| Message ID | 20221121170530.1196341-1-yhs@fb.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | BPF |
| Headers | show |
| Series | bpf: Add bpf_rcu_read_lock() support | expand |
On 11/21/22 9:05 AM, Yonghong Song wrote: > @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env, > return -EACCES; > } > > + /* Access rcu protected memory */ > + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && > + !env->cur_state->active_rcu_lock) { > + verbose(env, > + "R%d is ptr_%s access rcu-protected memory with off=%d, not rcu protected\n", > + regno, tname, off); > + return -EACCES; > + } > + > if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { > if (!btf_is_kernel(reg->btf)) { > verbose(env, "verifier internal error: reg->btf must be kernel btf\n"); > @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env, > if (ret < 0) > return ret; > > + /* The value is a rcu pointer. The load needs to be in a rcu lock region, > + * similar to rcu_dereference(). > + */ > + if ((flag & MEM_RCU) && env->prog->aux->sleepable && !env->cur_state->active_rcu_lock) { > + verbose(env, > + "R%d is rcu dereference ptr_%s with off=%d, not in rcu_read_lock region\n", > + regno, tname, off); > + return -EACCES; > + } Would this make the existing rdonly use case fail? SEC("fentry.s/" SYS_PREFIX "sys_getpgid") int task_real_parent(void *ctx) { struct task_struct *task, *real_parent; task = bpf_get_current_task_btf(); real_parent = task->real_parent; bpf_printk("pid %u\n", real_parent->pid); return 0; }
On 11/21/22 11:41 AM, Martin KaFai Lau wrote: > On 11/21/22 9:05 AM, Yonghong Song wrote: >> @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct >> bpf_verifier_env *env, >> return -EACCES; >> } >> + /* Access rcu protected memory */ >> + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && >> + !env->cur_state->active_rcu_lock) { >> + verbose(env, >> + "R%d is ptr_%s access rcu-protected memory with off=%d, >> not rcu protected\n", >> + regno, tname, off); >> + return -EACCES; >> + } >> + >> if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { >> if (!btf_is_kernel(reg->btf)) { >> verbose(env, "verifier internal error: reg->btf must be >> kernel btf\n"); >> @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct >> bpf_verifier_env *env, >> if (ret < 0) >> return ret; >> + /* The value is a rcu pointer. The load needs to be in a rcu lock >> region, >> + * similar to rcu_dereference(). >> + */ >> + if ((flag & MEM_RCU) && env->prog->aux->sleepable && >> !env->cur_state->active_rcu_lock) { >> + verbose(env, >> + "R%d is rcu dereference ptr_%s with off=%d, not in >> rcu_read_lock region\n", >> + regno, tname, off); >> + return -EACCES; >> + } > > Would this make the existing rdonly use case fail? > > SEC("fentry.s/" SYS_PREFIX "sys_getpgid") > int task_real_parent(void *ctx) > { > struct task_struct *task, *real_parent; > > task = bpf_get_current_task_btf(); > real_parent = task->real_parent; > bpf_printk("pid %u\n", real_parent->pid); > return 0; > } Right, it will fail. To fix the issue, user can do bpf_rcu_read_lock(); real_parent = task->real_parent; bpf_printk("pid %u\n", real_parent->pid); bpf_rcu_read_unlock(); But this raised a good question. How do we deal with legacy sleepable programs with newly-added rcu tagging capabilities. My current option is to error out if rcu usage is not right. But this might break existing sleepable programs. Another option intends to not break existing, like above, codes. In this case, MEM_RCU will not tagged if it is not inside bpf_rcu_read_lock() region. In this case, the above non-rcu-protected code should work. And the following should work as well although it is a little bit awkward. real_parent = task->real_parent; // real_parent not tagged with rcu bpf_rcu_read_lock(); bpf_printk("pid %u\n", real_parent->pid); bpf_rcu_read_unlock(); Maybe we should take the second choice in the above instead? >
On 11/21/22 12:01 PM, Yonghong Song wrote: > > > On 11/21/22 11:41 AM, Martin KaFai Lau wrote: >> On 11/21/22 9:05 AM, Yonghong Song wrote: >>> @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct >>> bpf_verifier_env *env, >>> return -EACCES; >>> } >>> + /* Access rcu protected memory */ >>> + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && >>> + !env->cur_state->active_rcu_lock) { >>> + verbose(env, >>> + "R%d is ptr_%s access rcu-protected memory with off=%d, not rcu >>> protected\n", >>> + regno, tname, off); >>> + return -EACCES; >>> + } >>> + >>> if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { >>> if (!btf_is_kernel(reg->btf)) { >>> verbose(env, "verifier internal error: reg->btf must be kernel >>> btf\n"); >>> @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct >>> bpf_verifier_env *env, >>> if (ret < 0) >>> return ret; >>> + /* The value is a rcu pointer. The load needs to be in a rcu lock region, >>> + * similar to rcu_dereference(). >>> + */ >>> + if ((flag & MEM_RCU) && env->prog->aux->sleepable && >>> !env->cur_state->active_rcu_lock) { >>> + verbose(env, >>> + "R%d is rcu dereference ptr_%s with off=%d, not in rcu_read_lock >>> region\n", >>> + regno, tname, off); >>> + return -EACCES; >>> + } >> >> Would this make the existing rdonly use case fail? >> >> SEC("fentry.s/" SYS_PREFIX "sys_getpgid") >> int task_real_parent(void *ctx) >> { >> struct task_struct *task, *real_parent; >> >> task = bpf_get_current_task_btf(); >> real_parent = task->real_parent; >> bpf_printk("pid %u\n", real_parent->pid); >> return 0; >> } > > Right, it will fail. To fix the issue, user can do > bpf_rcu_read_lock(); > real_parent = task->real_parent; > bpf_printk("pid %u\n", real_parent->pid); > bpf_rcu_read_unlock(); > > But this raised a good question. How do we deal with > legacy sleepable programs with newly-added rcu tagging > capabilities. > > My current option is to error out if rcu usage is not right. > But this might break existing sleepable programs. > > Another option intends to not break existing, like above, > codes. In this case, MEM_RCU will not tagged if it is > not inside bpf_rcu_read_lock() region. hmm.... it is to make MEM_RCU to mean a reg is protected by the current active_rcu_lock or not? > In this case, the above non-rcu-protected code should work. And the > following should work as well although it is a little > bit awkward. > real_parent = task->real_parent; // real_parent not tagged with rcu > bpf_rcu_read_lock(); > bpf_printk("pid %u\n", real_parent->pid); > bpf_rcu_read_unlock(); I think it should be fine. bpf_rcu_read_lock() just not useful in this example but nothing break or crash. Also, after bpf_rcu_read_unlock(), real_parent will continue to be readable because the MEM_RCU is not set? On top of the active_rcu_lock, should MEM_RCU be set only when it is dereferenced from a PTR_TRUSTED ptr (or with ref_obj_id != 0)? I am thinking about the following more common case: /* bpf_get_current_task_btf() may need to be changed * to set PTR_TRUSTED at the retval? */ /* task: PTR_TO_BTF_ID | PTR_TRUSTED */ task = bpf_get_current_task_btf(); bpf_rcu_read_lock(); /* real_parent: PTR_TO_BTF_ID | PTR_TRUSTED | MEM_RCU */ real_parent = task->real_parent; /* bpf_task_acquire() needs to change to use refcount_inc_not_zero */ real_parent = bpf_task_acquire(real_parent); bpf_rcu_read_unlock(); /* real_parent is accessible here (after checking NULL) and * can be passed to kfunc */
On 11/21/22 2:56 PM, Martin KaFai Lau wrote: > On 11/21/22 12:01 PM, Yonghong Song wrote: >> >> >> On 11/21/22 11:41 AM, Martin KaFai Lau wrote: >>> On 11/21/22 9:05 AM, Yonghong Song wrote: >>>> @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct >>>> bpf_verifier_env *env, >>>> return -EACCES; >>>> } >>>> + /* Access rcu protected memory */ >>>> + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && >>>> + !env->cur_state->active_rcu_lock) { >>>> + verbose(env, >>>> + "R%d is ptr_%s access rcu-protected memory with off=%d, >>>> not rcu protected\n", >>>> + regno, tname, off); >>>> + return -EACCES; >>>> + } >>>> + >>>> if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { >>>> if (!btf_is_kernel(reg->btf)) { >>>> verbose(env, "verifier internal error: reg->btf must >>>> be kernel btf\n"); >>>> @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct >>>> bpf_verifier_env *env, >>>> if (ret < 0) >>>> return ret; >>>> + /* The value is a rcu pointer. The load needs to be in a rcu >>>> lock region, >>>> + * similar to rcu_dereference(). >>>> + */ >>>> + if ((flag & MEM_RCU) && env->prog->aux->sleepable && >>>> !env->cur_state->active_rcu_lock) { >>>> + verbose(env, >>>> + "R%d is rcu dereference ptr_%s with off=%d, not in >>>> rcu_read_lock region\n", >>>> + regno, tname, off); >>>> + return -EACCES; >>>> + } >>> >>> Would this make the existing rdonly use case fail? >>> >>> SEC("fentry.s/" SYS_PREFIX "sys_getpgid") >>> int task_real_parent(void *ctx) >>> { >>> struct task_struct *task, *real_parent; >>> >>> task = bpf_get_current_task_btf(); >>> real_parent = task->real_parent; >>> bpf_printk("pid %u\n", real_parent->pid); >>> return 0; >>> } >> >> Right, it will fail. To fix the issue, user can do >> bpf_rcu_read_lock(); >> real_parent = task->real_parent; >> bpf_printk("pid %u\n", real_parent->pid); >> bpf_rcu_read_unlock(); >> >> But this raised a good question. How do we deal with >> legacy sleepable programs with newly-added rcu tagging >> capabilities. >> >> My current option is to error out if rcu usage is not right. >> But this might break existing sleepable programs. >> >> Another option intends to not break existing, like above, >> codes. In this case, MEM_RCU will not tagged if it is >> not inside bpf_rcu_read_lock() region. > > hmm.... it is to make MEM_RCU to mean a reg is protected by the current > active_rcu_lock or not? Yes, for example, in 'real_parent = task->real_parent' where 'real_parent' in task_struct is tagged with __rcu in the struct definition. So the 'real_parent' variable in the above assignment will be tagged with MEM_RCU. > >> In this case, the above non-rcu-protected code should work. And the >> following should work as well although it is a little >> bit awkward. >> real_parent = task->real_parent; // real_parent not tagged with rcu >> bpf_rcu_read_lock(); >> bpf_printk("pid %u\n", real_parent->pid); >> bpf_rcu_read_unlock(); > > I think it should be fine. bpf_rcu_read_lock() just not useful in this > example but nothing break or crash. Also, after bpf_rcu_read_unlock(), > real_parent will continue to be readable because the MEM_RCU is not set? That is correct. the variable real_parent is not tagged with MEM_RCU and it will stay that way for the rest of its life cycle. With new PTR_TRUSTED mechanism, real_parent will be marked as normal PTR_TO_BTF_ID and it is not marked as PTR_UNTRUSTED for backward compatibility. So in the above code, real_parent->pid is just a normal load (not related to rcu/trusted/untrusted). People may think it is okay, but actually it does not okay. Verifier could add more state to issue proper warnings, but I am not sure whether it is worthwhile or not. As you mentioned, nothing breaks. It is just the current existing way. So we should be able to live with this. > > On top of the active_rcu_lock, should MEM_RCU be set only when it is > dereferenced from a PTR_TRUSTED ptr (or with ref_obj_id != 0)? I didn't consider PTR_TRUSTED because it is just introduced yesterday... My current implementation inherits the old ptr_to_btf_id way where by default any ptr_to_btf_id is trusted. But since we have PTR_TRUSTED we should be able to use it for a stronger guarantee. > I am thinking about the following more common case: > > /* bpf_get_current_task_btf() may need to be changed > * to set PTR_TRUSTED at the retval? > */ > /* task: PTR_TO_BTF_ID | PTR_TRUSTED */ > task = bpf_get_current_task_btf(); > > bpf_rcu_read_lock(); > > /* real_parent: PTR_TO_BTF_ID | PTR_TRUSTED | MEM_RCU */ > real_parent = task->real_parent; > > /* bpf_task_acquire() needs to change to use > refcount_inc_not_zero */ > real_parent = bpf_task_acquire(real_parent); > > bpf_rcu_read_unlock(); > > /* real_parent is accessible here (after checking NULL) and > * can be passed to kfunc > */ > Yes, the above is a typical use case. Or alternatively after real_parent = task->real_parent; /* use real_parent inside the bpf_rcu_read_lock() region */ I will try to utilize PTR_TRUSTED concept in the next revision.
Yonghong Song wrote: > > > On 11/21/22 2:56 PM, Martin KaFai Lau wrote: > > On 11/21/22 12:01 PM, Yonghong Song wrote: > >> > >> > >> On 11/21/22 11:41 AM, Martin KaFai Lau wrote: > >>> On 11/21/22 9:05 AM, Yonghong Song wrote: > >>>> @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct > >>>> bpf_verifier_env *env, > >>>> return -EACCES; > >>>> } > >>>> + /* Access rcu protected memory */ > >>>> + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && > >>>> + !env->cur_state->active_rcu_lock) { > >>>> + verbose(env, > >>>> + "R%d is ptr_%s access rcu-protected memory with off=%d, > >>>> not rcu protected\n", > >>>> + regno, tname, off); > >>>> + return -EACCES; > >>>> + } > >>>> + > >>>> if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { > >>>> if (!btf_is_kernel(reg->btf)) { > >>>> verbose(env, "verifier internal error: reg->btf must > >>>> be kernel btf\n"); > >>>> @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct > >>>> bpf_verifier_env *env, > >>>> if (ret < 0) > >>>> return ret; > >>>> + /* The value is a rcu pointer. The load needs to be in a rcu > >>>> lock region, > >>>> + * similar to rcu_dereference(). > >>>> + */ > >>>> + if ((flag & MEM_RCU) && env->prog->aux->sleepable && > >>>> !env->cur_state->active_rcu_lock) { > >>>> + verbose(env, > >>>> + "R%d is rcu dereference ptr_%s with off=%d, not in > >>>> rcu_read_lock region\n", > >>>> + regno, tname, off); > >>>> + return -EACCES; > >>>> + } > >>> > >>> Would this make the existing rdonly use case fail? > >>> > >>> SEC("fentry.s/" SYS_PREFIX "sys_getpgid") > >>> int task_real_parent(void *ctx) > >>> { > >>> struct task_struct *task, *real_parent; > >>> > >>> task = bpf_get_current_task_btf(); > >>> real_parent = task->real_parent; > >>> bpf_printk("pid %u\n", real_parent->pid); > >>> return 0; > >>> } > >> > >> Right, it will fail. To fix the issue, user can do > >> bpf_rcu_read_lock(); > >> real_parent = task->real_parent; > >> bpf_printk("pid %u\n", real_parent->pid); > >> bpf_rcu_read_unlock(); > >> > >> But this raised a good question. How do we deal with > >> legacy sleepable programs with newly-added rcu tagging > >> capabilities. > >> > >> My current option is to error out if rcu usage is not right. > >> But this might break existing sleepable programs. > >> > >> Another option intends to not break existing, like above, > >> codes. In this case, MEM_RCU will not tagged if it is > >> not inside bpf_rcu_read_lock() region. > > > > hmm.... it is to make MEM_RCU to mean a reg is protected by the current > > active_rcu_lock or not? > > Yes, for example, in 'real_parent = task->real_parent' where > 'real_parent' in task_struct is tagged with __rcu in the struct > definition. So the 'real_parent' variable in the above assignment > will be tagged with MEM_RCU. > > > > >> In this case, the above non-rcu-protected code should work. And the > >> following should work as well although it is a little > >> bit awkward. > >> real_parent = task->real_parent; // real_parent not tagged with rcu > >> bpf_rcu_read_lock(); > >> bpf_printk("pid %u\n", real_parent->pid); > >> bpf_rcu_read_unlock(); > > > > I think it should be fine. bpf_rcu_read_lock() just not useful in this > > example but nothing break or crash. Also, after bpf_rcu_read_unlock(), > > real_parent will continue to be readable because the MEM_RCU is not set? > > That is correct. the variable real_parent is not tagged with MEM_RCU > and it will stay that way for the rest of its life cycle. > > With new PTR_TRUSTED mechanism, real_parent will be marked as normal > PTR_TO_BTF_ID and it is not marked as PTR_UNTRUSTED for backward > compatibility. So in the above code, real_parent->pid is just a normal > load (not related to rcu/trusted/untrusted). People may think it > is okay, but actually it does not okay. Verifier could add more state > to issue proper warnings, but I am not sure whether it is worthwhile > or not. As you mentioned, nothing breaks. It is just the current > existing way. So we should be able to live with this. > > > > > On top of the active_rcu_lock, should MEM_RCU be set only when it is > > dereferenced from a PTR_TRUSTED ptr (or with ref_obj_id != 0)? > > I didn't consider PTR_TRUSTED because it is just introduced yesterday... > > My current implementation inherits the old ptr_to_btf_id way where by > default any ptr_to_btf_id is trusted. But since we have PTR_TRUSTED > we should be able to use it for a stronger guarantee. > > > I am thinking about the following more common case: > > > > /* bpf_get_current_task_btf() may need to be changed > > * to set PTR_TRUSTED at the retval? > > */ > > /* task: PTR_TO_BTF_ID | PTR_TRUSTED */ > > task = bpf_get_current_task_btf(); > > > > bpf_rcu_read_lock(); > > > > /* real_parent: PTR_TO_BTF_ID | PTR_TRUSTED | MEM_RCU */ > > real_parent = task->real_parent; > > > > /* bpf_task_acquire() needs to change to use > > refcount_inc_not_zero */ > > real_parent = bpf_task_acquire(real_parent); > > > > bpf_rcu_read_unlock(); > > > > /* real_parent is accessible here (after checking NULL) and > > * can be passed to kfunc > > */ > > > > Yes, the above is a typical use case. Or alternatively after > real_parent = task->real_parent; > /* use real_parent inside the bpf_rcu_read_lock() region */ > > I will try to utilize PTR_TRUSTED concept in the next revision. Also perhaps interesting is when task is read out of a map with reference already pinned. I think you should clear the MEM_RCU tag on all referenced objects?
On 11/21/22 6:03 PM, John Fastabend wrote: > Yonghong Song wrote: >> >> >> On 11/21/22 2:56 PM, Martin KaFai Lau wrote: >>> On 11/21/22 12:01 PM, Yonghong Song wrote: >>>> >>>> >>>> On 11/21/22 11:41 AM, Martin KaFai Lau wrote: >>>>> On 11/21/22 9:05 AM, Yonghong Song wrote: >>>>>> @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct >>>>>> bpf_verifier_env *env, >>>>>> return -EACCES; >>>>>> } >>>>>> + /* Access rcu protected memory */ >>>>>> + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && >>>>>> + !env->cur_state->active_rcu_lock) { >>>>>> + verbose(env, >>>>>> + "R%d is ptr_%s access rcu-protected memory with off=%d, >>>>>> not rcu protected\n", >>>>>> + regno, tname, off); >>>>>> + return -EACCES; >>>>>> + } >>>>>> + >>>>>> if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { >>>>>> if (!btf_is_kernel(reg->btf)) { >>>>>> verbose(env, "verifier internal error: reg->btf must >>>>>> be kernel btf\n"); >>>>>> @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct >>>>>> bpf_verifier_env *env, >>>>>> if (ret < 0) >>>>>> return ret; >>>>>> + /* The value is a rcu pointer. The load needs to be in a rcu >>>>>> lock region, >>>>>> + * similar to rcu_dereference(). >>>>>> + */ >>>>>> + if ((flag & MEM_RCU) && env->prog->aux->sleepable && >>>>>> !env->cur_state->active_rcu_lock) { >>>>>> + verbose(env, >>>>>> + "R%d is rcu dereference ptr_%s with off=%d, not in >>>>>> rcu_read_lock region\n", >>>>>> + regno, tname, off); >>>>>> + return -EACCES; >>>>>> + } >>>>> >>>>> Would this make the existing rdonly use case fail? >>>>> >>>>> SEC("fentry.s/" SYS_PREFIX "sys_getpgid") >>>>> int task_real_parent(void *ctx) >>>>> { >>>>> struct task_struct *task, *real_parent; >>>>> >>>>> task = bpf_get_current_task_btf(); >>>>> real_parent = task->real_parent; >>>>> bpf_printk("pid %u\n", real_parent->pid); >>>>> return 0; >>>>> } >>>> >>>> Right, it will fail. To fix the issue, user can do >>>> bpf_rcu_read_lock(); >>>> real_parent = task->real_parent; >>>> bpf_printk("pid %u\n", real_parent->pid); >>>> bpf_rcu_read_unlock(); >>>> >>>> But this raised a good question. How do we deal with >>>> legacy sleepable programs with newly-added rcu tagging >>>> capabilities. >>>> >>>> My current option is to error out if rcu usage is not right. >>>> But this might break existing sleepable programs. >>>> >>>> Another option intends to not break existing, like above, >>>> codes. In this case, MEM_RCU will not tagged if it is >>>> not inside bpf_rcu_read_lock() region. >>> >>> hmm.... it is to make MEM_RCU to mean a reg is protected by the current >>> active_rcu_lock or not? >> >> Yes, for example, in 'real_parent = task->real_parent' where >> 'real_parent' in task_struct is tagged with __rcu in the struct >> definition. So the 'real_parent' variable in the above assignment >> will be tagged with MEM_RCU. >> >>> >>>> In this case, the above non-rcu-protected code should work. And the >>>> following should work as well although it is a little >>>> bit awkward. >>>> real_parent = task->real_parent; // real_parent not tagged with rcu >>>> bpf_rcu_read_lock(); >>>> bpf_printk("pid %u\n", real_parent->pid); >>>> bpf_rcu_read_unlock(); >>> >>> I think it should be fine. bpf_rcu_read_lock() just not useful in this >>> example but nothing break or crash. Also, after bpf_rcu_read_unlock(), >>> real_parent will continue to be readable because the MEM_RCU is not set? >> >> That is correct. the variable real_parent is not tagged with MEM_RCU >> and it will stay that way for the rest of its life cycle. >> >> With new PTR_TRUSTED mechanism, real_parent will be marked as normal >> PTR_TO_BTF_ID and it is not marked as PTR_UNTRUSTED for backward >> compatibility. So in the above code, real_parent->pid is just a normal >> load (not related to rcu/trusted/untrusted). People may think it >> is okay, but actually it does not okay. Verifier could add more state >> to issue proper warnings, but I am not sure whether it is worthwhile >> or not. As you mentioned, nothing breaks. It is just the current >> existing way. So we should be able to live with this. >> >>> >>> On top of the active_rcu_lock, should MEM_RCU be set only when it is >>> dereferenced from a PTR_TRUSTED ptr (or with ref_obj_id != 0)? >> >> I didn't consider PTR_TRUSTED because it is just introduced yesterday... >> >> My current implementation inherits the old ptr_to_btf_id way where by >> default any ptr_to_btf_id is trusted. But since we have PTR_TRUSTED >> we should be able to use it for a stronger guarantee. >> >>> I am thinking about the following more common case: >>> >>> /* bpf_get_current_task_btf() may need to be changed >>> * to set PTR_TRUSTED at the retval? >>> */ >>> /* task: PTR_TO_BTF_ID | PTR_TRUSTED */ >>> task = bpf_get_current_task_btf(); >>> >>> bpf_rcu_read_lock(); >>> >>> /* real_parent: PTR_TO_BTF_ID | PTR_TRUSTED | MEM_RCU */ >>> real_parent = task->real_parent; >>> >>> /* bpf_task_acquire() needs to change to use >>> refcount_inc_not_zero */ >>> real_parent = bpf_task_acquire(real_parent); >>> >>> bpf_rcu_read_unlock(); >>> >>> /* real_parent is accessible here (after checking NULL) and >>> * can be passed to kfunc >>> */ >>> >> >> Yes, the above is a typical use case. Or alternatively after >> real_parent = task->real_parent; >> /* use real_parent inside the bpf_rcu_read_lock() region */ >> >> I will try to utilize PTR_TRUSTED concept in the next revision. > > Also perhaps interesting is when task is read out of a map > with reference already pinned. I think you should clear > the MEM_RCU tag on all referenced objects? The register tagged with MEM_RCU will not be a referenced obj. MEM_RCU tag only appears to a register inside the rcu read lock region as the rcu_reference() result. So the obj tagged with MEM_RCU is protected with rcu read lock and it is valid and trusted and there is no need to acquire additional reference. If user calls another kfunc to acquire a reference, then the resulted ptr will not have MEM_RCU tag but with non-zero ref_obj_id. The MEM_RCU reg will be invalidated when seeing bpf_rcu_read_unlock() to prevent rcu-protected ptr to leak out of the rcu read lock region.
On 11/21/22 9:05 AM, Yonghong Song wrote: > + if (env->cur_state->active_rcu_lock) { > + if (bpf_lsm_sleepable_func_proto(func_id) || > + bpf_tracing_sleepable_func_proto(func_id)) { > + verbose(env, "sleepable helper %s#%din rcu_read_lock region\n", > + func_id_name(func_id), func_id); > + return -EINVAL; > + } > + Even after patch 2 refactoring the above bit is still quite fragile. Ex: bpf_d_path is not included, but it should be. How about we add 'bool might_sleep' to bpf_func_proto and mark existing 5 functions with it and refactor patch 2 differently. We won't be doing prog->aux->sleepable ? in bpf_tracing_func_proto() anymore. Those cbs will be returning func_proto-s, but the verifier later will check might_sleep flag.
On 11/21/22 9:48 PM, Alexei Starovoitov wrote: > On 11/21/22 9:05 AM, Yonghong Song wrote: >> + if (env->cur_state->active_rcu_lock) { >> + if (bpf_lsm_sleepable_func_proto(func_id) || >> + bpf_tracing_sleepable_func_proto(func_id)) { >> + verbose(env, "sleepable helper %s#%din rcu_read_lock >> region\n", >> + func_id_name(func_id), func_id); >> + return -EINVAL; >> + } >> + > > Even after patch 2 refactoring the above bit is still quite fragile. > Ex: bpf_d_path is not included, but it should be. > > How about we add 'bool might_sleep' to bpf_func_proto and mark existing > 5 functions with it and refactor patch 2 differently. > We won't be doing prog->aux->sleepable ? in bpf_tracing_func_proto() > anymore. > Those cbs will be returning func_proto-s, > but the verifier later will check might_sleep flag. Ya, bpf_func_proto->might_sleep indeed better. I could do that. The only problem is bpf_d_path. static bool bpf_d_path_allowed(const struct bpf_prog *prog) { if (prog->type == BPF_PROG_TYPE_TRACING && prog->expected_attach_type == BPF_TRACE_ITER) return true; if (prog->type == BPF_PROG_TYPE_LSM) return bpf_lsm_is_sleepable_hook(prog->aux->attach_btf_id); return btf_id_set_contains(&btf_allowlist_d_path, prog->aux->attach_btf_id); } If I understand correctly, bpf_d_path helper doesn't mean the helper itself will be sleepable. For example, bpf_d_path can only appear in sleepable programs if program type is BPF_PROG_TYPE_LSM, from 6f100640ca5b ("bpf: Expose bpf_d_path helper to sleepable LSM hooks") it looks like the reason is those sleepable lsm programs provide better context so bpf_d_path won't have potential lock or other issues. So essentially, bpf_d_path helper itself won't be a helper causing the prog to sleep, right? If this is the case, we only assign might_sleepable to the other 4 helpers.
diff --git a/include/linux/bpf.h b/include/linux/bpf.h index c9eafa67f2a2..d0fbb7f0f13b 100644 --- a/include/linux/bpf.h +++ b/include/linux/bpf.h @@ -572,6 +572,9 @@ enum bpf_type_flag { */ PTR_TRUSTED = BIT(12 + BPF_BASE_TYPE_BITS), + /* MEM is tagged with rcu and memory access needs rcu_read_lock protection. */ + MEM_RCU = BIT(13 + BPF_BASE_TYPE_BITS), + __BPF_TYPE_FLAG_MAX, __BPF_TYPE_LAST_FLAG = __BPF_TYPE_FLAG_MAX - 1, }; diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h index 545152ac136c..748431578902 100644 --- a/include/linux/bpf_verifier.h +++ b/include/linux/bpf_verifier.h @@ -344,6 +344,7 @@ struct bpf_verifier_state { u32 id; } active_lock; bool speculative; + bool active_rcu_lock; /* first and last insn idx of this verifier state */ u32 first_insn_idx; @@ -445,6 +446,7 @@ struct bpf_insn_aux_data { u32 seen; /* this insn was processed by the verifier at env->pass_cnt */ bool sanitize_stack_spill; /* subject to Spectre v4 sanitation */ bool zext_dst; /* this insn zero extends dst reg */ + bool storage_get_func_atomic; /* bpf_*_storage_get() with atomic memory alloc */ u8 alu_state; /* used in combination with alu_limit */ /* below fields are initialized once */ diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c index 1a59cc7ad730..68df0df27302 100644 --- a/kernel/bpf/btf.c +++ b/kernel/bpf/btf.c @@ -6237,6 +6237,9 @@ static int btf_struct_walk(struct bpf_verifier_log *log, const struct btf *btf, /* check __percpu tag */ if (strcmp(tag_value, "percpu") == 0) tmp_flag = MEM_PERCPU; + /* check __rcu tag */ + if (strcmp(tag_value, "rcu") == 0) + tmp_flag = MEM_RCU; } stype = btf_type_skip_modifiers(btf, mtype->type, &id); diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c index 9e8a5557ea8d..3268be36a319 100644 --- a/kernel/bpf/helpers.c +++ b/kernel/bpf/helpers.c @@ -1889,6 +1889,16 @@ void *bpf_rdonly_cast(void *obj__ign, u32 btf_id__k) return obj__ign; } +void bpf_rcu_read_lock(void) +{ + rcu_read_lock(); +} + +void bpf_rcu_read_unlock(void) +{ + rcu_read_unlock(); +} + __diag_pop(); BTF_SET8_START(generic_btf_ids) @@ -1919,6 +1929,8 @@ BTF_ID(func, bpf_task_release) BTF_SET8_START(common_btf_ids) BTF_ID_FLAGS(func, bpf_cast_to_kern_ctx) BTF_ID_FLAGS(func, bpf_rdonly_cast) +BTF_ID_FLAGS(func, bpf_rcu_read_lock) +BTF_ID_FLAGS(func, bpf_rcu_read_unlock) BTF_SET8_END(common_btf_ids) static const struct btf_kfunc_id_set common_kfunc_set = { diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 9528a066cfa5..d39ab0e969dd 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -23,6 +23,7 @@ #include <linux/error-injection.h> #include <linux/bpf_lsm.h> #include <linux/btf_ids.h> +#include <linux/trace_events.h> #include <linux/poison.h> #include "disasm.h" @@ -527,6 +528,14 @@ static bool is_callback_calling_function(enum bpf_func_id func_id) func_id == BPF_FUNC_user_ringbuf_drain; } +static bool is_storage_get_function(enum bpf_func_id func_id) +{ + return func_id == BPF_FUNC_sk_storage_get || + func_id == BPF_FUNC_inode_storage_get || + func_id == BPF_FUNC_task_storage_get || + func_id == BPF_FUNC_cgrp_storage_get; +} + static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id, const struct bpf_map *map) { @@ -589,11 +598,12 @@ static const char *reg_type_str(struct bpf_verifier_env *env, strncpy(postfix, "_or_null", 16); } - snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s", + snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s", type & MEM_RDONLY ? "rdonly_" : "", type & MEM_RINGBUF ? "ringbuf_" : "", type & MEM_USER ? "user_" : "", type & MEM_PERCPU ? "percpu_" : "", + type & MEM_RCU ? "rcu_" : "", type & PTR_UNTRUSTED ? "untrusted_" : "", type & PTR_TRUSTED ? "trusted_" : "" ); @@ -1220,6 +1230,7 @@ static int copy_verifier_state(struct bpf_verifier_state *dst_state, dst_state->frame[i] = NULL; } dst_state->speculative = src->speculative; + dst_state->active_rcu_lock = src->active_rcu_lock; dst_state->curframe = src->curframe; dst_state->active_lock.ptr = src->active_lock.ptr; dst_state->active_lock.id = src->active_lock.id; @@ -4704,6 +4715,15 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env, return -EACCES; } + /* Access rcu protected memory */ + if ((reg->type & MEM_RCU) && env->prog->aux->sleepable && + !env->cur_state->active_rcu_lock) { + verbose(env, + "R%d is ptr_%s access rcu-protected memory with off=%d, not rcu protected\n", + regno, tname, off); + return -EACCES; + } + if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) { if (!btf_is_kernel(reg->btf)) { verbose(env, "verifier internal error: reg->btf must be kernel btf\n"); @@ -4731,12 +4751,27 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env, if (ret < 0) return ret; + /* The value is a rcu pointer. The load needs to be in a rcu lock region, + * similar to rcu_dereference(). + */ + if ((flag & MEM_RCU) && env->prog->aux->sleepable && !env->cur_state->active_rcu_lock) { + verbose(env, + "R%d is rcu dereference ptr_%s with off=%d, not in rcu_read_lock region\n", + regno, tname, off); + return -EACCES; + } + /* If this is an untrusted pointer, all pointers formed by walking it * also inherit the untrusted flag. */ if (type_flag(reg->type) & PTR_UNTRUSTED) flag |= PTR_UNTRUSTED; + /* Mark the pointee of the rcu protected memory access as PTR_UNTRUSTED */ + if (env->cur_state->active_rcu_lock && !(flag & PTR_UNTRUSTED) && + (reg->type & MEM_RCU) && !(flag & MEM_RCU)) + flag |= PTR_UNTRUSTED; + /* Any pointer obtained from walking a trusted pointer is no longer trusted. */ flag &= ~PTR_TRUSTED; @@ -5896,6 +5931,7 @@ static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_T static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID, + PTR_TO_BTF_ID | MEM_RCU, PTR_TO_BTF_ID | PTR_TRUSTED, }, }; @@ -5976,6 +6012,20 @@ static int check_reg_type(struct bpf_verifier_env *env, u32 regno, if (arg_type & PTR_MAYBE_NULL) type &= ~PTR_MAYBE_NULL; + /* If the reg type is marked as MEM_RCU, ensure the usage is in the rcu_read_lock + * region, and remove MEM_RCU from the type since the arg_type won't encode + * MEM_RCU. + */ + if (type & MEM_RCU) { + if (env->prog->aux->sleepable && !env->cur_state->active_rcu_lock) { + verbose(env, + "R%d is arg type %s needs rcu protection\n", + regno, reg_type_str(env, reg->type)); + return -EACCES; + } + type &= ~MEM_RCU; + } + for (i = 0; i < ARRAY_SIZE(compatible->types); i++) { expected = compatible->types[i]; if (expected == NOT_INIT) @@ -5992,7 +6042,8 @@ static int check_reg_type(struct bpf_verifier_env *env, u32 regno, return -EACCES; found: - if (reg->type == PTR_TO_BTF_ID || reg->type & PTR_TRUSTED) { + /* reg is already protected by rcu_read_lock(). Peel off MEM_RCU from reg->type. */ + if ((reg->type & ~MEM_RCU) == PTR_TO_BTF_ID || reg->type & PTR_TRUSTED) { /* For bpf_sk_release, it needs to match against first member * 'struct sock_common', hence make an exception for it. This * allows bpf_sk_release to work for multiple socket types. @@ -6073,6 +6124,7 @@ int check_func_arg_reg_off(struct bpf_verifier_env *env, * fixed offset. */ case PTR_TO_BTF_ID: + case PTR_TO_BTF_ID | MEM_RCU: case PTR_TO_BTF_ID | MEM_ALLOC: case PTR_TO_BTF_ID | PTR_TRUSTED: case PTR_TO_BTF_ID | MEM_ALLOC | PTR_TRUSTED: @@ -7534,6 +7586,18 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn return err; } + if (env->cur_state->active_rcu_lock) { + if (bpf_lsm_sleepable_func_proto(func_id) || + bpf_tracing_sleepable_func_proto(func_id)) { + verbose(env, "sleepable helper %s#%din rcu_read_lock region\n", + func_id_name(func_id), func_id); + return -EINVAL; + } + + if (env->prog->aux->sleepable && is_storage_get_function(func_id)) + env->insn_aux_data[insn_idx].storage_get_func_atomic = true; + } + meta.func_id = func_id; /* check args */ for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) { @@ -8158,6 +8222,8 @@ enum special_kfunc_type { KF_bpf_list_pop_back, KF_bpf_cast_to_kern_ctx, KF_bpf_rdonly_cast, + KF_bpf_rcu_read_lock, + KF_bpf_rcu_read_unlock, }; BTF_SET_START(special_kfunc_set) @@ -8180,6 +8246,18 @@ BTF_ID(func, bpf_list_pop_front) BTF_ID(func, bpf_list_pop_back) BTF_ID(func, bpf_cast_to_kern_ctx) BTF_ID(func, bpf_rdonly_cast) +BTF_ID(func, bpf_rcu_read_lock) +BTF_ID(func, bpf_rcu_read_unlock) + +static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta) +{ + return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock]; +} + +static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta) +{ + return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock]; +} static enum kfunc_ptr_arg_type get_kfunc_ptr_arg_type(struct bpf_verifier_env *env, @@ -8812,6 +8890,7 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, const struct btf_type *t, *func, *func_proto, *ptr_type; struct bpf_reg_state *regs = cur_regs(env); const char *func_name, *ptr_type_name; + bool sleepable, rcu_lock, rcu_unlock; struct bpf_kfunc_call_arg_meta meta; u32 i, nargs, func_id, ptr_type_id; int err, insn_idx = *insn_idx_p; @@ -8853,11 +8932,38 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, return -EACCES; } - if (is_kfunc_sleepable(&meta) && !env->prog->aux->sleepable) { + sleepable = is_kfunc_sleepable(&meta); + if (sleepable && !env->prog->aux->sleepable) { verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name); return -EACCES; } + rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta); + rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta); + if (env->cur_state->active_rcu_lock) { + struct bpf_func_state *state; + struct bpf_reg_state *reg; + + if (rcu_lock) { + verbose(env, "nested rcu read lock (kernel function %s)\n", func_name); + return -EINVAL; + } else if (rcu_unlock) { + bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({ + if (reg->type & MEM_RCU) + __mark_reg_unknown(env, reg); + })); + env->cur_state->active_rcu_lock = false; + } else if (sleepable) { + verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name); + return -EACCES; + } + } else if (rcu_lock) { + env->cur_state->active_rcu_lock = true; + } else if (rcu_unlock) { + verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name); + return -EINVAL; + } + /* Check the arguments */ err = check_kfunc_args(env, &meta); if (err < 0) @@ -11749,6 +11855,11 @@ static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn) return -EINVAL; } + if (env->prog->aux->sleepable && env->cur_state->active_rcu_lock) { + verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n"); + return -EINVAL; + } + if (regs[ctx_reg].type != PTR_TO_CTX) { verbose(env, "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n"); @@ -13014,6 +13125,9 @@ static bool states_equal(struct bpf_verifier_env *env, old->active_lock.id != cur->active_lock.id) return false; + if (old->active_rcu_lock != cur->active_rcu_lock) + return false; + /* for states to be equal callsites have to be the same * and all frame states need to be equivalent */ @@ -13427,6 +13541,11 @@ static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev) !reg_type_mismatch_ok(prev)); } +static bool reg_type_mismatch_ignore_rcu(enum bpf_reg_type src, enum bpf_reg_type prev) +{ + return reg_type_mismatch(src & ~MEM_RCU, prev & ~MEM_RCU); +} + static int do_check(struct bpf_verifier_env *env) { bool pop_log = !(env->log.level & BPF_LOG_LEVEL2); @@ -13552,6 +13671,17 @@ static int do_check(struct bpf_verifier_env *env) prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type; + /* For code like below, + * struct foo *f; + * if (...) + * f = ...; // f with MEM_RCU type tag. + * else + * f = ...; // f without MEM_RCU type tag. + * ... f ... // Here f could be with/without MEM_RCU + * + * It is safe to ignore MEM_RCU type tag here since + * base types are the same. + */ if (*prev_src_type == NOT_INIT) { /* saw a valid insn * dst_reg = *(u32 *)(src_reg + off) @@ -13559,7 +13689,7 @@ static int do_check(struct bpf_verifier_env *env) */ *prev_src_type = src_reg_type; - } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) { + } else if (reg_type_mismatch_ignore_rcu(src_reg_type, *prev_src_type)) { /* ABuser program is trying to use the same insn * dst_reg = *(u32*) (src_reg + off) * with different pointer types: @@ -13701,6 +13831,11 @@ static int do_check(struct bpf_verifier_env *env) return -EINVAL; } + if (env->cur_state->active_rcu_lock) { + verbose(env, "bpf_rcu_read_unlock is missing\n"); + return -EINVAL; + } + /* We must do check_reference_leak here before * prepare_func_exit to handle the case when * state->curframe > 0, it may be a callback @@ -14795,6 +14930,7 @@ static int convert_ctx_accesses(struct bpf_verifier_env *env) convert_ctx_access = bpf_xdp_sock_convert_ctx_access; break; case PTR_TO_BTF_ID: + case PTR_TO_BTF_ID | MEM_RCU: case PTR_TO_BTF_ID | PTR_UNTRUSTED: case PTR_TO_BTF_ID | PTR_TRUSTED: /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike @@ -15489,14 +15625,12 @@ static int do_misc_fixups(struct bpf_verifier_env *env) goto patch_call_imm; } - if (insn->imm == BPF_FUNC_task_storage_get || - insn->imm == BPF_FUNC_sk_storage_get || - insn->imm == BPF_FUNC_inode_storage_get || - insn->imm == BPF_FUNC_cgrp_storage_get) { - if (env->prog->aux->sleepable) - insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL); - else + if (is_storage_get_function(insn->imm)) { + if (!env->prog->aux->sleepable || + env->insn_aux_data[i + delta].storage_get_func_atomic) insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC); + else + insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL); insn_buf[1] = *insn; cnt = 2;
Add two kfunc's bpf_rcu_read_lock() and bpf_rcu_read_unlock(). These two kfunc's can be used for all program types. The following is an example about how rcu pointer are used w.r.t. bpf_rcu_read_lock()/bpf_rcu_read_unlock(). struct task_struct { ... struct task_struct *last_wakee; struct task_struct __rcu *real_parent; ... }; Let us say prog does 'task = bpf_get_current_task_btf()' to get a 'task' pointer. The basic rules are: - 'real_parent = task->real_parent' should be inside bpf_rcu_read_lock region. this is to simulate rcu_dereference() operation. - 'last_wakee = real_parent->last_wakee' should be iinside bpf_rcu_read_lock region since it tries to access rcu protected memory. - the ptr 'last_wakee' will be marked as PTR_UNTRUSTED since in general it is not clear whether the object pointed by 'last_wakee' is valid or not even inside bpf_rcu_read_lock region. Note that for non-sleepable progs, it is permitted that rcu pointer load and accessing rcu protected memory can be outside rcu read lock region. But verification will be reported for sleepable progs. To prevent rcu pointer leaks outside the rcu read lock region. The verifier will clear all rcu pointer register state to unknown, i.e., scalar_value, so later dereference becomes impossible. Signed-off-by: Yonghong Song <yhs@fb.com> --- include/linux/bpf.h | 3 + include/linux/bpf_verifier.h | 2 + kernel/bpf/btf.c | 3 + kernel/bpf/helpers.c | 12 +++ kernel/bpf/verifier.c | 156 ++++++++++++++++++++++++++++++++--- 5 files changed, 165 insertions(+), 11 deletions(-)