RE: [U-Boot-Users] uploading OS over network instead of u-boot do wnloading it from a server.
Wolfgang,
Sorry for the delay, but I've been unexpectedly out of the office. You make some good points that I'd like to address.
- You are assuming that Brian's customers are at least as sophisticated
as
Abatron customers. Brian's original statement leads me to believe this
is
not so.
Indeed I think they are. Typing "tftpsrv" (or which other ways there are to start the supplied application "tftpsrv.exe") is IMHO no a big challenge. And this is all you need to do.
Typing "tftpd" or whatever is not what I'd worry about. Many systems are installed and operated these days by folks with very limited technical skills. The problem I see is that 6 months after installation the customer's MIS group makes network changes and the target device stops working. At this point, the operator needs to diagnose the problem and determine what needs to be done. No matter how trivial that act of typing "tftpd" is, the person diagnosing the problem must first determine that the trivial act is required. While this diagnosis task is not hard for programmers and consultants, it may be quite a challenge for the customer.
Now, some might say that the appropriate thing to do in such a situation is to call MIS or the help desk. My experience with MIS departments is they do a good job of networking together Windows (or Solaris or Linux) boxes and printers, but they want nothing to do with other TCP/IP devices. The customer is on their own. I've seen this problem kill a project. Remember, the customer doesn't have to be embarrassed by the triviality of typing "tftpd", they can just switch to a competitor's product. They don't even have to give your sales person a reason.
U-boot supports the target-as-client method of downloading but does not support the target-as-server method. I've used both methods a number of times over the years and both have their advantages and disadvantages.
Your
Can you please explain the advantages of the boot loader providing server function?
It is a unavoidable fact of standard client/server pairs that the client has a-priori knowledge about the server, the server has no a-priori knowledge about the client. If the target is a multicast server, there is NO NETWORK SETUP AT ALL. In the case of target-as-client, the client must know at least the server's IP address (or DNS name). That knowledge increases the amount of setup. Further, during the course of maintenance and expansion, this information may become invalid. You can make the required information as small as possible, but you have to go through the analysis to test whether this is acceptable to the customer. My experience is that making the amount of information used by the client (that can be unilaterally changed by the MIS department) nil is the best approach.
The key to this is that my download server uses standard UDP over multicast IP. This has a number of advantages, including the fact that the download server is receive-only and has no traditional TCP/IP stack (I won't go into error control here, but that's taken care of as well).
Another advantage is that this technique can be used very early in the manufacturing process to load the software (via UDP/IP) that programs the MAC address (as well as model and serial numbers) into EEPROM. When this software is loaded, the target has no IP address or MAC address. I can even load many units in parallel on a test rack.
U-Boot is a Boot Loader, not an OS.
U-Boot shall be powerful to use, yet simple in design to make it easily portable to new boards and architectures.
This means U-Boot will be strictly single-tasking; it will not use virtual memory; it will use polling drivers (versus interrupt driven ones) whenever possible.
U-Boot shall use standard protocols and interfaces whenever possible.
My server-loader meets all of these criteria, except the "powerful to use" one. It is not only extremely simple in design but also extremely simple in implementation (which explains its small size of less than 8k). It only does two things, it loads programs in to memory and jumps to that program. Occum's razor seemed appropriate since we have to strongly avoid updating the boot code in the field. Note that my boot code does have a FLASH boot component that I haven't discussed here.
<> Standard Protocols and Custom Clients
My loader uses standard protocols as well. It uses UDP over multicast IP. Since I wanted a passive, receive-only server loader, I rejected TFTP early because it wouldn't solve all my problems.
You mentioned in a subsequent email to Brian that he would have to write a custom client. You are quite correct. My version of this client is written as a command-line tool that does the following: 0) create a standard multicast UDP socket and bind it appropriately, 1) open a file for binary read, 2) read 1k blocks from the file, 3) prefix a small payload header, and 4) send the block to the socket. Since we already had a socket library, this task took an afternoon and the result was just 5-6 pages of C code. This client hasn't been maintained after it's original creation and testing (although it has been extended with new features; e.g., alternate binary file formats, etc.).
It's one thing to say that you have found a boot loader (u-boot) and an operating system (Linux) for free off the web and made them work. In fact, this is very well received by many managers. But, I do have to develop my own embedded-target and host applications. For this I need programmers. The overhead of writing this custom client program (using standard protocols) was tiny compared to the entire project. Further, we were able to code this functionality into the host-based Visual Basic client used by our field reps. with very little effort. This is, I believe one of the features that Brian is thinking about.
I'm sorry, but IMHO there is no advantage running a server in the boot loader.
I'm sorry, but IMHO YHO is wrong. :-)
Perhaps you might be so kind as to elucidate.
Perhaps I have.
<> Summary
Ultimately, I believe the difference in the two approaches is in where the control is. I prefer to view the host as "in control" and the target as subservient. It has simply been my experience, over many projects, that this way works out better.
This whole discussion may not be appropriate on this list. Since I did this work as a paid employee, I must honor my employeer's policies that prevent sharing source code. I can, however, freely discuss the details and answer questions If Brian or anyone else cares to find out more about this technique, they are welcome to contact me directly.
Regards, Charlie
Charles.Wells@nielsenmedia.com c.wells@ieee.org
Dear Charles,
in message DF2B720CF774D21189EE00805FA7FA220B909B75@nmrusdunsx3.nielsenmedia.com you wrote:
It is a unavoidable fact of standard client/server pairs that the client has a-priori knowledge about the server, the server has no a-priori knowledge about the client. If the target is a multicast server, there is NO NETWORK
This is wrong. The only a-priori knowledge needed by the client to ruin any of the standard network boot protocols (like BOOTP or DHCP) is it's own MAC address, which should have been programmed by the manufacturer and which will never need to be changed.
SETUP AT ALL. In the case of target-as-client, the client must know at least the server's IP address (or DNS name). That knowledge increases the amount of setup. Further, during the course of maintenance and expansion,
I disagree. For standard network boot protocols no such knowl;edge is required.
whether this is acceptable to the customer. My experience is that making the amount of information used by the client (that can be unilaterally changed by the MIS department) nil is the best approach.
The use BOOTP or DHCP.
The key to this is that my download server uses standard UDP over multicast IP. This has a number of advantages, including the fact that the download
You mention "standard UDP" - but are you using a standard download protocol?
Another advantage is that this technique can be used very early in the manufacturing process to load the software (via UDP/IP) that programs the MAC address (as well as model and serial numbers) into EEPROM. When this software is loaded, the target has no IP address or MAC address. I can even load many units in parallel on a test rack.
OK, this obviously means that you are NOT using any standard protocols. I understand that such a solution might make sense for you (although I think that error handling is not easy to implement), but IMHO there are already more than enough proprietary protocols and devices that cannot talk to each other just because the manufacturer tried to be especially clever.
As I wrote before:
U-Boot shall use standard protocols and interfaces whenever possible.
My server-loader meets all of these criteria, except the "powerful to use"
Which is the standard protocol you are using, then? [Note: UDP alone is _not_ a download protocol.]
My loader uses standard protocols as well. It uses UDP over multicast IP.
C'me on. UDP is _NOT_ a file transfer protocol.
You mentioned in a subsequent email to Brian that he would have to write a custom client. You are quite correct. My version of this client is written as a command-line tool that does the following: 0) create a standard multicast UDP socket and bind it appropriately, 1) open a file for binary read, 2) read 1k blocks from the file, 3) prefix a small payload header, and 4) send the block to the socket. Since we already had a socket library, this task took an afternoon and the result was just 5-6 pages of C code.
This is OK, and there is no reason for you not to use such a proprietary approach.
This client hasn't been maintained after it's original creation and testing (although it has been extended with new features; e.g., alternate binary file formats, etc.).
This requirement can be also met with U-Boot using widely available standard protocols (like DHCP or BOOTP).
I'm sorry, but IMHO there is no advantage running a server in the boot loader.
I'm sorry, but IMHO YHO is wrong. :-)
Well, YMMV ;-)
Ultimately, I believe the difference in the two approaches is in where the control is. I prefer to view the host as "in control" and the target as subservient. It has simply been my experience, over many projects, that this way works out better.
The only argument I want to reopeart here is that in my experience the use of open standards is a very strong benefit for everybody, and if there is a solution for a problem based on standard protocols and interfaces I do not want to use or invent some proprietary solution instead.
This whole discussion may not be appropriate on this list. Since I did this
I don't think so. It may be beneficial to many people, one way or another. To be honest, the idea of running a server on the target never even crossed my mind before. It is always a good thing to discuss alternative solutions - and if it's only to come to the conclusion that everything was done correctly :-)
work as a paid employee, I must honor my employeer's policies that prevent sharing source code. I can, however, freely discuss the details and answer questions If Brian or anyone else cares to find out more about this technique, they are welcome to contact me directly.
Thanks!
Best regards,
Wolfgang Denk
In message 20030627142446.9EC7FC592A@atlas.denx.de I wrote:
This is wrong. The only a-priori knowledge needed by the client to ruin any of the standard network boot protocols (like BOOTP or DHCP) is it's own MAC address, which should have been programmed by the manufacturer and which will never need to be changed.
Nice typo, but of course I meant "...to _run_ any of the standard network boot protocols...".
Best regards,
Wolfgang Denk
participants (2)
-
Wells, Charles -
Wolfgang Denk