vmalloc: bpf: introduce VM_ALLOW_HUGE_VMAP

Huge page backed vmalloc memory could benefit performance in many cases. Since some users of vmalloc may not be ready to handle huge pages, VM_NO_HUGE_VMAP was introduced to allow vmalloc users to opt-out huge pages. However, it is not easy to add VM_NO_HUGE_VMAP to all the users that may try to allocate >= PMD_SIZE pages, but are not ready to handle huge pages properly.

Replace VM_NO_HUGE_VMAP with an opt-in flag, VM_ALLOW_HUGE_VMAP, so that users that benefit from huge pages could ask specificially.

Also, replace vmalloc_no_huge() with opt-in helper vmalloc_huge().

+extern void *vmalloc_huge(unsigned long size) __alloc_size(1);

+EXPORT_SYMBOL(vmalloc_huge);

Quoted 6 lines · Song Liu

Huge page backed vmalloc memory could benefit performance in many cases. Since some users of vmalloc may not be ready to handle huge pages, VM_NO_HUGE_VMAP was introduced to allow vmalloc users to opt-out huge pages. However, it is not easy to add VM_NO_HUGE_VMAP to all the users that may try to allocate >= PMD_SIZE pages, but are not ready to handle huge pages properly.

This is a good place to document what the problems are, and how they are hard to track down (e.g. because the allocations are passed down I/O stacks)

Quoted 6 lines

Replace VM_NO_HUGE_VMAP with an opt-in flag, VM_ALLOW_HUGE_VMAP, so that users that benefit from huge pages could ask specificially.

Also, replace vmalloc_no_huge() with opt-in helper vmalloc_huge().

We still need to find out what the primary users of the large vmalloc hashes was and convert them.

Quoted 2 lines

+extern void *vmalloc_huge(unsigned long size) __alloc_size(1);

No need for the extern.

Quoted 1 line

+EXPORT_SYMBOL(vmalloc_huge);

EXPORT_SYMBOL_GPL for all advanced vmalloc functionality, please.

Quoted 9 lines · Song Liu

Quoted 6 lines

Huge page backed vmalloc memory could benefit performance in many cases. Since some users of vmalloc may not be ready to handle huge pages, VM_NO_HUGE_VMAP was introduced to allow vmalloc users to opt-out huge pages. However, it is not easy to add VM_NO_HUGE_VMAP to all the users that may try to allocate >= PMD_SIZE pages, but are not ready to handle huge pages properly.

Quoted 3 lines

This is a good place to document what the problems are, and how they are hard to track down (e.g. because the allocations are passed down I/O stacks)

Will add it in v3.

Quoted 7 lines

Quoted 6 lines

Replace VM_NO_HUGE_VMAP with an opt-in flag, VM_ALLOW_HUGE_VMAP, so that users that benefit from huge pages could ask specificially.

Also, replace vmalloc_no_huge() with opt-in helper vmalloc_huge().

Quoted 2 lines

We still need to find out what the primary users of the large vmalloc hashes was and convert them.

@ Claudio and Nicholas,

Could you please help identify users of large vmalloc? So far, I found alloc_large_system_hash(), and something like the following seems to work:

Excerpts from Song Liu's message of April 12, 2022 4:00 pm:

Quoted 28 lines · Christoph Hellwig

Quoted 9 lines · Song Liu

Quoted 6 lines

Huge page backed vmalloc memory could benefit performance in many cases. Since some users of vmalloc may not be ready to handle huge pages, VM_NO_HUGE_VMAP was introduced to allow vmalloc users to opt-out huge pages. However, it is not easy to add VM_NO_HUGE_VMAP to all the users that may try to allocate >= PMD_SIZE pages, but are not ready to handle huge pages properly.

Quoted 3 lines

This is a good place to document what the problems are, and how they are hard to track down (e.g. because the allocations are passed down I/O stacks)

Will add it in v3.

Quoted 9 lines

Quoted 6 lines

Replace VM_NO_HUGE_VMAP with an opt-in flag, VM_ALLOW_HUGE_VMAP, so that users that benefit from huge pages could ask specificially.

Also, replace vmalloc_no_huge() with opt-in helper vmalloc_huge().

We still need to find out what the primary users of the large vmalloc hashes was and convert them.

@ Claudio and Nicholas,

Quoted 3 lines

Could you please help identify users of large vmalloc? So far, I found alloc_large_system_hash(), and something like the following seems to work:

The large system hashes were the main ones I was interested in. IIRC there was a few more in some drivers or tracing things depending on config but those are less important (to me at least).

Introduce module_alloc_huge, which allocates huge page backed memory in module memory space. The primary user of this memory is bpf_prog_pack (multiple BPF programs sharing a huge page).

Quoted 3 lines · Song Liu

Introduce module_alloc_huge, which allocates huge page backed memory in module memory space. The primary user of this memory is bpf_prog_pack (multiple BPF programs sharing a huge page).

I kow I lead you downthis road first, but I wonder if we just want to pass a flag to module_alloc instead. This avoids duplicating all the arch overrides.

Use __vmalloc_node_range with VM_ALLOW_HUGE_VMAP for bpf_prog_pack so that

Quoted 1 line · Song Liu

Use __vmalloc_node_range with VM_ALLOW_HUGE_VMAP for bpf_prog_pack so that

That is only very indirectly true now.

(quoted diff) +0 -1

@@ -889,7 +889,6 @@ static struct bpf_prog_pack *alloc_new_pack(void)
        bitmap_zero(pack->bitmap, bpf_prog_pack_size /
 BPF_PROG_CHUNK_SIZE);
        list_add_tail(&pack->list, &pack_list);
 
-       set_vm_flush_reset_perms(pack->ptr);
        set_memory_ro((unsigned long)pack->ptr, bpf_prog_pack_size /
 PAGE_SIZE);

        set_memory_x((unsigned long)pack->ptr, bpf_prog_pack_size /
PAGE_SIZE);
        return pack;

(quoted diff) +1 -1

@@ -970,7 +969,9 @@ static void bpf_prog_pack_free(struct
 bpf_binary_header *hdr)
        if (bitmap_find_next_zero_area(pack->bitmap,
 bpf_prog_chunk_count(), 0,
                                       bpf_prog_chunk_count(), 0) ==
 0) {
                list_del(&pack->list);
-               module_memfree(pack->ptr);
 
 
+               set_memory_nx((unsigned long)pack->ptr,

bpf_prog_pack_size / PAGE_SIZE);
+ set_memory_rw((unsigned long)pack->ptr,
bpf_prog_pack_size / PAGE_SIZE);

+               vfree(pack->ptr);
                kfree(pack);

Quoted 11 lines · Song Liu

(quoted diff) +0 -1

@@ -889,7 +889,6 @@ static struct bpf_prog_pack *alloc_new_pack(void)
        bitmap_zero(pack->bitmap, bpf_prog_pack_size /
 BPF_PROG_CHUNK_SIZE);
        list_add_tail(&pack->list, &pack_list);
 
-       set_vm_flush_reset_perms(pack->ptr);
        set_memory_ro((unsigned long)pack->ptr, bpf_prog_pack_size /
 PAGE_SIZE);

        set_memory_x((unsigned long)pack->ptr, bpf_prog_pack_size /
PAGE_SIZE);
        return pack;

Dropping set_vm_flush_reset_perms() is not mentioned in the commit log. It is kind of a fix for a different issue.

Now that x86 supports vmalloc huge pages, but VM_FLUSH_RESET_PERMS does not work with them, we should have some comments or warnings to that effect somewhere. Someone may try to pass the flags in together.

Quoted 15 lines

(quoted diff) +1 -1

@@ -970,7 +969,9 @@ static void bpf_prog_pack_free(struct
 bpf_binary_header *hdr)
        if (bitmap_find_next_zero_area(pack->bitmap,
 bpf_prog_chunk_count(), 0,
                                       bpf_prog_chunk_count(), 0) ==
 0) {
                list_del(&pack->list);
-               module_memfree(pack->ptr);
 
 
+               set_memory_nx((unsigned long)pack->ptr,

bpf_prog_pack_size / PAGE_SIZE);
+ set_memory_rw((unsigned long)pack->ptr,
bpf_prog_pack_size / PAGE_SIZE);

Quoted 2 lines

+               vfree(pack->ptr);
                kfree(pack);

Now that it calls module_alloc_huge() instead of vmalloc_node_range(), should it call module_memfree() instead of vfree()?

Since there are bugs, simple, immediate fixes seem like the right thing to do, but I had a couple long term focused comments on this new feature:

It would be nice if bpf and the other module_alloc() callers could share the same large pages. Meaning, ultimately that this whole thing should probably live outside of bpf. BPF tracing usages might benefit for example, and kprobes and ftrace are not too different than bpf progs from a text allocation perspective.

I agree that the module's part is non-trivial. A while back I had tried to do something like bpf_prog_pack() that worked for all the module_alloc() callers. It had some modules changes to allow different permissions to go to different allocations so they could be made to share large pages:

https://lore.kernel.org/lkml/[email protected]/

I thought the existing kernel special permission allocation methods were just too brittle and intertwined to improve without a new interface. The hope was the new interface could wrap all the arch intricacies instead of leaving them exposed in the cross-arch callers.

I wonder what you think of that general direction or if you have any follow up plans for this?

Since I am still learning the vmalloc/module_alloc code, I think I am not really capable of commenting on the direction. From our use cases, we do see performance hit due to large number of BPF program fragmenting the page table. Kernel module, OTOH, is not too big an issue for us, as we usually build hot modules into the kernel. That being said, we are interested in making the huge page interface general for BPF program and kernel module. We can commit resources to this effort.

module_alloc_huge for bpf_prog_pack so that BPF programs sit on PMD_SIZE pages. This benefits system performance by reducing iTLB miss rate.

Signed-off-by Song Liu <[email protected]>

---
 kernel/bpf/core.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

kernel/bpf/core.c +5 -5

index 13e9dbeeedf3..fd45bdd80a75 100644
@@ -857,7 +857,7 @@ static size_t select_bpf_prog_pack_size(void)
        void *ptr;
 
        size = BPF_HPAGE_SIZE * num_online_nodes();
-       ptr = module_alloc(size);
+       ptr = module_alloc_huge(size);
 
        /* Test whether we can get huge pages. If not just use PAGE_SIZE
         * packs.

@@ -881,7 +881,7 @@ static struct bpf_prog_pack *alloc_new_pack(void)
                       GFP_KERNEL);
        if (!pack)
                return NULL;
-       pack->ptr = module_alloc(bpf_prog_pack_size);
+       pack->ptr = module_alloc_huge(bpf_prog_pack_size);
        if (!pack->ptr) {
                kfree(pack);
                return NULL;

@@ -889,7 +889,6 @@ static struct bpf_prog_pack *alloc_new_pack(void)
        bitmap_zero(pack->bitmap, bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE);
        list_add_tail(&pack->list, &pack_list);
 
-       set_vm_flush_reset_perms(pack->ptr);
        set_memory_ro((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE);
        set_memory_x((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE);
        return pack;

@@ -970,7 +969,9 @@ static void bpf_prog_pack_free(struct bpf_binary_header *hdr)
        if (bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0,
                                       bpf_prog_chunk_count(), 0) == 0) {
                list_del(&pack->list);
-               module_memfree(pack->ptr);
+               set_memory_nx((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE);
+               set_memory_rw((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE);
+               vfree(pack->ptr);
                kfree(pack);
        }
 out:
--

2.30.2