When I create a new PPP connection, I am seeing a hardfault (segfault)
coming from pbuf_free.
I traced the problem to an invalid in_head field of the pppos_pcb structure.
The field is invalid because the memory is never cleared to zero after the
pppos_pcb structure is created in pppos_create().
I was able to fix the issue by adding a memset after the memp_malloc call.
Signed-off-by: Sylvain Rochet <gradator@gradator.net>
Jiffies isn't really a humanly readable value and it means the default
PPP_MAXIDLEFLAG period depends on the platform "jiffies" frequency,
which isn't nice.
Change PPP_MAXIDLEFLAG to use ms instead of jiffies, the current
PPP_MAXIDLEFLAG default (100 ms), looks like a sane value and is
left unchanged.
The overall lwIP design on data flows (netif,udp,tcp) is to use a user
defined callback to get data from stack and a static function to send
data to stack, which makes perfect sense. The SIO port was an exception,
the PPP stack never really used the SIO port by only using the
sio_send() function (and the ignominious sio_read_abort() function a
while back).
The way the SIO port is currently designed adds a tight coupling between
the lwIP port and the user code if the user need to do specific user
code if the current uart used is the PPPoS uart, which is not nice,
especially because all the lwIP stack is quite clean at this subject.
While we are at stabilizing the PPP API, change this behavior before
it's too late by replacing the static sio_write() calls to a user
defined callback.
There is no point of calling magic_randomize() for each pppos_input()
call, making magic_randomize() potentially called for each serial input
byte which is quite a bad idea since magic_randomize() is quite
intensive in processing time (MD5 computation) compared to HDLC frame
parsing. There is no entropy added when being called for each input byte
rather than for each valid input packet because byte input is a
monotonic event at the packet level. Well, if packet arrival time is a
valid entropy source even so, which I doubt a lot, but we don't really
have anything else and we really need random for PPP authentication
layers.
PBUF_LINK_ENCAPSULATION_HLEN support was introduced by 6ef7563f and
missed the fact that header size calculation/reservation using
computation like PBUF_LINK_HLEN + PBUF_IP_HLEN + ... are used all over
the source code. Hopefully fixed all of them.
We need to do VJ compression before CCP/MPPE compression and VJ
decompression after CCP/MPPE decompression. This leads to a massive
rewrite of how we currently handled VJ only in the PPPoS lower protocol
handler.
Moved VJ structures from pppos to ppp_pcb because we need them back in
PPP core. This is a bit unfortunate because that's not necessary for
PPPoE or PPPoL2TP, but, hey!. Fixed CCP+MPPE+VJ order.
In PPP, we previously know if we are dealing with a IPv4 or a IPv6 packet,
we don't need to use the ip_input() dispatch function, removing a useless
if and reducing call stack by one.
ip_addr_t is used for all generic IP addresses for the API, ip(4/6)_addr_t are only used internally or when initializing netifs or when calling version-related functions
Follow-up of the #44565 bug fix, renamed the misnamed
PPP_INPROC_MULTITHREADED to PPP_INPROC_IRQ_SAFE because it is
IRQ safe but not thread safe.
Updated PPP documentation which now clearly state when and how
this feature can be used.
Renamed pppos_drop() → pppos_input_drop()
Renamed pppos_free_current_input_packet() → pppos_input_free_current_packet()
Moved pppos_output_last() after pppos_output_append()
Moved pppos_input_free_current_packet() before pppos_input_drop()
We actually allocated a pbuf chain only to iterate later the linked list
calling sio_write() for each pbuf, improved by calling sio_write() when
buffer is full and by recycling the pbuf, therefore only using one pbuf
for PPPoS output path.
Reworked pppos_write() and pppos_netif_output() to share more common
code into pppos_output_append() and pppos_output_last().
If PPP_INPROC_MULTITHREADED is true, then user does not what to use
the TCPIP API. Disabling the TCPIP API helps the user to understand
that PPP_INPROC_MULTITHREADED must not be used if he wish to use
the TCPIP API.
!NO_SYS users may now use as well the TCPIP API for PPPoS input data,
this way they can disable PPP_INPROC_MULTITHREADED and run pppos_input()
inside the lwIP thread, which fixes, at least for them, all the
threading issues related to PPP_INPROC_MULTITHREADED.
If PPP_INPROC_MULTITHREADED is not enabled, we can free unfinished
RX pbuf from the pppos_disconnect() function because pppos_input()
is running in the same context. Thanks to the pppos->open flags we
now only need to free remaining pbuf in the disconnect function
if PPP_INPROC_MULTITHREADED is not enabled.
Don't process input data if PPPoS is closed, it helps using
pppos_input() from a different context to prevent pppos_input() to
modify PPPoS RX machine state on a closed PPPoS session. It also
prevents allocating pbuf (which are going to be tossed out by PPP core)
and parsing serial input on a closed session.
It only mitigates the fact that this function is actually NOT thread
safe in absolutely all cases, it does not fix it but it helps for a low
cost.
For example user application should never call pppos_input() while
pppos_connect() or pppos_listen() is currently running because both of
them are freeing any input pbuf left over from the last session before
resetting the PPPoS state, they really have to to prevent pbuf leaks.
We cannot fix that easily because we don't have spinlock with an
irqsave/irqrestore helper for IRQ contexts. Mutex cannot be used in
interrupt contexts (or again, with an IRQ mutex helper).
We are going to improve the documentation on this point.
