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>
19 #include "xsk_queue.h"
21 #define XDP_UMEM_MIN_CHUNK_SIZE 2048
23 static DEFINE_IDA(umem_ida);
25 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
29 spin_lock_irqsave(&umem->xsk_list_lock, flags);
30 list_add_rcu(&xs->list, &umem->xsk_list);
31 spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
34 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
38 spin_lock_irqsave(&umem->xsk_list_lock, flags);
39 list_del_rcu(&xs->list);
40 spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
43 /* The umem is stored both in the _rx struct and the _tx struct as we do
44 * not know if the device has more tx queues than rx, or the opposite.
45 * This might also change during run time.
47 static void xdp_reg_umem_at_qid(struct net_device *dev, struct xdp_umem *umem,
50 if (queue_id < dev->real_num_rx_queues)
51 dev->_rx[queue_id].umem = umem;
52 if (queue_id < dev->real_num_tx_queues)
53 dev->_tx[queue_id].umem = umem;
56 struct xdp_umem *xdp_get_umem_from_qid(struct net_device *dev,
59 if (queue_id < dev->real_num_rx_queues)
60 return dev->_rx[queue_id].umem;
61 if (queue_id < dev->real_num_tx_queues)
62 return dev->_tx[queue_id].umem;
67 static void xdp_clear_umem_at_qid(struct net_device *dev, u16 queue_id)
69 if (queue_id < dev->real_num_rx_queues)
70 dev->_rx[queue_id].umem = NULL;
71 if (queue_id < dev->real_num_tx_queues)
72 dev->_tx[queue_id].umem = NULL;
75 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev,
76 u16 queue_id, u16 flags)
78 bool force_zc, force_copy;
79 struct netdev_bpf bpf;
82 force_zc = flags & XDP_ZEROCOPY;
83 force_copy = flags & XDP_COPY;
85 if (force_zc && force_copy)
89 if (xdp_get_umem_from_qid(dev, queue_id)) {
94 xdp_reg_umem_at_qid(dev, umem, queue_id);
96 umem->queue_id = queue_id;
98 /* For copy-mode, we are done. */
101 if (!dev->netdev_ops->ndo_bpf ||
102 !dev->netdev_ops->ndo_xsk_async_xmit) {
107 bpf.command = XDP_SETUP_XSK_UMEM;
109 bpf.xsk.queue_id = queue_id;
111 err = dev->netdev_ops->ndo_bpf(dev, &bpf);
121 xdp_clear_umem_at_qid(dev, queue_id);
123 err = 0; /* fallback to copy mode */
129 static void xdp_umem_clear_dev(struct xdp_umem *umem)
131 struct netdev_bpf bpf;
135 bpf.command = XDP_SETUP_XSK_UMEM;
137 bpf.xsk.queue_id = umem->queue_id;
140 err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
144 WARN(1, "failed to disable umem!\n");
149 xdp_clear_umem_at_qid(umem->dev, umem->queue_id);
159 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
163 for (i = 0; i < umem->npgs; i++) {
164 struct page *page = umem->pgs[i];
166 set_page_dirty_lock(page);
174 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
177 atomic_long_sub(umem->npgs, &umem->user->locked_vm);
178 free_uid(umem->user);
182 static void xdp_umem_release(struct xdp_umem *umem)
184 struct task_struct *task;
185 struct mm_struct *mm;
187 xdp_umem_clear_dev(umem);
189 ida_simple_remove(&umem_ida, umem->id);
192 xskq_destroy(umem->fq);
197 xskq_destroy(umem->cq);
201 xsk_reuseq_destroy(umem);
203 xdp_umem_unpin_pages(umem);
205 task = get_pid_task(umem->pid, PIDTYPE_PID);
209 mm = get_task_mm(task);
210 put_task_struct(task);
218 xdp_umem_unaccount_pages(umem);
223 static void xdp_umem_release_deferred(struct work_struct *work)
225 struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
227 xdp_umem_release(umem);
230 void xdp_get_umem(struct xdp_umem *umem)
232 refcount_inc(&umem->users);
235 void xdp_put_umem(struct xdp_umem *umem)
240 if (refcount_dec_and_test(&umem->users)) {
241 INIT_WORK(&umem->work, xdp_umem_release_deferred);
242 schedule_work(&umem->work);
246 static int xdp_umem_pin_pages(struct xdp_umem *umem)
248 unsigned int gup_flags = FOLL_WRITE;
252 umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
253 GFP_KERNEL | __GFP_NOWARN);
257 down_write(¤t->mm->mmap_sem);
258 npgs = get_user_pages(umem->address, umem->npgs,
259 gup_flags, &umem->pgs[0], NULL);
260 up_write(¤t->mm->mmap_sem);
262 if (npgs != umem->npgs) {
274 xdp_umem_unpin_pages(umem);
281 static int xdp_umem_account_pages(struct xdp_umem *umem)
283 unsigned long lock_limit, new_npgs, old_npgs;
285 if (capable(CAP_IPC_LOCK))
288 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
289 umem->user = get_uid(current_user());
292 old_npgs = atomic_long_read(&umem->user->locked_vm);
293 new_npgs = old_npgs + umem->npgs;
294 if (new_npgs > lock_limit) {
295 free_uid(umem->user);
299 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
300 new_npgs) != old_npgs);
304 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
306 u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
307 unsigned int chunks, chunks_per_page;
308 u64 addr = mr->addr, size = mr->len;
309 int size_chk, err, i;
311 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
312 /* Strictly speaking we could support this, if:
314 * - using an IOMMU, or
315 * - making sure the memory area is consecutive
316 * but for now, we simply say "computer says no".
321 if (!is_power_of_2(chunk_size))
324 if (!PAGE_ALIGNED(addr)) {
325 /* Memory area has to be page size aligned. For
326 * simplicity, this might change.
331 if ((addr + size) < addr)
334 chunks = (unsigned int)div_u64(size, chunk_size);
338 chunks_per_page = PAGE_SIZE / chunk_size;
339 if (chunks < chunks_per_page || chunks % chunks_per_page)
342 headroom = ALIGN(headroom, 64);
344 size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
348 umem->pid = get_task_pid(current, PIDTYPE_PID);
349 umem->address = (unsigned long)addr;
350 umem->chunk_mask = ~((u64)chunk_size - 1);
352 umem->headroom = headroom;
353 umem->chunk_size_nohr = chunk_size - headroom;
354 umem->npgs = size / PAGE_SIZE;
357 INIT_LIST_HEAD(&umem->xsk_list);
358 spin_lock_init(&umem->xsk_list_lock);
360 refcount_set(&umem->users, 1);
362 err = xdp_umem_account_pages(umem);
366 err = xdp_umem_pin_pages(umem);
370 umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL);
376 for (i = 0; i < umem->npgs; i++)
377 umem->pages[i].addr = page_address(umem->pgs[i]);
382 xdp_umem_unaccount_pages(umem);
388 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
390 struct xdp_umem *umem;
393 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
395 return ERR_PTR(-ENOMEM);
397 err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
404 err = xdp_umem_reg(umem, mr);
406 ida_simple_remove(&umem_ida, umem->id);
414 bool xdp_umem_validate_queues(struct xdp_umem *umem)
416 return umem->fq && umem->cq;