Re: [PATCH bpf-next v3 2/3] bpf: struct_ops supports more than one page for trampolines.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On 2/23/24 22:19, Martin KaFai Lau wrote:
On 2/23/24 7:03 PM, Kui-Feng Lee wrote:
+static void bpf_struct_ops_map_free_image(struct bpf_struct_ops_map *st_map)
+{
+    int i;
+    void *image;
+
+    bpf_jit_uncharge_modmem(PAGE_SIZE * st_map->image_pages_cnt);
+    for (i = 0; i < st_map->image_pages_cnt; i++) {
+        image = st_map->image_pages[i];
+        arch_free_bpf_trampoline(image, PAGE_SIZE);
+    }
+    st_map->image_pages_cnt = 0;
+}
+

[ ... ]

@@ -133,7 +128,8 @@ int bpf_struct_ops_test_run(struct bpf_prog *prog, const union bpf_attr *kattr,
      err = bpf_struct_ops_prepare_trampoline(tlinks, link,
                          &st_ops->func_models[op_idx],
                          &dummy_ops_test_ret_function,
-                        image, image + PAGE_SIZE);
+                        &image, &image_off,
+                        true);
      if (err < 0)
          goto out;
@@ -147,6 +143,8 @@ int bpf_struct_ops_test_run(struct bpf_prog *prog, const union bpf_attr *kattr,
          err = -EFAULT;
  out:
      kfree(args);
+    if (image)
+        bpf_jit_uncharge_modmem(PAGE_SIZE);
      arch_free_bpf_trampoline(image, PAGE_SIZE);

It seems my last reply on v2 has crossed over with v3.

The bpf_struct_ops_free_trampoline() highlighted in my last reply should
address your concern in v2 that the caller needs to remember
the bpf_jit_uncharge_modmem here.

I think the trampoline alloc(aka prepare here)/free pair that you also
suggested in v2 discussion is a nice match here and work as a
charge+alloc and uncharge+free pair.

sure


      if (link)
          bpf_link_put(&link->link);





[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux