From patchwork Tue May 11 11:22:36 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Weihang Li X-Patchwork-Id: 12250705 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-16.8 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,URIBL_BLOCKED,USER_AGENT_GIT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 65573C43460 for ; Tue, 11 May 2021 11:22:55 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id 3DC8F616E9 for ; Tue, 11 May 2021 11:22:55 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S231273AbhEKLYA (ORCPT ); Tue, 11 May 2021 07:24:00 -0400 Received: from szxga05-in.huawei.com ([45.249.212.191]:2624 "EHLO szxga05-in.huawei.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S230441AbhEKLX6 (ORCPT ); Tue, 11 May 2021 07:23:58 -0400 Received: from DGGEMS408-HUB.china.huawei.com (unknown [172.30.72.58]) by szxga05-in.huawei.com (SkyGuard) with ESMTP id 4Ffb6S2FqSzPwvH; Tue, 11 May 2021 19:19:28 +0800 (CST) Received: from localhost.localdomain (10.69.192.56) by DGGEMS408-HUB.china.huawei.com (10.3.19.208) with Microsoft SMTP Server id 14.3.498.0; Tue, 11 May 2021 19:22:42 +0800 From: Weihang Li To: , CC: , , , "Xi Wang" , Weihang Li Subject: [PATCH v2 for-next 2/7] RDMA/hns: Add method for shrinking DCA memory pool Date: Tue, 11 May 2021 19:22:36 +0800 Message-ID: <1620732161-27180-3-git-send-email-liweihang@huawei.com> X-Mailer: git-send-email 2.7.4 In-Reply-To: <1620732161-27180-1-git-send-email-liweihang@huawei.com> References: <1620732161-27180-1-git-send-email-liweihang@huawei.com> MIME-Version: 1.0 X-Originating-IP: [10.69.192.56] X-CFilter-Loop: Reflected Precedence: bulk List-ID: X-Mailing-List: linux-rdma@vger.kernel.org From: Xi Wang If no QP is using a DCA mem object, the userspace driver can destroy it. So add a new method 'HNS_IB_METHOD_DCA_MEM_SHRINK' to allow the userspace dirver to remove an object from DCA memory pool. If a DCA mem object has been shrunk, the userspace driver can destroy it by 'HNS_IB_METHOD_DCA_MEM_DEREG' method and free the buffer which is allocated in userspace. Signed-off-by: Xi Wang Signed-off-by: Weihang Li --- drivers/infiniband/hw/hns/hns_roce_dca.c | 142 ++++++++++++++++++++++++++++++- drivers/infiniband/hw/hns/hns_roce_dca.h | 7 ++ include/uapi/rdma/hns-abi.h | 9 ++ 3 files changed, 157 insertions(+), 1 deletion(-) diff --git a/drivers/infiniband/hw/hns/hns_roce_dca.c b/drivers/infiniband/hw/hns/hns_roce_dca.c index 2a03cf3..604d6cf 100644 --- a/drivers/infiniband/hw/hns/hns_roce_dca.c +++ b/drivers/infiniband/hw/hns/hns_roce_dca.c @@ -35,6 +35,11 @@ struct dca_mem_attr { u32 size; }; +static inline bool dca_page_is_free(struct hns_dca_page_state *state) +{ + return state->buf_id == HNS_DCA_INVALID_BUF_ID; +} + static inline bool dca_mem_is_free(struct dca_mem *mem) { return mem->flags == 0; @@ -60,6 +65,11 @@ static inline void clr_dca_mem_registered(struct dca_mem *mem) mem->flags &= ~DCA_MEM_FLAGS_REGISTERED; } +static inline bool dca_mem_is_available(struct dca_mem *mem) +{ + return mem->flags == (DCA_MEM_FLAGS_ALLOCED | DCA_MEM_FLAGS_REGISTERED); +} + static void free_dca_pages(void *pages) { ib_umem_release(pages); @@ -123,6 +133,41 @@ static struct hns_dca_page_state *alloc_dca_states(void *pages, int count) return states; } +#define DCA_MEM_STOP_ITERATE -1 +#define DCA_MEM_NEXT_ITERATE -2 +static void travel_dca_pages(struct hns_roce_dca_ctx *ctx, void *param, + int (*cb)(struct dca_mem *, int, void *)) +{ + struct dca_mem *mem, *tmp; + unsigned long flags; + bool avail; + int ret; + int i; + + spin_lock_irqsave(&ctx->pool_lock, flags); + list_for_each_entry_safe(mem, tmp, &ctx->pool, list) { + spin_unlock_irqrestore(&ctx->pool_lock, flags); + + spin_lock(&mem->lock); + avail = dca_mem_is_available(mem); + ret = 0; + for (i = 0; avail && i < mem->page_count; i++) { + ret = cb(mem, i, param); + if (ret == DCA_MEM_STOP_ITERATE || + ret == DCA_MEM_NEXT_ITERATE) + break; + } + spin_unlock(&mem->lock); + spin_lock_irqsave(&ctx->pool_lock, flags); + + if (ret == DCA_MEM_STOP_ITERATE) + goto done; + } + +done: + spin_unlock_irqrestore(&ctx->pool_lock, flags); +} + /* user DCA is managed by ucontext */ static inline struct hns_roce_dca_ctx * to_hr_dca_ctx(struct hns_roce_ucontext *uctx) @@ -194,6 +239,63 @@ static int register_dca_mem(struct hns_roce_dev *hr_dev, return 0; } +struct dca_mem_shrink_attr { + u64 shrink_key; + u32 shrink_mems; +}; + +static int shrink_dca_page_proc(struct dca_mem *mem, int index, void *param) +{ + struct dca_mem_shrink_attr *attr = param; + struct hns_dca_page_state *state; + int i, free_pages; + + free_pages = 0; + for (i = 0; i < mem->page_count; i++) { + state = &mem->states[i]; + if (dca_page_is_free(state)) + free_pages++; + } + + /* No pages are in use */ + if (free_pages == mem->page_count) { + /* unregister first empty DCA mem */ + if (!attr->shrink_mems) { + clr_dca_mem_registered(mem); + attr->shrink_key = mem->key; + } + + attr->shrink_mems++; + } + + if (attr->shrink_mems > 1) + return DCA_MEM_STOP_ITERATE; + else + return DCA_MEM_NEXT_ITERATE; +} + +static int shrink_dca_mem(struct hns_roce_dev *hr_dev, + struct hns_roce_ucontext *uctx, u64 reserved_size, + struct hns_dca_shrink_resp *resp) +{ + struct hns_roce_dca_ctx *ctx = to_hr_dca_ctx(uctx); + struct dca_mem_shrink_attr attr = {}; + unsigned long flags; + bool need_shink; + + spin_lock_irqsave(&ctx->pool_lock, flags); + need_shink = ctx->free_mems > 0 && ctx->free_size > reserved_size; + spin_unlock_irqrestore(&ctx->pool_lock, flags); + if (!need_shink) + return 0; + + travel_dca_pages(ctx, &attr, shrink_dca_page_proc); + resp->free_mems = attr.shrink_mems; + resp->free_key = attr.shrink_key; + + return 0; +} + static void init_dca_context(struct hns_roce_dca_ctx *ctx) { INIT_LIST_HEAD(&ctx->pool); @@ -361,10 +463,48 @@ DECLARE_UVERBS_NAMED_METHOD_DESTROY( UVERBS_ATTR_IDR(HNS_IB_ATTR_DCA_MEM_DEREG_HANDLE, HNS_IB_OBJECT_DCA_MEM, UVERBS_ACCESS_DESTROY, UA_MANDATORY)); +static int UVERBS_HANDLER(HNS_IB_METHOD_DCA_MEM_SHRINK)( + struct uverbs_attr_bundle *attrs) +{ + struct hns_roce_ucontext *uctx = uverbs_attr_to_hr_uctx(attrs); + struct hns_dca_shrink_resp resp = {}; + u64 reserved_size = 0; + int ret; + + if (uverbs_copy_from(&reserved_size, attrs, + HNS_IB_ATTR_DCA_MEM_SHRINK_RESERVED_SIZE)) + return -EFAULT; + + ret = shrink_dca_mem(to_hr_dev(uctx->ibucontext.device), uctx, + reserved_size, &resp); + if (ret) + return ret; + + if (uverbs_copy_to(attrs, HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_KEY, + &resp.free_key, sizeof(resp.free_key)) || + uverbs_copy_to(attrs, HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_MEMS, + &resp.free_mems, sizeof(resp.free_mems))) + return -EFAULT; + + return 0; +} + +DECLARE_UVERBS_NAMED_METHOD( + HNS_IB_METHOD_DCA_MEM_SHRINK, + UVERBS_ATTR_IDR(HNS_IB_ATTR_DCA_MEM_SHRINK_HANDLE, + HNS_IB_OBJECT_DCA_MEM, UVERBS_ACCESS_WRITE, + UA_MANDATORY), + UVERBS_ATTR_PTR_IN(HNS_IB_ATTR_DCA_MEM_SHRINK_RESERVED_SIZE, + UVERBS_ATTR_TYPE(u64), UA_MANDATORY), + UVERBS_ATTR_PTR_OUT(HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_KEY, + UVERBS_ATTR_TYPE(u64), UA_MANDATORY), + UVERBS_ATTR_PTR_OUT(HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_MEMS, + UVERBS_ATTR_TYPE(u32), UA_MANDATORY)); DECLARE_UVERBS_NAMED_OBJECT(HNS_IB_OBJECT_DCA_MEM, UVERBS_TYPE_ALLOC_IDR(dca_cleanup), &UVERBS_METHOD(HNS_IB_METHOD_DCA_MEM_REG), - &UVERBS_METHOD(HNS_IB_METHOD_DCA_MEM_DEREG)); + &UVERBS_METHOD(HNS_IB_METHOD_DCA_MEM_DEREG), + &UVERBS_METHOD(HNS_IB_METHOD_DCA_MEM_SHRINK)); static bool dca_is_supported(struct ib_device *device) { diff --git a/drivers/infiniband/hw/hns/hns_roce_dca.h b/drivers/infiniband/hw/hns/hns_roce_dca.h index cb3481f..97caf03 100644 --- a/drivers/infiniband/hw/hns/hns_roce_dca.h +++ b/drivers/infiniband/hw/hns/hns_roce_dca.h @@ -14,6 +14,13 @@ struct hns_dca_page_state { u32 head : 1; /* This page is the head in a continuous address range. */ }; +struct hns_dca_shrink_resp { + u64 free_key; /* free buffer's key which registered by the user */ + u32 free_mems; /* free buffer count which no any QP be using */ +}; + +#define HNS_DCA_INVALID_BUF_ID 0UL + void hns_roce_register_udca(struct hns_roce_dev *hr_dev, struct hns_roce_ucontext *uctx); void hns_roce_unregister_udca(struct hns_roce_dev *hr_dev, diff --git a/include/uapi/rdma/hns-abi.h b/include/uapi/rdma/hns-abi.h index c17ec91..bcca5be 100644 --- a/include/uapi/rdma/hns-abi.h +++ b/include/uapi/rdma/hns-abi.h @@ -109,6 +109,7 @@ enum hns_ib_objects { enum hns_ib_dca_mem_methods { HNS_IB_METHOD_DCA_MEM_REG = (1U << UVERBS_ID_NS_SHIFT), HNS_IB_METHOD_DCA_MEM_DEREG, + HNS_IB_METHOD_DCA_MEM_SHRINK, }; enum hns_ib_dca_mem_reg_attrs { @@ -121,4 +122,12 @@ enum hns_ib_dca_mem_reg_attrs { enum hns_ib_dca_mem_dereg_attrs { HNS_IB_ATTR_DCA_MEM_DEREG_HANDLE = (1U << UVERBS_ID_NS_SHIFT), }; + +enum hns_ib_dca_mem_shrink_attrs { + HNS_IB_ATTR_DCA_MEM_SHRINK_HANDLE = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_ATTR_DCA_MEM_SHRINK_RESERVED_SIZE, + HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_KEY, + HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_MEMS, +}; + #endif /* HNS_ABI_USER_H */