When flushing the packetlist, we must only add the request as a zombie
if no response has already been received. Otherwise we could wrongly
make it a zombie even though the response was already received and then
we'd get a zombie stuck there "forever"...
Since the sftp_packetlist_flush() function will move all the existing
FXP_READ requests in this handle to the zombie list we must first remove
this just received packet as it is clearly not a zombie.
Exactly as the comment in the code said, checking the return code from
sftp_packet_read() with <= was wrong and it should be < 0. With the new
filtering on incoming packets that are "zombies" we can now see this
getting zero returned.
In order to be fast, sftp_read sends many read requests at once. With a small
file, this can mean that when EOF is received back, many of these requests are
still outstanding. Responses arriving after we close the file and abandon the
file handle are queued in the SFTP packet queue and never collected. This
causes transfer speed to drop as a progressively longer queue must be searched
for every packet.
This change introduces a zombie request-ID list in the SFTP session that is
used to recognise these outstanding requests and prevent them being added to
the queue.
When calling _libssh2_channel_receive_window_adjust() internally, we now
always use the 'force' option to prevent libssh2 to avoid sending the
update if the update isn't big enough.
It isn't fully analyzed but we have seen corner cases which made a
necessary window update not get send due to this and then the other side
doesn't send data our side then sits waiting for forever.
The commit in 7194a9bd7ba45 wasn't complete. This change makes sure
variables are initialized properly before used in the EAGAIN and window
adjust cases.
Lots of places in the code translated the original error into the more
generic LIBSSH2_ERROR_SOCKET_TIMEOUT but this turns out to distort the
original error reason a lot and makes tracking down the real origin of a
problem really hard. This change makes the original error code be
preserved to a larger extent when return up to the parent function.
Commit 03ca9020756 tried to simplify the window sizing logic but broke
SFTP readdir as there was no window sizing code left there so large
directory listings no longer worked.
This change introduces window sizing logic to the sftp_packet_read()
function so that it now tells the remote about the local size having a
window size that suffice when it is about to ask for directory data.
Bug: http://www.libssh2.org/mail/libssh2-devel-archive-2012-03/0069.shtml
Reported by: Eric
Whenever we have acked data and is about to call a function that *MAY*
return EAGAIN we must return the number now and wait to get called
again. Our API only allows data *or* EAGAIN and we must never try to get
both.
Removed the total_read variable that originally must have tracked how
much data had been written to the buffer. With non-blocking reads, we
must return straight away once we have read data into the buffer so this
variable served not purpose.
I think it was still hanging around in case the initial processing of
'leftover' data meant we wrote to the buffer but this case, like the
others, must return immediately. Now that it does, the last remaining
need for the variable is gone.
Whenever we have data and is about to call a function that *MAY* return
EAGAIN we must return the data now and wait to get called again. Our API
only allows data *or* EAGAIN and we must never try to get both.
Some SFTP servers send SFTP packets larger than 40000. Since the limit
is only present to avoid insane sizes anyway, we can easily bump it.
The define was formerly in the public header libssh2_sftp.h but served
no external purpose and was moved into the source dir.
Bug: http://www.libssh2.org/mail/libssh2-devel-archive-2011-11/0004.shtml
Reported by: Michael Harris
Set read_state back to idle before trying to send anything so that if
the state somehow is wrongly set.
Also, avoid such a case of confusion by resetting the read_state when an
sftp handle is closed.
Removed the automatic window_size adjustments from
_libssh2_channel_read() and instead all channel readers must now make
sure to enlarge the window sizes properly themselves.
libssh2_channel_read_ex() - the public function, now grows the window
size according to the requested buffer size. Applications can still opt
to grow the window more on demand. Larger windows tend to give higher
performance.
sftp_read() now uses the read-ahead logic to figure out a window_size.
A returned READ packet that is short will now only reduce the
offset.
This is a temporary fix as it is slightly better than the previous
approach but still not very good.
When seeking to a new position, flush the packetlist and buffered data
to prevent already received or pending data to wrongly get used when
sftp-reading from the new offset within the file.
In the case where a read packet has been received from the server, but
the entire contents couldn't be copied to the user-buffer, the data is
instead buffered and copied to the user's buffer in the next invocation
of sftp_read(). When that "extra" copy is made, the 'offset' pointer was
not advanced accordingly.
The biggest impact of this flaw was that the 'already' variable at the
top of the function that figures out how much data "ahead" that has
already been asked for would slowly go more and more out of sync, which
could lead to the file not being read all the way to the end.
This problem was most noticable in cases where the application would
only try to read the exact file size amount, like curl does. In the
examples libssh2 provides the sftp read function is most often called
with a fixed size large buffer and then the bug would not appear as
easily.
This bug was introduced in the SFTP rewrite in 1.2.8.
Bug: http://curl.haxx.se/mail/lib-2011-08/0305.htmlhttp://www.libssh2.org/mail/libssh2-devel-archive-2011-08/0085.shtml
As pointed out in bug #206, if a second invoke of libssh2_sftp_read()
would shrink the buffer size, libssh2 would go nuts and send out read
requests like crazy. This was due to an unsigned variable turning
"negative" by some wrong math, and that value would be the amount of
data attempt to pre-buffer!
Bug: http://trac.libssh2.org/ticket/206
If asked to read data into a buffer and the buffer is too small to hold
the data, this function now returns an error instead of as previously
just copy as much as fits.
By using a new separate struct for incoming SFTP packets and not sharing
the generic packet struct, we can get rid of an unused field and add a
new one dedicated for holding the request_id for the incoming
package. As sftp_packet_ask() is called fairly often, a "mere" integer
comparison is MUCH faster than the previous memcmp() of (typically) 5
bytes.
Make sure that we cleanup remainders when the handle is closed and when
the subsystem is shutdown.
Existing flaw: if a single handle sends packets that haven't been
replied to yet at the time when the handle is closed, those packets will
arrive later and end up in the generic packet brigade queue and they
will remain in there until flushed. They will use unnecessary memory,
make things slower and they will ruin the SFTP handling if the
request_id counter ever wraps (highly unlikely to every happen).
The SFTP read function now does transfers the same way the SFTP write
function was made to recently: it creates a list of many outgoing
FXP_READ packets that each asks for a small data chunk. The code then
tries to keep sending read request while collecting the acks for the
previous requests and returns the received data.
The new SFTP write code caused a regression as the seek function no
longer worked as it didn't set the write position properly.
It should be noted that seeking is STRONGLY PROHIBITED during upload, as
the upload magic uses two different offset positions and the multiple
outstanding packets etc make them sensitive to change in the midst of
operations.
This functionality was just verified with the new example code
sftp_append. This bug was filed as bug #202:
Bug: http://trac.libssh2.org/ticket/202
The SFTP handle struct now buffers number of acked bytes that haven't
yet been returned. The way this is used is as following:
1. sftp_write() gets called with a buffer of let say size 32000. We
split 32000 into 8 smaller packets and send them off one by one. One of
them gets acked before the function returns so 4000 is returned.
2. sftp_write() gets called again a short while after the previous one,
now with a much smaller size passed in to the function. Lets say 8000.
In the mean-time, all of the remaining packets from the previous call
have been acked (7*4000 = 28000). This function then returns 8000 as all
data passed in are already sent and it can't return any more than what
it got passed in. But we have 28000 bytes acked. We now store the
remaining 20000 in the handle->u.file.acked struct field to add up in
the next call.
3. sftp_write() gets called again, and now there's a backlogged 20000
bytes to return as fine and that will get skipped from the beginning
of the buffer that is passed in.
This function now only returns EAGAIN if a lower layer actually returned
EAGAIN to it. If nothing was acked and no EAGAIN was received, it will
now instead return 0.
