1// SPDX-License-Identifier: GPL-2.0
  2/* XDP user-space packet buffer
  3 * Copyright(c) 2018 Intel Corporation.
  4 */
  5
  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>
 13#include <linux/mm.h>
 14#include <linux/netdevice.h>
 15#include <linux/rtnetlink.h>
 16#include <linux/idr.h>
 17#include <linux/vmalloc.h>
 18
 19#include "xdp_umem.h"
 20#include "xsk_queue.h"
 21
 22static DEFINE_IDA(umem_ida);
 23
 24static void xdp_umem_unpin_pages(struct xdp_umem *umem)
 25{
 26	unpin_user_pages_dirty_lock(umem->pgs, umem->npgs, true);
 27
 28	kvfree(umem->pgs);
 29	umem->pgs = NULL;
 30}
 31
 32static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
 33{
 34	if (umem->user) {
 35		atomic_long_sub(umem->npgs, &umem->user->locked_vm);
 36		free_uid(umem->user);
 37	}
 38}
 39
 40static void xdp_umem_addr_unmap(struct xdp_umem *umem)
 41{
 42	vunmap(umem->addrs);
 43	umem->addrs = NULL;
 44}
 45
 46static int xdp_umem_addr_map(struct xdp_umem *umem, struct page **pages,
 47			     u32 nr_pages)
 48{
 49	umem->addrs = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
 50	if (!umem->addrs)
 51		return -ENOMEM;
 52	return 0;
 53}
 54
 55static void xdp_umem_release(struct xdp_umem *umem)
 56{
 57	umem->zc = false;
 58	ida_free(&umem_ida, umem->id);
 59
 60	xdp_umem_addr_unmap(umem);
 61	xdp_umem_unpin_pages(umem);
 62
 63	xdp_umem_unaccount_pages(umem);
 64	kfree(umem);
 65}
 66
 67static void xdp_umem_release_deferred(struct work_struct *work)
 68{
 69	struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
 70
 71	xdp_umem_release(umem);
 72}
 73
 74void xdp_get_umem(struct xdp_umem *umem)
 75{
 76	refcount_inc(&umem->users);
 77}
 78
 79void xdp_put_umem(struct xdp_umem *umem, bool defer_cleanup)
 80{
 81	if (!umem)
 82		return;
 83
 84	if (refcount_dec_and_test(&umem->users)) {
 85		if (defer_cleanup) {
 86			INIT_WORK(&umem->work, xdp_umem_release_deferred);
 87			schedule_work(&umem->work);
 88		} else {
 89			xdp_umem_release(umem);
 90		}
 91	}
 92}
 93
 94static int xdp_umem_pin_pages(struct xdp_umem *umem, unsigned long address)
 95{
 96	unsigned int gup_flags = FOLL_WRITE;
 97	long npgs;
 98	int err;
 99
100	umem->pgs = kvcalloc(umem->npgs, sizeof(*umem->pgs), GFP_KERNEL | __GFP_NOWARN);
101	if (!umem->pgs)
102		return -ENOMEM;
103
104	mmap_read_lock(current->mm);
105	npgs = pin_user_pages(address, umem->npgs,
106			      gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
107	mmap_read_unlock(current->mm);
108
109	if (npgs != umem->npgs) {
110		if (npgs >= 0) {
111			umem->npgs = npgs;
112			err = -ENOMEM;
113			goto out_pin;
114		}
115		err = npgs;
116		goto out_pgs;
117	}
118	return 0;
119
120out_pin:
121	xdp_umem_unpin_pages(umem);
122out_pgs:
123	kvfree(umem->pgs);
124	umem->pgs = NULL;
125	return err;
126}
127
128static int xdp_umem_account_pages(struct xdp_umem *umem)
129{
130	unsigned long lock_limit, new_npgs, old_npgs;
131
132	if (capable(CAP_IPC_LOCK))
133		return 0;
134
135	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
136	umem->user = get_uid(current_user());
137
138	do {
139		old_npgs = atomic_long_read(&umem->user->locked_vm);
140		new_npgs = old_npgs + umem->npgs;
141		if (new_npgs > lock_limit) {
142			free_uid(umem->user);
143			umem->user = NULL;
144			return -ENOBUFS;
145		}
146	} while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
147				     new_npgs) != old_npgs);
148	return 0;
149}
150
151static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
152{
153	u32 npgs_rem, chunk_size = mr->chunk_size, headroom = mr->headroom;
154	bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
155	u64 npgs, addr = mr->addr, size = mr->len;
156	unsigned int chunks, chunks_rem;
157	int err;
158
159	if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
160		/* Strictly speaking we could support this, if:
161		 * - huge pages, or*
162		 * - using an IOMMU, or
163		 * - making sure the memory area is consecutive
164		 * but for now, we simply say "computer says no".
165		 */
166		return -EINVAL;
167	}
168
169	if (mr->flags & ~XDP_UMEM_UNALIGNED_CHUNK_FLAG)
170		return -EINVAL;
171
172	if (!unaligned_chunks && !is_power_of_2(chunk_size))
173		return -EINVAL;
174
175	if (!PAGE_ALIGNED(addr)) {
176		/* Memory area has to be page size aligned. For
177		 * simplicity, this might change.
178		 */
179		return -EINVAL;
180	}
181
182	if ((addr + size) < addr)
183		return -EINVAL;
184
185	npgs = div_u64_rem(size, PAGE_SIZE, &npgs_rem);
186	if (npgs_rem)
187		npgs++;
188	if (npgs > U32_MAX)
189		return -EINVAL;
190
191	chunks = (unsigned int)div_u64_rem(size, chunk_size, &chunks_rem);
192	if (chunks == 0)
193		return -EINVAL;
194
195	if (!unaligned_chunks && chunks_rem)
196		return -EINVAL;
197
198	if (headroom >= chunk_size - XDP_PACKET_HEADROOM)
199		return -EINVAL;
200
201	umem->size = size;
202	umem->headroom = headroom;
203	umem->chunk_size = chunk_size;
204	umem->chunks = chunks;
205	umem->npgs = (u32)npgs;
206	umem->pgs = NULL;
207	umem->user = NULL;
208	umem->flags = mr->flags;
209
210	INIT_LIST_HEAD(&umem->xsk_dma_list);
211	refcount_set(&umem->users, 1);
212
213	err = xdp_umem_account_pages(umem);
214	if (err)
215		return err;
216
217	err = xdp_umem_pin_pages(umem, (unsigned long)addr);
218	if (err)
219		goto out_account;
220
221	err = xdp_umem_addr_map(umem, umem->pgs, umem->npgs);
222	if (err)
223		goto out_unpin;
224
225	return 0;
226
227out_unpin:
228	xdp_umem_unpin_pages(umem);
229out_account:
230	xdp_umem_unaccount_pages(umem);
231	return err;
232}
233
234struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
235{
236	struct xdp_umem *umem;
237	int err;
238
239	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
240	if (!umem)
241		return ERR_PTR(-ENOMEM);
242
243	err = ida_alloc(&umem_ida, GFP_KERNEL);
244	if (err < 0) {
245		kfree(umem);
246		return ERR_PTR(err);
247	}
248	umem->id = err;
249
250	err = xdp_umem_reg(umem, mr);
251	if (err) {
252		ida_free(&umem_ida, umem->id);
253		kfree(umem);
254		return ERR_PTR(err);
255	}
256
257	return umem;
258}