commit 3b23a32a63 upstream.
dev_ifsioc_locked() is called with only RCU read lock, so when
there is a parallel writer changing the mac address, it could
get a partially updated mac address, as shown below:
Thread 1 Thread 2
// eth_commit_mac_addr_change()
memcpy(dev->dev_addr, addr->sa_data, ETH_ALEN);
// dev_ifsioc_locked()
memcpy(ifr->ifr_hwaddr.sa_data,
dev->dev_addr,...);
Close this race condition by guarding them with a RW semaphore,
like netdev_get_name(). We can not use seqlock here as it does not
allow blocking. The writers already take RTNL anyway, so this does
not affect the slow path. To avoid bothering existing
dev_set_mac_address() callers in drivers, introduce a new wrapper
just for user-facing callers on ioctl and rtnetlink paths.
Note, bonding also changes slave mac addresses but that requires
a separate patch due to the complexity of bonding code.
Fixes: 3710becf8a ("net: RCU locking for simple ioctl()")
Reported-by: "Gong, Sishuai" <sishuai@purdue.edu>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Cong Wang <cong.wang@bytedance.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a93dcaada2 upstream.
Currently, the psample netlink skb is allocated with a size that does
not account for the nested 'PSAMPLE_ATTR_TUNNEL' attribute and the
padding required for the 64-bit attribute 'PSAMPLE_TUNNEL_KEY_ATTR_ID'.
This can result in failure to add attributes to the netlink skb due
to insufficient tail room. The following error message is printed to
the kernel log: "Could not create psample log message".
Fix this by adjusting the allocation size to take into account the
nested attribute and the padding.
Fixes: d8bed686ab ("net: psample: Add tunnel support")
CC: Yotam Gigi <yotam.gi@gmail.com>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Chris Mi <cmi@nvidia.com>
Link: https://lore.kernel.org/r/20210225075145.184314-1-cmi@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 86dd9868b8 upstream.
Support also transmitting frames using the custom "8899 A"
4 byte tag.
Qingfang came up with the solution: we need to pad the
ethernet frame to 60 bytes using eth_skb_pad(), then the
switch will happily accept frames with custom tags.
Cc: Mauri Sandberg <sandberg@mailfence.com>
Reported-by: DENG Qingfang <dqfext@gmail.com>
Fixes: efd7fe68f0 ("net: dsa: tag_rtl4_a: Implement Realtek 4 byte A tag")
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 52557dbc75 upstream.
MPJ subflows are not exposed as fds to user spaces. As such,
incoming MPJ subflows are removed from the accept queue by
tcp_check_req()/tcp_get_cookie_sock().
Later tcp_child_process() invokes subflow_data_ready() on the
parent socket regardless of the subflow kind, leading to poll
wakeups even if the later accept will block.
Address the issue by double-checking the queue state before
waking the user-space.
Closes: https://github.com/multipath-tcp/mptcp_net-next/issues/164
Reported-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Fixes: f296234c98 ("mptcp: Add handling of incoming MP_JOIN requests")
Reviewed-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a1ba9da8f0 upstream.
The current code would unnecessarily expand the address range. Consider
one example, (start, end) = (1G-2M, 3G+2M), and (vm_start, vm_end) =
(1G-4M, 3G+4M), the expected adjustment should be keep (1G-2M, 3G+2M)
without expand. But the current result will be (1G-4M, 3G+4M). Actually,
the range (1G-4M, 1G) and (3G, 3G+4M) would never been involved in pmd
sharing.
After this patch, we will check that the vma span at least one PUD aligned
size and the start,end range overlap the aligned range of vma.
With above example, the aligned vma range is (1G, 3G), so if (start, end)
range is within (1G-4M, 1G), or within (3G, 3G+4M), then no adjustment to
both start and end. Otherwise, we will have chance to adjust start
downwards or end upwards without exceeding (vm_start, vm_end).
Mike:
: The 'adjusted range' is used for calls to mmu notifiers and cache(tlb)
: flushing. Since the current code unnecessarily expands the range in some
: cases, more entries than necessary would be flushed. This would/could
: result in performance degradation. However, this is highly dependent on
: the user runtime. Is there a combination of vma layout and calls to
: actually hit this issue? If the issue is hit, will those entries
: unnecessarily flushed be used again and need to be unnecessarily reloaded?
Link: https://lkml.kernel.org/r/20210104081631.2921415-1-lixinhai.lxh@gmail.com
Fixes: 75802ca663 ("mm/hugetlb: fix calculation of adjust_range_if_pmd_sharing_possible")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Suggested-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 0f02de4481 upstream.
At early boot stage, we have a whole PGDIR to map the kernel, so there
is no need to restrict the early mapping size to 128MB. Removing this
define also allows us to simplify some compile time logic.
This fixes large kernel mappings with a size greater than 128MB, as it
is the case for syzbot kernels whose size was just ~130MB.
Note that on rv64, for now, we are then limited to PGDIR size for early
mapping as we can't use PGD mappings (see [1]). That should be enough
given the relative small size of syzbot kernels compared to PGDIR_SIZE
which is 1GB.
[1] https://lore.kernel.org/lkml/20200603153608.30056-1-alex@ghiti.fr/
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
Tested-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Palmer Dabbelt <palmerdabbelt@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 7ef4c19d24 upstream.
syzbot found WARNINGs in several smackfs write operations where
bytes count is passed to memdup_user_nul which exceeds
GFP MAX_ORDER. Check count size if bigger than PAGE_SIZE.
Per smackfs doc, smk_write_net4addr accepts any label or -CIPSO,
smk_write_net6addr accepts any label or -DELETE. I couldn't find
any general rule for other label lengths except SMK_LABELLEN,
SMK_LONGLABEL, SMK_CIPSOMAX which are documented.
Let's constrain, in general, smackfs label lengths for PAGE_SIZE.
Although fuzzer crashes write to smackfs/netlabel on 0x400000 length.
Here is a quick way to reproduce the WARNING:
python -c "print('A' * 0x400000)" > /sys/fs/smackfs/netlabel
Reported-by: syzbot+a71a442385a0b2815497@syzkaller.appspotmail.com
Signed-off-by: Sabyrzhan Tasbolatov <snovitoll@gmail.com>
Signed-off-by: Casey Schaufler <casey@schaufler-ca.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 88a9e03bee upstream.
An assert failure is triggered by syzkaller test due to
ATTR_KILL_PRIV is not cleared before xfs_setattr_size.
As ATTR_KILL_PRIV is not checked/used by xfs_setattr_size,
just remove it from the assert.
Signed-off-by: Yumei Huang <yuhuang@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 3bef198f1b upstream.
syzbot is feeding invalid superblock data to JFS for mount testing.
JFS does not check several of the fields -- just assumes that they
are good since the JFS_MAGIC and version fields are good.
In this case (syzbot reproducer), we have s_l2bsize == 0xda0c,
pad == 0xf045, and s_state == 0x50, all of which are invalid IMO.
Having s_l2bsize == 0xda0c causes this UBSAN warning:
UBSAN: shift-out-of-bounds in fs/jfs/jfs_mount.c:373:25
shift exponent -9716 is negative
s_l2bsize can be tested for correctness. pad can be tested for non-0
and punted. s_state can be tested for its valid values and punted.
Do those 3 tests and if any of them fails, report the superblock as
invalid/corrupt and let fsck handle it.
With this patch, chkSuper() says this when JFS_DEBUG is enabled:
jfs_mount: Mount Failure: superblock is corrupt!
Mount JFS Failure: -22
jfs_mount failed w/return code = -22
The obvious problem with this method is that next week there could
be another syzbot test that uses different fields for invalid values,
this making this like a game of whack-a-mole.
syzkaller link: https://syzkaller.appspot.com/bug?extid=36315852ece4132ec193
Reported-by: syzbot+36315852ece4132ec193@syzkaller.appspotmail.com
Reported-by: kernel test robot <lkp@intel.com> # v2
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Dave Kleikamp <dave.kleikamp@oracle.com>
Cc: jfs-discussion@lists.sourceforge.net
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit bb73d07148 upstream.
This is similar to commit
b21ebf2fb4 ("x86: Treat R_X86_64_PLT32 as R_X86_64_PC32")
but for i386. As far as the kernel is concerned, R_386_PLT32 can be
treated the same as R_386_PC32.
R_386_PLT32/R_X86_64_PLT32 are PC-relative relocation types which
can only be used by branches. If the referenced symbol is defined
externally, a PLT will be used.
R_386_PC32/R_X86_64_PC32 are PC-relative relocation types which can be
used by address taking operations and branches. If the referenced symbol
is defined externally, a copy relocation/canonical PLT entry will be
created in the executable.
On x86-64, there is no PIC vs non-PIC PLT distinction and an
R_X86_64_PLT32 relocation is produced for both `call/jmp foo` and
`call/jmp foo@PLT` with newer (2018) GNU as/LLVM integrated assembler.
This avoids canonical PLT entries (st_shndx=0, st_value!=0).
On i386, there are 2 types of PLTs, PIC and non-PIC. Currently,
the GCC/GNU as convention is to use R_386_PC32 for non-PIC PLT and
R_386_PLT32 for PIC PLT. Copy relocations/canonical PLT entries
are possible ABI issues but GCC/GNU as will likely keep the status
quo because (1) the ABI is legacy (2) the change will drop a GNU
ld diagnostic for non-default visibility ifunc in shared objects.
clang-12 -fno-pic (since [1]) can emit R_386_PLT32 for compiler
generated function declarations, because preventing canonical PLT
entries is weighed over the rare ifunc diagnostic.
Further info for the more interested:
https://github.com/ClangBuiltLinux/linux/issues/1210https://sourceware.org/bugzilla/show_bug.cgi?id=27169a084c0388e [1]
[ bp: Massage commit message. ]
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Fangrui Song <maskray@google.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Reviewed-by: Nathan Chancellor <natechancellor@gmail.com>
Tested-by: Nick Desaulniers <ndesaulniers@google.com>
Tested-by: Nathan Chancellor <natechancellor@gmail.com>
Tested-by: Sedat Dilek <sedat.dilek@gmail.com>
Link: https://lkml.kernel.org/r/20210127205600.1227437-1-maskray@google.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit d349f99768 upstream.
tcf_action_init_1() loads tc action modules automatically with
request_module() after parsing the tc action names, and it drops RTNL
lock and re-holds it before and after request_module(). This causes a
lot of troubles, as discovered by syzbot, because we can be in the
middle of batch initializations when we create an array of tc actions.
One of the problem is deadlock:
CPU 0 CPU 1
rtnl_lock();
for (...) {
tcf_action_init_1();
-> rtnl_unlock();
-> request_module();
rtnl_lock();
for (...) {
tcf_action_init_1();
-> tcf_idr_check_alloc();
// Insert one action into idr,
// but it is not committed until
// tcf_idr_insert_many(), then drop
// the RTNL lock in the _next_
// iteration
-> rtnl_unlock();
-> rtnl_lock();
-> a_o->init();
-> tcf_idr_check_alloc();
// Now waiting for the same index
// to be committed
-> request_module();
-> rtnl_lock()
// Now waiting for RTNL lock
}
rtnl_unlock();
}
rtnl_unlock();
This is not easy to solve, we can move the request_module() before
this loop and pre-load all the modules we need for this netlink
message and then do the rest initializations. So the loop breaks down
to two now:
for (i = 1; i <= TCA_ACT_MAX_PRIO && tb[i]; i++) {
struct tc_action_ops *a_o;
a_o = tc_action_load_ops(name, tb[i]...);
ops[i - 1] = a_o;
}
for (i = 1; i <= TCA_ACT_MAX_PRIO && tb[i]; i++) {
act = tcf_action_init_1(ops[i - 1]...);
}
Although this looks serious, it only has been reported by syzbot, so it
seems hard to trigger this by humans. And given the size of this patch,
I'd suggest to make it to net-next and not to backport to stable.
This patch has been tested by syzbot and tested with tdc.py by me.
Fixes: 0fedc63fad ("net_sched: commit action insertions together")
Reported-and-tested-by: syzbot+82752bc5331601cf4899@syzkaller.appspotmail.com
Reported-and-tested-by: syzbot+b3b63b6bff456bd95294@syzkaller.appspotmail.com
Reported-by: syzbot+ba67b12b1ca729912834@syzkaller.appspotmail.com
Cc: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Cong Wang <cong.wang@bytedance.com>
Tested-by: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://lore.kernel.org/r/20210117005657.14810-1-xiyou.wangcong@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 8b5553ace8 upstream.
Having two ring buffers per-peer means that every peer results in two
massive ring allocations. On an 8-core x86_64 machine, this commit
reduces the per-peer allocation from 18,688 bytes to 1,856 bytes, which
is an 90% reduction. Ninety percent! With some single-machine
deployments approaching 500,000 peers, we're talking about a reduction
from 7 gigs of memory down to 700 megs of memory.
In order to get rid of these per-peer allocations, this commit switches
to using a list-based queueing approach. Currently GSO fragments are
chained together using the skb->next pointer (the skb_list_* singly
linked list approach), so we form the per-peer queue around the unused
skb->prev pointer (which sort of makes sense because the links are
pointing backwards). Use of skb_queue_* is not possible here, because
that is based on doubly linked lists and spinlocks. Multiple cores can
write into the queue at any given time, because its writes occur in the
start_xmit path or in the udp_recv path. But reads happen in a single
workqueue item per-peer, amounting to a multi-producer, single-consumer
paradigm.
The MPSC queue is implemented locklessly and never blocks. However, it
is not linearizable (though it is serializable), with a very tight and
unlikely race on writes, which, when hit (some tiny fraction of the
0.15% of partial adds on a fully loaded 16-core x86_64 system), causes
the queue reader to terminate early. However, because every packet sent
queues up the same workqueue item after it is fully added, the worker
resumes again, and stopping early isn't actually a problem, since at
that point the packet wouldn't have yet been added to the encryption
queue. These properties allow us to avoid disabling interrupts or
spinning. The design is based on Dmitry Vyukov's algorithm [1].
Performance-wise, ordinarily list-based queues aren't preferable to
ringbuffers, because of cache misses when following pointers around.
However, we *already* have to follow the adjacent pointers when working
through fragments, so there shouldn't actually be any change there. A
potential downside is that dequeueing is a bit more complicated, but the
ptr_ring structure used prior had a spinlock when dequeueing, so all and
all the difference appears to be a wash.
Actually, from profiling, the biggest performance hit, by far, of this
commit winds up being atomic_add_unless(count, 1, max) and atomic_
dec(count), which account for the majority of CPU time, according to
perf. In that sense, the previous ring buffer was superior in that it
could check if it was full by head==tail, which the list-based approach
cannot do.
But all and all, this enables us to get massive memory savings, allowing
WireGuard to scale for real world deployments, without taking much of a
performance hit.
[1] http://www.1024cores.net/home/lock-free-algorithms/queues/intrusive-mpsc-node-based-queue
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit d5a49aa6c3 upstream.
In order to test ndo_start_xmit being called in parallel, explicitly add
separate tests, which should all run on different cores. This should
help tease out bugs associated with queueing up packets from different
cores in parallel. Currently, it hasn't found those types of bugs, but
given future planned work, this is a useful regression to avoid.
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit ee576c47db upstream.
The icmp{,v6}_send functions make all sorts of use of skb->cb, casting
it with IPCB or IP6CB, assuming the skb to have come directly from the
inet layer. But when the packet comes from the ndo layer, especially
when forwarded, there's no telling what might be in skb->cb at that
point. As a result, the icmp sending code risks reading bogus memory
contents, which can result in nasty stack overflows such as this one
reported by a user:
panic+0x108/0x2ea
__stack_chk_fail+0x14/0x20
__icmp_send+0x5bd/0x5c0
icmp_ndo_send+0x148/0x160
In icmp_send, skb->cb is cast with IPCB and an ip_options struct is read
from it. The optlen parameter there is of particular note, as it can
induce writes beyond bounds. There are quite a few ways that can happen
in __ip_options_echo. For example:
// sptr/skb are attacker-controlled skb bytes
sptr = skb_network_header(skb);
// dptr/dopt points to stack memory allocated by __icmp_send
dptr = dopt->__data;
// sopt is the corrupt skb->cb in question
if (sopt->rr) {
optlen = sptr[sopt->rr+1]; // corrupt skb->cb + skb->data
soffset = sptr[sopt->rr+2]; // corrupt skb->cb + skb->data
// this now writes potentially attacker-controlled data, over
// flowing the stack:
memcpy(dptr, sptr+sopt->rr, optlen);
}
In the icmpv6_send case, the story is similar, but not as dire, as only
IP6CB(skb)->iif and IP6CB(skb)->dsthao are used. The dsthao case is
worse than the iif case, but it is passed to ipv6_find_tlv, which does
a bit of bounds checking on the value.
This is easy to simulate by doing a `memset(skb->cb, 0x41,
sizeof(skb->cb));` before calling icmp{,v6}_ndo_send, and it's only by
good fortune and the rarity of icmp sending from that context that we've
avoided reports like this until now. For example, in KASAN:
BUG: KASAN: stack-out-of-bounds in __ip_options_echo+0xa0e/0x12b0
Write of size 38 at addr ffff888006f1f80e by task ping/89
CPU: 2 PID: 89 Comm: ping Not tainted 5.10.0-rc7-debug+ #5
Call Trace:
dump_stack+0x9a/0xcc
print_address_description.constprop.0+0x1a/0x160
__kasan_report.cold+0x20/0x38
kasan_report+0x32/0x40
check_memory_region+0x145/0x1a0
memcpy+0x39/0x60
__ip_options_echo+0xa0e/0x12b0
__icmp_send+0x744/0x1700
Actually, out of the 4 drivers that do this, only gtp zeroed the cb for
the v4 case, while the rest did not. So this commit actually removes the
gtp-specific zeroing, while putting the code where it belongs in the
shared infrastructure of icmp{,v6}_ndo_send.
This commit fixes the issue by passing an empty IPCB or IP6CB along to
the functions that actually do the work. For the icmp_send, this was
already trivial, thanks to __icmp_send providing the plumbing function.
For icmpv6_send, this required a tiny bit of refactoring to make it
behave like the v4 case, after which it was straight forward.
Fixes: a2b78e9b2c ("sunvnet: generate ICMP PTMUD messages for smaller port MTUs")
Reported-by: SinYu <liuxyon@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/netdev/CAF=yD-LOF116aHub6RMe8vB8ZpnrrnoTdqhobEx+bvoA8AsP0w@mail.gmail.com/T/
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Link: https://lore.kernel.org/r/20210223131858.72082-1-Jason@zx2c4.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 1faba27f11 upstream.
The W=1 compilation of allmodconfig generates the following warning:
net/ipv6/icmp.c:448:6: warning: no previous prototype for 'icmp6_send' [-Wmissing-prototypes]
448 | void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
| ^~~~~~~~~~
Fix it by providing function declaration for builds with ipv6 as a module.
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit cca2c6aebe upstream.
Metadata resize shouldn't happen in the ctr. The ctr loads a temporary
(inactive) table that will only become active upon resume. That is why
resize should always be done in terms of resume. Otherwise a load (ctr)
whose inactive table never becomes active will incorrectly resize the
metadata.
Also, perform the resize directly in preresume, instead of using the
worker to do it.
The worker might run other metadata operations, e.g., it could start
digestion, before resizing the metadata. These operations will end up
using the old size.
This could lead to errors, like:
device-mapper: era: metadata_digest_transcribe_writeset: dm_array_set_value failed
device-mapper: era: process_old_eras: digest step failed, stopping digestion
The reason of the above error is that the worker started the digestion
of the archived writeset using the old, larger size.
As a result, metadata_digest_transcribe_writeset tried to write beyond
the end of the era array.
Fixes: eec40579d8 ("dm: add era target")
Cc: stable@vger.kernel.org # v3.15+
Signed-off-by: Nikos Tsironis <ntsironis@arrikto.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 2524933307 upstream.
In case of devices with at most 64 blocks, the digestion of consecutive
eras uses the writeset of the first era as the writeset of all eras to
digest, leading to lost writes. That is, we lose the information about
what blocks were written during the affected eras.
The digestion code uses a dm_disk_bitset object to access the archived
writesets. This structure includes a one word (64-bit) cache to reduce
the number of array lookups.
This structure is initialized only once, in metadata_digest_start(),
when we kick off digestion.
But, when we insert a new writeset into the writeset tree, before the
digestion of the previous writeset is done, or equivalently when there
are multiple writesets in the writeset tree to digest, then all these
writesets are digested using the same cache and the cache is not
re-initialized when moving from one writeset to the next.
For devices with more than 64 blocks, i.e., the size of the cache, the
cache is indirectly invalidated when we move to a next set of blocks, so
we avoid the bug.
But for devices with at most 64 blocks we end up using the same cached
data for digesting all archived writesets, i.e., the cache is loaded
when digesting the first writeset and it never gets reloaded, until the
digestion is done.
As a result, the writeset of the first era to digest is used as the
writeset of all the following archived eras, leading to lost writes.
Fix this by reinitializing the dm_disk_bitset structure, and thus
invalidating the cache, every time the digestion code starts digesting a
new writeset.
Fixes: eec40579d8 ("dm: add era target")
Cc: stable@vger.kernel.org # v3.15+
Signed-off-by: Nikos Tsironis <ntsironis@arrikto.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit c8e846ff93 upstream.
dm-era doesn't support changing the data block size of existing devices,
so check explicitly that the requested block size for a new target
matches the one stored in the metadata.
Fixes: eec40579d8 ("dm: add era target")
Cc: stable@vger.kernel.org # v3.15+
Signed-off-by: Nikos Tsironis <ntsironis@arrikto.com>
Reviewed-by: Ming-Hung Tsai <mtsai@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 2099b145d7 upstream.
In case of a system crash, dm-era might fail to mark blocks as written
in its metadata, although the corresponding writes to these blocks were
passed down to the origin device and completed successfully.
Consider the following sequence of events:
1. We write to a block that has not been yet written in the current era
2. era_map() checks the in-core bitmap for the current era and sees
that the block is not marked as written.
3. The write is deferred for submission after the metadata have been
updated and committed.
4. The worker thread processes the deferred write
(process_deferred_bios()) and marks the block as written in the
in-core bitmap, **before** committing the metadata.
5. The worker thread starts committing the metadata.
6. We do more writes that map to the same block as the write of step (1)
7. era_map() checks the in-core bitmap and sees that the block is marked
as written, **although the metadata have not been committed yet**.
8. These writes are passed down to the origin device immediately and the
device reports them as completed.
9. The system crashes, e.g., power failure, before the commit from step
(5) finishes.
When the system recovers and we query the dm-era target for the list of
written blocks it doesn't report the aforementioned block as written,
although the writes of step (6) completed successfully.
The issue is that era_map() decides whether to defer or not a write
based on non committed information. The root cause of the bug is that we
update the in-core bitmap, **before** committing the metadata.
Fix this by updating the in-core bitmap **after** successfully
committing the metadata.
Fixes: eec40579d8 ("dm: add era target")
Cc: stable@vger.kernel.org # v3.15+
Signed-off-by: Nikos Tsironis <ntsironis@arrikto.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit de89afc1e4 upstream.
Following a system crash, dm-era fails to recover the committed writeset
for the current era, leading to lost writes. That is, we lose the
information about what blocks were written during the affected era.
dm-era assumes that the writeset of the current era is archived when the
device is suspended. So, when resuming the device, it just moves on to
the next era, ignoring the committed writeset.
This assumption holds when the device is properly shut down. But, when
the system crashes, the code that suspends the target never runs, so the
writeset for the current era is not archived.
There are three issues that cause the committed writeset to get lost:
1. dm-era doesn't load the committed writeset when opening the metadata
2. The code that resizes the metadata wipes the information about the
committed writeset (assuming it was loaded at step 1)
3. era_preresume() starts a new era, without taking into account that
the current era might not have been archived, due to a system crash.
To fix this:
1. Load the committed writeset when opening the metadata
2. Fix the code that resizes the metadata to make sure it doesn't wipe
the loaded writeset
3. Fix era_preresume() to check for a loaded writeset and archive it,
before starting a new era.
Fixes: eec40579d8 ("dm: add era target")
Cc: stable@vger.kernel.org # v3.15+
Signed-off-by: Nikos Tsironis <ntsironis@arrikto.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
