linux.git/kernel/sys.c, branch v6.1-rc5 Clone of https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git Merge tag 'random-6.1-rc1-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/crng/random 2022-10-10T17:41:21+00:00 Linus Torvalds torvalds@linux-foundation.org 2022-10-10T17:41:21+00:00 8adc0486f3c85e3c1e40c1ce6884317a17c380d3 Pull random number generator updates from Jason Donenfeld: - Huawei reported that when they updated their kernel from 4.4 to something much newer, some userspace code they had broke, the culprit being the accidental removal of O_NONBLOCK from /dev/random way back in 5.6. It's been gone for over 2 years now and this is the first we've heard of it, but userspace breakage is userspace breakage, so O_NONBLOCK is now back. - Use randomness from hardware RNGs much more often during early boot, at the same interval that crng reseeds are done, from Dominik. - A semantic change in hardware RNG throttling, so that the hwrng framework can properly feed random.c with randomness from hardware RNGs that aren't specifically marked as creditable. A related patch coming to you via Herbert's hwrng tree depends on this one, not to compile, but just to function properly, so you may want to merge this PULL before that one. - A fix to clamp credited bits from the interrupts pool to the size of the pool sample. This is mainly just a theoretical fix, as it'd be pretty hard to exceed it in practice. - Oracle reported that InfiniBand TCP latency regressed by around 10-15% after a change a few cycles ago made at the request of the RT folks, in which we hoisted a somewhat rare operation (1 in 1024 times) out of the hard IRQ handler and into a workqueue, a pretty common and boring pattern. It turns out, though, that scheduling a worker from there has overhead of its own, whereas scheduling a timer on that same CPU for the next jiffy amortizes better and doesn't incur the same overhead. I also eliminated a cache miss by moving the work_struct (and subsequently, the timer_list) to below a critical cache line, so that the more critical members that are accessed on every hard IRQ aren't split between two cache lines. - The boot-time initialization of the RNG has been split into two approximate phases: what we can accomplish before timekeeping is possible and what we can accomplish after. This winds up being useful so that we can use RDRAND to seed the RNG before CONFIG_SLAB_FREELIST_RANDOM=y systems initialize slabs, in addition to other early uses of randomness. The effect is that systems with RDRAND (or a bootloader seed) will never see any warnings at all when setting CONFIG_WARN_ALL_UNSEEDED_RANDOM=y. And kfence benefits from getting a better seed of its own. - Small systems without much entropy sometimes wind up putting some truncated serial number read from flash into hostname, so contribute utsname changes to the RNG, without crediting. - Add smaller batches to serve requests for smaller integers, and make use of them when people ask for random numbers bounded by a given compile-time constant. This has positive effects all over the tree, most notably in networking and kfence. - The original jitter algorithm intended (I believe) to schedule the timer for the next jiffy, not the next-next jiffy, yet it used mod_timer(jiffies + 1), which will fire on the next-next jiffy, instead of what I believe was intended, mod_timer(jiffies), which will fire on the next jiffy. So fix that. - Fix a comment typo, from William. * tag 'random-6.1-rc1-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/crng/random: random: clear new batches when bringing new CPUs online random: fix typos in get_random_bytes() comment random: schedule jitter credit for next jiffy, not in two jiffies prandom: make use of smaller types in prandom_u32_max random: add 8-bit and 16-bit batches utsname: contribute changes to RNG random: use init_utsname() instead of utsname() kfence: use better stack hash seed random: split initialization into early step and later step random: use expired timer rather than wq for mixing fast pool random: avoid reading two cache lines on irq randomness random: clamp credited irq bits to maximum mixed random: throttle hwrng writes if no entropy is credited random: use hwgenerator randomness more frequently at early boot random: restore O_NONBLOCK support 
Pull random number generator updates from Jason Donenfeld:

 - Huawei reported that when they updated their kernel from 4.4 to
   something much newer, some userspace code they had broke, the culprit
   being the accidental removal of O_NONBLOCK from /dev/random way back
   in 5.6. It's been gone for over 2 years now and this is the first
   we've heard of it, but userspace breakage is userspace breakage, so
   O_NONBLOCK is now back.

 - Use randomness from hardware RNGs much more often during early boot,
   at the same interval that crng reseeds are done, from Dominik.

 - A semantic change in hardware RNG throttling, so that the hwrng
   framework can properly feed random.c with randomness from hardware
   RNGs that aren't specifically marked as creditable.

   A related patch coming to you via Herbert's hwrng tree depends on
   this one, not to compile, but just to function properly, so you may
   want to merge this PULL before that one.

 - A fix to clamp credited bits from the interrupts pool to the size of
   the pool sample. This is mainly just a theoretical fix, as it'd be
   pretty hard to exceed it in practice.

 - Oracle reported that InfiniBand TCP latency regressed by around
   10-15% after a change a few cycles ago made at the request of the RT
   folks, in which we hoisted a somewhat rare operation (1 in 1024
   times) out of the hard IRQ handler and into a workqueue, a pretty
   common and boring pattern.

   It turns out, though, that scheduling a worker from there has
   overhead of its own, whereas scheduling a timer on that same CPU for
   the next jiffy amortizes better and doesn't incur the same overhead.

   I also eliminated a cache miss by moving the work_struct (and
   subsequently, the timer_list) to below a critical cache line, so that
   the more critical members that are accessed on every hard IRQ aren't
   split between two cache lines.

 - The boot-time initialization of the RNG has been split into two
   approximate phases: what we can accomplish before timekeeping is
   possible and what we can accomplish after.

   This winds up being useful so that we can use RDRAND to seed the RNG
   before CONFIG_SLAB_FREELIST_RANDOM=y systems initialize slabs, in
   addition to other early uses of randomness. The effect is that
   systems with RDRAND (or a bootloader seed) will never see any
   warnings at all when setting CONFIG_WARN_ALL_UNSEEDED_RANDOM=y. And
   kfence benefits from getting a better seed of its own.

 - Small systems without much entropy sometimes wind up putting some
   truncated serial number read from flash into hostname, so contribute
   utsname changes to the RNG, without crediting.

 - Add smaller batches to serve requests for smaller integers, and make
   use of them when people ask for random numbers bounded by a given
   compile-time constant. This has positive effects all over the tree,
   most notably in networking and kfence.

 - The original jitter algorithm intended (I believe) to schedule the
   timer for the next jiffy, not the next-next jiffy, yet it used
   mod_timer(jiffies + 1), which will fire on the next-next jiffy,
   instead of what I believe was intended, mod_timer(jiffies), which
   will fire on the next jiffy. So fix that.

 - Fix a comment typo, from William.

* tag 'random-6.1-rc1-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/crng/random:
  random: clear new batches when bringing new CPUs online
  random: fix typos in get_random_bytes() comment
  random: schedule jitter credit for next jiffy, not in two jiffies
  prandom: make use of smaller types in prandom_u32_max
  random: add 8-bit and 16-bit batches
  utsname: contribute changes to RNG
  random: use init_utsname() instead of utsname()
  kfence: use better stack hash seed
  random: split initialization into early step and later step
  random: use expired timer rather than wq for mixing fast pool
  random: avoid reading two cache lines on irq randomness
  random: clamp credited irq bits to maximum mixed
  random: throttle hwrng writes if no entropy is credited
  random: use hwgenerator randomness more frequently at early boot
  random: restore O_NONBLOCK support
Merge tag 'ucount-rlimits-cleanups-for-v5.19' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace 2022-10-09T23:24:05+00:00 Linus Torvalds torvalds@linux-foundation.org 2022-10-09T23:24:05+00:00 493ffd6605b2d3d4dc7008ab927dba319f36671f Pull ucounts update from Eric Biederman: "Split rlimit and ucount values and max values After the ucount rlimit code was merged a bunch of small but siginificant bugs were found and fixed. At the time it was realized that part of the problem was that while the ucount rlimits were very similar to the oridinary ucounts (in being nested counts with limits) the semantics were slightly different and the code would be less error prone if there was less sharing. This is the long awaited cleanup that should hopefully keep things more comprehensible and less error prone for whoever needs to touch that code next" * tag 'ucount-rlimits-cleanups-for-v5.19' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace: ucounts: Split rlimit and ucount values and max values 
Pull ucounts update from Eric Biederman:
 "Split rlimit and ucount values and max values

  After the ucount rlimit code was merged a bunch of small but
  siginificant bugs were found and fixed. At the time it was realized
  that part of the problem was that while the ucount rlimits were very
  similar to the oridinary ucounts (in being nested counts with limits)
  the semantics were slightly different and the code would be less error
  prone if there was less sharing.

  This is the long awaited cleanup that should hopefully keep things
  more comprehensible and less error prone for whoever needs to touch
  that code next"

* tag 'ucount-rlimits-cleanups-for-v5.19' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace:
  ucounts: Split rlimit and ucount values and max values
utsname: contribute changes to RNG 2022-09-29T19:37:27+00:00 Jason A. Donenfeld Jason@zx2c4.com 2022-09-27T09:35:16+00:00 37608ba315a2b1b548aa5b1064e5559e029cb016 On some small machines with little entropy, a quasi-unique hostname is sometimes a relevant factor. I've seen, for example, 8 character alpha-numeric serial numbers. In addition, the time at which the hostname is set is usually a decent measurement of how long early boot took. So, call add_device_randomness() on new hostnames, which feeds its arguments to the RNG in addition to a fresh cycle counter. Low cost hooks like this never hurt and can only ever help, and since this costs basically nothing for an operation that is never a fast path, this is an overall easy win. Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Dominik Brodowski <linux@dominikbrodowski.net> Reviewed-by: Kees Cook <keescook@chromium.org> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> 
On some small machines with little entropy, a quasi-unique hostname is
sometimes a relevant factor. I've seen, for example, 8 character
alpha-numeric serial numbers. In addition, the time at which the hostname
is set is usually a decent measurement of how long early boot took. So,
call add_device_randomness() on new hostnames, which feeds its arguments
to the RNG in addition to a fresh cycle counter.

Low cost hooks like this never hurt and can only ever help, and since
this costs basically nothing for an operation that is never a fast path,
this is an overall easy win.

Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
ucounts: Split rlimit and ucount values and max values 2022-05-18T23:24:57+00:00 Alexey Gladkov legion@kernel.org 2022-05-18T17:17:30+00:00 de399236e240743ad2dd10d719c37b97ddf31996 Since the semantics of maximum rlimit values are different, it would be better not to mix ucount and rlimit values. This will prevent the error of using inc_count/dec_ucount for rlimit parameters. This patch also renames the functions to emphasize the lack of connection between rlimit and ucount. v3: - Fix BUG:KASAN:use-after-free_in_dec_ucount. v2: - Fix the array-index-out-of-bounds that was found by the lkp project. Reported-by: kernel test robot <oliver.sang@intel.com> Signed-off-by: Alexey Gladkov <legion@kernel.org> Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Link: https://lkml.kernel.org/r/20220518171730.l65lmnnjtnxnftpq@example.org Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> 
Since the semantics of maximum rlimit values are different, it would be
better not to mix ucount and rlimit values. This will prevent the error
of using inc_count/dec_ucount for rlimit parameters.

This patch also renames the functions to emphasize the lack of
connection between rlimit and ucount.

v3:
- Fix BUG:KASAN:use-after-free_in_dec_ucount.

v2:
- Fix the array-index-out-of-bounds that was found by the lkp project.

Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Alexey Gladkov <legion@kernel.org>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Link: https://lkml.kernel.org/r/20220518171730.l65lmnnjtnxnftpq@example.org
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
arm64/sme: Implement vector length configuration prctl()s 2022-04-22T17:50:54+00:00 Mark Brown broonie@kernel.org 2022-04-19T11:22:19+00:00 9e4ab6c89109472082616f8d2f6ada7deaffe161 As for SVE provide a prctl() interface which allows processes to configure their SME vector length. Signed-off-by: Mark Brown <broonie@kernel.org> Reviewed-by: Catalin Marinas <catalin.marinas@arm.com> Link: https://lore.kernel.org/r/20220419112247.711548-12-broonie@kernel.org Signed-off-by: Catalin Marinas <catalin.marinas@arm.com> 
As for SVE provide a prctl() interface which allows processes to
configure their SME vector length.

Signed-off-by: Mark Brown <broonie@kernel.org>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Link: https://lore.kernel.org/r/20220419112247.711548-12-broonie@kernel.org
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
prlimit: do not grab the tasklist_lock 2022-03-08T20:33:36+00:00 Barret Rhoden brho@google.com 2022-01-06T17:20:41+00:00 18c91bb2d87268d23868bf13508f5bc9cf04e89a Unnecessarily grabbing the tasklist_lock can be a scalability bottleneck for workloads that also must grab the tasklist_lock for waiting, killing, and cloning. The tasklist_lock was grabbed to protect tsk->sighand from disappearing (becoming NULL). tsk->signal was already protected by holding a reference to tsk. update_rlimit_cpu() assumed tsk->sighand != NULL. With this commit, it attempts to lock_task_sighand(). However, this means that update_rlimit_cpu() can fail. This only happens when a task is exiting. Note that during exec, sighand may *change*, but it will not be NULL. Prior to this commit, the do_prlimit() ensured that update_rlimit_cpu() would not fail by read locking the tasklist_lock and checking tsk->sighand != NULL. If update_rlimit_cpu() fails, there may be other tasks that are not exiting that share tsk->signal. However, the group_leader is the last task to be released, so if we cannot update_rlimit_cpu(group_leader), then the entire process is exiting. The only other caller of update_rlimit_cpu() is selinux_bprm_committing_creds(). It has tsk == current, so update_rlimit_cpu() cannot fail (current->sighand cannot disappear until current exits). This change resulted in a 14% speedup on a microbenchmark where parents kill and wait on their children, and children getpriority, setpriority, and getrlimit. Signed-off-by: Barret Rhoden <brho@google.com> Link: https://lkml.kernel.org/r/20220106172041.522167-4-brho@google.com Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> 
Unnecessarily grabbing the tasklist_lock can be a scalability bottleneck
for workloads that also must grab the tasklist_lock for waiting,
killing, and cloning.

The tasklist_lock was grabbed to protect tsk->sighand from disappearing
(becoming NULL).  tsk->signal was already protected by holding a
reference to tsk.

update_rlimit_cpu() assumed tsk->sighand != NULL.  With this commit, it
attempts to lock_task_sighand().  However, this means that
update_rlimit_cpu() can fail.  This only happens when a task is exiting.
Note that during exec, sighand may *change*, but it will not be NULL.

Prior to this commit, the do_prlimit() ensured that update_rlimit_cpu()
would not fail by read locking the tasklist_lock and checking tsk->sighand
!= NULL.

If update_rlimit_cpu() fails, there may be other tasks that are not
exiting that share tsk->signal.  However, the group_leader is the last
task to be released, so if we cannot update_rlimit_cpu(group_leader),
then the entire process is exiting.

The only other caller of update_rlimit_cpu() is
selinux_bprm_committing_creds().  It has tsk == current, so
update_rlimit_cpu() cannot fail (current->sighand cannot disappear
until current exits).

This change resulted in a 14% speedup on a microbenchmark where parents
kill and wait on their children, and children getpriority, setpriority,
and getrlimit.

Signed-off-by: Barret Rhoden <brho@google.com>
Link: https://lkml.kernel.org/r/20220106172041.522167-4-brho@google.com
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
prlimit: make do_prlimit() static 2022-03-08T20:33:36+00:00 Barret Rhoden brho@google.com 2022-01-06T17:20:40+00:00 c57bef0287dd5deeabaea5727950559fb9037cd9 There are no other callers in the kernel. Fixed up a comment format and whitespace issue when moving do_prlimit() higher in sys.c. Signed-off-by: Barret Rhoden <brho@google.com> Link: https://lkml.kernel.org/r/20220106172041.522167-3-brho@google.com Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> 
There are no other callers in the kernel.

Fixed up a comment format and whitespace issue when moving do_prlimit()
higher in sys.c.

Signed-off-by: Barret Rhoden <brho@google.com>
Link: https://lkml.kernel.org/r/20220106172041.522167-3-brho@google.com
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
mm: refactor vm_area_struct::anon_vma_name usage code 2022-03-05T19:08:32+00:00 Suren Baghdasaryan surenb@google.com 2022-03-05T04:28:51+00:00 5c26f6ac9416b63d093e29c30e79b3297e425472 Avoid mixing strings and their anon_vma_name referenced pointers by using struct anon_vma_name whenever possible. This simplifies the code and allows easier sharing of anon_vma_name structures when they represent the same name. [surenb@google.com: fix comment] Link: https://lkml.kernel.org/r/20220223153613.835563-1-surenb@google.com Link: https://lkml.kernel.org/r/20220224231834.1481408-1-surenb@google.com Signed-off-by: Suren Baghdasaryan <surenb@google.com> Suggested-by: Matthew Wilcox <willy@infradead.org> Suggested-by: Michal Hocko <mhocko@suse.com> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Colin Cross <ccross@google.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Kees Cook <keescook@chromium.org> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Christian Brauner <brauner@kernel.org> Cc: Alexey Gladkov <legion@kernel.org> Cc: Sasha Levin <sashal@kernel.org> Cc: Chris Hyser <chris.hyser@oracle.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Peter Collingbourne <pcc@google.com> Cc: Xiaofeng Cao <caoxiaofeng@yulong.com> Cc: David Hildenbrand <david@redhat.com> Cc: Cyrill Gorcunov <gorcunov@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> 
Avoid mixing strings and their anon_vma_name referenced pointers by
using struct anon_vma_name whenever possible.  This simplifies the code
and allows easier sharing of anon_vma_name structures when they
represent the same name.

[surenb@google.com: fix comment]

Link: https://lkml.kernel.org/r/20220223153613.835563-1-surenb@google.com
Link: https://lkml.kernel.org/r/20220224231834.1481408-1-surenb@google.com
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Colin Cross <ccross@google.com>
Cc: Sumit Semwal <sumit.semwal@linaro.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Alexey Gladkov <legion@kernel.org>
Cc: Sasha Levin <sashal@kernel.org>
Cc: Chris Hyser <chris.hyser@oracle.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Xiaofeng Cao <caoxiaofeng@yulong.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ucounts: Move RLIMIT_NPROC handling after set_user 2022-02-17T15:11:26+00:00 Eric W. Biederman ebiederm@xmission.com 2022-02-14T15:40:25+00:00 c923a8e7edb010da67424077cbf1a6f1396ebd2e During set*id() which cred->ucounts to charge the the current process to is not known until after set_cred_ucounts. So move the RLIMIT_NPROC checking into a new helper flag_nproc_exceeded and call flag_nproc_exceeded after set_cred_ucounts. This is very much an arbitrary subset of the places where we currently change the RLIMIT_NPROC accounting, designed to preserve the existing logic. Fixing the existing logic will be the subject of another series of changes. Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20220216155832.680775-4-ebiederm@xmission.com Fixes: 21d1c5e386bc ("Reimplement RLIMIT_NPROC on top of ucounts") Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> 
During set*id() which cred->ucounts to charge the the current process
to is not known until after set_cred_ucounts.  So move the
RLIMIT_NPROC checking into a new helper flag_nproc_exceeded and call
flag_nproc_exceeded after set_cred_ucounts.

This is very much an arbitrary subset of the places where we currently
change the RLIMIT_NPROC accounting, designed to preserve the existing
logic.

Fixing the existing logic will be the subject of another series of
changes.

Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/20220216155832.680775-4-ebiederm@xmission.com
Fixes: 21d1c5e386bc ("Reimplement RLIMIT_NPROC on top of ucounts")
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
rlimit: Fix RLIMIT_NPROC enforcement failure caused by capability calls in set_user 2022-02-17T15:08:05+00:00 Eric W. Biederman ebiederm@xmission.com 2022-02-11T19:57:44+00:00 c16bdeb5a39ffa3f32b32f812831a2092d2a3061 Solar Designer <solar@openwall.com> wrote: > I'm not aware of anyone actually running into this issue and reporting > it. The systems that I personally know use suexec along with rlimits > still run older/distro kernels, so would not yet be affected. > > So my mention was based on my understanding of how suexec works, and > code review. Specifically, Apache httpd has the setting RLimitNPROC, > which makes it set RLIMIT_NPROC: > > https://httpd.apache.org/docs/2.4/mod/core.html#rlimitnproc > > The above documentation for it includes: > > "This applies to processes forked from Apache httpd children servicing > requests, not the Apache httpd children themselves. This includes CGI > scripts and SSI exec commands, but not any processes forked from the > Apache httpd parent, such as piped logs." > > In code, there are: > > ./modules/generators/mod_cgid.c: ( (cgid_req.limits.limit_nproc_set) && ((rc = apr_procattr_limit_set(procattr, APR_LIMIT_NPROC, > ./modules/generators/mod_cgi.c: ((rc = apr_procattr_limit_set(procattr, APR_LIMIT_NPROC, > ./modules/filters/mod_ext_filter.c: rv = apr_procattr_limit_set(procattr, APR_LIMIT_NPROC, conf->limit_nproc); > > For example, in mod_cgi.c this is in run_cgi_child(). > > I think this means an httpd child sets RLIMIT_NPROC shortly before it > execs suexec, which is a SUID root program. suexec then switches to the > target user and execs the CGI script. > > Before 2863643fb8b9, the setuid() in suexec would set the flag, and the > target user's process count would be checked against RLIMIT_NPROC on > execve(). After 2863643fb8b9, the setuid() in suexec wouldn't set the > flag because setuid() is (naturally) called when the process is still > running as root (thus, has those limits bypass capabilities), and > accordingly execve() would not check the target user's process count > against RLIMIT_NPROC. In commit 2863643fb8b9 ("set_user: add capability check when rlimit(RLIMIT_NPROC) exceeds") capable calls were added to set_user to make it more consistent with fork. Unfortunately because of call site differences those capable calls were checking the credentials of the user before set*id() instead of after set*id(). This breaks enforcement of RLIMIT_NPROC for applications that set the rlimit and then call set*id() while holding a full set of capabilities. The capabilities are only changed in the new credential in security_task_fix_setuid(). The code in apache suexec appears to follow this pattern. Commit 909cc4ae86f3 ("[PATCH] Fix two bugs with process limits (RLIMIT_NPROC)") where this check was added describes the targes of this capability check as: 2/ When a root-owned process (e.g. cgiwrap) sets up process limits and then calls setuid, the setuid should fail if the user would then be running more than rlim_cur[RLIMIT_NPROC] processes, but it doesn't. This patch adds an appropriate test. With this patch, and per-user process limit imposed in cgiwrap really works. So the original use case of this check also appears to match the broken pattern. Restore the enforcement of RLIMIT_NPROC by removing the bad capable checks added in set_user. This unfortunately restores the inconsistent state the code has been in for the last 11 years, but dealing with the inconsistencies looks like a larger problem. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/all/20210907213042.GA22626@openwall.com/ Link: https://lkml.kernel.org/r/20220212221412.GA29214@openwall.com Link: https://lkml.kernel.org/r/20220216155832.680775-1-ebiederm@xmission.com Fixes: 2863643fb8b9 ("set_user: add capability check when rlimit(RLIMIT_NPROC) exceeds") History-Tree: https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git Reviewed-by: Solar Designer <solar@openwall.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> 
Solar Designer <solar@openwall.com> wrote:
> I'm not aware of anyone actually running into this issue and reporting
> it.  The systems that I personally know use suexec along with rlimits
> still run older/distro kernels, so would not yet be affected.
>
> So my mention was based on my understanding of how suexec works, and
> code review.  Specifically, Apache httpd has the setting RLimitNPROC,
> which makes it set RLIMIT_NPROC:
>
> https://httpd.apache.org/docs/2.4/mod/core.html#rlimitnproc
>
> The above documentation for it includes:
>
> "This applies to processes forked from Apache httpd children servicing
> requests, not the Apache httpd children themselves. This includes CGI
> scripts and SSI exec commands, but not any processes forked from the
> Apache httpd parent, such as piped logs."
>
> In code, there are:
>
> ./modules/generators/mod_cgid.c:        ( (cgid_req.limits.limit_nproc_set) && ((rc = apr_procattr_limit_set(procattr, APR_LIMIT_NPROC,
> ./modules/generators/mod_cgi.c:        ((rc = apr_procattr_limit_set(procattr, APR_LIMIT_NPROC,
> ./modules/filters/mod_ext_filter.c:    rv = apr_procattr_limit_set(procattr, APR_LIMIT_NPROC, conf->limit_nproc);
>
> For example, in mod_cgi.c this is in run_cgi_child().
>
> I think this means an httpd child sets RLIMIT_NPROC shortly before it
> execs suexec, which is a SUID root program.  suexec then switches to the
> target user and execs the CGI script.
>
> Before 2863643fb8b9, the setuid() in suexec would set the flag, and the
> target user's process count would be checked against RLIMIT_NPROC on
> execve().  After 2863643fb8b9, the setuid() in suexec wouldn't set the
> flag because setuid() is (naturally) called when the process is still
> running as root (thus, has those limits bypass capabilities), and
> accordingly execve() would not check the target user's process count
> against RLIMIT_NPROC.

In commit 2863643fb8b9 ("set_user: add capability check when
rlimit(RLIMIT_NPROC) exceeds") capable calls were added to set_user to
make it more consistent with fork.  Unfortunately because of call site
differences those capable calls were checking the credentials of the
user before set*id() instead of after set*id().

This breaks enforcement of RLIMIT_NPROC for applications that set the
rlimit and then call set*id() while holding a full set of
capabilities.  The capabilities are only changed in the new credential
in security_task_fix_setuid().

The code in apache suexec appears to follow this pattern.

Commit 909cc4ae86f3 ("[PATCH] Fix two bugs with process limits
(RLIMIT_NPROC)") where this check was added describes the targes of this
capability check as:

  2/ When a root-owned process (e.g. cgiwrap) sets up process limits and then
      calls setuid, the setuid should fail if the user would then be running
      more than rlim_cur[RLIMIT_NPROC] processes, but it doesn't.  This patch
      adds an appropriate test.  With this patch, and per-user process limit
      imposed in cgiwrap really works.

So the original use case of this check also appears to match the broken
pattern.

Restore the enforcement of RLIMIT_NPROC by removing the bad capable
checks added in set_user.  This unfortunately restores the
inconsistent state the code has been in for the last 11 years, but
dealing with the inconsistencies looks like a larger problem.

Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/all/20210907213042.GA22626@openwall.com/
Link: https://lkml.kernel.org/r/20220212221412.GA29214@openwall.com
Link: https://lkml.kernel.org/r/20220216155832.680775-1-ebiederm@xmission.com
Fixes: 2863643fb8b9 ("set_user: add capability check when rlimit(RLIMIT_NPROC) exceeds")
History-Tree: https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git
Reviewed-by: Solar Designer <solar@openwall.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>