]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/commit
KEYS: Don't write out to userspace while holding key semaphore
authorWaiman Long <longman@redhat.com>
Sun, 22 Mar 2020 01:11:24 +0000 (21:11 -0400)
committerDavid Howells <dhowells@redhat.com>
Sun, 29 Mar 2020 11:40:41 +0000 (12:40 +0100)
commitd3ec10aa95819bff18a0d936b18884c7816d0914
treeb313dd59ae2e25792066a38e5f083eb447928f33
parent1b649e0bcae71c118c1333e02249a7510ba7f70a
KEYS: Don't write out to userspace while holding key semaphore

A lockdep circular locking dependency report was seen when running a
keyutils test:

[12537.027242] ======================================================
[12537.059309] WARNING: possible circular locking dependency detected
[12537.088148] 4.18.0-147.7.1.el8_1.x86_64+debug #1 Tainted: G OE    --------- -  -
[12537.125253] ------------------------------------------------------
[12537.153189] keyctl/25598 is trying to acquire lock:
[12537.175087] 000000007c39f96c (&mm->mmap_sem){++++}, at: __might_fault+0xc4/0x1b0
[12537.208365]
[12537.208365] but task is already holding lock:
[12537.234507] 000000003de5b58d (&type->lock_class){++++}, at: keyctl_read_key+0x15a/0x220
[12537.270476]
[12537.270476] which lock already depends on the new lock.
[12537.270476]
[12537.307209]
[12537.307209] the existing dependency chain (in reverse order) is:
[12537.340754]
[12537.340754] -> #3 (&type->lock_class){++++}:
[12537.367434]        down_write+0x4d/0x110
[12537.385202]        __key_link_begin+0x87/0x280
[12537.405232]        request_key_and_link+0x483/0xf70
[12537.427221]        request_key+0x3c/0x80
[12537.444839]        dns_query+0x1db/0x5a5 [dns_resolver]
[12537.468445]        dns_resolve_server_name_to_ip+0x1e1/0x4d0 [cifs]
[12537.496731]        cifs_reconnect+0xe04/0x2500 [cifs]
[12537.519418]        cifs_readv_from_socket+0x461/0x690 [cifs]
[12537.546263]        cifs_read_from_socket+0xa0/0xe0 [cifs]
[12537.573551]        cifs_demultiplex_thread+0x311/0x2db0 [cifs]
[12537.601045]        kthread+0x30c/0x3d0
[12537.617906]        ret_from_fork+0x3a/0x50
[12537.636225]
[12537.636225] -> #2 (root_key_user.cons_lock){+.+.}:
[12537.664525]        __mutex_lock+0x105/0x11f0
[12537.683734]        request_key_and_link+0x35a/0xf70
[12537.705640]        request_key+0x3c/0x80
[12537.723304]        dns_query+0x1db/0x5a5 [dns_resolver]
[12537.746773]        dns_resolve_server_name_to_ip+0x1e1/0x4d0 [cifs]
[12537.775607]        cifs_reconnect+0xe04/0x2500 [cifs]
[12537.798322]        cifs_readv_from_socket+0x461/0x690 [cifs]
[12537.823369]        cifs_read_from_socket+0xa0/0xe0 [cifs]
[12537.847262]        cifs_demultiplex_thread+0x311/0x2db0 [cifs]
[12537.873477]        kthread+0x30c/0x3d0
[12537.890281]        ret_from_fork+0x3a/0x50
[12537.908649]
[12537.908649] -> #1 (&tcp_ses->srv_mutex){+.+.}:
[12537.935225]        __mutex_lock+0x105/0x11f0
[12537.954450]        cifs_call_async+0x102/0x7f0 [cifs]
[12537.977250]        smb2_async_readv+0x6c3/0xc90 [cifs]
[12538.000659]        cifs_readpages+0x120a/0x1e50 [cifs]
[12538.023920]        read_pages+0xf5/0x560
[12538.041583]        __do_page_cache_readahead+0x41d/0x4b0
[12538.067047]        ondemand_readahead+0x44c/0xc10
[12538.092069]        filemap_fault+0xec1/0x1830
[12538.111637]        __do_fault+0x82/0x260
[12538.129216]        do_fault+0x419/0xfb0
[12538.146390]        __handle_mm_fault+0x862/0xdf0
[12538.167408]        handle_mm_fault+0x154/0x550
[12538.187401]        __do_page_fault+0x42f/0xa60
[12538.207395]        do_page_fault+0x38/0x5e0
[12538.225777]        page_fault+0x1e/0x30
[12538.243010]
[12538.243010] -> #0 (&mm->mmap_sem){++++}:
[12538.267875]        lock_acquire+0x14c/0x420
[12538.286848]        __might_fault+0x119/0x1b0
[12538.306006]        keyring_read_iterator+0x7e/0x170
[12538.327936]        assoc_array_subtree_iterate+0x97/0x280
[12538.352154]        keyring_read+0xe9/0x110
[12538.370558]        keyctl_read_key+0x1b9/0x220
[12538.391470]        do_syscall_64+0xa5/0x4b0
[12538.410511]        entry_SYSCALL_64_after_hwframe+0x6a/0xdf
[12538.435535]
[12538.435535] other info that might help us debug this:
[12538.435535]
[12538.472829] Chain exists of:
[12538.472829]   &mm->mmap_sem --> root_key_user.cons_lock --> &type->lock_class
[12538.472829]
[12538.524820]  Possible unsafe locking scenario:
[12538.524820]
[12538.551431]        CPU0                    CPU1
[12538.572654]        ----                    ----
[12538.595865]   lock(&type->lock_class);
[12538.613737]                                lock(root_key_user.cons_lock);
[12538.644234]                                lock(&type->lock_class);
[12538.672410]   lock(&mm->mmap_sem);
[12538.687758]
[12538.687758]  *** DEADLOCK ***
[12538.687758]
[12538.714455] 1 lock held by keyctl/25598:
[12538.732097]  #0: 000000003de5b58d (&type->lock_class){++++}, at: keyctl_read_key+0x15a/0x220
[12538.770573]
[12538.770573] stack backtrace:
[12538.790136] CPU: 2 PID: 25598 Comm: keyctl Kdump: loaded Tainted: G
[12538.844855] Hardware name: HP ProLiant DL360 Gen9/ProLiant DL360 Gen9, BIOS P89 12/27/2015
[12538.881963] Call Trace:
[12538.892897]  dump_stack+0x9a/0xf0
[12538.907908]  print_circular_bug.isra.25.cold.50+0x1bc/0x279
[12538.932891]  ? save_trace+0xd6/0x250
[12538.948979]  check_prev_add.constprop.32+0xc36/0x14f0
[12538.971643]  ? keyring_compare_object+0x104/0x190
[12538.992738]  ? check_usage+0x550/0x550
[12539.009845]  ? sched_clock+0x5/0x10
[12539.025484]  ? sched_clock_cpu+0x18/0x1e0
[12539.043555]  __lock_acquire+0x1f12/0x38d0
[12539.061551]  ? trace_hardirqs_on+0x10/0x10
[12539.080554]  lock_acquire+0x14c/0x420
[12539.100330]  ? __might_fault+0xc4/0x1b0
[12539.119079]  __might_fault+0x119/0x1b0
[12539.135869]  ? __might_fault+0xc4/0x1b0
[12539.153234]  keyring_read_iterator+0x7e/0x170
[12539.172787]  ? keyring_read+0x110/0x110
[12539.190059]  assoc_array_subtree_iterate+0x97/0x280
[12539.211526]  keyring_read+0xe9/0x110
[12539.227561]  ? keyring_gc_check_iterator+0xc0/0xc0
[12539.249076]  keyctl_read_key+0x1b9/0x220
[12539.266660]  do_syscall_64+0xa5/0x4b0
[12539.283091]  entry_SYSCALL_64_after_hwframe+0x6a/0xdf

One way to prevent this deadlock scenario from happening is to not
allow writing to userspace while holding the key semaphore. Instead,
an internal buffer is allocated for getting the keys out from the
read method first before copying them out to userspace without holding
the lock.

That requires taking out the __user modifier from all the relevant
read methods as well as additional changes to not use any userspace
write helpers. That is,

  1) The put_user() call is replaced by a direct copy.
  2) The copy_to_user() call is replaced by memcpy().
  3) All the fault handling code is removed.

Compiling on a x86-64 system, the size of the rxrpc_read() function is
reduced from 3795 bytes to 2384 bytes with this patch.

Fixes: ^1da177e4c3f4 ("Linux-2.6.12-rc2")
Reviewed-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: David Howells <dhowells@redhat.com>
12 files changed:
include/keys/big_key-type.h
include/keys/user-type.h
include/linux/key-type.h
net/dns_resolver/dns_key.c
net/rxrpc/key.c
security/keys/big_key.c
security/keys/encrypted-keys/encrypted.c
security/keys/keyctl.c
security/keys/keyring.c
security/keys/request_key_auth.c
security/keys/trusted-keys/trusted_tpm1.c
security/keys/user_defined.c