This patch moves the call to ceph_osdc_build_request() out of
ceph_osdc_new_request() and into its caller.

This is in order to defer formatting osd operation information into
the request message until just before request is started.

The only unusual (ab)user of ceph_osdc_build_request() is
ceph_writepages_start(), where the final length of write request may
change (downward) based on the current inode size or the oldest
snapshot context with dirty data for the inode.

The remaining callers don't change anything in the request after has
been built.

This means the ops array is now supplied by the caller.  It also
means there is no need to pass the mtime to ceph_osdc_new_request()
(it gets provided to ceph_osdc_build_request()).  And rather than
passing a do_sync flag, have the number of ops in the ops array
supplied imply adding a second STARTSYNC operation after the READ or
WRITE requested.

This and some of the patches that follow are related to having the
messenger (only) be responsible for filling the content of the
message header, as described here:
    http://tracker.ceph.com/issues/4589Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Keep track of the length of the data portion for a message in a
separate field in the ceph_msg structure.  This information has
been maintained in wire byte order in the message header, but
that's going to change soon.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

When assigning a bio pointer to an osd request, we don't have an
efficient way of knowing the total length bytes in the bio list.
That information is available at the point it's set up by the rbd
code, so record it with the osd data when it's set.

This and the next patch are related to maintaining the length of a
message's data independent of the message header, as described here:
    http://tracker.ceph.com/issues/4589Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

A field in an osd request keeps track of whether a connection is
currently filling the request's reply message.  This patch gets rid
of that field.

An osd request includes two messages--a request and a reply--and
they're both associated with the connection that existed to its
the target osd at the time the request was created.

An osd request can be dropped early, even when it's in flight.
And at that time both messages are released.  It's possible the
reply message has been supplied to its connection to receive
an incoming response message at the time the osd request gets
dropped.  So ceph_osdc_release_request() revokes that message
from the connection before releasing it so things get cleaned up
properly.

Previously this may have caused a problem, because the connection
that a message was associated with might have gone away before the
revoke request.  And to avoid any problems using that connection,
the osd client held a reference to it when it supplies its response
message.

However since this commit:
    38941f80 libceph: have messages point to their connection
all messages hold a reference to the connection they are associated
with whenever the connection is actively operating on the message
(i.e. while the message is queued to send or sending, and when it
data is being received into it).  And if a message has no connection
associated with it, ceph_msg_revoke_incoming() won't do anything
when asked to revoke it.

As a result, there is no need to keep an additional reference to the
connection associated with a message when we hand the message to the
messenger when it calls our alloc_msg() method to receive something.
If the connection *were* operating on it, it would have its own
reference, and if not, there's no work to be done when we need to
revoke it.

So get rid of the osd request's r_con_filling_msg field.

This resolves:
    http://tracker.ceph.com/issues/4647Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

There's one spot in ceph_writepages_start() that open-codes what
page_offset() does safely.  Use the macro so we don't have to worry
about wrapping.

This resolves:
    http://tracker.ceph.com/issues/4648Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

There are two basically identical definitions of __decode_pgid()
in libceph, one in "net/ceph/osdmap.c" and the other in
"net/ceph/osd_client.c".  Get rid of both, and instead define
a single inline version in "include/linux/ceph/osdmap.h".
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The osd client mutex is acquired just before getting a reference to
a request in handle_reply().  However the error paths after that
don't drop the mutex before returning as they should.

Drop the mutex after dropping the request reference.  Also add a
bad_mutex label at that point and use it so the failed request
lookup case can be handled with the rest.

This resolves:
    http://tracker.ceph.com/issues/4615Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>

In create_fs_client() a memory pool is set up be used for arrays of
pages that might be needed in ceph_writepages_start() if memory is
tight.  There are two problems with the way it's initialized:
    - The size provided is the number of pages we want in the
      array, but it should be the number of bytes required for
      that many page pointers.
    - The number of pages computed can end up being 0, while we
      will always need at least one page.

This patch fixes both of these problems.

This resolves the two simple problems defined in:
    http://tracker.ceph.com/issues/4603Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Use osd_req_op_extent_init() in ceph_osdc_new_request() to
initialize the one or two ops built in that function.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

All callers of ceph_osd_new_request() pass either CEPH_OSD_OP_READ
or CEPH_OSD_OP_WRITE as the opcode value.  The function assumes it
by filling in the extent fields in the ops array it builds.  So just
assert that is the case, and don't bother calling op_has_extent()
before filling in the first osd operation in the array.

Define some local variables to gather the information to fill into
the first op, and then fill in the op array all in one place.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The ceph_osdc_new_request() an array of osd operations is built up
and filled in partially within that function and partially in the
called function calc_layout().  Move the latter part back out to
ceph_osdc_new_request() so it's all done in one place.  This makes
it unnecessary to pass the op pointer to calc_layout(), so get rid
of that parameter.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The purpose of calc_layout() is to determine, given a file offset
and length and a layout describing the placement of file data across
objects, where in "object space" that data resides.

Specifically, it determines which object should hold the first part
of the specified range of file data, and the offset and length of
data within that object.  The length will not exceed the bounds
of the object, and the caller is informed of that maximum length.

Add two parameters to calc_layout() to allow the object-relative
offset and length to be passed back to the caller.

This is the first steps toward having ceph_osdc_new_request() build
its osd op structure using osd_req_op_extent_init().
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The rbd code has a function that allocates and populates a
ceph_osd_req_op structure (the in-core version of an osd request
operation).  When reviewed, Josh suggested two things: that the
big varargs function might be better split into type-specific
functions; and that this functionality really belongs in the osd
client rather than rbd.

This patch implements both of Josh's suggestions.  It breaks
up the rbd function into separate functions and defines them
in the osd client module as exported interfaces.  Unlike the
rbd version, however, the functions don't allocate an osd_req_op
structure; they are provided the address of one and that is
initialized instead.

The rbd function has been eliminated and calls to it have been
replaced by calls to the new routines.  The rbd code now now use a
stack (struct) variable to hold the op rather than allocating and
freeing it each time.

For now only the capabilities used by rbd are implemented.
Implementing all the other osd op types, and making the rest of the
code use it will be done separately, in the next few patches.

Note that only the extent, cls, and watch portions of the
ceph_osd_req_op structure are currently used.  Delete the others
(xattr, pgls, and snap) from its definition so nobody thinks it's
actually implemented or needed.  We can add it back again later
if needed, when we know it's been tested.

This (and a few follow-on patches) resolves:
    http://tracker.ceph.com/issues/3861Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Define a separate function to determine the validity of an opcode,
and use it inside osd_req_encode_op() in order to unclutter that
function.

Don't update the destination op at all--and return zero--if an
unsupported or unrecognized opcode is seen in osd_req_encode_op().
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Move some definitions for max integer values out of the rbd code and
into the more central "decode.h" header file.  These really belong
in a Linux (or libc) header somewhere, but I haven't gotten around
to proposing that yet.

This is in preparation for moving some code out of rbd.c and into
the osd client.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

In ceph_osdc_build_request() there is a call to cpu_to_le16() which
provides a 64-bit value as its argument.  Because of the implied
byte swapping going on it looked pretty suspect to me.

At the moment it turns out the behavior is well defined, but masking
off those bottom bits explicitly eliminates this distraction, and is
in fact more directly related to the purpose of the message header's
data_off field.

This resolves:
    http://tracker.ceph.com/issues/4125Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

When a cursor for a page array data message is initialized it needs
to determine the initial value for cursor->last_piece.  Currently it
just checks if length is less than a page, but that's not correct.
The data in the first page in the array will be offset by a page
offset based on the alignment recorded for the data.  (All pages
thereafter will be aligned at the base of the page, so there's
no need to account for this except for the first page.)

Because this was wrong, there was a case where the length of a piece
would be calculated as all of the residual bytes in the message and
that plus the page offset could exceed the length of a page.

So fix this case.  Make sure the sum won't wrap.

This resolves a third issue described in:
    http://tracker.ceph.com/issues/4598Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>

Currently ceph_msg_data_pages_advance() allows the page offset value
to be PAGE_SIZE, apparently assuming ceph_msg_data_pages_next() will
treat it as 0.  But that doesn't happen, and the result led to a
helpful assertion failure.

Change ceph_msg_data_pages_advance() to truncate the offset to 0
before returning if it reaches PAGE_SIZE.

Make a few other minor adjustments in this area (comments and a
better assertion) while modifying it.

This resolves a second issue described in:
    http://tracker.ceph.com/issues/4598Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>

It's OK for the result of a read to come back with fewer bytes than
were requested.  So don't trigger a BUG() in that case when
initializing the data cursor.

This resolves the first problem described in:
    http://tracker.ceph.com/issues/4598Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>

Begin the transition from a single message data item to a list of
them by replacing the "data" structure in a message with a pointer
to a ceph_msg_data structure.

A null pointer will indicate the message has no data; replace the
use of ceph_msg_has_data() with a simple check for a null pointer.

Create functions ceph_msg_data_create() and ceph_msg_data_destroy()
to dynamically allocate and free a data item structure of a given type.

When a message has its data item "set," allocate one of these to
hold the data description, and free it when the last reference to
the message is dropped.

This partially resolves:
    http://tracker.ceph.com/issues/4429Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The *_msg_pos_next() functions do little more than call
ceph_msg_data_advance().  Replace those wrapper functions with
a simple call to ceph_msg_data_advance().

This cleanup is related to:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

In write_partial_message_data() we aggregate the crc for the data
portion of the message as each new piece of the data item is
encountered.  Because it was computed *before* sending the data, if
an attempt to send a new piece resulted in 0 bytes being sent, the
crc crc across that piece would erroneously get computed again and
added to the aggregate result.  This would occasionally happen in
the evnet of a connection failure.

The crc value isn't really needed until the complete value is known
after sending all data, so there's no need to compute it before
sending.

So don't calculate the crc for a piece until *after* we know at
least one byte of it has been sent.  That will avoid this problem.

This resolves:
    http://tracker.ceph.com/issues/4450Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>

The only remaining field in the ceph_msg_pos structure is
did_page_crc.  In the new cursor model of things that flag (or
something like it) belongs in the cursor.

Define a new field "need_crc" in the cursor (which applies to all
types of data) and initialize it to true whenever a cursor is
initialized.

In write_partial_message_data(), the data CRC still will be computed
as before, but it will check the cursor->need_crc field to determine
whether it's needed.  Any time the cursor is advanced to a new piece
of a data item, need_crc will be set, and this will cause the crc
for that entire piece to be accumulated into the data crc.

In write_partial_message_data() the intermediate crc value is now
held in a local variable so it doesn't have to be byte-swapped so
many times.  In read_partial_msg_data() we do something similar
(but mainly for consistency there).

With that, the ceph_msg_pos structure can go away,  and it no longer
needs to be passed as an argument to prepare_message_data().

This cleanup is related to:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

All but one of the fields in the ceph_msg_pos structure are now
never used (only assigned), so get rid of them.  This allows
several small blocks of code to go away.

This is cleanup of old code related to:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Use the "resid" field of a cursor rather than finding when the
message data position has moved up to meet the data length to
determine when all data has been sent or received in
write_partial_message_data() and read_partial_msg_data().

This is cleanup of old code related to:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

It turns out that only one of the data item types is ever used at
any one time in a single message (currently).
    - A page array is used by the osd client (on behalf of the file
      system) and by rbd.  Only one osd op (and therefore at most
      one data item) is ever used at a time by rbd.  And the only
      time the file system sends two, the second op contains no
      data.
    - A bio is only used by the rbd client (and again, only one
      data item per message)
    - A page list is used by the file system and by rbd for outgoing
      data, but only one op (and one data item) at a time.

We can therefore collapse all three of our data item fields into a
single field "data", and depend on the messenger code to properly
handle it based on its type.

This allows us to eliminate quite a bit of duplicated code.

This is related to:
    http://tracker.ceph.com/issues/4429Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Now that read_partial_message_pages() and read_partial_message_bio()
are literally identical functions we can factor them out.  They're
pretty simple as well, so just move their relevant content into
read_partial_msg_data().

This is and previous patches together resolve:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

There is handling in write_partial_message_data() for the case where
only the length of--and no other information about--the data to be
sent has been specified.  It uses the zero page as the source of
data to send in this case.

This case doesn't occur.  All message senders set up a page array,
pagelist, or bio describing the data to be sent.  So eliminate the
block of code that handles this (but check and issue a warning for
now, just in case it happens for some reason).

This resolves:
    http://tracker.ceph.com/issues/4426Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The cursor code for a page array selects the right page, page
offset, and length to use for a ceph_tcp_recvpage() call, so
we can use it to replace a block in read_partial_message_pages().

This partially resolves:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The bio_iter and bio_seg fields in a message are no longer used, we
use the cursor instead.  So get rid of them and the functions that
operate on them them.

This is related to:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

Replace the use of the information in con->in_msg_pos for incoming
bio data.  The old in_msg_pos and the new cursor mechanism do
basically the same thing, just slightly differently.

The main functional difference is that in_msg_pos keeps track of the
length of the complete bio list, and assumed it was fully consumed
when that many bytes had been transferred.  The cursor does not assume
a length, it simply consumes all bytes in the bio list.  Because the
only user of bio data is the rbd client, and because the length of a
bio list provided by rbd client always matches the number of bytes
in the list, both ways of tracking length are equivalent.

In addition, for in_msg_pos the initial bio vector is selected as
the initial value of the bio->bi_idx, while the cursor assumes this
is zero.  Again, the rbd client always passes 0 as the initial index
so the effect is the same.

Other than that, they basically match:
    in_msg_pos      cursor
    ----------      ------
    bio_iter        bio
    bio_seg         vec_index
    page_pos        page_offset

The in_msg_pos field is initialized by a call to init_bio_iter().
The bio cursor is initialized by ceph_msg_data_cursor_init().
Both now happen in the same spot, in prepare_message_data().

The in_msg_pos field is advanced by a call to in_msg_pos_next(),
which updates page_pos and calls iter_bio_next() to move to the next
bio vector, or to the next bio in the list.  The cursor is advanced
by ceph_msg_data_advance().  That isn't currently happening so
add a call to that in in_msg_pos_next().

Finally, the next piece of data to use for a read is determined
by a bunch of lines in read_partial_message_bio().  Those can be
replaced by an equivalent ceph_msg_data_bio_next() call.

This partially resolves:
    http://tracker.ceph.com/issues/4428Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

All of the data types can use this, not just the page array.  Until
now, only the bio type doesn't have it available, and only the
initiator of the request (the rbd client) is able to supply the
length of the full request without re-scanning the bio list.  Change
the cursor init routines so the length is supplied based on the
message header "data_len" field, and use that length to intiialize
the "resid" field of the cursor.

In addition, change the way "last_piece" is defined so it is based
on the residual number of bytes in the original request.  This is
necessary (at least for bio messages) because it is possible for
a read request to succeed without consuming all of the space
available in the data buffer.

This resolves:
    http://tracker.ceph.com/issues/4427Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

The value passed for "pages" in read_partial_message_pages() is
always the pages pointer from the incoming message, which can be
derived inside that function.  So just get rid of the parameter.
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

When the last reference to a ceph message is dropped,
ceph_msg_last_put() is called to clean things up.

For "normal" messages (allocated via ceph_msg_new() rather than
being allocated from a memory pool) it's sufficient to just release
resources.  But for a mempool-allocated message we actually have to
re-initialize the data fields in the message back to initial state
so they're ready to go in the event the message gets reused.

Some of this was already done; this fleshes it out so it's done
more completely.

This resolves:
    http://tracker.ceph.com/issues/4540Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>

An osd expects the transaction ids of arriving request messages from
a given client to a given osd to increase monotonically.  So the osd
client needs to send its requests in ascending tid order.

The transaction id for a request is set at the time it is
registered, in __register_request().  This is also where the request
gets placed at the end of the osd client's unsent messages list.

At the end of ceph_osdc_start_request(), the request message for a
newly-mapped osd request is supplied to the messenger to be sent
(via __send_request()).  If any other messages were present in the
osd client's unsent list at that point they would be sent *after*
this new request message.

Because those unsent messages have already been registered, their
tids would be lower than the newly-mapped request message, and
sending that message first can violate the tid ordering rule.

Rather than sending the new request only, send all queued requests
(including the new one) at that point in ceph_osdc_start_request().
This ensures the tid ordering property is preserved.

With this in place, all messages should now be sent in tid order
regardless of whether they're being sent for the first time or
re-sent as a result of a call to osd_reset().

This resolves:
    http://tracker.ceph.com/issues/4392Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>

In __map_request(), when adding a request to an osd client's unsent
list, add it to the tail rather than the head.  That way the newest
entries (with the highest tid value) will be last.

Maintain an osd's request list in order of increasing tid also.

Finally--to be consistent--maintain an osd client's "notarget" list
in that order as well.

This partially resolves:
    http://tracker.ceph.com/issues/4392Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>

The osd expects incoming requests for a given object from a given
client to arrive in order, with the tid for each request being
greater than the tid for requests that have already arrived.  This
patch fixes two places the osd client might not maintain that
ordering.

For the osd client, the connection fault method is osd_reset().
That function calls __reset_osd() to close and re-open the
connection, then calls __kick_osd_requests() to cause all
outstanding requests for the affected osd to be re-sent after
the connection has been re-established.

When an osd is reset, any in-flight messages will need to be
re-sent.  An osd client maintains distinct lists for unsent and
in-flight messages.  Meanwhile, an osd maintains a single list of
all its requests (both sent and un-sent).  (Each message is linked
into two lists--one for the osd client and one list for the osd.)

To process an osd "kick" operation, the request list for the *osd*
is traversed, and each request is moved off whichever osd *client*
list it was on (unsent or sent) and placed onto the osd client's
unsent list.  (It remains where it is on the osd's request list.)

When that is done, osd_reset() calls __send_queued() to cause each
of the osd client's unsent messages to be sent.

OK, with that background...

As the osd request list is traversed each request is prepended to
the osd client's unsent list in the order they're seen.  The effect
of this is to reverse the order of these requests as they are put
(back) onto the unsent list.

Instead, build up a list of only the requests for an osd that have
already been sent (by checking their r_sent flag values).  Once an
unsent request is found, stop examining requests and prepend the
requests that need re-sending to the osd client's unsent list.

Preserve the original order of requests in the process (previously
re-queued requests were reversed in this process).  Because they
have already been sent, they will have lower tids than any request
already present on the unsent list.

Just below that, traverse the linger list in forward order as
before, but add them to the *tail* of the list rather than the head.
These requests get re-registered, and in the process are give a new
(higher) tid, so the should go at the end.

This partially resolves:
    http://tracker.ceph.com/issues/4392Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>

Since we no longer drop the request mutex between registering and
mapping an osd request in ceph_osdc_start_request(), there is no
chance of a race with kick_requests().

We can now therefore map and send the new request unconditionally
(but we'll issue a warning should it ever occur).
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>

One of the first things ceph_osdc_start_request() does is register
the request.  It then acquires the osd client's map semaphore and
request mutex and proceeds to map and send the request.

There is no reason the request has to be registered before acquiring
the map semaphore.  So hold off doing so until after the map
semaphore is held.

Since register_request() is nothing more than a wrapper around
__register_request(), call the latter function instead, after
acquiring the request mutex.

That leaves register_request() unused, so get rid of it.

This partially resolves:
    http://tracker.ceph.com/issues/4392Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>

The auth code is called from a variety of contexts, include the mon_client
(protected by the monc's mutex) and the messenger callbacks (currently
protected by nothing).  Avoid chaos by protecting all auth state with a
mutex.  Nothing is blocking, so this should be simple and lightweight.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>