infiniband/core/umem.c

0001 /*
0002  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
0003  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
0004  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
0005  * Copyright (c) 2020 Intel Corporation. All rights reserved.
0006  *
0007  * This software is available to you under a choice of one of two
0008  * licenses.  You may choose to be licensed under the terms of the GNU
0009  * General Public License (GPL) Version 2, available from the file
0010  * COPYING in the main directory of this source tree, or the
0011  * OpenIB.org BSD license below:
0012  *
0013  *     Redistribution and use in source and binary forms, with or
0014  *     without modification, are permitted provided that the following
0015  *     conditions are met:
0016  *
0017  *      - Redistributions of source code must retain the above
0018  *        copyright notice, this list of conditions and the following
0019  *        disclaimer.
0020  *
0021  *      - Redistributions in binary form must reproduce the above
0022  *        copyright notice, this list of conditions and the following
0023  *        disclaimer in the documentation and/or other materials
0024  *        provided with the distribution.
0025  *
0026  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0027  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0028  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0029  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0030  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0031  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0032  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0033  * SOFTWARE.
0034  */
0035
0036 #include <linux/mm.h>
0037 #include <linux/dma-mapping.h>
0038 #include <linux/sched/signal.h>
0039 #include <linux/sched/mm.h>
0040 #include <linux/export.h>
0041 #include <linux/slab.h>
0042 #include <linux/pagemap.h>
0043 #include <linux/count_zeros.h>
0044 #include <rdma/ib_umem_odp.h>
0045
0046 #include "uverbs.h"
0047
0048 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
0049 {
0050     bool make_dirty = umem->writable && dirty;
0051     struct scatterlist *sg;
0052     unsigned int i;
0053
0054     if (dirty)
0055         ib_dma_unmap_sgtable_attrs(dev, &umem->sgt_append.sgt,
0056                        DMA_BIDIRECTIONAL, 0);
0057
0058     for_each_sgtable_sg(&umem->sgt_append.sgt, sg, i)
0059         unpin_user_page_range_dirty_lock(sg_page(sg),
0060             DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);
0061
0062     sg_free_append_table(&umem->sgt_append);
0063 }
0064
0065 /**
0066  * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
0067  *
0068  * @umem: umem struct
0069  * @pgsz_bitmap: bitmap of HW supported page sizes
0070  * @virt: IOVA
0071  *
0072  * This helper is intended for HW that support multiple page
0073  * sizes but can do only a single page size in an MR.
0074  *
0075  * Returns 0 if the umem requires page sizes not supported by
0076  * the driver to be mapped. Drivers always supporting PAGE_SIZE
0077  * or smaller will never see a 0 result.
0078  */
0079 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
0080                      unsigned long pgsz_bitmap,
0081                      unsigned long virt)
0082 {
0083     struct scatterlist *sg;
0084     unsigned long va, pgoff;
0085     dma_addr_t mask;
0086     int i;
0087
0088     if (umem->is_odp) {
0089         unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift);
0090
0091         /* ODP must always be self consistent. */
0092         if (!(pgsz_bitmap & page_size))
0093             return 0;
0094         return page_size;
0095     }
0096
0097     /* rdma_for_each_block() has a bug if the page size is smaller than the
0098      * page size used to build the umem. For now prevent smaller page sizes
0099      * from being returned.
0100      */
0101     pgsz_bitmap &= GENMASK(BITS_PER_LONG - 1, PAGE_SHIFT);
0102
0103     umem->iova = va = virt;
0104     /* The best result is the smallest page size that results in the minimum
0105      * number of required pages. Compute the largest page size that could
0106      * work based on VA address bits that don't change.
0107      */
0108     mask = pgsz_bitmap &
0109            GENMASK(BITS_PER_LONG - 1,
0110                bits_per((umem->length - 1 + virt) ^ virt));
0111     /* offset into first SGL */
0112     pgoff = umem->address & ~PAGE_MASK;
0113
0114     for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) {
0115         /* Walk SGL and reduce max page size if VA/PA bits differ
0116          * for any address.
0117          */
0118         mask |= (sg_dma_address(sg) + pgoff) ^ va;
0119         va += sg_dma_len(sg) - pgoff;
0120         /* Except for the last entry, the ending iova alignment sets
0121          * the maximum possible page size as the low bits of the iova
0122          * must be zero when starting the next chunk.
0123          */
0124         if (i != (umem->sgt_append.sgt.nents - 1))
0125             mask |= va;
0126         pgoff = 0;
0127     }
0128
0129     /* The mask accumulates 1's in each position where the VA and physical
0130      * address differ, thus the length of trailing 0 is the largest page
0131      * size that can pass the VA through to the physical.
0132      */
0133     if (mask)
0134         pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0);
0135     return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0;
0136 }
0137 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
0138
0139 /**
0140  * ib_umem_get - Pin and DMA map userspace memory.
0141  *
0142  * @device: IB device to connect UMEM
0143  * @addr: userspace virtual address to start at
0144  * @size: length of region to pin
0145  * @access: IB_ACCESS_xxx flags for memory being pinned
0146  */
0147 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
0148                 size_t size, int access)
0149 {
0150     struct ib_umem *umem;
0151     struct page **page_list;
0152     unsigned long lock_limit;
0153     unsigned long new_pinned;
0154     unsigned long cur_base;
0155     unsigned long dma_attr = 0;
0156     struct mm_struct *mm;
0157     unsigned long npages;
0158     int pinned, ret;
0159     unsigned int gup_flags = FOLL_WRITE;
0160
0161     /*
0162      * If the combination of the addr and size requested for this memory
0163      * region causes an integer overflow, return error.
0164      */
0165     if (((addr + size) < addr) ||
0166         PAGE_ALIGN(addr + size) < (addr + size))
0167         return ERR_PTR(-EINVAL);
0168
0169     if (!can_do_mlock())
0170         return ERR_PTR(-EPERM);
0171
0172     if (access & IB_ACCESS_ON_DEMAND)
0173         return ERR_PTR(-EOPNOTSUPP);
0174
0175     umem = kzalloc(sizeof(*umem), GFP_KERNEL);
0176     if (!umem)
0177         return ERR_PTR(-ENOMEM);
0178     umem->ibdev      = device;
0179     umem->length     = size;
0180     umem->address    = addr;
0181     /*
0182      * Drivers should call ib_umem_find_best_pgsz() to set the iova
0183      * correctly.
0184      */
0185     umem->iova = addr;
0186     umem->writable   = ib_access_writable(access);
0187     umem->owning_mm = mm = current->mm;
0188     mmgrab(mm);
0189
0190     page_list = (struct page **) __get_free_page(GFP_KERNEL);
0191     if (!page_list) {
0192         ret = -ENOMEM;
0193         goto umem_kfree;
0194     }
0195
0196     npages = ib_umem_num_pages(umem);
0197     if (npages == 0 || npages > UINT_MAX) {
0198         ret = -EINVAL;
0199         goto out;
0200     }
0201
0202     lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
0203
0204     new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
0205     if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
0206         atomic64_sub(npages, &mm->pinned_vm);
0207         ret = -ENOMEM;
0208         goto out;
0209     }
0210
0211     cur_base = addr & PAGE_MASK;
0212
0213     if (!umem->writable)
0214         gup_flags |= FOLL_FORCE;
0215
0216     while (npages) {
0217         cond_resched();
0218         pinned = pin_user_pages_fast(cur_base,
0219                       min_t(unsigned long, npages,
0220                         PAGE_SIZE /
0221                         sizeof(struct page *)),
0222                       gup_flags | FOLL_LONGTERM, page_list);
0223         if (pinned < 0) {
0224             ret = pinned;
0225             goto umem_release;
0226         }
0227
0228         cur_base += pinned * PAGE_SIZE;
0229         npages -= pinned;
0230         ret = sg_alloc_append_table_from_pages(
0231             &umem->sgt_append, page_list, pinned, 0,
0232             pinned << PAGE_SHIFT, ib_dma_max_seg_size(device),
0233             npages, GFP_KERNEL);
0234         if (ret) {
0235             unpin_user_pages_dirty_lock(page_list, pinned, 0);
0236             goto umem_release;
0237         }
0238     }
0239
0240     if (access & IB_ACCESS_RELAXED_ORDERING)
0241         dma_attr |= DMA_ATTR_WEAK_ORDERING;
0242
0243     ret = ib_dma_map_sgtable_attrs(device, &umem->sgt_append.sgt,
0244                        DMA_BIDIRECTIONAL, dma_attr);
0245     if (ret)
0246         goto umem_release;
0247     goto out;
0248
0249 umem_release:
0250     __ib_umem_release(device, umem, 0);
0251     atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
0252 out:
0253     free_page((unsigned long) page_list);
0254 umem_kfree:
0255     if (ret) {
0256         mmdrop(umem->owning_mm);
0257         kfree(umem);
0258     }
0259     return ret ? ERR_PTR(ret) : umem;
0260 }
0261 EXPORT_SYMBOL(ib_umem_get);
0262
0263 /**
0264  * ib_umem_release - release memory pinned with ib_umem_get
0265  * @umem: umem struct to release
0266  */
0267 void ib_umem_release(struct ib_umem *umem)
0268 {
0269     if (!umem)
0270         return;
0271     if (umem->is_dmabuf)
0272         return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
0273     if (umem->is_odp)
0274         return ib_umem_odp_release(to_ib_umem_odp(umem));
0275
0276     __ib_umem_release(umem->ibdev, umem, 1);
0277
0278     atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
0279     mmdrop(umem->owning_mm);
0280     kfree(umem);
0281 }
0282 EXPORT_SYMBOL(ib_umem_release);
0283
0284 /*
0285  * Copy from the given ib_umem's pages to the given buffer.
0286  *
0287  * umem - the umem to copy from
0288  * offset - offset to start copying from
0289  * dst - destination buffer
0290  * length - buffer length
0291  *
0292  * Returns 0 on success, or an error code.
0293  */
0294 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
0295               size_t length)
0296 {
0297     size_t end = offset + length;
0298     int ret;
0299
0300     if (offset > umem->length || length > umem->length - offset) {
0301         pr_err("%s not in range. offset: %zd umem length: %zd end: %zd\n",
0302                __func__, offset, umem->length, end);
0303         return -EINVAL;
0304     }
0305
0306     ret = sg_pcopy_to_buffer(umem->sgt_append.sgt.sgl,
0307                  umem->sgt_append.sgt.orig_nents, dst, length,
0308                  offset + ib_umem_offset(umem));
0309
0310     if (ret < 0)
0311         return ret;
0312     else if (ret != length)
0313         return -EINVAL;
0314     else
0315         return 0;
0316 }
0317 EXPORT_SYMBOL(ib_umem_copy_from);