xdp_umem.c 5.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261
  1. // SPDX-License-Identifier: GPL-2.0
  2. /* XDP user-space packet buffer
  3. * Copyright(c) 2018 Intel Corporation.
  4. */
  5. #include <linux/init.h>
  6. #include <linux/sched/mm.h>
  7. #include <linux/sched/signal.h>
  8. #include <linux/sched/task.h>
  9. #include <linux/uaccess.h>
  10. #include <linux/slab.h>
  11. #include <linux/bpf.h>
  12. #include <linux/mm.h>
  13. #include <linux/netdevice.h>
  14. #include <linux/rtnetlink.h>
  15. #include <linux/idr.h>
  16. #include <linux/vmalloc.h>
  17. #include "xdp_umem.h"
  18. #include "xsk_queue.h"
  19. #define XDP_UMEM_MIN_CHUNK_SIZE 2048
  20. static DEFINE_IDA(umem_ida);
  21. static void xdp_umem_unpin_pages(struct xdp_umem *umem)
  22. {
  23. unpin_user_pages_dirty_lock(umem->pgs, umem->npgs, true);
  24. kfree(umem->pgs);
  25. umem->pgs = NULL;
  26. }
  27. static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
  28. {
  29. if (umem->user) {
  30. atomic_long_sub(umem->npgs, &umem->user->locked_vm);
  31. free_uid(umem->user);
  32. }
  33. }
  34. static void xdp_umem_addr_unmap(struct xdp_umem *umem)
  35. {
  36. vunmap(umem->addrs);
  37. umem->addrs = NULL;
  38. }
  39. static int xdp_umem_addr_map(struct xdp_umem *umem, struct page **pages,
  40. u32 nr_pages)
  41. {
  42. umem->addrs = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
  43. if (!umem->addrs)
  44. return -ENOMEM;
  45. return 0;
  46. }
  47. static void xdp_umem_release(struct xdp_umem *umem)
  48. {
  49. umem->zc = false;
  50. ida_simple_remove(&umem_ida, umem->id);
  51. xdp_umem_addr_unmap(umem);
  52. xdp_umem_unpin_pages(umem);
  53. xdp_umem_unaccount_pages(umem);
  54. kfree(umem);
  55. }
  56. static void xdp_umem_release_deferred(struct work_struct *work)
  57. {
  58. struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
  59. xdp_umem_release(umem);
  60. }
  61. void xdp_get_umem(struct xdp_umem *umem)
  62. {
  63. refcount_inc(&umem->users);
  64. }
  65. void xdp_put_umem(struct xdp_umem *umem, bool defer_cleanup)
  66. {
  67. if (!umem)
  68. return;
  69. if (refcount_dec_and_test(&umem->users)) {
  70. if (defer_cleanup) {
  71. INIT_WORK(&umem->work, xdp_umem_release_deferred);
  72. schedule_work(&umem->work);
  73. } else {
  74. xdp_umem_release(umem);
  75. }
  76. }
  77. }
  78. static int xdp_umem_pin_pages(struct xdp_umem *umem, unsigned long address)
  79. {
  80. unsigned int gup_flags = FOLL_WRITE;
  81. long npgs;
  82. int err;
  83. umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
  84. GFP_KERNEL | __GFP_NOWARN);
  85. if (!umem->pgs)
  86. return -ENOMEM;
  87. mmap_read_lock(current->mm);
  88. npgs = pin_user_pages(address, umem->npgs,
  89. gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
  90. mmap_read_unlock(current->mm);
  91. if (npgs != umem->npgs) {
  92. if (npgs >= 0) {
  93. umem->npgs = npgs;
  94. err = -ENOMEM;
  95. goto out_pin;
  96. }
  97. err = npgs;
  98. goto out_pgs;
  99. }
  100. return 0;
  101. out_pin:
  102. xdp_umem_unpin_pages(umem);
  103. out_pgs:
  104. kfree(umem->pgs);
  105. umem->pgs = NULL;
  106. return err;
  107. }
  108. static int xdp_umem_account_pages(struct xdp_umem *umem)
  109. {
  110. unsigned long lock_limit, new_npgs, old_npgs;
  111. if (capable(CAP_IPC_LOCK))
  112. return 0;
  113. lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
  114. umem->user = get_uid(current_user());
  115. do {
  116. old_npgs = atomic_long_read(&umem->user->locked_vm);
  117. new_npgs = old_npgs + umem->npgs;
  118. if (new_npgs > lock_limit) {
  119. free_uid(umem->user);
  120. umem->user = NULL;
  121. return -ENOBUFS;
  122. }
  123. } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
  124. new_npgs) != old_npgs);
  125. return 0;
  126. }
  127. static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
  128. {
  129. u32 npgs_rem, chunk_size = mr->chunk_size, headroom = mr->headroom;
  130. bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
  131. u64 npgs, addr = mr->addr, size = mr->len;
  132. unsigned int chunks, chunks_rem;
  133. int err;
  134. if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
  135. /* Strictly speaking we could support this, if:
  136. * - huge pages, or*
  137. * - using an IOMMU, or
  138. * - making sure the memory area is consecutive
  139. * but for now, we simply say "computer says no".
  140. */
  141. return -EINVAL;
  142. }
  143. if (mr->flags & ~XDP_UMEM_UNALIGNED_CHUNK_FLAG)
  144. return -EINVAL;
  145. if (!unaligned_chunks && !is_power_of_2(chunk_size))
  146. return -EINVAL;
  147. if (!PAGE_ALIGNED(addr)) {
  148. /* Memory area has to be page size aligned. For
  149. * simplicity, this might change.
  150. */
  151. return -EINVAL;
  152. }
  153. if ((addr + size) < addr)
  154. return -EINVAL;
  155. npgs = div_u64_rem(size, PAGE_SIZE, &npgs_rem);
  156. if (npgs_rem)
  157. npgs++;
  158. if (npgs > U32_MAX)
  159. return -EINVAL;
  160. chunks = (unsigned int)div_u64_rem(size, chunk_size, &chunks_rem);
  161. if (chunks == 0)
  162. return -EINVAL;
  163. if (!unaligned_chunks && chunks_rem)
  164. return -EINVAL;
  165. if (headroom >= chunk_size - XDP_PACKET_HEADROOM)
  166. return -EINVAL;
  167. umem->size = size;
  168. umem->headroom = headroom;
  169. umem->chunk_size = chunk_size;
  170. umem->chunks = chunks;
  171. umem->npgs = (u32)npgs;
  172. umem->pgs = NULL;
  173. umem->user = NULL;
  174. umem->flags = mr->flags;
  175. INIT_LIST_HEAD(&umem->xsk_dma_list);
  176. refcount_set(&umem->users, 1);
  177. err = xdp_umem_account_pages(umem);
  178. if (err)
  179. return err;
  180. err = xdp_umem_pin_pages(umem, (unsigned long)addr);
  181. if (err)
  182. goto out_account;
  183. err = xdp_umem_addr_map(umem, umem->pgs, umem->npgs);
  184. if (err)
  185. goto out_unpin;
  186. return 0;
  187. out_unpin:
  188. xdp_umem_unpin_pages(umem);
  189. out_account:
  190. xdp_umem_unaccount_pages(umem);
  191. return err;
  192. }
  193. struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
  194. {
  195. struct xdp_umem *umem;
  196. int err;
  197. umem = kzalloc(sizeof(*umem), GFP_KERNEL);
  198. if (!umem)
  199. return ERR_PTR(-ENOMEM);
  200. err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
  201. if (err < 0) {
  202. kfree(umem);
  203. return ERR_PTR(err);
  204. }
  205. umem->id = err;
  206. err = xdp_umem_reg(umem, mr);
  207. if (err) {
  208. ida_simple_remove(&umem_ida, umem->id);
  209. kfree(umem);
  210. return ERR_PTR(err);
  211. }
  212. return umem;
  213. }