[PATCH v38 00/39] LSM: Module stacking for AppArmor
by Casey Schaufler
This patchset provides the changes required for
the AppArmor security module to stack safely with any other.
There are additional changes required for SELinux and Smack
to coexist. These are primarily in the networking code and
will be addressed after these changes are upstream.
v38: Rebase to 6.0-rc7
- Discard the implementation of /proc/self/attr/context
- Discard the implementation of /proc/self/attr/display_lsm
- Implement a system call lsm_self_attr() which provides
a complete list of LSM process attributes.
- Implement a system call lsm_module_list() which provides
a list of the security modules active on the system.
- Implement a pair of prctl() options PR_LSM_ATTR_SET
and PR_LSM_ATTR_GET. These are used to control the LSM
that provides the values in /proc/.../attr entries.
- Make more of the infrastructure data changes early in the
patch set in support of the new system calls.
- Add a user interface header uapi/linux/lsm.h which contains
the format of data provided by lsm_self_attr() and the
integer LSM identifier values.
v37: Rebase to 5.19-rc3
- Audit changes should be complete, all comments have been
addressed.
- Address indexing an empty array for the case where no
built in security modules require data in struct lsmblob.
- Fix a few checkpatch complaints.
v36: Rebase to 5.19-rc1
- Yet another rework of the audit changes. Rearranging how the
timestamp is managed allows auxiliary records to be generated
correctly with a minimum of fuss.
- In the end no LSM interface scaffolding remains. Secids have
been replaced with lsmblob structures in all cases, including
IMA and NetLabel.
v35: Rebase to 5.18-rc2
- Address the case where CONFIG_SECURITY is enabled but
no security modules that use secid slots are included.
The resulting blob.secid[0] instances, although never
present in a call path, raised concerns.
- Address the case in interface_lsm where the BPF module
returns -EINVAL, resulting in a failed setting of the
value that would be otherwise allowed.
v34: Rebase to 5.18-rc1
- Incorporate feedback on the audit generation.
v33:
- Rework the supplimental audit record generation once more,
this time taking pseudo-code provided by Paul Moore as a
basis. The resulting code is considerably simpler and fits
better with the existing code flow.
v32: Rebase to 5.17-rc2
- Incorporate additional feedback from v30.
v31: Rebase to 5.16-rc4
- Incorporate feedback from Paul Moore on the audit
component changes.
v30: Rebase to 5.16-rc1
- Replace the integrity sub-system reuse of the audit
subsystem event matching functions with IMA specific
functions. This is done because audit needs to maintain
information about multiple security modules in audit
rules while IMA to restricts the information to a single
security module.
- The binder hooks have been changed and are no longer
called with sufficient information to identify the
interface_lsm. Pass that information in the binder
message, and use that in the compatibility decision.
- Refactor the audit changes.
v29: Rebase to 5.15-rc1
- Rework the supplimental audit record generation. Attach
a list of supplimental data to the audit_buffer and
generate the auxiliary records as needed on event end.
This should be usable for other auxiliary data, such as
container IDs. There is other ongoing audit work that
will require integration with this.
v28: Rebase to 5.14-rc2
- Provide IMA rules bounds checking (patch 04)
- Quote contexts in MAC_TASK_CONTEXTS and MAC_OBJ_CONTEXTS
audit records because of AppArmor's use of '=' in context
values. (patch 22,23)
v27: Fixes for landlock (patch 02)
- Rework the subject audit record generation. This version is
simpler and reflects feedback from Paul Moore. (patch 22)
v26: Rebase to 5.13-rc1
- Include the landlock security module.
- Accomodate change from security_task_getsecid() to
security_task_getsecid_obj() and security_task_getsecid_subj().
v25: Rebase to 5.12-rc2
Incorporate feedback from v24
- The IMA team suggested improvements to the integrity rule
processing.
v24: Rebase to 5.11-rc1
Incorporate feedback from v23
- Address the IMA team's concerns about "label collisions".
A label collision occurs when there is ambiguity about
which of multiple LSMs is being targeted in the definition
of an integrity check rule. A system with Smack and
AppArmor would be unable to distinguish which LSM is
important to an integrity rule referrencing the label
"unconfined" as that label is meaningful to both.
Provide a boot option to specify which LSM will be used in
IMA rules when multiple LSMs are present. (patch 04)
Pull LSM "slot" identification from later audit patches in
in support of this (patch 03).
- Pick up a few audit events that need to include supplimental
subject context records that had been missed in the
previous version.
v23: Rebase to 5.10-rc4
Incorporate feedback from v22
- Change /proc/*/attr/display to /proc/*/attr/interface_lsm to
make the purpose clearer. (patch 0012)
- Include ABI documentation. (patch 0012, 0022)
- Introduce LSM documentation updates with the patches where
the interfaces are added rather than at the end. (patch 0012, 0022)
Include more maintainers and mail lists in To: and Cc: directives.
v22: Rebase to 5.10-rc1
v21: Rebase to 5.9-rc4
Incorporate feedback from v20
- Further revert UDS SO_PEERSEC to use scaffolding around
the interfaces that use lsmblobs and store only a single
secid. The possibility of multiple security modules
requiring data here is still a future problem.
- Incorporate Richard Guy Briggs' non-syscall auxiliary
records patch (patch 0019-0021) in place of my "supplimental"
records implementation. [I'm not sure I've given proper
attestation. I will correct as appropriate]
v20: Rebase to 5.9-rc1
Change the BPF security module to use the lsmblob data. (patch 0002)
Repair length logic in subject label processing (patch 0015)
Handle -EINVAL from the empty BPF setprocattr hook (patch 0020)
Correct length processing in append_ctx() (patch 0022)
v19: Rebase to 5.8-rc6
Incorporate feedback from v18
- Revert UDS SO_PEERSEC implementation to use lsmblobs
directly, rather than allocating as needed. The correct
treatment of out-of-memory conditions in the later case
is difficult to define. (patch 0005)
- Use a size_t in append_ctx() (patch 0021)
- Fix a memory leak when creating compound contexts. (patch 0021)
Fix build error when CONFIG_SECURITY isn't set (patch 0013)
Fix build error when CONFIG_SECURITY isn't set (patch 0020)
Fix build error when CONFIG_SECURITY isn't set (patch 0021)
v18: Rebase to 5.8-rc3
Incorporate feedback from v17
- Null pointer checking in UDS (patch 0005)
Match changes in IMA code (patch 0012)
Fix the behavior of LSM context supplimental audit
records so that there's always exactly one when it's
appropriate for there to be one. This is a substantial
change that requires extention of the audit_context beyond
syscall events. (patch 0020)
v17: Rebase to 5.7-rc4
v16: Rebase to 5.6
Incorporate feedback from v15 - Thanks Stephen, Mimi and Paul
- Generally improve commit messages WRT scaffolding
- Comment ima_lsm_isset() (patch 0002)
- Some question may remain on IMA warning (patch 0002)
- Mark lsm_slot as __lsm_ro_after_init not __init_data (patch 0002)
- Change name of lsmblob variable in ima_match_rules() (patch 0003)
- Instead of putting a struct lsmblob into the unix_skb_parms
structure put a pointer to an allocated instance. There is
currently only space for 5 u32's in unix_skb_parms and it is
likely to get even tighter. Fortunately, the lifecycle
management of the allocated lsmblob is simple. (patch 0005)
- Dropped Acks due to the above change (patch 0005)
- Improved commentary on secmark labeling scaffolding. (patch 0006)
- Reduced secmark related labeling scaffolding. (patch 0006)
- Replace use of the zeroth entry of an lsmblob in scaffolding
with a function lsmblob_value() to hopefully make it less
obscure. (patch 0006)
- Convert security_secmark_relabel_packet to use lsmblob as
this reduces much of the most contentious scaffolding. (patch 0006)
- Dropped Acks due to the above change (patch 0006)
- Added BUILD_BUG_ON() for CIPSO tag 6. (patch 0018)
- Reworked audit subject information. Instead of adding fields in
the middle of existing records add a new record to the event. When
a separate record is required use subj="?". (patch 0020)
- Dropped Acks due to the above change (patch 0020)
- Reworked audit object information. Instead of adding fields in
the middle of existing records add a new record to the event. When
a separate record is required use obj="?". (patch 0021)
- Dropped Acks due to the above change (patch 0021)
- Enhanced documentation (patch 0022)
- Removed unnecessary error code check in security_getprocattr()
(patch 0021)
v15: Rebase to 5.6-rc1
- Revise IMA data use (patch 0002)
Incorporate feedback from v14
- Fix lockdown module registration naming (patch 0002)
- Revise how /proc/self/attr/context is gathered. (patch 0022)
- Revise access modes on /proc/self/attr/context. (patch 0022)
- Revise documentation on LSM external interfaces. (patch 0022)
v14: Rebase to 5.5-rc5
Incorporate feedback from v13
- Use an array of audit rules (patch 0002)
- Significant change, removed Acks (patch 0002)
- Remove unneeded include (patch 0013)
- Use context.len correctly (patch 0015)
- Reorder code to be more sensible (patch 0016)
- Drop SO_PEERCONTEXT as it's not needed yet (patch 0023)
v13: Rebase to 5.5-rc2
Incorporate feedback from v12
- Print lsmblob size with %z (Patch 0002)
- Convert lockdown LSM initialization. (Patch 0002)
- Restore error check in nft_secmark_compute_secid (Patch 0006)
- Correct blob scaffolding in ima_must_appraise() (Patch 0009)
- Make security_setprocattr() clearer (Patch 0013)
- Use lsm_task_display more widely (Patch 0013)
- Use passed size in lsmcontext_init() (Patch 0014)
- Don't add a smack_release_secctx() hook (Patch 0014)
- Don't print warning in security_release_secctx() (Patch 0014)
- Don't duplicate the label in nfs4_label_init_security() (Patch 0016)
- Remove reviewed-by as code has significant change (Patch 0016)
- Send the entire lsmblob for Tag 6 (Patch 0019)
- Fix description of socket_getpeersec_stream parameters (Patch 0023)
- Retain LSMBLOB_FIRST. What was I thinking? (Patch 0023)
- Add compound context to LSM documentation (Patch 0023)
v12: Rebase to 5.5-rc1
Fixed a couple of incorrect contractions in the text.
v11: Rebase to 5.4-rc6
Incorporate feedback from v10
- Disambiguate reading /proc/.../attr/display by restricting
all use of the interface to the current process.
- Fix a merge error in AppArmor's display attribute check
v10: Ask the security modules if the display can be changed.
v9: There is no version 9
v8: Incorporate feedback from v7
- Minor clean-up in display value management
- refactor "compound" context creation to use a common
append_ctx() function.
v7: Incorporate feedback from v6
- Make setting the display a privileged operation. The
availability of compound contexts reduces the need for
setting the display.
v6: Incorporate feedback from v5
- Add subj_<lsm>= and obj_<lsm>= fields to audit records
- Add /proc/.../attr/context to get the full context in
lsmname\0value\0... format as suggested by Simon McVittie
- Add SO_PEERCONTEXT for getsockopt() to get the full context
in the same format, also suggested by Simon McVittie.
- Add /sys/kernel/security/lsm_display_default to provide
the display default value.
v5: Incorporate feedback from v4
- Initialize the lsmcontext in security_secid_to_secctx()
- Clear the lsmcontext in all security_release_secctx() cases
- Don't use the "display" on strictly internal context
interfaces.
- The SELinux binder hooks check for cases where the context
"display" isn't compatible with SELinux.
v4: Incorporate feedback from v3
- Mark new lsm_<blob>_alloc functions static
- Replace the lsm and slot fields of the security_hook_list
with a pointer to a LSM allocated lsm_id structure. The
LSM identifies if it needs a slot explicitly. Use the
lsm_id rather than make security_add_hooks return the
slot value.
- Validate slot values used in security.c
- Reworked the "display" process attribute handling so that
it works right and doesn't use goofy list processing.
- fix display value check in dentry_init_security
- Replace audit_log of secids with '?' instead of deleting
the audit log
v3: Incorporate feedback from v2
- Make lsmblob parameter and variable names more
meaningful, changing "le" and "l" to "blob".
- Improve consistency of constant naming.
- Do more sanity checking during LSM initialization.
- Be a bit clearer about what is temporary scaffolding.
- Rather than clutter security_getpeersec_dgram with
otherwise unnecessary checks remove the apparmor
stub, which does nothing useful.
Patch 01 changes the LSM registration interface to pass a
structure lsm_id containing the name of the module instead of
just the LSM name itself.
Patch 02 introduces uapi/linux/lsm.h and numeric LSM identifiers.
The numeric ID is added to the lsm_id structure.
Patch 03 Adds the information about which process attributes are
supported by each LSM to the lsm_id structure.
Patch 04 introduces a table of LSM attribute data.
Patch 05 changes security_[gs]etprocattr() to use the LSM ID rather
than the LSM name.
Patch 06 implements the lsm_self_attr() system call.
Patch 07 separates the audit rule processing from the
integrity rule processing. They were never really the
same, but void pointers could hide that. The changes
following use the rule pointers differently in audit
and IMA, so keeping the code common is not a good idea.
Patch 08 moves management of the sock security blob
from the individual modules to the infrastructure.
Patches 09-10 introduce a structure "lsmblob" that will gradually
replace the "secid" as a shorthand for security module information.
At this point lsmblob contains an array of u32 secids, one "slot"
for each of the security modules compiled into the kernel that
used secids. A "slot" is allocated when a security module requests
one.
Patch 11 provides mechanism for the IMA subsystem to identify
explicitly which LSM is subject to IMA policy. This includes
a boot option for specifying the default and an additional option
in IMA rules "lsm=".
Patches 12-19 change LSM interfaces to use the lsmblob instead
of secids. It is important that the lsmblob be a fixed size entity
that does not have to be allocated. Several of the places
where it is used would have performance and/or locking
issues with dynamic allocation.
Patch 20 provides a mechanism for a process to identify which
security module's hooks should be used when displaying or
converting a security context string. A new prctl() options
PR_LSM_ATTR_[GS]ET get and set which security module to show.
Setting the value requires that all modules using the /proc
interfaces allow the transition. The interface LSM of other
processess can be neither read nor written.
Patch 21 Starts the process of changing how a security
context is represented. Since it is possible for a
security context to have been generated by more than one
security module it is now necessary to note which module
created a security context so that the correct "release"
hook can be called. There are several places where the
module that created a security context cannot be inferred.
This is achieved by introducing a "lsmcontext" structure
which contains the context string, its length and the
"slot" number of the security module that created it.
The security_release_secctx() interface is changed,
replacing the (string,len) pointer pair with a lsmcontext
pointer.
Patches 22-25 convert the security interfaces from
(string,len) pointer pairs to a lsmcontext pointer.
The slot number identifying the creating module is
added by the infrastructure. Where the security context
is stored for extended periods the data type is changed.
Patch 26 converts the Netlabel code to save lsmblob structures
instead of secids. This is not strictly necessary as there can
only be one security module that uses Netlabel at this point.
Using a lsmblob is much cleaner, as the interfaces that use the
data have all been converted.
Patch 27 adds checks to the binder hooks which verify
that both ends of a transaction use the same interface LSM.
Patch 28 adds a parameter to security_secid_to_secctx()
that indicates which of the security modules should be used
to provide the context.
Patches 29-33 update the audit system to better handle auxiliary
records. This requires rearranging the timestamp and serial number
handling. The skb pointer used to collect audit data is replaced
by a list of skb pointers.
Patch 34 adds a supplimental audit record for subject
LSM data when there are multiple security modules with such data.
The AUDIT_MAC_TASK_CONTEXTS record is used in conjuction with a
"subj=?" field to identify the subject data. The
AUDIT_MAC_TASK_CONTEXTS record identifies the security module
with the data: subj_selinux=xyz_t subj_apparmor=abc.
An example of the MAC_TASK_CONTEXTS (1420) record is:
type=UNKNOWN[1420]
msg=audit(1600880931.832:113)
subj_apparmor==unconfined
subj_smack=_
Patch 35 adds a supplimental audit record for object
LSM data when there are multiple security modules with such data.
The AUDIT_MAC_OBJ_CONTEXTS record is used in conjuction The
with a "obj=?" field to identify the object data.
The AUDIT_MAC_OBJ_CONTEXTS record identifies the security module
with the data: obj_selinux="xyz_t obj_apparmor="abc". While
AUDIT_MAC_TASK_CONTEXTS records will always contain an entry
for each possible security modules, AUDIT_MAC_OBJ_CONTEXTS
records will only contain entries for security modules for
which the object in question has data.
An example of the MAC_OBJ_CONTEXTS (1421) record is:
type=UNKNOWN[1421]
msg=audit(1601152467.009:1050):
obj_selinux=unconfined_u:object_r:user_home_t:s0
Patches 36-37 clean out scaffolding code.
With all interference on the AppArmor hooks removed,
Patch 38 removes the exclusive bit from AppArmor. An unnecessary
stub hook was also removed.
Patch 39 adds a system call lsm_module_list() that provides a set
of integer LSM ID values describing what is active on the system.
The Ubuntu project is using an earlier version of this patchset in
their distribution to enable stacking for containers.
Performance measurements to date have the change within the "noise".
The sockperf and dbench results are on the order of 0.2% to 0.8%
difference, with better performance being as common as worse. The
benchmarks were run with AppArmor and Smack on Ubuntu.
https://github.com/cschaufler/lsm-stacking.git#stack-6.0-rc7-38
Casey Schaufler (39):
LSM: Identify modules by more than name
LSM: Add an LSM identifier for external use
LSM: Identify the process attributes for each module
LSM: Maintain a table of LSM attribute data
proc: Use lsmids instead of lsm names for attrs
LSM: lsm_self_attr syscall for LSM self attributes
integrity: disassociate ima_filter_rule from security_audit_rule
LSM: Infrastructure management of the sock security
LSM: Add the lsmblob data structure.
LSM: provide lsm name and id slot mappings
IMA: avoid label collisions with stacked LSMs
LSM: Use lsmblob in security_audit_rule_match
LSM: Use lsmblob in security_kernel_act_as
LSM: Use lsmblob in security_secctx_to_secid
LSM: Use lsmblob in security_secid_to_secctx
LSM: Use lsmblob in security_ipc_getsecid
LSM: Use lsmblob in security_current_getsecid
LSM: Use lsmblob in security_inode_getsecid
LSM: Use lsmblob in security_cred_getsecid
LSM: Specify which LSM to display
LSM: Ensure the correct LSM context releaser
LSM: Use lsmcontext in security_secid_to_secctx
LSM: Use lsmcontext in security_inode_getsecctx
Use lsmcontext in security_dentry_init_security
LSM: security_secid_to_secctx in netlink netfilter
NET: Store LSM netlabel data in a lsmblob
binder: Pass LSM identifier for confirmation
LSM: security_secid_to_secctx module selection
Audit: Keep multiple LSM data in audit_names
Audit: Create audit_stamp structure
LSM: Add a function to report multiple LSMs
Audit: Allow multiple records in an audit_buffer
Audit: Add record for multiple task security contexts
audit: multiple subject lsm values for netlabel
Audit: Add record for multiple object contexts
netlabel: Use a struct lsmblob in audit data
LSM: Removed scaffolding function lsmcontext_init
AppArmor: Remove the exclusive flag
LSM: Create lsm_module_list system call
Documentation/ABI/testing/ima_policy | 8 +-
arch/x86/entry/syscalls/syscall_64.tbl | 2 +
drivers/android/binder.c | 47 ++-
drivers/android/binder_internal.h | 1 +
fs/ceph/super.h | 3 +-
fs/ceph/xattr.c | 15 +-
fs/fuse/dir.c | 35 +-
fs/nfs/dir.c | 2 +-
fs/nfs/inode.c | 17 +-
fs/nfs/internal.h | 8 +-
fs/nfs/nfs4proc.c | 14 +-
fs/nfs/nfs4xdr.c | 22 +-
fs/nfsd/nfs4xdr.c | 20 +-
fs/proc/base.c | 29 +-
fs/proc/internal.h | 2 +-
include/linux/audit.h | 21 +-
include/linux/cred.h | 3 +-
include/linux/lsm_hooks.h | 15 +-
include/linux/nfs4.h | 8 +-
include/linux/nfs_fs.h | 2 +-
include/linux/security.h | 292 +++++++++++--
include/linux/syscalls.h | 2 +
include/net/netlabel.h | 10 +-
include/net/scm.h | 15 +-
include/uapi/asm-generic/unistd.h | 8 +-
include/uapi/linux/audit.h | 2 +
include/uapi/linux/lsm.h | 67 +++
include/uapi/linux/prctl.h | 4 +
kernel/audit.c | 261 +++++++++---
kernel/audit.h | 18 +-
kernel/auditfilter.c | 29 +-
kernel/auditsc.c | 160 +++----
kernel/cred.c | 12 +-
kernel/sys_ni.c | 4 +
net/ipv4/cipso_ipv4.c | 26 +-
net/ipv4/ip_sockglue.c | 12 +-
net/netfilter/nf_conntrack_netlink.c | 24 +-
net/netfilter/nf_conntrack_standalone.c | 11 +-
net/netfilter/nfnetlink_queue.c | 38 +-
net/netfilter/nft_meta.c | 10 +-
net/netfilter/xt_SECMARK.c | 7 +-
net/netlabel/netlabel_kapi.c | 6 +-
net/netlabel/netlabel_unlabeled.c | 98 +++--
net/netlabel/netlabel_unlabeled.h | 2 +-
net/netlabel/netlabel_user.c | 10 +-
net/netlabel/netlabel_user.h | 2 +-
security/Makefile | 1 +
security/apparmor/include/apparmor.h | 3 +-
security/apparmor/include/net.h | 6 +-
security/apparmor/lsm.c | 110 ++---
security/apparmor/net.c | 2 +-
security/bpf/hooks.c | 14 +-
security/commoncap.c | 9 +-
security/integrity/ima/ima.h | 32 +-
security/integrity/ima/ima_api.c | 6 +-
security/integrity/ima/ima_appraise.c | 11 +-
security/integrity/ima/ima_main.c | 62 +--
security/integrity/ima/ima_policy.c | 92 +++-
security/landlock/cred.c | 2 +-
security/landlock/fs.c | 2 +-
security/landlock/ptrace.c | 2 +-
security/landlock/setup.c | 7 +
security/landlock/setup.h | 1 +
security/loadpin/loadpin.c | 10 +-
security/lockdown/lockdown.c | 9 +-
security/lsm_syscalls.c | 206 +++++++++
security/safesetid/lsm.c | 10 +-
security/security.c | 539 +++++++++++++++++++++---
security/selinux/hooks.c | 115 +++--
security/selinux/include/classmap.h | 3 +-
security/selinux/include/objsec.h | 5 +
security/selinux/include/security.h | 1 +
security/selinux/netlabel.c | 25 +-
security/selinux/ss/services.c | 4 +-
security/smack/smack.h | 6 +
security/smack/smack_access.c | 2 +-
security/smack/smack_lsm.c | 85 ++--
security/smack/smack_netfilter.c | 4 +-
security/smack/smackfs.c | 12 +-
security/tomoyo/tomoyo.c | 10 +-
security/yama/yama_lsm.c | 9 +-
81 files changed, 2063 insertions(+), 758 deletions(-)
create mode 100644 include/uapi/linux/lsm.h
create mode 100644 security/lsm_syscalls.c
--
2.37.3
1 year, 9 months
- 7
- 67
[PATCH] audit: fix undefined behavior in bit shift for AUDIT_BIT
by Gaosheng Cui
Shifting signed 32-bit value by 31 bits is undefined, so changing
significant bit to unsigned. The UBSAN warning calltrace like below:
UBSAN: shift-out-of-bounds in kernel/auditfilter.c:179:23
left shift of 1 by 31 places cannot be represented in type 'int'
Call Trace:
<TASK>
dump_stack_lvl+0x7d/0xa5
dump_stack+0x15/0x1b
ubsan_epilogue+0xe/0x4e
__ubsan_handle_shift_out_of_bounds+0x1e7/0x20c
audit_register_class+0x9d/0x137
audit_classes_init+0x4d/0xb8
do_one_initcall+0x76/0x430
kernel_init_freeable+0x3b3/0x422
kernel_init+0x24/0x1e0
ret_from_fork+0x1f/0x30
</TASK>
Fixes: 607ca46e97a1 ("UAPI: (Scripted) Disintegrate include/linux")
Signed-off-by: Gaosheng Cui <cuigaosheng1(a)huawei.com>
---
include/uapi/linux/audit.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/uapi/linux/audit.h b/include/uapi/linux/audit.h
index 7c1dc818b1d5..d676ed2b246e 100644
--- a/include/uapi/linux/audit.h
+++ b/include/uapi/linux/audit.h
@@ -187,7 +187,7 @@
#define AUDIT_MAX_KEY_LEN 256
#define AUDIT_BITMASK_SIZE 64
#define AUDIT_WORD(nr) ((__u32)((nr)/32))
-#define AUDIT_BIT(nr) (1 << ((nr) - AUDIT_WORD(nr)*32))
+#define AUDIT_BIT(nr) (1U << ((nr) - AUDIT_WORD(nr)*32))
#define AUDIT_SYSCALL_CLASSES 16
#define AUDIT_CLASS_DIR_WRITE 0
--
2.25.1
2 years, 1 month
- 4
- 4
LSM stacking in next for 6.1?
by Casey Schaufler
I would like very much to get v38 or v39 of the LSM stacking for Apparmor
patch set in the LSM next branch for 6.1. The audit changes have polished
up nicely and I believe that all comments on the integrity code have been
addressed. The interface_lsm mechanism has been beaten to a frothy peak.
There are serious binder changes, but I think they address issues beyond
the needs of stacking. Changes outside these areas are pretty well limited
to LSM interface improvements.
2 years, 1 month
- 7
- 73
[PATCH] Properly pass struct timeval.tv_sec to snprintf under -D_TIME_BITS=64
by gkabe＠vega.pgw.jp
Without this patch, auditd coredumps on 32bit when compiled with
-D_TIME_BITS=64:
Thread 1 "auditd" received signal SIGSEGV, Segmentation fault.
0xb7c7528f in strlen () from /lib/libc.so.6
(gdb) where
#0 0xb7c7528f in strlen () from /lib/libc.so.6
#1 0xb7cbe3c9 in __vfprintf_internal () from /lib/libc.so.6
#2 0xb7cce3a5 in __vsnprintf_internal () from /lib/libc.so.6
#3 0xb7d72180 in __snprintf_chk () from /lib/libc.so.6
#4 0x0040b3b5 in send_audit_event ()
#5 0x004047ff in main ()
diff -up ./src/auditd-event.c.t64 ./src/auditd-event.c
--- ./src/auditd-event.c.t64 2022-01-24 04:36:56.000000000 +0900
+++ ./src/auditd-event.c 2022-10-28 22:43:32.103854794 +0900
@@ -104,7 +104,7 @@ void write_logging_state(FILE *f)
int rc;
struct statfs buf;
- fprintf(f, "current log size = %lu KB\n", log_size/1024);
+ fprintf(f, "current log size = %llu KB\n", (unsigned long long)log_size/1024);
fprintf(f, "max log size = %lu KB\n",
config->max_log_size * (MEGABYTE/1024));
fprintf(f,"logs detected last rotate/shift = %u\n", known_logs);
@@ -112,8 +112,8 @@ void write_logging_state(FILE *f)
fs_space_left ? "yes" : "no");
rc = fstatfs(log_fd, &buf);
if (rc == 0) {
- fprintf(f, "Logging partition free space %lu MB\n",
- (buf.f_bavail * buf.f_bsize)/MEGABYTE);
+ fprintf(f, "Logging partition free space %llu MB\n",
+ (unsigned long long)(buf.f_bavail * buf.f_bsize)/MEGABYTE);
fprintf(f, "space_left setting %lu MB\n",
config->space_left);
fprintf(f, "admin_space_left setting %lu MB\n",
@@ -1632,7 +1632,7 @@ static void reconfigure(struct auditd_ev
srand(time(NULL));
seq_num = rand()%10000;
if (gettimeofday(&tv, NULL) == 0) {
- snprintf(date, sizeof(date), "audit(%lu.%03u:%u)", tv.tv_sec,
+ snprintf(date, sizeof(date), "audit(%llu.%03u:%u)", (unsigned long long)tv.tv_sec,
(unsigned)(tv.tv_usec/1000), seq_num);
} else {
snprintf(date, sizeof(date),
diff -up ./src/auditd.c.t64 ./src/auditd.c
--- ./src/auditd.c.t64 2022-01-24 04:36:56.000000000 +0900
+++ ./src/auditd.c 2022-10-28 22:38:40.817495792 +0900
@@ -312,12 +312,12 @@ int send_audit_event(int type, const cha
// Write event into netlink area like normal events
if (gettimeofday(&tv, NULL) == 0) {
e->reply.len = snprintf((char *)e->reply.msg.data,
- DMSG_SIZE, "audit(%lu.%03u:%u): %s",
- tv.tv_sec, (unsigned)(tv.tv_usec/1000), seq_num, str);
+ DMSG_SIZE, "audit(%llu.%03u:%u): %s",
+ (long long)tv.tv_sec, (unsigned)(tv.tv_usec/1000), seq_num, str);
} else {
e->reply.len = snprintf((char *)e->reply.msg.data,
- DMSG_SIZE, "audit(%lu.%03d:%u): %s",
- (unsigned long)time(NULL), 0, seq_num, str);
+ DMSG_SIZE, "audit(%llu.%03d:%u): %s",
+ (long long)time(NULL), 0, seq_num, str);
}
// Point message at the netlink buffer like normal events
e->reply.message = e->reply.msg.data;
diff -up ./src/ausearch-checkpt.c.t64 ./src/ausearch-checkpt.c
--- ./src/ausearch-checkpt.c.t64 2022-01-24 04:36:56.000000000 +0900
+++ ./src/ausearch-checkpt.c 2022-10-28 22:50:20.558201108 +0900
@@ -147,6 +147,7 @@ void save_ChkPt(const char *fn)
static int parse_checkpt_event(char *lbuf, int ndix, event *e)
{
char *rest;
+ unsigned long long t;
/*
* Find the space after the node, then make it '\0' so
@@ -173,13 +174,14 @@ static int parse_checkpt_event(char *lbu
return 1;
}
}
- if (sscanf(rest, "%lu.%03u:%lu 0x%X", &e->sec, &e->milli,
+ if (sscanf(rest, "%llu.%03u:%lu 0x%X", &t, &e->milli,
&e->serial, &e->type) != 4) {
fprintf(stderr, "Malformed output/event checkpoint line "
"after node - [%s]\n", lbuf);
checkpt_failure |= CP_STATUSBAD;
return 1;
}
+ e->sec = t;
return 0;
}
diff -up ./src/ausearch.c.t64 ./src/ausearch.c
--- ./src/ausearch.c.t64 2022-01-24 04:36:56.000000000 +0900
+++ ./src/ausearch.c 2022-10-28 22:48:03.676082763 +0900
@@ -472,10 +472,10 @@ static int process_log_fd(void)
output_event(entries);
} else if (do_output == 3) {
fprintf(stderr,
- "Corrupted checkpoint file. Inode match, but newer complete event (%lu.%03u:%lu) found before loaded checkpoint %lu.%03u:%lu\n",
- entries->e.sec, entries->e.milli,
+ "Corrupted checkpoint file. Inode match, but newer complete event (%llu.%03u:%lu) found before loaded checkpoint %llu.%03u:%lu\n",
+ (unsigned long long)entries->e.sec, entries->e.milli,
entries->e.serial,
- chkpt_input_levent.sec,
+ (unsigned long long)chkpt_input_levent.sec,
chkpt_input_levent.milli,
chkpt_input_levent.serial);
checkpt_failure |= CP_CORRUPTED;
diff -up ./tools/auvirt/auvirt.c.t64 ./tools/auvirt/auvirt.c
--- ./tools/auvirt/auvirt.c.t64 2022-01-24 04:36:56.000000000 +0900
+++ ./tools/auvirt/auvirt.c 2022-10-28 22:52:31.081335459 +0900
@@ -340,8 +340,8 @@ static int extract_virt_fields(auparse_s
error:
if (debug) {
fprintf(stderr, "Failed to get field \"%s\" for record "
- "%ld.%03u:%lu\n", field,
- auparse_get_time(au),
+ "%lld.%03u:%lu\n", field,
+ (long long)auparse_get_time(au),
auparse_get_milli(au),
auparse_get_serial(au));
}
2 years, 1 month
- 1
- 0
[PATCH 0/3] improve audit syscall-exit latency
by Ankur Arora
This series attempts to reduce syscall-exit latency in the audit path,
especially for cases where there are a lot of audit exit rules.
>From profiling, audit_filter_syscall() takes the largest chunk of time,
specifically in this list-walk while processing the AUDIT_FILTER_EXIT
list:
list = &audit_filter_list[AUDIT_FILTER_EXIT];
list_for_each_entry_rcu(e, list, list) {
if (audit_in_mask(&e->rule, ctx->major) &&
audit_filter_rules(tsk, &e->rule, ctx, NULL,
&state, false, x)) {
rcu_read_unlock();
ctx->current_state = state;
return state;
}
}
(Note that almost identical logic exists in audit_filter_uring(),
audit_filter_inode_name().)
Comparing baseline performance with audit=0/audit=1 with a user-space
getpid() loop (executes 10^7 times.) For audit=1, this uses an
audit-rule set where the audit_filter_syscall() loop iterates over
42 AUDIT_FILTER_EXIT rules which, for getpid(), calls audit_filter_rules()
for 5 of them (we use this set of rules in production.)
Test system:
Server: ORACLE SERVER X8-2L
CPU: Intel Skylakex (6:85:6)
Microcode: 0x400320a
# v6.0.0-rc5.baseline, audit=0
Performance counter stats for 'bin/getpid' (3 runs):
734.10 msec task-clock # 0.999 CPUs utilized ( +- 0.03% )
1 context-switches # 1.361 /sec ( +- 66.67% )
0 cpu-migrations # 0.000 /sec
53 page-faults # 72.152 /sec ( +- 0.63% )
2,838,869,156 cycles # 3.865 GHz ( +- 0.01% ) (62.40%)
4,174,224,305 instructions # 1.47 insn per cycle ( +- 0.01% ) (74.93%)
890,798,133 branches # 1.213 G/sec ( +- 0.01% ) (74.93%)
5,015,118 branch-misses # 0.56% of all branches ( +- 0.05% ) (74.93%)
1,231,150,558 L1-dcache-loads # 1.676 G/sec ( +- 0.01% ) (74.94%)
418,297 L1-dcache-load-misses # 0.03% of all L1-dcache accesses ( +- 0.68% ) (75.07%)
3,937 LLC-loads # 5.360 K/sec ( +- 3.76% ) (50.13%)
510 LLC-load-misses # 13.39% of all LL-cache accesses ( +- 79.89% ) (50.00%)
0.735018 +- 0.000275 seconds time elapsed ( +- 0.04% )
# v6.0.0-rc5.baseline, audit=1
Performance counter stats for 'bin/getpid' (3 runs):
2,158.81 msec task-clock # 0.998 CPUs utilized ( +- 0.13% )
2 context-switches # 0.925 /sec ( +- 28.87% )
0 cpu-migrations # 0.000 /sec
52 page-faults # 24.056 /sec ( +- 0.64% )
8,364,119,898 cycles # 3.869 GHz ( +- 0.11% ) (62.48%)
19,996,521,678 instructions # 2.39 insn per cycle ( +- 0.01% ) (74.98%)
4,302,068,252 branches # 1.990 G/sec ( +- 0.00% ) (74.98%)
15,135,694 branch-misses # 0.35% of all branches ( +- 0.16% ) (74.99%)
4,714,694,841 L1-dcache-loads # 2.181 G/sec ( +- 0.01% ) (74.99%)
66,407,857 L1-dcache-load-misses # 1.41% of all L1-dcache accesses ( +- 1.50% ) (75.01%)
6,785 LLC-loads # 3.139 K/sec ( +- 12.49% ) (50.03%)
3,235 LLC-load-misses # 41.29% of all LL-cache accesses ( +- 6.08% ) (50.01%)
2.16213 +- 0.00288 seconds time elapsed ( +- 0.13% )
perf stat numbers for each getpid() iteration:
baseline baseline
audit=0 audit=1
cycles 283 836
instructions 417 1999
IPC 1.47 2.39
branches 89 430
branch-misses 0.50 1.51
L1-loads 123 471
L1-load-misses ~0 ~6.6*
* the L1-load-misses are largely stable for runs across a single
boot, but vary across boots due to vagaries of the SLAB allocator.
baseline audit=1 spends a significant amount of time executing in audit
code and incurs a new branch-miss and a few new L1-load-misses. Also
note that computed audit-only IPC is 2.86 so the baseline is not
ill-performant code.
Patches
==
Patch 1 "audit: cache ctx->major in audit_filter_syscall()", caches
ctx->major in a local variable. This gets rid of a persistent entry
from L1-dcache (audit_context::major) that had similar alignment
constraints as a potentially busy cache-set (audit_entry::list) and
allows some of the error checking in audit_in_mask() to be hoisted out
of the loop.
Patch 2: "audit: annotate branch direction for audit_in_mask()", so
the compiler can pessimize the call to audit_filter_rules().
Patch 3, "audit: unify audit_filter_{uring(),inode_name(),syscall()}"
centralizes this logic in a common function.
with these changes:
Performance counter stats for 'bin/getpid' (3 runs):
1,750.21 msec task-clock # 0.994 CPUs utilized ( +- 0.45% )
3 context-switches # 1.705 /sec ( +- 11.11% )
0 cpu-migrations # 0.000 /sec
52 page-faults # 29.548 /sec ( +- 0.64% )
6,770,976,590 cycles # 3.848 GHz ( +- 0.40% ) (27.74%)
16,588,372,024 instructions # 2.44 insn per cycle ( +- 0.03% ) (33.34%)
4,322,555,829 branches # 2.456 G/sec ( +- 0.02% ) (33.40%)
2,803,286 branch-misses # 0.06% of all branches ( +- 26.45% ) (33.46%)
4,449,426,336 L1-dcache-loads # 2.528 G/sec ( +- 0.01% ) (27.71%)
63,612,108 L1-dcache-load-misses # 1.43% of all L1-dcache accesses ( +- 0.50% ) (27.71%)
6,123 LLC-loads # 3.479 K/sec ( +- 8.68% ) (27.71%)
1,890 LLC-load-misses # 26.69% of all LL-cache accesses ( +- 46.99% ) (27.71%)
1.76033 +- 0.00791 seconds time elapsed ( +- 0.45% )
And, overall getpid() latency numbers (aggregated over 12 boots for each):
Min Mean Median Max pstdev
(ns) (ns) (ns) (ns)
baseline 201.30 216.14 216.22 228.46 (+- 1.45%)
patch1 196.63 207.86 206.60 230.98 (+- 3.92%)
patch1-2 173.11 182.51 179.65 202.09 (+- 4.34%)
patch1-3 162.11 175.26 173.71 190.56 (+- 4.33%)
This gives a reasonable speedup. My testing was on Intel Skylake, but I
suspect this should translate to other archs as well (especially on less
wide architectures.)
Please review.
Thanks
Ankur
Ankur Arora (3):
audit: cache ctx->major in audit_filter_syscall()
audit: annotate branch direction for audit_in_mask()
audit: unify audit_filter_{uring(),inode_name(),syscall()}
kernel/auditsc.c | 86 +++++++++++++++++++++++++-----------------------
1 file changed, 45 insertions(+), 41 deletions(-)
--
2.31.1
2 years, 2 months
- 3
- 19
IORING_OP_SENDMSG_ZC should enable auditing just like IORING_OP_SENDMSG
by Paul Moore
Commit 493108d95f14 ("io_uring/net: zerocopy sendmsg") added a new
zerocopy sendmsg command to io_uring, but it enabled the
io_op_def:audit_skip flag, causing the operation to bypass auditing.
As far as I can see, the SENDMSG_ZC operation should be subject to
auditing just as SENDMSG.
--
paul-moore.com
2 years, 2 months
- 2
- 1
[GIT PULL] Audit patches for v6.1
by Paul Moore
Hi Linus,
Six audit patches for v6.1, most are pretty trivial, but a quick list
of the highlights are below:
* Only free the audit proctitle information on task exit. This allows
us to cache the information and improve performance slightly.
* Use the time_after() macro to do time comparisons instead of doing
it directly and potentially causing ourselves problems when the timer
wraps.
* Convert an audit_context state comparison from a relative enum
comparison, e.g. (x < y), to a not-equal comparison to ensure that we
are not caught out at some unknown point in the future by an enum
shuffle.
* A handful of small cleanups such as tidying up comments and removing
unused declarations.
-Paul
--
The following changes since commit 568035b01cfb107af8d2e4bd2fb9aea22cf5b868:
Linux 6.0-rc1 (2022-08-14 15:50:18 -0700)
are available in the Git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/pcmoore/audit.git
tags/audit-pr-20221003
for you to fetch changes up to 934f70d9d498a7fffd611edf48356698f1d72e91:
audit: remove selinux_audit_rule_update() declaration
(2022-09-07 11:30:15 -0400)
----------------------------------------------------------------
audit/stable-6.1 PR 20221003
----------------------------------------------------------------
Jilin Yuan (1):
audit: fix repeated words in comments
Richard Guy Briggs (3):
audit: audit_context pid unused, context enum comment fix
audit: explicitly check audit_context->context enum value
audit: free audit_proctitle only on task exit
Xiu Jianfeng (1):
audit: remove selinux_audit_rule_update() declaration
wuchi (1):
audit: use time_after to compare time
kernel/audit.c | 8 +++-----
kernel/audit.h | 4 +---
kernel/auditsc.c | 10 +++++-----
3 files changed, 9 insertions(+), 13 deletions(-)
--
paul-moore.com
2 years, 2 months
- 2
- 1
- ← Newer
- 1
- Older →