = wd_blk_head(pool, va); + if (hd->blk_tag != TAG_USED || + (uintptr_t)va < (uintptr_t)hd->blk) + return NULL; + + return (void *)((uintptr_t)hd->blk_dma + ((uintptr_t)va - + (uintptr_t)hd->blk)); +} + +int wd_blkpool_get_free_blk_num(void *pool, __u32 *free_num) +{ + struct wd_blkpool *p = pool; + + if (!p || !free_num) { + WD_ERR("get_free_blk_num err, parameter err!\n"); + return -WD_EINVAL; + } + + *free_num = __atomic_load_n(&p->free_blk_num, __ATOMIC_RELAXED); + + return WD_SUCCESS; +} + +int wd_blkpool_alloc_failures(void *pool, __u32 *fail_num) +{ + struct wd_blkpool *p = pool; + + if (!p || !fail_num) { + WD_ERR("get_blk_alloc_failure err, pool is NULL!\n"); + return -WD_EINVAL; + } + + *fail_num = __atomic_load_n(&p->alloc_failures, __ATOMIC_RELAXED); + + return WD_SUCCESS; +} + +__u32 wd_blkpool_blksize(void *pool) +{ + struct wd_blkpool *p = pool; + + if (!p) { + WD_ERR("get blk_size pool is null!\n"); + return 0; + } + + return p->blk_sz; +} + +void wd_blkpool_destroy_mem(void *pool) +{ + struct wd_blkpool_setup *setup; + struct wd_blkpool *p = pool; + + if (!p) { + WD_ERR("pool destroy err, pool is NULL.\n"); + return; + } + + pthread_spin_lock(&p->lock); + if (p->mem) { + setup = &p->setup; + if (setup->br.free) { + setup->br.free(setup->br.usr, p->mem); + } else { + drv_free_slice(p); + munmap(p->mem, p->size); + } + p->mem = NULL; + p->size = 0; + } + pthread_spin_unlock(&p->lock); +} + +void wd_blkpool_delete(void *pool) +{ + struct wd_blkpool *p = pool; + + if (!p) + return; + + wd_blkpool_destroy_mem(pool); + pthread_spin_destroy(&p->lock); + free(p); +} + +handle_t wd_blkpool_create_sglpool(void *pool) +{ + struct wd_blkpool *p = pool; + struct hisi_sgl_pool *sgl_pool; + struct hisi_sgl *sgl_align; + size_t sgl_size = calculate_sgl_size(); + size_t extra_sgl_size = calculate_extra_sgl_size(); + struct wd_blkpool_setup *sp; + void *base; + + if (!p) + return 0; + + sgl_pool = calloc(1, sizeof(struct hisi_sgl_pool)); + if (!sgl_pool) { + WD_ERR("failed to alloc memory for sgl_pool!\n"); + return 0; + } + + sgl_pool->sgl_align = calloc(SGL_NUM_IN_BD, sizeof(void *)); + if (!sgl_pool->sgl_align) { + WD_ERR("failed to alloc memory for sgl align!\n"); + goto err_out; + } + + sgl_pool->phys = calloc(SGL_NUM_IN_BD, sizeof(void *)); + if (!sgl_pool->phys) { + WD_ERR("failed to alloc memory for phys!\n"); + goto err_out; + } + + base = (void *)((uintptr_t)p->mem + p->size - extra_sgl_size); + sp = &p->setup; + + for (int i = 0; i < SGL_NUM_IN_BD; i++) { + sgl_align = (struct hisi_sgl *)ALIGN(base + sgl_size * i, DEFAULT_ALIGN_SIZE); + sgl_align->entry_sum_in_chain = SGE_NUM_IN_SGL; + sgl_align->entry_sum_in_sgl = 0; + sgl_align->entry_length_in_sgl = SGE_NUM_IN_SGL; + sgl_align->next_dma = 0; + sgl_pool->sgl_align[i] = sgl_align; + if (sp->br.iova_map) + sgl_pool->phys[i] = sp->br.iova_map(sp->br.usr, sgl_align, sgl_size); + else + sgl_pool->phys[i] = wd_get_phys(p, sgl_align); + } + + if (pthread_spin_init(&sgl_pool->lock, PTHREAD_PROCESS_SHARED) != 0) { + WD_ERR("failed to init sgl pool lock!\n"); + goto err_out; + } + + sgl_pool->sgl_num = SGL_NUM_IN_BD; + sgl_pool->sge_num = SGE_NUM_IN_SGL; + sgl_pool->depth = SGL_NUM_IN_BD; + sgl_pool->top = SGL_NUM_IN_BD; + p->sgl_pool = sgl_pool; + p->sgl_size = sgl_size; + p->sgl_mem = (void *)ALIGN(base, DEFAULT_ALIGN_SIZE); + + return (handle_t)sgl_pool; + +err_out: + if (sgl_pool->phys) + free(sgl_pool->phys); + if (sgl_pool->sgl_align) + free(sgl_pool->sgl_align); + free(sgl_pool); + return (handle_t)0; +} + +void wd_blkpool_destroy_sglpool(void *pool, handle_t h_sgl_pool) +{ + struct hisi_sgl_pool *sgl_pool = (struct hisi_sgl_pool *)h_sgl_pool; + + if (!h_sgl_pool) + return; + + pthread_spin_destroy(&sgl_pool->lock); + if (sgl_pool->phys) + free(sgl_pool->phys); + if (sgl_pool->sgl_align) + free(sgl_pool->sgl_align); + free(sgl_pool); +} -- 2.25.1

Add api wd_comp_setup_blkpool. Other alg.c will need wd_xxx_setup_blkpool as well. The reason is app does not know ctx. It will setup blkpool for ctx[0] and sglpool for sgl mode. The blkpool will be used by app and driver. App need call wd_xxx_setup_blkpool for user pointer mode and sgl mode. The returned blkpool will be used for wd_blkpool_alloc/free. Alloc_sess will call wd_xxx_setup_blkpool if it is not called by app. Then uadk library will alloc blkpool and memcpy to user memory, with poorer performance. The driver will translate va to pa when configure register. Signed-off-by: Zhangfei Gao <zhangfei.gao@linaro.org> --- include/wd_comp.h | 2 ++ libwd_comp.map | 1 + wd_comp.c | 21 +++++++++++++++++++++ 3 files changed, 24 insertions(+) diff --git a/include/wd_comp.h b/include/wd_comp.h index 45994ff..a957021 100644 --- a/include/wd_comp.h +++ b/include/wd_comp.h @@ -10,6 +10,7 @@ #include <numa.h> #include "wd_alg_common.h" +#include "wd_bmm.h" #ifdef __cplusplus extern "C" { @@ -256,6 +257,7 @@ void wd_comp_ctx_num_uninit(void); int wd_comp_get_env_param(__u32 node, __u32 type, __u32 mode, __u32 *num, __u8 *is_enable); +void *wd_comp_setup_blkpool(struct wd_blkpool_setup *setup); #ifdef __cplusplus } #endif diff --git a/libwd_comp.map b/libwd_comp.map index 6b1f8c2..033b476 100644 --- a/libwd_comp.map +++ b/libwd_comp.map @@ -22,6 +22,7 @@ global: wd_comp_get_driver; wd_comp_get_msg; wd_comp_reset_sess; + wd_comp_setup_blkpool; wd_sched_rr_instance; wd_sched_rr_alloc; diff --git a/wd_comp.c b/wd_comp.c index 647c320..692abf4 100644 --- a/wd_comp.c +++ b/wd_comp.c @@ -588,6 +588,27 @@ static int wd_comp_check_params(struct wd_comp_sess *sess, return 0; } +void *wd_comp_setup_blkpool(struct wd_blkpool_setup *setup) +{ + struct wd_ctx_config_internal *config = &wd_comp_setting.config; + struct wd_ctx_internal *ctx = config->ctxs; + int ret; + + if (setup->block_size < HW_CTX_SIZE) + setup->block_size = HW_CTX_SIZE; + + ret = wd_blkpool_setup(ctx->blkpool, setup); + if (ret) + return NULL; + + ctx->blkpool_mode = BLKPOOL_MODE_USER; + pthread_spin_lock(&ctx->lock); + if (ctx->h_sgl_pool == 0) + ctx->h_sgl_pool = wd_blkpool_create_sglpool(ctx->blkpool); + pthread_spin_unlock(&ctx->lock); + return ctx->blkpool; +} + static int wd_comp_sync_job(struct wd_comp_sess *sess, struct wd_comp_req *req, struct wd_comp_msg *msg) -- 2.25.1

sync: mempcy mode ./uadk_tool/uadk_tool benchmark --alg zlib --mode sva \ --opt 0 --sync --pktlen 1024 user pointer mode, --user ./uadk_tool/uadk_tool benchmark --alg zlib --mode sva \ --user --opt 0 --sync --pktlen 1024 sgl mode, --sgl ./uadk_tool/uadk_tool benchmark --alg zlib --mode sva \ --sgl --opt 0 --sync --pktlen 1024 async: memcpy mode ./uadk_tool/uadk_tool benchmark --alg zlib --mode sva \ --opt 0 --async --pktlen 1024 user pointer mode, --user ./uadk_tool/uadk_tool benchmark --alg zlib --mode sva \ --user --opt 0 --async --pktlen 1024 sgl mode, --sgl ./uadk_tool/uadk_tool benchmark --alg zlib --mode sva \ --sgl --opt 0 --async --pktlen 1024 Signed-off-by: Zhangfei Gao <zhangfei.gao@linaro.org> --- uadk_tool/benchmark/uadk_benchmark.c | 8 + uadk_tool/benchmark/uadk_benchmark.h | 2 + uadk_tool/benchmark/zip_uadk_benchmark.c | 249 ++++++++++++++++++++--- 3 files changed, 226 insertions(+), 33 deletions(-) diff --git a/uadk_tool/benchmark/uadk_benchmark.c b/uadk_tool/benchmark/uadk_benchmark.c index 1698061..064affc 100644 --- a/uadk_tool/benchmark/uadk_benchmark.c +++ b/uadk_tool/benchmark/uadk_benchmark.c @@ -718,6 +718,8 @@ int acc_cmd_parse(int argc, char *argv[], struct acc_option *option) {"complevel", required_argument, 0, 16}, {"init2", no_argument, 0, 17}, {"device", required_argument, 0, 18}, + {"user", no_argument, 0, 19}, + {"sgl", no_argument, 0, 20}, {0, 0, 0, 0} }; @@ -789,6 +791,12 @@ int acc_cmd_parse(int argc, char *argv[], struct acc_option *option) } strcpy(option->device, optarg); break; + case 19: + option->user = true; + break; + case 20: + option->sgl = true; + break; default: ACC_TST_PRT("invalid: bad input parameter!\n"); print_benchmark_help(); diff --git a/uadk_tool/benchmark/uadk_benchmark.h b/uadk_tool/benchmark/uadk_benchmark.h index 2739a0e..7fde341 100644 --- a/uadk_tool/benchmark/uadk_benchmark.h +++ b/uadk_tool/benchmark/uadk_benchmark.h @@ -81,6 +81,8 @@ struct acc_option { bool latency; u32 sched_type; int task_type; + bool user; + bool sgl; }; enum acc_type { diff --git a/uadk_tool/benchmark/zip_uadk_benchmark.c b/uadk_tool/benchmark/zip_uadk_benchmark.c index cad8016..cfb1a54 100644 --- a/uadk_tool/benchmark/zip_uadk_benchmark.c +++ b/uadk_tool/benchmark/zip_uadk_benchmark.c @@ -8,6 +8,8 @@ #include "include/wd_sched.h" #include "include/fse.h" +#define HW_CTX_SIZE (64 * 1024) + #define ZIP_TST_PRT printf #define PATH_SIZE 64 #define ZIP_FILE "./zip" @@ -22,6 +24,8 @@ struct uadk_bd { u8 *dst; u32 src_len; u32 dst_len; + void *pool_src; + void *pool_dst; }; struct bd_pool { @@ -32,6 +36,8 @@ struct thread_pool { struct bd_pool *pool; } g_zip_pool; +void *g_blkpool; + enum ZIP_OP_MODE { BLOCK_MODE, STREAM_MODE @@ -62,6 +68,7 @@ typedef struct uadk_thread_res { struct zip_async_tag *tag; COMP_TUPLE_TAG *ftuple; char *hw_buff_out; + bool sgl; } thread_data; struct zip_file_head { @@ -145,7 +152,11 @@ static int save_file_data(const char *alg, u32 pkg_len, u32 optype) // write data for one buffer one buffer to file line. for (j = 0; j < MAX_POOL_LENTH_COMP; j++) { - size = write(fd, g_zip_pool.pool[0].bds[j].dst, + if (g_blkpool) + size = write(fd, g_zip_pool.pool[0].bds[j].pool_dst, + fhead->blk_sz[j]); + else + size = write(fd, g_zip_pool.pool[0].bds[j].dst, fhead->blk_sz[j]); if (size < 0) { ZIP_TST_PRT("compress write data error size: %lu!\n", size); @@ -297,6 +308,17 @@ static void uninit_ctx_config2(void) wd_comp_uninit2(); } +static void init_blkpool(struct acc_option *options) +{ + struct wd_blkpool_setup setup; + + memset(&setup, 0, sizeof(setup)); + setup.block_size = HW_CTX_SIZE; + setup.block_num = DEFAULT_BLOCK_NM; + setup.align_size = DEFAULT_ALIGN_SIZE; + g_blkpool = wd_comp_setup_blkpool(&setup); +} + static int init_ctx_config2(struct acc_option *options) { struct wd_ctx_params cparams = {0}; @@ -333,6 +355,10 @@ static int init_ctx_config2(struct acc_option *options) ZIP_TST_PRT("failed to do comp init2!\n"); free(ctx_set_num); + + if (options->user || options->sgl) + init_blkpool(options); + return ret; } @@ -506,6 +532,8 @@ static int init_ctx_config(struct acc_option *options) goto free_sched; } + if (options->user || options->sgl) + init_blkpool(options); return 0; free_sched: @@ -614,6 +642,13 @@ static void free_uadk_bd_pool(void) for (j = 0; j < MAX_POOL_LENTH_COMP; j++) { free(g_zip_pool.pool[i].bds[j].src); free(g_zip_pool.pool[i].bds[j].dst); + + if (g_blkpool) { + wd_blkpool_free(g_blkpool, + g_zip_pool.pool[i].bds[j].pool_src); + wd_blkpool_free(g_blkpool, + g_zip_pool.pool[i].bds[j].pool_dst); + } } } free(g_zip_pool.pool[i].bds); @@ -975,6 +1010,7 @@ static void *zip_uadk_blk_sync_run(void *arg) { thread_data *pdata = (thread_data *)arg; struct wd_comp_sess_setup comp_setup = {0}; + struct wd_datalist *list_src = NULL, *list_dst = NULL; struct bd_pool *uadk_pool; struct wd_comp_req creq; handle_t h_sess; @@ -1003,23 +1039,71 @@ static void *zip_uadk_blk_sync_run(void *arg) out_len = uadk_pool->bds[0].dst_len; creq.cb = NULL; - creq.data_fmt = 0; + if (pdata->sgl) + creq.data_fmt = WD_SGL_BUF; + else + creq.data_fmt = 0; creq.priv = 0; creq.status = 0; + if (pdata->sgl) { + struct wd_datalist *src, *dst; + + list_src = calloc(MAX_POOL_LENTH_COMP, sizeof(struct wd_datalist)); + list_dst = calloc(MAX_POOL_LENTH_COMP, sizeof(struct wd_datalist)); + + for (i = 0; i < MAX_POOL_LENTH_COMP; i++) { + src = &list_src[i]; + dst = &list_dst[i]; + if (g_blkpool) { + src->data = uadk_pool->bds[i].pool_src; + dst->data = uadk_pool->bds[i].pool_dst; + } else { + src->data = uadk_pool->bds[i].src; + dst->data = uadk_pool->bds[i].dst; + } + src->len = uadk_pool->bds[i].src_len; + dst->len = uadk_pool->bds[i].dst_len; + src->next = (i < MAX_POOL_LENTH_COMP-1) ? &list_src[i+1] : NULL; + dst->next = (i < MAX_POOL_LENTH_COMP-1) ? &list_dst[i+1] : NULL; + } + } + while(1) { - i = count % MAX_POOL_LENTH_COMP; - creq.src = uadk_pool->bds[i].src; - creq.dst = uadk_pool->bds[i].dst; - creq.src_len = uadk_pool->bds[i].src_len; - creq.dst_len = out_len; + if (pdata->sgl) { + creq.list_src = list_src; + creq.list_dst = list_dst; - ret = wd_do_comp_sync(h_sess, &creq); - if (ret || creq.status) - break; + creq.src_len = uadk_pool->bds[0].src_len * MAX_POOL_LENTH_COMP; + creq.dst_len = out_len * MAX_POOL_LENTH_COMP; - count++; - uadk_pool->bds[i].dst_len = creq.dst_len; + ret = wd_do_comp_sync(h_sess, &creq); + if (ret || creq.status) + break; + count++; + uadk_pool->bds[0].dst_len = creq.dst_len; + if (get_run_state() == 0) + break; + + } else { + i = count % MAX_POOL_LENTH_COMP; + if (g_blkpool) { + creq.src = uadk_pool->bds[i].pool_src; + creq.dst = uadk_pool->bds[i].pool_dst; + } else { + creq.src = uadk_pool->bds[i].src; + creq.dst = uadk_pool->bds[i].dst; + } + creq.src_len = uadk_pool->bds[i].src_len; + creq.dst_len = out_len; + + ret = wd_do_comp_sync(h_sess, &creq); + if (ret || creq.status) + break; + + count++; + uadk_pool->bds[i].dst_len = creq.dst_len; + } if (get_run_state() == 0) break; } @@ -1028,6 +1112,10 @@ static void *zip_uadk_blk_sync_run(void *arg) cal_avg_latency(count); add_recv_data(count, g_pktlen); + if (pdata->sgl) { + free(list_src); + free(list_dst); + } return NULL; } @@ -1099,6 +1187,7 @@ static void *zip_uadk_blk_async_run(void *arg) { thread_data *pdata = (thread_data *)arg; struct wd_comp_sess_setup comp_setup = {0}; + struct wd_datalist *list_src = NULL, *list_dst = NULL; struct bd_pool *uadk_pool; struct wd_comp_req creq; handle_t h_sess; @@ -1128,35 +1217,94 @@ static void *zip_uadk_blk_async_run(void *arg) out_len = uadk_pool->bds[0].dst_len; creq.cb = zip_async_cb; - creq.data_fmt = 0; + if (pdata->sgl) + creq.data_fmt = WD_SGL_BUF; + else + creq.data_fmt = 0; creq.priv = 0; creq.status = 0; + if (pdata->sgl) { + struct wd_datalist *src, *dst; + + list_src = calloc(MAX_POOL_LENTH_COMP, sizeof(struct wd_datalist)); + list_dst = calloc(MAX_POOL_LENTH_COMP, sizeof(struct wd_datalist)); + + for (i = 0; i < MAX_POOL_LENTH_COMP; i++) { + src = &list_src[i]; + dst = &list_dst[i]; + if (g_blkpool) { + src->data = uadk_pool->bds[i].pool_src; + dst->data = uadk_pool->bds[i].pool_dst; + } else { + src->data = uadk_pool->bds[i].src; + dst->data = uadk_pool->bds[i].dst; + } + src->len = uadk_pool->bds[i].src_len; + dst->len = uadk_pool->bds[i].dst_len; + src->next = (i < MAX_POOL_LENTH_COMP-1) ? &list_src[i+1] : NULL; + dst->next = (i < MAX_POOL_LENTH_COMP-1) ? &list_dst[i+1] : NULL; + } + } + while(1) { if (get_run_state() == 0) break; - i = count % MAX_POOL_LENTH_COMP; - creq.src = uadk_pool->bds[i].src; - creq.dst = uadk_pool->bds[i].dst; - creq.src_len = uadk_pool->bds[i].src_len; - creq.dst_len = out_len; - - pdata->tag[i].td_id = pdata->td_id; - pdata->tag[i].bd_idx = i; - creq.cb_param = &pdata->tag[i]; - - ret = wd_do_comp_async(h_sess, &creq); - if (ret == -WD_EBUSY) { - usleep(SEND_USLEEP * try_cnt); - try_cnt++; - if (try_cnt > MAX_TRY_CNT) { - ZIP_TST_PRT("Test compress send fail %d times!\n", MAX_TRY_CNT); - try_cnt = 0; + if (pdata->sgl) { + creq.list_src = list_src; + creq.list_dst = list_dst; + + creq.src_len = uadk_pool->bds[0].src_len * MAX_POOL_LENTH_COMP; + creq.dst_len = out_len * MAX_POOL_LENTH_COMP; + + pdata->tag[0].td_id = pdata->td_id; + pdata->tag[0].bd_idx = i; + creq.cb_param = &pdata->tag[0]; + + ret = wd_do_comp_async(h_sess, &creq); + if (ret == -WD_EBUSY) { + usleep(SEND_USLEEP * try_cnt); + try_cnt++; + if (try_cnt > MAX_TRY_CNT) { + ZIP_TST_PRT("Test compress send fail %d times!\n", + MAX_TRY_CNT); + try_cnt = 0; + } + continue; + } else if (ret || creq.status) { + break; + } + uadk_pool->bds[0].dst_len = creq.dst_len; + } else { + i = count % MAX_POOL_LENTH_COMP; + if (g_blkpool) { + creq.src = uadk_pool->bds[i].pool_src; + creq.dst = uadk_pool->bds[i].pool_dst; + } else { + creq.src = uadk_pool->bds[i].src; + creq.dst = uadk_pool->bds[i].dst; + } + creq.src_len = uadk_pool->bds[i].src_len; + creq.dst_len = out_len; + + pdata->tag[i].td_id = pdata->td_id; + pdata->tag[i].bd_idx = i; + creq.cb_param = &pdata->tag[i]; + + ret = wd_do_comp_async(h_sess, &creq); + if (ret == -WD_EBUSY) { + usleep(SEND_USLEEP * try_cnt); + try_cnt++; + if (try_cnt > MAX_TRY_CNT) { + ZIP_TST_PRT("Test compress send fail %d times!\n", + MAX_TRY_CNT); + try_cnt = 0; + } + continue; + } else if (ret || creq.status) { + break; } - continue; - } else if (ret || creq.status) { - break; } try_cnt = 0; count++; @@ -1202,6 +1350,7 @@ static int zip_uadk_sync_threads(struct acc_option *options) threads_args[i].optype = threads_option.optype; threads_args[i].win_sz = threads_option.win_sz; threads_args[i].comp_lv = threads_option.comp_lv; + threads_args[i].sgl = options->sgl; threads_args[i].td_id = i; ret = pthread_create(&tdid[i], NULL, uadk_zip_sync_run, &threads_args[i]); if (ret) { @@ -1293,6 +1442,7 @@ static int zip_uadk_async_threads(struct acc_option *options) } threads_args[i].tag->recv_cnt = 0; threads_args[i].send_cnt = 0; + threads_args[i].sgl = options->sgl; ret = pthread_create(&tdid[i], NULL, uadk_zip_async_run, &threads_args[i]); if (ret) { ZIP_TST_PRT("Create async thread fail!\n"); @@ -1344,6 +1494,35 @@ async_error: return ret; } +static int load_blkpool_data(void) +{ + int i, j; + int src_len, dst_len; + + if (!g_blkpool) + return 0; + + for (i = 0; i < g_thread_num; i++) { + for (j = 0; j < MAX_POOL_LENTH_COMP; j++) { + src_len = g_zip_pool.pool[i].bds[j].src_len; + g_zip_pool.pool[i].bds[j].pool_src = + wd_blkpool_alloc(g_blkpool, src_len); + + dst_len = g_zip_pool.pool[i].bds[j].dst_len; + g_zip_pool.pool[i].bds[j].pool_dst = + wd_blkpool_alloc(g_blkpool, dst_len); + + if (!g_zip_pool.pool[i].bds[j].pool_src || + !g_zip_pool.pool[i].bds[j].pool_dst) + return -EINVAL; + + memcpy(g_zip_pool.pool[i].bds[j].pool_src, + g_zip_pool.pool[0].bds[j].src, src_len); + } + } + return 0; +} + int zip_uadk_benchmark(struct acc_option *options) { u32 ptime; @@ -1375,6 +1554,10 @@ int zip_uadk_benchmark(struct acc_option *options) if (ret) return ret; + ret = load_blkpool_data(); + if (ret) + return ret; + get_pid_cpu_time(&ptime); time_start(options->times); if (options->syncmode) -- 2.25.1

test: digest: uadk_tool benchmark --alg aes-128-ecb --mode sva --opt 0 \ --sync --pktlen 1024 --seconds 1 --multi 1 --thread 1 uadk_tool benchmark --alg aes-128-ecb --mode sva --opt 0 \ --async --pktlen 1024 --seconds 1 --multi 1 --thread 1 cipher: uadk_tool benchmark --alg sm4-128-cbc --mode sva --opt 0 \ --sync --pktlen 1024 --seconds 1 --multi 1 --thread 1 uadk_tool benchmark --alg sm4-128-cbc --mode sva --opt 0 \ --async --pktlen 1024 --seconds 1 --multi 1 --thread 1 Signed-off-by: Zhangfei Gao <zhangfei.gao@linaro.org> --- drv/hisi_sec.c | 57 ++++++++++++++++++++++++++----------- include/drv/wd_cipher_drv.h | 2 ++ include/drv/wd_digest_drv.h | 2 ++ include/wd_cipher.h | 2 ++ include/wd_digest.h | 2 ++ libwd_crypto.map | 2 ++ wd_cipher.c | 52 +++++++++++++++++++++++++++++++-- wd_digest.c | 45 +++++++++++++++++++++++++++-- 8 files changed, 142 insertions(+), 22 deletions(-) diff --git a/drv/hisi_sec.c b/drv/hisi_sec.c index a305985..1e5d7b2 100644 --- a/drv/hisi_sec.c +++ b/drv/hisi_sec.c @@ -916,10 +916,17 @@ static int fill_cipher_bd2_mode(struct wd_cipher_msg *msg, static void fill_cipher_bd2_addr(struct wd_cipher_msg *msg, struct hisi_sec_sqe *sqe) { - sqe->type2.data_src_addr = (__u64)(uintptr_t)msg->in; - sqe->type2.data_dst_addr = (__u64)(uintptr_t)msg->out; - sqe->type2.c_ivin_addr = (__u64)(uintptr_t)msg->iv; - sqe->type2.c_key_addr = (__u64)(uintptr_t)msg->key; + if (msg->blkpool) { + sqe->type2.data_src_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->in); + sqe->type2.data_dst_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->out); + sqe->type2.c_ivin_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->iv); + sqe->type2.c_key_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->key); + } else { + sqe->type2.data_src_addr = (__u64)(uintptr_t)msg->in; + sqe->type2.data_dst_addr = (__u64)(uintptr_t)msg->out; + sqe->type2.c_ivin_addr = (__u64)(uintptr_t)msg->iv; + sqe->type2.c_key_addr = (__u64)(uintptr_t)msg->key; + } } static void parse_cipher_bd2(struct hisi_qp *qp, struct hisi_sec_sqe *sqe, @@ -944,8 +951,6 @@ static void parse_cipher_bd2(struct hisi_qp *qp, struct hisi_sec_sqe *sqe, if (qp->q_info.qp_mode == CTX_MODE_ASYNC) { recv_msg->alg_type = WD_CIPHER; recv_msg->data_fmt = get_data_fmt_v2(sqe->sds_sa_type); - recv_msg->in = (__u8 *)(uintptr_t)sqe->type2.data_src_addr; - recv_msg->out = (__u8 *)(uintptr_t)sqe->type2.data_dst_addr; temp_msg = wd_cipher_get_msg(qp->q_info.idx, tag); if (!temp_msg) { recv_msg->result = WD_IN_EPARA; @@ -953,6 +958,8 @@ static void parse_cipher_bd2(struct hisi_qp *qp, struct hisi_sec_sqe *sqe, qp->q_info.idx, tag); return; } + recv_msg->in = temp_msg->in; + recv_msg->out = temp_msg->out; } else { /* The synchronization mode uses the same message */ temp_msg = recv_msg; @@ -1318,10 +1325,17 @@ static int fill_cipher_bd3_mode(struct wd_cipher_msg *msg, static void fill_cipher_bd3_addr(struct wd_cipher_msg *msg, struct hisi_sec_sqe3 *sqe) { - sqe->data_src_addr = (__u64)(uintptr_t)msg->in; - sqe->data_dst_addr = (__u64)(uintptr_t)msg->out; - sqe->no_scene.c_ivin_addr = (__u64)(uintptr_t)msg->iv; - sqe->c_key_addr = (__u64)(uintptr_t)msg->key; + if (msg->blkpool) { + sqe->data_src_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->in); + sqe->data_dst_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->out); + sqe->no_scene.c_ivin_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->iv); + sqe->c_key_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->key); + } else { + sqe->data_src_addr = (__u64)(uintptr_t)msg->in; + sqe->data_dst_addr = (__u64)(uintptr_t)msg->out; + sqe->no_scene.c_ivin_addr = (__u64)(uintptr_t)msg->iv; + sqe->c_key_addr = (__u64)(uintptr_t)msg->key; + } } static int fill_cipher_bd3(struct wd_cipher_msg *msg, struct hisi_sec_sqe3 *sqe) @@ -1429,8 +1443,6 @@ static void parse_cipher_bd3(struct hisi_qp *qp, struct hisi_sec_sqe3 *sqe, if (qp->q_info.qp_mode == CTX_MODE_ASYNC) { recv_msg->alg_type = WD_CIPHER; recv_msg->data_fmt = get_data_fmt_v3(sqe->bd_param); - recv_msg->in = (__u8 *)(uintptr_t)sqe->data_src_addr; - recv_msg->out = (__u8 *)(uintptr_t)sqe->data_dst_addr; temp_msg = wd_cipher_get_msg(qp->q_info.idx, tag); if (!temp_msg) { recv_msg->result = WD_IN_EPARA; @@ -1438,6 +1450,8 @@ static void parse_cipher_bd3(struct hisi_qp *qp, struct hisi_sec_sqe3 *sqe, qp->q_info.idx, tag); return; } + recv_msg->in = temp_msg->in; + recv_msg->out = temp_msg->out; } else { /* The synchronization mode uses the same message */ temp_msg = recv_msg; @@ -1597,7 +1611,6 @@ static void parse_digest_bd2(struct hisi_qp *qp, struct hisi_sec_sqe *sqe, if (qp->q_info.qp_mode == CTX_MODE_ASYNC) { recv_msg->alg_type = WD_DIGEST; recv_msg->data_fmt = get_data_fmt_v2(sqe->sds_sa_type); - recv_msg->in = (__u8 *)(uintptr_t)sqe->type2.data_src_addr; temp_msg = wd_digest_get_msg(qp->q_info.idx, recv_msg->tag); if (!temp_msg) { recv_msg->result = WD_IN_EPARA; @@ -1742,8 +1755,13 @@ static int hisi_sec_digest_send(struct wd_alg_driver *drv, handle_t ctx, void *w sqe.sds_sa_type |= (__u8)(de | scene); sqe.type2.alen_ivllen |= (__u32)msg->in_bytes; - sqe.type2.data_src_addr = (__u64)(uintptr_t)msg->in; - sqe.type2.mac_addr = (__u64)(uintptr_t)msg->out; + if (msg->blkpool) { + sqe.type2.data_src_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->in); + sqe.type2.mac_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->out); + } else { + sqe.type2.data_src_addr = (__u64)(uintptr_t)msg->in; + sqe.type2.mac_addr = (__u64)(uintptr_t)msg->out; + } ret = fill_digest_bd2_alg(msg, &sqe); if (ret) @@ -1977,8 +1995,13 @@ static int hisi_sec_digest_send_v3(struct wd_alg_driver *drv, handle_t ctx, void } sqe.a_len_key = (__u32)msg->in_bytes; - sqe.data_src_addr = (__u64)(uintptr_t)msg->in; - sqe.mac_addr = (__u64)(uintptr_t)msg->out; + if (msg->blkpool) { + sqe.data_src_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->in); + sqe.mac_addr = (__u64)(uintptr_t)wd_blkpool_phy(msg->blkpool, msg->out); + } else { + sqe.data_src_addr = (__u64)(uintptr_t)msg->in; + sqe.mac_addr = (__u64)(uintptr_t)msg->out; + } ret = fill_digest_bd3_alg(msg, &sqe); if (ret) diff --git a/include/drv/wd_cipher_drv.h b/include/drv/wd_cipher_drv.h index c6d8ddf..b0da45a 100644 --- a/include/drv/wd_cipher_drv.h +++ b/include/drv/wd_cipher_drv.h @@ -48,6 +48,8 @@ struct wd_cipher_msg { __u8 *in; /* output data pointer */ __u8 *out; + void *blkpool; + handle_t h_sgl_pool; }; struct wd_cipher_msg *wd_cipher_get_msg(__u32 idx, __u32 tag); diff --git a/include/drv/wd_digest_drv.h b/include/drv/wd_digest_drv.h index a55ef5b..5e693ae 100644 --- a/include/drv/wd_digest_drv.h +++ b/include/drv/wd_digest_drv.h @@ -59,6 +59,8 @@ struct wd_digest_msg { __u8 *partial_block; /* total of data for stream mode */ __u64 long_data_len; + void *blkpool; + handle_t h_sgl_pool; }; static inline enum hash_block_type get_hash_block_type(struct wd_digest_msg *msg) diff --git a/include/wd_cipher.h b/include/wd_cipher.h index d54f7fe..acb99a5 100644 --- a/include/wd_cipher.h +++ b/include/wd_cipher.h @@ -10,6 +10,7 @@ #include <dlfcn.h> #include <asm/types.h> #include "wd_alg_common.h" +#include "wd_bmm.h" #ifdef __cplusplus extern "C" { @@ -220,6 +221,7 @@ void wd_cipher_ctx_num_uninit(void); int wd_cipher_get_env_param(__u32 node, __u32 type, __u32 mode, __u32 *num, __u8 *is_enable); +void *wd_cipher_setup_blkpool(struct wd_blkpool_setup *setup); #ifdef __cplusplus } #endif diff --git a/include/wd_digest.h b/include/wd_digest.h index 6ce31f2..e130ea0 100644 --- a/include/wd_digest.h +++ b/include/wd_digest.h @@ -9,6 +9,7 @@ #include <dlfcn.h> #include "wd_alg_common.h" +#include "wd_bmm.h" #ifdef __cplusplus extern "C" { @@ -281,6 +282,7 @@ void wd_digest_ctx_num_uninit(void); int wd_digest_get_env_param(__u32 node, __u32 type, __u32 mode, __u32 *num, __u8 *is_enable); +void *wd_digest_setup_blkpool(struct wd_blkpool_setup *setup); #ifdef __cplusplus } #endif diff --git a/libwd_crypto.map b/libwd_crypto.map index e8555c9..b430886 100644 --- a/libwd_crypto.map +++ b/libwd_crypto.map @@ -20,6 +20,7 @@ global: wd_cipher_set_driver; wd_cipher_get_driver; wd_cipher_get_msg; + wd_cipher_setup_blkpool; wd_aead_init; wd_aead_uninit; @@ -66,6 +67,7 @@ global: wd_digest_set_driver; wd_digest_get_driver; wd_digest_get_msg; + wd_digest_setup_blkpool; wd_rsa_is_crt; wd_rsa_get_key_bits; diff --git a/wd_cipher.c b/wd_cipher.c index 95d6eb6..4b0e637 100644 --- a/wd_cipher.c +++ b/wd_cipher.c @@ -66,6 +66,9 @@ struct wd_cipher_sess { unsigned char key[MAX_CIPHER_KEY_SIZE]; __u32 key_bytes; void *sched_key; + void *blkpool; + unsigned char *pool_key; + handle_t h_sgl_pool; }; struct wd_env_config wd_cipher_env_config; @@ -244,13 +247,36 @@ int wd_cipher_set_key(handle_t h_sess, const __u8 *key, __u32 key_len) } sess->key_bytes = key_len; - memcpy(sess->key, key, key_len); + if (sess->pool_key) + memcpy(sess->pool_key, key, key_len); + else + memcpy(sess->key, key, key_len); return 0; } +void *wd_cipher_setup_blkpool(struct wd_blkpool_setup *setup) +{ + struct wd_ctx_config_internal *config = &wd_cipher_setting.config; + struct wd_ctx_internal *ctx = config->ctxs; + int ret; + + ret = wd_blkpool_setup(ctx->blkpool, setup); + if (ret) + return NULL; + + ctx->blkpool_mode = BLKPOOL_MODE_USER; + pthread_spin_lock(&ctx->lock); + if (ctx->h_sgl_pool == 0) + ctx->h_sgl_pool = wd_blkpool_create_sglpool(ctx->blkpool); + pthread_spin_unlock(&ctx->lock); + return ctx->blkpool; +} + handle_t wd_cipher_alloc_sess(struct wd_cipher_sess_setup *setup) { + struct wd_ctx_config_internal *config = &wd_cipher_setting.config; + struct wd_ctx_internal *ctx = config->ctxs; struct wd_cipher_sess *sess = NULL; bool ret; @@ -289,6 +315,15 @@ handle_t wd_cipher_alloc_sess(struct wd_cipher_sess_setup *setup) goto free_sess; } + if (ctx->blkpool) { + sess->blkpool = ctx->blkpool; + sess->h_sgl_pool = ctx->h_sgl_pool; + + sess->pool_key = wd_blkpool_alloc(sess->blkpool, MAX_CIPHER_KEY_SIZE); + if (!sess->pool_key) + goto free_sess; + } + return (handle_t)sess; free_sess: @@ -539,7 +574,10 @@ static void fill_request_msg(struct wd_cipher_msg *msg, msg->in_bytes = req->in_bytes; msg->out = req->dst; msg->out_bytes = req->out_bytes; - msg->key = sess->key; + if (sess->blkpool) + msg->key = sess->pool_key; + else + msg->key = sess->key; msg->key_bytes = sess->key_bytes; msg->iv = req->iv; msg->iv_bytes = req->iv_bytes; @@ -719,6 +757,11 @@ int wd_do_cipher_sync(handle_t h_sess, struct wd_cipher_req *req) wd_dfx_msg_cnt(config, WD_CTX_CNT_NUM, idx); ctx = config->ctxs + idx; + if (sess->blkpool) { + msg.blkpool = sess->blkpool; + msg.h_sgl_pool = sess->h_sgl_pool; + } + ret = send_recv_sync(ctx, &msg); req->state = msg.result; @@ -759,6 +802,11 @@ int wd_do_cipher_async(handle_t h_sess, struct wd_cipher_req *req) fill_request_msg(msg, req, sess); msg->tag = msg_id; + if (sess->blkpool) { + msg->blkpool = sess->blkpool; + msg->h_sgl_pool = sess->h_sgl_pool; + } + ret = wd_alg_driver_send(wd_cipher_setting.driver, ctx->ctx, msg); if (unlikely(ret < 0)) { if (ret != -WD_EBUSY) diff --git a/wd_digest.c b/wd_digest.c index 4c1b0ca..696f1b0 100644 --- a/wd_digest.c +++ b/wd_digest.c @@ -68,6 +68,9 @@ struct wd_digest_sess { __u32 key_bytes; void *sched_key; struct wd_digest_stream_data stream_data; + void *blkpool; + unsigned char *pool_key; + handle_t h_sgl_pool; }; struct wd_env_config wd_digest_env_config; @@ -180,14 +183,38 @@ int wd_digest_set_key(handle_t h_sess, const __u8 *key, __u32 key_len) } sess->key_bytes = key_len; - if (key_len) - memcpy(sess->key, key, key_len); + if (key_len) { + if (sess->pool_key) + memcpy(sess->pool_key, key, key_len); + else + memcpy(sess->key, key, key_len); + } return 0; } +void *wd_digest_setup_blkpool(struct wd_blkpool_setup *setup) +{ + struct wd_ctx_config_internal *config = &wd_digest_setting.config; + struct wd_ctx_internal *ctx = config->ctxs; + int ret; + + ret = wd_blkpool_setup(ctx->blkpool, setup); + if (ret) + return NULL; + + ctx->blkpool_mode = BLKPOOL_MODE_USER; + pthread_spin_lock(&ctx->lock); + if (ctx->h_sgl_pool == 0) + ctx->h_sgl_pool = wd_blkpool_create_sglpool(ctx->blkpool); + pthread_spin_unlock(&ctx->lock); + return ctx->blkpool; +} + handle_t wd_digest_alloc_sess(struct wd_digest_sess_setup *setup) { + struct wd_ctx_config_internal *config = &wd_digest_setting.config; + struct wd_ctx_internal *ctx = config->ctxs; struct wd_digest_sess *sess = NULL; bool ret; @@ -222,6 +249,15 @@ handle_t wd_digest_alloc_sess(struct wd_digest_sess_setup *setup) goto err_sess; } + if (ctx->blkpool) { + sess->blkpool = ctx->blkpool; + sess->h_sgl_pool = ctx->h_sgl_pool; + + sess->pool_key = wd_blkpool_alloc(sess->blkpool, MAX_HMAC_KEY_SIZE); + if (!sess->pool_key) + goto err_sess; + } + return (handle_t)sess; err_sess: @@ -589,7 +625,10 @@ static void fill_request_msg(struct wd_digest_msg *msg, msg->alg_type = WD_DIGEST; msg->alg = sess->alg; msg->mode = sess->mode; - msg->key = sess->key; + if (sess->blkpool) + msg->key = sess->pool_key; + else + msg->key = sess->key; msg->key_bytes = sess->key_bytes; msg->iv = req->iv; msg->in = req->in; -- 2.25.1

在 2025/8/26 11:21, Zhangfei Gao 写道:

这个方案3 我有点忘了

[PATCH 4/4] uacce: ensure safe queue release with state management

这个是解决这个问题的么

---------- Forwarded message --------- From: Zhangfei Gao <zhangfei.gao@linaro.org> Date: Mon, 21 Oct 2024 at 17:04 Subject: uacce ioctl UACCE_CMD_PUT_Q To: Yang Shen <shenyang39@huawei.com>, Hao Fang <fanghao11@huawei.com>, liulongfang <liulongfang@huawei.com>

讨论个uacce ioctl 问题： 进程1 put_q 之后，还没unmap 进程2 get_q 获取到这个q，mmap, 可能有风险访问/修改到进程1 映射到的内容

如何解决 方案1: uacce_put_queue + if (q->mapping) { + unmap_mapping_range(q->mapping, 0, 0, 1); + q->mapping = NULL; + }

这个方案不行，会同时unmap掉其他共享该filep->f_mappig的映射 比如多个线程，打开同一个设备，尽管fd/filep不同，但filep->f_mappig相同， 做了实验会同时unmap所有的映射。

hack了一个测试程序，2 threads，2个ctx 然后释放其中一个ctx，put_queue, 另一个也被调用了，然后core dump. https://github.com/gaozhangfei/uadk/tree/hack-10.17

方案2 想办法put_queue只unmap该fd的mapping

2.1. 目前还不知道怎么实现，没找到类似例子，估计比较hack，因为remove_vma不对外开放 2.2 假设能实现，应用程序addr = mmap, 调用ioctl(put_q), 继续访问 addr, 挂了，这个能否被接受 今天vpn连不上，没做实验，不知道啥错误，估计类似segmentation fault 之类 我理解这个不能被接受，use after free, 也得驱动自己去检查避免，用户程序没有unmap, 继续访问addr, 不能算错吧。 并行性也可能有问题，比如多线程一个在操作addr，另一个来着么一下，也没法上锁。 kernel不开放remove_vma 肯定是有原因的。

方案3 修改ioctl uacce_put_queue, 把uacce->ops->put_queue(q)挪到release 这个方案比较简单，也没啥风险。 问题，1. 和 UACCE_CMD_PUT_Q语义不匹配，变成stop_q， 2 如何兼容 觉得只要解释清楚，语义不匹配也能接受吧， 考虑兼容，可以再加一个UACCE_CMD_STOP_Q，两个行为一致。

put_q 本身也只是一个优化

/* * UACCE_CMD_PUT_Q: * User actively stop queue and free queue resource immediately * Optimization method since close fd may delay */ #define UACCE_CMD_PUT_Q _IO('W', 1)

大家什么建议

也见过其他修改ioctl 的例子 1, https://lore.kernel.org/all/1-v3-e2e16cd7467f+2a6a1-smmuv3_nesting_jgg@nvidi... jason 之前修改过一个ioctl --- a/include/uapi/linux/vfio.h +++ b/include/uapi/linux/vfio.h -#define VFIO_TYPE1_NESTING_IOMMU 6 /* Implies v2 */ +#define __VFIO_RESERVED_TYPE1_NESTING_IOMMU 6 /* Implies v2 */

git log --oneline include/uapi/linux

2. commit 636119af94f2fbf3e4458be66a1bc740ba69ce6d Author: Jens Axboe <axboe@kernel.dk> Date: Sat Sep 14 08:51:15 2024 -0600 io_uring: rename "copy buffers" to "clone buffers"

Re: Fwd: uacce ioctl UACCE_CMD_PUT_Q 问题: 应用先调用mmap获取地址（p1），再调用remap获取地址p2，接下来调用munmap解除p1，然后对p2区域做读写操作 mremap后，munmap原地址会成功（导致释放vma），然后使用新remap后地址读写成功 复现方法： #define _GNU_SOURCE         /* See feature_test_macros(7) */ #include <stdio.h> #include <string.h> #include <stdint.h> #include <unistd.h> #include <inttypes.h> #include <sys/types.h> #include <sys/stat.h> #include <fcntl.h> #include <sys/mman.h> static const char *char_dev = "/dev/hisi_sec2-0"; static const uint32_t size = 151552; static const uint32_t sleep_time = 10; #define RETURN_IF_CHECK_FAIL(exp) do { \         if (exp) { \                 printf("exit due to check failed, %u %s\n", __LINE__, #exp); \                 return; \         } \ } while(0) static void do_memset(void *addr, uint32_t size, const char *place) {         printf("memset for place: %s\n", place);         fflush(stdout);         memset(addr, 0x11, size); } void main() {         uint32_t remap_size = size / 2;         int fd = open(char_dev, O_RDWR | O_CLOEXEC);         RETURN_IF_CHECK_FAIL(fd < 0);         void *mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 1 * getpagesize());         RETURN_IF_CHECK_FAIL(mem == (void *)-1);         printf("mmap ret address = %p\n", mem);         do_memset(mem, size, "after mmap");         void *remap = mremap(mem, size, remap_size, MREMAP_MAYMOVE | MREMAP_FIXED, 0x400000000);         RETURN_IF_CHECK_FAIL(remap == (void *)-1);         printf("mremap ret address = %p\n", remap);         do_memset(remap, remap_size, "after remap");         // 实测访问再访问mem会出现段错误         printf("start munmap old!\n");         munmap(mem, size);         sleep(sleep_time);         printf("start munmap new!\n");         do_memset(remap, remap_size, "before munmap remap");         munmap(remap, size);         sleep(sleep_time);         printf("exit!\n"); } 此外在uacce加入打印以跟踪qfr： --- a/drivers/misc/uacce/uacce.c +++ b/drivers/misc/uacce/uacce.c @@ -189,6 +189,8 @@ static int uacce_fops_open(struct inode *inode, struct file *filep)         list_add(&q->list, &uacce->queues);         mutex_unlock(&uacce->mutex); +       printk("uacce_fops_open: q=%p\n", q); +         return 0;  out_with_bond: @@ -204,6 +206,8 @@ static int uacce_fops_release(struct inode *inode, struct file *filep)         struct uacce_queue *q = filep->private_data;         struct uacce_device *uacce = q->uacce; +       printk("uacce_fops_release: q=%p\n", q); +         mutex_lock(&uacce->mutex);         uacce_put_queue(q);         uacce_unbind_queue(q); @@ -221,6 +225,9 @@ static void uacce_vma_close(struct vm_area_struct *vma)         if (vma->vm_pgoff < UACCE_MAX_REGION) {                 struct uacce_qfile_region *qfr = q->qfrs[vma->vm_pgoff]; +               printk("uacce_vma_close: q=%p vm_pgoff=%u\n", q, vma->vm_pgoff); +               printk("uacce_vma_close: qfr=%p\n", qfr); +                 mutex_lock(&q->mutex);                 q->qfrs[vma->vm_pgoff] = NULL;                 mutex_unlock(&q->mutex); @@ -246,6 +253,8 @@ static int uacce_fops_mmap(struct file *filep, struct vm_area_struct *vma)         enum uacce_qfrt type = UACCE_MAX_REGION;         int ret = 0; +       printk("uacce_fops_mmap: q=%p vm_pgoff=%u\n", q, vma->vm_pgoff); +         if (vma->vm_pgoff < UACCE_MAX_REGION)                 type = vma->vm_pgoff;         else @@ -290,6 +299,7 @@ static int uacce_fops_mmap(struct file *filep, struct vm_area_struct *vma)         }         q->qfrs[type] = qfr; +       printk("uacce_fops_mmap: type=%u qfr=%p\n", type, qfr);         mutex_unlock(&q->mutex);         return ret;