For some reason, the ip_len field in a reassembled IP datagram is set
to just the size of the payload, but it should be set to the value it
would have had if the datagram had never been fragmented in the first
place, i.e. size of payload plus size of IP header.
That latter value is currently returned correctly via the "len"
variable. And before entering net_defragment(), len does have the
value ntohs(ip->ip_len), so if we're not dealing with a
fragment (so net_defragment leaves *len alone), that relationship of
course also holds after the net_defragment() call.
The only use I can find of ip->ip_len after the net_defragment call is
the ntohs(ip->udp_len) > ntohs(ip->ip_len) sanity check - none of the
functions that are passed the "ip" pointer themselves inspect ->ip_len
but instead use the passed len.
But that sanity check is a bit odd, since the RHS really should be
"ntohs(ip->ip_len) - 20", i.e. the IP payload size.
Now that we've fixed things so that len == ntohs(ip->ip_len) in all
cases, change that sanity check to use len-20 as the RHS.
Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
I hit a strange problem with v2022.10: Sometimes my tftp transfer
would seemingly just hang. It only happened for some files. Moreover,
changing tftpblocksize from 65464 to 65460 or 65000 made it work again
for all the files I tried. So I started suspecting it had something to
do with the file sizes and in particular the way the tftp blocks get
fragmented and reassembled.
v2022.01 showed no problems with any of the files or any value of
tftpblocksize.
Looking at what had changed in net.c or tftp.c since January showed
only one remotely interesting thing, b85d130ea0.
So I fired up wireshark on my host to see if somehow one of the
packets would be too small. But no, with both v2022.01 and v2022.10,
the exact same sequence of packets were sent, all but the last of size
1500, and the last being 1280 bytes.
But then it struck me that 1280 is 5*256, so one of the two bytes
on-the-wire is 0 and the other is 5, and when then looking at the code
again the lack of endianness conversion becomes obvious. [ntohs is
both applied to ip->ip_off just above, as well as to ip->ip_len just a
little further down when the "len" is actually computed].
IOWs the current code would falsely reject any packet which happens to
be a multiple of 256 bytes in size, breaking tftp transfers somewhat
randomly, and if it did get one of those "malicious" packets with
ip_len set to, say, 27, it would be seen by this check as being 6912
and hence not rejected.
====
Now, just adding the missing ntohs() would make my initial problem go
away, in that I can now download the file where the last fragment ends
up being 1280 bytes. But there's another bug in the code and/or
analysis: The right-hand side is too strict, in that it is ok for the
last fragment not to have a multiple of 8 bytes as payload - it really
must be ok, because nothing in the IP spec says that IP datagrams must
have a multiple of 8 bytes as payload. And comments in the code also
mention this.
To fix that, replace the comparison with <= IP_HDR_SIZE and add
another check that len is actually a multiple of 8 when the "more
fragments" bit is set - which it necessarily is for the case where
offset8 ends up being 0, since we're only called when
(ip_off & (IP_OFFS | IP_FLAGS_MFRAG)).
====
So, does this fix CVE-2022-30790 for real? It certainly correctly
rejects the POC code which relies on sending a packet of size 27 with
the MFRAG flag set. Can the attack be carried out with a size 27
packet that doesn't set MFRAG (hence must set a non-zero fragment
offset)? I dunno. If we get a packet without MFRAG, we update
h->last_byte in the hole we've found to be start+len, hence we'd enter
one of
if ((h >= thisfrag) && (h->last_byte <= start + len)) {
or
} else if (h->last_byte <= start + len) {
and thus won't reach any of the
/* overlaps with initial part of the hole: move this hole */
newh = thisfrag + (len / 8);
/* fragment sits in the middle: split the hole */
newh = thisfrag + (len / 8);
IOW these division are now guaranteed to be exact, and thus I think
the scenario in CVE-2022-30790 cannot happen anymore.
====
However, there's a big elephant in the room, which has always been
spelled out in the comments, and which makes me believe that one can
still cause mayhem even with packets whose payloads are all 8-byte
aligned:
This code doesn't deal with a fragment that overlaps with two
different holes (thus being a superset of a previously-received
fragment).
Suppose each character below represents 8 bytes, with D being already
received data, H being a hole descriptor (struct hole), h being
non-populated chunks, and P representing where the payload of a just
received packet should go:
DDDHhhhhDDDDHhhhDDDD
PPPPPPPPP
I'm pretty sure in this case we'd end up with h being the first hole,
enter the simple
} else if (h->last_byte <= start + len) {
/* overlaps with final part of the hole: shorten this hole */
h->last_byte = start;
case, and thus in the memcpy happily overwrite the second H with our
chosen payload. This is probably worth fixing...
Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
While the code mostly/only handles UDP packets, it's possible for the
last fragment of a fragmented UDP packet to be smaller than 28 bytes;
it can be as small as 21 bytes (an IP header plus one byte of
payload). So until we've performed the defragmentation step and thus
know whether we're now holding a full packet, we should only check for
the existence of the fields in the ip header, i.e. that there are at
least 20 bytes present.
In practice, we always seem to be handed a "len" of minimum 60 from the
device layer, i.e. minimal ethernet frame length minus FCS, so this is
mostly theoretical.
After we've fetched the header's claimed length and used that to
update the len variable, check that the header itself claims to be the
minimal possible length.
This is probably how CVE-2022-30552 should have been dealt with in the
first place, because net_defragment() is not the only place that wants
to know the size of the IP datagram payload: If we receive a
non-fragmented ICMP packet, we pass "len" to receive_icmp() which in
turn may pass it to ping_receive() which does
compute_ip_checksum(icmph, len - IP_HDR_SIZE)
and due to the signature of compute_ip_checksum(), that would then
lead to accessing ~4G of address space, very likely leading to a
crash.
Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
There's no reason we should accept an IP packet with a malformed IHL
field. So ensure that it is exactly 5, not just <= 5.
Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
Reviewed-by: Ramon Fried <rfried.dev@gmail.com>
Add new desc_per_cacheline property which lets a platform run RX descriptor
cleanup after every power-of-2 - 1 received packets instead of every packet.
This is useful on platforms where (axi_bus_width EQOS_AXI_WIDTH_n * DMA DSL
inter-descriptor word skip count + DMA descriptor size) is less than cache
line size, which necessitates packing multiple DMA descriptors into single
cache line.
In case of TX descriptors, this is not a problem, since the driver always
does synchronous TX, i.e. the TX descriptor is always written, flushed and
polled for completion in eqos_send().
In case of RX descriptors, it is necessary to update their status in bulk,
i.e. after the entire cache line worth of RX descriptors has been used up
to receive data.
Signed-off-by: Marek Vasut <marex@denx.de>
Reviewed-by: Patrice Chotard <patrice.chotard@foss.st.com>
Reviewed-by: Ramon Fried <rfried.dev@gmail.com>
Separate TX and RX DMA rings to make their handling slightly clearer.
This is a preparatory patch for bulk RX descriptor flushing.
Signed-off-by: Marek Vasut <marex@denx.de>
Reviewed-by: Patrice Chotard <patrice.chotard@foss.st.com>
Reviewed-by: Ramon Fried <rfried.dev@gmail.com>
LiteX is a soft system-on-chip that targets FPGAs. LiteETH is a basic
network device that is commonly used in LiteX designs.
Signed-off-by: Joel Stanley <joel@jms.id.au>
Reviewed-by: Ramon Fried <rfried.dev@gmail.com>
Drop support for quickly deprecated DT property "snps,ref-clock-period-ns"
to prevent its proliferation.
Reviewed-by: Sean Anderson <seanga2@gmail.com>
Signed-off-by: Marek Vasut <marex@denx.de>
GUCTL.REFCLKPER can only account for clock frequencies with integer
periods. To address this, program REFCLK_FLADJ with the relative error
caused by period truncation. The formula given in the register reference
has been rearranged to allow calculation based on rate (instead of
period), and to allow for fixed-point arithmetic.
Additionally, calculate a value for 240MHZDECR. This configures a
simulated 240Mhz clock using a counter with one fractional bit (PLS1).
This register is programmed only for versions >= 2.50a, since this is
the check also used by commit db2be4e9e30c ("usb: dwc3: Add frame length
adjustment quirk").
[ marek: Ported from Linux kernel commit
596c87856e08d ("usb: dwc3: Program GFLADJ") ]
Reviewed-by: Sean Anderson <seanga2@gmail.com>
Signed-off-by: Sean Anderson <sean.anderson@seco.com>
Signed-off-by: Marek Vasut <marex@denx.de> # Port from Linux
Instead of using a special property to determine the reference clock
period, use the rate of the reference clock. When we have a legacy
snps,ref-clock-period-ns property and no reference clock, use it
instead. Fractional clocks are not currently supported, and will be
dealt with in the next commit.
[ marek: Ported from Linux kernel commit
5114c3ee24875 ("usb: dwc3: Calculate REFCLKPER based on reference clock") ]
Reviewed-by: Sean Anderson <seanga2@gmail.com>
Signed-off-by: Sean Anderson <sean.anderson@seco.com>
Signed-off-by: Marek Vasut <marex@denx.de> # Port from Linux
Set reference clock period when it differs from dwc3 default hardware
set.
We could calculate clock period based on reference clock frequency. But
this information is not always available. This is the case of PCI bus
attached USB host. For that reason we use a custom property.
Tested (USB2 only) on IPQ6010 SoC based board with 24 MHz reference
clock while hardware default is 19.2 MHz.
[ baruch: rewrite commit message; drop GFLADJ code; remove 'quirk-' from
property name; mention tested hardware ]
[ marek: Ported from Linux kernel commit
7bee318838890 ("usb: dwc3: reference clock period configuration") ]
Reviewed-by: Sean Anderson <seanga2@gmail.com>
Signed-off-by: Balaji Prakash J <bjagadee@codeaurora.org>
Signed-off-by: Baruch Siach <baruch@tkos.co.il>
Signed-off-by: Marek Vasut <marex@denx.de> # Port from Linux
Cache ref_clk clock pointer in struct dwc3 . This is a preparatory
patch for subsequent backports from Linux kernel which configure
GFLADJ register content based on the ref_clk rate and therefore need
access to the ref_clk pointer.
It is possible to extract the clock pointer from existing clk_bulk
list of already claimed clock, no need to call clk_get*() again.
Reviewed-by: Sean Anderson <seanga2@gmail.com>
Signed-off-by: Marek Vasut <marex@denx.de>
The "generic_bus_%x_dev_%x" string which is printed into this buffer
can be up to 34 characters long ("generic_bus_12345678_dev_12345678").
The buffer would be clipped by snprintf() if both %x were at maximum
range. Make sure the buffer is long enough to cover such possibility.
Signed-off-by: Marek Vasut <marex@denx.de>
Reviewed-by: Simon Glass <sjg@chromium.org>
It seems this symbol was missed when renaming DM_VIDEO -> VIDEO. Update
it.
Fixes: b86986c7b3 ("video: Rename CONFIG_DM_VIDEO to CONFIG_VIDEO")
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Signed-off-by: John Keeping <john@metanate.com>
We have a number of jobs that will have git complain about needing to
set safe.directory and this being untrue as a fatal error, but then
complete. Set this flag correctly now as it should be used, and may
prevent a future failure.
Signed-off-by: Tom Rini <trini@konsulko.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
Use macro to represent the RL and WL setting to ensure the PHY and
controller setting are aligned.
Review-by: Ryan Chen <ryan_chen@aspeedtech.com>
Signed-off-by: Dylan Hung <dylan_hung@aspeedtech.com>
Adjust the following settings to get better timing and signal quality.
1. write DQS/DQ delay
- 1e6e2304[0]
- 1e6e2304[15:8]
2. read DQS/DQ delay
- 0x1e6e0298[0]
- 0x1e6e0298[15:8]
3. CLK/CA timing
- 0x1e6e01a8[31]
4. Read and write termination
- change RTT_ROM from 40 ohm to 48 ohm (MR1[10:8])
- change RTT_PARK from disable to 48 ohm (MR5[8:6])
- change RTT_WR from 120 ohm to disable (MR2[11:9])
- change PHY ODT from 40 ohm to 80 ohm (0x1e6e0130[10:8])
Note1: Both DDR-PHY and DDR controller have their own registers for DDR4
Mode Registers (MR0~MR6). This patch introduces macros to synchronize
the MR value on both sides.
Note2: the waveform meansurement can be found in item #21 of Aspeed
AST26x0 Application note (AP note).
Review-by: Ryan Chen <ryan_chen@aspeedtech.com>
Signed-off-by: Dylan Hung <dylan_hung@aspeedtech.com>
The condition "~data" in the if-statement is a typo. The original
intention is to poll if SDRAM_PHYCTRL0_INIT bit equals to 0. So use
"data == 0" for instead.
Besides, the bit[1] of "phy_status" register is hardwired to
SDRAM_PHYCTRL0_INIT (with inverse logic). Since SDRAM_PHYCTRL0_INIT has
already done, remove the unnecessary checking of phy_status[1].
Fixes: fde9314346 ("ram: aspeed: Add AST2600 DRAM control support")
Review-by: Ryan Chen <ryan_chen@aspeedtech.com>
Signed-off-by: Dylan Hung <dylan_hung@aspeedtech.com>
This header was already included just above version.h,
do not include it twice.
Fixes: 3db7110857 ("crc32: Use the crc.h header for crc functions")
Signed-off-by: Marek Vasut <marex@denx.de>
Reviewed-by: Simon Glass <sjg@chromium.org>
Return value -1 cause U-Boot to print usage message. Return value
1 (CMD_RET_FAILURE) indicates failure. So fix return value when ubifs
command starts it execution and fails.
Signed-off-by: Pali Rohár <pali@kernel.org>
Reviewed-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Dhruva Gole <d-gole@ti.com>
Debug dump logs are not always required. Add a new config option
UBIFS_SILENCE_DEBUG_DUMP to silence all debug dumps. On powerpc/mpc85xx
when enabled this will decrease size of U-Boot binary by 11 kB.
Signed-off-by: Pali Rohár <pali@kernel.org>
Tested-by: Tony Dinh <mibodhi@gmail.com>
- Re-add the old PHY reset binding for nanopi-k2
-----BEGIN PGP SIGNATURE-----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=yEaP
-----END PGP SIGNATURE-----
Merge tag 'u-boot-amlogic-20221122' of https://source.denx.de/u-boot/custodians/u-boot-amlogic
- Implement setbrg op to meson serial device
- Re-add the old PHY reset binding for nanopi-k2
Add test case for 'fdt get value' sub command.
The test case can be triggered using:
"
./u-boot -d u-boot.dtb -c 'ut fdt'
"
Acked-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Signed-off-by: Marek Vasut <marex@denx.de>
Always increment both the iterator and pointer into the string
property value by length of the current element + 1 (to cater
for the string delimiter), otherwise the element extracted from
the string property value would be extracted from an offset that
is multiple of the length of the first element, instead of sum
of element lengths until select index.
This fixes 'fdt get value' operation for index above 1 (counting
from index 0).
Acked-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Fixes: 13982ced2c ("cmd: fdt: Add support for reading stringlist property values")
Signed-off-by: Marek Vasut <marex@denx.de>
The sandbox config file is to be removed. Move the GUID declarations
needed for capsule update functionality to the board file where they
are used.
Signed-off-by: Sughosh Ganu <sughosh.ganu@linaro.org>
Reviewed-by: Simon Glass <sjg@chromium.org>
Invoking lseek() may result in an error. Handle it.
Addresses-Coverity-ID: 376212 ("Error handling issues (CHECKED_RETURN)")
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Add a new flag to buildman so that we will in turn pass
BINMAN_ALLOW_MISSING=1 to 'make'. Make use of this flag in CI.
Allow the settings file to control this.
Cc: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
Cc: Simon Glass <sjg@chromium.org>
Signed-off-by: Tom Rini <trini@konsulko.com>
Signed-off-by: Simon Glass <sjg@chromium.org>
This was dropped my mistake. Reinstate it.
Signed-off-by: Simon Glass <sjg@chromium.org>
Fixes: d829f1217c ("bulidman: Add support for a simple build")
While it is possible and documented on how to re-run buildman to replace
faked required binary files after the fact, this behavior ends up being
more confusing than helpful in practice. Switch to requiring
BINMAN_ALLOW_MISSING=1 to be passed on the 'make' line to enable this
behavior.
Cc: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
Cc: Simon Glass <sjg@chromium.org>
Signed-off-by: Tom Rini <trini@konsulko.com>
Signed-off-by: Simon Glass <sjg@chromium.org>
These are documented in various several sections. Add a new section that
mentions them all in one place so it is easier to see what environment
variables can be used to control U-Boot's use of binman.
Signed-off-by: Simon Glass <sjg@chromium.org>
Suggested-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Buildman should consider a build as a success (with warnings) if missing
blobs have been dealt with by binman, even though buildman itself returns
and error code overall. This is how other warnings are dealt with.
We cannot easily access the 103 exit code, so detect the problem in the
output.
With this change, missing blobs result in an exit code of 101, although
they still indicate failure.
Signed-off-by: Simon Glass <sjg@chromium.org>
Support for some architectures has been removed since buildman was first
written. Also all toolchains are now available at kernel.org so we don't
need the links, except for arc where the kernel.org toolchain fails to
build all boards.
Signed-off-by: Simon Glass <sjg@chromium.org>
The settings file omits a few lines which are useful for getting every
board building. Add these and update the documentation tool.
Signed-off-by: Simon Glass <sjg@chromium.org>
This script was removed about 6 years ago so most people should be aware
that it is not needed anymore. Drop mention of it.
Signed-off-by: Simon Glass <sjg@chromium.org>
Convert the buildman documentation to rST format and include it in the
'build' section.
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Quentin Schulz <foss+uboot@0leil.net>
At present binman returns success when told to handle missing/faked blobs
or missing bintools. This is confusing since in fact the resulting image
cannot work.
Use exit code 103 to signal this problem, with a -W option to convert
it to a warning.
Rename the flag to --ignore-missing since it controls bintools also.
Add documentation about exit codes while we are here.
Signed-off-by: Simon Glass <sjg@chromium.org>
This currently uses if_changed on a phony target. Use a real file as the
target and add FORCE at the end, as required. Drop the 'inputs' phony
since it is not needed.
Signed-off-by: Simon Glass <sjg@chromium.org>
gcc 12 seems to warn on strncpy() as a matter of course. Rewrite the code
a different way to do the same thing, to avoid the warning.
Signed-off-by: Simon Glass <sjg@chromium.org>
This reverts commit daa2da754a.
This commit is not needed anymore since the btool_ prefix is
automatically stripped by bintool.
Cc: Quentin Schulz <foss+uboot@0leil.net>
Signed-off-by: Quentin Schulz <quentin.schulz@theobroma-systems.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
The binary is looked on the system by the suffix of the packer class.
This means binman was looking for btool_gzip on the system and not gzip.
Since a btool can have its btool_ prefix missing but its module and
binary presence on the system appropriately found, there's no need to
actually keep this prefix after listing all possible btools, so let's
remove it.
This fixes gzip btool by letting Bintool.find_bintool_class handle the
missing prefix and still return the correct class which is then init
with gzip name instead of btool_gzip.
Additionally, there was an issue with the cached module global variable.
The variable only stores the module and not the associated class name
when calling find_bintool_class.
This means that when caching the module on the first call to
find_bintool_class, class_name would be set to Bintoolbtool_gzip but the
module_name gzip only, adding the module in the gzip key in the module
dictionary. When hitting the cache on next calls, the gzip key would be
found, so its value (the module) is used. However the default class_name
(Bintoolgzip) is used, failing the getattr call.
Instead, let's enforce the same class name: Bintool<packer>, whatever
the filename it is contained in.
Cc: Quentin Schulz <foss+uboot@0leil.net>
Signed-off-by: Quentin Schulz <quentin.schulz@theobroma-systems.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
The sync of device-tree/bindings in 11a48a5a18c6 ("Linux 5.6-rc2") causes
Ethernet to break on some GXBB boards; the PHY seems to need proper reset
timing to function in u-boot and Linux. Re-add the old PHY reset binding
for dwmac until we support new bindings in the PHY node. This borrows the
same fix applied to the Odroid C2 board [0].
[0] https://lists.denx.de/pipermail/u-boot/2021-April/446658.html
Fixes: dd5f2351e9 ("arm64: dts: meson: sync dt and bindings from v5.6-rc2")
Signed-off-by: Christian Hewitt <christianshewitt@gmail.com>
Link: https://lore.kernel.org/r/20221025143205.14470-1-christianshewitt@gmail.com
Signed-off-by: Neil Armstrong <neil.armstrong@linaro.org>
