Re: [PATCH v2] riscv: Safely remove entries from relocation list
From: Samuel Holland
Date: Tue Nov 21 2023 - 19:04:22 EST
Hi Charlie,
On 2023-11-21 4:50 PM, Charlie Jenkins wrote:
> Use the safe versions of list and hlist iteration to safely remove
> entries from the module relocation lists. To allow mutliple threads to
> load modules concurrently, move relocation list pointers onto the stack
> rather than using global variables.
>
> Fixes: 8fd6c5142395 ("riscv: Add remaining module relocations")
> Reported-by: Ron Economos <re@xxxxxxxx>
> Closes: https://lore.kernel.org/linux-riscv/444de86a-7e7c-4de7-5d1d-c1c40eefa4ba@xxxxxxxx
> Signed-off-by: Charlie Jenkins <charlie@xxxxxxxxxxxx>
> ---
> Changes in v2:
> - Support linking modules concurrently across threads.
> - Link to v1: https://lore.kernel.org/r/20231120-module_linking_freeing-v1-1-fff81d7289fc@xxxxxxxxxxxx
> ---
> arch/riscv/kernel/module.c | 76 +++++++++++++++++++++++++++++++---------------
> 1 file changed, 51 insertions(+), 25 deletions(-)
>
> diff --git a/arch/riscv/kernel/module.c b/arch/riscv/kernel/module.c
> index 56a8c78e9e21..f53e82b70dff 100644
> --- a/arch/riscv/kernel/module.c
> +++ b/arch/riscv/kernel/module.c
> @@ -40,14 +40,17 @@ struct relocation_handlers {
> long buffer);
> };
>
> -unsigned int initialize_relocation_hashtable(unsigned int num_relocations);
> -void process_accumulated_relocations(struct module *me);
> +unsigned int
> +initialize_relocation_hashtable(unsigned int num_relocations,
> + struct hlist_head **relocation_hashtable,
> + struct list_head *used_buckets_list);
> +void process_accumulated_relocations(struct module *me,
> + struct hlist_head **relocation_hashtable,
> + struct list_head *used_buckets_list);
> int add_relocation_to_accumulate(struct module *me, int type, void *location,
> - unsigned int hashtable_bits, Elf_Addr v);
> -
> -struct hlist_head *relocation_hashtable;
> -
> -struct list_head used_buckets_list;
This hunk conflicts with your other patch, which is still needed for the __le16
change. Since they are both fixes, do you intend to rebase and send them together?
> + unsigned int hashtable_bits, Elf_Addr v,
> + struct hlist_head **relocation_hashtable,
> + struct list_head *used_buckets_list);
minor: the indentation is off by one here.
>
> /*
> * The auipc+jalr instruction pair can reach any PC-relative offset
> @@ -604,7 +607,9 @@ static const struct relocation_handlers reloc_handlers[] = {
> /* 192-255 nonstandard ABI extensions */
> };
>
> -void process_accumulated_relocations(struct module *me)
> +void process_accumulated_relocations(struct module *me,
> + struct hlist_head **relocation_hashtable,
You only need the double pointer in initialize_relocation_hashtable(). If you
pass the single pointer here and in add_relocation_to_accumulate(), you can
avoid the extra dereference operations.
> + struct list_head *used_buckets_list)
> {
> /*
> * Only ADD/SUB/SET/ULEB128 should end up here.
> @@ -624,18 +629,25 @@ void process_accumulated_relocations(struct module *me)
> * - Each relocation entry for a location address
> */
> struct used_bucket *bucket_iter;
> + struct used_bucket *bucket_iter_tmp;
> struct relocation_head *rel_head_iter;
> + struct hlist_node *rel_head_iter_tmp;
> struct relocation_entry *rel_entry_iter;
> + struct relocation_entry *rel_entry_iter_tmp;
> int curr_type;
> void *location;
> long buffer;
>
> - list_for_each_entry(bucket_iter, &used_buckets_list, head) {
> - hlist_for_each_entry(rel_head_iter, bucket_iter->bucket, node) {
> + list_for_each_entry_safe(bucket_iter, bucket_iter_tmp,
> + used_buckets_list, head) {
> + hlist_for_each_entry_safe(rel_head_iter, rel_head_iter_tmp,
> + bucket_iter->bucket, node) {
> buffer = 0;
> location = rel_head_iter->location;
> - list_for_each_entry(rel_entry_iter,
> - rel_head_iter->rel_entry, head) {
> + list_for_each_entry_safe(rel_entry_iter,
> + rel_entry_iter_tmp,
> + rel_head_iter->rel_entry,
> + head) {
> curr_type = rel_entry_iter->type;
> reloc_handlers[curr_type].reloc_handler(
> me, &buffer, rel_entry_iter->value);
> @@ -648,11 +660,13 @@ void process_accumulated_relocations(struct module *me)
> kfree(bucket_iter);
> }
>
> - kfree(relocation_hashtable);
> + kfree(*relocation_hashtable);
> }
>
> int add_relocation_to_accumulate(struct module *me, int type, void *location,
> - unsigned int hashtable_bits, Elf_Addr v)
> + unsigned int hashtable_bits, Elf_Addr v,
> + struct hlist_head **relocation_hashtable,
> + struct list_head *used_buckets_list)
> {
> struct relocation_entry *entry;
> struct relocation_head *rel_head;
> @@ -667,7 +681,7 @@ int add_relocation_to_accumulate(struct module *me, int type, void *location,
>
> hash = hash_min((uintptr_t)location, hashtable_bits);
>
> - current_head = &relocation_hashtable[hash];
> + current_head = &((*relocation_hashtable)[hash]);
>
> /* Find matching location (if any) */
> bool found = false;
> @@ -693,7 +707,7 @@ int add_relocation_to_accumulate(struct module *me, int type, void *location,
> kmalloc(sizeof(struct used_bucket), GFP_KERNEL);
> INIT_LIST_HEAD(&bucket->head);
> bucket->bucket = current_head;
> - list_add(&bucket->head, &used_buckets_list);
> + list_add(&bucket->head, used_buckets_list);
> }
> hlist_add_head(&rel_head->node, current_head);
> }
> @@ -704,7 +718,10 @@ int add_relocation_to_accumulate(struct module *me, int type, void *location,
> return 0;
> }
>
> -unsigned int initialize_relocation_hashtable(unsigned int num_relocations)
> +unsigned int
> +initialize_relocation_hashtable(unsigned int num_relocations,
> + struct hlist_head **relocation_hashtable,
> + struct list_head *used_buckets_list)
> {
> /* Can safely assume that bits is not greater than sizeof(long) */
> unsigned long hashtable_size = roundup_pow_of_two(num_relocations);
> @@ -720,12 +737,12 @@ unsigned int initialize_relocation_hashtable(unsigned int num_relocations)
>
> hashtable_size <<= should_double_size;
>
> - relocation_hashtable = kmalloc_array(hashtable_size,
> - sizeof(*relocation_hashtable),
> - GFP_KERNEL);
> - __hash_init(relocation_hashtable, hashtable_size);
> + *relocation_hashtable = kmalloc_array(hashtable_size,
> + sizeof(*relocation_hashtable),
> + GFP_KERNEL);
You need to check for allocation failure here and inside
add_relocation_to_accumulate(). Module loading under memory pressure is a
reasonably likely scenario.
> + __hash_init(*relocation_hashtable, hashtable_size);
>
> - INIT_LIST_HEAD(&used_buckets_list);
> + INIT_LIST_HEAD(used_buckets_list);
This is the only place used_buckets_list is used in this function. If you move
this line out to apply_relocate_add, you can drop the parameter.
Regards,
Samuel
>
> return hashtable_bits;
> }
> @@ -742,7 +759,13 @@ int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,
> Elf_Addr v;
> int res;
> unsigned int num_relocations = sechdrs[relsec].sh_size / sizeof(*rel);
> - unsigned int hashtable_bits = initialize_relocation_hashtable(num_relocations);
> + struct hlist_head *relocation_hashtable;
> + struct list_head used_buckets_list;
> + unsigned int hashtable_bits;
> +
> + hashtable_bits = initialize_relocation_hashtable(num_relocations,
> + &relocation_hashtable,
> + &used_buckets_list);
>
> pr_debug("Applying relocate section %u to %u\n", relsec,
> sechdrs[relsec].sh_info);
> @@ -823,14 +846,17 @@ int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,
> }
>
> if (reloc_handlers[type].accumulate_handler)
> - res = add_relocation_to_accumulate(me, type, location, hashtable_bits, v);
> + res = add_relocation_to_accumulate(
> + me, type, location, hashtable_bits, v,
> + &relocation_hashtable, &used_buckets_list);
> else
> res = handler(me, location, v);
> if (res)
> return res;
> }
>
> - process_accumulated_relocations(me);
> + process_accumulated_relocations(me, &relocation_hashtable,
> + &used_buckets_list);
>
> return 0;
> }
>
> ---
> base-commit: 98b1cc82c4affc16f5598d4fa14b1858671b2263
> change-id: 20231120-module_linking_freeing-2b5a3b255b5e