1 // SPDX-License-Identifier: GPL-2.0
2 /* XDP user-space packet buffer
3 * Copyright(c) 2018 Intel Corporation.
6 #include <linux/init.h>
7 #include <linux/sched/mm.h>
8 #include <linux/sched/signal.h>
9 #include <linux/sched/task.h>
10 #include <linux/uaccess.h>
11 #include <linux/slab.h>
12 #include <linux/bpf.h>
14 #include <linux/netdevice.h>
15 #include <linux/rtnetlink.h>
16 #include <linux/idr.h>
17 #include <linux/vmalloc.h>
20 #include "xsk_queue.h"
22 #define XDP_UMEM_MIN_CHUNK_SIZE 2048
24 static DEFINE_IDA(umem_ida);
26 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
28 unpin_user_pages_dirty_lock(umem->pgs, umem->npgs, true);
34 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
37 atomic_long_sub(umem->npgs, &umem->user->locked_vm);
42 static void xdp_umem_addr_unmap(struct xdp_umem *umem)
48 static int xdp_umem_addr_map(struct xdp_umem *umem, struct page **pages,
51 umem->addrs = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
57 static void xdp_umem_release(struct xdp_umem *umem)
60 ida_simple_remove(&umem_ida, umem->id);
62 xdp_umem_addr_unmap(umem);
63 xdp_umem_unpin_pages(umem);
65 xdp_umem_unaccount_pages(umem);
69 static void xdp_umem_release_deferred(struct work_struct *work)
71 struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
73 xdp_umem_release(umem);
76 void xdp_get_umem(struct xdp_umem *umem)
78 refcount_inc(&umem->users);
81 void xdp_put_umem(struct xdp_umem *umem, bool defer_cleanup)
86 if (refcount_dec_and_test(&umem->users)) {
88 INIT_WORK(&umem->work, xdp_umem_release_deferred);
89 schedule_work(&umem->work);
91 xdp_umem_release(umem);
96 static int xdp_umem_pin_pages(struct xdp_umem *umem, unsigned long address)
98 unsigned int gup_flags = FOLL_WRITE;
102 umem->pgs = kvcalloc(umem->npgs, sizeof(*umem->pgs), GFP_KERNEL | __GFP_NOWARN);
106 mmap_read_lock(current->mm);
107 npgs = pin_user_pages(address, umem->npgs,
108 gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
109 mmap_read_unlock(current->mm);
111 if (npgs != umem->npgs) {
123 xdp_umem_unpin_pages(umem);
130 static int xdp_umem_account_pages(struct xdp_umem *umem)
132 unsigned long lock_limit, new_npgs, old_npgs;
134 if (capable(CAP_IPC_LOCK))
137 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
138 umem->user = get_uid(current_user());
141 old_npgs = atomic_long_read(&umem->user->locked_vm);
142 new_npgs = old_npgs + umem->npgs;
143 if (new_npgs > lock_limit) {
144 free_uid(umem->user);
148 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
149 new_npgs) != old_npgs);
153 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
155 bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
156 u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
157 u64 addr = mr->addr, size = mr->len;
158 u32 chunks_rem, npgs_rem;
162 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
163 /* Strictly speaking we could support this, if:
165 * - using an IOMMU, or
166 * - making sure the memory area is consecutive
167 * but for now, we simply say "computer says no".
172 if (mr->flags & ~XDP_UMEM_UNALIGNED_CHUNK_FLAG)
175 if (!unaligned_chunks && !is_power_of_2(chunk_size))
178 if (!PAGE_ALIGNED(addr)) {
179 /* Memory area has to be page size aligned. For
180 * simplicity, this might change.
185 if ((addr + size) < addr)
188 npgs = div_u64_rem(size, PAGE_SIZE, &npgs_rem);
194 chunks = div_u64_rem(size, chunk_size, &chunks_rem);
195 if (!chunks || chunks > U32_MAX)
198 if (!unaligned_chunks && chunks_rem)
201 if (headroom >= chunk_size - XDP_PACKET_HEADROOM)
205 umem->headroom = headroom;
206 umem->chunk_size = chunk_size;
207 umem->chunks = chunks;
211 umem->flags = mr->flags;
213 INIT_LIST_HEAD(&umem->xsk_dma_list);
214 refcount_set(&umem->users, 1);
216 err = xdp_umem_account_pages(umem);
220 err = xdp_umem_pin_pages(umem, (unsigned long)addr);
224 err = xdp_umem_addr_map(umem, umem->pgs, umem->npgs);
231 xdp_umem_unpin_pages(umem);
233 xdp_umem_unaccount_pages(umem);
237 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
239 struct xdp_umem *umem;
242 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
244 return ERR_PTR(-ENOMEM);
246 err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
253 err = xdp_umem_reg(umem, mr);
255 ida_simple_remove(&umem_ida, umem->id);