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>
18 #include "xsk_queue.h"
20 #define XDP_UMEM_MIN_CHUNK_SIZE 2048
22 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
26 spin_lock_irqsave(&umem->xsk_list_lock, flags);
27 list_add_rcu(&xs->list, &umem->xsk_list);
28 spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
31 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
36 spin_lock_irqsave(&umem->xsk_list_lock, flags);
37 list_del_rcu(&xs->list);
38 spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
45 int xdp_umem_query(struct net_device *dev, u16 queue_id)
47 struct netdev_bpf bpf;
51 memset(&bpf, 0, sizeof(bpf));
52 bpf.command = XDP_QUERY_XSK_UMEM;
53 bpf.xsk.queue_id = queue_id;
55 if (!dev->netdev_ops->ndo_bpf)
57 return dev->netdev_ops->ndo_bpf(dev, &bpf) ?: !!bpf.xsk.umem;
60 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev,
61 u32 queue_id, u16 flags)
63 bool force_zc, force_copy;
64 struct netdev_bpf bpf;
67 force_zc = flags & XDP_ZEROCOPY;
68 force_copy = flags & XDP_COPY;
70 if (force_zc && force_copy)
76 if (!dev->netdev_ops->ndo_bpf || !dev->netdev_ops->ndo_xsk_async_xmit)
77 return force_zc ? -EOPNOTSUPP : 0; /* fail or fallback */
79 bpf.command = XDP_QUERY_XSK_UMEM;
82 err = xdp_umem_query(dev, queue_id);
84 err = err < 0 ? -EOPNOTSUPP : -EBUSY;
88 bpf.command = XDP_SETUP_XSK_UMEM;
90 bpf.xsk.queue_id = queue_id;
92 err = dev->netdev_ops->ndo_bpf(dev, &bpf);
99 umem->queue_id = queue_id;
105 return force_zc ? err : 0; /* fail or fallback */
108 static void xdp_umem_clear_dev(struct xdp_umem *umem)
110 struct netdev_bpf bpf;
114 bpf.command = XDP_SETUP_XSK_UMEM;
116 bpf.xsk.queue_id = umem->queue_id;
119 err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
123 WARN(1, "failed to disable umem!\n");
130 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
134 for (i = 0; i < umem->npgs; i++) {
135 struct page *page = umem->pgs[i];
137 set_page_dirty_lock(page);
145 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
148 atomic_long_sub(umem->npgs, &umem->user->locked_vm);
149 free_uid(umem->user);
153 static void xdp_umem_release(struct xdp_umem *umem)
155 struct task_struct *task;
156 struct mm_struct *mm;
158 xdp_umem_clear_dev(umem);
161 xskq_destroy(umem->fq);
166 xskq_destroy(umem->cq);
170 xdp_umem_unpin_pages(umem);
172 task = get_pid_task(umem->pid, PIDTYPE_PID);
176 mm = get_task_mm(task);
177 put_task_struct(task);
185 xdp_umem_unaccount_pages(umem);
190 static void xdp_umem_release_deferred(struct work_struct *work)
192 struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
194 xdp_umem_release(umem);
197 void xdp_get_umem(struct xdp_umem *umem)
199 refcount_inc(&umem->users);
202 void xdp_put_umem(struct xdp_umem *umem)
207 if (refcount_dec_and_test(&umem->users)) {
208 INIT_WORK(&umem->work, xdp_umem_release_deferred);
209 schedule_work(&umem->work);
213 static int xdp_umem_pin_pages(struct xdp_umem *umem)
215 unsigned int gup_flags = FOLL_WRITE;
219 umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
220 GFP_KERNEL | __GFP_NOWARN);
224 down_write(¤t->mm->mmap_sem);
225 npgs = get_user_pages(umem->address, umem->npgs,
226 gup_flags, &umem->pgs[0], NULL);
227 up_write(¤t->mm->mmap_sem);
229 if (npgs != umem->npgs) {
241 xdp_umem_unpin_pages(umem);
248 static int xdp_umem_account_pages(struct xdp_umem *umem)
250 unsigned long lock_limit, new_npgs, old_npgs;
252 if (capable(CAP_IPC_LOCK))
255 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
256 umem->user = get_uid(current_user());
259 old_npgs = atomic_long_read(&umem->user->locked_vm);
260 new_npgs = old_npgs + umem->npgs;
261 if (new_npgs > lock_limit) {
262 free_uid(umem->user);
266 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
267 new_npgs) != old_npgs);
271 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
273 u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
274 unsigned int chunks, chunks_per_page;
275 u64 addr = mr->addr, size = mr->len;
276 int size_chk, err, i;
278 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
279 /* Strictly speaking we could support this, if:
281 * - using an IOMMU, or
282 * - making sure the memory area is consecutive
283 * but for now, we simply say "computer says no".
288 if (!is_power_of_2(chunk_size))
291 if (!PAGE_ALIGNED(addr)) {
292 /* Memory area has to be page size aligned. For
293 * simplicity, this might change.
298 if ((addr + size) < addr)
301 chunks = (unsigned int)div_u64(size, chunk_size);
305 chunks_per_page = PAGE_SIZE / chunk_size;
306 if (chunks < chunks_per_page || chunks % chunks_per_page)
309 headroom = ALIGN(headroom, 64);
311 size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
315 umem->pid = get_task_pid(current, PIDTYPE_PID);
316 umem->address = (unsigned long)addr;
317 umem->props.chunk_mask = ~((u64)chunk_size - 1);
318 umem->props.size = size;
319 umem->headroom = headroom;
320 umem->chunk_size_nohr = chunk_size - headroom;
321 umem->npgs = size / PAGE_SIZE;
324 INIT_LIST_HEAD(&umem->xsk_list);
325 spin_lock_init(&umem->xsk_list_lock);
327 refcount_set(&umem->users, 1);
329 err = xdp_umem_account_pages(umem);
333 err = xdp_umem_pin_pages(umem);
337 umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL);
343 for (i = 0; i < umem->npgs; i++)
344 umem->pages[i].addr = page_address(umem->pgs[i]);
349 xdp_umem_unaccount_pages(umem);
355 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
357 struct xdp_umem *umem;
360 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
362 return ERR_PTR(-ENOMEM);
364 err = xdp_umem_reg(umem, mr);
373 bool xdp_umem_validate_queues(struct xdp_umem *umem)
375 return umem->fq && umem->cq;