Page MenuHomeFreeBSD
Differential D42189

bpf: Make dead_bpf_if const
ClosedPublic
Actions

Authored by zlei on Oct 13 2023, 2:00 PM.
Tags
None
Referenced Files
Unknown Object (File)
Sat, Apr 5, 3:34 AM
Unknown Object (File)
Feb 10 2025, 7:47 AM
Unknown Object (File)
Feb 9 2025, 9:54 AM
Unknown Object (File)
Feb 5 2025, 2:21 PM
Unknown Object (File)
Feb 5 2025, 4:47 AM
Unknown Object (File)
Feb 5 2025, 2:33 AM
Unknown Object (File)
Feb 3 2025, 11:12 AM
Unknown Object (File)
Jan 28 2025, 11:16 PM
View All 61 Files
Subscribers
glebius
imp
melifaro

Details

Reviewers
ae
markj
Group Reviewers
network
Commits
rG281f61d211b3: bpf: Make dead_bpf_if const
rG1e147980497b: bpf: Make dead_bpf_if const
rG5d031613affa: bpf: Make dead_bpf_if const
rG7a974a649848: bpf: Make dead_bpf_if const
Summary

The dead_bpf_if is not subjected to be written. Make it const so that the kernel can panic on destructive writing to it instead of silently memory corruption.

No functional change intended.

MFC after: 1 week

Diff Detail

Repository
rG FreeBSD src repository
Lint
Lint Skipped
Unit
Tests Skipped

Event Timeline

zlei created this revision.Oct 13 2023, 2:00 PM
Herald added subscribers: glebius, melifaro, imp. · View Herald TranscriptOct 13 2023, 2:00 PM
zlei requested review of this revision.Oct 13 2023, 2:00 PM
zlei edited the summary of this revision. (Show Details)
markj added inline comments.Oct 13 2023, 2:08 PM
sys/net/bpf.c
2862

Rather than casting away the const (which generally should be done with __DECONST), why not change the type of bif_bpf?

zlei added inline comments.Oct 13 2023, 2:23 PM
sys/net/bpf.c
2862

Emm, clang complains with

error: passing 'struct bpf_if ' to parameter of type 'const struct bpf_if ' discards qualifiers in nested pointer types [-Werror,-Wincompatible-pointer-types-discards-qualifiers]

if bif_bpf is const struct bpf_if ** .

zlei added inline comments.Oct 13 2023, 3:26 PM
sys/net/bpf.c
2862

Found a good explaining https://c-faq.com/ansi/constmismatch.html .

OK, a cast is much simpler. It make troubles to change the type of bif_bpf to const struct bpf_if ** .

zlei updated this revision to Diff 128728.Oct 13 2023, 3:34 PM

Use macro __DECONST.

markj added inline comments.Oct 13 2023, 3:35 PM
sys/net/bpf.c
2862

It make troubles to change the type of bif_bpf to const struct bpf_if ** .

It should be done anyway. Only bpf.c should be potentially modifying bpf_if fields anyway, so external references ought to be const: https://people.freebsd.org/~markj/patches/bpf_const.diff

In any case, please use __DECONST at least.

markj accepted this revision as: markj.Oct 13 2023, 3:35 PM
This revision is now accepted and ready to land.Oct 13 2023, 3:35 PM
zlei added inline comments.Oct 13 2023, 4:44 PM
sys/net/bpf.c
2862

It should be done anyway. Only bpf.c should be potentially modifying bpf_if fields anyway, so external references ought to be const

Reasonable.

zlei added a comment.Oct 20 2023, 11:40 AM

This change is quite simple. I'm going to commit this firstly.

Closed by commit rG7a974a649848: bpf: Make dead_bpf_if const (authored by zlei). · Explain WhyOct 21 2023, 4:55 AM
This revision was automatically updated to reflect the committed changes.
zlei added a commit: rG7a974a649848: bpf: Make dead_bpf_if const.
zlei added a commit: rG5d031613affa: bpf: Make dead_bpf_if const.Nov 13 2023, 3:58 AM
zlei added a commit: rG1e147980497b: bpf: Make dead_bpf_if const.
zlei added a commit: rG281f61d211b3: bpf: Make dead_bpf_if const.Nov 13 2023, 10:25 AM

Revision Contents
Changeset List

PathSize
sys/
net/
6 lines

Diff 128728

View Options

sys/net/bpf.c

Show First 20 Lines • Show All 97 Lines • ▼ Show 20 Lines
#endif #endif
int old_mlock = 0; int old_mlock = 0;
SYSCTL_INT(_vm, OID_AUTO, old_mlock, CTLFLAG_RWTUN, &old_mlock, 0, SYSCTL_INT(_vm, OID_AUTO, old_mlock, CTLFLAG_RWTUN, &old_mlock, 0,
"Do not apply RLIMIT_MEMLOCK on mlockall"); "Do not apply RLIMIT_MEMLOCK on mlockall");
static int mincore_mapped = 1; static int mincore_mapped = 1;
SYSCTL_INT(_vm, OID_AUTO, mincore_mapped, CTLFLAG_RWTUN, &mincore_mapped, 0, SYSCTL_INT(_vm, OID_AUTO, mincore_mapped, CTLFLAG_RWTUN, &mincore_mapped, 0,
"mincore reports mappings, not residency"); "mincore reports mappings, not residency");
static int imply_prot_max = 0;
SYSCTL_INT(_vm, OID_AUTO, imply_prot_max, CTLFLAG_RWTUN, &imply_prot_max, 0,
emasteUnsubmitted
Done
Inline Actions

s/mincore_mapped/imply_prot_max/

emaste: `s/mincore_mapped/imply_prot_max/`
"Imply maximum page permissions in mmap() when none are specified.");
markjUnsubmitted
Done
Inline Actions

sysctl description strings usually don't contain a period.

markj: sysctl description strings usually don't contain a period.
#ifdef MAP_32BIT #ifdef MAP_32BIT
#define MAP_32BIT_MAX_ADDR ((vm_offset_t)1 << 31) #define MAP_32BIT_MAX_ADDR ((vm_offset_t)1 << 31)
#endif #endif
#ifndef _SYS_SYSPROTO_H_ #ifndef _SYS_SYSPROTO_H_
struct sbrk_args { struct sbrk_args {
int incr; int incr;
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
kern_mmap(struct thread *td, uintptr_t addr0, size_t size, int prot, int flags, kern_mmap(struct thread *td, uintptr_t addr0, size_t size, int prot, int flags,
int fd, off_t pos) int fd, off_t pos)
{ {
struct vmspace *vms; struct vmspace *vms;
struct file *fp; struct file *fp;
vm_offset_t addr; vm_offset_t addr;
vm_size_t pageoff; vm_size_t pageoff;
vm_prot_t cap_maxprot; vm_prot_t cap_maxprot;
int align, error; int align, error, max_prot;
cap_rights_t rights; cap_rights_t rights;
if ((prot & ~(_PROT_ALL | PROT_MAX(_PROT_ALL))) != 0)
return (EINVAL);
emasteUnsubmitted
Done
Inline Actions

This has the effect of masking invalid flags (e.g. PROT_READ | PROT_WRITE | 0x100000 or 0xffff in tests/sys/vm/mmap_test.c).

emaste: This has the effect of masking invalid flags (e.g. `PROT_READ | PROT_WRITE | 0x100000` or…
max_prot = PROT_MAX_EXTRACT(prot);
prot = PROT_EXTRACT(prot);
markjUnsubmitted
Done
Inline Actions

Don't we also need to verify that prot is a subset of max_prot?

markj: Don't we also need to verify that prot is a subset of max_prot?
/*
* Always honor PROT_MAX if set. If not, default to all
* permissions unless we're implying maximum permissions.
*
* XXX: should be tunable per process and ABI.
*/
if (max_prot == 0)
max_prot = (imply_prot_max && prot != PROT_NONE) ?
prot : _PROT_ALL;
kibUnsubmitted
Done
Inline Actions

I would write this as
`
max_prot = imply_prot_max ? prot : _PROT_ALL
`

kib: I would write this as ` max_prot = imply_prot_max ? prot : _PROT_ALL `
vms = td->td_proc->p_vmspace; vms = td->td_proc->p_vmspace;
fp = NULL; fp = NULL;
AUDIT_ARG_FD(fd); AUDIT_ARG_FD(fd);
addr = addr0; addr = addr0;
/* /*
* Ignore old flags that used to be defined but did not do anything. * Ignore old flags that used to be defined but did not do anything.
*/ */
▲ Show 20 Lines • Show All 126 Lines • ▼ Show 20 Lines error = vm_mmap_object(&vms->vm_map, &addr, size, VM_PROT_NONE,
VM_PROT_NONE, flags, NULL, pos, FALSE, td); VM_PROT_NONE, flags, NULL, pos, FALSE, td);
} else if ((flags & MAP_ANON) != 0) { } else if ((flags & MAP_ANON) != 0) {
/* /*
* Mapping blank space is trivial. * Mapping blank space is trivial.
* *
* This relies on VM_PROT_* matching PROT_*. * This relies on VM_PROT_* matching PROT_*.
*/ */
error = vm_mmap_object(&vms->vm_map, &addr, size, prot, error = vm_mmap_object(&vms->vm_map, &addr, size, prot,
VM_PROT_ALL, flags, NULL, pos, FALSE, td); max_prot, flags, NULL, pos, FALSE, td);
} else { } else {
/* /*
* Mapping file, get fp for validation and don't let the * Mapping file, get fp for validation and don't let the
* descriptor disappear on us if we block. Check capability * descriptor disappear on us if we block. Check capability
* rights, but also return the maximum rights to be combined * rights, but also return the maximum rights to be combined
* with maxprot later. * with maxprot later.
*/ */
cap_rights_init(&rights, CAP_MMAP); cap_rights_init(&rights, CAP_MMAP);
if (prot & PROT_READ) if (prot & PROT_READ)
cap_rights_set(&rights, CAP_MMAP_R); cap_rights_set(&rights, CAP_MMAP_R);
if ((flags & MAP_SHARED) != 0) { if ((flags & MAP_SHARED) != 0) {
if (prot & PROT_WRITE) if (prot & PROT_WRITE)
cap_rights_set(&rights, CAP_MMAP_W); cap_rights_set(&rights, CAP_MMAP_W);
} }
if (prot & PROT_EXEC) if (prot & PROT_EXEC)
cap_rights_set(&rights, CAP_MMAP_X); cap_rights_set(&rights, CAP_MMAP_X);
error = fget_mmap(td, fd, &rights, &cap_maxprot, &fp); error = fget_mmap(td, fd, &rights, &cap_maxprot, &fp);
if (error != 0) if (error != 0)
goto done; goto done;
if ((flags & (MAP_SHARED | MAP_PRIVATE)) == 0 && if ((flags & (MAP_SHARED | MAP_PRIVATE)) == 0 &&
markjUnsubmitted
Not Done
Inline Actions

Shouldn't it be (max_prot & cap_maxprot) == max_prot? Otherwise userland can trigger this assertion by requesting a PROT_MAX(PROT_READ) | PROT_READ mapping using a descriptor with read and write rights.

markj: Shouldn't it be `(max_prot & cap_maxprot) == max_prot`? Otherwise userland can trigger this…
brooksAuthorUnsubmitted
Done
Inline Actions

My belief was that this can't happen because of the way fget_mmap is implemented, but rereading, it looks like that's only true when CAPABILITIES is defined (in which case cap_maxprot == prot so this is safe because prot must be a subset of max_prot per the check above). In the non-CAPABILITIES case cap_maxprot is VM_PROT_ALL so this will fail. Dropping the assertion may be the best fix.

brooks: My belief was that this can't happen because of the way fget_mmap is implemented, but rereading…
markjUnsubmitted
Done
Inline Actions

Sorry, I don't see why that assertion worked even with CAPABILITIES defined. fget_mmap() returns the maximum protections permitted by capabilities on the descriptor.

markj: Sorry, I don't see why that assertion worked even with CAPABILITIES defined. fget_mmap()…
brooksAuthorUnsubmitted
Done
Inline Actions

On further reading, you're correct.

brooks: On further reading, you're correct.
td->td_proc->p_osrel >= P_OSREL_MAP_FSTRICT) { td->td_proc->p_osrel >= P_OSREL_MAP_FSTRICT) {
error = EINVAL; error = EINVAL;
goto done; goto done;
} }
/* This relies on VM_PROT_* matching PROT_*. */ /* This relies on VM_PROT_* matching PROT_*. */
error = fo_mmap(fp, &vms->vm_map, &addr, size, prot, error = fo_mmap(fp, &vms->vm_map, &addr, size, prot,
cap_maxprot, flags, pos, td); max_prot & cap_maxprot, flags, pos, td);
kibUnsubmitted
Not Done
Inline Actions

Should we deny the request if (max_prot & cap_maxprot) != max_prot, so that user knows that the call was not really satisfied ?

kib: Should we deny the request if `(max_prot & cap_maxprot) != max_prot`, so that user knows that…
brooksAuthorUnsubmitted
Done
Inline Actions

I'm leaning toward not. I think we'd only want to error when we'd explicitly passed a PROT_MAX(FOO) expression so we'd have to add state for that since we don't want to fail when max_prot defaults to _PROT_ALL or when it's implied. We'd also just add yet another inexplicable EINVAL for a rare case.

brooks: I'm leaning toward not. I think we'd only want to error when we'd explicitly passed a PROT_MAX…
brooksAuthorUnsubmitted
Done
Inline Actions

I've added a KASSERT of this because I'm convinced that with the (prot & max_prot) == prot check above that this will always be safe (cap_maxprot ends up set to something derived directly from prot).

brooks: I've added a KASSERT of this because I'm convinced that with the `(prot & max_prot) == prot`…
} }
if (error == 0) if (error == 0)
td->td_retval[0] = (register_t) (addr + pageoff); td->td_retval[0] = (register_t) (addr + pageoff);
done: done:
if (fp) if (fp)
fdrop(fp, td); fdrop(fp, td);
▲ Show 20 Lines • Show All 214 Lines • ▼ Show 20 Linessys_mprotect(struct thread *td, struct mprotect_args *uap)
return (kern_mprotect(td, (uintptr_t)uap->addr, uap->len, uap->prot)); return (kern_mprotect(td, (uintptr_t)uap->addr, uap->len, uap->prot));
} }
int int
kern_mprotect(struct thread *td, uintptr_t addr0, size_t size, int prot) kern_mprotect(struct thread *td, uintptr_t addr0, size_t size, int prot)
{ {
vm_offset_t addr; vm_offset_t addr;
vm_size_t pageoff; vm_size_t pageoff;
int vm_error, max_prot;
addr = addr0; addr = addr0;
if ((prot & ~(_PROT_ALL | PROT_MAX(_PROT_ALL))) != 0)
return (EINVAL);
brooksAuthorUnsubmitted
Done
Inline Actions

This check is interesting. Linux documents this behavior (EINVAL for invalid protections), but POSIX seems to endorse our silent stripping. I'm inclined to think Linux is more correct here.

brooks: This check is interesting. Linux documents this behavior (EINVAL for invalid protections), but…
markjUnsubmitted
Done
Inline Actions

We should probably document this error in the man page too.

markj: We should probably document this error in the man page too.
max_prot = PROT_MAX_EXTRACT(prot);
prot = (prot & VM_PROT_ALL); prot = (prot & VM_PROT_ALL);
alcUnsubmitted
Done
Inline Actions

While not strictly necessary for correctness, I would advocate changing this to

prot = PROT_EXTRACT(prot);
alc: While not strictly necessary for correctness, I would advocate changing this to ``` prot =…
pageoff = (addr & PAGE_MASK); pageoff = (addr & PAGE_MASK);
addr -= pageoff; addr -= pageoff;
size += pageoff; size += pageoff;
size = (vm_size_t) round_page(size); size = (vm_size_t) round_page(size);
#ifdef COMPAT_FREEBSD32 #ifdef COMPAT_FREEBSD32
if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) { if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
if (((addr + size) & 0xffffffff) < addr) if (((addr + size) & 0xffffffff) < addr)
return (EINVAL); return (EINVAL);
} else } else
#endif #endif
if (addr + size < addr) if (addr + size < addr)
return (EINVAL); return (EINVAL);
switch (vm_map_protect(&td->td_proc->p_vmspace->vm_map, addr, if (max_prot != 0) {
addr + size, prot, FALSE)) { if ((max_prot & prot) != prot)
return (EINVAL);
vm_error = vm_map_protect(&td->td_proc->p_vmspace->vm_map,
kibUnsubmitted
Not Done
Inline Actions

I do not understand this.

Suppose you have a readonly shared mapping of the file, which was created using O_RDONLY file descriptor. Isn't this addition allows to add VM_PROT_WRITE to the corresponding vm_map_entry max_prot ? Then on the next step user can mprotect(2) the mapping to modify the file.

kib: I do not understand this. Suppose you have a readonly shared mapping of the file, which was…
brooksAuthorUnsubmitted
Done
Inline Actions

I think this does the right thing, but I had to read the implementation a couple times to convince myself. In vm_map_protect, when set_max is true, the new max_prot is checked against existing entries max_protection in the first loop over the region:

if ((new_prot & current->max_protection) != new_prot) {
  vm_map_unlock(map);
  return (KERN_PROTECTION_FAILURE);
}

This rejects attempts to upgrade max_prot.

brooks: I think this does the right thing, but I had to read the implementation a couple times to…
addr, addr + size, max_prot, TRUE);
if (vm_error != KERN_SUCCESS)
goto error_out;
markjUnsubmitted
Done
Inline Actions

You can avoid a goto by initializing vm_error to KERN_SUCCESS and skipping the final vm_map_protect() if vm_error != KERN_SUCCESS.

markj: You can avoid a goto by initializing `vm_error` to `KERN_SUCCESS` and skipping the final…
}
vm_error = vm_map_protect(&td->td_proc->p_vmspace->vm_map, addr,
addr + size, prot, FALSE);
error_out:
switch (vm_error) {
case KERN_SUCCESS: case KERN_SUCCESS:
return (0); return (0);
case KERN_PROTECTION_FAILURE: case KERN_PROTECTION_FAILURE:
return (EACCES); return (EACCES);
case KERN_RESOURCE_SHORTAGE: case KERN_RESOURCE_SHORTAGE:
return (ENOMEM); return (ENOMEM);
} }
return (EINVAL); return (EINVAL);
▲ Show 20 Lines • Show All 977 Lines • Show Last 20 Lines